/*
- * Copyright (c) 2000-2001 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@
* Virtual memory object module.
*/
-#ifdef MACH_BSD
-/* remove as part of compoenent support merge */
-extern int vnode_pager_workaround;
-#endif
-
+#include <debug.h>
#include <mach_pagemap.h>
#include <task_swapper.h>
#include <mach/memory_object_default.h>
#include <mach/memory_object_control_server.h>
#include <mach/vm_param.h>
+
+#include <mach/sdt.h>
+
+#include <ipc/ipc_types.h>
#include <ipc/ipc_port.h>
+
+#include <kern/kern_types.h>
#include <kern/assert.h>
-#include <kern/lock.h>
#include <kern/queue.h>
#include <kern/xpr.h>
+#include <kern/kalloc.h>
#include <kern/zalloc.h>
#include <kern/host.h>
#include <kern/host_statistics.h>
#include <kern/processor.h>
+#include <kern/misc_protos.h>
+
#include <vm/memory_object.h>
+#include <vm/vm_compressor_pager.h>
#include <vm/vm_fault.h>
#include <vm/vm_map.h>
#include <vm/vm_object.h>
#include <vm/vm_page.h>
#include <vm/vm_pageout.h>
-#include <kern/misc_protos.h>
+#include <vm/vm_protos.h>
+#include <vm/vm_purgeable_internal.h>
+
+#include <vm/vm_compressor.h>
+
+#if CONFIG_PHANTOM_CACHE
+#include <vm/vm_phantom_cache.h>
+#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;
+decl_simple_lock_data(static,vm_object_tracking_lock_data);
+btlog_t *vm_object_tracking_btlog;
+static void
+vm_object_tracking_lock(void *context)
+{
+ simple_lock((simple_lock_t)context);
+}
+static void
+vm_object_tracking_unlock(void *context)
+{
+ simple_unlock((simple_lock_t)context);
+}
+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) {
+ simple_lock_init(&vm_object_tracking_lock_data, 0);
+ vm_object_tracking_btlog = btlog_create(
+ 50000,
+ VM_OBJECT_TRACKING_BTDEPTH,
+ vm_object_tracking_lock,
+ vm_object_tracking_unlock,
+ &vm_object_tracking_lock_data);
+ assert(vm_object_tracking_btlog);
+ vm_object_tracking_inited = TRUE;
+ }
+}
+#endif /* VM_OBJECT_TRACKING */
/*
* Virtual memory objects maintain the actual data
*/
/* Forward declarations for internal functions. */
-static void _vm_object_allocate(
- vm_object_size_t size,
- vm_object_t object);
-
static kern_return_t vm_object_terminate(
vm_object_t object);
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 void vm_object_abort_activity(
- vm_object_t object);
-
static kern_return_t vm_object_copy_call(
vm_object_t src_object,
vm_object_offset_t src_offset,
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 */
* memory object (kernel_object) to avoid wasting data structures.
*/
static struct vm_object kernel_object_store;
-__private_extern__ vm_object_t kernel_object = &kernel_object_store;
+vm_object_t kernel_object;
+
+static struct vm_object compressor_object_store;
+vm_object_t compressor_object = &compressor_object_store;
/*
* The submap object is used as a placeholder for vm_map_submap
*/
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
* 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 */
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);
+
+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 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);
}
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
vm_object_hash_entry_free(
vm_object_hash_entry_t entry)
{
- zfree(vm_object_hash_zone, (vm_offset_t)entry);
+ zfree(vm_object_hash_zone, entry);
}
/*
* Returns a new object with the given size.
*/
-static void
+__private_extern__ void
_vm_object_allocate(
vm_object_size_t size,
vm_object_t object)
{
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);
queue_init(&object->msr_q);
-#ifdef UBC_DEBUG
+#if UPL_DEBUG || CONFIG_IOSCHED
queue_init(&object->uplq);
-#endif /* UBC_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
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:
*
__private_extern__ void
vm_object_bootstrap(void)
{
- register i;
+ register int i;
vm_object_zone = zinit((vm_size_t) sizeof(struct vm_object),
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, ETAP_VM_OBJ_CACHE);
+
+ 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(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.frozen_size = 0;
+ vm_object_template.memq.prev = NULL;
+ vm_object_template.memq.next = NULL;
+#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
+ 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 = 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_request = PAGER_REQUEST_NULL;
- /* msr_q; init after allocation */
-
+ vm_object_template.pager_control = MEMORY_OBJECT_CONTROL_NULL;
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 */
vm_object_template.private = FALSE;
vm_object_template.pageout = FALSE;
vm_object_template.alive = TRUE;
- vm_object_template.lock_in_progress = FALSE;
- vm_object_template.lock_restart = FALSE;
- 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;
+ 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
#if MACH_PAGEMAP
vm_object_template.existence_map = VM_EXTERNAL_NULL;
#endif /* MACH_PAGEMAP */
+ 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;
+
+#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"
*/
/*
* Note that in the following size specifications, we need to add 1 because
- * VM_MAX_KERNEL_ADDRESS is a maximum address, not a size.
+ * VM_MAX_KERNEL_ADDRESS (vm_last_addr) is a maximum address, not a size.
*/
- _vm_object_allocate((VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) + 1,
- kernel_object);
+
+ _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
*/
vm_submap_object = &vm_submap_object_store;
- _vm_object_allocate((VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) + 1,
- vm_submap_object);
+ _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1,
+ vm_submap_object);
+ vm_submap_object->copy_strategy = MEMORY_OBJECT_COPY_NONE;
+
/*
* Create an "extra" reference to this object so that we never
* try to deallocate it; zfree doesn't like to be called with
#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)
+{
+ kern_return_t kr;
+ thread_t thread;
+
+ kr = kernel_thread_start_priority(
+ (thread_continue_t) vm_object_reaper_thread,
+ NULL,
+ BASEPRI_PREEMPT - 1,
+ &thread);
+ if (kr != KERN_SUCCESS) {
+ panic("failed to launch vm_object_reaper_thread kr=0x%x", kr);
+ }
+ thread_deallocate(thread);
+}
+
__private_extern__ void
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:
*
* 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)
{
- boolean_t retry_cache_trim = FALSE;
- vm_object_t shadow;
+#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) {
/*
- * The cache holds a reference (uncounted) to
- * the object; we must lock it before removing
- * the object.
+ * 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 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.
+ */
+ 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) {
+ /*
+ * ref_count was updated atomically !
+ */
+ 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_cache_lock();
vm_object_lock(object);
assert(object->ref_count > 0);
/* 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);
- vm_object_cache_lock();
+ memory_object_last_unmap(pager);
+
vm_object_lock(object);
- assert(object->ref_count > 0);
+ vm_object_mapping_end(object);
}
+ assert(object->ref_count > 0);
}
/*
/* 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_unlock(object);
- vm_object_cache_unlock();
+
+ if (object->ref_count == 1 &&
+ object->shadow != VM_OBJECT_NULL) {
+ /*
+ * There's only one reference left on this
+ * VM object. We can't tell if it's a valid
+ * one (from a mapping for example) or if this
+ * object is just part of a possibly stale and
+ * useless shadow chain.
+ * We would like to try and collapse it into
+ * its parent, but we don't have any pointers
+ * back to this parent object.
+ * But we can try and collapse this object with
+ * its own shadows, in case these are useless
+ * too...
+ * We can't bypass this object though, since we
+ * don't know if this last reference on it is
+ * meaningful or not.
+ */
+ 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;
}
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.
* 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)) {
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);
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%lX ref %d\n",
- (integer_t)object, object->resident_page_count,
- object->paging_in_progress,
- (natural_t)current_thread(),object->ref_count);
+ "vm_o_deallocate: !cacheable 0x%X res %d paging_ops %d thread 0x%p ref %d\n",
+ object, object->resident_page_count,
+ object->paging_in_progress,
+ (void *)current_thread(),object->ref_count);
VM_OBJ_RES_DECR(object); /* XXX ? */
/*
* 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)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 (!queue_end(&object->memq, (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)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;
+ if (p->pmapped || p->dirty || p->precious) {
+ vm_page_lockspin_queues();
+
+ if (p->pmapped) {
+ int refmod_state;
+
+ vm_object_page_grab_pmapped++;
+
+ if (p->reference == FALSE || p->dirty == FALSE) {
+
+ refmod_state = pmap_get_refmod(p->phys_page);
+
+ 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(p->phys_page);
+
+ 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->inactive && p->reference == TRUE) {
+ vm_page_activate(p);
+
+ VM_STAT_INCR(reactivations);
+ vm_object_page_grab_reactivations++;
+ }
+ vm_page_unlock_queues();
+move_page_in_obj:
+ queue_remove(&object->memq, p, vm_page_t, listq);
+ 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)
+{
+ 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;
+
+ 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);
+
+ 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 (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)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 (!queue_end(&object->memq, (queue_entry_t)next_p) && object->vo_cache_pages_to_scan && ep_count < ep_limit) {
+
+ p = next_p;
+ next_p = (vm_page_t)queue_next(&next_p->listq);
+
+ object->vo_cache_pages_to_scan--;
+
+ if (VM_PAGE_WIRED(p) || p->busy || p->cleaning || p->laundry) {
+ queue_remove(&object->memq, p, vm_page_t, listq);
+ queue_enter(&object->memq, p, vm_page_t, listq);
+
+ ep_skipped++;
+ continue;
+ }
+ if (p->wpmapped || p->dirty || p->precious) {
+ queue_remove(&object->memq, p, vm_page_t, listq);
+ queue_enter(&object->memq, p, vm_page_t, listq);
+
+ pmap_clear_reference(p->phys_page);
+ }
+ 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
+ */
+ assert(!p->pageout_queue);
+
+ vm_page_queues_remove(p);
+ 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 == NULL && p->pageq.prev == NULL);
+ /*
+ * Add this page to our list of reclaimed pages,
+ * to be freed later.
+ */
+ p->pageq.next = (queue_entry_t) 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)
+{
+ register 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.
assert(object->pager_initialized);
assert(object->ref_count == 0);
+ vm_object_lock_assert_exclusive(object);
object->ref_count++;
/*
* (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;
}
}
}
+#endif
-boolean_t vm_object_terminate_remove_all = FALSE;
/*
* Routine: vm_object_terminate
*/
static kern_return_t
vm_object_terminate(
- register vm_object_t object)
+ vm_object_t object)
{
- memory_object_t pager;
- 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)) {
+ 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
*/
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();
- 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_addr);
-
- if ((p->dirty || p->precious) && !p->error && object->alive) {
- p->busy = TRUE;
- vm_object_paging_begin(object);
- /* protect the object from re-use/caching while it */
- /* is unlocked */
- vm_object_unlock(object);
- vm_pageout_cluster(p); /* flush page */
- vm_object_lock(object);
- 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;
}
* 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;
}
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.
+ * copy. The reference we hold on the shadow must be dropped
+ * by our caller.
