X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/0c530ab8987f0ae6a1a3d9284f40182b88852816..d190cdc3f5544636abb56dc1874be391d3e1b148:/osfmk/vm/vm_object.c diff --git a/osfmk/vm/vm_object.c b/osfmk/vm/vm_object.c index 221565271..72b525fdd 100644 --- a/osfmk/vm/vm_object.c +++ b/osfmk/vm/vm_object.c @@ -1,23 +1,29 @@ /* - * Copyright (c) 2000-2004 Apple Computer, Inc. All rights reserved. + * Copyright (c) 2000-2007 Apple Inc. All rights reserved. * - * @APPLE_LICENSE_HEADER_START@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * - * The contents of this file constitute Original Code as defined in and - * are subject to the Apple Public Source License Version 1.1 (the - * "License"). You may not use this file except in compliance with the - * License. Please obtain a copy of the License at - * http://www.apple.com/publicsource and read it before using this file. + * This file contains Original Code and/or Modifications of Original Code + * as defined in and that are subject to the Apple Public Source License + * Version 2.0 (the 'License'). You may not use this file except in + * compliance with the License. The rights granted to you under the License + * may not be used to create, or enable the creation or redistribution of, + * unlawful or unlicensed copies of an Apple operating system, or to + * circumvent, violate, or enable the circumvention or violation of, any + * terms of an Apple operating system software license agreement. * - * This Original Code and all software distributed under the License are - * distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, EITHER + * Please obtain a copy of the License at + * http://www.opensource.apple.com/apsl/ and read it before using this file. + * + * The Original Code and all software distributed under the License are + * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE OR NON-INFRINGEMENT. Please see the - * License for the specific language governing rights and limitations - * under the License. + * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. + * Please see the License for the specific language governing rights and + * limitations under the License. * - * @APPLE_LICENSE_HEADER_END@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ /* * @OSF_COPYRIGHT@ @@ -56,6 +62,7 @@ * Virtual memory object module. */ +#include #include #include @@ -65,27 +72,66 @@ #include #include +#include + #include #include #include #include -#include #include #include +#include #include #include #include #include #include +#include #include +#include #include #include #include #include #include #include +#include + +#include + +#if CONFIG_PHANTOM_CACHE +#include +#endif + +boolean_t vm_object_collapse_compressor_allowed = TRUE; + +struct vm_counters vm_counters; + +#if VM_OBJECT_TRACKING +boolean_t vm_object_tracking_inited = FALSE; +btlog_t *vm_object_tracking_btlog; + +void +vm_object_tracking_init(void) +{ + int vm_object_tracking; + + vm_object_tracking = 1; + PE_parse_boot_argn("vm_object_tracking", &vm_object_tracking, + sizeof (vm_object_tracking)); + + if (vm_object_tracking) { + vm_object_tracking_btlog = btlog_create( + VM_OBJECT_TRACKING_NUM_RECORDS, + VM_OBJECT_TRACKING_BTDEPTH, + TRUE /* caller_will_remove_entries_for_element? */); + assert(vm_object_tracking_btlog); + vm_object_tracking_inited = TRUE; + } +} +#endif /* VM_OBJECT_TRACKING */ /* * Virtual memory objects maintain the actual data @@ -167,12 +213,6 @@ static kern_return_t vm_object_terminate( extern void vm_object_remove( vm_object_t object); -static vm_object_t vm_object_cache_trim( - boolean_t called_from_vm_object_deallocate); - -static void vm_object_deactivate_all_pages( - vm_object_t object); - static kern_return_t vm_object_copy_call( vm_object_t src_object, vm_object_offset_t src_offset, @@ -188,7 +228,8 @@ static void vm_object_do_bypass( vm_object_t backing_object); static void vm_object_release_pager( - memory_object_t pager); + memory_object_t pager, + boolean_t hashed); static zone_t vm_object_zone; /* vm backing store zone */ @@ -196,8 +237,11 @@ static zone_t vm_object_zone; /* vm backing store zone */ * All wired-down kernel memory belongs to a single virtual * memory object (kernel_object) to avoid wasting data structures. */ -static struct vm_object kernel_object_store; -vm_object_t kernel_object; +static struct vm_object kernel_object_store __attribute__((aligned(VM_PACKED_POINTER_ALIGNMENT))); +vm_object_t kernel_object; + +static struct vm_object compressor_object_store __attribute__((aligned(VM_PACKED_POINTER_ALIGNMENT))); +vm_object_t compressor_object = &compressor_object_store; /* * The submap object is used as a placeholder for vm_map_submap @@ -205,7 +249,7 @@ vm_object_t kernel_object; * is exported by the vm_map module. The storage is declared * here because it must be initialized here. */ -static struct vm_object vm_submap_object_store; +static struct vm_object vm_submap_object_store __attribute__((aligned(VM_PACKED_POINTER_ALIGNMENT))); /* * Virtual memory objects are initialized from @@ -217,6 +261,11 @@ static struct vm_object vm_submap_object_store; */ static struct vm_object vm_object_template; +unsigned int vm_page_purged_wired = 0; +unsigned int vm_page_purged_busy = 0; +unsigned int vm_page_purged_others = 0; + +#if VM_OBJECT_CACHE /* * Virtual memory objects that are not referenced by * any address maps, but that are allowed to persist @@ -245,23 +294,53 @@ static struct vm_object vm_object_template; * from the reference mechanism, so that the lock need * not be held to make simple references. */ -static queue_head_t vm_object_cached_list; -static int vm_object_cached_count=0; +static vm_object_t vm_object_cache_trim( + boolean_t called_from_vm_object_deallocate); + +static void vm_object_deactivate_all_pages( + vm_object_t object); + static int vm_object_cached_high; /* highest # cached objects */ static int vm_object_cached_max = 512; /* may be patched*/ -static decl_mutex_data(,vm_object_cached_lock_data) - #define vm_object_cache_lock() \ - mutex_lock(&vm_object_cached_lock_data) -#define vm_object_cache_lock_try() \ - mutex_try(&vm_object_cached_lock_data) + lck_mtx_lock(&vm_object_cached_lock_data) +#define vm_object_cache_lock_try() \ + lck_mtx_try_lock(&vm_object_cached_lock_data) + +#endif /* VM_OBJECT_CACHE */ + +static queue_head_t vm_object_cached_list; +static uint32_t vm_object_cache_pages_freed = 0; +static uint32_t vm_object_cache_pages_moved = 0; +static uint32_t vm_object_cache_pages_skipped = 0; +static uint32_t vm_object_cache_adds = 0; +static uint32_t vm_object_cached_count = 0; +static lck_mtx_t vm_object_cached_lock_data; +static lck_mtx_ext_t vm_object_cached_lock_data_ext; + +static uint32_t vm_object_page_grab_failed = 0; +static uint32_t vm_object_page_grab_skipped = 0; +static uint32_t vm_object_page_grab_returned = 0; +static uint32_t vm_object_page_grab_pmapped = 0; +static uint32_t vm_object_page_grab_reactivations = 0; + +#define vm_object_cache_lock_spin() \ + lck_mtx_lock_spin(&vm_object_cached_lock_data) #define vm_object_cache_unlock() \ - mutex_unlock(&vm_object_cached_lock_data) + lck_mtx_unlock(&vm_object_cached_lock_data) + +static void vm_object_cache_remove_locked(vm_object_t); + #define VM_OBJECT_HASH_COUNT 1024 +#define VM_OBJECT_HASH_LOCK_COUNT 512 + +static lck_mtx_t vm_object_hashed_lock_data[VM_OBJECT_HASH_LOCK_COUNT]; +static lck_mtx_ext_t vm_object_hashed_lock_data_ext[VM_OBJECT_HASH_LOCK_COUNT]; + static queue_head_t vm_object_hashtable[VM_OBJECT_HASH_COUNT]; -static struct zone *vm_object_hash_zone; +static struct zone *vm_object_hash_zone; struct vm_object_hash_entry { queue_chain_t hash_link; /* hash chain link */ @@ -274,9 +353,12 @@ struct vm_object_hash_entry { typedef struct vm_object_hash_entry *vm_object_hash_entry_t; #define VM_OBJECT_HASH_ENTRY_NULL ((vm_object_hash_entry_t) 0) -#define VM_OBJECT_HASH_SHIFT 8 +#define VM_OBJECT_HASH_SHIFT 5 #define vm_object_hash(pager) \ - ((((unsigned)pager) >> VM_OBJECT_HASH_SHIFT) % VM_OBJECT_HASH_COUNT) + ((int)((((uintptr_t)pager) >> VM_OBJECT_HASH_SHIFT) % VM_OBJECT_HASH_COUNT)) + +#define vm_object_lock_hash(pager) \ + ((int)((((uintptr_t)pager) >> VM_OBJECT_HASH_SHIFT) % VM_OBJECT_HASH_LOCK_COUNT)) void vm_object_hash_entry_free( vm_object_hash_entry_t entry); @@ -284,38 +366,101 @@ void vm_object_hash_entry_free( static void vm_object_reap(vm_object_t object); static void vm_object_reap_async(vm_object_t object); static void vm_object_reaper_thread(void); -static queue_head_t vm_object_reaper_queue; /* protected by vm_object_cache_lock() */ + +static lck_mtx_t vm_object_reaper_lock_data; +static lck_mtx_ext_t vm_object_reaper_lock_data_ext; + +static queue_head_t vm_object_reaper_queue; /* protected by vm_object_reaper_lock() */ unsigned int vm_object_reap_count = 0; unsigned int vm_object_reap_count_async = 0; +#define vm_object_reaper_lock() \ + lck_mtx_lock(&vm_object_reaper_lock_data) +#define vm_object_reaper_lock_spin() \ + lck_mtx_lock_spin(&vm_object_reaper_lock_data) +#define vm_object_reaper_unlock() \ + lck_mtx_unlock(&vm_object_reaper_lock_data) + +#if CONFIG_IOSCHED +/* I/O Re-prioritization request list */ +queue_head_t io_reprioritize_list; +lck_spin_t io_reprioritize_list_lock; + +#define IO_REPRIORITIZE_LIST_LOCK() \ + lck_spin_lock(&io_reprioritize_list_lock) +#define IO_REPRIORITIZE_LIST_UNLOCK() \ + lck_spin_unlock(&io_reprioritize_list_lock) + +#define MAX_IO_REPRIORITIZE_REQS 8192 +zone_t io_reprioritize_req_zone; + +/* I/O Re-prioritization thread */ +int io_reprioritize_wakeup = 0; +static void io_reprioritize_thread(void *param __unused, wait_result_t wr __unused); + +#define IO_REPRIO_THREAD_WAKEUP() thread_wakeup((event_t)&io_reprioritize_wakeup) +#define IO_REPRIO_THREAD_CONTINUATION() \ +{ \ + assert_wait(&io_reprioritize_wakeup, THREAD_UNINT); \ + thread_block(io_reprioritize_thread); \ +} + +void vm_page_request_reprioritize(vm_object_t, uint64_t, uint32_t, int); +void vm_page_handle_prio_inversion(vm_object_t, vm_page_t); +void vm_decmp_upl_reprioritize(upl_t, int); +#endif + +#if 0 +#undef KERNEL_DEBUG +#define KERNEL_DEBUG KERNEL_DEBUG_CONSTANT +#endif + + +static lck_mtx_t * +vm_object_hash_lock_spin( + memory_object_t pager) +{ + int index; + + index = vm_object_lock_hash(pager); + + lck_mtx_lock_spin(&vm_object_hashed_lock_data[index]); + + return (&vm_object_hashed_lock_data[index]); +} + +static void +vm_object_hash_unlock(lck_mtx_t *lck) +{ + lck_mtx_unlock(lck); +} + + /* * vm_object_hash_lookup looks up a pager in the hashtable * and returns the corresponding entry, with optional removal. */ - static vm_object_hash_entry_t vm_object_hash_lookup( memory_object_t pager, boolean_t remove_entry) { - register queue_t bucket; - register vm_object_hash_entry_t entry; + queue_t bucket; + vm_object_hash_entry_t entry; bucket = &vm_object_hashtable[vm_object_hash(pager)]; entry = (vm_object_hash_entry_t)queue_first(bucket); while (!queue_end(bucket, (queue_entry_t)entry)) { - if (entry->pager == pager && !remove_entry) - return(entry); - else if (entry->pager == pager) { - queue_remove(bucket, entry, - vm_object_hash_entry_t, hash_link); + if (entry->pager == pager) { + if (remove_entry) { + queue_remove(bucket, entry, + vm_object_hash_entry_t, hash_link); + } return(entry); } - entry = (vm_object_hash_entry_t)queue_next(&entry->hash_link); } - return(VM_OBJECT_HASH_ENTRY_NULL); } @@ -326,13 +471,31 @@ vm_object_hash_lookup( static void vm_object_hash_insert( - vm_object_hash_entry_t entry) + vm_object_hash_entry_t entry, + vm_object_t object) { - register queue_t bucket; + queue_t bucket; + + assert(vm_object_hash_lookup(entry->pager, FALSE) == NULL); bucket = &vm_object_hashtable[vm_object_hash(entry->pager)]; queue_enter(bucket, entry, vm_object_hash_entry_t, hash_link); + + if (object->hashed) { + /* + * "hashed" was pre-set on this (new) object to avoid + * locking issues in vm_object_enter() (can't attempt to + * grab the object lock while holding the hash lock as + * a spinlock), so no need to set it here (and no need to + * hold the object's lock). + */ + } else { + vm_object_lock_assert_exclusive(object); + object->hashed = TRUE; + } + + entry->object = object; } static vm_object_hash_entry_t @@ -369,23 +532,37 @@ _vm_object_allocate( { XPR(XPR_VM_OBJECT, "vm_object_allocate, object 0x%X size 0x%X\n", - (integer_t)object, size, 0,0,0); + object, size, 0,0,0); *object = vm_object_template; - queue_init(&object->memq); + vm_page_queue_init(&object->memq); queue_init(&object->msr_q); -#ifdef UPL_DEBUG +#if UPL_DEBUG || CONFIG_IOSCHED queue_init(&object->uplq); -#endif /* UPL_DEBUG */ +#endif vm_object_lock_init(object); - object->size = size; + object->vo_size = size; + +#if VM_OBJECT_TRACKING_OP_CREATED + if (vm_object_tracking_inited) { + void *bt[VM_OBJECT_TRACKING_BTDEPTH]; + int numsaved = 0; + + numsaved = OSBacktrace(bt, VM_OBJECT_TRACKING_BTDEPTH); + btlog_add_entry(vm_object_tracking_btlog, + object, + VM_OBJECT_TRACKING_OP_CREATED, + bt, + numsaved); + } +#endif /* VM_OBJECT_TRACKING_OP_CREATED */ } __private_extern__ vm_object_t vm_object_allocate( vm_object_size_t size) { - register vm_object_t object; + vm_object_t object; object = (vm_object_t) zalloc(vm_object_zone); @@ -397,6 +574,14 @@ vm_object_allocate( return object; } + +lck_grp_t vm_object_lck_grp; +lck_grp_t vm_object_cache_lck_grp; +lck_grp_attr_t vm_object_lck_grp_attr; +lck_attr_t vm_object_lck_attr; +lck_attr_t kernel_object_lck_attr; +lck_attr_t compressor_object_lck_attr; + /* * vm_object_bootstrap: * @@ -405,53 +590,95 @@ vm_object_allocate( __private_extern__ void vm_object_bootstrap(void) { - register int i; + int i; + vm_size_t vm_object_size; - vm_object_zone = zinit((vm_size_t) sizeof(struct vm_object), - round_page_32(512*1024), - round_page_32(12*1024), - "vm objects"); + vm_object_size = (sizeof(struct vm_object) + (VM_PACKED_POINTER_ALIGNMENT-1)) & ~(VM_PACKED_POINTER_ALIGNMENT - 1); - queue_init(&vm_object_reaper_queue); + vm_object_zone = zinit(vm_object_size, + round_page(512*1024), + round_page(12*1024), + "vm objects"); + zone_change(vm_object_zone, Z_CALLERACCT, FALSE); /* don't charge caller */ + zone_change(vm_object_zone, Z_NOENCRYPT, TRUE); + + vm_object_init_lck_grp(); queue_init(&vm_object_cached_list); - mutex_init(&vm_object_cached_lock_data, 0); + + lck_mtx_init_ext(&vm_object_cached_lock_data, + &vm_object_cached_lock_data_ext, + &vm_object_cache_lck_grp, + &vm_object_lck_attr); + + queue_init(&vm_object_reaper_queue); + + for (i = 0; i < VM_OBJECT_HASH_LOCK_COUNT; i++) { + lck_mtx_init_ext(&vm_object_hashed_lock_data[i], + &vm_object_hashed_lock_data_ext[i], + &vm_object_lck_grp, + &vm_object_lck_attr); + } + lck_mtx_init_ext(&vm_object_reaper_lock_data, + &vm_object_reaper_lock_data_ext, + &vm_object_lck_grp, + &vm_object_lck_attr); vm_object_hash_zone = zinit((vm_size_t) sizeof (struct vm_object_hash_entry), - round_page_32(512*1024), - round_page_32(12*1024), + round_page(512*1024), + round_page(12*1024), "vm object hash entries"); + zone_change(vm_object_hash_zone, Z_CALLERACCT, FALSE); + zone_change(vm_object_hash_zone, Z_NOENCRYPT, TRUE); for (i = 0; i < VM_OBJECT_HASH_COUNT; i++) queue_init(&vm_object_hashtable[i]); + /* * Fill in a template object, for quick initialization */ /* memq; Lock; init after allocation */ - vm_object_template.size = 0; + + + vm_object_template.memq.prev = 0; + vm_object_template.memq.next = 0; +#if 0 + /* + * We can't call vm_object_lock_init() here because that will + * allocate some memory and VM is not fully initialized yet. + * The lock will be initialized for each allocated object in + * _vm_object_allocate(), so we don't need to initialize it in + * the vm_object_template. + */ + vm_object_lock_init(&vm_object_template); +#endif +#if DEVELOPMENT || DEBUG + vm_object_template.Lock_owner = 0; +#endif + vm_object_template.vo_size = 0; vm_object_template.memq_hint = VM_PAGE_NULL; vm_object_template.ref_count = 1; #if TASK_SWAPPER vm_object_template.res_count = 1; #endif /* TASK_SWAPPER */ vm_object_template.resident_page_count = 0; + vm_object_template.wired_page_count = 0; + vm_object_template.reusable_page_count = 0; vm_object_template.copy = VM_OBJECT_NULL; vm_object_template.shadow = VM_OBJECT_NULL; - vm_object_template.shadow_offset = (vm_object_offset_t) 0; - vm_object_template.cow_hint = ~(vm_offset_t)0; - vm_object_template.true_share = FALSE; - + vm_object_template.vo_shadow_offset = (vm_object_offset_t) 0; vm_object_template.pager = MEMORY_OBJECT_NULL; vm_object_template.paging_offset = 0; vm_object_template.pager_control = MEMORY_OBJECT_CONTROL_NULL; - /* msr_q; init after allocation */ - vm_object_template.copy_strategy = MEMORY_OBJECT_COPY_SYMMETRIC; - vm_object_template.absent_count = 0; vm_object_template.paging_in_progress = 0; +#if __LP64__ + vm_object_template.__object1_unused_bits = 0; +#endif /* __LP64__ */ + vm_object_template.activity_in_progress = 0; /* Begin bitfields */ vm_object_template.all_wanted = 0; /* all bits FALSE */ @@ -465,29 +692,86 @@ vm_object_bootstrap(void) vm_object_template.private = FALSE; vm_object_template.pageout = FALSE; vm_object_template.alive = TRUE; - vm_object_template.purgable = VM_OBJECT_NONPURGABLE; - vm_object_template.silent_overwrite = FALSE; - vm_object_template.advisory_pageout = FALSE; + vm_object_template.purgable = VM_PURGABLE_DENY; + vm_object_template.purgeable_when_ripe = FALSE; vm_object_template.shadowed = FALSE; + vm_object_template.advisory_pageout = FALSE; + vm_object_template.true_share = FALSE; vm_object_template.terminating = FALSE; + vm_object_template.named = FALSE; vm_object_template.shadow_severed = FALSE; vm_object_template.phys_contiguous = FALSE; vm_object_template.nophyscache = FALSE; /* End bitfields */ - /* cache bitfields */ - vm_object_template.wimg_bits = VM_WIMG_DEFAULT; - - /* cached_list; init after allocation */ + vm_object_template.cached_list.prev = NULL; + vm_object_template.cached_list.next = NULL; + vm_object_template.msr_q.prev = NULL; + vm_object_template.msr_q.next = NULL; + vm_object_template.last_alloc = (vm_object_offset_t) 0; - vm_object_template.cluster_size = 0; -#if MACH_PAGEMAP - vm_object_template.existence_map = VM_EXTERNAL_NULL; -#endif /* MACH_PAGEMAP */ + vm_object_template.sequential = (vm_object_offset_t) 0; + vm_object_template.pages_created = 0; + vm_object_template.pages_used = 0; + vm_object_template.scan_collisions = 0; +#if CONFIG_PHANTOM_CACHE + vm_object_template.phantom_object_id = 0; +#endif + vm_object_template.cow_hint = ~(vm_offset_t)0; #if MACH_ASSERT vm_object_template.paging_object = VM_OBJECT_NULL; #endif /* MACH_ASSERT */ + /* cache bitfields */ + vm_object_template.wimg_bits = VM_WIMG_USE_DEFAULT; + vm_object_template.set_cache_attr = FALSE; + vm_object_template.object_slid = FALSE; + vm_object_template.code_signed = FALSE; + vm_object_template.hashed = FALSE; + vm_object_template.transposed = FALSE; + vm_object_template.mapping_in_progress = FALSE; + vm_object_template.phantom_isssd = FALSE; + vm_object_template.volatile_empty = FALSE; + vm_object_template.volatile_fault = FALSE; + vm_object_template.all_reusable = FALSE; + vm_object_template.blocked_access = FALSE; + vm_object_template.__object2_unused_bits = 0; +#if CONFIG_IOSCHED || UPL_DEBUG + vm_object_template.uplq.prev = NULL; + vm_object_template.uplq.next = NULL; +#endif /* UPL_DEBUG */ +#ifdef VM_PIP_DEBUG + bzero(&vm_object_template.pip_holders, + sizeof (vm_object_template.pip_holders)); +#endif /* VM_PIP_DEBUG */ + + vm_object_template.objq.next = NULL; + vm_object_template.objq.prev = NULL; + + vm_object_template.purgeable_queue_type = PURGEABLE_Q_TYPE_MAX; + vm_object_template.purgeable_queue_group = 0; + + vm_object_template.vo_cache_ts = 0; + + vm_object_template.wire_tag = VM_KERN_MEMORY_NONE; + + vm_object_template.io_tracking = FALSE; + +#if CONFIG_SECLUDED_MEMORY + vm_object_template.eligible_for_secluded = FALSE; + vm_object_template.can_grab_secluded = FALSE; +#else /* CONFIG_SECLUDED_MEMORY */ + vm_object_template.__object3_unused_bits = 0; +#endif /* CONFIG_SECLUDED_MEMORY */ + +#if DEBUG + bzero(&vm_object_template.purgeable_owner_bt[0], + sizeof (vm_object_template.purgeable_owner_bt)); + vm_object_template.vo_purgeable_volatilizer = NULL; + bzero(&vm_object_template.purgeable_volatilizer_bt[0], + sizeof (vm_object_template.purgeable_volatilizer_bt)); +#endif /* DEBUG */ + /* * Initialize the "kernel object" */ @@ -499,14 +783,13 @@ vm_object_bootstrap(void) * VM_MAX_KERNEL_ADDRESS (vm_last_addr) is a maximum address, not a size. */ -#ifdef ppc - _vm_object_allocate((vm_last_addr - VM_MIN_KERNEL_ADDRESS) + 1, - kernel_object); -#else - _vm_object_allocate((VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) + 1, - kernel_object); -#endif + _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1, + kernel_object); + + _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1, + compressor_object); kernel_object->copy_strategy = MEMORY_OBJECT_COPY_NONE; + compressor_object->copy_strategy = MEMORY_OBJECT_COPY_NONE; /* * Initialize the "submap object". Make it as large as the @@ -514,13 +797,8 @@ vm_object_bootstrap(void) */ vm_submap_object = &vm_submap_object_store; -#ifdef ppc - _vm_object_allocate((vm_last_addr - VM_MIN_KERNEL_ADDRESS) + 1, - vm_submap_object); -#else - _vm_object_allocate((VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) + 1, - vm_submap_object); -#endif + _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1, + vm_submap_object); vm_submap_object->copy_strategy = MEMORY_OBJECT_COPY_NONE; /* @@ -529,11 +807,31 @@ vm_object_bootstrap(void) * non-zone memory. */ vm_object_reference(vm_submap_object); +} -#if MACH_PAGEMAP - vm_external_module_initialize(); -#endif /* MACH_PAGEMAP */ +#if CONFIG_IOSCHED +void +vm_io_reprioritize_init(void) +{ + kern_return_t result; + thread_t thread = THREAD_NULL; + + /* Initialze the I/O reprioritization subsystem */ + lck_spin_init(&io_reprioritize_list_lock, &vm_object_lck_grp, &vm_object_lck_attr); + queue_init(&io_reprioritize_list); + + io_reprioritize_req_zone = zinit(sizeof(struct io_reprioritize_req), + MAX_IO_REPRIORITIZE_REQS * sizeof(struct io_reprioritize_req), + 4096, "io_reprioritize_req"); + + result = kernel_thread_start_priority(io_reprioritize_thread, NULL, 95 /* MAXPRI_KERNEL */, &thread); + if (result == KERN_SUCCESS) { + thread_deallocate(thread); + } else { + panic("Could not create io_reprioritize_thread"); + } } +#endif void vm_object_reaper_init(void) @@ -547,7 +845,7 @@ vm_object_reaper_init(void) BASEPRI_PREEMPT - 1, &thread); if (kr != KERN_SUCCESS) { - panic("failed to launch vm_object_reaper_thread kr=0x%x\n", kr); + panic("failed to launch vm_object_reaper_thread kr=0x%x", kr); } thread_deallocate(thread); } @@ -560,18 +858,29 @@ vm_object_init(void) */ } -/* remove the typedef below when emergency work-around is taken out */ -typedef struct vnode_pager { - memory_object_t pager; - memory_object_t pager_handle; /* pager */ - memory_object_control_t control_handle; /* memory object's control handle */ - void *vnode_handle; /* vnode handle */ -} *vnode_pager_t; +__private_extern__ void +vm_object_init_lck_grp(void) +{ + /* + * initialze the vm_object lock world + */ + lck_grp_attr_setdefault(&vm_object_lck_grp_attr); + lck_grp_init(&vm_object_lck_grp, "vm_object", &vm_object_lck_grp_attr); + lck_grp_init(&vm_object_cache_lck_grp, "vm_object_cache", &vm_object_lck_grp_attr); + lck_attr_setdefault(&vm_object_lck_attr); + lck_attr_setdefault(&kernel_object_lck_attr); + lck_attr_cleardebug(&kernel_object_lck_attr); + lck_attr_setdefault(&compressor_object_lck_attr); + lck_attr_cleardebug(&compressor_object_lck_attr); +} + +#if VM_OBJECT_CACHE #define MIGHT_NOT_CACHE_SHADOWS 1 #if MIGHT_NOT_CACHE_SHADOWS static int cache_shadows = TRUE; #endif /* MIGHT_NOT_CACHE_SHADOWS */ +#endif /* * vm_object_deallocate: @@ -584,39 +893,124 @@ static int cache_shadows = TRUE; * * No object may be locked. */ +unsigned long vm_object_deallocate_shared_successes = 0; +unsigned long vm_object_deallocate_shared_failures = 0; +unsigned long vm_object_deallocate_shared_swap_failures = 0; + __private_extern__ void vm_object_deallocate( - register vm_object_t object) + vm_object_t object) { - boolean_t retry_cache_trim = FALSE; - vm_object_t shadow = VM_OBJECT_NULL; +#if VM_OBJECT_CACHE + boolean_t retry_cache_trim = FALSE; + uint32_t try_failed_count = 0; +#endif + vm_object_t shadow = VM_OBJECT_NULL; // if(object)dbgLog(object, object->ref_count, object->can_persist, 3); /* (TEST/DEBUG) */ // else dbgLog(object, 0, 0, 3); /* (TEST/DEBUG) */ + if (object == VM_OBJECT_NULL) + return; - while (object != VM_OBJECT_NULL) { + if (object == kernel_object || object == compressor_object) { + vm_object_lock_shared(object); + + OSAddAtomic(-1, &object->ref_count); + + if (object->ref_count == 0) { + if (object == kernel_object) + panic("vm_object_deallocate: losing kernel_object\n"); + else + panic("vm_object_deallocate: losing compressor_object\n"); + } + vm_object_unlock(object); + return; + } + + if (object->ref_count == 2 && + object->named) { + /* + * This "named" object's reference count is about to + * drop from 2 to 1: + * we'll need to call memory_object_last_unmap(). + */ + } else if (object->ref_count == 2 && + object->internal && + object->shadow != VM_OBJECT_NULL) { + /* + * This internal object's reference count is about to + * drop from 2 to 1 and it has a shadow object: + * we'll want to try and collapse this object with its + * shadow. + */ + } else if (object->ref_count >= 2) { + UInt32 original_ref_count; + volatile UInt32 *ref_count_p; + Boolean atomic_swap; /* - * The cache holds a reference (uncounted) to - * the object; we must lock it before removing - * the object. + * The object currently looks like it is not being + * kept alive solely by the reference we're about to release. + * Let's try and release our reference without taking + * all the locks we would need if we had to terminate the + * object (cache lock + exclusive object lock). + * Lock the object "shared" to make sure we don't race with + * anyone holding it "exclusive". + */ + vm_object_lock_shared(object); + ref_count_p = (volatile UInt32 *) &object->ref_count; + original_ref_count = object->ref_count; + /* + * Test again as "ref_count" could have changed. + * "named" shouldn't change. */ - for (;;) { - vm_object_cache_lock(); + if (original_ref_count == 2 && + object->named) { + /* need to take slow path for m_o_last_unmap() */ + atomic_swap = FALSE; + } else if (original_ref_count == 2 && + object->internal && + object->shadow != VM_OBJECT_NULL) { + /* need to take slow path for vm_object_collapse() */ + atomic_swap = FALSE; + } else if (original_ref_count < 2) { + /* need to take slow path for vm_object_terminate() */ + atomic_swap = FALSE; + } else { + /* try an atomic update with the shared lock */ + atomic_swap = OSCompareAndSwap( + original_ref_count, + original_ref_count - 1, + (UInt32 *) &object->ref_count); + if (atomic_swap == FALSE) { + vm_object_deallocate_shared_swap_failures++; + /* fall back to the slow path... */ + } + } + + vm_object_unlock(object); + if (atomic_swap) { /* - * if we try to take a regular lock here - * we risk deadlocking against someone - * holding a lock on this object while - * trying to vm_object_deallocate a different - * object + * ref_count was updated atomically ! */ - if (vm_object_lock_try(object)) - break; - vm_object_cache_unlock(); - mutex_pause(); /* wait a bit */ + vm_object_deallocate_shared_successes++; + return; } + + /* + * Someone else updated the ref_count at the same + * time and we lost the race. Fall back to the usual + * slow but safe path... + */ + vm_object_deallocate_shared_failures++; + } + + while (object != VM_OBJECT_NULL) { + + vm_object_lock(object); + assert(object->ref_count > 0); /* @@ -631,28 +1025,16 @@ vm_object_deallocate( /* more mappers for this object */ if (pager != MEMORY_OBJECT_NULL) { + vm_object_mapping_wait(object, THREAD_UNINT); + vm_object_mapping_begin(object); vm_object_unlock(object); - vm_object_cache_unlock(); - - memory_object_unmap(pager); - for (;;) { - vm_object_cache_lock(); + memory_object_last_unmap(pager); - /* - * if we try to take a regular lock here - * we risk deadlocking against someone - * holding a lock on this object while - * trying to vm_object_deallocate a different - * object - */ - if (vm_object_lock_try(object)) - break; - vm_object_cache_unlock(); - mutex_pause(); /* wait a bit */ - } - assert(object->ref_count > 0); + vm_object_lock(object); + vm_object_mapping_end(object); } + assert(object->ref_count > 0); } /* @@ -668,9 +1050,9 @@ vm_object_deallocate( /* terminate again. */ if ((object->ref_count > 1) || object->terminating) { + vm_object_lock_assert_exclusive(object); object->ref_count--; vm_object_res_deallocate(object); - vm_object_cache_unlock(); if (object->ref_count == 1 && object->shadow != VM_OBJECT_NULL) { @@ -692,13 +1074,14 @@ vm_object_deallocate( */ vm_object_collapse(object, 0, FALSE); } - vm_object_unlock(object); +#if VM_OBJECT_CACHE if (retry_cache_trim && ((object = vm_object_cache_trim(TRUE)) != VM_OBJECT_NULL)) { continue; } +#endif return; } @@ -713,11 +1096,12 @@ vm_object_deallocate( VM_OBJECT_EVENT_INITIALIZED, THREAD_UNINT); vm_object_unlock(object); - vm_object_cache_unlock(); + thread_block(THREAD_CONTINUE_NULL); continue; } +#if VM_OBJECT_CACHE /* * If this object can persist, then enter it in * the cache. Otherwise, terminate it. @@ -733,10 +1117,12 @@ vm_object_deallocate( * Now it is safe to decrement reference count, * and to return if reference count is > 0. */ + + vm_object_lock_assert_exclusive(object); if (--object->ref_count > 0) { vm_object_res_deallocate(object); vm_object_unlock(object); - vm_object_cache_unlock(); + if (retry_cache_trim && ((object = vm_object_cache_trim(TRUE)) != VM_OBJECT_NULL)) { @@ -765,16 +1151,38 @@ vm_object_deallocate( VM_OBJ_RES_DECR(object); XPR(XPR_VM_OBJECT, "vm_o_deallocate: adding %x to cache, queue = (%x, %x)\n", - (integer_t)object, - (integer_t)vm_object_cached_list.next, - (integer_t)vm_object_cached_list.prev,0,0); + object, + vm_object_cached_list.next, + vm_object_cached_list.prev,0,0); + + + vm_object_unlock(object); + + try_failed_count = 0; + for (;;) { + vm_object_cache_lock(); + /* + * if we try to take a regular lock here + * we risk deadlocking against someone + * holding a lock on this object while + * trying to vm_object_deallocate a different + * object + */ + if (vm_object_lock_try(object)) + break; + vm_object_cache_unlock(); + try_failed_count++; + + mutex_pause(try_failed_count); /* wait a bit */ + } vm_object_cached_count++; if (vm_object_cached_count > vm_object_cached_high) vm_object_cached_high = vm_object_cached_count; queue_enter(&vm_object_cached_list, object, vm_object_t, cached_list); vm_object_cache_unlock(); + vm_object_deactivate_all_pages(object); vm_object_unlock(object); @@ -804,14 +1212,15 @@ vm_object_deallocate( return; } retry_cache_trim = TRUE; - - } else { + } else +#endif /* VM_OBJECT_CACHE */ + { /* * This object is not cachable; terminate it. */ XPR(XPR_VM_OBJECT, "vm_o_deallocate: !cacheable 0x%X res %d paging_ops %d thread 0x%p ref %d\n", - (integer_t)object, object->resident_page_count, + object, object->resident_page_count, object->paging_in_progress, (void *)current_thread(),object->ref_count); @@ -825,84 +1234,467 @@ vm_object_deallocate( * a normal reference. */ shadow = object->pageout?VM_OBJECT_NULL:object->shadow; - if(vm_object_terminate(object) != KERN_SUCCESS) { + + if (vm_object_terminate(object) != KERN_SUCCESS) { return; } if (shadow != VM_OBJECT_NULL) { object = shadow; continue; } +#if VM_OBJECT_CACHE if (retry_cache_trim && ((object = vm_object_cache_trim(TRUE)) != VM_OBJECT_NULL)) { continue; } +#endif return; } } +#if VM_OBJECT_CACHE assert(! retry_cache_trim); +#endif } -/* - * Check to see whether we really need to trim - * down the cache. If so, remove an object from - * the cache, terminate it, and repeat. - * - * Called with, and returns with, cache lock unlocked. - */ -vm_object_t -vm_object_cache_trim( - boolean_t called_from_vm_object_deallocate) + + +vm_page_t +vm_object_page_grab( + vm_object_t object) { - register vm_object_t object = VM_OBJECT_NULL; - vm_object_t shadow; + vm_page_t p, next_p; + int p_limit = 0; + int p_skipped = 0; - for (;;) { + vm_object_lock_assert_exclusive(object); - /* - * If we no longer need to trim the cache, - * then we are done. - */ + next_p = (vm_page_t)vm_page_queue_first(&object->memq); + p_limit = MIN(50, object->resident_page_count); - vm_object_cache_lock(); - if (vm_object_cached_count <= vm_object_cached_max) { - vm_object_cache_unlock(); - return VM_OBJECT_NULL; - } + while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t)next_p) && --p_limit > 0) { - /* - * We must trim down the cache, so remove - * the first object in the cache. - */ - XPR(XPR_VM_OBJECT, - "vm_object_cache_trim: removing from front of cache (%x, %x)\n", - (integer_t)vm_object_cached_list.next, - (integer_t)vm_object_cached_list.prev, 0, 0, 0); + p = next_p; + next_p = (vm_page_t)vm_page_queue_next(&next_p->listq); - object = (vm_object_t) queue_first(&vm_object_cached_list); - if(object == (vm_object_t) &vm_object_cached_list) { - /* something's wrong with the calling parameter or */ - /* the value of vm_object_cached_count, just fix */ - /* and return */ - if(vm_object_cached_max < 0) - vm_object_cached_max = 0; - vm_object_cached_count = 0; - vm_object_cache_unlock(); - return VM_OBJECT_NULL; - } - vm_object_lock(object); - queue_remove(&vm_object_cached_list, object, vm_object_t, - cached_list); - vm_object_cached_count--; + if (VM_PAGE_WIRED(p) || p->busy || p->cleaning || p->laundry || p->fictitious) + goto move_page_in_obj; - /* - * Since this object is in the cache, we know - * that it is initialized and has no references. - * Take a reference to avoid recursive deallocations. - */ + if (p->pmapped || p->dirty || p->precious) { + vm_page_lockspin_queues(); - assert(object->pager_initialized); + if (p->pmapped) { + int refmod_state; + + vm_object_page_grab_pmapped++; + + if (p->reference == FALSE || p->dirty == FALSE) { + + refmod_state = pmap_get_refmod(VM_PAGE_GET_PHYS_PAGE(p)); + + if (refmod_state & VM_MEM_REFERENCED) + p->reference = TRUE; + if (refmod_state & VM_MEM_MODIFIED) { + SET_PAGE_DIRTY(p, FALSE); + } + } + if (p->dirty == FALSE && p->precious == FALSE) { + + refmod_state = pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(p)); + + if (refmod_state & VM_MEM_REFERENCED) + p->reference = TRUE; + if (refmod_state & VM_MEM_MODIFIED) { + SET_PAGE_DIRTY(p, FALSE); + } + + if (p->dirty == FALSE) + goto take_page; + } + } + if ((p->vm_page_q_state != VM_PAGE_ON_ACTIVE_Q) && p->reference == TRUE) { + vm_page_activate(p); + + VM_STAT_INCR(reactivations); + vm_object_page_grab_reactivations++; + } + vm_page_unlock_queues(); +move_page_in_obj: + vm_page_queue_remove(&object->memq, p, vm_page_t, listq); + vm_page_queue_enter(&object->memq, p, vm_page_t, listq); + + p_skipped++; + continue; + } + vm_page_lockspin_queues(); +take_page: + vm_page_free_prepare_queues(p); + vm_object_page_grab_returned++; + vm_object_page_grab_skipped += p_skipped; + + vm_page_unlock_queues(); + + vm_page_free_prepare_object(p, TRUE); + + return (p); + } + vm_object_page_grab_skipped += p_skipped; + vm_object_page_grab_failed++; + + return (NULL); +} + + + +#define EVICT_PREPARE_LIMIT 64 +#define EVICT_AGE 10 + +static clock_sec_t vm_object_cache_aging_ts = 0; + +static void +vm_object_cache_remove_locked( + vm_object_t object) +{ + assert(object->purgable == VM_PURGABLE_DENY); + assert(object->wired_page_count == 0); + + queue_remove(&vm_object_cached_list, object, vm_object_t, objq); + object->objq.next = NULL; + object->objq.prev = NULL; + + vm_object_cached_count--; +} + +void +vm_object_cache_remove( + vm_object_t object) +{ + vm_object_cache_lock_spin(); + + if (object->objq.next || object->objq.prev) + vm_object_cache_remove_locked(object); + + vm_object_cache_unlock(); +} + +void +vm_object_cache_add( + vm_object_t object) +{ + clock_sec_t sec; + clock_nsec_t nsec; + + assert(object->purgable == VM_PURGABLE_DENY); + assert(object->wired_page_count == 0); + + if (object->resident_page_count == 0) + return; + clock_get_system_nanotime(&sec, &nsec); + + vm_object_cache_lock_spin(); + + if (object->objq.next == NULL && object->objq.prev == NULL) { + queue_enter(&vm_object_cached_list, object, vm_object_t, objq); + object->vo_cache_ts = sec + EVICT_AGE; + object->vo_cache_pages_to_scan = object->resident_page_count; + + vm_object_cached_count++; + vm_object_cache_adds++; + } + vm_object_cache_unlock(); +} + +int +vm_object_cache_evict( + int num_to_evict, + int max_objects_to_examine) +{ + vm_object_t object = VM_OBJECT_NULL; + vm_object_t next_obj = VM_OBJECT_NULL; + vm_page_t local_free_q = VM_PAGE_NULL; + vm_page_t p; + vm_page_t next_p; + int object_cnt = 0; + vm_page_t ep_array[EVICT_PREPARE_LIMIT]; + int ep_count; + int ep_limit; + int ep_index; + int ep_freed = 0; + int ep_moved = 0; + uint32_t ep_skipped = 0; + clock_sec_t sec; + clock_nsec_t nsec; + + KERNEL_DEBUG(0x13001ec | DBG_FUNC_START, 0, 0, 0, 0, 0); + /* + * do a couple of quick checks to see if it's + * worthwhile grabbing the lock + */ + if (queue_empty(&vm_object_cached_list)) { + KERNEL_DEBUG(0x13001ec | DBG_FUNC_END, 0, 0, 0, 0, 0); + return (0); + } + clock_get_system_nanotime(&sec, &nsec); + + /* + * the object on the head of the queue has not + * yet sufficiently aged + */ + if (sec < vm_object_cache_aging_ts) { + KERNEL_DEBUG(0x13001ec | DBG_FUNC_END, 0, 0, 0, 0, 0); + return (0); + } + /* + * don't need the queue lock to find + * and lock an object on the cached list + */ + vm_page_unlock_queues(); + + vm_object_cache_lock_spin(); + + for (;;) { + next_obj = (vm_object_t)queue_first(&vm_object_cached_list); + + while (!queue_end(&vm_object_cached_list, (queue_entry_t)next_obj) && object_cnt++ < max_objects_to_examine) { + + object = next_obj; + next_obj = (vm_object_t)queue_next(&next_obj->objq); + + assert(object->purgable == VM_PURGABLE_DENY); + assert(object->wired_page_count == 0); + + if (sec < object->vo_cache_ts) { + KERNEL_DEBUG(0x130020c, object, object->resident_page_count, object->vo_cache_ts, sec, 0); + + vm_object_cache_aging_ts = object->vo_cache_ts; + object = VM_OBJECT_NULL; + break; + } + if (!vm_object_lock_try_scan(object)) { + /* + * just skip over this guy for now... if we find + * an object to steal pages from, we'll revist in a bit... + * hopefully, the lock will have cleared + */ + KERNEL_DEBUG(0x13001f8, object, object->resident_page_count, 0, 0, 0); + + object = VM_OBJECT_NULL; + continue; + } + if (vm_page_queue_empty(&object->memq) || object->vo_cache_pages_to_scan == 0) { + /* + * this case really shouldn't happen, but it's not fatal + * so deal with it... if we don't remove the object from + * the list, we'll never move past it. + */ + KERNEL_DEBUG(0x13001fc, object, object->resident_page_count, ep_freed, ep_moved, 0); + + vm_object_cache_remove_locked(object); + vm_object_unlock(object); + object = VM_OBJECT_NULL; + continue; + } + /* + * we have a locked object with pages... + * time to start harvesting + */ + break; + } + vm_object_cache_unlock(); + + if (object == VM_OBJECT_NULL) + break; + + /* + * object is locked at this point and + * has resident pages + */ + next_p = (vm_page_t)vm_page_queue_first(&object->memq); + + /* + * break the page scan into 2 pieces to minimize the time spent + * behind the page queue lock... + * the list of pages on these unused objects is likely to be cold + * w/r to the cpu cache which increases the time to scan the list + * tenfold... and we may have a 'run' of pages we can't utilize that + * needs to be skipped over... + */ + if ((ep_limit = num_to_evict - (ep_freed + ep_moved)) > EVICT_PREPARE_LIMIT) + ep_limit = EVICT_PREPARE_LIMIT; + ep_count = 0; + + while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t)next_p) && object->vo_cache_pages_to_scan && ep_count < ep_limit) { + + p = next_p; + next_p = (vm_page_t)vm_page_queue_next(&next_p->listq); + + object->vo_cache_pages_to_scan--; + + if (VM_PAGE_WIRED(p) || p->busy || p->cleaning || p->laundry) { + vm_page_queue_remove(&object->memq, p, vm_page_t, listq); + vm_page_queue_enter(&object->memq, p, vm_page_t, listq); + + ep_skipped++; + continue; + } + if (p->wpmapped || p->dirty || p->precious) { + vm_page_queue_remove(&object->memq, p, vm_page_t, listq); + vm_page_queue_enter(&object->memq, p, vm_page_t, listq); + + pmap_clear_reference(VM_PAGE_GET_PHYS_PAGE(p)); + } + ep_array[ep_count++] = p; + } + KERNEL_DEBUG(0x13001f4 | DBG_FUNC_START, object, object->resident_page_count, ep_freed, ep_moved, 0); + + vm_page_lockspin_queues(); + + for (ep_index = 0; ep_index < ep_count; ep_index++) { + + p = ep_array[ep_index]; + + if (p->wpmapped || p->dirty || p->precious) { + p->reference = FALSE; + p->no_cache = FALSE; + + /* + * we've already filtered out pages that are in the laundry + * so if we get here, this page can't be on the pageout queue + */ + vm_page_queues_remove(p, FALSE); + vm_page_enqueue_inactive(p, TRUE); + + ep_moved++; + } else { +#if CONFIG_PHANTOM_CACHE + vm_phantom_cache_add_ghost(p); +#endif + vm_page_free_prepare_queues(p); + + assert(p->pageq.next == 0 && p->pageq.prev == 0); + /* + * Add this page to our list of reclaimed pages, + * to be freed later. + */ + p->snext = local_free_q; + local_free_q = p; + + ep_freed++; + } + } + vm_page_unlock_queues(); + + KERNEL_DEBUG(0x13001f4 | DBG_FUNC_END, object, object->resident_page_count, ep_freed, ep_moved, 0); + + if (local_free_q) { + vm_page_free_list(local_free_q, TRUE); + local_free_q = VM_PAGE_NULL; + } + if (object->vo_cache_pages_to_scan == 0) { + KERNEL_DEBUG(0x1300208, object, object->resident_page_count, ep_freed, ep_moved, 0); + + vm_object_cache_remove(object); + + KERNEL_DEBUG(0x13001fc, object, object->resident_page_count, ep_freed, ep_moved, 0); + } + /* + * done with this object + */ + vm_object_unlock(object); + object = VM_OBJECT_NULL; + + /* + * at this point, we are not holding any locks + */ + if ((ep_freed + ep_moved) >= num_to_evict) { + /* + * we've reached our target for the + * number of pages to evict + */ + break; + } + vm_object_cache_lock_spin(); + } + /* + * put the page queues lock back to the caller's + * idea of it + */ + vm_page_lock_queues(); + + vm_object_cache_pages_freed += ep_freed; + vm_object_cache_pages_moved += ep_moved; + vm_object_cache_pages_skipped += ep_skipped; + + KERNEL_DEBUG(0x13001ec | DBG_FUNC_END, ep_freed, 0, 0, 0, 0); + return (ep_freed); +} + + +#if VM_OBJECT_CACHE +/* + * Check to see whether we really need to trim + * down the cache. If so, remove an object from + * the cache, terminate it, and repeat. + * + * Called with, and returns with, cache lock unlocked. + */ +vm_object_t +vm_object_cache_trim( + boolean_t called_from_vm_object_deallocate) +{ + vm_object_t object = VM_OBJECT_NULL; + vm_object_t shadow; + + for (;;) { + + /* + * If we no longer need to trim the cache, + * then we are done. + */ + if (vm_object_cached_count <= vm_object_cached_max) + return VM_OBJECT_NULL; + + vm_object_cache_lock(); + if (vm_object_cached_count <= vm_object_cached_max) { + vm_object_cache_unlock(); + return VM_OBJECT_NULL; + } + + /* + * We must trim down the cache, so remove + * the first object in the cache. + */ + XPR(XPR_VM_OBJECT, + "vm_object_cache_trim: removing from front of cache (%x, %x)\n", + vm_object_cached_list.next, + vm_object_cached_list.prev, 0, 0, 0); + + object = (vm_object_t) queue_first(&vm_object_cached_list); + if(object == (vm_object_t) &vm_object_cached_list) { + /* something's wrong with the calling parameter or */ + /* the value of vm_object_cached_count, just fix */ + /* and return */ + if(vm_object_cached_max < 0) + vm_object_cached_max = 0; + vm_object_cached_count = 0; + vm_object_cache_unlock(); + return VM_OBJECT_NULL; + } + vm_object_lock(object); + queue_remove(&vm_object_cached_list, object, vm_object_t, + cached_list); + vm_object_cached_count--; + + vm_object_cache_unlock(); + /* + * Since this object is in the cache, we know + * that it is initialized and has no references. + * Take a reference to avoid recursive deallocations. + */ + + assert(object->pager_initialized); assert(object->ref_count == 0); + vm_object_lock_assert_exclusive(object); object->ref_count++; /* @@ -914,8 +1706,10 @@ vm_object_cache_trim( * (We are careful here to limit recursion.) */ shadow = object->pageout?VM_OBJECT_NULL:object->shadow; + if(vm_object_terminate(object) != KERN_SUCCESS) continue; + if (shadow != VM_OBJECT_NULL) { if (called_from_vm_object_deallocate) { return shadow; @@ -925,8 +1719,8 @@ vm_object_cache_trim( } } } +#endif -boolean_t vm_object_terminate_remove_all = FALSE; /* * Routine: vm_object_terminate @@ -947,18 +1741,17 @@ boolean_t vm_object_terminate_remove_all = FALSE; */ static kern_return_t vm_object_terminate( - register vm_object_t object) + vm_object_t object) { - register vm_page_t p; - vm_object_t shadow_object; + vm_object_t shadow_object; XPR(XPR_VM_OBJECT, "vm_object_terminate, object 0x%X ref %d\n", - (integer_t)object, object->ref_count, 0, 0, 0); + object, object->ref_count, 0, 0, 0); - if (!object->pageout && (!object->temporary || object->can_persist) - && (object->pager != NULL || object->shadow_severed)) { - vm_object_cache_unlock(); - while (!queue_empty(&object->memq)) { + vm_object_lock_assert_exclusive(object); + + if (!object->pageout && (!object->temporary || object->can_persist) && + (object->pager != NULL || object->shadow_severed)) { /* * Clear pager_trusted bit so that the pages get yanked * out of the object instead of cleaned in place. This @@ -966,66 +1759,15 @@ vm_object_terminate( */ object->pager_trusted = FALSE; - p = (vm_page_t) queue_first(&object->memq); - - VM_PAGE_CHECK(p); - - if (p->busy || p->cleaning) { - if(p->cleaning || p->absent) { - vm_object_paging_wait(object, THREAD_UNINT); - continue; - } else { - panic("vm_object_terminate.3 0x%x 0x%x", object, p); - } - } - - vm_page_lock_queues(); - p->busy = TRUE; - VM_PAGE_QUEUES_REMOVE(p); - vm_page_unlock_queues(); - - if (p->absent || p->private) { - - /* - * For private pages, VM_PAGE_FREE just - * leaves the page structure around for - * its owner to clean up. For absent - * pages, the structure is returned to - * the appropriate pool. - */ - - goto free_page; - } - - if (p->fictitious) - panic("vm_object_terminate.4 0x%x 0x%x", object, p); - - if (!p->dirty) - p->dirty = pmap_is_modified(p->phys_page); - - if ((p->dirty || p->precious) && !p->error && object->alive) { - vm_pageout_cluster(p); /* flush page */ - vm_object_paging_wait(object, THREAD_UNINT); - XPR(XPR_VM_OBJECT, - "vm_object_terminate restart, object 0x%X ref %d\n", - (integer_t)object, object->ref_count, 0, 0, 0); - } else { - free_page: - VM_PAGE_FREE(p); - } - } - vm_object_unlock(object); - vm_object_cache_lock(); - vm_object_lock(object); + vm_object_reap_pages(object, REAP_TERMINATE); } - /* * Make sure the object isn't already being terminated */ - if(object->terminating) { - object->ref_count -= 1; + if (object->terminating) { + vm_object_lock_assert_exclusive(object); + object->ref_count--; assert(object->ref_count > 0); - vm_object_cache_unlock(); vm_object_unlock(object); return KERN_FAILURE; } @@ -1034,11 +1776,11 @@ vm_object_terminate( * Did somebody get a reference to the object while we were * cleaning it? */ - if(object->ref_count != 1) { - object->ref_count -= 1; + if (object->ref_count != 1) { + vm_object_lock_assert_exclusive(object); + object->ref_count--; assert(object->ref_count > 0); vm_object_res_deallocate(object); - vm_object_cache_unlock(); vm_object_unlock(object); return KERN_FAILURE; } @@ -1049,8 +1791,17 @@ vm_object_terminate( object->terminating = TRUE; object->alive = FALSE; - vm_object_remove(object); + if ( !object->internal && (object->objq.next || object->objq.prev)) + vm_object_cache_remove(object); + + if (object->hashed) { + lck_mtx_t *lck; + + lck = vm_object_hash_lock_spin(object->pager); + vm_object_remove(object); + vm_object_hash_unlock(lck); + } /* * Detach the object from its shadow if we are the shadow's * copy. The reference we hold on the shadow must be dropped @@ -1064,7 +1815,8 @@ vm_object_terminate( vm_object_unlock(shadow_object); } - if (object->paging_in_progress != 0) { + if (object->paging_in_progress != 0 || + object->activity_in_progress != 0) { /* * There are still some paging_in_progress references * on this object, meaning that there are some paging @@ -1087,7 +1839,6 @@ vm_object_terminate( * VM object is "terminating" and not "alive". */ vm_object_reap_async(object); - vm_object_cache_unlock(); vm_object_unlock(object); /* * Return KERN_FAILURE to let the caller know that we @@ -1098,62 +1849,65 @@ vm_object_terminate( */ return KERN_FAILURE; } - - /* complete the VM object termination */ + /* + * complete the VM object termination + */ vm_object_reap(object); object = VM_OBJECT_NULL; - /* cache lock and object lock were released by vm_object_reap() */ + /* + * the object lock was released by vm_object_reap() + * + * KERN_SUCCESS means that this object has been terminated + * and no longer needs its shadow object but still holds a + * reference on it. + * The caller is responsible for dropping that reference. + * We can't call vm_object_deallocate() here because that + * would create a recursion. + */ return KERN_SUCCESS; } + /* * vm_object_reap(): * * Complete the termination of a VM object after it's been marked * as "terminating" and "!alive" by vm_object_terminate(). * - * The VM object cache and the VM object must be locked by caller. - * The locks will be released on return and the VM object is no longer valid. + * The VM object must be locked by caller. + * The lock will be released on return and the VM object is no longer valid. */ + void vm_object_reap( vm_object_t object) { memory_object_t pager; - vm_page_t p; -#if DEBUG - mutex_assert(&vm_object_cached_lock_data, MA_OWNED); - mutex_assert(&object->Lock, MA_OWNED); -#endif /* DEBUG */ + vm_object_lock_assert_exclusive(object); + assert(object->paging_in_progress == 0); + assert(object->activity_in_progress == 0); vm_object_reap_count++; /* - * The pageout daemon might be playing with our pages. - * Now that the object is dead, it won't touch any more - * pages, but some pages might already be on their way out. - * Hence, we wait until the active paging activities have - * ceased before we break the association with the pager - * itself. + * Disown this purgeable object to cleanup its owner's purgeable + * ledgers. We need to do this before disconnecting the object + * from its pager, to properly account for compressed pages. */ - while (object->paging_in_progress != 0) { - vm_object_cache_unlock(); - vm_object_wait(object, - VM_OBJECT_EVENT_PAGING_IN_PROGRESS, - THREAD_UNINT); - vm_object_cache_lock(); - vm_object_lock(object); + if (object->internal && + object->purgable != VM_PURGABLE_DENY) { + vm_purgeable_accounting(object, + object->purgable, + TRUE); /* disown */ } - assert(object->paging_in_progress == 0); pager = object->pager; object->pager = MEMORY_OBJECT_NULL; if (pager != MEMORY_OBJECT_NULL) memory_object_control_disable(object->pager_control); - vm_object_cache_unlock(); object->ref_count--; #if TASK_SWAPPER @@ -1162,6 +1916,75 @@ vm_object_reap( assert (object->ref_count == 0); + /* + * remove from purgeable queue if it's on + */ + if (object->internal) { + task_t owner; + + owner = object->vo_purgeable_owner; + + VM_OBJECT_UNWIRED(object); + + if (object->purgable == VM_PURGABLE_DENY) { + /* not purgeable: nothing to do */ + } else if (object->purgable == VM_PURGABLE_VOLATILE) { + purgeable_q_t queue; + + assert(object->vo_purgeable_owner == NULL); + + queue = vm_purgeable_object_remove(object); + assert(queue); + + if (object->purgeable_when_ripe) { + /* + * Must take page lock for this - + * using it to protect token queue + */ + vm_page_lock_queues(); + vm_purgeable_token_delete_first(queue); + + assert(queue->debug_count_objects>=0); + vm_page_unlock_queues(); + } + + /* + * Update "vm_page_purgeable_count" in bulk and mark + * object as VM_PURGABLE_EMPTY to avoid updating + * "vm_page_purgeable_count" again in vm_page_remove() + * when reaping the pages. + */ + unsigned int delta; + assert(object->resident_page_count >= + object->wired_page_count); + delta = (object->resident_page_count - + object->wired_page_count); + if (delta != 0) { + assert(vm_page_purgeable_count >= delta); + OSAddAtomic(-delta, + (SInt32 *)&vm_page_purgeable_count); + } + if (object->wired_page_count != 0) { + assert(vm_page_purgeable_wired_count >= + object->wired_page_count); + OSAddAtomic(-object->wired_page_count, + (SInt32 *)&vm_page_purgeable_wired_count); + } + object->purgable = VM_PURGABLE_EMPTY; + } + else if (object->purgable == VM_PURGABLE_NONVOLATILE || + object->purgable == VM_PURGABLE_EMPTY) { + /* remove from nonvolatile queue */ + assert(object->vo_purgeable_owner == TASK_NULL); + vm_purgeable_nonvolatile_dequeue(object); + } else { + panic("object %p in unexpected purgeable state 0x%x\n", + object, object->purgable); + } + assert(object->objq.next == NULL); + assert(object->objq.prev == NULL); + } + /* * Clean or free the pages, as appropriate. * It is possible for us to find busy/absent pages, @@ -1172,19 +1995,13 @@ vm_object_reap( vm_pageout_object_terminate(object); - } else if ((object->temporary && !object->can_persist) || - (pager == MEMORY_OBJECT_NULL)) { - while (!queue_empty(&object->memq)) { - p = (vm_page_t) queue_first(&object->memq); + } else if (((object->temporary && !object->can_persist) || (pager == MEMORY_OBJECT_NULL))) { - VM_PAGE_CHECK(p); - VM_PAGE_FREE(p); - } - } else if (!queue_empty(&object->memq)) { - panic("vm_object_reap: queue just emptied isn't"); + vm_object_reap_pages(object, REAP_REAP); } - + assert(vm_page_queue_empty(&object->memq)); assert(object->paging_in_progress == 0); + assert(object->activity_in_progress == 0); assert(object->ref_count == 0); /* @@ -1194,7 +2011,7 @@ vm_object_reap( */ if (pager != MEMORY_OBJECT_NULL) { vm_object_unlock(object); - vm_object_release_pager(pager); + vm_object_release_pager(pager, object->hashed); vm_object_lock(object); } @@ -1204,12 +2021,16 @@ vm_object_reap( vm_object_paging_end(object); vm_object_unlock(object); -#if MACH_PAGEMAP - vm_external_destroy(object->existence_map, object->size); -#endif /* MACH_PAGEMAP */ - object->shadow = VM_OBJECT_NULL; +#if VM_OBJECT_TRACKING + if (vm_object_tracking_inited) { + btlog_remove_entries_for_element(vm_object_tracking_btlog, + object); + } +#endif /* VM_OBJECT_TRACKING */ + + vm_object_lock_destroy(object); /* * Free the space for the object. */ @@ -1217,39 +2038,307 @@ vm_object_reap( object = VM_OBJECT_NULL; } + +unsigned int vm_max_batch = 256; + +#define V_O_R_MAX_BATCH 128 + +#define BATCH_LIMIT(max) (vm_max_batch >= max ? max : vm_max_batch) + + +#define VM_OBJ_REAP_FREELIST(_local_free_q, do_disconnect) \ + MACRO_BEGIN \ + if (_local_free_q) { \ + if (do_disconnect) { \ + vm_page_t m; \ + for (m = _local_free_q; \ + m != VM_PAGE_NULL; \ + m = m->snext) { \ + if (m->pmapped) { \ + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m)); \ + } \ + } \ + } \ + vm_page_free_list(_local_free_q, TRUE); \ + _local_free_q = VM_PAGE_NULL; \ + } \ + MACRO_END + + +void +vm_object_reap_pages( + vm_object_t object, + int reap_type) +{ + vm_page_t p; + vm_page_t next; + vm_page_t local_free_q = VM_PAGE_NULL; + int loop_count; + boolean_t disconnect_on_release; + pmap_flush_context pmap_flush_context_storage; + + if (reap_type == REAP_DATA_FLUSH) { + /* + * We need to disconnect pages from all pmaps before + * releasing them to the free list + */ + disconnect_on_release = TRUE; + } else { + /* + * Either the caller has already disconnected the pages + * from all pmaps, or we disconnect them here as we add + * them to out local list of pages to be released. + * No need to re-disconnect them when we release the pages + * to the free list. + */ + disconnect_on_release = FALSE; + } + +restart_after_sleep: + if (vm_page_queue_empty(&object->memq)) + return; + loop_count = BATCH_LIMIT(V_O_R_MAX_BATCH); + + if (reap_type == REAP_PURGEABLE) + pmap_flush_context_init(&pmap_flush_context_storage); + + vm_page_lockspin_queues(); + + next = (vm_page_t)vm_page_queue_first(&object->memq); + + while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t)next)) { + + p = next; + next = (vm_page_t)vm_page_queue_next(&next->listq); + + if (--loop_count == 0) { + + vm_page_unlock_queues(); + + if (local_free_q) { + + if (reap_type == REAP_PURGEABLE) { + pmap_flush(&pmap_flush_context_storage); + pmap_flush_context_init(&pmap_flush_context_storage); + } + /* + * Free the pages we reclaimed so far + * and take a little break to avoid + * hogging the page queue lock too long + */ + VM_OBJ_REAP_FREELIST(local_free_q, + disconnect_on_release); + } else + mutex_pause(0); + + loop_count = BATCH_LIMIT(V_O_R_MAX_BATCH); + + vm_page_lockspin_queues(); + } + if (reap_type == REAP_DATA_FLUSH || reap_type == REAP_TERMINATE) { + + if (p->busy || p->cleaning) { + + vm_page_unlock_queues(); + /* + * free the pages reclaimed so far + */ + VM_OBJ_REAP_FREELIST(local_free_q, + disconnect_on_release); + + PAGE_SLEEP(object, p, THREAD_UNINT); + + goto restart_after_sleep; + } + if (p->laundry) + vm_pageout_steal_laundry(p, TRUE); + } + switch (reap_type) { + + case REAP_DATA_FLUSH: + if (VM_PAGE_WIRED(p)) { + /* + * this is an odd case... perhaps we should + * zero-fill this page since we're conceptually + * tossing its data at this point, but leaving + * it on the object to honor the 'wire' contract + */ + continue; + } + break; + + case REAP_PURGEABLE: + if (VM_PAGE_WIRED(p)) { + /* + * can't purge a wired page + */ + vm_page_purged_wired++; + continue; + } + if (p->laundry && !p->busy && !p->cleaning) + vm_pageout_steal_laundry(p, TRUE); + + if (p->cleaning || p->laundry || p->absent) { + /* + * page is being acted upon, + * so don't mess with it + */ + vm_page_purged_others++; + continue; + } + if (p->busy) { + /* + * We can't reclaim a busy page but we can + * make it more likely to be paged (it's not wired) to make + * sure that it gets considered by + * vm_pageout_scan() later. + */ + if (VM_PAGE_PAGEABLE(p)) + vm_page_deactivate(p); + vm_page_purged_busy++; + continue; + } + + assert(VM_PAGE_OBJECT(p) != kernel_object); + + /* + * we can discard this page... + */ + if (p->pmapped == TRUE) { + /* + * unmap the page + */ + pmap_disconnect_options(VM_PAGE_GET_PHYS_PAGE(p), PMAP_OPTIONS_NOFLUSH | PMAP_OPTIONS_NOREFMOD, (void *)&pmap_flush_context_storage); + } + vm_page_purged_count++; + + break; + + case REAP_TERMINATE: + if (p->absent || p->private) { + /* + * For private pages, VM_PAGE_FREE just + * leaves the page structure around for + * its owner to clean up. For absent + * pages, the structure is returned to + * the appropriate pool. + */ + break; + } + if (p->fictitious) { + assert (VM_PAGE_GET_PHYS_PAGE(p) == vm_page_guard_addr); + break; + } + if (!p->dirty && p->wpmapped) + p->dirty = pmap_is_modified(VM_PAGE_GET_PHYS_PAGE(p)); + + if ((p->dirty || p->precious) && !p->error && object->alive) { + + assert(!object->internal); + + p->free_when_done = TRUE; + + if (!p->laundry) { + vm_page_queues_remove(p, TRUE); + /* + * flush page... page will be freed + * upon completion of I/O + */ + (void)vm_pageout_cluster(p, FALSE, FALSE); + } + vm_page_unlock_queues(); + /* + * free the pages reclaimed so far + */ + VM_OBJ_REAP_FREELIST(local_free_q, + disconnect_on_release); + + vm_object_paging_wait(object, THREAD_UNINT); + + goto restart_after_sleep; + } + break; + + case REAP_REAP: + break; + } + vm_page_free_prepare_queues(p); + assert(p->pageq.next == 0 && p->pageq.prev == 0); + /* + * Add this page to our list of reclaimed pages, + * to be freed later. + */ + p->snext = local_free_q; + local_free_q = p; + } + vm_page_unlock_queues(); + + /* + * Free the remaining reclaimed pages + */ + if (reap_type == REAP_PURGEABLE) + pmap_flush(&pmap_flush_context_storage); + + VM_OBJ_REAP_FREELIST(local_free_q, + disconnect_on_release); +} + + void vm_object_reap_async( vm_object_t object) { -#if DEBUG - mutex_assert(&vm_object_cached_lock_data, MA_OWNED); - mutex_assert(&object->Lock, MA_OWNED); -#endif /* DEBUG */ + vm_object_lock_assert_exclusive(object); + + vm_object_reaper_lock_spin(); vm_object_reap_count_async++; /* enqueue the VM object... */ queue_enter(&vm_object_reaper_queue, object, vm_object_t, cached_list); + + vm_object_reaper_unlock(); + /* ... and wake up the reaper thread */ thread_wakeup((event_t) &vm_object_reaper_queue); } + void vm_object_reaper_thread(void) { vm_object_t object, shadow_object; - vm_object_cache_lock(); + vm_object_reaper_lock_spin(); while (!queue_empty(&vm_object_reaper_queue)) { queue_remove_first(&vm_object_reaper_queue, object, vm_object_t, cached_list); + + vm_object_reaper_unlock(); vm_object_lock(object); + assert(object->terminating); assert(!object->alive); + + /* + * The pageout daemon might be playing with our pages. + * Now that the object is dead, it won't touch any more + * pages, but some pages might already be on their way out. + * Hence, we wait until the active paging activities have + * ceased before we break the association with the pager + * itself. + */ + while (object->paging_in_progress != 0 || + object->activity_in_progress != 0) { + vm_object_wait(object, + VM_OBJECT_EVENT_PAGING_IN_PROGRESS, + THREAD_UNINT); + vm_object_lock(object); + } shadow_object = object->pageout ? VM_OBJECT_NULL : object->shadow; @@ -1266,13 +2355,14 @@ vm_object_reaper_thread(void) vm_object_deallocate(shadow_object); shadow_object = VM_OBJECT_NULL; } - - vm_object_cache_lock(); + vm_object_reaper_lock_spin(); } /* wait for more work... */ assert_wait((event_t) &vm_object_reaper_queue, THREAD_UNINT); - vm_object_cache_unlock(); + + vm_object_reaper_unlock(); + thread_block((thread_continue_t) vm_object_reaper_thread); /*NOTREACHED*/ } @@ -1288,16 +2378,18 @@ vm_object_pager_wakeup( { vm_object_hash_entry_t entry; boolean_t waiting = FALSE; + lck_mtx_t *lck; /* * If anyone was waiting for the memory_object_terminate * to be queued, wake