*/
if (((shadow_object = object->shadow) != VM_OBJECT_NULL) &&
!(object->pageout)) {
vm_object_lock(shadow_object);
- assert((shadow_object->copy == object) ||
- (shadow_object->copy == VM_OBJECT_NULL));
- shadow_object->copy = VM_OBJECT_NULL;
+ if (shadow_object->copy == object)
+ shadow_object->copy = VM_OBJECT_NULL;
vm_object_unlock(shadow_object);
}
+ 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
+ * or other I/O operations in progress for this VM object.
+ * Such operations take some paging_in_progress references
+ * up front to ensure that the object doesn't go away, but
+ * they may also need to acquire a reference on the VM object,
+ * to map it in kernel space, for example. That means that
+ * they may end up releasing the last reference on the VM
+ * object, triggering its termination, while still holding
+ * paging_in_progress references. Waiting for these
+ * pending paging_in_progress references to go away here would
+ * deadlock.
+ *
+ * To avoid deadlocking, we'll let the vm_object_reaper_thread
+ * complete the VM object termination if it still holds
+ * paging_in_progress references at this point.
+ *
+ * No new paging_in_progress should appear now that the
+ * VM object is "terminating" and not "alive".
+ */
+ vm_object_reap_async(object);
+ vm_object_unlock(object);
+ /*
+ * Return KERN_FAILURE to let the caller know that we
+ * haven't completed the termination and it can't drop this
+ * object's reference on its shadow object yet.
+ * The reaper thread will take care of that once it has
+ * completed this object's termination.
+ */
+ return KERN_FAILURE;
+ }
/*
- * 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.
+ * complete the VM object termination
*/
- 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);
- }
+ vm_object_reap(object);
+ object = VM_OBJECT_NULL;
+
+ /*
+ * 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 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_object_lock_assert_exclusive(object);
+ assert(object->paging_in_progress == 0);
+ assert(object->activity_in_progress == 0);
+
+ vm_object_reap_count++;
+
+ /*
+ * 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.
+ */
+ if (object->internal &&
+ object->purgable != VM_PURGABLE_DENY) {
+ vm_purgeable_accounting(object,
+ object->purgable,
+ TRUE); /* disown */
+ }
pager = object->pager;
object->pager = MEMORY_OBJECT_NULL;
if (pager != MEMORY_OBJECT_NULL)
- memory_object_control_disable(object->pager_request);
- vm_object_cache_unlock();
+ memory_object_control_disable(object->pager_control);
object->ref_count--;
#if TASK_SWAPPER
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,
* if some faults on this object were aborted.
*/
if (object->pageout) {
- assert(shadow_object != VM_OBJECT_NULL);
- assert(shadow_object == object->shadow);
+ assert(object->shadow != VM_OBJECT_NULL);
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_terminate: queue just emptied isn't");
+ vm_object_reap_pages(object, REAP_REAP);
}
-
+ assert(queue_empty(&object->memq));
assert(object->paging_in_progress == 0);
+ assert(object->activity_in_progress == 0);
assert(object->ref_count == 0);
/*
*/
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);
}
vm_object_unlock(object);
#if MACH_PAGEMAP
- vm_external_destroy(object->existence_map, object->size);
+ vm_external_destroy(object->existence_map, object->vo_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.
*/
- zfree(vm_object_zone, (vm_offset_t) object);
- return KERN_SUCCESS;
+ zfree(vm_object_zone, object);
+ 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 = (vm_page_t) m->pageq.next) { \
+ if (m->pmapped) { \
+ pmap_disconnect(m->phys_page); \
+ } \
+ } \
+ } \
+ 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 (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)queue_first(&object->memq);
+
+ while (!queue_end(&object->memq, (queue_entry_t)next)) {
+
+ p = next;
+ next = (vm_page_t)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) {
+ p->pageout = FALSE;
+
+ 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) {
+ p->pageout = FALSE;
+
+ 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.
+ */
+ vm_page_deactivate(p);
+ vm_page_purged_busy++;
+ continue;
+ }
+
+ assert(p->object != kernel_object);
+
+ /*
+ * we can discard this page...
+ */
+ if (p->pmapped == TRUE) {
+ /*
+ * unmap the page
+ */
+ pmap_disconnect_options(p->phys_page, 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 (p->phys_page == vm_page_guard_addr);
+ break;
+ }
+ if (!p->dirty && p->wpmapped)
+ p->dirty = pmap_is_modified(p->phys_page);
+
+ if ((p->dirty || p->precious) && !p->error && object->alive) {
+
+ assert(!object->internal);
+
+ if (!p->laundry) {
+ vm_page_queues_remove(p);
+ /*
+ * flush page... page will be freed
+ * upon completion of I/O
+ */
+ (void)vm_pageout_cluster(p, TRUE, 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 == NULL && p->pageq.prev == NULL);
+ /*
+ * Add this page to our list of reclaimed pages,
+ * to be freed later.
+ */
+ p->pageq.next = (queue_entry_t) 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)
+{
+ 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_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;
+
+ vm_object_reap(object);
+ /* cache is unlocked and object is no longer valid */
+ object = VM_OBJECT_NULL;
+
+ if (shadow_object != VM_OBJECT_NULL) {
+ /*
+ * Drop the reference "object" was holding on
+ * its shadow object.
+ */
+ vm_object_deallocate(shadow_object);
+ shadow_object = VM_OBJECT_NULL;
+ }
+ vm_object_reaper_lock_spin();
+ }
+
+ /* wait for more work... */
+ assert_wait((event_t) &vm_object_reaper_queue, THREAD_UNINT);
+
+ vm_object_reaper_unlock();
+
+ thread_block((thread_continue_t) vm_object_reaper_thread);
+ /*NOTREACHED*/
}
/*
{
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);
*/
static void
vm_object_release_pager(
- memory_object_t pager)
+ memory_object_t pager,
+ boolean_t hashed)
{
/*
(void) memory_object_terminate(pager);
- /*
- * Wakeup anyone waiting for this terminate
- */
- vm_object_pager_wakeup(pager);
-
- /*
- * Release reference to pager.
- */
- memory_object_deallocate(pager);
-}
-
-/*
- * 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 (hashed == TRUE) {
/*
- * If it's being paged in, destroy it.
- * If an unlock has been requested, start it again.
+ * Wakeup anyone waiting for this terminate
+ * and remove the entry from the hash
*/
-
- 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;
+ vm_object_pager_wakeup(pager);
}
-
/*
- * Wake up threads waiting for the memory object to
- * become ready.
+ * Release reference to pager.
*/
-
- object->pager_ready = TRUE;
- vm_object_wakeup(object, VM_OBJECT_EVENT_PAGER_READY);
+ memory_object_deallocate(pager);
}
/*
kern_return_t
vm_object_destroy(
vm_object_t object,
- kern_return_t reason)
+ __unused kern_return_t reason)
{
memory_object_t old_pager;
* 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_request);
- vm_object_cache_unlock();
+ memory_object_control_disable(object->pager_control);
/*
* Wait for the existing paging activity (that got
* 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
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.)
register vm_object_t object)
{
register 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();
queue_iterate(&object->memq, p, vm_page_t, listq) {
- vm_page_lock_queues();
- if (!p->busy)
+ 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->throttled) {
+#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(
+
+
+/*
+ * 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.
+ */
+
+#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. */
+
+#define CHUNK_SIZE (PAGES_IN_A_CHUNK * PAGE_SIZE_64) /* Size of a chunk in bytes */
+
+typedef uint64_t chunk_state_t;
+
+/*
+ * 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.
+ */
+
+#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
+
+
+/*
+ * Return true if all pages in the chunk have not yet been processed.
+ */
+
+#define CHUNK_NOT_COMPLETE(c) ((c) != 0)
+
+/*
+ * 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.
+ */
+
+#define PAGE_ALREADY_HANDLED(c, p) (((c) & (1LL << (p))) == 0)
+
+/*
+ * Mark the page at offset 'p' in the bit map as having been processed.
+ */
+
+#define MARK_PAGE_HANDLED(c, p) \
+MACRO_BEGIN \
+ (c) = (c) & ~(1LL << (p)); \
+MACRO_END
+
+
+/*
+ * Return true if the page at the given offset has been paged out. Object is
+ * locked upon entry and returned locked.
+ */
+
+static boolean_t
+page_is_paged_out(
+ vm_object_t object,
+ vm_object_offset_t offset)
+{
+ kern_return_t kr;
+ memory_object_t pager;
+
+ /*
+ * Check the existence map for the page if we have one, otherwise
+ * ask the pager about this page.
+ */
+
+#if MACH_PAGEMAP
+ if (object->existence_map) {
+ if (vm_external_state_get(object->existence_map, offset)
+ == VM_EXTERNAL_STATE_EXISTS) {
+ /*
+ * We found the page
+ */
+
+ return TRUE;
+ }
+ } else
+#endif /* MACH_PAGEMAP */
+ if (object->internal &&
+ object->alive &&
+ !object->terminating &&
+ object->pager_ready) {
+
+ if (COMPRESSED_PAGER_IS_ACTIVE || DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE) {
+ if (VM_COMPRESSOR_PAGER_STATE_GET(object, offset)
+ == VM_EXTERNAL_STATE_EXISTS) {
+ return TRUE;
+ } else {
+ return FALSE;
+ }
+ }
+
+ /*
+ * We're already holding a "paging in progress" reference
+ * so the object can't disappear when we release the lock.
+ */
+
+ assert(object->paging_in_progress);
+ pager = object->pager;
+ vm_object_unlock(object);
+
+ kr = memory_object_data_request(
+ pager,
+ offset + object->paging_offset,
+ 0, /* just poke the pager */
+ VM_PROT_READ,
+ NULL);
+
+ vm_object_lock(object);
+
+ if (kr == KERN_SUCCESS) {
+
+ /*
+ * We found the page
+ */
+
+ return TRUE;
+ }
+ }
+
+ return FALSE;
+}
+
+
+
+/*
+ * madvise_free_debug
+ *
+ * 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 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_object_t orig_object;
- int pages_moved = 0;
- int pages_found = 0;
+ 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;
/*
- * entered with object lock held, acquire a paging reference to
- * prevent the memory_object and control ports from
- * being destroyed.
+ * 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.
*/
- orig_object = object;
-
- for (;;) {
- register vm_page_t m;
- vm_object_offset_t toffset;
- vm_object_size_t tsize;
- vm_object_paging_begin(object);
- vm_page_lock_queues();
+ dwp = &dw_array[0];
+ dw_count = 0;
+ dw_limit = DELAYED_WORK_LIMIT(DEFAULT_DELAYED_WORK_LIMIT);
- for (tsize = size, toffset = offset; tsize; tsize -= PAGE_SIZE, toffset += PAGE_SIZE) {
+ for(p = 0; size && CHUNK_NOT_COMPLETE(*chunk_state); p++, size -= PAGE_SIZE_64, offset += PAGE_SIZE_64, pmap_offset += PAGE_SIZE_64) {
- if ((m = vm_page_lookup(object, toffset)) != VM_PAGE_NULL) {
+ /*
+ * 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.