them up now. */ - vm_object_cache_lock(); + lck = vm_object_hash_lock_spin(pager); entry = vm_object_hash_lookup(pager, TRUE); if (entry != VM_OBJECT_HASH_ENTRY_NULL) waiting = entry->waiting; - vm_object_cache_unlock(); + vm_object_hash_unlock(lck); + if (entry != VM_OBJECT_HASH_ENTRY_NULL) { if (waiting) thread_wakeup((event_t) pager); @@ -1316,7 +2408,8 @@ vm_object_pager_wakeup( */ static void vm_object_release_pager( - memory_object_t pager) + memory_object_t pager, + boolean_t hashed) { /* @@ -1325,11 +2418,13 @@ vm_object_release_pager( (void) memory_object_terminate(pager); - /* - * Wakeup anyone waiting for this terminate - */ - vm_object_pager_wakeup(pager); - + if (hashed == TRUE) { + /* + * Wakeup anyone waiting for this terminate + * and remove the entry from the hash + */ + vm_object_pager_wakeup(pager); + } /* * Release reference to pager. */ @@ -1362,22 +2457,24 @@ vm_object_destroy( * the destroy call.] */ - vm_object_cache_lock(); vm_object_lock(object); object->can_persist = FALSE; object->named = FALSE; object->alive = FALSE; - /* - * Rip out the pager from the vm_object now... - */ - - vm_object_remove(object); + if (object->hashed) { + lck_mtx_t *lck; + /* + * Rip out the pager from the vm_object now... + */ + lck = vm_object_hash_lock_spin(object->pager); + vm_object_remove(object); + vm_object_hash_unlock(lck); + } old_pager = object->pager; object->pager = MEMORY_OBJECT_NULL; if (old_pager != MEMORY_OBJECT_NULL) memory_object_control_disable(object->pager_control); - vm_object_cache_unlock(); /* * Wait for the existing paging activity (that got @@ -1391,7 +2488,7 @@ vm_object_destroy( * Terminate the object now. */ if (old_pager != MEMORY_OBJECT_NULL) { - vm_object_release_pager(old_pager); + vm_object_release_pager(old_pager, object->hashed); /* * JMM - Release the caller's reference. This assumes the @@ -1406,8 +2503,16 @@ vm_object_destroy( return(KERN_SUCCESS); } + +#if VM_OBJECT_CACHE + +#define VM_OBJ_DEACT_ALL_STATS DEBUG +#if VM_OBJ_DEACT_ALL_STATS +uint32_t vm_object_deactivate_all_pages_batches = 0; +uint32_t vm_object_deactivate_all_pages_pages = 0; +#endif /* VM_OBJ_DEACT_ALL_STATS */ /* - * vm_object_deactivate_pages + * vm_object_deactivate_all_pages * * Deactivate all pages in the specified object. (Keep its pages * in memory even though it is no longer referenced.) @@ -1416,125 +2521,634 @@ vm_object_destroy( */ static void vm_object_deactivate_all_pages( - register vm_object_t object) + vm_object_t object) { - register vm_page_t p; - - queue_iterate(&object->memq, p, vm_page_t, listq) { - vm_page_lock_queues(); - if (!p->busy) + vm_page_t p; + int loop_count; +#if VM_OBJ_DEACT_ALL_STATS + int pages_count; +#endif /* VM_OBJ_DEACT_ALL_STATS */ +#define V_O_D_A_P_MAX_BATCH 256 + + loop_count = BATCH_LIMIT(V_O_D_A_P_MAX_BATCH); +#if VM_OBJ_DEACT_ALL_STATS + pages_count = 0; +#endif /* VM_OBJ_DEACT_ALL_STATS */ + vm_page_lock_queues(); + vm_page_queue_iterate(&object->memq, p, vm_page_t, listq) { + if (--loop_count == 0) { +#if VM_OBJ_DEACT_ALL_STATS + hw_atomic_add(&vm_object_deactivate_all_pages_batches, + 1); + hw_atomic_add(&vm_object_deactivate_all_pages_pages, + pages_count); + pages_count = 0; +#endif /* VM_OBJ_DEACT_ALL_STATS */ + lck_mtx_yield(&vm_page_queue_lock); + loop_count = BATCH_LIMIT(V_O_D_A_P_MAX_BATCH); + } + if (!p->busy && (p->vm_page_q_state != VM_PAGE_ON_THROTTLED_Q)) { +#if VM_OBJ_DEACT_ALL_STATS + pages_count++; +#endif /* VM_OBJ_DEACT_ALL_STATS */ vm_page_deactivate(p); - vm_page_unlock_queues(); + } } +#if VM_OBJ_DEACT_ALL_STATS + if (pages_count) { + hw_atomic_add(&vm_object_deactivate_all_pages_batches, 1); + hw_atomic_add(&vm_object_deactivate_all_pages_pages, + pages_count); + pages_count = 0; + } +#endif /* VM_OBJ_DEACT_ALL_STATS */ + vm_page_unlock_queues(); } +#endif /* VM_OBJECT_CACHE */ -__private_extern__ void -vm_object_deactivate_pages( - vm_object_t object, - vm_object_offset_t offset, - vm_object_size_t size, - boolean_t kill_page) -{ - vm_object_t orig_object; - int pages_moved = 0; - int pages_found = 0; - /* - * entered with object lock held, acquire a paging reference to - * prevent the memory_object and control ports from - * being destroyed. - */ - orig_object = object; - for (;;) { - register vm_page_t m; - vm_object_offset_t toffset; - vm_object_size_t tsize; +/* + * The "chunk" macros are used by routines below when looking for pages to deactivate. These + * exist because of the need to handle shadow chains. When deactivating pages, we only + * want to deactive the ones at the top most level in the object chain. In order to do + * this efficiently, the specified address range is divided up into "chunks" and we use + * a bit map to keep track of which pages have already been processed as we descend down + * the shadow chain. These chunk macros hide the details of the bit map implementation + * as much as we can. + * + * For convenience, we use a 64-bit data type as the bit map, and therefore a chunk is + * set to 64 pages. The bit map is indexed from the low-order end, so that the lowest + * order bit represents page 0 in the current range and highest order bit represents + * page 63. + * + * For further convenience, we also use negative logic for the page state in the bit map. + * The bit is set to 1 to indicate it has not yet been seen, and to 0 to indicate it has + * been processed. This way we can simply test the 64-bit long word to see if it's zero + * to easily tell if the whole range has been processed. Therefore, the bit map starts + * out with all the bits set. The macros below hide all these details from the caller. + */ - vm_object_paging_begin(object); - vm_page_lock_queues(); +#define PAGES_IN_A_CHUNK 64 /* The number of pages in the chunk must */ + /* be the same as the number of bits in */ + /* the chunk_state_t type. We use 64 */ + /* just for convenience. */ - for (tsize = size, toffset = offset; tsize; tsize -= PAGE_SIZE, toffset += PAGE_SIZE) { +#define CHUNK_SIZE (PAGES_IN_A_CHUNK * PAGE_SIZE_64) /* Size of a chunk in bytes */ - if ((m = vm_page_lookup(object, toffset)) != VM_PAGE_NULL) { +typedef uint64_t chunk_state_t; - pages_found++; +/* + * The bit map uses negative logic, so we start out with all 64 bits set to indicate + * that no pages have been processed yet. Also, if len is less than the full CHUNK_SIZE, + * then we mark pages beyond the len as having been "processed" so that we don't waste time + * looking at pages in that range. This can save us from unnecessarily chasing down the + * shadow chain. + */ - if ((m->wire_count == 0) && (!m->private) && (!m->gobbled) && (!m->busy)) { +#define CHUNK_INIT(c, len) \ + MACRO_BEGIN \ + uint64_t p; \ + \ + (c) = 0xffffffffffffffffLL; \ + \ + for (p = (len) / PAGE_SIZE_64; p < PAGES_IN_A_CHUNK; p++) \ + MARK_PAGE_HANDLED(c, p); \ + MACRO_END - assert(!m->laundry); - m->reference = FALSE; - pmap_clear_reference(m->phys_page); +/* + * Return true if all pages in the chunk have not yet been processed. + */ - if ((kill_page) && (object->internal)) { - m->precious = FALSE; - m->dirty = FALSE; - pmap_clear_modify(m->phys_page); - vm_external_state_clr(object->existence_map, offset); - } - VM_PAGE_QUEUES_REMOVE(m); - - assert(!m->laundry); - assert(m->object != kernel_object); - assert(m->pageq.next == NULL && - m->pageq.prev == NULL); - if(m->zero_fill) { - queue_enter_first( - &vm_page_queue_zf, - m, vm_page_t, pageq); - } else { - queue_enter_first( - &vm_page_queue_inactive, - m, vm_page_t, pageq); - } +#define CHUNK_NOT_COMPLETE(c) ((c) != 0) - m->inactive = TRUE; - if (!m->fictitious) - vm_page_inactive_count++; +/* + * Return true if the page at offset 'p' in the bit map has already been handled + * while processing a higher level object in the shadow chain. + */ - pages_moved++; - } - } - } - vm_page_unlock_queues(); - vm_object_paging_end(object); +#define PAGE_ALREADY_HANDLED(c, p) (((c) & (1LL << (p))) == 0) + +/* + * Mark the page at offset 'p' in the bit map as having been processed. + */ - if (object->shadow) { - vm_object_t tmp_object; +#define MARK_PAGE_HANDLED(c, p) \ +MACRO_BEGIN \ + (c) = (c) & ~(1LL << (p)); \ +MACRO_END - kill_page = 0; - offset += object->shadow_offset; +/* + * Return true if the page at the given offset has been paged out. Object is + * locked upon entry and returned locked. + */ - tmp_object = object->shadow; - vm_object_lock(tmp_object); +static boolean_t +page_is_paged_out( + vm_object_t object, + vm_object_offset_t offset) +{ + if (object->internal && + object->alive && + !object->terminating && + object->pager_ready) { - if (object != orig_object) - vm_object_unlock(object); - object = tmp_object; - } else - break; + if (VM_COMPRESSOR_PAGER_STATE_GET(object, offset) + == VM_EXTERNAL_STATE_EXISTS) { + return TRUE; + } } - if (object != orig_object) - vm_object_unlock(object); + return FALSE; } + + /* - * Routine: vm_object_pmap_protect - * - * Purpose: - * Reduces the permission for all physical - * pages in the specified object range. + * madvise_free_debug * - * If removing write permission only, it is - * sufficient to protect only the pages in - * the top-level object; only those pages may - * have write permission. - * - * If removing all access, we must follow the - * shadow chain from the top-level object to - * remove access to all pages in shadowed objects. + * To help debug madvise(MADV_FREE*) mis-usage, this triggers a + * zero-fill as soon as a page is affected by a madvise(MADV_FREE*), to + * simulate the loss of the page's contents as if the page had been + * reclaimed and then re-faulted. + */ +#if DEVELOPMENT || DEBUG +int madvise_free_debug = 1; +#else /* DEBUG */ +int madvise_free_debug = 0; +#endif /* DEBUG */ + +/* + * Deactivate the pages in the specified object and range. If kill_page is set, also discard any + * page modified state from the pmap. Update the chunk_state as we go along. The caller must specify + * a size that is less than or equal to the CHUNK_SIZE. + */ + +static void +deactivate_pages_in_object( + vm_object_t object, + vm_object_offset_t offset, + vm_object_size_t size, + boolean_t kill_page, + boolean_t reusable_page, + boolean_t all_reusable, + chunk_state_t *chunk_state, + pmap_flush_context *pfc, + struct pmap *pmap, + vm_map_offset_t pmap_offset) +{ + vm_page_t m; + int p; + struct vm_page_delayed_work dw_array[DEFAULT_DELAYED_WORK_LIMIT]; + struct vm_page_delayed_work *dwp; + int dw_count; + int dw_limit; + unsigned int reusable = 0; + + /* + * Examine each page in the chunk. The variable 'p' is the page number relative to the start of the + * chunk. Since this routine is called once for each level in the shadow chain, the chunk_state may + * have pages marked as having been processed already. We stop the loop early if we find we've handled + * all the pages in the chunk. + */ + + dwp = &dw_array[0]; + dw_count = 0; + dw_limit = DELAYED_WORK_LIMIT(DEFAULT_DELAYED_WORK_LIMIT); + + for(p = 0; size && CHUNK_NOT_COMPLETE(*chunk_state); p++, size -= PAGE_SIZE_64, offset += PAGE_SIZE_64, pmap_offset += PAGE_SIZE_64) { + + /* + * If this offset has already been found and handled in a higher level object, then don't + * do anything with it in the current shadow object. + */ + + if (PAGE_ALREADY_HANDLED(*chunk_state, p)) + continue; + + /* + * See if the page at this offset is around. First check to see if the page is resident, + * then if not, check the existence map or with the pager. + */ + + if ((m = vm_page_lookup(object, offset)) != VM_PAGE_NULL) { + + /* + * We found a page we were looking for. Mark it as "handled" now in the chunk_state + * so that we won't bother looking for a page at this offset again if there are more + * shadow objects. Then deactivate the page. + */ + + MARK_PAGE_HANDLED(*chunk_state, p); + + if (( !VM_PAGE_WIRED(m)) && (!m->private) && (!m->gobbled) && (!m->busy) && (!m->laundry)) { + int clear_refmod; + int pmap_options; + + dwp->dw_mask = 0; + + pmap_options = 0; + clear_refmod = VM_MEM_REFERENCED; + dwp->dw_mask |= DW_clear_reference; + + if ((kill_page) && (object->internal)) { + if (madvise_free_debug) { + /* + * zero-fill the page now + * to simulate it being + * reclaimed and re-faulted. + */ + pmap_zero_page(VM_PAGE_GET_PHYS_PAGE(m)); + } + m->precious = FALSE; + m->dirty = FALSE; + + clear_refmod |= VM_MEM_MODIFIED; + if (m->vm_page_q_state == VM_PAGE_ON_THROTTLED_Q) { + /* + * This page is now clean and + * reclaimable. Move it out + * of the throttled queue, so + * that vm_pageout_scan() can + * find it. + */ + dwp->dw_mask |= DW_move_page; + } + + VM_COMPRESSOR_PAGER_STATE_CLR(object, offset); + + if (reusable_page && !m->reusable) { + assert(!all_reusable); + assert(!object->all_reusable); + m->reusable = TRUE; + object->reusable_page_count++; + assert(object->resident_page_count >= object->reusable_page_count); + reusable++; + /* + * Tell pmap this page is now + * "reusable" (to update pmap + * stats for all mappings). + */ + pmap_options |= PMAP_OPTIONS_SET_REUSABLE; + } + } + pmap_options |= PMAP_OPTIONS_NOFLUSH; + pmap_clear_refmod_options(VM_PAGE_GET_PHYS_PAGE(m), + clear_refmod, + pmap_options, + (void *)pfc); + + if ((m->vm_page_q_state != VM_PAGE_ON_THROTTLED_Q) && !(reusable_page || all_reusable)) + dwp->dw_mask |= DW_move_page; + + if (dwp->dw_mask) + VM_PAGE_ADD_DELAYED_WORK(dwp, m, + dw_count); + + if (dw_count >= dw_limit) { + if (reusable) { + OSAddAtomic(reusable, + &vm_page_stats_reusable.reusable_count); + vm_page_stats_reusable.reusable += reusable; + reusable = 0; + } + vm_page_do_delayed_work(object, VM_KERN_MEMORY_NONE, &dw_array[0], dw_count); + + dwp = &dw_array[0]; + dw_count = 0; + } + } + + } else { + + /* + * The page at this offset isn't memory resident, check to see if it's + * been paged out. If so, mark it as handled so we don't bother looking + * for it in the shadow chain. + */ + + if (page_is_paged_out(object, offset)) { + MARK_PAGE_HANDLED(*chunk_state, p); + + /* + * If we're killing a non-resident page, then clear the page in the existence + * map so we don't bother paging it back in if it's touched again in the future. + */ + + if ((kill_page) && (object->internal)) { + + VM_COMPRESSOR_PAGER_STATE_CLR(object, offset); + + if (pmap != PMAP_NULL) { + /* + * Tell pmap that this page + * is no longer mapped, to + * adjust the footprint ledger + * because this page is no + * longer compressed. + */ + pmap_remove_options( + pmap, + pmap_offset, + (pmap_offset + + PAGE_SIZE), + PMAP_OPTIONS_REMOVE); + } + } + } + } + } + + if (reusable) { + OSAddAtomic(reusable, &vm_page_stats_reusable.reusable_count); + vm_page_stats_reusable.reusable += reusable; + reusable = 0; + } + + if (dw_count) + vm_page_do_delayed_work(object, VM_KERN_MEMORY_NONE, &dw_array[0], dw_count); +} + + +/* + * Deactive a "chunk" of the given range of the object starting at offset. A "chunk" + * will always be less than or equal to the given size. The total range is divided up + * into chunks for efficiency and performance related to the locks and handling the shadow + * chain. This routine returns how much of the given "size" it actually processed. It's + * up to the caler to loop and keep calling this routine until the entire range they want + * to process has been done. + */ + +static vm_object_size_t +deactivate_a_chunk( + vm_object_t orig_object, + vm_object_offset_t offset, + vm_object_size_t size, + boolean_t kill_page, + boolean_t reusable_page, + boolean_t all_reusable, + pmap_flush_context *pfc, + struct pmap *pmap, + vm_map_offset_t pmap_offset) +{ + vm_object_t object; + vm_object_t tmp_object; + vm_object_size_t length; + chunk_state_t chunk_state; + + + /* + * Get set to do a chunk. We'll do up to CHUNK_SIZE, but no more than the + * remaining size the caller asked for. + */ + + length = MIN(size, CHUNK_SIZE); + + /* + * The chunk_state keeps track of which pages we've already processed if there's + * a shadow chain on this object. At this point, we haven't done anything with this + * range of pages yet, so initialize the state to indicate no pages processed yet. + */ + + CHUNK_INIT(chunk_state, length); + object = orig_object; + + /* + * Start at the top level object and iterate around the loop once for each object + * in the shadow chain. We stop processing early if we've already found all the pages + * in the range. Otherwise we stop when we run out of shadow objects. + */ + + while (object && CHUNK_NOT_COMPLETE(chunk_state)) { + vm_object_paging_begin(object); + + deactivate_pages_in_object(object, offset, length, kill_page, reusable_page, all_reusable, &chunk_state, pfc, pmap, pmap_offset); + + vm_object_paging_end(object); + + /* + * We've finished with this object, see if there's a shadow object. If + * there is, update the offset and lock the new object. We also turn off + * kill_page at this point since we only kill pages in the top most object. + */ + + tmp_object = object->shadow; + + if (tmp_object) { + kill_page = FALSE; + reusable_page = FALSE; + all_reusable = FALSE; + offset += object->vo_shadow_offset; + vm_object_lock(tmp_object); + } + + if (object != orig_object) + vm_object_unlock(object); + + object = tmp_object; + } + + if (object && object != orig_object) + vm_object_unlock(object); + + return length; +} + + + +/* + * Move any resident pages in the specified range to the inactive queue. If kill_page is set, + * we also clear the modified status of the page and "forget" any changes that have been made + * to the page. + */ + +__private_extern__ void +vm_object_deactivate_pages( + vm_object_t object, + vm_object_offset_t offset, + vm_object_size_t size, + boolean_t kill_page, + boolean_t reusable_page, + struct pmap *pmap, + vm_map_offset_t pmap_offset) +{ + vm_object_size_t length; + boolean_t all_reusable; + pmap_flush_context pmap_flush_context_storage; + + /* + * We break the range up into chunks and do one chunk at a time. This is for + * efficiency and performance while handling the shadow chains and the locks. + * The deactivate_a_chunk() function returns how much of the range it processed. + * We keep calling this routine until the given size is exhausted. + */ + + + all_reusable = FALSE; +#if 11 + /* + * For the sake of accurate "reusable" pmap stats, we need + * to tell pmap about each page that is no longer "reusable", + * so we can't do the "all_reusable" optimization. + */ +#else + if (reusable_page && + object->internal && + object->vo_size != 0 && + object->vo_size == size && + object->reusable_page_count == 0) { + all_reusable = TRUE; + reusable_page = FALSE; + } +#endif + + if ((reusable_page || all_reusable) && object->all_reusable) { + /* This means MADV_FREE_REUSABLE has been called twice, which + * is probably illegal. */ + return; + } + + pmap_flush_context_init(&pmap_flush_context_storage); + + while (size) { + length = deactivate_a_chunk(object, offset, size, kill_page, reusable_page, all_reusable, &pmap_flush_context_storage, pmap, pmap_offset); + + size -= length; + offset += length; + pmap_offset += length; + } + pmap_flush(&pmap_flush_context_storage); + + if (all_reusable) { + if (!object->all_reusable) { + unsigned int reusable; + + object->all_reusable = TRUE; + assert(object->reusable_page_count == 0); + /* update global stats */ + reusable = object->resident_page_count; + OSAddAtomic(reusable, + &vm_page_stats_reusable.reusable_count); + vm_page_stats_reusable.reusable += reusable; + vm_page_stats_reusable.all_reusable_calls++; + } + } else if (reusable_page) { + vm_page_stats_reusable.partial_reusable_calls++; + } +} + +void +vm_object_reuse_pages( + vm_object_t object, + vm_object_offset_t start_offset, + vm_object_offset_t end_offset, + boolean_t allow_partial_reuse) +{ + vm_object_offset_t cur_offset; + vm_page_t m; + unsigned int reused, reusable; + +#define VM_OBJECT_REUSE_PAGE(object, m, reused) \ + MACRO_BEGIN \ + if ((m) != VM_PAGE_NULL && \ + (m)->reusable) { \ + assert((object)->reusable_page_count <= \ + (object)->resident_page_count); \ + assert((object)->reusable_page_count > 0); \ + (object)->reusable_page_count--; \ + (m)->reusable = FALSE; \ + (reused)++; \ + /* \ + * Tell pmap that this page is no longer \ + * "reusable", to update the "reusable" stats \ + * for all the pmaps that have mapped this \ + * page. \ + */ \ + pmap_clear_refmod_options(VM_PAGE_GET_PHYS_PAGE((m)), \ + 0, /* refmod */ \ + (PMAP_OPTIONS_CLEAR_REUSABLE \ + | PMAP_OPTIONS_NOFLUSH), \ + NULL); \ + } \ + MACRO_END + + reused = 0; + reusable = 0; + + vm_object_lock_assert_exclusive(object); + + if (object->all_reusable) { + panic("object %p all_reusable: can't update pmap stats\n", + object); + assert(object->reusable_page_count == 0); + object->all_reusable = FALSE; + if (end_offset - start_offset == object->vo_size || + !allow_partial_reuse) { + vm_page_stats_reusable.all_reuse_calls++; + reused = object->resident_page_count; + } else { + vm_page_stats_reusable.partial_reuse_calls++; + vm_page_queue_iterate(&object->memq, m, vm_page_t, listq) { + if (m->offset < start_offset || + m->offset >= end_offset) { + m->reusable = TRUE; + object->reusable_page_count++; + assert(object->resident_page_count >= object->reusable_page_count); + continue; + } else { + assert(!m->reusable); + reused++; + } + } + } + } else if (object->resident_page_count > + ((end_offset - start_offset) >> PAGE_SHIFT)) { + vm_page_stats_reusable.partial_reuse_calls++; + for (cur_offset = start_offset; + cur_offset < end_offset; + cur_offset += PAGE_SIZE_64) { + if (object->reusable_page_count == 0) { + break; + } + m = vm_page_lookup(object, cur_offset); + VM_OBJECT_REUSE_PAGE(object, m, reused); + } + } else { + vm_page_stats_reusable.partial_reuse_calls++; + vm_page_queue_iterate(&object->memq, m, vm_page_t, listq) { + if (object->reusable_page_count == 0) { + break; + } + if (m->offset < start_offset || + m->offset >= end_offset) { + continue; + } + VM_OBJECT_REUSE_PAGE(object, m, reused); + } + } + + /* update global stats */ + OSAddAtomic(reusable-reused, &vm_page_stats_reusable.reusable_count); + vm_page_stats_reusable.reused += reused; + vm_page_stats_reusable.reusable += reusable; +} + +/* + * Routine: vm_object_pmap_protect + * + * Purpose: + * Reduces the permission for all physical + * pages in the specified object range. + * + * If removing write permission only, it is + * sufficient to protect only the pages in + * the top-level object; only those pages may + * have write permission. + * + * If removing all access, we must follow the + * shadow chain from the top-level object to + * remove access to all pages in shadowed objects. * * The object must *not* be locked. The object must * be temporary/internal. @@ -1546,58 +3160,122 @@ vm_object_deactivate_pages( __private_extern__ void vm_object_pmap_protect( - register vm_object_t object, - register vm_object_offset_t offset, + vm_object_t object, + vm_object_offset_t offset, vm_object_size_t size, pmap_t pmap, vm_map_offset_t pmap_start, vm_prot_t prot) { + vm_object_pmap_protect_options(object, offset, size, + pmap, pmap_start, prot, 0); +} + +__private_extern__ void +vm_object_pmap_protect_options( + vm_object_t object, + vm_object_offset_t offset, + vm_object_size_t size, + pmap_t pmap, + vm_map_offset_t pmap_start, + vm_prot_t prot, + int options) +{ + pmap_flush_context pmap_flush_context_storage; + boolean_t delayed_pmap_flush = FALSE; + if (object == VM_OBJECT_NULL) - return; + return; size = vm_object_round_page(size); offset = vm_object_trunc_page(offset); vm_object_lock(object); + if (object->phys_contiguous) { + if (pmap != NULL) { + vm_object_unlock(object); + pmap_protect_options(pmap, + pmap_start, + pmap_start + size, + prot, + options & ~PMAP_OPTIONS_NOFLUSH, + NULL); + } else { + vm_object_offset_t phys_start, phys_end, phys_addr; + + phys_start = object->vo_shadow_offset + offset; + phys_end = phys_start + size; + assert(phys_start <= phys_end); + assert(phys_end <= object->vo_shadow_offset + object->vo_size); + vm_object_unlock(object); + + pmap_flush_context_init(&pmap_flush_context_storage); + delayed_pmap_flush = FALSE; + + for (phys_addr = phys_start; + phys_addr < phys_end; + phys_addr += PAGE_SIZE_64) { + pmap_page_protect_options( + (ppnum_t) (phys_addr >> PAGE_SHIFT), + prot, + options | PMAP_OPTIONS_NOFLUSH, + (void *)&pmap_flush_context_storage); + delayed_pmap_flush = TRUE; + } + if (delayed_pmap_flush == TRUE) + pmap_flush(&pmap_flush_context_storage); + } + return; + } + assert(object->internal); while (TRUE) { if (ptoa_64(object->resident_page_count) > size/2 && pmap != PMAP_NULL) { vm_object_unlock(object); - pmap_protect(pmap, pmap_start, pmap_start + size, prot); + pmap_protect_options(pmap, pmap_start, pmap_start + size, prot, + options & ~PMAP_OPTIONS_NOFLUSH, NULL); return; } - /* if we are doing large ranges with respect to resident */ - /* page count then we should interate over pages otherwise */ - /* inverse page look-up will be faster */ + pmap_flush_context_init(&pmap_flush_context_storage); + delayed_pmap_flush = FALSE; + + /* + * if we are doing large ranges with respect to resident + * page count then we should interate over pages otherwise + * inverse page look-up will be faster + */ if (ptoa_64(object->resident_page_count / 4) < size) { vm_page_t p; vm_object_offset_t end; end = offset + size; - if (pmap != PMAP_NULL) { - queue_iterate(&object->memq, p, vm_page_t, listq) { - if (!p->fictitious && - (offset <= p->offset) && (p->offset < end)) { - vm_map_offset_t start; - - start = pmap_start + p->offset - offset; - pmap_protect(pmap, start, start + PAGE_SIZE_64, prot); - } - } - } else { - queue_iterate(&object->memq, p, vm_page_t, listq) { - if (!p->fictitious && - (offset <= p->offset) && (p->offset < end)) { - - pmap_page_protect(p->phys_page, - prot & ~p->page_lock); - } - } + vm_page_queue_iterate(&object->memq, p, vm_page_t, listq) { + if (!p->fictitious && (offset <= p->offset) && (p->offset < end)) { + vm_map_offset_t start; + + start = pmap_start + p->offset - offset; + + if (pmap != PMAP_NULL) + pmap_protect_options( + pmap, + start, + start + PAGE_SIZE_64, + prot, + options | PMAP_OPTIONS_NOFLUSH, + &pmap_flush_context_storage); + else + pmap_page_protect_options( + VM_PAGE_GET_PHYS_PAGE(p), + prot, + options | PMAP_OPTIONS_NOFLUSH, + &pmap_flush_context_storage); + delayed_pmap_flush = TRUE; + } } + } else { vm_page_t p; vm_object_offset_t end; @@ -1605,41 +3283,47 @@ vm_object_pmap_protect( end = offset + size; - if (pmap != PMAP_NULL) { - for(target_off = offset; - target_off < end; - target_off += PAGE_SIZE) { - p = vm_page_lookup(object, target_off); - if (p != VM_PAGE_NULL) { - vm_offset_t start; - start = pmap_start + - (vm_offset_t)(p->offset - offset); - pmap_protect(pmap, start, - start + PAGE_SIZE, prot); - } - } - } else { - for(target_off = offset; - target_off < end; target_off += PAGE_SIZE) { - p = vm_page_lookup(object, target_off); - if (p != VM_PAGE_NULL) { - pmap_page_protect(p->phys_page, - prot & ~p->page_lock); - } + for (target_off = offset; + target_off < end; target_off += PAGE_SIZE) { + + p = vm_page_lookup(object, target_off); + + if (p != VM_PAGE_NULL) { + vm_object_offset_t start; + + start = pmap_start + (p->offset - offset); + + if (pmap != PMAP_NULL) + pmap_protect_options( + pmap, + start, + start + PAGE_SIZE_64, + prot, + options | PMAP_OPTIONS_NOFLUSH, + &pmap_flush_context_storage); + else + pmap_page_protect_options( + VM_PAGE_GET_PHYS_PAGE(p), + prot, + options | PMAP_OPTIONS_NOFLUSH, + &pmap_flush_context_storage); + delayed_pmap_flush = TRUE; } } - } + } + if (delayed_pmap_flush == TRUE) + pmap_flush(&pmap_flush_context_storage); if (prot == VM_PROT_NONE) { /* * Must follow shadow chain to remove access * to pages in shadowed objects. */ - register vm_object_t next_object; + vm_object_t next_object; next_object = object->shadow; if (next_object != VM_OBJECT_NULL) { - offset += object->shadow_offset; + offset += object->vo_shadow_offset; vm_object_lock(next_object); vm_object_unlock(object); object = next_object; @@ -1695,7 +3379,7 @@ vm_object_pmap_protect( */ __private_extern__ kern_return_t vm_object_copy_slowly( - register vm_object_t src_object, + vm_object_t src_object, vm_object_offset_t src_offset, vm_object_size_t size, boolean_t interruptible, @@ -1704,8 +3388,7 @@ vm_object_copy_slowly( vm_object_t new_object; vm_object_offset_t new_offset; - vm_object_offset_t src_lo_offset = src_offset; - vm_object_offset_t src_hi_offset = src_offset + size; + struct vm_object_fault_info fault_info; XPR(XPR_VM_OBJECT, "v_o_c_slowly obj 0x%x off 0x%x size 0x%x\n", src_object, src_offset, size, 0, 0); @@ -1720,9 +3403,7 @@ vm_object_copy_slowly( * Prevent destruction of the source object while we copy. */ - assert(src_object->ref_count > 0); - src_object->ref_count++; - VM_OBJ_RES_INCR(src_object); + vm_object_reference_locked(src_object); vm_object_unlock(src_object); /* @@ -1736,10 +3417,22 @@ vm_object_copy_slowly( new_object = vm_object_allocate(size); new_offset = 0; - vm_object_lock(new_object); assert(size == trunc_page_64(size)); /* Will the loop terminate? */ + fault_info.interruptible = interruptible; + fault_info.behavior = VM_BEHAVIOR_SEQUENTIAL; + fault_info.user_tag = 0; + fault_info.pmap_options = 0; + fault_info.lo_offset = src_offset; + fault_info.hi_offset = src_offset + size; + fault_info.no_cache = FALSE; + fault_info.stealth = TRUE; + fault_info.io_sync = FALSE; + fault_info.cs_bypass = FALSE; + fault_info.mark_zf_absent = FALSE; + fault_info.batch_pmap_op = FALSE; + for ( ; size != 0 ; src_offset += PAGE_SIZE_64, @@ -1748,123 +3441,172 @@ vm_object_copy_slowly( vm_page_t new_page; vm_fault_return_t result; + vm_object_lock(new_object); + while ((new_page = vm_page_alloc(new_object, new_offset)) == VM_PAGE_NULL) { + + vm_object_unlock(new_object); + if (!vm_page_wait(interruptible)) { - vm_object_unlock(new_object); vm_object_deallocate(new_object); vm_object_deallocate(src_object); *_result_object = VM_OBJECT_NULL; return(MACH_SEND_INTERRUPTED); } + vm_object_lock(new_object); } + vm_object_unlock(new_object); do { vm_prot_t prot = VM_PROT_READ; vm_page_t _result_page; vm_page_t top_page; - register vm_page_t result_page; kern_return_t error_code; + vm_object_t result_page_object; + vm_object_lock(src_object); + + if (src_object->internal && + src_object->shadow == VM_OBJECT_NULL && + (vm_page_lookup(src_object, + src_offset) == VM_PAGE_NULL) && + (src_object->pager == NULL || + (VM_COMPRESSOR_PAGER_STATE_GET(src_object, + src_offset) == + VM_EXTERNAL_STATE_ABSENT))) { + /* + * This page is neither resident nor compressed + * and there's no shadow object below + * "src_object", so this page is really missing. + * There's no need to zero-fill it just to copy + * it: let's leave it missing in "new_object" + * and get zero-filled on demand. + */ + vm_object_unlock(src_object); + /* free the unused "new_page"... */ + vm_object_lock(new_object); + VM_PAGE_FREE(new_page); + new_page = VM_PAGE_NULL; + vm_object_unlock(new_object); + /* ...and go to next page in "src_object" */ + result = VM_FAULT_SUCCESS; + break; + } + vm_object_paging_begin(src_object); + if (size > (vm_size_t) -1) { + /* 32-bit overflow */ + fault_info.cluster_size = (vm_size_t) (0 - PAGE_SIZE); + } else { + fault_info.cluster_size = (vm_size_t) size; + assert(fault_info.cluster_size == size); + } + XPR(XPR_VM_FAULT,"vm_object_copy_slowly -> vm_fault_page",0,0,0,0,0); + _result_page = VM_PAGE_NULL; result = vm_fault_page(src_object, src_offset, - VM_PROT_READ, FALSE, interruptible, - src_lo_offset, src_hi_offset, - VM_BEHAVIOR_SEQUENTIAL, + VM_PROT_READ, FALSE, + FALSE, /* page not looked up */ &prot, &_result_page, &top_page, (int *)0, - &error_code, FALSE, FALSE, NULL, 0); + &error_code, FALSE, FALSE, &fault_info); switch(result) { - case VM_FAULT_SUCCESS: - result_page = _result_page; + case VM_FAULT_SUCCESS: + result_page = _result_page; + result_page_object = VM_PAGE_OBJECT(result_page); - /* - * We don't need to hold the object - * lock -- the busy page will be enough. - * [We don't care about picking up any - * new modifications.] - * - * Copy the page to the new object. - * - * POLICY DECISION: - * If result_page is clean, - * we could steal it instead - * of copying. - */ + /* + * Copy the page to the new object. + * + * POLICY DECISION: + * If result_page is clean, + * we could steal it instead + * of copying. + */ - vm_object_unlock(result_page->object); - vm_page_copy(result_page, new_page); + vm_page_copy(result_page, new_page); + vm_object_unlock(result_page_object); - /* - * Let go of both pages (make them - * not busy, perform wakeup, activate). - */ + /* + * Let go of both pages (make them + * not busy, perform wakeup, activate). + */ + vm_object_lock(new_object); + SET_PAGE_DIRTY(new_page, FALSE); + PAGE_WAKEUP_DONE(new_page); + vm_object_unlock(new_object); + + vm_object_lock(result_page_object); + PAGE_WAKEUP_DONE(result_page); + + vm_page_lockspin_queues(); + if ((result_page->vm_page_q_state == VM_PAGE_ON_SPECULATIVE_Q) || + (result_page->vm_page_q_state == VM_PAGE_NOT_ON_Q)) { + vm_page_activate(result_page); + } + vm_page_activate(new_page); + vm_page_unlock_queues(); - new_page->busy = FALSE; - new_page->dirty = TRUE; - vm_object_lock(result_page->object); - PAGE_WAKEUP_DONE(result_page); + /* + * Release paging references and + * top-level placeholder page, if any. + */ - vm_page_lock_queues(); - if (!result_page->active && - !result_page->inactive) - vm_page_activate(result_page); - vm_page_activate(new_page); - vm_page_unlock_queues(); - - /* - * Release paging references and - * top-level placeholder page, if any. - */ + vm_fault_cleanup(result_page_object, + top_page); - vm_fault_cleanup(result_page->object, - top_page); - - break; + break; - case VM_FAULT_RETRY: - break; + case VM_FAULT_RETRY: + break; - case VM_FAULT_FICTITIOUS_SHORTAGE: - vm_page_more_fictitious(); + case VM_FAULT_MEMORY_SHORTAGE: + if (vm_page_wait(interruptible)) break; + /* fall thru */ - case VM_FAULT_MEMORY_SHORTAGE: - if (vm_page_wait(interruptible)) - break; - /* fall thru */ + case VM_FAULT_INTERRUPTED: + vm_object_lock(new_object); + VM_PAGE_FREE(new_page); + vm_object_unlock(new_object); + + vm_object_deallocate(new_object); + vm_object_deallocate(src_object); + *_result_object = VM_OBJECT_NULL; + return(MACH_SEND_INTERRUPTED); + + case VM_FAULT_SUCCESS_NO_VM_PAGE: + /* success but no VM page: fail */ + vm_object_paging_end(src_object); + vm_object_unlock(src_object); + /*FALLTHROUGH*/ + case VM_FAULT_MEMORY_ERROR: + /* + * A policy choice: + * (a) ignore pages that we can't + * copy + * (b) return the null object if + * any page fails [chosen] + */ - case VM_FAULT_INTERRUPTED: - vm_page_free(new_page); - vm_object_unlock(new_object); - vm_object_deallocate(new_object); - vm_object_deallocate(src_object); - *_result_object = VM_OBJECT_NULL; - return(MACH_SEND_INTERRUPTED); + vm_object_lock(new_object); + VM_PAGE_FREE(new_page); + vm_object_unlock(new_object); - case VM_FAULT_MEMORY_ERROR: - /* - * A policy choice: - * (a) ignore pages that we can't - * copy - * (b) return the null object if - * any page fails [chosen] - */ + vm_object_deallocate(new_object); + vm_object_deallocate(src_object); + *_result_object = VM_OBJECT_NULL; + return(error_code ? error_code: + KERN_MEMORY_ERROR); - vm_page_lock_queues(); - vm_page_free(new_page); - vm_page_unlock_queues(); - vm_object_unlock(new_object); - vm_object_deallocate(new_object); - vm_object_deallocate(src_object); - *_result_object = VM_OBJECT_NULL; - return(error_code ? error_code: - KERN_MEMORY_ERROR); + default: + panic("vm_object_copy_slowly: unexpected error" + " 0x%x from vm_fault_page()\n", result); } } while (result != VM_FAULT_SUCCESS); } @@ -1872,8 +3614,6 @@ vm_object_copy_slowly( /* * Lose the extra reference, and return our object. */ - - vm_object_unlock(new_object); vm_object_deallocate(src_object); *_result_object = new_object; return(KERN_SUCCESS); @@ -1929,9 +3669,7 @@ vm_object_copy_quickly( * Leave object/offset unchanged. */ - assert(object->ref_count > 0); - object->ref_count++; - vm_object_res_reference(object); + vm_object_reference_locked(object); object->shadowed = TRUE; vm_object_unlock(object); @@ -1989,6 +3727,7 @@ vm_object_copy_call( kern_return_t kr; vm_object_t copy; boolean_t check_ready = FALSE; + uint32_t try_failed_count = 0; /* * If a copy is already in progress, wait and retry. @@ -2048,12 +3787,15 @@ Retry: copy = src_object->copy; if (!vm_object_lock_try(copy)) { vm_object_unlock(src_object); - mutex_pause(); /* wait a bit */ + + try_failed_count++; + mutex_pause(try_failed_count); /* wait a bit */ + vm_object_lock(src_object); goto Retry; } - if (copy->size < src_offset+size) - copy->size = src_offset+size; + if (copy->vo_size < src_offset+size) + copy->vo_size = src_offset+size; if (!copy->pager_ready) check_ready = TRUE; @@ -2100,12 +3842,16 @@ __private_extern__ vm_object_t vm_object_copy_delayed( vm_object_t src_object, vm_object_offset_t src_offset, - vm_object_size_t size) + vm_object_size_t size, + boolean_t src_object_shared) { vm_object_t new_copy = VM_OBJECT_NULL; vm_object_t old_copy; vm_page_t p; vm_object_size_t copy_size = src_offset + size; + pmap_flush_context pmap_flush_context_storage; + boolean_t delayed_pmap_flush = FALSE; + int collisions = 0; /* @@ -2147,14 +3893,23 @@ vm_object_copy_delayed( * the original object must be done carefully, to avoid deadlock. */ + copy_size = vm_object_round_page(copy_size); Retry: /* * Wait for paging in progress. */ - if (!src_object->true_share) + if (!src_object->true_share && + (src_object->paging_in_progress != 0 || + src_object->activity_in_progress != 0)) { + if (src_object_shared == TRUE) { + vm_object_unlock(src_object); + vm_object_lock(src_object); + src_object_shared = FALSE; + goto Retry; + } vm_object_paging_wait(src_object, THREAD_UNINT); - + } /* * See whether we can reuse the result of a previous * copy operation. @@ -2162,22 +3917,34 @@ vm_object_copy_delayed( old_copy = src_object->copy; if (old_copy != VM_OBJECT_NULL) { + int lock_granted; + /* * Try to get the locks (out of order) */ - if (!vm_object_lock_try(old_copy)) { + if (src_object_shared == TRUE) + lock_granted = vm_object_lock_try_shared(old_copy); + else + lock_granted = vm_object_lock_try(old_copy); + + if (!lock_granted) { vm_object_unlock(src_object); - mutex_pause(); - /* Heisenberg Rules */ - copy_delayed_lock_collisions++; if (collisions++ == 0) copy_delayed_lock_contention++; + mutex_pause(collisions); + + /* Heisenberg Rules */ + copy_delayed_lock_collisions++; if (collisions > copy_delayed_max_collisions) copy_delayed_max_collisions = collisions; - vm_object_lock(src_object); + if (src_object_shared == TRUE) + vm_object_lock_shared(src_object); + else + vm_object_lock(src_object); + goto Retry; } @@ -2197,7 +3964,15 @@ vm_object_copy_delayed( * needed). */ - if (old_copy->size < copy_size) { + if (old_copy->vo_size < copy_size) { + if (src_object_shared == TRUE) { + vm_object_unlock(old_copy); + vm_object_unlock(src_object); + + vm_object_lock(src_object); + src_object_shared = FALSE; + goto Retry; + } /* * We can't perform a delayed copy if any of the * pages in the extended range are wired (because @@ -2206,11 +3981,15 @@ vm_object_copy_delayed( * go ahead and protect them. */ copy_delayed_protect_iterate++; - queue_iterate(&src_object->memq, p, vm_page_t, listq) { + + pmap_flush_context_init(&pmap_flush_context_storage); + delayed_pmap_flush = FALSE; + + vm_page_queue_iterate(&src_object->memq, p, vm_page_t, listq) { if (!p->fictitious && - p->offset >= old_copy->size && + p->offset >= old_copy->vo_size && p->offset < copy_size) { - if (p->wire_count > 0) { + if (VM_PAGE_WIRED(p)) { vm_object_unlock(old_copy); vm_object_unlock(src_object); @@ -2218,19 +3997,26 @@ vm_object_copy_delayed( vm_object_unlock(new_copy); vm_object_deallocate(new_copy); } + if (delayed_pmap_flush == TRUE) + pmap_flush(&pmap_flush_context_storage); return VM_OBJECT_NULL; } else { - pmap_page_protect(p->phys_page, - (VM_PROT_ALL & ~VM_PROT_WRITE & - ~p->page_lock)); + pmap_page_protect_options(VM_PAGE_GET_PHYS_PAGE(p), (VM_PROT_ALL & ~VM_PROT_WRITE), + PMAP_OPTIONS_NOFLUSH, (void *)&pmap_flush_context_storage); + delayed_pmap_flush = TRUE; } } } - old_copy->size = copy_size; + if (delayed_pmap_flush == TRUE) + pmap_flush(&pmap_flush_context_storage); + + old_copy->vo_size = copy_size; } - - vm_object_reference_locked(old_copy); + if (src_object_shared == TRUE) + vm_object_reference_shared(old_copy); + else + vm_object_reference_locked(old_copy); vm_object_unlock(old_copy); vm_object_unlock(src_object); @@ -2238,17 +4024,18 @@ vm_object_copy_delayed( vm_object_unlock(new_copy); vm_object_deallocate(new_copy); } - return(old_copy); } + + /* * Adjust the size argument so that the newly-created * copy object will be large enough to back either the * old copy object or the new mapping. */ - if (old_copy->size > copy_size) - copy_size = old_copy->size; + if (old_copy->vo_size > copy_size) + copy_size = old_copy->vo_size; if (new_copy == VM_OBJECT_NULL) { vm_object_unlock(old_copy); @@ -2256,9 +4043,11 @@ vm_object_copy_delayed( new_copy = vm_object_allocate(copy_size); vm_object_lock(src_object); vm_object_lock(new_copy); + + src_object_shared = FALSE; goto Retry; } - new_copy->size = copy_size; + new_copy->vo_size = copy_size; /* * The copy-object is always made large enough to @@ -2268,13 +4057,15 @@ vm_object_copy_delayed( */ assert((old_copy->shadow == src_object) && - (old_copy->shadow_offset == (vm_object_offset_t) 0)); + (old_copy->vo_shadow_offset == (vm_object_offset_t) 0)); } else if (new_copy == VM_OBJECT_NULL) { vm_object_unlock(src_object); new_copy = vm_object_allocate(copy_size); vm_object_lock(src_object); vm_object_lock(new_copy); + + src_object_shared = FALSE; goto Retry; } @@ -2290,22 +4081,32 @@ vm_object_copy_delayed( * wired, then go ahead and protect them. */ copy_delayed_protect_iterate++; - queue_iterate(&src_object->memq, p, vm_page_t, listq) { + + pmap_flush_context_init(&pmap_flush_context_storage); + delayed_pmap_flush = FALSE; + + vm_page_queue_iterate(&src_object->memq, p, vm_page_t, listq) { if (!p->fictitious && p->offset < copy_size) { - if (p->wire_count > 0) { + if (VM_PAGE_WIRED(p)) { if (old_copy) vm_object_unlock(old_copy); vm_object_unlock(src_object); vm_object_unlock(new_copy); vm_object_deallocate(new_copy); + + if (delayed_pmap_flush == TRUE) + pmap_flush(&pmap_flush_context_storage); + return VM_OBJECT_NULL; } else { - pmap_page_protect(p->phys_page, - (VM_PROT_ALL & ~VM_PROT_WRITE & - ~p->page_lock)); + pmap_page_protect_options(VM_PAGE_GET_PHYS_PAGE(p), (VM_PROT_ALL & ~VM_PROT_WRITE), + PMAP_OPTIONS_NOFLUSH, (void *)&pmap_flush_context_storage); + delayed_pmap_flush = TRUE; } } } + if (delayed_pmap_flush == TRUE) + pmap_flush(&pmap_flush_context_storage); if (old_copy != VM_OBJECT_NULL) { /* @@ -2314,9 +4115,13 @@ vm_object_copy_delayed( * object. */ - src_object->ref_count--; /* remove ref. from old_copy */ + /* remove ref. from old_copy */ + vm_object_lock_assert_exclusive(src_object); + src_object->ref_count--; assert(src_object->ref_count > 0); + vm_object_lock_assert_exclusive(old_copy); old_copy->shadow = new_copy; + vm_object_lock_assert_exclusive(new_copy); assert(new_copy->ref_count > 0); new_copy->ref_count++; /* for old_copy->shadow ref. */ @@ -2333,21 +4138,22 @@ vm_object_copy_delayed( /* * Point the new copy at the existing object. */ + vm_object_lock_assert_exclusive(new_copy); new_copy->shadow = src_object; - new_copy->shadow_offset = 0; + new_copy->vo_shadow_offset = 0; new_copy->shadowed = TRUE; /* caller must set needs_copy */ - assert(src_object->ref_count > 0); - src_object->ref_count++; - VM_OBJ_RES_INCR(src_object); + + vm_object_lock_assert_exclusive(src_object); + vm_object_reference_locked(src_object); src_object->copy = new_copy; vm_object_unlock(src_object); vm_object_unlock(new_copy); XPR(XPR_VM_OBJECT, "vm_object_copy_delayed: used copy object %X for source %X\n", - (integer_t)new_copy, (integer_t)src_object, 0, 0, 0); + new_copy, src_object, 0, 0, 0); - return(new_copy); + return new_copy; } /* @@ -2360,7 +4166,7 @@ vm_object_copy_delayed( */ __private_extern__ kern_return_t vm_object_copy_strategically( - register vm_object_t src_object, + vm_object_t src_object, vm_object_offset_t src_offset, vm_object_size_t size, vm_object_t *dst_object, /* OUT */ @@ -2369,11 +4175,18 @@ vm_object_copy_strategically( { boolean_t result; boolean_t interruptible = THREAD_ABORTSAFE; /* XXX */ + boolean_t object_lock_shared = FALSE; memory_object_copy_strategy_t copy_strategy; assert(src_object != VM_OBJECT_NULL); - vm_object_lock(src_object); + copy_strategy = src_object->copy_strategy; + + if (copy_strategy == MEMORY_OBJECT_COPY_DELAY) { + vm_object_lock_shared(src_object); + object_lock_shared = TRUE; + } else + vm_object_lock(src_object); /* * The copy strategy is only valid if the memory manager @@ -2383,6 +4196,12 @@ vm_object_copy_strategically( while (!src_object->internal && !src_object->pager_ready) { wait_result_t wait_result; + if (object_lock_shared == TRUE) { + vm_object_unlock(src_object); + vm_object_lock(src_object); + object_lock_shared = FALSE; + continue; + } wait_result = vm_object_sleep( src_object, VM_OBJECT_EVENT_PAGER_READY, interruptible); @@ -2395,8 +4214,6 @@ vm_object_copy_strategically( } } - copy_strategy = src_object->copy_strategy; - /* * Use the appropriate copy strategy. */ @@ -2404,7 +4221,7 @@ vm_object_copy_strategically( switch (copy_strategy) { case MEMORY_OBJECT_COPY_DELAY: *dst_object = vm_object_copy_delayed(src_object, - src_offset, size); + src_offset, size, object_lock_shared); if (*dst_object != VM_OBJECT_NULL) { *dst_offset = src_offset; *dst_needs_copy = TRUE; @@ -2433,7 +4250,7 @@ vm_object_copy_strategically( break; case MEMORY_OBJECT_COPY_SYMMETRIC: - XPR(XPR_VM_OBJECT, "v_o_c_strategically obj 0x%x off 0x%x size 0x%x\n",(natural_t)src_object, src_offset, size, 0, 0); + XPR(XPR_VM_OBJECT, "v_o_c_strategically obj 0x%x off 0x%x size 0x%x\n", src_object, src_offset, size, 0, 0); vm_object_unlock(src_object); result = KERN_MEMORY_RESTART_COPY; break; @@ -2455,7 +4272,7 @@ vm_object_copy_strategically( * The new object and offset into that object * are returned in the source parameters. */ -boolean_t vm_object_shadow_check = FALSE; +boolean_t vm_object_shadow_check = TRUE; __private_extern__ boolean_t vm_object_shadow( @@ -2463,22 +4280,63 @@ vm_object_shadow( vm_object_offset_t *offset, /* IN/OUT */ vm_object_size_t length) { - register vm_object_t source; - register vm_object_t result; + vm_object_t source; + vm_object_t result; source = *object; + assert(source != VM_OBJECT_NULL); + if (source == VM_OBJECT_NULL) + return FALSE; + +#if 0 + /* + * XXX FBDP + * This assertion is valid but it gets triggered by Rosetta for example + * due to a combination of vm_remap() that changes a VM object's + * copy_strategy from SYMMETRIC to DELAY and vm_protect(VM_PROT_COPY) + * that then sets "needs_copy" on its map entry. This creates a + * mapping situation that VM should never see and doesn't know how to + * handle. + * It's not clear if this can create any real problem but we should + * look into fixing this, probably by having vm_protect(VM_PROT_COPY) + * do more than just set "needs_copy" to handle the copy-on-write... + * In the meantime, let's disable the assertion. + */ assert(source->copy_strategy == MEMORY_OBJECT_COPY_SYMMETRIC); +#endif /* * Determine if we really need a shadow. + * + * If the source object is larger than what we are trying + * to create, then force the shadow creation even if the + * ref count is 1. This will allow us to [potentially] + * collapse the underlying object away in the future + * (freeing up the extra data it might contain and that + * we don't need). */ - if (vm_object_shadow_check && source->ref_count == 1 && + assert(source->copy_strategy != MEMORY_OBJECT_COPY_NONE); /* Purgeable objects shouldn't have shadow objects. */ + + if (vm_object_shadow_check && + source->vo_size == length && + source->ref_count == 1 && (source->shadow == VM_OBJECT_NULL || - source->shadow->copy == VM_OBJECT_NULL)) + source->shadow->copy == VM_OBJECT_NULL) ) { - source->shadowed = FALSE; - return FALSE; + /* lock the object and check again */ + vm_object_lock(source); + if (source->vo_size == length && + source->ref_count == 1 && + (source->shadow == VM_OBJECT_NULL || + source->shadow->copy == VM_OBJECT_NULL)) + { + source->shadowed = FALSE; + vm_object_unlock(source); + return FALSE; + } + /* things changed while we were locking "source"... */ + vm_object_unlock(source); } /* @@ -2502,7 +4360,7 @@ vm_object_shadow( * and fix up the offset into the new object. */ - result->shadow_offset = *offset; + result->vo_shadow_offset = *offset; /* * Return the new things @@ -2518,7 +4376,7 @@ vm_object_shadow( * the memory_object requires careful synchronization. * * All associations are created by memory_object_create_named - * for external pagers and vm_object_pager_create for internal + * for external pagers and vm_object_compressor_pager_create for internal * objects as follows: * * pager: the memory_object itself, supplied by @@ -2572,9 +4430,9 @@ vm_object_shadow( * [Furthermore, each routine must cope with the simultaneous * or previous operations of the others.] * - * In addition to the lock on the object, the vm_object_cache_lock + * In addition to the lock on the object, the vm_object_hash_lock * governs the associations. References gained through the - * association require use of the cache lock. + * association require use of the hash lock. * * Because the pager field may be cleared spontaneously, it * cannot be used to determine whether a memory object has @@ -2592,162 +4450,6 @@ vm_object_shadow( * termination routines and vm_object_collapse.] */ -#if 0 -static void vm_object_abort_activity( - vm_object_t object); - -/* - * Routine: vm_object_abort_activity [internal use only] - * Purpose: - * Abort paging requests pending on this object. - * In/out conditions: - * The object is locked on entry and exit. - */ -static void -vm_object_abort_activity( - vm_object_t object) -{ - register - vm_page_t p; - vm_page_t next; - - XPR(XPR_VM_OBJECT, "vm_object_abort_activity, object 0x%X\n", - (integer_t)object, 0, 0, 0, 0); - - /* - * Abort all activity that would be waiting - * for a result on this memory object. - * - * We could also choose to destroy all pages - * that we have in memory for this object, but - * we don't. - */ - - p = (vm_page_t) queue_first(&object->memq); - while (!queue_end(&object->memq, (queue_entry_t) p)) { - next = (vm_page_t) queue_next(&p->listq); - - /* - * If it's being paged in, destroy it. - * If an unlock has been requested, start it again. - */ - - if (p->busy && p->absent) { - VM_PAGE_FREE(p); - } - else { - if (p->unlock_request != VM_PROT_NONE) - p->unlock_request = VM_PROT_NONE; - PAGE_WAKEUP(p); - } - - p = next; - } - - /* - * Wake up threads waiting for the memory object to - * become ready. - */ - - object->pager_ready = TRUE; - vm_object_wakeup(object, VM_OBJECT_EVENT_PAGER_READY); -} - -/* - * Routine: vm_object_pager_dead - * - * Purpose: - * A port is being destroy, and the IPC kobject code - * can't tell if it represents a pager port or not. - * So this function is called each time it sees a port - * die. - * THIS IS HORRIBLY INEFFICIENT. We should only call - * this routine if we had requested a notification on - * the port. - */ - -__private_extern__ void -vm_object_pager_dead( - ipc_port_t pager) -{ - vm_object_t object; - vm_object_hash_entry_t entry; - - /* - * Perform essentially the same operations as in vm_object_lookup, - * except that this time we look up based on the memory_object - * port, not the control port. - */ - vm_object_cache_lock(); - entry = vm_object_hash_lookup(pager, FALSE); - if (entry == VM_OBJECT_HASH_ENTRY_NULL || - entry->object == VM_OBJECT_NULL) { - vm_object_cache_unlock(); - return; - } - - object = entry->object; - entry->object = VM_OBJECT_NULL; - - vm_object_lock(object); - if (object->ref_count == 0) { - XPR(XPR_VM_OBJECT_CACHE, - "vm_object_destroy: removing %x from cache, head (%x, %x)\n", - (integer_t)object, - (integer_t)vm_object_cached_list.next, - (integer_t)vm_object_cached_list.prev, 0,0); - - queue_remove(&vm_object_cached_list, object, - vm_object_t, cached_list); - vm_object_cached_count--; - } - object->ref_count++; - vm_object_res_reference(object); - - object->can_persist = FALSE; - - assert(object->pager == pager); - - /* - * Remove the pager association. - * - * Note that the memory_object itself is dead, so - * we don't bother with it. - */ - - object->pager = MEMORY_OBJECT_NULL; - - vm_object_unlock(object); - vm_object_cache_unlock(); - - vm_object_pager_wakeup(pager); - - /* - * Release the pager reference. Note that there's no - * point in trying the memory_object_terminate call - * because the memory_object itself is dead. Also - * release the memory_object_control reference, since - * the pager didn't do that either. - */ - - memory_object_deallocate(pager); - memory_object_control_deallocate(object->pager_request); - - - /* - * Restart pending page requests - */ - vm_object_lock(object); - vm_object_abort_activity(object); - vm_object_unlock(object); - - /* - * Lose the object reference. - */ - - vm_object_deallocate(object); -} -#endif /* * Routine: vm_object_enter @@ -2764,10 +4466,12 @@ vm_object_enter( boolean_t init, boolean_t named) { - register vm_object_t object; + vm_object_t object; vm_object_t new_object; boolean_t must_init; vm_object_hash_entry_t entry, new_entry; + uint32_t try_failed_count = 0; + lck_mtx_t *lck; if (pager == MEMORY_OBJECT_NULL) return(vm_object_allocate(size)); @@ -2779,8 +4483,8 @@ vm_object_enter( /* * Look for an object associated with this port. */ - - vm_object_cache_lock(); +Retry: + lck = vm_object_hash_lock_spin(pager); do { entry = vm_object_hash_lookup(pager, FALSE); @@ -2790,19 +4494,37 @@ vm_object_enter( * We must unlock to create a new object; * if we do so, we must try the lookup again. */ - vm_object_cache_unlock(); + vm_object_hash_unlock(lck); assert(new_entry == VM_OBJECT_HASH_ENTRY_NULL); new_entry = vm_object_hash_entry_alloc(pager); new_object = vm_object_allocate(size); - vm_object_cache_lock(); + /* + * Set new_object->hashed now, while noone + * knows about this object yet and we + * don't need to lock it. Once it's in + * the hash table, we would have to lock + * the object to set its "hashed" bit and + * we can't lock the object while holding + * the hash lock as a spinlock... + */ + new_object->hashed = TRUE; + lck = vm_object_hash_lock_spin(pager); } else { /* * Lookup failed twice, and we have something * to insert; set the object. */ - vm_object_hash_insert(new_entry); + /* + * We can't lock the object here since we're + * holding the hash lock as a spin lock. + * We've already pre-set "new_object->hashed" + * when we created "new_object" above, so we + * won't need to modify the object in + * vm_object_hash_insert(). + */ + assert(new_object->hashed); + vm_object_hash_insert(new_entry, new_object); entry = new_entry; - entry->object = new_object; new_entry = VM_OBJECT_HASH_ENTRY_NULL; new_object = VM_OBJECT_NULL; must_init = TRUE; @@ -2816,9 +4538,10 @@ vm_object_enter( entry->waiting = TRUE; entry = VM_OBJECT_HASH_ENTRY_NULL; assert_wait((event_t) pager, THREAD_UNINT); - vm_object_cache_unlock(); + vm_object_hash_unlock(lck); + thread_block(THREAD_CONTINUE_NULL); - vm_object_cache_lock(); + lck = vm_object_hash_lock_spin(pager); } } while (entry == VM_OBJECT_HASH_ENTRY_NULL); @@ -2826,45 +4549,76 @@ vm_object_enter( assert(object != VM_OBJECT_NULL); if (!must_init) { - vm_object_lock(object); - assert(!internal || object->internal); - if (named) { - assert(!object->named); - object->named = TRUE; + if ( !vm_object_lock_try(object)) { + + vm_object_hash_unlock(lck); + + try_failed_count++; + mutex_pause(try_failed_count); /* wait a bit */ + goto Retry; } + assert(!internal || object->internal); +#if VM_OBJECT_CACHE if (object->ref_count == 0) { + if ( !vm_object_cache_lock_try()) { + + vm_object_hash_unlock(lck); + vm_object_unlock(object); + + try_failed_count++; + mutex_pause(try_failed_count); /* wait a bit */ + goto Retry; + } XPR(XPR_VM_OBJECT_CACHE, - "vm_object_enter: removing %x from cache, head (%x, %x)\n", - (integer_t)object, - (integer_t)vm_object_cached_list.next, - (integer_t)vm_object_cached_list.prev, 0,0); + "vm_object_enter: removing %x from cache, head (%x, %x)\n", + object, + vm_object_cached_list.next, + vm_object_cached_list.prev, 0,0); queue_remove(&vm_object_cached_list, object, vm_object_t, cached_list); vm_object_cached_count--; + + vm_object_cache_unlock(); + } +#endif + if (named) { + assert(!object->named); + object->named = TRUE; } + vm_object_lock_assert_exclusive(object); object->ref_count++; vm_object_res_reference(object); + + vm_object_hash_unlock(lck); vm_object_unlock(object); - VM_STAT(hits++); - } - assert(object->ref_count > 0); + VM_STAT_INCR(hits); + } else + vm_object_hash_unlock(lck); - VM_STAT(lookups++); + assert(object->ref_count > 0); - vm_object_cache_unlock(); + VM_STAT_INCR(lookups); XPR(XPR_VM_OBJECT, "vm_o_enter: pager 0x%x obj 0x%x must_init %d\n", - (integer_t)pager, (integer_t)object, must_init, 0, 0); + pager, object, must_init, 0, 0); /* * If we raced to create a vm_object but lost, let's * throw away ours. */ - if (new_object != VM_OBJECT_NULL) + if (new_object != VM_OBJECT_NULL) { + /* + * Undo the pre-setting of "new_object->hashed" before + * deallocating "new_object", since we did not insert it + * into the hash table after all. + */ + assert(new_object->hashed); + new_object->hashed = FALSE; vm_object_deallocate(new_object); + } if (new_entry != VM_OBJECT_HASH_ENTRY_NULL) vm_object_hash_entry_free(new_entry); @@ -2912,6 +4666,7 @@ vm_object_enter( if (named) object->named = TRUE; if (internal) { + vm_object_lock_assert_exclusive(object); object->pager_ready = TRUE; vm_object_wakeup(object, VM_OBJECT_EVENT_PAGER_READY); } @@ -2940,12 +4695,12 @@ vm_object_enter( XPR(XPR_VM_OBJECT, "vm_object_enter: vm_object %x, memory_object %x, internal %d\n", - (integer_t)object, (integer_t)object->pager, internal, 0,0); + object, object->pager, internal, 0,0); return(object); } /* - * Routine: vm_object_pager_create + * Routine: vm_object_compressor_pager_create * Purpose: * Create a memory object for an internal object. * In/out conditions: @@ -2953,30 +4708,22 @@ vm_object_enter( * it may be unlocked within this call. * Limitations: * Only one thread may be performing a - * vm_object_pager_create on an object at + * vm_object_compressor_pager_create on an object at * a time. Presumably, only the pageout * daemon will be using this routine. */ void -vm_object_pager_create( - register vm_object_t object) +vm_object_compressor_pager_create( + vm_object_t object) { memory_object_t pager; vm_object_hash_entry_t entry; -#if MACH_PAGEMAP - vm_object_size_t size; - vm_external_map_t map; -#endif /* MACH_PAGEMAP */ - - XPR(XPR_VM_OBJECT, "vm_object_pager_create, object 0x%X\n", - (integer_t)object, 0,0,0,0); + lck_mtx_t *lck; + vm_object_t pager_object = VM_OBJECT_NULL; assert(object != kernel_object); - if (memory_manager_default_check() != KERN_SUCCESS) - return; - /* * Prevent collapse or termination by holding a paging reference */ @@ -3004,18 +4751,16 @@ vm_object_pager_create( object->pager_created = TRUE; object->paging_offset = 0; -#if MACH_PAGEMAP - size = object->size; -#endif /* MACH_PAGEMAP */ vm_object_unlock(object); -#if MACH_PAGEMAP - map = vm_external_create(size); - vm_object_lock(object); - assert(object->size == size); - object->existence_map = map; - vm_object_unlock(object); -#endif /* MACH_PAGEMAP */ + if ((uint32_t) (object->vo_size/PAGE_SIZE) != + (object->vo_size/PAGE_SIZE)) { + panic("vm_object_compressor_pager_create(%p): " + "object size 0x%llx >= 0x%llx\n", + object, + (uint64_t) object->vo_size, + 0x0FFFFFFFFULL*PAGE_SIZE); + } /* * Create the [internal] pager, and associate it with this object. @@ -3025,29 +4770,28 @@ vm_object_pager_create( * user will ever map this object. */ { - memory_object_default_t dmm; - vm_size_t cluster_size; - - /* acquire a reference for the default memory manager */ - dmm = memory_manager_default_reference(&cluster_size); - assert(cluster_size >= PAGE_SIZE); - - object->cluster_size = cluster_size; /* XXX ??? */ assert(object->temporary); /* create our new memory object */ - (void) memory_object_create(dmm, object->size, &pager); - - memory_object_default_deallocate(dmm); + assert((uint32_t) (object->vo_size/PAGE_SIZE) == + (object->vo_size/PAGE_SIZE)); + (void) compressor_memory_object_create( + (memory_object_size_t) object->vo_size, + &pager); + if (pager == NULL) { + panic("vm_object_compressor_pager_create(): " + "no pager for object %p size 0x%llx\n", + object, (uint64_t) object->vo_size); + } } entry = vm_object_hash_entry_alloc(pager); - vm_object_cache_lock(); - vm_object_hash_insert(entry); - - entry->object = object; - vm_object_cache_unlock(); + vm_object_lock(object); + lck = vm_object_hash_lock_spin(pager); + vm_object_hash_insert(entry, object); + vm_object_hash_unlock(lck); + vm_object_unlock(object); /* * A reference was returned by @@ -3055,8 +4799,11 @@ vm_object_pager_create( * copied by vm_object_enter(). */ - if (vm_object_enter(pager, object->size, TRUE, TRUE, FALSE) != object) - panic("vm_object_pager_create: mismatch"); + pager_object = vm_object_enter(pager, object->vo_size, TRUE, TRUE, FALSE); + + if (pager_object != object) { + panic("vm_object_compressor_pager_create: mismatch (pager: %p, pager_object: %p, orig_object: %p, orig_object size: 0x%llx)\n", pager, pager_object, object, (uint64_t) object->vo_size); + } /* * Drop the reference we were passed. @@ -3107,33 +4854,113 @@ static long object_bypasses = 0; static boolean_t vm_object_collapse_allowed = TRUE; static boolean_t vm_object_bypass_allowed = TRUE; -static int vm_external_discarded; -static int vm_external_collapsed; - unsigned long vm_object_collapse_encrypted = 0; -/* - * Routine: vm_object_do_collapse - * Purpose: - * Collapse an object with the object backing it. - * Pages in the backing object are moved into the - * parent, and the backing object is deallocated. - * Conditions: - * Both objects and the cache are locked; the page - * queues are