+ */
- pages_found++;
+ 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->wire_count == 0) && (!m->private) && (!m->gobbled) && (!m->busy)) {
+ if ((m = vm_page_lookup(object, offset)) != VM_PAGE_NULL) {
- m->reference = FALSE;
- pmap_clear_reference(m->phys_addr);
+ /*
+ * 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.
+ */
- if ((kill_page) && (object->internal)) {
- m->precious = FALSE;
- m->dirty = FALSE;
- pmap_clear_modify(m->phys_addr);
- vm_external_state_clr(object->existence_map, offset);
+ 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(m->phys_page);
+ }
+ m->precious = FALSE;
+ m->dirty = FALSE;
+
+ clear_refmod |= VM_MEM_MODIFIED;
+ if (m->throttled) {
+ /*
+ * 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;
+ }
+#if MACH_PAGEMAP
+ vm_external_state_clr(object->existence_map, offset);
+#endif /* MACH_PAGEMAP */
+ 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;
}
- VM_PAGE_QUEUES_REMOVE(m);
-
- 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);
+ }
+ pmap_options |= PMAP_OPTIONS_NOFLUSH;
+ pmap_clear_refmod_options(m->phys_page,
+ clear_refmod,
+ pmap_options,
+ (void *)pfc);
+
+ if (!m->throttled && !(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);
- m->inactive = TRUE;
- if (!m->fictitious)
- vm_page_inactive_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);
- pages_moved++;
+ /*
+ * 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)) {
+#if MACH_PAGEMAP
+ vm_external_state_clr(object->existence_map, offset);
+#endif /* MACH_PAGEMAP */
+ VM_COMPRESSOR_PAGER_STATE_CLR(object,
+ offset);
+ if (pmap != PMAP_NULL &&
+ (COMPRESSED_PAGER_IS_ACTIVE ||
+ DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE)) {
+ /*
+ * 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);
+ }
}
}
}
- vm_page_unlock_queues();
- vm_object_paging_end(object);
+ }
+
+ 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);
- if (object->shadow) {
- vm_object_t tmp_object;
+ vm_object_paging_end(object);
- kill_page = 0;
+ /*
+ * 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.
+ */
- offset += object->shadow_offset;
+ tmp_object = object->shadow;
- 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;
- } else
- break;
+ if (object != orig_object)
+ vm_object_unlock(object);
+
+ object = tmp_object;
}
- if (object != orig_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((m)->phys_page, \
+ 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++;
+ 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++;
+ 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;
}
/*
vm_object_pmap_protect(
register vm_object_t object,
register vm_object_offset_t offset,
- vm_size_t size,
+ vm_object_size_t size,
pmap_t pmap,
- vm_offset_t pmap_start,
+ 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(
+ register vm_object_t object,
+ register 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;
- size = round_page_64(size);
- offset = trunc_page_64(offset);
+ return;
+ size = vm_object_round_page(size);
+ offset = vm_object_trunc_page(offset);
vm_object_lock(object);
- assert(object->copy_strategy == MEMORY_OBJECT_COPY_SYMMETRIC);
+ 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 (object->resident_page_count > atop(size) / 2 &&
- pmap != PMAP_NULL) {
+ 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 */
- if ((object->resident_page_count / 4) < atop(size)) {
+ 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_offset_t start = pmap_start +
- (vm_offset_t)(p->offset - offset);
-
- pmap_protect(pmap, start, start + PAGE_SIZE, 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_addr,
- prot & ~p->page_lock);
- }
- }
+ 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(
+ p->phys_page,
+ prot,
+ options | PMAP_OPTIONS_NOFLUSH,
+ &pmap_flush_context_storage);
+ delayed_pmap_flush = TRUE;
+ }
}
+
} else {
vm_page_t p;
vm_object_offset_t end;
end = offset + size;
- if (pmap != PMAP_NULL) {
- for(target_off = offset;
- target_off < end; target_off += PAGE_SIZE) {
- if(p = vm_page_lookup(object, target_off)) {
- vm_offset_t 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) {
- if(p = vm_page_lookup(object, target_off)) {
- pmap_page_protect(p->phys_addr,
- 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(
+ p->phys_page,
+ 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) {
/*
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;
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);
* 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);
/*
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,
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_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;
kern_return_t error_code;
vm_object_lock(src_object);
- vm_object_paging_begin(src_object);
- XPR(XPR_VM_FAULT,"vm_object_copy_slowly -> vm_fault_page",0,0,0,0,0);
- result = vm_fault_page(src_object, src_offset,
- VM_PROT_READ, FALSE, interruptible,
- src_lo_offset, src_hi_offset,
- VM_BEHAVIOR_SEQUENTIAL,
- &prot, &_result_page, &top_page,
- (int *)0,
- &error_code, FALSE, FALSE, NULL, 0);
+ 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;
+ }
- switch(result) {
- case VM_FAULT_SUCCESS:
- result_page = _result_page;
+ vm_object_paging_begin(src_object);
- /*
- * 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.
- */
+ 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);
+ }
- vm_object_unlock(result_page->object);
- vm_page_copy(result_page, new_page);
+ 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,
+ FALSE, /* page not looked up */
+ &prot, &_result_page, &top_page,
+ (int *)0,
+ &error_code, FALSE, FALSE, &fault_info);
- /*
- * Let go of both pages (make them
- * not busy, perform wakeup, activate).
- */
+ switch(result) {
+ case VM_FAULT_SUCCESS:
+ result_page = _result_page;
- new_page->busy = FALSE;
- new_page->dirty = TRUE;
- vm_object_lock(result_page->object);
- PAGE_WAKEUP_DONE(result_page);
+ /*
+ * Copy the page to the new object.
+ *
+ * POLICY DECISION:
+ * If result_page is clean,
+ * we could steal it instead
+ * of copying.
+ */
- 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();
+ vm_page_copy(result_page, new_page);
+ vm_object_unlock(result_page->object);
- /*
- * Release paging references and
- * top-level placeholder page, if any.
- */
+ /*
+ * 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->active &&
+ !result_page->inactive &&
+ !result_page->throttled)
+ 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_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_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);
}
/*
* Lose the extra reference, and return our object.
*/
-
vm_object_deallocate(src_object);
*_result_object = new_object;
return(KERN_SUCCESS);
__private_extern__ boolean_t
vm_object_copy_quickly(
vm_object_t *_object, /* INOUT */
- vm_object_offset_t offset, /* IN */
- vm_object_size_t size, /* IN */
+ __unused vm_object_offset_t offset, /* IN */
+ __unused vm_object_size_t size, /* IN */
boolean_t *_src_needs_copy, /* OUT */
boolean_t *_dst_needs_copy) /* OUT */
{
* 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);
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.
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;
static int copy_delayed_max_collisions = 0;
static int copy_delayed_lock_contention = 0;
static int copy_delayed_protect_iterate = 0;
-static int copy_delayed_protect_lookup = 0;
-static int copy_delayed_protect_lookup_wait = 0;
/*
* Routine: vm_object_copy_delayed [internal]
* the asymmetric copy-on-write algorithm.
*
* In/out conditions:
- * The object must be unlocked on entry.
+ * The src_object must be locked on entry. It will be unlocked
+ * on exit - so the caller must also hold a reference to it.
*
* This routine will not block waiting for user-generated
* events. It is not interruptible.
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;
+ 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;
/*
* the original object must be done carefully, to avoid deadlock.
*/
+ copy_size = vm_object_round_page(copy_size);
Retry:
- vm_object_lock(src_object);
+ /*
+ * Wait for paging in progress.
+ */
+ 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.
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;
+ if (src_object_shared == TRUE)
+ vm_object_lock_shared(src_object);
+ else
+ vm_object_lock(src_object);
+
goto Retry;
}
* It has not been modified.
*
* Return another reference to
- * the existing copy-object.
+ * the existing copy-object if
+ * we can safely grow it (if
+ * needed).
*/
- assert(old_copy->ref_count > 0);
- old_copy->ref_count++;
- if (old_copy->size < src_offset+size)
- old_copy->size = src_offset+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
+ * we can't safely take write permission away from
+ * wired pages). If the pages aren't wired, then
+ * go ahead and protect them.
+ */
+ copy_delayed_protect_iterate++;
+
+ pmap_flush_context_init(&pmap_flush_context_storage);
+ delayed_pmap_flush = FALSE;
+
+ queue_iterate(&src_object->memq, p, vm_page_t, listq) {
+ if (!p->fictitious &&
+ p->offset >= old_copy->vo_size &&
+ p->offset < copy_size) {
+ if (VM_PAGE_WIRED(p)) {
+ vm_object_unlock(old_copy);
+ vm_object_unlock(src_object);
+
+ if (new_copy != VM_OBJECT_NULL) {
+ 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_options(p->phys_page, (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 TASK_SWAPPER
- /*
- * We have to reproduce some of the code from
- * vm_object_res_reference because we've taken
- * the locks out of order here, and deadlock
- * would result if we simply called that function.
- */
- if (++old_copy->res_count == 1) {
- assert(old_copy->shadow == src_object);
- vm_object_res_reference(src_object);
+ old_copy->vo_size = copy_size;
}
-#endif /* TASK_SWAPPER */
-
+ 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);
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->vo_size > copy_size)
+ copy_size = old_copy->vo_size;
+
if (new_copy == VM_OBJECT_NULL) {
vm_object_unlock(old_copy);
vm_object_unlock(src_object);
- new_copy = vm_object_allocate(src_offset + size);
+ new_copy = vm_object_allocate(copy_size);
+ vm_object_lock(src_object);
vm_object_lock(new_copy);
+
+ src_object_shared = FALSE;
goto Retry;
}
-
- /*
- * Adjust the size argument so that the newly-created
- * copy object will be large enough to back either the
- * new old copy object or the new mapping.
- */
- if (old_copy->size > src_offset+size)
- size = old_copy->size - src_offset;
+ new_copy->vo_size = copy_size;
/*
* The copy-object is always made large enough to
*/
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;
+ }
+
+ /*
+ * We now have the src object locked, and the new copy object
+ * allocated and locked (and potentially the old copy locked).
+ * Before we go any further, make sure we can still perform
+ * a delayed copy, as the situation may have changed.
+ *
+ * Specifically, we can't perform a delayed copy if any of the
+ * pages in the range are wired (because we can't safely take
+ * write permission away from wired pages). If the pages aren't
+ * wired, then go ahead and protect them.
+ */
+ copy_delayed_protect_iterate++;
+
+ pmap_flush_context_init(&pmap_flush_context_storage);
+ delayed_pmap_flush = FALSE;
+
+ queue_iterate(&src_object->memq, p, vm_page_t, listq) {
+ if (!p->fictitious && p->offset < copy_size) {
+ 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_options(p->phys_page, (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) {
/*
* Make the old copy-object shadow the new one.
* It will receive no more pages from the original
* 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. */
#endif
vm_object_unlock(old_copy); /* done with old_copy */
- } else if (new_copy == VM_OBJECT_NULL) {
- vm_object_unlock(src_object);
- new_copy = vm_object_allocate(src_offset + size);
- vm_object_lock(new_copy);
- goto Retry;
- }
-
- /*
- * Readjust the copy-object size if necessary.
- */
- copy_size = new_copy->size;
- if (copy_size < src_offset+size) {
- copy_size = src_offset+size;
- new_copy->size = copy_size;
}
/*
* 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);
- /*
- * Mark all (current) pages of the existing object copy-on-write.
- * This object may have a shadow chain below it, but
- * those pages will already be marked copy-on-write.
- */
-
- vm_object_paging_wait(src_object, THREAD_UNINT);
- copy_delayed_protect_iterate++;
- queue_iterate(&src_object->memq, p, vm_page_t, listq) {
- if (!p->fictitious)
- pmap_page_protect(p->phys_addr,
- (VM_PROT_ALL & ~VM_PROT_WRITE &
- ~p->page_lock));
- }
- vm_object_unlock(src_object);
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;
}
/*
{
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
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);
}
}
- copy_strategy = src_object->copy_strategy;
-
/*
* Use the appropriate copy strategy.
*/
switch (copy_strategy) {
+ case MEMORY_OBJECT_COPY_DELAY:
+ *dst_object = vm_object_copy_delayed(src_object,
+ src_offset, size, object_lock_shared);
+ if (*dst_object != VM_OBJECT_NULL) {
+ *dst_offset = src_offset;
+ *dst_needs_copy = TRUE;
+ result = KERN_SUCCESS;
+ break;
+ }
+ vm_object_lock(src_object);
+ /* fall thru when delayed copy not allowed */
+
case MEMORY_OBJECT_COPY_NONE:
result = vm_object_copy_slowly(src_object, src_offset, size,
interruptible, dst_object);
}
break;
- case MEMORY_OBJECT_COPY_DELAY:
- vm_object_unlock(src_object);
- *dst_object = vm_object_copy_delayed(src_object,
- src_offset, size);
- *dst_offset = src_offset;
- *dst_needs_copy = TRUE;
- result = KERN_SUCCESS;
- 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;
* 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(
register 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 &&
+ 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;
* and fix up the offset into the new object.
*/
- result->shadow_offset = *offset;
+ result->vo_shadow_offset = *offset;
/*
* Return the new things
* [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
* termination routines and vm_object_collapse.]
*/
-#if 0
-/*
- * 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
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));
/*
* Look for an object associated with this port.
*/
-
-restart:
- vm_object_cache_lock();
- for (;;) {
+Retry:
+ lck = vm_object_hash_lock_spin(pager);
+ do {
entry = vm_object_hash_lookup(pager, FALSE);
- /*
- * If a previous object is being terminated,
- * we must wait for the termination message
- * to be queued.
- *
- * We set kobject to a non-null value to let the
- * terminator know that someone is waiting.
- * Among the possibilities is that the port
- * could die while we're waiting. Must restart
- * instead of continuing the loop.
- */
-
- if (entry != VM_OBJECT_HASH_ENTRY_NULL) {
- if (entry->object != VM_OBJECT_NULL)
- break;
-
+ if (entry == VM_OBJECT_HASH_ENTRY_NULL) {
+ if (new_object == VM_OBJECT_NULL) {
+ /*
+ * We must unlock to create a new object;
+ * if we do so, we must try the lookup again.
+ */
+ 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);
+ /*
+ * 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.
+ */
+ /*
+ * 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;
+ new_entry = VM_OBJECT_HASH_ENTRY_NULL;
+ new_object = VM_OBJECT_NULL;
+ must_init = TRUE;
+ }
+ } else if (entry->object == VM_OBJECT_NULL) {
+ /*
+ * If a previous object is being terminated,
+ * we must wait for the termination message
+ * to be queued (and lookup the entry again).
+ */
entry->waiting = TRUE;
+ entry = VM_OBJECT_HASH_ENTRY_NULL;
assert_wait((event_t) pager, THREAD_UNINT);
- vm_object_cache_unlock();
- thread_block((void (*)(void))0);
- goto restart;
- }
-
- /*
- * We must unlock to create a new object;
- * if we do so, we must try the lookup again.
- */
-
- if (new_object == VM_OBJECT_NULL) {
- vm_object_cache_unlock();
- 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();
- } else {
- /*
- * Lookup failed twice, and we have something
- * to insert; set the object.
- */
-
- if (entry == VM_OBJECT_HASH_ENTRY_NULL) {
- vm_object_hash_insert(new_entry);
- entry = new_entry;
- new_entry = VM_OBJECT_HASH_ENTRY_NULL;
- }
+ vm_object_hash_unlock(lck);
- entry->object = new_object;
- new_object = VM_OBJECT_NULL;
- must_init = TRUE;
+ thread_block(THREAD_CONTINUE_NULL);
+ lck = vm_object_hash_lock_spin(pager);
}
- }
+ } while (entry == VM_OBJECT_HASH_ENTRY_NULL);
object = entry->object;
assert(object != VM_OBJECT_NULL);
if (!must_init) {
- vm_object_lock(object);
- assert(object->pager_created);
- 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);
if (must_init) {
- pager_request_t pager_request;
+ memory_object_control_t control;
/*
* Allocate request port.
*/
- pager_request = memory_object_control_allocate(object);
- assert (pager_request != PAGER_REQUEST_NULL);
+ control = memory_object_control_allocate(object);
+ assert (control != MEMORY_OBJECT_CONTROL_NULL);
vm_object_lock(object);
+ assert(object != kernel_object);
/*
* Copy the reference we were given.
/* copy strategy invalid until set by memory manager */
object->copy_strategy = MEMORY_OBJECT_COPY_INVALID;
}
- object->pager_request = pager_request;
+ object->pager_control = control;
object->pager_ready = FALSE;
vm_object_unlock(object);
*/
(void) memory_object_init(pager,
- object->pager_request,
+ object->pager_control,
PAGE_SIZE);
vm_object_lock(object);
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);
}
{
memory_object_t pager;
vm_object_hash_entry_t entry;
+ lck_mtx_t *lck;
#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);
+ object, 0,0,0,0);
+
+ assert(object != kernel_object);
if (memory_manager_default_check() != KERN_SUCCESS)
return;
object->paging_offset = 0;
#if MACH_PAGEMAP
- size = object->size;
+ size = object->vo_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);
+ if (DEFAULT_PAGER_IS_ACTIVE) {
+ map = vm_external_create(size);
+ vm_object_lock(object);
+ assert(object->vo_size == size);
+ object->existence_map = map;
+ vm_object_unlock(object);
+ }
#endif /* MACH_PAGEMAP */
+ if ((uint32_t) object->vo_size != object->vo_size) {
+ panic("vm_object_pager_create(): object size 0x%llx >= 4GB\n",
+ (uint64_t) object->vo_size);
+ }
+
/*
* Create the [internal] pager, and associate it with 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);
+ dmm = memory_manager_default_reference();
- object->cluster_size = cluster_size; /* XXX ??? */
assert(object->temporary);
/* create our new memory object */
- (void) memory_object_create(dmm, object->size, &pager);
+ assert((vm_size_t) object->vo_size == object->vo_size);
+ (void) memory_object_create(dmm, (vm_size_t) object->vo_size,
+ &pager);
memory_object_default_deallocate(dmm);
}
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
* copied by vm_object_enter().
*/
- if (vm_object_enter(pager, object->size, TRUE, TRUE, FALSE) != object)
+ if (vm_object_enter(pager, object->vo_size, TRUE, TRUE, FALSE) != object)
panic("vm_object_pager_create: mismatch");
/*
vm_object_paging_end(object);
}
+void
+vm_object_compressor_pager_create(
+ register vm_object_t object)
+{
+ memory_object_t pager;
+ vm_object_hash_entry_t entry;
+ lck_mtx_t *lck;
+ vm_object_t pager_object = VM_OBJECT_NULL;
+
+ assert(object != kernel_object);
+
+ /*
+ * Prevent collapse or termination by holding a paging reference
+ */
+
+ vm_object_paging_begin(object);
+ if (object->pager_created) {
+ /*
+ * Someone else got to it first...
+ * wait for them to finish initializing the ports
+ */
+ while (!object->pager_initialized) {
+ vm_object_sleep(object,
+ VM_OBJECT_EVENT_INITIALIZED,
+ THREAD_UNINT);
+ }
+ vm_object_paging_end(object);
+ return;
+ }
+
+ /*
+ * Indicate that a memory object has been assigned
+ * before dropping the lock, to prevent a race.
+ */
+
+ object->pager_created = TRUE;
+ object->paging_offset = 0;
+
+ vm_object_unlock(object);
+
+ 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.
+ *
+ * We make the association here so that vm_object_enter()
+ * can look up the object to complete initializing it. No
+ * user will ever map this object.
+ */
+ {
+ assert(object->temporary);
+
+ /* create our new memory object */
+ 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_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
+ * memory_object_create(), and it is
+ * copied by vm_object_enter().
+ */
+
+ 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.
+ */
+ memory_object_deallocate(pager);
+
+ vm_object_lock(object);
+
+ /*
+ * Release the paging reference
+ */
+ vm_object_paging_end(object);
+}
+
/*
* Routine: vm_object_remove
* Purpose:
vm_object_t object)
{
memory_object_t pager;
- pager_request_t pager_request;
if ((pager = object->pager) != MEMORY_OBJECT_NULL) {
vm_object_hash_entry_t entry;
static boolean_t vm_object_collapse_allowed = TRUE;
static boolean_t vm_object_bypass_allowed = TRUE;
+#if MACH_PAGEMAP
static int vm_external_discarded;
static int vm_external_collapsed;
+#endif
+
+unsigned long vm_object_collapse_encrypted = 0;
+
+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_object_offset_t new_offset, backing_offset;
+ vm_object_size_t 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
vm_object_offset_t new_offset, backing_offset;
vm_object_size_t size;
- backing_offset = object->shadow_offset;
- size = object->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
new_offset = (p->offset - backing_offset);
assert(!p->busy || p->absent);
-
+
/*
* If the parent has a page here, or if
* this page falls outside the parent,
if (p->offset < backing_offset || new_offset >= size) {
VM_PAGE_FREE(p);
} else {
+ /*
+ * ENCRYPTED SWAP:
+ * The encryption key includes the "pager" and the
+ * "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.
+ */
+ 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.
- */
+ 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);
+ }
- vm_page_rename(p, object, new_offset);
#if MACH_PAGEMAP
} else if (pp->absent) {
*/
VM_PAGE_FREE(pp);
- vm_page_rename(p, object, new_offset);
+ vm_page_rename(p, object, new_offset, TRUE);
#endif /* MACH_PAGEMAP */
} else {
assert(! pp->absent);
}
}
}
-
- assert(object->pager == MEMORY_OBJECT_NULL ||
- backing_object->pager == MEMORY_OBJECT_NULL);
- 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;
+#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 */
+
/*
* Move the pager from backing_object to object.