unlocked. - * - */ -static void -vm_object_do_collapse( +void vm_object_do_collapse_compressor(vm_object_t object, + vm_object_t backing_object); +void +vm_object_do_collapse_compressor( vm_object_t object, vm_object_t backing_object) { - vm_page_t p, pp; vm_object_offset_t new_offset, backing_offset; vm_object_size_t size; - backing_offset = object->shadow_offset; - size = object->size; + vm_counters.do_collapse_compressor++; + + vm_object_lock_assert_exclusive(object); + vm_object_lock_assert_exclusive(backing_object); + + size = object->vo_size; + + /* + * Move all compressed pages from backing_object + * to the parent. + */ + + for (backing_offset = object->vo_shadow_offset; + backing_offset < object->vo_shadow_offset + object->vo_size; + backing_offset += PAGE_SIZE) { + memory_object_offset_t backing_pager_offset; + + /* find the next compressed page at or after this offset */ + backing_pager_offset = (backing_offset + + backing_object->paging_offset); + backing_pager_offset = vm_compressor_pager_next_compressed( + backing_object->pager, + backing_pager_offset); + if (backing_pager_offset == (memory_object_offset_t) -1) { + /* no more compressed pages */ + break; + } + backing_offset = (backing_pager_offset - + backing_object->paging_offset); + + new_offset = backing_offset - object->vo_shadow_offset; + + if (new_offset >= object->vo_size) { + /* we're out of the scope of "object": done */ + break; + } + + if ((vm_page_lookup(object, new_offset) != VM_PAGE_NULL) || + (vm_compressor_pager_state_get(object->pager, + (new_offset + + object->paging_offset)) == + VM_EXTERNAL_STATE_EXISTS)) { + /* + * This page already exists in object, resident or + * compressed. + * We don't need this compressed page in backing_object + * and it will be reclaimed when we release + * backing_object. + */ + continue; + } + + /* + * backing_object has this page in the VM compressor and + * we need to transfer it to object. + */ + vm_counters.do_collapse_compressor_pages++; + vm_compressor_pager_transfer( + /* destination: */ + object->pager, + (new_offset + object->paging_offset), + /* source: */ + backing_object->pager, + (backing_offset + backing_object->paging_offset)); + } +} + +/* + * Routine: vm_object_do_collapse + * Purpose: + * Collapse an object with the object backing it. + * Pages in the backing object are moved into the + * parent, and the backing object is deallocated. + * Conditions: + * Both objects and the cache are locked; the page + * queues are unlocked. + * + */ +static void +vm_object_do_collapse( + vm_object_t object, + vm_object_t backing_object) +{ + vm_page_t p, pp; + vm_object_offset_t new_offset, backing_offset; + vm_object_size_t size; + + vm_object_lock_assert_exclusive(object); + vm_object_lock_assert_exclusive(backing_object); + + assert(object->purgable == VM_PURGABLE_DENY); + assert(backing_object->purgable == VM_PURGABLE_DENY); + + backing_offset = object->vo_shadow_offset; + size = object->vo_size; /* * Move all in-memory pages from backing_object @@ -3142,9 +4969,9 @@ vm_object_do_collapse( * pages that shadow them. */ - while (!queue_empty(&backing_object->memq)) { + while (!vm_page_queue_empty(&backing_object->memq)) { - p = (vm_page_t) queue_first(&backing_object->memq); + p = (vm_page_t) vm_page_queue_first(&backing_object->memq); new_offset = (p->offset - backing_offset); @@ -3164,43 +4991,37 @@ vm_object_do_collapse( /* * ENCRYPTED SWAP: * The encryption key includes the "pager" and the - * "paging_offset". These might not be the same in - * the new object, so we can't just move an encrypted - * page from one object to the other. We can't just - * decrypt the page here either, because that would drop + * "paging_offset". These will not change during the + * object collapse, so we can just move an encrypted + * page from one object to the other in this case. + * We can't decrypt the page here, since we can't drop * the object lock. - * The caller should check for encrypted pages before - * attempting to collapse. */ - ASSERT_PAGE_DECRYPTED(p); - + if (p->encrypted) { + vm_object_collapse_encrypted++; + } pp = vm_page_lookup(object, new_offset); if (pp == VM_PAGE_NULL) { - /* - * Parent now has no page. - * Move the backing object's page up. - */ - - vm_page_rename(p, object, new_offset); -#if MACH_PAGEMAP - } else if (pp->absent) { - - /* - * Parent has an absent page... - * it's not being paged in, so - * it must really be missing from - * the parent. - * - * Throw out the absent page... - * any faults looking for that - * page will restart with the new - * one. - */ - - VM_PAGE_FREE(pp); - vm_page_rename(p, object, new_offset); -#endif /* MACH_PAGEMAP */ + if (VM_COMPRESSOR_PAGER_STATE_GET(object, + new_offset) + == VM_EXTERNAL_STATE_EXISTS) { + /* + * Parent object has this page + * in the VM compressor. + * Throw away the backing + * object's page. + */ + VM_PAGE_FREE(p); + } else { + /* + * Parent now has no page. + * Move the backing object's page + * up. + */ + vm_page_rename(p, object, new_offset, + TRUE); + } } else { assert(! pp->absent); @@ -3213,18 +5034,21 @@ vm_object_do_collapse( } } } - -#if !MACH_PAGEMAP - assert(!object->pager_created && object->pager == MEMORY_OBJECT_NULL - || (!backing_object->pager_created - && backing_object->pager == MEMORY_OBJECT_NULL)); -#else - assert(!object->pager_created && object->pager == MEMORY_OBJECT_NULL); -#endif /* !MACH_PAGEMAP */ - if (backing_object->pager != MEMORY_OBJECT_NULL) { + if (vm_object_collapse_compressor_allowed && + object->pager != MEMORY_OBJECT_NULL && + backing_object->pager != MEMORY_OBJECT_NULL) { + + /* move compressed pages from backing_object to object */ + vm_object_do_collapse_compressor(object, backing_object); + + } else if (backing_object->pager != MEMORY_OBJECT_NULL) { vm_object_hash_entry_t entry; + assert((!object->pager_created && + (object->pager == MEMORY_OBJECT_NULL)) || + (!backing_object->pager_created && + (backing_object->pager == MEMORY_OBJECT_NULL))); /* * Move the pager from backing_object to object. * @@ -3234,51 +5058,39 @@ vm_object_do_collapse( */ assert(!object->paging_in_progress); + assert(!object->activity_in_progress); + assert(!object->pager_created); + assert(object->pager == NULL); object->pager = backing_object->pager; - entry = vm_object_hash_lookup(object->pager, FALSE); - assert(entry != VM_OBJECT_HASH_ENTRY_NULL); - entry->object = object; + + if (backing_object->hashed) { + lck_mtx_t *lck; + + lck = vm_object_hash_lock_spin(backing_object->pager); + entry = vm_object_hash_lookup(object->pager, FALSE); + assert(entry != VM_OBJECT_HASH_ENTRY_NULL); + entry->object = object; + vm_object_hash_unlock(lck); + + object->hashed = TRUE; + } object->pager_created = backing_object->pager_created; object->pager_control = backing_object->pager_control; object->pager_ready = backing_object->pager_ready; object->pager_initialized = backing_object->pager_initialized; - object->cluster_size = backing_object->cluster_size; object->paging_offset = backing_object->paging_offset + backing_offset; if (object->pager_control != MEMORY_OBJECT_CONTROL_NULL) { memory_object_control_collapse(object->pager_control, object); } + /* the backing_object has lost its pager: reset all fields */ + backing_object->pager_created = FALSE; + backing_object->pager_control = NULL; + backing_object->pager_ready = FALSE; + backing_object->paging_offset = 0; + backing_object->pager = NULL; } - - vm_object_cache_unlock(); - -#if MACH_PAGEMAP - /* - * If the shadow offset is 0, the use the existence map from - * the backing object if there is one. If the shadow offset is - * not zero, toss it. - * - * XXX - If the shadow offset is not 0 then a bit copy is needed - * if the map is to be salvaged. For now, we just just toss the - * old map, giving the collapsed object no map. This means that - * the pager is invoked for zero fill pages. If analysis shows - * that this happens frequently and is a performance hit, then - * this code should be fixed to salvage the map. - */ - assert(object->existence_map == VM_EXTERNAL_NULL); - if (backing_offset || (size != backing_object->size)) { - vm_external_discarded++; - vm_external_destroy(backing_object->existence_map, - backing_object->size); - } - else { - vm_external_collapsed++; - object->existence_map = backing_object->existence_map; - } - backing_object->existence_map = VM_EXTERNAL_NULL; -#endif /* MACH_PAGEMAP */ - /* * Object now shadows whatever backing_object did. * Note that the reference to backing_object->shadow @@ -3289,10 +5101,13 @@ vm_object_do_collapse( assert(!backing_object->phys_contiguous); object->shadow = backing_object->shadow; if (object->shadow) { - object->shadow_offset += backing_object->shadow_offset; + object->vo_shadow_offset += backing_object->vo_shadow_offset; + /* "backing_object" gave its shadow to "object" */ + backing_object->shadow = VM_OBJECT_NULL; + backing_object->vo_shadow_offset = 0; } else { /* no shadow, therefore no shadow offset... */ - object->shadow_offset = 0; + object->vo_shadow_offset = 0; } assert((object->shadow == VM_OBJECT_NULL) || (object->shadow->copy != backing_object)); @@ -3304,20 +5119,43 @@ vm_object_do_collapse( * pager left, and no object references within it, * all that is necessary is to dispose of it. */ + object_collapses++; - assert((backing_object->ref_count == 1) && - (backing_object->resident_page_count == 0) && - (backing_object->paging_in_progress == 0)); + assert(backing_object->ref_count == 1); + assert(backing_object->resident_page_count == 0); + assert(backing_object->paging_in_progress == 0); + assert(backing_object->activity_in_progress == 0); + assert(backing_object->shadow == VM_OBJECT_NULL); + assert(backing_object->vo_shadow_offset == 0); + + if (backing_object->pager != MEMORY_OBJECT_NULL) { + /* ... unless it has a pager; need to terminate pager too */ + vm_counters.do_collapse_terminate++; + if (vm_object_terminate(backing_object) != KERN_SUCCESS) { + vm_counters.do_collapse_terminate_failure++; + } + return; + } + + assert(backing_object->pager == NULL); backing_object->alive = FALSE; vm_object_unlock(backing_object); XPR(XPR_VM_OBJECT, "vm_object_collapse, collapsed 0x%X\n", - (integer_t)backing_object, 0,0,0,0); + backing_object, 0,0,0,0); + +#if VM_OBJECT_TRACKING + if (vm_object_tracking_inited) { + btlog_remove_entries_for_element(vm_object_tracking_btlog, + backing_object); + } +#endif /* VM_OBJECT_TRACKING */ + + vm_object_lock_destroy(backing_object); zfree(vm_object_zone, backing_object); - object_collapses++; } static void @@ -3330,6 +5168,9 @@ vm_object_do_bypass( * in the chain. */ + vm_object_lock_assert_exclusive(object); + vm_object_lock_assert_exclusive(backing_object); + #if TASK_SWAPPER /* * Do object reference in-line to @@ -3340,6 +5181,7 @@ vm_object_do_bypass( */ if (backing_object->shadow != VM_OBJECT_NULL) { vm_object_lock(backing_object->shadow); + vm_object_lock_assert_exclusive(backing_object->shadow); backing_object->shadow->ref_count++; if (object->res_count != 0) vm_object_res_reference(backing_object->shadow); @@ -3353,10 +5195,10 @@ vm_object_do_bypass( assert(!backing_object->phys_contiguous); object->shadow = backing_object->shadow; if (object->shadow) { - object->shadow_offset += backing_object->shadow_offset; + object->vo_shadow_offset += backing_object->vo_shadow_offset; } else { /* no shadow, therefore no shadow offset... */ - object->shadow_offset = 0; + object->vo_shadow_offset = 0; } /* @@ -3373,7 +5215,7 @@ vm_object_do_bypass( * Since its ref_count was at least 2, it * will not vanish; so we don't need to call * vm_object_deallocate. - * [FBDP: that doesn't seem to be true any more] + * [with a caveat for "named" objects] * * The res_count on the backing object is * conditionally decremented. It's possible @@ -3391,7 +5233,9 @@ vm_object_do_bypass( * is temporary and cachable. #endif */ - if (backing_object->ref_count > 1) { + if (backing_object->ref_count > 2 || + (!backing_object->named && backing_object->ref_count > 1)) { + vm_object_lock_assert_exclusive(backing_object); backing_object->ref_count--; #if TASK_SWAPPER if (object->res_count != 0) @@ -3412,7 +5256,19 @@ vm_object_do_bypass( vm_object_res_reference(backing_object); } #endif /* TASK_SWAPPER */ + /* + * vm_object_collapse (the caller of this function) is + * now called from contexts that may not guarantee that a + * valid reference is held on the object... w/o a valid + * reference, it is unsafe and unwise (you will definitely + * regret it) to unlock the object and then retake the lock + * since the object may be terminated and recycled in between. + * The "activity_in_progress" reference will keep the object + * 'stable'. + */ + vm_object_activity_begin(object); vm_object_unlock(object); + vm_object_unlock(backing_object); vm_object_deallocate(backing_object); @@ -3424,6 +5280,7 @@ vm_object_do_bypass( */ vm_object_lock(object); + vm_object_activity_end(object); } object_bypasses++; @@ -3444,19 +5301,19 @@ static unsigned long vm_object_collapse_calls = 0; static unsigned long vm_object_collapse_objects = 0; static unsigned long vm_object_collapse_do_collapse = 0; static unsigned long vm_object_collapse_do_bypass = 0; + __private_extern__ void vm_object_collapse( - register vm_object_t object, - register vm_object_offset_t hint_offset, + vm_object_t object, + vm_object_offset_t hint_offset, boolean_t can_bypass) { - register vm_object_t backing_object; - register unsigned int rcount; - register unsigned int size; - vm_object_offset_t collapse_min_offset; - vm_object_offset_t collapse_max_offset; - vm_page_t page; + vm_object_t backing_object; + unsigned int rcount; + unsigned int size; vm_object_t original_object; + int object_lock_type; + int backing_object_lock_type; vm_object_collapse_calls++; @@ -3466,13 +5323,26 @@ vm_object_collapse( } XPR(XPR_VM_OBJECT, "vm_object_collapse, obj 0x%X\n", - (integer_t)object, 0,0,0,0); + object, 0,0,0,0); if (object == VM_OBJECT_NULL) return; original_object = object; + /* + * The top object was locked "exclusive" by the caller. + * In the first pass, to determine if we can collapse the shadow chain, + * take a "shared" lock on the shadow objects. If we can collapse, + * we'll have to go down the chain again with exclusive locks. + */ + object_lock_type = OBJECT_LOCK_EXCLUSIVE; + backing_object_lock_type = OBJECT_LOCK_SHARED; + +retry: + object = original_object; + vm_object_lock_assert_exclusive(object); + while (TRUE) { vm_object_collapse_objects++; /* @@ -3491,24 +5361,27 @@ vm_object_collapse( } return; } - + if (backing_object_lock_type == OBJECT_LOCK_SHARED) { + vm_object_lock_shared(backing_object); + } else { + vm_object_lock(backing_object); + } + /* * No pages in the object are currently * being paged out, and */ if (object->paging_in_progress != 0 || - object->absent_count != 0) { + object->activity_in_progress != 0) { /* try and collapse the rest of the shadow chain */ - vm_object_lock(backing_object); if (object != original_object) { vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; continue; } - vm_object_lock(backing_object); - /* * ... * The backing object is not read_only, @@ -3519,12 +5392,37 @@ vm_object_collapse( */ if (!backing_object->internal || - backing_object->paging_in_progress != 0) { + backing_object->paging_in_progress != 0 || + backing_object->activity_in_progress != 0) { + /* try and collapse the rest of the shadow chain */ + if (object != original_object) { + vm_object_unlock(object); + } + object = backing_object; + object_lock_type = backing_object_lock_type; + continue; + } + + /* + * Purgeable objects are not supposed to engage in + * copy-on-write activities, so should not have + * any shadow objects or be a shadow object to another + * object. + * Collapsing a purgeable object would require some + * updates to the purgeable compressed ledgers. + */ + if (object->purgable != VM_PURGABLE_DENY || + backing_object->purgable != VM_PURGABLE_DENY) { + panic("vm_object_collapse() attempting to collapse " + "purgeable object: %p(%d) %p(%d)\n", + object, object->purgable, + backing_object, backing_object->purgable); /* try and collapse the rest of the shadow chain */ if (object != original_object) { vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; continue; } @@ -3545,6 +5443,7 @@ vm_object_collapse( vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; continue; } @@ -3558,75 +5457,46 @@ vm_object_collapse( * object, we may be able to collapse it into the * parent. * - * If MACH_PAGEMAP is defined: - * The parent must not have a pager created for it, - * since collapsing a backing_object dumps new pages - * into the parent that its pager doesn't know about - * (and the collapse code can't merge the existence - * maps). - * Otherwise: * As long as one of the objects is still not known * to the pager, we can collapse them. */ if (backing_object->ref_count == 1 && - (!object->pager_created -#if !MACH_PAGEMAP - || !backing_object->pager_created -#endif /*!MACH_PAGEMAP */ + (vm_object_collapse_compressor_allowed || + !object->pager_created + || (!backing_object->pager_created) ) && vm_object_collapse_allowed) { - XPR(XPR_VM_OBJECT, - "vm_object_collapse: %x to %x, pager %x, pager_control %x\n", - (integer_t)backing_object, (integer_t)object, - (integer_t)backing_object->pager, - (integer_t)backing_object->pager_control, 0); - /* - * We need the cache lock for collapsing, - * but we must not deadlock. + * We need the exclusive lock on the VM objects. */ - - if (! vm_object_cache_lock_try()) { - if (object != original_object) { - vm_object_unlock(object); - } + if (backing_object_lock_type != OBJECT_LOCK_EXCLUSIVE) { + /* + * We have an object and its shadow locked + * "shared". We can't just upgrade the locks + * to "exclusive", as some other thread might + * also have these objects locked "shared" and + * attempt to upgrade one or the other to + * "exclusive". The upgrades would block + * forever waiting for the other "shared" locks + * to get released. + * So we have to release the locks and go + * down the shadow chain again (since it could + * have changed) with "exclusive" locking. + */ vm_object_unlock(backing_object); - return; + if (object != original_object) + vm_object_unlock(object); + object_lock_type = OBJECT_LOCK_EXCLUSIVE; + backing_object_lock_type = OBJECT_LOCK_EXCLUSIVE; + goto retry; } - /* - * ENCRYPTED SWAP - * We can't collapse the object if it contains - * any encypted page, because the encryption key - * includes the info. We can't - * drop the object lock in vm_object_do_collapse() - * so we can't decrypt the page there either. - */ - if (vm_pages_encrypted) { - collapse_min_offset = object->shadow_offset; - collapse_max_offset = - object->shadow_offset + object->size; - queue_iterate(&backing_object->memq, - page, vm_page_t, listq) { - if (page->encrypted && - (page->offset >= - collapse_min_offset) && - (page->offset < - collapse_max_offset)) { - /* - * We found an encrypted page - * in the backing object, - * within the range covered - * by the parent object: we can - * not collapse them. - */ - vm_object_collapse_encrypted++; - vm_object_cache_unlock(); - goto try_bypass; - } - } - } - + XPR(XPR_VM_OBJECT, + "vm_object_collapse: %x to %x, pager %x, pager_control %x\n", + backing_object, object, + backing_object->pager, + backing_object->pager_control, 0); + /* * Collapse the object with its backing * object, and try again with the object's @@ -3638,7 +5508,6 @@ vm_object_collapse( continue; } - try_bypass: /* * Collapsing the backing object was not possible * or permitted, so let's try bypassing it. @@ -3650,6 +5519,7 @@ vm_object_collapse( vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; continue; } @@ -3659,29 +5529,26 @@ vm_object_collapse( * we have to make sure no pages in the backing object * "show through" before bypassing it. */ - size = atop(object->size); + size = (unsigned int)atop(object->vo_size); rcount = object->resident_page_count; + if (rcount != size) { vm_object_offset_t offset; vm_object_offset_t backing_offset; unsigned int backing_rcount; - unsigned int lookups = 0; /* * If the backing object has a pager but no pagemap, * then we cannot bypass it, because we don't know * what pages it has. */ - if (backing_object->pager_created -#if MACH_PAGEMAP - && (backing_object->existence_map == VM_EXTERNAL_NULL) -#endif /* MACH_PAGEMAP */ - ) { + if (backing_object->pager_created) { /* try and collapse the rest of the shadow chain */ if (object != original_object) { vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; continue; } @@ -3690,16 +5557,31 @@ vm_object_collapse( * then we cannot bypass it, because we don't know * what pages it has. */ - if (object->pager_created -#if MACH_PAGEMAP - && (object->existence_map == VM_EXTERNAL_NULL) -#endif /* MACH_PAGEMAP */ - ) { + if (object->pager_created) { /* try and collapse the rest of the shadow chain */ if (object != original_object) { vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; + continue; + } + + backing_offset = object->vo_shadow_offset; + backing_rcount = backing_object->resident_page_count; + + if ( (int)backing_rcount - (int)(atop(backing_object->vo_size) - size) > (int)rcount) { + /* + * we have enough pages in the backing object to guarantee that + * at least 1 of them must be 'uncovered' by a resident page + * in the object we're evaluating, so move on and + * try to collapse the rest of the shadow chain + */ + if (object != original_object) { + vm_object_unlock(object); + } + object = backing_object; + object_lock_type = backing_object_lock_type; continue; } @@ -3716,13 +5598,10 @@ vm_object_collapse( * */ - backing_offset = object->shadow_offset; - backing_rcount = backing_object->resident_page_count; - -#define EXISTS_IN_OBJECT(obj, off, rc) \ - (vm_external_state_get((obj)->existence_map, \ - (vm_offset_t)(off)) == VM_EXTERNAL_STATE_EXISTS || \ - ((rc) && ++lookups && vm_page_lookup((obj), (off)) != VM_PAGE_NULL && (rc)--)) +#define EXISTS_IN_OBJECT(obj, off, rc) \ + ((VM_COMPRESSOR_PAGER_STATE_GET((obj), (off)) \ + == VM_EXTERNAL_STATE_EXISTS) || \ + ((rc) && vm_page_lookup((obj), (off)) != VM_PAGE_NULL && (rc)--)) /* * Check the hint location first @@ -3738,12 +5617,13 @@ vm_object_collapse( backing_offset, backing_rcount) && !EXISTS_IN_OBJECT(object, hint_offset, rcount)) { /* dependency right at the hint */ - object->cow_hint = (vm_offset_t)hint_offset; + object->cow_hint = (vm_offset_t) hint_offset; /* atomic */ /* try and collapse the rest of the shadow chain */ if (object != original_object) { vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; continue; } @@ -3753,38 +5633,32 @@ vm_object_collapse( * pages in the backing object, it makes sense to * walk the backing_object's resident pages first. * - * NOTE: Pages may be in both the existence map and - * resident. So, we can't permanently decrement - * the rcount here because the second loop may - * find the same pages in the backing object' - * existence map that we found here and we would - * double-decrement the rcount. We also may or - * may not have found the + * NOTE: Pages may be in both the existence map and/or + * resident, so if we don't find a dependency while + * walking the backing object's resident page list + * directly, and there is an existence map, we'll have + * to run the offset based 2nd pass. Because we may + * have to run both passes, we need to be careful + * not to decrement 'rcount' in the 1st pass */ - if (backing_rcount && size > - ((backing_object->existence_map) ? - backing_rcount : (backing_rcount >> 1))) { + if (backing_rcount && backing_rcount < (size / 8)) { unsigned int rc = rcount; vm_page_t p; backing_rcount = backing_object->resident_page_count; - p = (vm_page_t)queue_first(&backing_object->memq); + p = (vm_page_t)vm_page_queue_first(&backing_object->memq); do { - /* Until we get more than one lookup lock */ - if (lookups > 256) { - lookups = 0; - delay(1); - } - offset = (p->offset - backing_offset); - if (offset < object->size && + + if (offset < object->vo_size && offset != hint_offset && !EXISTS_IN_OBJECT(object, offset, rc)) { /* found a dependency */ - object->cow_hint = (vm_offset_t)offset; + object->cow_hint = (vm_offset_t) offset; /* atomic */ + break; } - p = (vm_page_t) queue_next(&p->listq); + p = (vm_page_t) vm_page_queue_next(&p->listq); } while (--backing_rcount); if (backing_rcount != 0 ) { @@ -3793,6 +5667,7 @@ vm_object_collapse( vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; continue; } } @@ -3801,24 +5676,18 @@ vm_object_collapse( * Walk through the offsets looking for pages in the * backing object that show through to the object. */ - if (backing_rcount || backing_object->existence_map) { + if (backing_rcount) { offset = hint_offset; while((offset = - (offset + PAGE_SIZE_64 < object->size) ? + (offset + PAGE_SIZE_64 < object->vo_size) ? (offset + PAGE_SIZE_64) : 0) != hint_offset) { - /* Until we get more than one lookup lock */ - if (lookups > 256) { - lookups = 0; - delay(1); - } - if (EXISTS_IN_OBJECT(backing_object, offset + backing_offset, backing_rcount) && !EXISTS_IN_OBJECT(object, offset, rcount)) { /* found a dependency */ - object->cow_hint = (vm_offset_t)offset; + object->cow_hint = (vm_offset_t) offset; /* atomic */ break; } } @@ -3828,11 +5697,24 @@ vm_object_collapse( vm_object_unlock(object); } object = backing_object; + object_lock_type = backing_object_lock_type; continue; } } } + /* + * We need "exclusive" locks on the 2 VM objects. + */ + if (backing_object_lock_type != OBJECT_LOCK_EXCLUSIVE) { + vm_object_unlock(backing_object); + if (object != original_object) + vm_object_unlock(object); + object_lock_type = OBJECT_LOCK_EXCLUSIVE; + backing_object_lock_type = OBJECT_LOCK_EXCLUSIVE; + goto retry; + } + /* reset the offset hint for any objects deeper in the chain */ object->cow_hint = (vm_offset_t)0; @@ -3852,9 +5734,12 @@ vm_object_collapse( continue; } + /* NOT REACHED */ + /* if (object != original_object) { vm_object_unlock(object); } + */ } /* @@ -3873,11 +5758,11 @@ unsigned int vm_object_page_remove_iterate = 0; __private_extern__ void vm_object_page_remove( - register vm_object_t object, - register vm_object_offset_t start, - register vm_object_offset_t end) + vm_object_t object, + vm_object_offset_t start, + vm_object_offset_t end) { - register vm_page_t p, next; + vm_page_t p, next; /* * One and two page removals are most popular. @@ -3891,22 +5776,22 @@ vm_object_page_remove( for (; start < end; start += PAGE_SIZE_64) { p = vm_page_lookup(object, start); if (p != VM_PAGE_NULL) { - assert(!p->cleaning && !p->pageout); - if (!p->fictitious) - pmap_disconnect(p->phys_page); + assert(!p->cleaning && !p->laundry); + if (!p->fictitious && p->pmapped) + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(p)); VM_PAGE_FREE(p); } } } else { vm_object_page_remove_iterate++; - p = (vm_page_t) queue_first(&object->memq); - while (!queue_end(&object->memq, (queue_entry_t) p)) { - next = (vm_page_t) queue_next(&p->listq); + p = (vm_page_t) vm_page_queue_first(&object->memq); + while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t) p)) { + next = (vm_page_t) vm_page_queue_next(&p->listq); if ((start <= p->offset) && (p->offset < end)) { - assert(!p->cleaning && !p->pageout); - if (!p->fictitious) - pmap_disconnect(p->phys_page); + assert(!p->cleaning && !p->laundry); + if (!p->fictitious && p->pmapped) + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(p)); VM_PAGE_FREE(p); } p = next; @@ -3942,7 +5827,7 @@ static int vm_object_coalesce_count = 0; __private_extern__ boolean_t vm_object_coalesce( - register vm_object_t prev_object, + vm_object_t prev_object, vm_object_t next_object, vm_object_offset_t prev_offset, __unused vm_object_offset_t next_offset, @@ -3965,7 +5850,7 @@ vm_object_coalesce( XPR(XPR_VM_OBJECT, "vm_object_coalesce: 0x%X prev_off 0x%X prev_size 0x%X next_size 0x%X\n", - (integer_t)prev_object, prev_offset, prev_size, next_size, 0); + prev_object, prev_offset, prev_size, next_size, 0); vm_object_lock(prev_object); @@ -3981,7 +5866,7 @@ vm_object_coalesce( * . paged out * . shadows another object * . has a copy elsewhere - * . is purgable + * . is purgeable * . paging references (pages might be in page-list) */ @@ -3990,8 +5875,9 @@ vm_object_coalesce( (prev_object->shadow != VM_OBJECT_NULL) || (prev_object->copy != VM_OBJECT_NULL) || (prev_object->true_share != FALSE) || - (prev_object->purgable != VM_OBJECT_NONPURGABLE) || - (prev_object->paging_in_progress != 0)) { + (prev_object->purgable != VM_PURGABLE_DENY) || + (prev_object->paging_in_progress != 0) || + (prev_object->activity_in_progress != 0)) { vm_object_unlock(prev_object); return(FALSE); } @@ -4010,502 +5896,113 @@ vm_object_coalesce( * Extend the object if necessary. */ newsize = prev_offset + prev_size + next_size; - if (newsize > prev_object->size) { -#if MACH_PAGEMAP - /* - * We cannot extend an object that has existence info, - * since the existence info might then fail to cover - * the entire object. - * - * This assertion must be true because the object - * has no pager, and we only create existence info - * for objects with pagers. - */ - assert(prev_object->existence_map == VM_EXTERNAL_NULL); -#endif /* MACH_PAGEMAP */ - prev_object->size = newsize; + if (newsize > prev_object->vo_size) { + prev_object->vo_size = newsize; } vm_object_unlock(prev_object); return(TRUE); } -/* - * Attach a set of physical pages to an object, so that they can - * be mapped by mapping the object. Typically used to map IO memory. - * - * The mapping function and its private data are used to obtain the - * physical addresses for each page to be mapped. - */ -void -vm_object_page_map( - vm_object_t object, - vm_object_offset_t offset, - vm_object_size_t size, - vm_object_offset_t (*map_fn)(void *map_fn_data, - vm_object_offset_t offset), - void *map_fn_data) /* private to map_fn */ +kern_return_t +vm_object_populate_with_private( + vm_object_t object, + vm_object_offset_t offset, + ppnum_t phys_page, + vm_size_t size) { - int num_pages; - int i; - vm_page_t m; - vm_page_t old_page; - vm_object_offset_t addr; - - num_pages = atop_64(size); - - for (i = 0; i < num_pages; i++, offset += PAGE_SIZE_64) { + ppnum_t base_page; + vm_object_offset_t base_offset; - addr = (*map_fn)(map_fn_data, offset); - while ((m = vm_page_grab_fictitious()) == VM_PAGE_NULL) - vm_page_more_fictitious(); + if (!object->private) + return KERN_FAILURE; - vm_object_lock(object); - if ((old_page = vm_page_lookup(object, offset)) - != VM_PAGE_NULL) - { - vm_page_lock_queues(); - vm_page_free(old_page); - vm_page_unlock_queues(); - } + base_page = phys_page; - vm_page_init(m, addr); - /* private normally requires lock_queues but since we */ - /* are initializing the page, its not necessary here */ - m->private = TRUE; /* don`t free page */ - m->wire_count = 1; - vm_page_insert(m, object, offset); + vm_object_lock(object); - PAGE_WAKEUP_DONE(m); - vm_object_unlock(object); - } -} + if (!object->phys_contiguous) { + vm_page_t m; -#include + if ((base_offset = trunc_page_64(offset)) != offset) { + vm_object_unlock(object); + return KERN_FAILURE; + } + base_offset += object->paging_offset; -#if MACH_KDB -#include -#include + while (size) { + m = vm_page_lookup(object, base_offset); -#define printf kdbprintf + if (m != VM_PAGE_NULL) { + if (m->fictitious) { + if (VM_PAGE_GET_PHYS_PAGE(m) != vm_page_guard_addr) { -extern boolean_t vm_object_cached( - vm_object_t object); + vm_page_lockspin_queues(); + m->private = TRUE; + vm_page_unlock_queues(); -extern void print_bitstring( - char byte); + m->fictitious = FALSE; + VM_PAGE_SET_PHYS_PAGE(m, base_page); + } + } else if (VM_PAGE_GET_PHYS_PAGE(m) != base_page) { -boolean_t vm_object_print_pages = FALSE; + if ( !m->private) { + /* + * we'd leak a real page... that can't be right + */ + panic("vm_object_populate_with_private - %p not private", m); + } + if (m->pmapped) { + /* + * pmap call to clear old mapping + */ + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m)); + } + VM_PAGE_SET_PHYS_PAGE(m, base_page); + } + if (m->encrypted) { + /* + * we should never see this on a ficticious or private page + */ + panic("vm_object_populate_with_private - %p encrypted", m); + } -void -print_bitstring( - char byte) -{ - printf("%c%c%c%c%c%c%c%c", - ((byte & (1 << 0)) ? '1' : '0'), - ((byte & (1 << 1)) ? '1' : '0'), - ((byte & (1 << 2)) ? '1' : '0'), - ((byte & (1 << 3)) ? '1' : '0'), - ((byte & (1 << 4)) ? '1' : '0'), - ((byte & (1 << 5)) ? '1' : '0'), - ((byte & (1 << 6)) ? '1' : '0'), - ((byte & (1 << 7)) ? '1' : '0')); -} + } else { + while ((m = vm_page_grab_fictitious()) == VM_PAGE_NULL) + vm_page_more_fictitious(); -boolean_t -vm_object_cached( - register vm_object_t object) -{ - register vm_object_t o; + /* + * private normally requires lock_queues but since we + * are initializing the page, its not necessary here + */ + m->private = TRUE; + m->fictitious = FALSE; + VM_PAGE_SET_PHYS_PAGE(m, base_page); + m->unusual = TRUE; + m->busy = FALSE; - queue_iterate(&vm_object_cached_list, o, vm_object_t, cached_list) { - if (object == o) { - return TRUE; + vm_page_insert(m, object, base_offset); + } + base_page++; /* Go to the next physical page */ + base_offset += PAGE_SIZE; + size -= PAGE_SIZE; } - } - return FALSE; -} - -#if MACH_PAGEMAP -/* - * vm_external_print: [ debug ] - */ -void -vm_external_print( - vm_external_map_t emap, - vm_size_t size) -{ - if (emap == VM_EXTERNAL_NULL) { - printf("0 "); } else { - vm_size_t existence_size = stob(size); - printf("{ size=%d, map=[", existence_size); - if (existence_size > 0) { - print_bitstring(emap[0]); - } - if (existence_size > 1) { - print_bitstring(emap[1]); - } - if (existence_size > 2) { - printf("..."); - print_bitstring(emap[existence_size-1]); - } - printf("] }\n"); - } - return; -} -#endif /* MACH_PAGEMAP */ - -int -vm_follow_object( - vm_object_t object) -{ - int count = 0; - int orig_db_indent = db_indent; - - while (TRUE) { - if (object == VM_OBJECT_NULL) { - db_indent = orig_db_indent; - return count; - } - - count += 1; - - iprintf("object 0x%x", object); - printf(", shadow=0x%x", object->shadow); - printf(", copy=0x%x", object->copy); - printf(", pager=0x%x", object->pager); - printf(", ref=%d\n", object->ref_count); + /* NOTE: we should check the original settings here */ + /* if we have a size > zero a pmap call should be made */ + /* to disable the range */ - db_indent += 2; - object = object->shadow; + /* pmap_? */ + + /* shadows on contiguous memory are not allowed */ + /* we therefore can use the offset field */ + object->vo_shadow_offset = (vm_object_offset_t)phys_page << PAGE_SHIFT; + object->vo_size = size; } + vm_object_unlock(object); -} - -/* - * vm_object_print: [ debug ] - */ -void -vm_object_print( - db_addr_t db_addr, - __unused boolean_t have_addr, - __unused int arg_count, - __unused char *modif) -{ - vm_object_t object; - register vm_page_t p; - const char *s; - - register int count; - - object = (vm_object_t) (long) db_addr; - if (object == VM_OBJECT_NULL) - return; - - iprintf("object 0x%x\n", object); - - db_indent += 2; - - iprintf("size=0x%x", object->size); - printf(", cluster=0x%x", object->cluster_size); - printf(", memq_hint=%p", object->memq_hint); - printf(", ref_count=%d\n", object->ref_count); - iprintf(""); -#if TASK_SWAPPER - printf("res_count=%d, ", object->res_count); -#endif /* TASK_SWAPPER */ - printf("resident_page_count=%d\n", object->resident_page_count); - - iprintf("shadow=0x%x", object->shadow); - if (object->shadow) { - register int i = 0; - vm_object_t shadow = object; - while((shadow = shadow->shadow)) - i++; - printf(" (depth %d)", i); - } - printf(", copy=0x%x", object->copy); - printf(", shadow_offset=0x%x", object->shadow_offset); - printf(", last_alloc=0x%x\n", object->last_alloc); - - iprintf("pager=0x%x", object->pager); - printf(", paging_offset=0x%x", object->paging_offset); - printf(", pager_control=0x%x\n", object->pager_control); - - iprintf("copy_strategy=%d[", object->copy_strategy); - switch (object->copy_strategy) { - case MEMORY_OBJECT_COPY_NONE: - printf("copy_none"); - break; - - case MEMORY_OBJECT_COPY_CALL: - printf("copy_call"); - break; - - case MEMORY_OBJECT_COPY_DELAY: - printf("copy_delay"); - break; - - case MEMORY_OBJECT_COPY_SYMMETRIC: - printf("copy_symmetric"); - break; - - case MEMORY_OBJECT_COPY_INVALID: - printf("copy_invalid"); - break; - - default: - printf("?"); - } - printf("]"); - printf(", absent_count=%d\n", object->absent_count); - - iprintf("all_wanted=0x%x<", object->all_wanted); - s = ""; - if (vm_object_wanted(object, VM_OBJECT_EVENT_INITIALIZED)) { - printf("%sinit", s); - s = ","; - } - if (vm_object_wanted(object, VM_OBJECT_EVENT_PAGER_READY)) { - printf("%sready", s); - s = ","; - } - if (vm_object_wanted(object, VM_OBJECT_EVENT_PAGING_IN_PROGRESS)) { - printf("%spaging", s); - s = ","; - } - if (vm_object_wanted(object, VM_OBJECT_EVENT_ABSENT_COUNT)) { - printf("%sabsent", s); - s = ","; - } - if (vm_object_wanted(object, VM_OBJECT_EVENT_LOCK_IN_PROGRESS)) { - printf("%slock", s); - s = ","; - } - if (vm_object_wanted(object, VM_OBJECT_EVENT_UNCACHING)) { - printf("%suncaching", s); - s = ","; - } - if (vm_object_wanted(object, VM_OBJECT_EVENT_COPY_CALL)) { - printf("%scopy_call", s); - s = ","; - } - if (vm_object_wanted(object, VM_OBJECT_EVENT_CACHING)) { - printf("%scaching", s); - s = ","; - } - printf(">"); - printf(", paging_in_progress=%d\n", object->paging_in_progress); - - iprintf("%screated, %sinit, %sready, %spersist, %strusted, %spageout, %s, %s\n", - (object->pager_created ? "" : "!"), - (object->pager_initialized ? "" : "!"), - (object->pager_ready ? "" : "!"), - (object->can_persist ? "" : "!"), - (object->pager_trusted ? "" : "!"), - (object->pageout ? "" : "!"), - (object->internal ? "internal" : "external"), - (object->temporary ? "temporary" : "permanent")); - iprintf("%salive, %spurgable, %spurgable_volatile, %spurgable_empty, %sshadowed, %scached, %sprivate\n", - (object->alive ? "" : "!"), - ((object->purgable != VM_OBJECT_NONPURGABLE) ? "" : "!"), - ((object->purgable == VM_OBJECT_PURGABLE_VOLATILE) ? "" : "!"), - ((object->purgable == VM_OBJECT_PURGABLE_EMPTY) ? "" : "!"), - (object->shadowed ? "" : "!"), - (vm_object_cached(object) ? "" : "!"), - (object->private ? "" : "!")); - iprintf("%sadvisory_pageout, %ssilent_overwrite\n", - (object->advisory_pageout ? "" : "!"), - (object->silent_overwrite ? "" : "!")); - -#if MACH_PAGEMAP - iprintf("existence_map="); - vm_external_print(object->existence_map, object->size); -#endif /* MACH_PAGEMAP */ -#if MACH_ASSERT - iprintf("paging_object=0x%x\n", object->paging_object); -#endif /* MACH_ASSERT */ - - if (vm_object_print_pages) { - count = 0; - p = (vm_page_t) queue_first(&object->memq); - while (!queue_end(&object->memq, (queue_entry_t) p)) { - if (count == 0) { - iprintf("memory:="); - } else if (count == 2) { - printf("\n"); - iprintf(" ..."); - count = 0; - } else { - printf(","); - } - count++; - - printf("(off=0x%llX,page=%p)", p->offset, p); - p = (vm_page_t) queue_next(&p->listq); - } - if (count != 0) { - printf("\n"); - } - } - db_indent -= 2; -} - - -/* - * vm_object_find [ debug ] - * - * Find all tasks which reference the given vm_object. - */ - -boolean_t vm_object_find(vm_object_t object); -boolean_t vm_object_print_verbose = FALSE; - -boolean_t -vm_object_find( - vm_object_t object) -{ - task_t task; - vm_map_t map; - vm_map_entry_t entry; - processor_set_t pset = &default_pset; - boolean_t found = FALSE; - - queue_iterate(&pset->tasks, task, task_t, pset_tasks) { - map = task->map; - for (entry = vm_map_first_entry(map); - entry && entry != vm_map_to_entry(map); - entry = entry->vme_next) { - - vm_object_t obj; - - /* - * For the time being skip submaps, - * only the kernel can have submaps, - * and unless we are interested in - * kernel objects, we can simply skip - * submaps. See sb/dejan/nmk18b7/src/mach_kernel/vm - * for a full solution. - */ - if (entry->is_sub_map) - continue; - if (entry) - obj = entry->object.vm_object; - else - continue; - - while (obj != VM_OBJECT_NULL) { - if (obj == object) { - if (!found) { - printf("TASK\t\tMAP\t\tENTRY\n"); - found = TRUE; - } - printf("0x%x\t0x%x\t0x%x\n", - task, map, entry); - } - obj = obj->shadow; - } - } - } - - return(found); -} - -#endif /* MACH_KDB */ - -kern_return_t -vm_object_populate_with_private( - vm_object_t object, - vm_object_offset_t offset, - ppnum_t phys_page, - vm_size_t size) -{ - ppnum_t base_page; - vm_object_offset_t base_offset; - - - if(!object->private) - return KERN_FAILURE; - - base_page = phys_page; - - vm_object_lock(object); - if(!object->phys_contiguous) { - vm_page_t m; - if((base_offset = trunc_page_64(offset)) != offset) { - vm_object_unlock(object); - return KERN_FAILURE; - } - base_offset += object->paging_offset; - while(size) { - m = vm_page_lookup(object, base_offset); - if(m != VM_PAGE_NULL) { - if(m->fictitious) { - vm_page_lock_queues(); - m->fictitious = FALSE; - m->private = TRUE; - m->phys_page = base_page; - if(!m->busy) { - m->busy = TRUE; - } - if(!m->absent) { - m->absent = TRUE; - object->absent_count++; - } - m->list_req_pending = TRUE; - vm_page_unlock_queues(); - } else if (m->phys_page != base_page) { - /* pmap call to clear old mapping */ - pmap_disconnect(m->phys_page); - m->phys_page = base_page; - } - - /* - * ENCRYPTED SWAP: - * We're not pointing to the same - * physical page any longer and the - * contents of the new one are not - * supposed to be encrypted. - * XXX What happens to the original - * physical page. Is it lost ? - */ - m->encrypted = FALSE; - - } else { - while ((m = vm_page_grab_fictitious()) - == VM_PAGE_NULL) - vm_page_more_fictitious(); - vm_page_lock_queues(); - m->fictitious = FALSE; - m->private = TRUE; - m->phys_page = base_page; - m->list_req_pending = TRUE; - m->absent = TRUE; - m->unusual = TRUE; - object->absent_count++; - vm_page_unlock_queues(); - vm_page_insert(m, object, base_offset); - } - base_page++; /* Go to the next physical page */ - base_offset += PAGE_SIZE; - size -= PAGE_SIZE; - } - } else { - /* NOTE: we should check the original settings here */ - /* if we have a size > zero a pmap call should be made */ - /* to disable the range */ - - /* pmap_? */ - - /* shadows on contiguous memory are not allowed */ - /* we therefore can use the offset field */ - object->shadow_offset = (vm_object_offset_t)(phys_page << 12); - object->size = size; - } - vm_object_unlock(object); - return KERN_SUCCESS; + return KERN_SUCCESS; } /* @@ -4523,13 +6020,13 @@ vm_object_populate_with_private( __private_extern__ kern_return_t memory_object_free_from_cache( __unused host_t host, - memory_object_pager_ops_t pager_ops, + __unused memory_object_pager_ops_t pager_ops, int *count) { - +#if VM_OBJECT_CACHE int object_released = 0; - register vm_object_t object = VM_OBJECT_NULL; + vm_object_t object = VM_OBJECT_NULL; vm_object_t shadow; /* @@ -4549,6 +6046,7 @@ memory_object_free_from_cache( vm_object_t, cached_list); vm_object_cached_count--; + vm_object_cache_unlock(); /* * Since this object is in the cache, we know * that it is initialized and has only a pager's @@ -4558,6 +6056,7 @@ memory_object_free_from_cache( assert(object->pager_initialized); assert(object->ref_count == 0); + vm_object_lock_assert_exclusive(object); object->ref_count++; /* @@ -4570,6 +6069,7 @@ memory_object_free_from_cache( * (We are careful here to limit recursion.) */ shadow = object->pageout?VM_OBJECT_NULL:object->shadow; + if ((vm_object_terminate(object) == KERN_SUCCESS) && (shadow != VM_OBJECT_NULL)) { vm_object_deallocate(shadow); @@ -4582,6 +6082,9 @@ memory_object_free_from_cache( } vm_object_cache_unlock(); *count = object_released; +#else + *count = 0; +#endif return KERN_SUCCESS; } @@ -4595,21 +6098,22 @@ memory_object_create_named( { vm_object_t object; vm_object_hash_entry_t entry; + lck_mtx_t *lck; *control = MEMORY_OBJECT_CONTROL_NULL; if (pager == MEMORY_OBJECT_NULL) return KERN_INVALID_ARGUMENT; - vm_object_cache_lock(); + lck = vm_object_hash_lock_spin(pager); entry = vm_object_hash_lookup(pager, FALSE); + if ((entry != VM_OBJECT_HASH_ENTRY_NULL) && (entry->object != VM_OBJECT_NULL)) { if (entry->object->named == TRUE) panic("memory_object_create_named: caller already holds the right"); } + vm_object_hash_unlock(lck); - vm_object_cache_unlock(); - if ((object = vm_object_enter(pager, size, FALSE, FALSE, TRUE)) - == VM_OBJECT_NULL) { + if ((object = vm_object_enter(pager, size, FALSE, FALSE, TRUE)) == VM_OBJECT_NULL) { return(KERN_INVALID_OBJECT); } @@ -4648,51 +6152,49 @@ memory_object_recover_named( { vm_object_t object; - vm_object_cache_lock(); object = memory_object_control_to_vm_object(control); if (object == VM_OBJECT_NULL) { - vm_object_cache_unlock(); return (KERN_INVALID_ARGUMENT); } - restart: vm_object_lock(object); if (object->terminating && wait_on_terminating) { - vm_object_cache_unlock(); vm_object_wait(object, VM_OBJECT_EVENT_PAGING_IN_PROGRESS, THREAD_UNINT); - vm_object_cache_lock(); goto restart; } if (!object->alive) { - vm_object_cache_unlock(); vm_object_unlock(object); return KERN_FAILURE; } if (object->named == TRUE) { - vm_object_cache_unlock(); vm_object_unlock(object); return KERN_SUCCESS; } - - if((object->ref_count == 0) && (!object->terminating)){ +#if VM_OBJECT_CACHE + if ((object->ref_count == 0) && (!object->terminating)) { + if (!vm_object_cache_lock_try()) { + vm_object_unlock(object); + goto restart; + } queue_remove(&vm_object_cached_list, object, vm_object_t, cached_list); - vm_object_cached_count--; - XPR(XPR_VM_OBJECT_CACHE, - "memory_object_recover_named: removing %X, head (%X, %X)\n", - (integer_t)object, - (integer_t)vm_object_cached_list.next, - (integer_t)vm_object_cached_list.prev, 0,0); + vm_object_cached_count--; + XPR(XPR_VM_OBJECT_CACHE, + "memory_object_recover_named: removing %X, head (%X, %X)\n", + object, + vm_object_cached_list.next, + vm_object_cached_list.prev, 0,0); + + vm_object_cache_unlock(); } - - vm_object_cache_unlock(); - +#endif object->named = TRUE; + vm_object_lock_assert_exclusive(object); object->ref_count++; vm_object_res_reference(object); while (!object->pager_ready) { @@ -4732,17 +6234,10 @@ vm_object_release_name( while (object != VM_OBJECT_NULL) { - /* - * The cache holds a reference (uncounted) to - * the object. We must locke it before removing - * the object. - * - */ - - vm_object_cache_lock(); vm_object_lock(object); + assert(object->alive); - if(original_object) + if (original_object) assert(object->named); assert(object->ref_count > 0); @@ -4757,7 +6252,6 @@ vm_object_release_name( VM_OBJECT_EVENT_INITIALIZED, THREAD_UNINT); vm_object_unlock(object); - vm_object_cache_unlock(); thread_block(THREAD_CONTINUE_NULL); continue; } @@ -4766,22 +6260,19 @@ vm_object_release_name( && (flags & MEMORY_OBJECT_TERMINATE_IDLE)) || (object->terminating)) { vm_object_unlock(object); - vm_object_cache_unlock(); return KERN_FAILURE; } else { if (flags & MEMORY_OBJECT_RELEASE_NO_OP) { vm_object_unlock(object); - vm_object_cache_unlock(); return KERN_SUCCESS; } } if ((flags & MEMORY_OBJECT_RESPECT_CACHE) && (object->ref_count == 1)) { - if(original_object) + if (original_object) object->named = FALSE; vm_object_unlock(object); - vm_object_cache_unlock(); /* let vm_object_deallocate push this thing into */ /* the cache, if that it is where it is bound */ vm_object_deallocate(object); @@ -4789,9 +6280,10 @@ vm_object_release_name( } VM_OBJ_RES_DECR(object); shadow = object->pageout?VM_OBJECT_NULL:object->shadow; - if(object->ref_count == 1) { - if(vm_object_terminate(object) != KERN_SUCCESS) { - if(original_object) { + + if (object->ref_count == 1) { + if (vm_object_terminate(object) != KERN_SUCCESS) { + if (original_object) { return KERN_FAILURE; } else { return KERN_SUCCESS; @@ -4804,12 +6296,12 @@ vm_object_release_name( } return KERN_SUCCESS; } else { + vm_object_lock_assert_exclusive(object); object->ref_count--; assert(object->ref_count > 0); if(original_object) object->named = FALSE; vm_object_unlock(object); - vm_object_cache_unlock(); return KERN_SUCCESS; } } @@ -4834,7 +6326,7 @@ vm_object_lock_request( XPR(XPR_MEMORY_OBJECT, "vm_o_lock_request, obj 0x%X off 0x%X size 0x%X flags %X prot %X\n", - (integer_t)object, offset, size, + object, offset, size, (((should_return&1)<<1)|should_flush), prot); /* @@ -4865,214 +6357,203 @@ vm_object_lock_request( } /* - * Empty a purgable object by grabbing the physical pages assigned to it and + * Empty a purgeable object by grabbing the physical pages assigned to it and * putting them on the free queue without writing them to backing store, etc. * When the pages are next touched they will be demand zero-fill pages. We * skip pages which are busy, being paged in/out, wired, etc. We do _not_ * skip referenced/dirty pages, pages on the active queue, etc. We're more - * than happy to grab these since this is a purgable object. We mark the + * than happy to grab these since this is a purgeable object. We mark the * object as "empty" after reaping its pages. * - * On entry the object and page queues are locked, the object must be a - * purgable object with no delayed copies pending. + * On entry the object must be locked and it must be + * purgeable with no delayed copies pending. */ -unsigned int -vm_object_purge(vm_object_t object) +void +vm_object_purge(vm_object_t object, int flags) { - vm_page_t p, next; - unsigned int num_purged_pages; - vm_page_t local_freeq; - unsigned long local_freed; - int purge_loop_quota; -/* free pages as soon as we gather PURGE_BATCH_FREE_LIMIT pages to free */ -#define PURGE_BATCH_FREE_LIMIT 50 -/* release page queues lock every PURGE_LOOP_QUOTA iterations */ -#define PURGE_LOOP_QUOTA 100 + unsigned int object_page_count = 0; + unsigned int pgcount = 0; + boolean_t skipped_object = FALSE; - num_purged_pages = 0; - if (object->purgable == VM_OBJECT_NONPURGABLE) - return num_purged_pages; + vm_object_lock_assert_exclusive(object); - object->purgable = VM_OBJECT_PURGABLE_EMPTY; + if (object->purgable == VM_PURGABLE_DENY) + return; assert(object->copy == VM_OBJECT_NULL); assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE); - purge_loop_quota = PURGE_LOOP_QUOTA; - - local_freeq = VM_PAGE_NULL; - local_freed = 0; /* - * Go through the object's resident pages and try and discard them. + * We need to set the object's state to VM_PURGABLE_EMPTY *before* + * reaping its pages. We update vm_page_purgeable_count in bulk + * and we don't want vm_page_remove() to update it again for each + * page we reap later. + * + * For the purgeable ledgers, pages from VOLATILE and EMPTY objects + * are all accounted for in the "volatile" ledgers, so this does not + * make any difference. + * If we transitioned directly from NONVOLATILE to EMPTY, + * vm_page_purgeable_count must have been updated when the object + * was dequeued from its volatile queue and the purgeable ledgers + * must have also been updated accordingly at that time (in + * vm_object_purgable_control()). */ - next = (vm_page_t)queue_first(&object->memq); - while (!queue_end(&object->memq, (queue_entry_t)next)) { - p = next; - next = (vm_page_t)queue_next(&next->listq); - - if (purge_loop_quota-- == 0) { - /* - * Avoid holding the page queues lock for too long. - * Let someone else take it for a while if needed. - * Keep holding the object's lock to guarantee that - * the object's page list doesn't change under us - * while we yield. - */ - if (local_freeq != VM_PAGE_NULL) { - /* - * Flush our queue of pages to free. - */ - vm_page_free_list(local_freeq); - local_freeq = VM_PAGE_NULL; - local_freed = 0; - } - vm_page_unlock_queues(); - mutex_pause(); - vm_page_lock_queues(); - - /* resume with the current page and a new quota */ - purge_loop_quota = PURGE_LOOP_QUOTA; - } - - - if (p->busy || p->cleaning || p->laundry || - p->list_req_pending) { - /* page is being acted upon, so don't mess with it */ - continue; - } - if (p->wire_count) { - /* don't discard a wired page */ - continue; - } - - if (p->tabled) { - /* clean up the object/offset table */ - vm_page_remove(p); + if (object->purgable == VM_PURGABLE_VOLATILE) { + unsigned int delta; + assert(object->resident_page_count >= + object->wired_page_count); + delta = (object->resident_page_count - + object->wired_page_count); + if (delta != 0) { + assert(vm_page_purgeable_count >= + delta); + OSAddAtomic(-delta, + (SInt32 *)&vm_page_purgeable_count); } - if (p->absent) { - /* update the object's count of absent pages */ - vm_object_absent_release(object); + if (object->wired_page_count != 0) { + assert(vm_page_purgeable_wired_count >= + object->wired_page_count); + OSAddAtomic(-object->wired_page_count, + (SInt32 *)&vm_page_purgeable_wired_count); } + object->purgable = VM_PURGABLE_EMPTY; + } + assert(object->purgable == VM_PURGABLE_EMPTY); + + object_page_count = object->resident_page_count; - /* we can discard this page */ + vm_object_reap_pages(object, REAP_PURGEABLE); - /* advertize that this page is in a transition state */ - p->busy = TRUE; + if (object->pager != NULL) { - if (p->no_isync == TRUE) { - /* the page hasn't been mapped yet */ - /* (optimization to delay the i-cache sync) */ - } else { - /* unmap the page */ - int refmod_state; + assert(VM_CONFIG_COMPRESSOR_IS_PRESENT); - refmod_state = pmap_disconnect(p->phys_page); - if (refmod_state & VM_MEM_MODIFIED) { - p->dirty = TRUE; + if (object->activity_in_progress == 0 && + object->paging_in_progress == 0) { + /* + * Also reap any memory coming from this object + * in the VM compressor. + * + * There are no operations in progress on the VM object + * and no operation can start while we're holding the + * VM object lock, so it's safe to reap the compressed + * pages and update the page counts. + */ + pgcount = vm_compressor_pager_get_count(object->pager); + if (pgcount) { + pgcount = vm_compressor_pager_reap_pages(object->pager, flags); + vm_compressor_pager_count(object->pager, + -pgcount, + FALSE, /* shared */ + object); + vm_purgeable_compressed_update(object, + -pgcount); } - } - - if (p->dirty || p->precious) { - /* we saved the cost of cleaning this page ! */ - num_purged_pages++; - vm_page_purged_count++; - } - - /* remove page from active or inactive queue... */ - VM_PAGE_QUEUES_REMOVE(p); - - /* ... and put it on our queue of pages to free */ - assert(!p->laundry); - assert(p->object != kernel_object); - assert(p->pageq.next == NULL && - p->pageq.prev == NULL); - p->pageq.next = (queue_entry_t) local_freeq; - local_freeq = p; - if (++local_freed >= PURGE_BATCH_FREE_LIMIT) { - /* flush our queue of pages to free */ - vm_page_free_list(local_freeq); - local_freeq = VM_PAGE_NULL; - local_freed = 0; + if ( !(flags & C_DONT_BLOCK)) { + assert(vm_compressor_pager_get_count(object->pager) + == 0); + } + } else { + /* + * There's some kind of paging activity in progress + * for this object, which could result in a page + * being compressed or decompressed, possibly while + * the VM object is not locked, so it could race + * with us. + * + * We can't really synchronize this without possibly + * causing a deadlock when the compressor needs to + * allocate or free memory while compressing or + * decompressing a page from a purgeable object + * mapped in the kernel_map... + * + * So let's not attempt to purge the compressor + * pager if there's any kind of operation in + * progress on the VM object. + */ + skipped_object = TRUE; } } - /* flush our local queue of pages to free one last time */ - if (local_freeq != VM_PAGE_NULL) { - vm_page_free_list(local_freeq); - local_freeq = VM_PAGE_NULL; - local_freed = 0; - } + vm_object_lock_assert_exclusive(object); + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, (MACHDBG_CODE(DBG_MACH_VM, OBJECT_PURGE_ONE)), + VM_KERNEL_UNSLIDE_OR_PERM(object), /* purged object */ + object_page_count, + pgcount, + skipped_object, + 0); - return num_purged_pages; } + /* - * vm_object_purgable_control() allows the caller to control and investigate the - * state of a purgable object. A purgable object is created via a call to - * vm_allocate() with VM_FLAGS_PURGABLE specified. A purgable object will - * never be coalesced with any other object -- even other purgable objects -- - * and will thus always remain a distinct object. A purgable object has + * vm_object_purgeable_control() allows the caller to control and investigate the + * state of a purgeable object. A purgeable object is created via a call to + * vm_allocate() with VM_FLAGS_PURGABLE specified. A purgeable object will + * never be coalesced with any other object -- even other purgeable objects -- + * and will thus always remain a distinct object. A purgeable object has * special semantics when its reference count is exactly 1. If its reference - * count is greater than 1, then a purgable object will behave like a normal + * count is greater than 1, then a purgeable object will behave like a normal * object and attempts to use this interface will result in an error return * of KERN_INVALID_ARGUMENT. * - * A purgable object may be put into a "volatile" state which will make the + * A purgeable object may be put into a "volatile" state which will make the * object's pages elligable for being reclaimed without paging to backing * store if the system runs low on memory. If the pages in a volatile - * purgable object are reclaimed, the purgable object is said to have been - * "emptied." When a purgable object is emptied the system will reclaim as + * purgeable object are reclaimed, the purgeable object is said to have been + * "emptied." When a purgeable object is emptied the system will reclaim as * many pages from the object as it can in a convenient manner (pages already * en route to backing store or busy for other reasons are left as is). When - * a purgable object is made volatile, its pages will generally be reclaimed + * a purgeable object is made volatile, its pages will generally be reclaimed * before other pages in the application's working set. This semantic is * generally used by applications which can recreate the data in the object * faster than it can be paged in. One such example might be media assets * which can be reread from a much faster RAID volume. * - * A purgable object may be designated as "non-volatile" which means it will + * A purgeable object may be designated as "non-volatile" which means it will * behave like all other objects in the system with pages being written to and * read from backing store as needed to satisfy system memory needs. If the * object was emptied before the object was made non-volatile, that fact will - * be returned as the old state of the purgable object (see + * be returned as the old state of the purgeable object (see * VM_PURGABLE_SET_STATE below). In this case, any pages of the object which * were reclaimed as part of emptying the object will be refaulted in as * zero-fill on demand. It is up to the application to note that an object * was emptied and recreate the objects contents if necessary. When a - * purgable object is made non-volatile, its pages will generally not be paged - * out to backing store in the immediate future. A purgable object may also + * purgeable object is made non-volatile, its pages will generally not be paged + * out to backing store in the immediate future. A purgeable object may also * be manually emptied. * * Finally, the current state (non-volatile, volatile, volatile & empty) of a - * volatile purgable object may be queried at any time. This information may + * volatile purgeable object may be queried at any time. This information may * be used as a control input to let the application know when the system is * experiencing memory pressure and is reclaiming memory. * - * The specified address may be any address within the purgable object. If + * The specified address may be any address within the purgeable object. If * the specified address does not represent any object in the target task's * virtual address space, then KERN_INVALID_ADDRESS will be returned. If the - * object containing the specified address is not a purgable object, then + * object containing the specified address is not a purgeable object, then * KERN_INVALID_ARGUMENT will be returned. Otherwise, KERN_SUCCESS will be * returned. * * The control parameter may be any one of VM_PURGABLE_SET_STATE or * VM_PURGABLE_GET_STATE. For VM_PURGABLE_SET_STATE, the in/out parameter - * state is used to set the new state of the purgable object and return its - * old state. For VM_PURGABLE_GET_STATE, the current state of the purgable + * state is used to set the new state of the purgeable object and return its + * old state. For VM_PURGABLE_GET_STATE, the current state of the purgeable * object is returned in the parameter state. * * The in/out parameter state may be one of VM_PURGABLE_NONVOLATILE, * VM_PURGABLE_VOLATILE or VM_PURGABLE_EMPTY. These, respectively, represent * the non-volatile, volatile and volatile/empty states described above. - * Setting the state of a purgable object to VM_PURGABLE_EMPTY will + * Setting the state of a purgeable object to VM_PURGABLE_EMPTY will * immediately reclaim as many pages in the object as can be conveniently * collected (some may have already been written to backing store or be * otherwise busy). * - * The process of making a purgable object non-volatile and determining its - * previous state is atomic. Thus, if a purgable object is made + * The process of making a purgeable object non-volatile and determining its + * previous state is atomic. Thus, if a purgeable object is made * VM_PURGABLE_NONVOLATILE and the old state is returned as - * VM_PURGABLE_VOLATILE, then the purgable object's previous contents are + * VM_PURGABLE_VOLATILE, then the purgeable object's previous contents are * completely intact and will remain so until the object is made volatile * again. If the old state is returned as VM_PURGABLE_EMPTY then the object * was reclaimed while it was in a volatile state and its previous contents @@ -5088,42 +6569,25 @@ vm_object_purgable_control( int *state) { int old_state; - vm_page_t p; + int new_state; if (object == VM_OBJECT_NULL) { /* - * Object must already be present or it can't be purgable. + * Object must already be present or it can't be purgeable. */ return KERN_INVALID_ARGUMENT; } + vm_object_lock_assert_exclusive(object); + /* - * Get current state of the purgable object. + * Get current state of the purgeable object. */ - switch (object->purgable) { - case VM_OBJECT_NONPURGABLE: + old_state = object->purgable; + if (old_state == VM_PURGABLE_DENY) return KERN_INVALID_ARGUMENT; - case VM_OBJECT_PURGABLE_NONVOLATILE: - old_state = VM_PURGABLE_NONVOLATILE; - break; - - case VM_OBJECT_PURGABLE_VOLATILE: - old_state = VM_PURGABLE_VOLATILE; - break; - - case VM_OBJECT_PURGABLE_EMPTY: - old_state = VM_PURGABLE_EMPTY; - break; - - default: - old_state = VM_PURGABLE_NONVOLATILE; - panic("Bad state (%d) for purgable object!