*
* unused portion.
*/
+ 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_request = backing_object->pager_request;
+ 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_request != PAGER_REQUEST_NULL) {
- memory_object_control_collapse(object->pager_request,
+ 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();
-
- object->paging_offset = backing_object->paging_offset + backing_offset;
-
#if MACH_PAGEMAP
/*
* If the shadow offset is 0, the use the existence map from
* this code should be fixed to salvage the map.
*/
assert(object->existence_map == VM_EXTERNAL_NULL);
- if (backing_offset || (size != backing_object->size)) {
+ if (backing_offset || (size != backing_object->vo_size)) {
vm_external_discarded++;
vm_external_destroy(backing_object->existence_map,
- backing_object->size);
+ backing_object->vo_size);
}
else {
vm_external_collapsed++;
* moves from within backing_object to within object.
*/
+ assert(!object->phys_contiguous);
+ assert(!backing_object->phys_contiguous);
object->shadow = backing_object->shadow;
- object->shadow_offset += backing_object->shadow_offset;
+ if (object->shadow) {
+ 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->vo_shadow_offset = 0;
+ }
assert((object->shadow == VM_OBJECT_NULL) ||
- (object->shadow->copy == VM_OBJECT_NULL));
+ (object->shadow->copy != backing_object));
/*
* Discard backing_object.
* 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 */
- zfree(vm_object_zone, (vm_offset_t) backing_object);
+ vm_object_lock_destroy(backing_object);
+
+ zfree(vm_object_zone, backing_object);
- object_collapses++;
}
static void
* 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
*/
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);
vm_object_reference(backing_object->shadow);
#endif /* TASK_SWAPPER */
+ assert(!object->phys_contiguous);
+ assert(!backing_object->phys_contiguous);
object->shadow = backing_object->shadow;
- object->shadow_offset += backing_object->shadow_offset;
+ if (object->shadow) {
+ object->vo_shadow_offset += backing_object->vo_shadow_offset;
+ } else {
+ /* no shadow, therefore no shadow offset... */
+ object->vo_shadow_offset = 0;
+ }
/*
* Backing object might have had a copy pointer
* 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
* 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)
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);
*/
vm_object_lock(object);
+ vm_object_activity_end(object);
}
object_bypasses++;
* Requires that the object be locked and the page queues be unlocked.
*
*/
+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_t object,
+ register vm_object_offset_t hint_offset,
+ boolean_t can_bypass)
{
register vm_object_t backing_object;
- register vm_object_offset_t backing_offset;
- register vm_object_size_t size;
- register vm_object_offset_t new_offset;
- register vm_page_t p;
+ register unsigned int rcount;
+ register unsigned int size;
+ vm_object_t original_object;
+ int object_lock_type;
+ int backing_object_lock_type;
- vm_offset_t current_offset;
+ vm_object_collapse_calls++;
- if (! vm_object_collapse_allowed && ! vm_object_bypass_allowed) {
+ if (! vm_object_collapse_allowed &&
+ ! (can_bypass && vm_object_bypass_allowed)) {
return;
}
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++;
/*
* Verify that the conditions are right for either
* collapse or bypass:
- *
- * The object exists and no pages in it are currently
- * being paged out, and
*/
- if (object == VM_OBJECT_NULL ||
- object->paging_in_progress != 0 ||
- object->absent_count != 0)
- return;
/*
* There is a backing object, and
*/
- if ((backing_object = object->shadow) == VM_OBJECT_NULL)
+ backing_object = object->shadow;
+ if (backing_object == VM_OBJECT_NULL) {
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
return;
-
- vm_object_lock(backing_object);
+ }
+ 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->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;
+ }
/*
* ...
*/
if (!backing_object->internal ||
- backing_object->paging_in_progress != 0) {
- vm_object_unlock(backing_object);
- return;
+ 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;
}
/*
* parent object.
*/
if (backing_object->shadow != VM_OBJECT_NULL &&
- backing_object->shadow->copy != VM_OBJECT_NULL) {
- vm_object_unlock(backing_object);
- return;
+ backing_object->shadow->copy == backing_object) {
+ /* 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;
}
/*
* object, we may be able to collapse it into the
* parent.
*
- * The backing object must not have a pager
- * created for it, since collapsing an object
- * into a backing_object dumps new pages into
- * the backing_object that its pager doesn't
- * know about.
+ * 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 &&
- vm_object_collapse_allowed) {
-
- XPR(XPR_VM_OBJECT,
- "vm_object_collapse: %x to %x, pager %x, pager_request %x\n",
- (integer_t)backing_object, (integer_t)object,
- (integer_t)backing_object->pager,
- (integer_t)backing_object->pager_request, 0);
+ (vm_object_collapse_compressor_allowed ||
+ !object->pager_created
+#if !MACH_PAGEMAP
+ || (!backing_object->pager_created)
+#endif /*!MACH_PAGEMAP */
+ ) && vm_object_collapse_allowed) {
/*
- * 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 (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;
}
+ 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
*/
vm_object_do_collapse(object, backing_object);
+ vm_object_collapse_do_collapse++;
continue;
}
-
/*
* Collapsing the backing object was not possible
* or permitted, so let's try bypassing it.
*/
- if (! vm_object_bypass_allowed) {
- vm_object_unlock(backing_object);
- return;
+ if (! (can_bypass && vm_object_bypass_allowed)) {
+ /* 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;
}
/*
- * 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 the object doesn't have all its pages present,
+ * we have to make sure no pages in the backing object
+ * "show through" before bypassing it.
*/
- if (backing_object->pager_created
+ 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;
+
+ /*
+ * 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)
+ && (backing_object->existence_map == VM_EXTERNAL_NULL)
#endif /* MACH_PAGEMAP */
- ) {
- vm_object_unlock(backing_object);
- return;
- }
+ ) {
+ /* 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;
+ }
- /*
- * If the object has a pager but no pagemap,
- * then we cannot bypass it, because we don't know
- * what pages it has.
- */
- if (object->pager_created
+ /*
+ * If the object has a pager but no pagemap,
+ * 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)
+ && (object->existence_map == VM_EXTERNAL_NULL)
#endif /* MACH_PAGEMAP */
- ) {
- vm_object_unlock(backing_object);
- return;
- }
+ ) {
+ /* 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->shadow_offset;
- size = object->size;
+ backing_offset = object->vo_shadow_offset;
+ backing_rcount = backing_object->resident_page_count;
- /*
- * If all of the pages in the backing object are
- * shadowed by the parent object, the parent
- * object no longer has to shadow the backing
- * object; it can shadow the next one in the
- * chain.
- *
- * If the backing object has existence info,
- * we must check examine its existence info
- * as well.
- *
- */
+ 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;
+ }
- if(object->cow_hint >= size)
- object->cow_hint = 0;
- current_offset = object->cow_hint;
- while(TRUE) {
- if (vm_page_lookup(object,
- (vm_object_offset_t)current_offset)
- != VM_PAGE_NULL) {
- current_offset+=PAGE_SIZE;
- } else if ((object->pager_created) &&
- (object->existence_map != NULL) &&
- (vm_external_state_get(object->existence_map,
- current_offset)
- != VM_EXTERNAL_STATE_ABSENT)) {
- current_offset+=PAGE_SIZE;
- } else if (vm_page_lookup(backing_object,
- (vm_object_offset_t)current_offset
- + backing_offset)!= VM_PAGE_NULL) {
- /* found a dependency */
- object->cow_hint = current_offset;
- vm_object_unlock(backing_object);
- return;
- } else if ((backing_object->pager_created) &&
- (backing_object->existence_map != NULL) &&
- (vm_external_state_get(
- backing_object->existence_map,
- current_offset + backing_offset)
- != VM_EXTERNAL_STATE_ABSENT)) {
- /* found a dependency */
- object->cow_hint = current_offset;
- vm_object_unlock(backing_object);
- return;
- } else {
- current_offset+=PAGE_SIZE;
+ /*
+ * If all of the pages in the backing object are
+ * shadowed by the parent object, the parent
+ * object no longer has to shadow the backing
+ * object; it can shadow the next one in the
+ * chain.
+ *
+ * If the backing object has existence info,
+ * we must check examine its existence info
+ * as well.
+ *
+ */
+
+#if MACH_PAGEMAP
+#define EXISTS_IN_OBJECT(obj, off, rc) \
+ ((vm_external_state_get((obj)->existence_map, \
+ (vm_offset_t)(off)) \
+ == VM_EXTERNAL_STATE_EXISTS) || \
+ (VM_COMPRESSOR_PAGER_STATE_GET((obj), (off)) \
+ == VM_EXTERNAL_STATE_EXISTS) || \
+ ((rc) && vm_page_lookup((obj), (off)) != VM_PAGE_NULL && (rc)--))
+#else /* MACH_PAGEMAP */
+#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)--))
+#endif /* MACH_PAGEMAP */
+
+ /*
+ * Check the hint location first
+ * (since it is often the quickest way out of here).
+ */
+ if (object->cow_hint != ~(vm_offset_t)0)
+ hint_offset = (vm_object_offset_t)object->cow_hint;
+ else
+ hint_offset = (hint_offset > 8 * PAGE_SIZE_64) ?
+ (hint_offset - 8 * PAGE_SIZE_64) : 0;
+
+ if (EXISTS_IN_OBJECT(backing_object, hint_offset +
+ backing_offset, backing_rcount) &&
+ !EXISTS_IN_OBJECT(object, hint_offset, rcount)) {
+ /* dependency right at the hint */
+ 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;
}
- if(current_offset >= size) {
- /* wrap at end of object */
- current_offset = 0;
+
+ /*
+ * If the object's window onto the backing_object
+ * is large compared to the number of resident
+ * 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/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 && 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);
+ do {
+ offset = (p->offset - backing_offset);
+
+ if (offset < object->vo_size &&
+ offset != hint_offset &&
+ !EXISTS_IN_OBJECT(object, offset, rc)) {
+ /* found a dependency */
+ object->cow_hint = (vm_offset_t) offset; /* atomic */
+
+ break;
+ }
+ p = (vm_page_t) queue_next(&p->listq);
+
+ } while (--backing_rcount);
+ if (backing_rcount != 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;
+ }
}
- if(current_offset == object->cow_hint) {
- /* we are free of shadow influence */
- break;
+
+ /*
+ * Walk through the offsets looking for pages in the
+ * backing object that show through to the object.
+ */
+ if (backing_rcount
+#if MACH_PAGEMAP
+ || backing_object->existence_map
+#endif /* MACH_PAGEMAP */
+ ) {
+ offset = hint_offset;
+
+ while((offset =
+ (offset + PAGE_SIZE_64 < object->vo_size) ?
+ (offset + PAGE_SIZE_64) : 0) != hint_offset) {
+
+ 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; /* atomic */
+ break;
+ }
+ }
+ if (offset != hint_offset) {
+ /* 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;
+ }
}
}
- /* reset the cow_hint for any objects deeper in the chain */
- object->cow_hint = 0;
+ /*
+ * 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;
/*
* All interesting pages in the backing object
*/
vm_object_do_bypass(object, backing_object);
+ vm_object_collapse_do_bypass++;
/*
* Try again with this object's new backing object.
continue;
}
+
+ /* NOT REACHED */
+ /*
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ */
}
/*
* It balances vm_object_lookup vs iteration.
*/
- if (atop(end - start) < (unsigned)object->resident_page_count/16) {
+ if (atop_64(end - start) < (unsigned)object->resident_page_count/16) {
vm_object_page_remove_lookup++;
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_page_protect(p->phys_addr,
- VM_PROT_NONE);
+ assert(!p->cleaning && !p->pageout && !p->laundry);
+ if (!p->fictitious && p->pmapped)
+ pmap_disconnect(p->phys_page);
VM_PAGE_FREE(p);
}
}
while (!queue_end(&object->memq, (queue_entry_t) p)) {
next = (vm_page_t) queue_next(&p->listq);
if ((start <= p->offset) && (p->offset < end)) {
- assert(!p->cleaning && !p->pageout);
- if (!p->fictitious)
- pmap_page_protect(p->phys_addr,
- VM_PROT_NONE);
+ assert(!p->cleaning && !p->pageout && !p->laundry);
+ if (!p->fictitious && p->pmapped)
+ pmap_disconnect(p->phys_page);
VM_PAGE_FREE(p);
}
p = next;
register vm_object_t prev_object,
vm_object_t next_object,
vm_object_offset_t prev_offset,
- vm_object_offset_t next_offset,
+ __unused vm_object_offset_t next_offset,
vm_object_size_t prev_size,
vm_object_size_t next_size)
{
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);
/*
* Try to collapse the object first
*/
- vm_object_collapse(prev_object);
+ vm_object_collapse(prev_object, prev_offset, TRUE);
/*
* Can't coalesce if pages not mapped to
* . paged out
* . shadows another object
* . has a copy elsewhere
+ * . is purgeable
* . paging references (pages might be in page-list)
*/
(prev_object->shadow != VM_OBJECT_NULL) ||
(prev_object->copy != VM_OBJECT_NULL) ||
(prev_object->true_share != FALSE) ||
- (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);
}
* Extend the object if necessary.
*/
newsize = prev_offset + prev_size + next_size;
- if (newsize > prev_object->size) {
+ if (newsize > prev_object->vo_size) {
#if MACH_PAGEMAP
/*
* We cannot extend an object that has existence info,
*/
assert(prev_object->existence_map == VM_EXTERNAL_NULL);
#endif /* MACH_PAGEMAP */
- prev_object->size = newsize;
+ 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 */
-{
- int num_pages;
- int i;
- vm_page_t m;
- vm_page_t old_page;
- vm_object_offset_t addr;
-
- num_pages = atop(size);
-
- for (i = 0; i < num_pages; i++, offset += PAGE_SIZE_64) {
-
- addr = (*map_fn)(map_fn_data, offset);
-
- while ((m = vm_page_grab_fictitious()) == VM_PAGE_NULL)
- vm_page_more_fictitious();
-
- 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();
- }
-
- 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);
-
- PAGE_WAKEUP_DONE(m);
- vm_object_unlock(object);
- }
-}
-
-#include <mach_kdb.h>
-
-#if MACH_KDB
-#include <ddb/db_output.h>
-#include <vm/vm_print.h>
-
-#define printf kdbprintf
-
-extern boolean_t vm_object_cached(
- vm_object_t object);
-
-extern void print_bitstring(
- char byte);
-
-boolean_t vm_object_print_pages = FALSE;
-
-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'));
-}
-
-boolean_t
-vm_object_cached(
- register vm_object_t object)
-{
- register vm_object_t o;
-
- queue_iterate(&vm_object_cached_list, o, vm_object_t, cached_list) {
- if (object == o) {
- return TRUE;
- }
- }
- return FALSE;
-}
-
-#if MACH_PAGEMAP
-/*
- * vm_external_print: [ debug ]
- */
-void
-vm_external_print(
- vm_external_map_t map,
- vm_size_t size)
-{
- if (map == 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(map[0]);
- }
- if (existence_size > 1) {
- print_bitstring(map[1]);
- }
- if (existence_size > 2) {
- printf("...");
- print_bitstring(map[existence_size-1]);
- }
- printf("] }\n");
- }
- return;
-}
-#endif /* MACH_PAGEMAP */
-
-int
-vm_follow_object(
- vm_object_t object)
-{
- extern db_indent;
-
- 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);
-
- db_indent += 2;
- object = object->shadow;
- }
-
-}
-
-/*
- * vm_object_print: [ debug ]
- */
-void
-vm_object_print(
- vm_object_t object,
- boolean_t have_addr,
- int arg_count,
- char *modif)
-{
- register vm_page_t p;
- extern db_indent;
- char *s;
-
- register int count;
-
- 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(", frozen=0x%x", object->frozen_size);
- 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_request=0x%x\n", object->pager_request);
-
- 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, %slock_in_progress, %slock_restart, %sshadowed, %scached, %sprivate\n",
- (object->alive ? "" : "!"),
- (object->lock_in_progress ? "" : "!"),
- (object->lock_restart ? "" : "!"),
- (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%X,page=0x%X)", p->offset, (integer_t) 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_t object,
vm_object_offset_t offset,
- vm_offset_t phys_addr,
- vm_size_t size)
+ ppnum_t phys_page,
+ vm_size_t size)
{
- vm_offset_t base_addr;
+ ppnum_t base_page;
vm_object_offset_t base_offset;
- if(!object->private)
- return KERN_FAILURE;
-
- if((base_addr = trunc_page(phys_addr)) != phys_addr) {
+ if (!object->private)
return KERN_FAILURE;
- }
+ base_page = phys_page;
vm_object_lock(object);
- if(!object->phys_contiguous) {
+
+ if (!object->phys_contiguous) {
vm_page_t m;
- if((base_offset = trunc_page(offset)) != offset) {
+
+ if ((base_offset = trunc_page_64(offset)) != offset) {
vm_object_unlock(object);
return KERN_FAILURE;
}
base_offset += object->paging_offset;
- while(size) {
+
+ 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_addr = base_addr;
- if(!m->busy) {
- m->busy = TRUE;
+
+ if (m != VM_PAGE_NULL) {
+ if (m->fictitious) {
+ if (m->phys_page != vm_page_guard_addr) {
+
+ vm_page_lockspin_queues();
+ m->private = TRUE;
+ vm_page_unlock_queues();
+
+ m->fictitious = FALSE;
+ m->phys_page = base_page;
}
- if(!m->absent) {
- m->absent = TRUE;
- object->absent_count++;
+ } else if (m->phys_page != base_page) {
+
+ if ( !m->private) {
+ /*
+ * we'd leak a real page... that can't be right
+ */
+ panic("vm_object_populate_with_private - %p not private", m);
}
- m->list_req_pending = TRUE;
- vm_page_unlock_queues();
- } else if (m->phys_addr != base_addr) {
- /* pmap call to clear old mapping */
- pmap_page_protect(m->phys_addr,
- VM_PROT_NONE);
- m->phys_addr = base_addr;
+ if (m->pmapped) {
+ /*
+ * pmap call to clear old mapping
+ */
+ pmap_disconnect(m->phys_page);
+ }
+ m->phys_page = 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);
}
+
} else {
- while ((m = vm_page_grab_fictitious())
- == VM_PAGE_NULL)
+ while ((m = vm_page_grab_fictitious()) == VM_PAGE_NULL)
vm_page_more_fictitious();
- vm_page_lock_queues();
- m->fictitious = FALSE;
+
+ /*
+ * private normally requires lock_queues but since we
+ * are initializing the page, its not necessary here
+ */
m->private = TRUE;
- m->phys_addr = base_addr;
- m->list_req_pending = TRUE;
- m->absent = TRUE;
+ m->fictitious = FALSE;
+ m->phys_page = base_page;
m->unusual = TRUE;
- object->absent_count++;
- vm_page_unlock_queues();
+ m->busy = FALSE;
+
vm_page_insert(m, object, base_offset);
}
- base_addr += PAGE_SIZE;
+ base_page++; /* Go to the next physical page */
base_offset += PAGE_SIZE;
size -= PAGE_SIZE;
}
/* shadows on contiguous memory are not allowed */
/* we therefore can use the offset field */
- object->shadow_offset = (vm_object_offset_t)phys_addr;
- object->size = size;
+ object->vo_shadow_offset = (vm_object_offset_t)phys_page << PAGE_SHIFT;
+ object->vo_size = size;
}
vm_object_unlock(object);
+
return KERN_SUCCESS;
}
* memory_object_free_from_cache:
*
* Walk the vm_object cache list, removing and freeing vm_objects
- * which are backed by the pager identified by the caller, (pager_id).
+ * which are backed by the pager identified by the caller, (pager_ops).
* Remove up to "count" objects, if there are that may available
* in the cache.
*
__private_extern__ kern_return_t
memory_object_free_from_cache(
- host_t host,
- int *pager_id,
+ __unused host_t host,
+ __unused memory_object_pager_ops_t pager_ops,
int *count)
{
-
+#if VM_OBJECT_CACHE
int object_released = 0;
- int i;
register vm_object_t object = VM_OBJECT_NULL;
vm_object_t shadow;
queue_iterate(&vm_object_cached_list, object,
vm_object_t, cached_list) {
- if (object->pager && (pager_id == object->pager->pager)) {
+ if (object->pager &&
+ (pager_ops == object->pager->mo_pager_ops)) {
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 only a pager's
assert(object->pager_initialized);
assert(object->ref_count == 0);
+ vm_object_lock_assert_exclusive(object);
object->ref_count++;
/*
* (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);
}
vm_object_cache_unlock();
*count = object_released;
- return KERN_SUCCESS;
+#else
+ *count = 0;
+#endif
+ return KERN_SUCCESS;
}
{
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);
}
VM_OBJECT_EVENT_PAGER_READY,
THREAD_UNINT);
}
- *control = object->pager_request;
+ *control = object->pager_control;
vm_object_unlock(object);
}
return (KERN_SUCCESS);
{
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) {
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);
VM_OBJECT_EVENT_INITIALIZED,
THREAD_UNINT);
vm_object_unlock(object);
- vm_object_cache_unlock();
thread_block(THREAD_CONTINUE_NULL);
continue;
}
&& (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);
}
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;
}
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;
}
}
+ /*NOTREACHED*/
+ assert(0);
+ return KERN_FAILURE;
}
int flags,
vm_prot_t prot)
{
- vm_object_offset_t original_offset = offset;
- boolean_t should_flush=flags & MEMORY_OBJECT_DATA_FLUSH;
+ __unused boolean_t should_flush;
+
+ should_flush = flags & MEMORY_OBJECT_DATA_FLUSH;
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);
/*
if ((prot & ~VM_PROT_ALL) != 0 && prot != VM_PROT_NO_CHANGE)
return (KERN_INVALID_ARGUMENT);
- size = round_page(size);
+ size = round_page_64(size);
/*
* Lock the object, and acquire a paging reference to
*/
vm_object_lock(object);
vm_object_paging_begin(object);
- offset -= object->paging_offset;
(void)vm_object_update(object,
- offset, size, should_return, flags, prot);
+ offset, size, NULL, NULL, should_return, flags, prot);
vm_object_paging_end(object);
vm_object_unlock(object);
return (KERN_SUCCESS);
}
-
-
-#if TASK_SWAPPER
/*
- * vm_object_res_deallocate
- *
- * (recursively) decrement residence counts on vm objects and their shadows.