\n", - object->purgable); - /*NOTREACHED*/ - } - - /* purgable cant have delayed copies - now or in the future */ + /* purgeable cant have delayed copies - now or in the future */ assert(object->copy == VM_OBJECT_NULL); assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE); @@ -5135,94 +6599,382 @@ vm_object_purgable_control( return KERN_SUCCESS; } - switch (*state) { + if ((*state) & VM_PURGABLE_DEBUG_EMPTY) { + object->volatile_empty = TRUE; + } + if ((*state) & VM_PURGABLE_DEBUG_FAULT) { + object->volatile_fault = TRUE; + } + + new_state = *state & VM_PURGABLE_STATE_MASK; + if (new_state == VM_PURGABLE_VOLATILE && + object->volatile_empty) { + new_state = VM_PURGABLE_EMPTY; + } + + switch (new_state) { + case VM_PURGABLE_DENY: case VM_PURGABLE_NONVOLATILE: - vm_page_lock_queues(); - if (object->purgable != VM_OBJECT_PURGABLE_NONVOLATILE) { - assert(vm_page_purgeable_count >= - object->resident_page_count); - vm_page_purgeable_count -= object->resident_page_count; - } + object->purgable = new_state; - object->purgable = VM_OBJECT_PURGABLE_NONVOLATILE; + if (old_state == VM_PURGABLE_VOLATILE) { + unsigned int delta; - /* - * If the object wasn't emptied, then mark all pages of the - * object as referenced in order to give them a complete turn - * of the virtual memory "clock" before becoming candidates - * for paging out (if the system is suffering from memory - * pressure). We don't really need to set the pmap reference - * bits (which would be expensive) since the software copies - * are believed if they're set to true ... - */ - if (old_state != VM_PURGABLE_EMPTY) { - for (p = (vm_page_t)queue_first(&object->memq); - !queue_end(&object->memq, (queue_entry_t)p); - p = (vm_page_t)queue_next(&p->listq)) - p->reference = TRUE; - } + assert(object->resident_page_count >= + object->wired_page_count); + delta = (object->resident_page_count - + object->wired_page_count); - vm_page_unlock_queues(); + assert(vm_page_purgeable_count >= delta); - break; + if (delta != 0) { + OSAddAtomic(-delta, + (SInt32 *)&vm_page_purgeable_count); + } + if (object->wired_page_count != 0) { + assert(vm_page_purgeable_wired_count >= + object->wired_page_count); + OSAddAtomic(-object->wired_page_count, + (SInt32 *)&vm_page_purgeable_wired_count); + } - case VM_PURGABLE_VOLATILE: - vm_page_lock_queues(); + vm_page_lock_queues(); - if (object->purgable != VM_OBJECT_PURGABLE_VOLATILE && - object->purgable != VM_OBJECT_PURGABLE_EMPTY) { - vm_page_purgeable_count += object->resident_page_count; - } + /* object should be on a queue */ + assert(object->objq.next != NULL && + object->objq.prev != NULL); + purgeable_q_t queue; - object->purgable = VM_OBJECT_PURGABLE_VOLATILE; + /* + * Move object from its volatile queue to the + * non-volatile queue... + */ + queue = vm_purgeable_object_remove(object); + assert(queue); - /* - * We want the newly volatile purgable object to be a - * candidate for the pageout scan before other pages in the - * application if the system is suffering from memory - * pressure. To do this, we move a page of the object from - * the active queue onto the inactive queue in order to - * promote the object for early reclaim. We only need to move - * a single page since the pageout scan will reap the entire - * purgable object if it finds a single page in a volatile - * state. Obviously we don't do this if there are no pages - * associated with the object or we find a page of the object - * already on the inactive queue. - */ - for (p = (vm_page_t)queue_first(&object->memq); - !queue_end(&object->memq, (queue_entry_t)p); - p = (vm_page_t)queue_next(&p->listq)) { - if (p->inactive) { - /* already a page on the inactive queue */ - break; - } - if (p->active && !p->busy) { - /* found one we can move */ - vm_page_deactivate(p); - break; + if (object->purgeable_when_ripe) { + vm_purgeable_token_delete_last(queue); } + assert(queue->debug_count_objects>=0); + + vm_page_unlock_queues(); + } + if (old_state == VM_PURGABLE_VOLATILE || + old_state == VM_PURGABLE_EMPTY) { + /* + * Transfer the object's pages from the volatile to + * non-volatile ledgers. + */ + vm_purgeable_accounting(object, VM_PURGABLE_VOLATILE, + FALSE); } - vm_page_unlock_queues(); break; - - case VM_PURGABLE_EMPTY: - vm_page_lock_queues(); - if (object->purgable != VM_OBJECT_PURGABLE_VOLATILE && - object->purgable != VM_OBJECT_PURGABLE_EMPTY) { - vm_page_purgeable_count += object->resident_page_count; + case VM_PURGABLE_VOLATILE: + if (object->volatile_fault) { + vm_page_t p; + int refmod; + + vm_page_queue_iterate(&object->memq, p, vm_page_t, listq) { + if (p->busy || + VM_PAGE_WIRED(p) || + p->fictitious) { + continue; + } + refmod = pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(p)); + if ((refmod & VM_MEM_MODIFIED) && + !p->dirty) { + SET_PAGE_DIRTY(p, FALSE); + } + } } - (void) vm_object_purge(object); - vm_page_unlock_queues(); - break; + + if (old_state == VM_PURGABLE_EMPTY && + object->resident_page_count == 0 && + object->pager == NULL) + break; - } - *state = old_state; + purgeable_q_t queue; + + /* find the correct queue */ + if ((*state&VM_PURGABLE_ORDERING_MASK) == VM_PURGABLE_ORDERING_OBSOLETE) + queue = &purgeable_queues[PURGEABLE_Q_TYPE_OBSOLETE]; + else { + if ((*state&VM_PURGABLE_BEHAVIOR_MASK) == VM_PURGABLE_BEHAVIOR_FIFO) + queue = &purgeable_queues[PURGEABLE_Q_TYPE_FIFO]; + else + queue = &purgeable_queues[PURGEABLE_Q_TYPE_LIFO]; + } + + if (old_state == VM_PURGABLE_NONVOLATILE || + old_state == VM_PURGABLE_EMPTY) { + unsigned int delta; + + if ((*state & VM_PURGABLE_NO_AGING_MASK) == + VM_PURGABLE_NO_AGING) { + object->purgeable_when_ripe = FALSE; + } else { + object->purgeable_when_ripe = TRUE; + } + + if (object->purgeable_when_ripe) { + kern_return_t result; + + /* try to add token... this can fail */ + vm_page_lock_queues(); + + result = vm_purgeable_token_add(queue); + if (result != KERN_SUCCESS) { + vm_page_unlock_queues(); + return result; + } + vm_page_unlock_queues(); + } + + assert(object->resident_page_count >= + object->wired_page_count); + delta = (object->resident_page_count - + object->wired_page_count); + + if (delta != 0) { + OSAddAtomic(delta, + &vm_page_purgeable_count); + } + if (object->wired_page_count != 0) { + OSAddAtomic(object->wired_page_count, + &vm_page_purgeable_wired_count); + } + + object->purgable = new_state; + + /* object should be on "non-volatile" queue */ + assert(object->objq.next != NULL); + assert(object->objq.prev != NULL); + } + else if (old_state == VM_PURGABLE_VOLATILE) { + purgeable_q_t old_queue; + boolean_t purgeable_when_ripe; + + /* + * if reassigning priorities / purgeable groups, we don't change the + * token queue. So moving priorities will not make pages stay around longer. + * Reasoning is that the algorithm gives most priority to the most important + * object. If a new token is added, the most important object' priority is boosted. + * This biases the system already for purgeable queues that move a lot. + * It doesn't seem more biasing is neccessary in this case, where no new object is added. + */ + assert(object->objq.next != NULL && object->objq.prev != NULL); /* object should be on a queue */ + + old_queue = vm_purgeable_object_remove(object); + assert(old_queue); + + if ((*state & VM_PURGABLE_NO_AGING_MASK) == + VM_PURGABLE_NO_AGING) { + purgeable_when_ripe = FALSE; + } else { + purgeable_when_ripe = TRUE; + } + + if (old_queue != queue || + (purgeable_when_ripe != + object->purgeable_when_ripe)) { + kern_return_t result; + + /* Changing queue. Have to move token. */ + vm_page_lock_queues(); + if (object->purgeable_when_ripe) { + vm_purgeable_token_delete_last(old_queue); + } + object->purgeable_when_ripe = purgeable_when_ripe; + if (object->purgeable_when_ripe) { + result = vm_purgeable_token_add(queue); + assert(result==KERN_SUCCESS); /* this should never fail since we just freed a token */ + } + vm_page_unlock_queues(); + + } + }; + vm_purgeable_object_add(object, queue, (*state&VM_VOLATILE_GROUP_MASK)>>VM_VOLATILE_GROUP_SHIFT ); + if (old_state == VM_PURGABLE_NONVOLATILE) { + vm_purgeable_accounting(object, VM_PURGABLE_NONVOLATILE, + FALSE); + } + + assert(queue->debug_count_objects>=0); + + break; + + + case VM_PURGABLE_EMPTY: + if (object->volatile_fault) { + vm_page_t p; + int refmod; + + vm_page_queue_iterate(&object->memq, p, vm_page_t, listq) { + if (p->busy || + VM_PAGE_WIRED(p) || + p->fictitious) { + continue; + } + refmod = pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(p)); + if ((refmod & VM_MEM_MODIFIED) && + !p->dirty) { + SET_PAGE_DIRTY(p, FALSE); + } + } + } + + if (old_state == new_state) { + /* nothing changes */ + break; + } + + assert(old_state == VM_PURGABLE_NONVOLATILE || + old_state == VM_PURGABLE_VOLATILE); + if (old_state == VM_PURGABLE_VOLATILE) { + purgeable_q_t old_queue; + + /* object should be on a queue */ + assert(object->objq.next != NULL && + object->objq.prev != NULL); + + old_queue = vm_purgeable_object_remove(object); + assert(old_queue); + if (object->purgeable_when_ripe) { + vm_page_lock_queues(); + vm_purgeable_token_delete_first(old_queue); + vm_page_unlock_queues(); + } + } + + if (old_state == VM_PURGABLE_NONVOLATILE) { + /* + * This object's pages were previously accounted as + * "non-volatile" and now need to be accounted as + * "volatile". + */ + vm_purgeable_accounting(object, VM_PURGABLE_NONVOLATILE, + FALSE); + /* + * Set to VM_PURGABLE_EMPTY because the pages are no + * longer accounted in the "non-volatile" ledger + * and are also not accounted for in + * "vm_page_purgeable_count". + */ + object->purgable = VM_PURGABLE_EMPTY; + } + + (void) vm_object_purge(object, 0); + assert(object->purgable == VM_PURGABLE_EMPTY); + + break; + } + + *state = old_state; + + vm_object_lock_assert_exclusive(object); return KERN_SUCCESS; } +kern_return_t +vm_object_get_page_counts( + vm_object_t object, + vm_object_offset_t offset, + vm_object_size_t size, + unsigned int *resident_page_count, + unsigned int *dirty_page_count) +{ + + kern_return_t kr = KERN_SUCCESS; + boolean_t count_dirty_pages = FALSE; + vm_page_t p = VM_PAGE_NULL; + unsigned int local_resident_count = 0; + unsigned int local_dirty_count = 0; + vm_object_offset_t cur_offset = 0; + vm_object_offset_t end_offset = 0; + + if (object == VM_OBJECT_NULL) + return KERN_INVALID_ARGUMENT; + + + cur_offset = offset; + + end_offset = offset + size; + + vm_object_lock_assert_exclusive(object); + + if (dirty_page_count != NULL) { + + count_dirty_pages = TRUE; + } + + if (resident_page_count != NULL && count_dirty_pages == FALSE) { + /* + * Fast path when: + * - we only want the resident page count, and, + * - the entire object is exactly covered by the request. + */ + if (offset == 0 && (object->vo_size == size)) { + + *resident_page_count = object->resident_page_count; + goto out; + } + } + + if (object->resident_page_count <= (size >> PAGE_SHIFT)) { + + vm_page_queue_iterate(&object->memq, p, vm_page_t, listq) { + + if (p->offset >= cur_offset && p->offset < end_offset) { + + local_resident_count++; + + if (count_dirty_pages) { + + if (p->dirty || (p->wpmapped && pmap_is_modified(VM_PAGE_GET_PHYS_PAGE(p)))) { + + local_dirty_count++; + } + } + } + } + } else { + + for (cur_offset = offset; cur_offset < end_offset; cur_offset += PAGE_SIZE_64) { + + p = vm_page_lookup(object, cur_offset); + + if (p != VM_PAGE_NULL) { + + local_resident_count++; + + if (count_dirty_pages) { + + if (p->dirty || (p->wpmapped && pmap_is_modified(VM_PAGE_GET_PHYS_PAGE(p)))) { + + local_dirty_count++; + } + } + } + } + + } + + if (resident_page_count != NULL) { + *resident_page_count = local_resident_count; + } + + if (dirty_page_count != NULL) { + *dirty_page_count = local_dirty_count; + } + +out: + return kr; +} + + #if TASK_SWAPPER /* * vm_object_res_deallocate @@ -5315,7 +7067,7 @@ vm_object_res_reference( #endif __private_extern__ void vm_object_reference( - register vm_object_t object) + vm_object_t object) { if (object == VM_OBJECT_NULL) return; @@ -5340,10 +7092,12 @@ vm_object_reference( kern_return_t adjust_vm_object_cache( __unused vm_size_t oval, - vm_size_t nval) + __unused vm_size_t nval) { +#if VM_OBJECT_CACHE vm_object_cached_max = nval; vm_object_cache_trim(FALSE); +#endif return (KERN_SUCCESS); } #endif /* MACH_BSD */ @@ -5359,6 +7113,7 @@ adjust_vm_object_cache( * * The VM objects must not be locked by caller. */ +unsigned int vm_object_transpose_count = 0; kern_return_t vm_object_transpose( vm_object_t object1, @@ -5368,13 +7123,13 @@ vm_object_transpose( vm_object_t tmp_object; kern_return_t retval; boolean_t object1_locked, object2_locked; - boolean_t object1_paging, object2_paging; vm_page_t page; vm_object_offset_t page_offset; + lck_mtx_t *hash_lck; + vm_object_hash_entry_t hash_entry; tmp_object = VM_OBJECT_NULL; object1_locked = FALSE; object2_locked = FALSE; - object1_paging = FALSE; object2_paging = FALSE; if (object1 == object2 || object1 == VM_OBJECT_NULL || @@ -5387,10 +7142,34 @@ vm_object_transpose( goto done; } + /* + * Since we need to lock both objects at the same time, + * make sure we always lock them in the same order to + * avoid deadlocks. + */ + if (object1 > object2) { + tmp_object = object1; + object1 = object2; + object2 = tmp_object; + } + + /* + * Allocate a temporary VM object to hold object1's contents + * while we copy object2 to object1. + */ + tmp_object = vm_object_allocate(transpose_size); + vm_object_lock(tmp_object); + tmp_object->can_persist = FALSE; + + + /* + * Grab control of the 1st VM object. + */ vm_object_lock(object1); object1_locked = TRUE; - if (object1->copy || object1->shadow || object1->shadowed || - object1->purgable != VM_OBJECT_NONPURGABLE) { + if (!object1->alive || object1->terminating || + object1->copy || object1->shadow || object1->shadowed || + object1->purgable != VM_PURGABLE_DENY) { /* * We don't deal with copy or shadow objects (yet). */ @@ -5398,59 +7177,36 @@ vm_object_transpose( goto done; } /* - * Since we're about to mess with the object's backing store, - * mark it as "paging_in_progress". Note that this is not enough + * We're about to mess with the object's backing store and + * taking a "paging_in_progress" reference wouldn't be enough * to prevent any paging activity on this object, so the caller should * have "quiesced" the objects beforehand, via a UPL operation with * UPL_SET_IO_WIRE (to make sure all the pages are there and wired) * and UPL_BLOCK_ACCESS (to mark the pages "busy"). + * + * Wait for any paging operation to complete (but only paging, not + * other kind of activities not linked to the pager). After we're + * statisfied that there's no more paging in progress, we keep the + * object locked, to guarantee that no one tries to access its pager. */ - vm_object_paging_begin(object1); - object1_paging = TRUE; - vm_object_unlock(object1); - object1_locked = FALSE; + vm_object_paging_only_wait(object1, THREAD_UNINT); /* * Same as above for the 2nd object... */ vm_object_lock(object2); object2_locked = TRUE; - if (object2->copy || object2->shadow || object2->shadowed || - object2->purgable != VM_OBJECT_NONPURGABLE) { + if (! object2->alive || object2->terminating || + object2->copy || object2->shadow || object2->shadowed || + object2->purgable != VM_PURGABLE_DENY) { retval = KERN_INVALID_VALUE; goto done; } - vm_object_paging_begin(object2); - object2_paging = TRUE; - vm_object_unlock(object2); - object2_locked = FALSE; - - /* - * Allocate a temporary VM object to hold object1's contents - * while we copy object2 to object1. - */ - tmp_object = vm_object_allocate(transpose_size); - vm_object_lock(tmp_object); - vm_object_paging_begin(tmp_object); - tmp_object->can_persist = FALSE; + vm_object_paging_only_wait(object2, THREAD_UNINT); - /* - * Since we need to lock both objects at the same time, - * make sure we always lock them in the same order to - * avoid deadlocks. - */ - if (object1 < object2) { - vm_object_lock(object1); - vm_object_lock(object2); - } else { - vm_object_lock(object2); - vm_object_lock(object1); - } - object1_locked = TRUE; - object2_locked = TRUE; - if (object1->size != object2->size || - object1->size != transpose_size) { + if (object1->vo_size != object2->vo_size || + object1->vo_size != transpose_size) { /* * If the 2 objects don't have the same size, we can't * exchange their backing stores or one would overflow. @@ -5466,60 +7222,56 @@ vm_object_transpose( /* * Transpose the lists of resident pages. + * This also updates the resident_page_count and the memq_hint. */ - if (object1->phys_contiguous || queue_empty(&object1->memq)) { + if (object1->phys_contiguous || vm_page_queue_empty(&object1->memq)) { /* * No pages in object1, just transfer pages * from object2 to object1. No need to go through * an intermediate object. */ - while (!queue_empty(&object2->memq)) { - page = (vm_page_t) queue_first(&object2->memq); - vm_page_rename(page, object1, page->offset); + while (!vm_page_queue_empty(&object2->memq)) { + page = (vm_page_t) vm_page_queue_first(&object2->memq); + vm_page_rename(page, object1, page->offset, FALSE); } - assert(queue_empty(&object2->memq)); - } else if (object2->phys_contiguous || queue_empty(&object2->memq)) { + assert(vm_page_queue_empty(&object2->memq)); + } else if (object2->phys_contiguous || vm_page_queue_empty(&object2->memq)) { /* * No pages in object2, just transfer pages * from object1 to object2. No need to go through * an intermediate object. */ - while (!queue_empty(&object1->memq)) { - page = (vm_page_t) queue_first(&object1->memq); - vm_page_rename(page, object2, page->offset); + while (!vm_page_queue_empty(&object1->memq)) { + page = (vm_page_t) vm_page_queue_first(&object1->memq); + vm_page_rename(page, object2, page->offset, FALSE); } - assert(queue_empty(&object1->memq)); + assert(vm_page_queue_empty(&object1->memq)); } else { /* transfer object1's pages to tmp_object */ - vm_page_lock_queues(); - while (!queue_empty(&object1->memq)) { - page = (vm_page_t) queue_first(&object1->memq); + while (!vm_page_queue_empty(&object1->memq)) { + page = (vm_page_t) vm_page_queue_first(&object1->memq); page_offset = page->offset; - vm_page_remove(page); + vm_page_remove(page, TRUE); page->offset = page_offset; - queue_enter(&tmp_object->memq, page, vm_page_t, listq); + vm_page_queue_enter(&tmp_object->memq, page, vm_page_t, listq); } - vm_page_unlock_queues(); - assert(queue_empty(&object1->memq)); + assert(vm_page_queue_empty(&object1->memq)); /* transfer object2's pages to object1 */ - while (!queue_empty(&object2->memq)) { - page = (vm_page_t) queue_first(&object2->memq); - vm_page_rename(page, object1, page->offset); - } - assert(queue_empty(&object2->memq)); - /* transfer tmp_object's pages to object1 */ - while (!queue_empty(&tmp_object->memq)) { - page = (vm_page_t) queue_first(&tmp_object->memq); - queue_remove(&tmp_object->memq, page, - vm_page_t, listq); + while (!vm_page_queue_empty(&object2->memq)) { + page = (vm_page_t) vm_page_queue_first(&object2->memq); + vm_page_rename(page, object1, page->offset, FALSE); + } + assert(vm_page_queue_empty(&object2->memq)); + /* transfer tmp_object's pages to object2 */ + while (!vm_page_queue_empty(&tmp_object->memq)) { + page = (vm_page_t) vm_page_queue_first(&tmp_object->memq); + vm_page_queue_remove(&tmp_object->memq, page, + vm_page_t, listq); vm_page_insert(page, object2, page->offset); } - assert(queue_empty(&tmp_object->memq)); + assert(vm_page_queue_empty(&tmp_object->memq)); } - /* no need to transpose the size: they should be identical */ - assert(object1->size == object2->size); - #define __TRANSPOSE_FIELD(field) \ MACRO_BEGIN \ tmp_object->field = object1->field; \ @@ -5527,16 +7279,26 @@ MACRO_BEGIN \ object2->field = tmp_object->field; \ MACRO_END + /* "Lock" refers to the object not its contents */ + /* "size" should be identical */ + assert(object1->vo_size == object2->vo_size); + /* "memq_hint" was updated above when transposing pages */ + /* "ref_count" refers to the object not its contents */ +#if TASK_SWAPPER + /* "res_count" refers to the object not its contents */ +#endif + /* "resident_page_count" was updated above when transposing pages */ + /* "wired_page_count" was updated above when transposing pages */ + /* "reusable_page_count" was updated above when transposing pages */ + /* there should be no "copy" */ assert(!object1->copy); assert(!object2->copy); - + /* there should be no "shadow" */ assert(!object1->shadow); assert(!object2->shadow); - - __TRANSPOSE_FIELD(shadow_offset); /* used by phys_contiguous objects */ + __TRANSPOSE_FIELD(vo_shadow_offset); /* used by phys_contiguous objects */ __TRANSPOSE_FIELD(pager); __TRANSPOSE_FIELD(paging_offset); - __TRANSPOSE_FIELD(pager_control); /* update the memory_objects' pointers back to the VM objects */ if (object1->pager_control != MEMORY_OBJECT_CONTROL_NULL) { @@ -5547,29 +7309,92 @@ MACRO_END memory_object_control_collapse(object2->pager_control, object2); } - - __TRANSPOSE_FIELD(absent_count); - - assert(object1->paging_in_progress); - assert(object2->paging_in_progress); - + __TRANSPOSE_FIELD(copy_strategy); + /* "paging_in_progress" refers to the object not its contents */ + assert(!object1->paging_in_progress); + assert(!object2->paging_in_progress); + assert(object1->activity_in_progress); + assert(object2->activity_in_progress); + /* "all_wanted" refers to the object not its contents */ __TRANSPOSE_FIELD(pager_created); __TRANSPOSE_FIELD(pager_initialized); __TRANSPOSE_FIELD(pager_ready); __TRANSPOSE_FIELD(pager_trusted); + __TRANSPOSE_FIELD(can_persist); __TRANSPOSE_FIELD(internal); __TRANSPOSE_FIELD(temporary); __TRANSPOSE_FIELD(private); __TRANSPOSE_FIELD(pageout); + /* "alive" should be set */ + assert(object1->alive); + assert(object2->alive); + /* "purgeable" should be non-purgeable */ + assert(object1->purgable == VM_PURGABLE_DENY); + assert(object2->purgable == VM_PURGABLE_DENY); + /* "shadowed" refers to the the object not its contents */ + __TRANSPOSE_FIELD(purgeable_when_ripe); + __TRANSPOSE_FIELD(advisory_pageout); __TRANSPOSE_FIELD(true_share); + /* "terminating" should not be set */ + assert(!object1->terminating); + assert(!object2->terminating); + __TRANSPOSE_FIELD(named); + /* "shadow_severed" refers to the object not its contents */ __TRANSPOSE_FIELD(phys_contiguous); __TRANSPOSE_FIELD(nophyscache); + /* "cached_list.next" points to transposed object */ + object1->cached_list.next = (queue_entry_t) object2; + object2->cached_list.next = (queue_entry_t) object1; + /* "cached_list.prev" should be NULL */ + assert(object1->cached_list.prev == NULL); + assert(object2->cached_list.prev == NULL); + /* "msr_q" is linked to the object not its contents */ + assert(queue_empty(&object1->msr_q)); + assert(queue_empty(&object2->msr_q)); __TRANSPOSE_FIELD(last_alloc); __TRANSPOSE_FIELD(sequential); - __TRANSPOSE_FIELD(cluster_size); - __TRANSPOSE_FIELD(existence_map); + __TRANSPOSE_FIELD(pages_created); + __TRANSPOSE_FIELD(pages_used); + __TRANSPOSE_FIELD(scan_collisions); __TRANSPOSE_FIELD(cow_hint); +#if MACH_ASSERT + __TRANSPOSE_FIELD(paging_object); +#endif __TRANSPOSE_FIELD(wimg_bits); + __TRANSPOSE_FIELD(set_cache_attr); + __TRANSPOSE_FIELD(code_signed); + if (object1->hashed) { + hash_lck = vm_object_hash_lock_spin(object2->pager); + hash_entry = vm_object_hash_lookup(object2->pager, FALSE); + assert(hash_entry != VM_OBJECT_HASH_ENTRY_NULL); + hash_entry->object = object2; + vm_object_hash_unlock(hash_lck); + } + if (object2->hashed) { + hash_lck = vm_object_hash_lock_spin(object1->pager); + hash_entry = vm_object_hash_lookup(object1->pager, FALSE); + assert(hash_entry != VM_OBJECT_HASH_ENTRY_NULL); + hash_entry->object = object1; + vm_object_hash_unlock(hash_lck); + } + __TRANSPOSE_FIELD(hashed); + object1->transposed = TRUE; + object2->transposed = TRUE; + __TRANSPOSE_FIELD(mapping_in_progress); + __TRANSPOSE_FIELD(volatile_empty); + __TRANSPOSE_FIELD(volatile_fault); + __TRANSPOSE_FIELD(all_reusable); + assert(object1->blocked_access); + assert(object2->blocked_access); + assert(object1->__object2_unused_bits == 0); + assert(object2->__object2_unused_bits == 0); +#if UPL_DEBUG + /* "uplq" refers to the object not its contents (see upl_transpose()) */ +#endif + assert((object1->purgable == VM_PURGABLE_DENY) || (object1->objq.next == NULL)); + assert((object1->purgable == VM_PURGABLE_DENY) || (object1->objq.prev == NULL)); + assert((object2->purgable == VM_PURGABLE_DENY) || (object2->objq.next == NULL)); + assert((object2->purgable == VM_PURGABLE_DENY) || (object2->objq.prev == NULL)); #undef __TRANSPOSE_FIELD @@ -5580,7 +7405,6 @@ done: * Cleanup. */ if (tmp_object != VM_OBJECT_NULL) { - vm_object_paging_end(tmp_object); vm_object_unlock(tmp_object); /* * Re-initialize the temporary object to avoid @@ -5599,109 +7423,450 @@ done: vm_object_unlock(object2); object2_locked = FALSE; } - if (object1_paging) { - vm_object_lock(object1); - vm_object_paging_end(object1); - vm_object_unlock(object1); - object1_paging = FALSE; - } - if (object2_paging) { - vm_object_lock(object2); - vm_object_paging_end(object2); - vm_object_unlock(object2); - object2_paging = FALSE; - } + + vm_object_transpose_count++; return retval; } -/* Allow manipulation of individual page state. This is actually part of */ -/* the UPL regimen but takes place on the VM object rather than on a UPL */ +/* + * vm_object_cluster_size + * + * Determine how big a cluster we should issue an I/O for... + * + * Inputs: *start == offset of page needed + * *length == maximum cluster pager can handle + * Outputs: *start == beginning offset of cluster + * *length == length of cluster to try + * + * The original *start will be encompassed by the cluster + * + */ +extern int speculative_reads_disabled; +extern int ignore_is_ssd; -kern_return_t -vm_object_page_op( - vm_object_t object, - vm_object_offset_t offset, - int ops, - ppnum_t *phys_entry, - int *flags) +/* + * Try to always keep these values an even multiple of PAGE_SIZE. We use these values + * to derive min_ph_bytes and max_ph_bytes (IMP: bytes not # of pages) and expect those values to + * always be page-aligned. The derivation could involve operations (e.g. division) + * that could give us non-page-size aligned values if we start out with values that + * are odd multiples of PAGE_SIZE. + */ + unsigned int preheat_max_bytes = MAX_UPL_TRANSFER_BYTES; +unsigned int preheat_min_bytes = (1024 * 32); + + +__private_extern__ void +vm_object_cluster_size(vm_object_t object, vm_object_offset_t *start, + vm_size_t *length, vm_object_fault_info_t fault_info, uint32_t *io_streaming) { - vm_page_t dst_page; + vm_size_t pre_heat_size; + vm_size_t tail_size; + vm_size_t head_size; + vm_size_t max_length; + vm_size_t cluster_size; + vm_object_offset_t object_size; + vm_object_offset_t orig_start; + vm_object_offset_t target_start; + vm_object_offset_t offset; + vm_behavior_t behavior; + boolean_t look_behind = TRUE; + boolean_t look_ahead = TRUE; + boolean_t isSSD = FALSE; + uint32_t throttle_limit; + int sequential_run; + int sequential_behavior = VM_BEHAVIOR_SEQUENTIAL; + vm_size_t max_ph_size; + vm_size_t min_ph_size; + + assert( !(*length & PAGE_MASK)); + assert( !(*start & PAGE_MASK_64)); - vm_object_lock(object); + /* + * remember maxiumum length of run requested + */ + max_length = *length; + /* + * we'll always return a cluster size of at least + * 1 page, since the original fault must always + * be processed + */ + *length = PAGE_SIZE; + *io_streaming = 0; - if(ops & UPL_POP_PHYSICAL) { - if(object->phys_contiguous) { - if (phys_entry) { - *phys_entry = (ppnum_t) - (object->shadow_offset >> 12); - } - vm_object_unlock(object); - return KERN_SUCCESS; - } else { - vm_object_unlock(object); - return KERN_INVALID_OBJECT; - } - } - if(object->phys_contiguous) { - vm_object_unlock(object); - return KERN_INVALID_OBJECT; + if (speculative_reads_disabled || fault_info == NULL) { + /* + * no cluster... just fault the page in + */ + return; } + orig_start = *start; + target_start = orig_start; + cluster_size = round_page(fault_info->cluster_size); + behavior = fault_info->behavior; - while(TRUE) { - if((dst_page = vm_page_lookup(object,offset)) == VM_PAGE_NULL) { - vm_object_unlock(object); - return KERN_FAILURE; + vm_object_lock(object); + + if (object->pager == MEMORY_OBJECT_NULL) + goto out; /* pager is gone for this object, nothing more to do */ + + if (!ignore_is_ssd) + vnode_pager_get_isSSD(object->pager, &isSSD); + + min_ph_size = round_page(preheat_min_bytes); + max_ph_size = round_page(preheat_max_bytes); + + if (isSSD) { + min_ph_size /= 2; + max_ph_size /= 8; + + if (min_ph_size & PAGE_MASK_64) { + min_ph_size = trunc_page(min_ph_size); } - /* Sync up on getting the busy bit */ - if((dst_page->busy || dst_page->cleaning) && - (((ops & UPL_POP_SET) && - (ops & UPL_POP_BUSY)) || (ops & UPL_POP_DUMP))) { - /* someone else is playing with the page, we will */ - /* have to wait */ - PAGE_SLEEP(object, dst_page, THREAD_UNINT); - continue; + if (max_ph_size & PAGE_MASK_64) { + max_ph_size = trunc_page(max_ph_size); } + } - if (ops & UPL_POP_DUMP) { - vm_page_lock_queues(); + if (min_ph_size < PAGE_SIZE) + min_ph_size = PAGE_SIZE; - if (dst_page->no_isync == FALSE) - pmap_disconnect(dst_page->phys_page); - vm_page_free(dst_page); + if (max_ph_size < PAGE_SIZE) + max_ph_size = PAGE_SIZE; + else if (max_ph_size > MAX_UPL_TRANSFER_BYTES) + max_ph_size = MAX_UPL_TRANSFER_BYTES; - vm_page_unlock_queues(); - break; - } + if (max_length > max_ph_size) + max_length = max_ph_size; - if (flags) { - *flags = 0; + if (max_length <= PAGE_SIZE) + goto out; - /* Get the condition of flags before requested ops */ - /* are undertaken */ + if (object->internal) + object_size = object->vo_size; + else + vnode_pager_get_object_size(object->pager, &object_size); - if(dst_page->dirty) *flags |= UPL_POP_DIRTY; - if(dst_page->pageout) *flags |= UPL_POP_PAGEOUT; - if(dst_page->precious) *flags |= UPL_POP_PRECIOUS; - if(dst_page->absent) *flags |= UPL_POP_ABSENT; - if(dst_page->busy) *flags |= UPL_POP_BUSY; - } + object_size = round_page_64(object_size); - /* The caller should have made a call either contingent with */ - /* or prior to this call to set UPL_POP_BUSY */ - if(ops & UPL_POP_SET) { - /* The protection granted with this assert will */ - /* not be complete. If the caller violates the */ - /* convention and attempts to change page state */ - /* without first setting busy we may not see it */ - /* because the page may already be busy. However */ + if (orig_start >= object_size) { + /* + * fault occurred beyond the EOF... + * we need to punt w/o changing the + * starting offset + */ + goto out; + } + if (object->pages_used > object->pages_created) { + /* + * must have wrapped our 32 bit counters + * so reset + */ + object->pages_used = object->pages_created = 0; + } + if ((sequential_run = object->sequential)) { + if (sequential_run < 0) { + sequential_behavior = VM_BEHAVIOR_RSEQNTL; + sequential_run = 0 - sequential_run; + } else { + sequential_behavior = VM_BEHAVIOR_SEQUENTIAL; + } + + } + switch (behavior) { + + default: + behavior = VM_BEHAVIOR_DEFAULT; + + case VM_BEHAVIOR_DEFAULT: + if (object->internal && fault_info->user_tag == VM_MEMORY_STACK) + goto out; + + if (sequential_run >= (3 * PAGE_SIZE)) { + pre_heat_size = sequential_run + PAGE_SIZE; + + if (sequential_behavior == VM_BEHAVIOR_SEQUENTIAL) + look_behind = FALSE; + else + look_ahead = FALSE; + + *io_streaming = 1; + } else { + + if (object->pages_created < (20 * (min_ph_size >> PAGE_SHIFT))) { + /* + * prime the pump + */ + pre_heat_size = min_ph_size; + } else { + /* + * Linear growth in PH size: The maximum size is max_length... + * this cacluation will result in a size that is neither a + * power of 2 nor a multiple of PAGE_SIZE... so round + * it up to the nearest PAGE_SIZE boundary + */ + pre_heat_size = (max_length * (uint64_t)object->pages_used) / object->pages_created; + + if (pre_heat_size < min_ph_size) + pre_heat_size = min_ph_size; + else + pre_heat_size = round_page(pre_heat_size); + } + } + break; + + case VM_BEHAVIOR_RANDOM: + if ((pre_heat_size = cluster_size) <= PAGE_SIZE) + goto out; + break; + + case VM_BEHAVIOR_SEQUENTIAL: + if ((pre_heat_size = cluster_size) == 0) + pre_heat_size = sequential_run + PAGE_SIZE; + look_behind = FALSE; + *io_streaming = 1; + + break; + + case VM_BEHAVIOR_RSEQNTL: + if ((pre_heat_size = cluster_size) == 0) + pre_heat_size = sequential_run + PAGE_SIZE; + look_ahead = FALSE; + *io_streaming = 1; + + break; + + } + throttle_limit = (uint32_t) max_length; + assert(throttle_limit == max_length); + + if (vnode_pager_get_throttle_io_limit(object->pager, &throttle_limit) == KERN_SUCCESS) { + if (max_length > throttle_limit) + max_length = throttle_limit; + } + if (pre_heat_size > max_length) + pre_heat_size = max_length; + + if (behavior == VM_BEHAVIOR_DEFAULT && (pre_heat_size > min_ph_size)) { + + unsigned int consider_free = vm_page_free_count + vm_page_cleaned_count; + + if (consider_free < vm_page_throttle_limit) { + pre_heat_size = trunc_page(pre_heat_size / 16); + } else if (consider_free < vm_page_free_target) { + pre_heat_size = trunc_page(pre_heat_size / 4); + } + + if (pre_heat_size < min_ph_size) + pre_heat_size = min_ph_size; + } + if (look_ahead == TRUE) { + if (look_behind == TRUE) { + /* + * if we get here its due to a random access... + * so we want to center the original fault address + * within the cluster we will issue... make sure + * to calculate 'head_size' as a multiple of PAGE_SIZE... + * 'pre_heat_size' is a multiple of PAGE_SIZE but not + * necessarily an even number of pages so we need to truncate + * the result to a PAGE_SIZE boundary + */ + head_size = trunc_page(pre_heat_size / 2); + + if (target_start > head_size) + target_start -= head_size; + else + target_start = 0; + + /* + * 'target_start' at this point represents the beginning offset + * of the cluster we are considering... 