- * Called from vm_object_deallocate and when swapping out an object.
- *
- * The object is locked, and remains locked throughout the function,
- * even as we iterate down the shadow chain. Locks on intermediate objects
- * will be dropped, but not the original object.
+ * 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 purgeable object. We mark the
+ * object as "empty" after reaping its pages.
*
- * NOTE: this function used to use recursion, rather than iteration.
+ * On entry the object must be locked and it must be
+ * purgeable with no delayed copies pending.
*/
-
-__private_extern__ void
-vm_object_res_deallocate(
- vm_object_t object)
+void
+vm_object_purge(vm_object_t object, int flags)
{
- vm_object_t orig_object = object;
+ vm_object_lock_assert_exclusive(object);
+
+ if (object->purgable == VM_PURGABLE_DENY)
+ return;
+
+ assert(object->copy == VM_OBJECT_NULL);
+ assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE);
+
/*
- * Object is locked so it can be called directly
- * from vm_object_deallocate. Original object is never
- * unlocked.
+ * 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()).
*/
- assert(object->res_count > 0);
- while (--object->res_count == 0) {
- assert(object->ref_count >= object->res_count);
- vm_object_deactivate_all_pages(object);
- /* iterate on shadow, if present */
- if (object->shadow != VM_OBJECT_NULL) {
- vm_object_t tmp_object = object->shadow;
- vm_object_lock(tmp_object);
- if (object != orig_object)
- vm_object_unlock(object);
- object = tmp_object;
- assert(object->res_count > 0);
- } else
- break;
+ 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 (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;
}
- if (object != orig_object)
- vm_object_unlock(object);
+ assert(object->purgable == VM_PURGABLE_EMPTY);
+
+ vm_object_reap_pages(object, REAP_PURGEABLE);
+
+ if (object->pager != NULL &&
+ COMPRESSED_PAGER_IS_ACTIVE) {
+ unsigned int pgcount;
+
+ 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 ( !(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.
+ */
+ }
+ }
+
+ vm_object_lock_assert_exclusive(object);
}
+
/*
- * vm_object_res_reference
+ * 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 purgeable object will behave like a normal
+ * object and attempts to use this interface will result in an error return
+ * of KERN_INVALID_ARGUMENT.
*
- * Internal function to increment residence count on a vm object
- * and its shadows. It is called only from vm_object_reference, and
- * when swapping in a vm object, via vm_map_swap.
+ * 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
+ * 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 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.
*
- * The object is locked, and remains locked throughout the function,
- * even as we iterate down the shadow chain. Locks on intermediate objects
- * will be dropped, but not the original object.
+ * 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 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
+ * 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.
*
- * NOTE: this function used to use recursion, rather than iteration.
+ * Finally, the current state (non-volatile, volatile, volatile & empty) of a
+ * 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 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 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 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 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 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 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
+ * have been lost.
*/
-
-__private_extern__ void
-vm_object_res_reference(
- vm_object_t object)
+/*
+ * The object must be locked.
+ */
+kern_return_t
+vm_object_purgable_control(
+ vm_object_t object,
+ vm_purgable_t control,
+ int *state)
{
- vm_object_t orig_object = object;
- /*
- * Object is locked, so this can be called directly
- * from vm_object_reference. This lock is never released.
+ int old_state;
+ int new_state;
+
+ if (object == VM_OBJECT_NULL) {
+ /*
+ * 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 purgeable object.
*/
- while ((++object->res_count == 1) &&
- (object->shadow != VM_OBJECT_NULL)) {
- vm_object_t tmp_object = object->shadow;
+ old_state = object->purgable;
+ if (old_state == VM_PURGABLE_DENY)
+ return KERN_INVALID_ARGUMENT;
+
+ /* purgeable cant have delayed copies - now or in the future */
+ assert(object->copy == VM_OBJECT_NULL);
+ assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE);
- assert(object->ref_count >= object->res_count);
- vm_object_lock(tmp_object);
- if (object != orig_object)
- vm_object_unlock(object);
- object = tmp_object;
+ /*
+ * Execute the desired operation.
+ */
+ if (control == VM_PURGABLE_GET_STATE) {
+ *state = old_state;
+ return KERN_SUCCESS;
}
- if (object != orig_object)
- vm_object_unlock(object);
- assert(orig_object->ref_count >= orig_object->res_count);
-}
-#endif /* TASK_SWAPPER */
-/*
- * vm_object_reference:
- *
- * Gets another reference to the given object.
- */
-#ifdef vm_object_reference
-#undef vm_object_reference
-#endif
-__private_extern__ void
-vm_object_reference(
- register vm_object_t object)
-{
- if (object == VM_OBJECT_NULL)
- return;
+ if ((*state) & VM_PURGABLE_DEBUG_EMPTY) {
+ object->volatile_empty = TRUE;
+ }
+ if ((*state) & VM_PURGABLE_DEBUG_FAULT) {
+ object->volatile_fault = TRUE;
+ }
- vm_object_lock(object);
- assert(object->ref_count > 0);
- vm_object_reference_locked(object);
- vm_object_unlock(object);
-}
+ new_state = *state & VM_PURGABLE_STATE_MASK;
+ if (new_state == VM_PURGABLE_VOLATILE &&
+ object->volatile_empty) {
+ new_state = VM_PURGABLE_EMPTY;
+ }
-#ifdef MACH_BSD
-/*
- * Scale the vm_object_cache
- * This is required to make sure that the vm_object_cache is big
- * enough to effectively cache the mapped file.
- * This is really important with UBC as all the regular file vnodes
- * have memory object associated with them. Havving this cache too
- * small results in rapid reclaim of vnodes and hurts performance a LOT!
- *
- * This is also needed as number of vnodes can be dynamically scaled.
- */
-kern_return_t
-adjust_vm_object_cache(vm_size_t oval, vm_size_t nval)
-{
- vm_object_cached_max = nval;
- vm_object_cache_trim(FALSE);
- return (KERN_SUCCESS);
-}
-#endif /* MACH_BSD */
+ switch (new_state) {
+ case VM_PURGABLE_DENY:
+ case VM_PURGABLE_NONVOLATILE:
+ object->purgable = new_state;
+
+ if (old_state == VM_PURGABLE_VOLATILE) {
+ unsigned int delta;
+
+ assert(object->resident_page_count >=
+ object->wired_page_count);
+ delta = (object->resident_page_count -
+ object->wired_page_count);
+
+ assert(vm_page_purgeable_count >= delta);
+
+ 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);
+ }
+
+ vm_page_lock_queues();
+
+ /* object should be on a queue */
+ assert(object->objq.next != NULL &&
+ object->objq.prev != NULL);
+ purgeable_q_t queue;
+
+ /*
+ * Move object from its volatile queue to the
+ * non-volatile queue...
+ */
+ queue = vm_purgeable_object_remove(object);
+ assert(queue);
+
+ 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);
+ }
+
+ break;
+
+ case VM_PURGABLE_VOLATILE:
+ if (object->volatile_fault) {
+ vm_page_t p;
+ int refmod;
+
+ queue_iterate(&object->memq, p, vm_page_t, listq) {
+ if (p->busy ||
+ VM_PAGE_WIRED(p) ||
+ p->fictitious) {
+ continue;
+ }
+ refmod = pmap_disconnect(p->phys_page);
+ if ((refmod & VM_MEM_MODIFIED) &&
+ !p->dirty) {
+ SET_PAGE_DIRTY(p, FALSE);
+ }
+ }
+ }
+
+ if (old_state == VM_PURGABLE_EMPTY &&
+ object->resident_page_count == 0 &&
+ object->pager == NULL)
+ break;
+
+ 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;
+
+ queue_iterate(&object->memq, p, vm_page_t, listq) {
+ if (p->busy ||
+ VM_PAGE_WIRED(p) ||
+ p->fictitious) {
+ continue;
+ }
+ refmod = pmap_disconnect(p->phys_page);
+ 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)) {
+
+ 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(p->phys_page))) {
+
+ 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(p->phys_page))) {
+
+ 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
+ *
+ * (recursively) decrement residence counts on vm objects and their shadows.
+ * Called from vm_object_deallocate and when swapping out an object.
+ *
+ * The object is locked, and remains locked throughout the function,
+ * even as we iterate down the shadow chain. Locks on intermediate objects
+ * will be dropped, but not the original object.
+ *
+ * NOTE: this function used to use recursion, rather than iteration.
+ */
+
+__private_extern__ void
+vm_object_res_deallocate(
+ vm_object_t object)
+{
+ vm_object_t orig_object = object;
+ /*
+ * Object is locked so it can be called directly
+ * from vm_object_deallocate. Original object is never
+ * unlocked.
+ */
+ assert(object->res_count > 0);
+ while (--object->res_count == 0) {
+ assert(object->ref_count >= object->res_count);
+ vm_object_deactivate_all_pages(object);
+ /* iterate on shadow, if present */
+ if (object->shadow != VM_OBJECT_NULL) {
+ vm_object_t tmp_object = object->shadow;
+ vm_object_lock(tmp_object);
+ if (object != orig_object)
+ vm_object_unlock(object);
+ object = tmp_object;
+ assert(object->res_count > 0);
+ } else
+ break;
+ }
+ if (object != orig_object)
+ vm_object_unlock(object);
+}
+
+/*
+ * vm_object_res_reference
+ *
+ * Internal function to increment residence count on a vm object
+ * and its shadows. It is called only from vm_object_reference, and
+ * when swapping in a vm object, via vm_map_swap.
+ *
+ * The object is locked, and remains locked throughout the function,
+ * even as we iterate down the shadow chain. Locks on intermediate objects
+ * will be dropped, but not the original object.
+ *
+ * NOTE: this function used to use recursion, rather than iteration.
+ */
+
+__private_extern__ void
+vm_object_res_reference(
+ vm_object_t object)
+{
+ vm_object_t orig_object = object;
+ /*
+ * Object is locked, so this can be called directly
+ * from vm_object_reference. This lock is never released.