'orig_start' will be in + * the center of this cluster if we didn't have to clip the start + * due to running into the start of the file + */ + } + if ((target_start + pre_heat_size) > object_size) + pre_heat_size = (vm_size_t)(round_page_64(object_size - target_start)); + /* + * at this point caclulate the number of pages beyond the original fault + * address that we want to consider... this is guaranteed not to extend beyond + * the current EOF... + */ + assert((vm_size_t)(orig_start - target_start) == (orig_start - target_start)); + tail_size = pre_heat_size - (vm_size_t)(orig_start - target_start) - PAGE_SIZE; + } else { + if (pre_heat_size > target_start) { + /* + * since pre_heat_size is always smaller then 2^32, + * if it is larger then target_start (a 64 bit value) + * it is safe to clip target_start to 32 bits + */ + pre_heat_size = (vm_size_t) target_start; + } + tail_size = 0; + } + assert( !(target_start & PAGE_MASK_64)); + assert( !(pre_heat_size & PAGE_MASK_64)); + + if (pre_heat_size <= PAGE_SIZE) + goto out; + + if (look_behind == TRUE) { + /* + * take a look at the pages before the original + * faulting offset... recalculate this in case + * we had to clip 'pre_heat_size' above to keep + * from running past the EOF. + */ + head_size = pre_heat_size - tail_size - PAGE_SIZE; + + for (offset = orig_start - PAGE_SIZE_64; head_size; offset -= PAGE_SIZE_64, head_size -= PAGE_SIZE) { + /* + * don't poke below the lowest offset + */ + if (offset < fault_info->lo_offset) + break; + /* + * for external objects or internal objects w/o a pager, + * VM_COMPRESSOR_PAGER_STATE_GET will return VM_EXTERNAL_STATE_UNKNOWN + */ + if (VM_COMPRESSOR_PAGER_STATE_GET(object, offset) == VM_EXTERNAL_STATE_ABSENT) { + break; + } + if (vm_page_lookup(object, offset) != VM_PAGE_NULL) { + /* + * don't bridge resident pages + */ + break; + } + *start = offset; + *length += PAGE_SIZE; + } + } + if (look_ahead == TRUE) { + for (offset = orig_start + PAGE_SIZE_64; tail_size; offset += PAGE_SIZE_64, tail_size -= PAGE_SIZE) { + /* + * don't poke above the highest offset + */ + if (offset >= fault_info->hi_offset) + break; + assert(offset < object_size); + + /* + * for external objects or internal objects w/o a pager, + * VM_COMPRESSOR_PAGER_STATE_GET will return VM_EXTERNAL_STATE_UNKNOWN + */ + if (VM_COMPRESSOR_PAGER_STATE_GET(object, offset) == VM_EXTERNAL_STATE_ABSENT) { + break; + } + if (vm_page_lookup(object, offset) != VM_PAGE_NULL) { + /* + * don't bridge resident pages + */ + break; + } + *length += PAGE_SIZE; + } + } +out: + if (*length > max_length) + *length = max_length; + + vm_object_unlock(object); + + DTRACE_VM1(clustersize, vm_size_t, *length); +} + + +/* + * Allow manipulation of individual page state. This is actually part of + * the UPL regimen but takes place on the VM object rather than on a UPL + */ + +kern_return_t +vm_object_page_op( + vm_object_t object, + vm_object_offset_t offset, + int ops, + ppnum_t *phys_entry, + int *flags) +{ + vm_page_t dst_page; + + vm_object_lock(object); + + if(ops & UPL_POP_PHYSICAL) { + if(object->phys_contiguous) { + if (phys_entry) { + *phys_entry = (ppnum_t) + (object->vo_shadow_offset >> PAGE_SHIFT); + } + vm_object_unlock(object); + return KERN_SUCCESS; + } else { + vm_object_unlock(object); + return KERN_INVALID_OBJECT; + } + } + if(object->phys_contiguous) { + vm_object_unlock(object); + return KERN_INVALID_OBJECT; + } + + while(TRUE) { + if((dst_page = vm_page_lookup(object,offset)) == VM_PAGE_NULL) { + vm_object_unlock(object); + return KERN_FAILURE; + } + + /* Sync up on getting the busy bit */ + if((dst_page->busy || dst_page->cleaning) && + (((ops & UPL_POP_SET) && + (ops & UPL_POP_BUSY)) || (ops & UPL_POP_DUMP))) { + /* someone else is playing with the page, we will */ + /* have to wait */ + PAGE_SLEEP(object, dst_page, THREAD_UNINT); + continue; + } + + if (ops & UPL_POP_DUMP) { + if (dst_page->pmapped == TRUE) + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(dst_page)); + + VM_PAGE_FREE(dst_page); + break; + } + + if (flags) { + *flags = 0; + + /* Get the condition of flags before requested ops */ + /* are undertaken */ + + if(dst_page->dirty) *flags |= UPL_POP_DIRTY; + if(dst_page->free_when_done) *flags |= UPL_POP_PAGEOUT; + if(dst_page->precious) *flags |= UPL_POP_PRECIOUS; + if(dst_page->absent) *flags |= UPL_POP_ABSENT; + if(dst_page->busy) *flags |= UPL_POP_BUSY; + } + + /* The caller should have made a call either contingent with */ + /* or prior to this call to set UPL_POP_BUSY */ + if(ops & UPL_POP_SET) { + /* The protection granted with this assert will */ + /* not be complete. If the caller violates the */ + /* convention and attempts to change page state */ + /* without first setting busy we may not see it */ + /* because the page may already be busy. However */ /* if such violations occur we will assert sooner */ /* or later. */ assert(dst_page->busy || (ops & UPL_POP_BUSY)); - if (ops & UPL_POP_DIRTY) dst_page->dirty = TRUE; - if (ops & UPL_POP_PAGEOUT) dst_page->pageout = TRUE; + if (ops & UPL_POP_DIRTY) { + SET_PAGE_DIRTY(dst_page, FALSE); + } + if (ops & UPL_POP_PAGEOUT) dst_page->free_when_done = TRUE; if (ops & UPL_POP_PRECIOUS) dst_page->precious = TRUE; if (ops & UPL_POP_ABSENT) dst_page->absent = TRUE; if (ops & UPL_POP_BUSY) dst_page->busy = TRUE; @@ -5710,7 +7875,7 @@ vm_object_page_op( if(ops & UPL_POP_CLR) { assert(dst_page->busy); if (ops & UPL_POP_DIRTY) dst_page->dirty = FALSE; - if (ops & UPL_POP_PAGEOUT) dst_page->pageout = FALSE; + if (ops & UPL_POP_PAGEOUT) dst_page->free_when_done = FALSE; if (ops & UPL_POP_PRECIOUS) dst_page->precious = FALSE; if (ops & UPL_POP_ABSENT) dst_page->absent = FALSE; if (ops & UPL_POP_BUSY) { @@ -5759,7 +7924,7 @@ vm_object_page_op( */ assert(dst_page->busy); assert(!dst_page->encrypted); - *phys_entry = dst_page->phys_page; + *phys_entry = VM_PAGE_GET_PHYS_PAGE(dst_page); } break; @@ -5786,17 +7951,23 @@ vm_object_range_op( vm_object_offset_t offset_beg, vm_object_offset_t offset_end, int ops, - int *range) + uint32_t *range) { vm_object_offset_t offset; vm_page_t dst_page; + if (offset_end - offset_beg > (uint32_t) -1) { + /* range is too big and would overflow "*range" */ + return KERN_INVALID_ARGUMENT; + } if (object->resident_page_count == 0) { if (range) { - if (ops & UPL_ROP_PRESENT) + if (ops & UPL_ROP_PRESENT) { *range = 0; - else - *range = offset_end - offset_beg; + } else { + *range = (uint32_t) (offset_end - offset_beg); + assert(*range == (offset_end - offset_beg)); + } } return KERN_SUCCESS; } @@ -5807,19 +7978,18 @@ vm_object_range_op( return KERN_INVALID_OBJECT; } - offset = offset_beg; + offset = offset_beg & ~PAGE_MASK_64; while (offset < offset_end) { dst_page = vm_page_lookup(object, offset); if (dst_page != VM_PAGE_NULL) { if (ops & UPL_ROP_DUMP) { if (dst_page->busy || dst_page->cleaning) { - /* + /* * someone else is playing with the * page, we will have to wait */ - PAGE_SLEEP(object, - dst_page, THREAD_UNINT); + PAGE_SLEEP(object, dst_page, THREAD_UNINT); /* * need to relook the page up since it's * state may have changed while we slept @@ -5828,15 +7998,18 @@ vm_object_range_op( */ continue; } - vm_page_lock_queues(); + if (dst_page->laundry) + vm_pageout_steal_laundry(dst_page, FALSE); - if (dst_page->no_isync == FALSE) - pmap_disconnect(dst_page->phys_page); - vm_page_free(dst_page); + if (dst_page->pmapped == TRUE) + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(dst_page)); - vm_page_unlock_queues(); - } else if (ops & UPL_ROP_ABSENT) - break; + VM_PAGE_FREE(dst_page); + + } else if ((ops & UPL_ROP_ABSENT) + && (!dst_page->absent || dst_page->busy)) { + break; + } } else if (ops & UPL_ROP_PRESENT) break; @@ -5844,8 +8017,812 @@ vm_object_range_op( } vm_object_unlock(object); - if (range) - *range = offset - offset_beg; - + if (range) { + if (offset > offset_end) + offset = offset_end; + if(offset > offset_beg) { + *range = (uint32_t) (offset - offset_beg); + assert(*range == (offset - offset_beg)); + } else { + *range = 0; + } + } return KERN_SUCCESS; } + +/* + * Used to point a pager directly to a range of memory (when the pager may be associated + * with a non-device vnode). Takes a virtual address, an offset, and a size. We currently + * expect that the virtual address will denote the start of a range that is physically contiguous. + */ +kern_return_t pager_map_to_phys_contiguous( + memory_object_control_t object, + memory_object_offset_t offset, + addr64_t base_vaddr, + vm_size_t size) +{ + ppnum_t page_num; + boolean_t clobbered_private; + kern_return_t retval; + vm_object_t pager_object; + + page_num = pmap_find_phys(kernel_pmap, base_vaddr); + + if (!page_num) { + retval = KERN_FAILURE; + goto out; + } + + pager_object = memory_object_control_to_vm_object(object); + + if (!pager_object) { + retval = KERN_FAILURE; + goto out; + } + + clobbered_private = pager_object->private; + if (pager_object->private != TRUE) { + vm_object_lock(pager_object); + pager_object->private = TRUE; + vm_object_unlock(pager_object); + } + retval = vm_object_populate_with_private(pager_object, offset, page_num, size); + + if (retval != KERN_SUCCESS) { + if (pager_object->private != clobbered_private) { + vm_object_lock(pager_object); + pager_object->private = clobbered_private; + vm_object_unlock(pager_object); + } + } + +out: + return retval; +} + +uint32_t scan_object_collision = 0; + +void +vm_object_lock(vm_object_t object) +{ + if (object == vm_pageout_scan_wants_object) { + scan_object_collision++; + mutex_pause(2); + } + lck_rw_lock_exclusive(&object->Lock); +#if DEVELOPMENT || DEBUG + object->Lock_owner = current_thread(); +#endif +} + +boolean_t +vm_object_lock_avoid(vm_object_t object) +{ + if (object == vm_pageout_scan_wants_object) { + scan_object_collision++; + return TRUE; + } + return FALSE; +} + +boolean_t +_vm_object_lock_try(vm_object_t object) +{ + boolean_t retval; + + retval = lck_rw_try_lock_exclusive(&object->Lock); +#if DEVELOPMENT || DEBUG + if (retval == TRUE) + object->Lock_owner = current_thread(); +#endif + return (retval); +} + +boolean_t +vm_object_lock_try(vm_object_t object) +{ + /* + * Called from hibernate path so check before blocking. + */ + if (vm_object_lock_avoid(object) && ml_get_interrupts_enabled() && get_preemption_level()==0) { + mutex_pause(2); + } + return _vm_object_lock_try(object); +} + +void +vm_object_lock_shared(vm_object_t object) +{ + if (vm_object_lock_avoid(object)) { + mutex_pause(2); + } + lck_rw_lock_shared(&object->Lock); +} + +boolean_t +vm_object_lock_try_shared(vm_object_t object) +{ + if (vm_object_lock_avoid(object)) { + mutex_pause(2); + } + return (lck_rw_try_lock_shared(&object->Lock)); +} + +boolean_t +vm_object_lock_upgrade(vm_object_t object) +{ boolean_t retval; + + retval = lck_rw_lock_shared_to_exclusive(&object->Lock); +#if DEVELOPMENT || DEBUG + if (retval == TRUE) + object->Lock_owner = current_thread(); +#endif + return (retval); +} + +void +vm_object_unlock(vm_object_t object) +{ +#if DEVELOPMENT || DEBUG + if (object->Lock_owner) { + if (object->Lock_owner != current_thread()) + panic("vm_object_unlock: not owner - %p\n", object); + object->Lock_owner = 0; + } +#endif + lck_rw_done(&object->Lock); +} + + +unsigned int vm_object_change_wimg_mode_count = 0; + +/* + * The object must be locked + */ +void +vm_object_change_wimg_mode(vm_object_t object, unsigned int wimg_mode) +{ + vm_page_t p; + + vm_object_lock_assert_exclusive(object); + + vm_object_paging_wait(object, THREAD_UNINT); + + vm_page_queue_iterate(&object->memq, p, vm_page_t, listq) { + + if (!p->fictitious) + pmap_set_cache_attributes(VM_PAGE_GET_PHYS_PAGE(p), wimg_mode); + } + if (wimg_mode == VM_WIMG_USE_DEFAULT) + object->set_cache_attr = FALSE; + else + object->set_cache_attr = TRUE; + + object->wimg_bits = wimg_mode; + + vm_object_change_wimg_mode_count++; +} + +#if CONFIG_FREEZE + +/* + * This routine does the "relocation" of previously + * compressed pages belonging to this object that are + * residing in a number of compressed segments into + * a set of compressed segments dedicated to hold + * compressed pages belonging to this object. + */ + +extern void *freezer_chead; +extern char *freezer_compressor_scratch_buf; +extern int c_freezer_compression_count; +extern AbsoluteTime c_freezer_last_yield_ts; + +#define MAX_FREE_BATCH 32 +#define FREEZER_DUTY_CYCLE_ON_MS 5 +#define FREEZER_DUTY_CYCLE_OFF_MS 5 + +static int c_freezer_should_yield(void); + + +static int +c_freezer_should_yield() +{ + AbsoluteTime cur_time; + uint64_t nsecs; + + assert(c_freezer_last_yield_ts); + clock_get_uptime(&cur_time); + + SUB_ABSOLUTETIME(&cur_time, &c_freezer_last_yield_ts); + absolutetime_to_nanoseconds(cur_time, &nsecs); + + if (nsecs > 1000 * 1000 * FREEZER_DUTY_CYCLE_ON_MS) + return (1); + return (0); +} + + +void +vm_object_compressed_freezer_done() +{ + vm_compressor_finished_filling(&freezer_chead); +} + + +void +vm_object_compressed_freezer_pageout( + vm_object_t object) +{ + vm_page_t p; + vm_page_t local_freeq = NULL; + int local_freed = 0; + kern_return_t retval = KERN_SUCCESS; + int obj_resident_page_count_snapshot = 0; + + assert(object != VM_OBJECT_NULL); + assert(object->internal); + + vm_object_lock(object); + + if (!object->pager_initialized || object->pager == MEMORY_OBJECT_NULL) { + + if (!object->pager_initialized) { + + vm_object_collapse(object, (vm_object_offset_t) 0, TRUE); + + if (!object->pager_initialized) + vm_object_compressor_pager_create(object); + } + + if (!object->pager_initialized || object->pager == MEMORY_OBJECT_NULL) { + vm_object_unlock(object); + return; + } + } + + if (VM_CONFIG_FREEZER_SWAP_IS_ACTIVE) { + vm_object_offset_t curr_offset = 0; + + /* + * Go through the object and make sure that any + * previously compressed pages are relocated into + * a compressed segment associated with our "freezer_chead". + */ + while (curr_offset < object->vo_size) { + + curr_offset = vm_compressor_pager_next_compressed(object->pager, curr_offset); + + if (curr_offset == (vm_object_offset_t) -1) + break; + + retval = vm_compressor_pager_relocate(object->pager, curr_offset, &freezer_chead); + + if (retval != KERN_SUCCESS) + break; + + curr_offset += PAGE_SIZE_64; + } + } + + /* + * We can't hold the object lock while heading down into the compressed pager + * layer because we might need the kernel map lock down there to allocate new + * compressor data structures. And if this same object is mapped in the kernel + * and there's a fault on it, then that thread will want the object lock while + * holding the kernel map lock. + * + * Since we are going to drop/grab the object lock repeatedly, we must make sure + * we won't be stuck in an infinite loop if the same page(s) keep getting + * decompressed. So we grab a snapshot of the number of pages in the object and + * we won't process any more than that number of pages. + */ + + obj_resident_page_count_snapshot = object->resident_page_count; + + vm_object_activity_begin(object); + + while ((obj_resident_page_count_snapshot--) && !vm_page_queue_empty(&object->memq)) { + + p = (vm_page_t)vm_page_queue_first(&object->memq); + + KERNEL_DEBUG(0xe0430004 | DBG_FUNC_START, object, local_freed, 0, 0, 0); + + vm_page_lockspin_queues(); + + if (p->cleaning || p->fictitious || p->busy || p->absent || p->unusual || p->error || VM_PAGE_WIRED(p)) { + + vm_page_unlock_queues(); + + KERNEL_DEBUG(0xe0430004 | DBG_FUNC_END, object, local_freed, 1, 0, 0); + + vm_page_queue_remove(&object->memq, p, vm_page_t, listq); + vm_page_queue_enter(&object->memq, p, vm_page_t, listq); + + continue; + } + + if (p->pmapped == TRUE) { + int refmod_state, pmap_flags; + + if (p->dirty || p->precious) { + pmap_flags = PMAP_OPTIONS_COMPRESSOR; + } else { + pmap_flags = PMAP_OPTIONS_COMPRESSOR_IFF_MODIFIED; + } + + refmod_state = pmap_disconnect_options(VM_PAGE_GET_PHYS_PAGE(p), pmap_flags, NULL); + if (refmod_state & VM_MEM_MODIFIED) { + SET_PAGE_DIRTY(p, FALSE); + } + } + + if (p->dirty == FALSE && p->precious == FALSE) { + /* + * Clean and non-precious page. + */ + vm_page_unlock_queues(); + VM_PAGE_FREE(p); + + KERNEL_DEBUG(0xe0430004 | DBG_FUNC_END, object, local_freed, 2, 0, 0); + continue; + } + + if (p->laundry) + vm_pageout_steal_laundry(p, TRUE); + + vm_page_queues_remove(p, TRUE); + + vm_page_unlock_queues(); + + + /* + * In case the compressor fails to compress this page, we need it at + * the back of the object memq so that we don't keep trying to process it. + * Make the move here while we have the object lock held. + */ + + vm_page_queue_remove(&object->memq, p, vm_page_t, listq); + vm_page_queue_enter(&object->memq, p, vm_page_t, listq); + + /* + * Grab an activity_in_progress here for vm_pageout_compress_page() to consume. + * + * Mark the page busy so no one messes with it while we have the object lock dropped. + */ + + p->busy = TRUE; + + vm_object_activity_begin(object); + + vm_object_unlock(object); + + /* + * arg3 == FALSE tells vm_pageout_compress_page that we don't hold the object lock and the pager may not be initialized. + */ + if (vm_pageout_compress_page(&freezer_chead, freezer_compressor_scratch_buf, p, FALSE) == KERN_SUCCESS) { + /* + * page has already been un-tabled from the object via 'vm_page_remove' + */ + p->snext = local_freeq; + local_freeq = p; + local_freed++; + + if (local_freed >= MAX_FREE_BATCH) { + + vm_page_free_list(local_freeq, TRUE); + + local_freeq = NULL; + local_freed = 0; + } + c_freezer_compression_count++; + } + KERNEL_DEBUG(0xe0430004 | DBG_FUNC_END, object, local_freed, 0, 0, 0); + + if (local_freed == 0 && c_freezer_should_yield()) { + + thread_yield_internal(FREEZER_DUTY_CYCLE_OFF_MS); + clock_get_uptime(&c_freezer_last_yield_ts); + } + + vm_object_lock(object); + } + + if (local_freeq) { + vm_page_free_list(local_freeq, TRUE); + + local_freeq = NULL; + local_freed = 0; + } + + vm_object_activity_end(object); + + vm_object_unlock(object); + + if (c_freezer_should_yield()) { + + thread_yield_internal(FREEZER_DUTY_CYCLE_OFF_MS); + clock_get_uptime(&c_freezer_last_yield_ts); + } +} + +#endif /* CONFIG_FREEZE */ + + +void +vm_object_pageout( + vm_object_t object) +{ + vm_page_t p, next; + struct vm_pageout_queue *iq; + boolean_t need_unlock = TRUE; + + if (!VM_CONFIG_COMPRESSOR_IS_PRESENT) + return; + + iq = &vm_pageout_queue_internal; + + assert(object != VM_OBJECT_NULL ); + + vm_object_lock(object); + + if (!object->internal || + object->terminating || + !object->alive) { + vm_object_unlock(object); + return; + } + + if (!object->pager_initialized || object->pager == MEMORY_OBJECT_NULL) { + + if (!object->pager_initialized) { + + vm_object_collapse(object, (vm_object_offset_t) 0, TRUE); + + if (!object->pager_initialized) + vm_object_compressor_pager_create(object); + } + + if (!object->pager_initialized || object->pager == MEMORY_OBJECT_NULL) { + vm_object_unlock(object); + return; + } + } + +ReScan: + next = (vm_page_t)vm_page_queue_first(&object->memq); + + while (!vm_page_queue_end(&object->memq, (vm_page_queue_entry_t)next)) { + p = next; + next = (vm_page_t)vm_page_queue_next(&next->listq); + + assert(p->vm_page_q_state != VM_PAGE_ON_FREE_Q); + + if ((p->vm_page_q_state == VM_PAGE_ON_THROTTLED_Q) || + p->encrypted_cleaning || + p->cleaning || + p->laundry || + p->busy || + p->absent || + p->error || + p->fictitious || + VM_PAGE_WIRED(p)) { + /* + * Page is already being cleaned or can't be cleaned. + */ + continue; + } + + /* Throw to the pageout queue */ + + vm_page_lockspin_queues(); + need_unlock = TRUE; + + if (vm_compressor_low_on_space()) { + vm_page_unlock_queues(); + break; + } + + if (VM_PAGE_Q_THROTTLED(iq)) { + + iq->pgo_draining = TRUE; + + assert_wait((event_t) (&iq->pgo_laundry + 1), + THREAD_INTERRUPTIBLE); + vm_page_unlock_queues(); + vm_object_unlock(object); + + thread_block(THREAD_CONTINUE_NULL); + + vm_object_lock(object); + goto ReScan; + } + + assert(!p->fictitious); + assert(!p->busy); + assert(!p->absent); + assert(!p->unusual); + assert(!p->error); + assert(!VM_PAGE_WIRED(p)); + assert(!p->cleaning); + + if (p->pmapped == TRUE) { + int refmod_state; + int pmap_options; + + /* + * Tell pmap the page should be accounted + * for as "compressed" if it's been modified. + */ + pmap_options = + PMAP_OPTIONS_COMPRESSOR_IFF_MODIFIED; + if (p->dirty || p->precious) { + /* + * We already know it's been modified, + * so tell pmap to account for it + * as "compressed". + */ + pmap_options = PMAP_OPTIONS_COMPRESSOR; + } + refmod_state = pmap_disconnect_options(VM_PAGE_GET_PHYS_PAGE(p), + pmap_options, + NULL); + if (refmod_state & VM_MEM_MODIFIED) { + SET_PAGE_DIRTY(p, FALSE); + } + } + + if (!p->dirty && !p->precious) { + vm_page_unlock_queues(); + VM_PAGE_FREE(p); + continue; + } + + vm_page_queues_remove(p, TRUE); + + if (vm_pageout_cluster(p, FALSE, TRUE)) + need_unlock = FALSE; + + if (need_unlock == TRUE) + vm_page_unlock_queues(); + } + + vm_object_unlock(object); +} + + +#if CONFIG_IOSCHED +void +vm_page_request_reprioritize(vm_object_t o, uint64_t blkno, uint32_t len, int prio) +{ + io_reprioritize_req_t req; + struct vnode *devvp = NULL; + + if(vnode_pager_get_object_devvp(o->pager, (uintptr_t *)&devvp) != KERN_SUCCESS) + return; + + /* + * Create the request for I/O reprioritization. + * We use the noblock variant of zalloc because we're holding the object + * lock here and we could cause a deadlock in low memory conditions. + */ + req = (io_reprioritize_req_t)zalloc_noblock(io_reprioritize_req_zone); + if (req == NULL) + return; + req->blkno = blkno; + req->len = len; + req->priority = prio; + req->devvp = devvp; + + /* Insert request into the reprioritization list */ + IO_REPRIORITIZE_LIST_LOCK(); + queue_enter(&io_reprioritize_list, req, io_reprioritize_req_t, io_reprioritize_list); + IO_REPRIORITIZE_LIST_UNLOCK(); + + /* Wakeup reprioritize thread */ + IO_REPRIO_THREAD_WAKEUP(); + + return; +} + +void +vm_decmp_upl_reprioritize(upl_t upl, int prio) +{ + int offset; + vm_object_t object; + io_reprioritize_req_t req; + struct vnode *devvp = NULL; + uint64_t blkno; + uint32_t len; + upl_t io_upl; + uint64_t *io_upl_reprio_info; + int io_upl_size; + + if ((upl->flags & UPL_TRACKED_BY_OBJECT) == 0 || (upl->flags & UPL_EXPEDITE_SUPPORTED) == 0) + return; + + /* + * We dont want to perform any allocations with the upl lock held since that might + * result in a deadlock. If the system is low on memory, the pageout thread would + * try to pageout stuff and might wait on this lock. If we are waiting for the memory to + * be freed up by the pageout thread, it would be a deadlock. + */ + + + /* First step is just to get the size of the upl to find out how big the reprio info is */ + if(!upl_try_lock(upl)) + return; + + if (upl->decmp_io_upl == NULL) { + /* The real I/O upl was destroyed by the time we came in here. Nothing to do. */ + upl_unlock(upl); + return; + } + + io_upl = upl->decmp_io_upl; + assert((io_upl->flags & UPL_DECMP_REAL_IO) != 0); + io_upl_size = io_upl->size; + upl_unlock(upl); + + /* Now perform the allocation */ + io_upl_reprio_info = (uint64_t *)kalloc(sizeof(uint64_t) * (io_upl_size / PAGE_SIZE)); + if (io_upl_reprio_info == NULL) + return; + + /* Now again take the lock, recheck the state and grab out the required info */ + if(!upl_try_lock(upl)) + goto out; + + if (upl->decmp_io_upl == NULL || upl->decmp_io_upl != io_upl) { + /* The real I/O upl was destroyed by the time we came in here. Nothing to do. */ + upl_unlock(upl); + goto out; + } + memcpy(io_upl_reprio_info, io_upl->upl_reprio_info, sizeof(uint64_t) * (io_upl_size / PAGE_SIZE)); + + /* Get the VM object for this UPL */ + if (io_upl->flags & UPL_SHADOWED) { + object = io_upl->map_object->shadow; + } else { + object = io_upl->map_object; + } + + /* Get the dev vnode ptr for this object */ + if(!object || !object->pager || + vnode_pager_get_object_devvp(object->pager, (uintptr_t *)&devvp) != KERN_SUCCESS) { + upl_unlock(upl); + goto out; + } + + upl_unlock(upl); + + /* Now we have all the information needed to do the expedite */ + + offset = 0; + while (offset < io_upl_size) { + blkno = io_upl_reprio_info[(offset / PAGE_SIZE)] & UPL_REPRIO_INFO_MASK; + len = (io_upl_reprio_info[(offset / PAGE_SIZE)] >> UPL_REPRIO_INFO_SHIFT) & UPL_REPRIO_INFO_MASK; + + /* + * This implementation may cause some spurious expedites due to the + * fact that we dont cleanup the blkno & len from the upl_reprio_info + * even after the I/O is complete. + */ + + if (blkno != 0 && len != 0) { + /* Create the request for I/O reprioritization */ + req = (io_reprioritize_req_t)zalloc(io_reprioritize_req_zone); + assert(req != NULL); + req->blkno = blkno; + req->len = len; + req->priority = prio; + req->devvp = devvp; + + /* Insert request into the reprioritization list */ + IO_REPRIORITIZE_LIST_LOCK(); + queue_enter(&io_reprioritize_list, req, io_reprioritize_req_t, io_reprioritize_list); + IO_REPRIORITIZE_LIST_UNLOCK(); + + offset += len; + } else { + offset += PAGE_SIZE; + } + } + + /* Wakeup reprioritize thread */ + IO_REPRIO_THREAD_WAKEUP(); + +out: + kfree(io_upl_reprio_info, sizeof(uint64_t) * (io_upl_size / PAGE_SIZE)); + return; +} + +void +vm_page_handle_prio_inversion(vm_object_t o, vm_page_t m) +{ + upl_t upl; + upl_page_info_t *pl; + unsigned int i, num_pages; + int cur_tier; + + cur_tier = proc_get_effective_thread_policy(current_thread(), TASK_POLICY_IO); + + /* + Scan through all UPLs associated with the object to find the + UPL containing the contended page. + */ + queue_iterate(&o->uplq, upl, upl_t, uplq) { + if (((upl->flags & UPL_EXPEDITE_SUPPORTED) == 0) || upl->upl_priority <= cur_tier) + continue; + pl = UPL_GET_INTERNAL_PAGE_LIST(upl); + num_pages = (upl->size / PAGE_SIZE); + + /* + For each page in the UPL page list, see if it matches the contended + page and was issued as a low prio I/O. + */ + for(i=0; i < num_pages; i++) { + if(UPL_PAGE_PRESENT(pl,i) && VM_PAGE_GET_PHYS_PAGE(m) == pl[i].phys_addr) { + if ((upl->flags & UPL_DECMP_REQ) && upl->decmp_io_upl) { + KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_PAGE_EXPEDITE)) | DBG_FUNC_NONE, VM_KERNEL_UNSLIDE_OR_PERM(upl->upl_creator), VM_KERNEL_UNSLIDE_OR_PERM(m), + VM_KERNEL_UNSLIDE_OR_PERM(upl), upl->upl_priority, 0); + vm_decmp_upl_reprioritize(upl, cur_tier); + break; + } + KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_PAGE_EXPEDITE)) | DBG_FUNC_NONE, VM_KERNEL_UNSLIDE_OR_PERM(upl->upl_creator), VM_KERNEL_UNSLIDE_OR_PERM(m), + upl->upl_reprio_info[i], upl->upl_priority, 0); + if (UPL_REPRIO_INFO_BLKNO(upl, i) != 0 && UPL_REPRIO_INFO_LEN(upl, i) != 0) + vm_page_request_reprioritize(o, UPL_REPRIO_INFO_BLKNO(upl, i), UPL_REPRIO_INFO_LEN(upl, i), cur_tier); + break; + } + } + /* Check if we found any hits */ + if (i != num_pages) + break; + } + + return; +} + +wait_result_t +vm_page_sleep(vm_object_t o, vm_page_t m, int interruptible) +{ + wait_result_t ret; + + KERNEL_DEBUG((MACHDBG_CODE(DBG_MACH_VM, VM_PAGE_SLEEP)) | DBG_FUNC_START, o, m, 0, 0, 0); + + if (o->io_tracking && ((m->busy == TRUE) || (m->cleaning == TRUE) || VM_PAGE_WIRED(m))) { + /* + Indicates page is busy due to an I/O. Issue a reprioritize request if necessary. + */ + vm_page_handle_prio_inversion(o,m); + } + m->wanted = TRUE; + ret = thread_sleep_vm_object(o, m, interruptible); + KERNEL_DEBUG((MACHDBG_CODE(DBG_MACH_VM, VM_PAGE_SLEEP)) | DBG_FUNC_END, o, m, 0, 0, 0); + return ret; +} + +static void +io_reprioritize_thread(void *param __unused, wait_result_t wr __unused) +{ + io_reprioritize_req_t req = NULL; + + while(1) { + + IO_REPRIORITIZE_LIST_LOCK(); + if (queue_empty(&io_reprioritize_list)) { + IO_REPRIORITIZE_LIST_UNLOCK(); + break; + } + + queue_remove_first(&io_reprioritize_list, req, io_reprioritize_req_t, io_reprioritize_list); + IO_REPRIORITIZE_LIST_UNLOCK(); + + vnode_pager_issue_reprioritize_io(req->devvp, req->blkno, req->len, req->priority); + zfree(io_reprioritize_req_zone, req); + } + + IO_REPRIO_THREAD_CONTINUATION(); +} +#endif