+ */
+ while ((++object->res_count == 1) &&
+ (object->shadow != VM_OBJECT_NULL)) {
+ vm_object_t tmp_object = object->shadow;
+
+ assert(object->ref_count >= object->res_count);
+ vm_object_lock(tmp_object);
+ if (object != orig_object)
+ vm_object_unlock(object);
+ object = tmp_object;
+ }
+ if (object != orig_object)
+ vm_object_unlock(object);
+ assert(orig_object->ref_count >= orig_object->res_count);
+}
+#endif /* TASK_SWAPPER */
+
+/*
+ * vm_object_reference:
+ *
+ * Gets another reference to the given object.
+ */
+#ifdef vm_object_reference
+#undef vm_object_reference
+#endif
+__private_extern__ void
+vm_object_reference(
+ register vm_object_t object)
+{
+ if (object == VM_OBJECT_NULL)
+ return;
+
+ vm_object_lock(object);
+ assert(object->ref_count > 0);
+ vm_object_reference_locked(object);
+ vm_object_unlock(object);
+}
+
+#ifdef MACH_BSD
+/*
+ * Scale the vm_object_cache
+ * This is required to make sure that the vm_object_cache is big
+ * enough to effectively cache the mapped file.
+ * This is really important with UBC as all the regular file vnodes
+ * have memory object associated with them. Havving this cache too
+ * small results in rapid reclaim of vnodes and hurts performance a LOT!
+ *
+ * This is also needed as number of vnodes can be dynamically scaled.
+ */
+kern_return_t
+adjust_vm_object_cache(
+ __unused vm_size_t oval,
+ __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 */
+
+
+/*
+ * vm_object_transpose
+ *
+ * This routine takes two VM objects of the same size and exchanges
+ * their backing store.
+ * The objects should be "quiesced" via a UPL operation with UPL_SET_IO_WIRE
+ * and UPL_BLOCK_ACCESS if they are referenced anywhere.
+ *
+ * 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,
+ vm_object_t object2,
+ vm_object_size_t transpose_size)
+{
+ vm_object_t tmp_object;
+ kern_return_t retval;
+ boolean_t object1_locked, object2_locked;
+ 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;
+
+ if (object1 == object2 ||
+ object1 == VM_OBJECT_NULL ||
+ object2 == VM_OBJECT_NULL) {
+ /*
+ * If the 2 VM objects are the same, there's
+ * no point in exchanging their backing store.
+ */
+ retval = KERN_INVALID_VALUE;
+ 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->alive || object1->terminating ||
+ object1->copy || object1->shadow || object1->shadowed ||
+ object1->purgable != VM_PURGABLE_DENY) {
+ /*
+ * We don't deal with copy or shadow objects (yet).
+ */
+ retval = KERN_INVALID_VALUE;
+ goto done;
+ }
+ /*
+ * 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_only_wait(object1, THREAD_UNINT);
+
+ /*
+ * Same as above for the 2nd object...
+ */
+ vm_object_lock(object2);
+ object2_locked = TRUE;
+ 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_only_wait(object2, THREAD_UNINT);
+
+
+ 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.
+ * If their size doesn't match the caller's
+ * "transpose_size", we can't do it either because the
+ * transpose operation will affect the entire span of
+ * the objects.
+ */
+ retval = KERN_INVALID_VALUE;
+ goto done;
+ }
+
+
+ /*
+ * 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)) {
+ /*
+ * 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, FALSE);
+ }
+ assert(queue_empty(&object2->memq));
+ } else if (object2->phys_contiguous || 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, FALSE);
+ }
+ assert(queue_empty(&object1->memq));
+ } else {
+ /* transfer object1's pages to tmp_object */
+ while (!queue_empty(&object1->memq)) {
+ page = (vm_page_t) queue_first(&object1->memq);
+ page_offset = page->offset;
+ vm_page_remove(page, TRUE);
+ page->offset = page_offset;
+ queue_enter(&tmp_object->memq, page, vm_page_t, listq);
+ }
+ assert(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, FALSE);
+ }
+ assert(queue_empty(&object2->memq));
+ /* transfer tmp_object's pages to object2 */
+ 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);
+ vm_page_insert(page, object2, page->offset);
+ }
+ assert(queue_empty(&tmp_object->memq));
+ }
+
+#define __TRANSPOSE_FIELD(field) \
+MACRO_BEGIN \
+ tmp_object->field = object1->field; \
+ object1->field = object2->field; \
+ 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(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) {
+ memory_object_control_collapse(object1->pager_control,
+ object1);
+ }
+ if (object2->pager_control != MEMORY_OBJECT_CONTROL_NULL) {
+ memory_object_control_collapse(object2->pager_control,
+ object2);
+ }
+ __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(pages_created);
+ __TRANSPOSE_FIELD(pages_used);
+ __TRANSPOSE_FIELD(scan_collisions);
+#if MACH_PAGEMAP
+ __TRANSPOSE_FIELD(existence_map);
+#endif
+ __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
+
+ retval = KERN_SUCCESS;
+
+done:
+ /*
+ * Cleanup.
+ */
+ if (tmp_object != VM_OBJECT_NULL) {
+ vm_object_unlock(tmp_object);
+ /*
+ * Re-initialize the temporary object to avoid
+ * deallocating a real pager.
+ */
+ _vm_object_allocate(transpose_size, tmp_object);
+ vm_object_deallocate(tmp_object);
+ tmp_object = VM_OBJECT_NULL;
+ }
+
+ if (object1_locked) {
+ vm_object_unlock(object1);
+ object1_locked = FALSE;
+ }
+ if (object2_locked) {
+ vm_object_unlock(object2);
+ object2_locked = FALSE;
+ }
+
+ vm_object_transpose_count++;
+
+ return retval;
+}
+
+
+/*
+ * 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;
+
+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_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));
+
+ /*
+ * 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 (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;
+
+ 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_SIZE)
+ min_ph_size = PAGE_SIZE;
+
+ 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;
+
+ if (max_length > max_ph_size)
+ max_length = max_ph_size;
+
+ if (max_length <= PAGE_SIZE)
+ goto out;
+
+ if (object->internal)
+ object_size = object->vo_size;
+ else
+ vnode_pager_get_object_size(object->pager, &object_size);
+
+ object_size = round_page_64(object_size);
+
+ 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));
+
+ 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 and internal objects w/o an existence map
+ * vm_externl_state_get will return VM_EXTERNAL_STATE_UNKNOWN
+ */
+#if MACH_PAGEMAP
+ if (vm_external_state_get(object->existence_map, offset) == VM_EXTERNAL_STATE_ABSENT) {
+ /*
+ * we know for a fact that the pager can't provide the page
+ * so don't include it or any pages beyond it in this cluster
+ */
+ break;
+ }
+#endif /* MACH_PAGEMAP */
+ 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 and internal objects w/o an existence map
+ * vm_externl_state_get will return VM_EXTERNAL_STATE_UNKNOWN
+ */
+#if MACH_PAGEMAP
+ if (vm_external_state_get(object->existence_map, offset) == VM_EXTERNAL_STATE_ABSENT) {
+ /*
+ * we know for a fact that the pager can't provide the page
+ * so don't include it or any pages beyond it in this cluster
+ */
+ break;
+ }
+#endif /* MACH_PAGEMAP */
+ 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(dst_page->phys_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->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;
+ }
+
+ /* 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) {
+ SET_PAGE_DIRTY(dst_page, FALSE);
+ }
+ if (ops & UPL_POP_PAGEOUT) dst_page->pageout = 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;
+ }
+
+ 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_PRECIOUS) dst_page->precious = FALSE;
+ if (ops & UPL_POP_ABSENT) dst_page->absent = FALSE;
+ if (ops & UPL_POP_BUSY) {
+ dst_page->busy = FALSE;
+ PAGE_WAKEUP(dst_page);
+ }
+ }
+
+ if (dst_page->encrypted) {
+ /*
+ * ENCRYPTED SWAP:
+ * We need to decrypt this encrypted page before the
+ * caller can access its contents.
+ * But if the caller really wants to access the page's
+ * contents, they have to keep the page "busy".
+ * Otherwise, the page could get recycled or re-encrypted
+ * at any time.
+ */
+ if ((ops & UPL_POP_SET) && (ops & UPL_POP_BUSY) &&
+ dst_page->busy) {
+ /*
+ * The page is stable enough to be accessed by
+ * the caller, so make sure its contents are
+ * not encrypted.
+ */
+ vm_page_decrypt(dst_page, 0);
+ } else {
+ /*
+ * The page is not busy, so don't bother
+ * decrypting it, since anything could
+ * happen to it between now and when the
+ * caller wants to access it.
+ * We should not give the caller access
+ * to this page.
+ */
+ assert(!phys_entry);
+ }
+ }
+
+ if (phys_entry) {
+ /*
+ * The physical page number will remain valid
+ * only if the page is kept busy.
+ * ENCRYPTED SWAP: make sure we don't let the
+ * caller access an encrypted page.
+ */
+ assert(dst_page->busy);
+ assert(!dst_page->encrypted);
+ *phys_entry = dst_page->phys_page;
+ }
+
+ break;
+ }
+
+ vm_object_unlock(object);
+ return KERN_SUCCESS;
+
+}
+
+/*
+ * vm_object_range_op offers performance enhancement over
+ * vm_object_page_op for page_op functions which do not require page
+ * level state to be returned from the call. Page_op was created to provide
+ * a low-cost alternative to page manipulation via UPLs when only a single
+ * page was involved. The range_op call establishes the ability in the _op
+ * family of functions to work on multiple pages where the lack of page level
+ * state handling allows the caller to avoid the overhead of the upl structures.
+ */
+
+kern_return_t
+vm_object_range_op(
+ vm_object_t object,
+ vm_object_offset_t offset_beg,
+ vm_object_offset_t offset_end,
+ int ops,
+ 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) {
+ *range = 0;
+ } else {
+ *range = (uint32_t) (offset_end - offset_beg);
+ assert(*range == (offset_end - offset_beg));
+ }
+ }
+ return KERN_SUCCESS;
+ }
+ vm_object_lock(object);
+
+ if (object->phys_contiguous) {
+ vm_object_unlock(object);
+ return KERN_INVALID_OBJECT;
+ }
+
+ 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);
+ /*
+ * need to relook the page up since it's
+ * state may have changed while we slept
+ * it might even belong to a different object
+ * at this point
+ */
+ continue;
+ }
+ if (dst_page->laundry) {
+ dst_page->pageout = FALSE;
+
+ vm_pageout_steal_laundry(dst_page, FALSE);
+ }
+ if (dst_page->pmapped == TRUE)
+ pmap_disconnect(dst_page->phys_page);
+
+ 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;
+
+ offset += PAGE_SIZE;
+ }
+ vm_object_unlock(object);
+
+ 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;
+ pager_object->private = TRUE;
+ retval = vm_object_populate_with_private(pager_object, offset, page_num, size);
+
+ if (retval != KERN_SUCCESS)
+ pager_object->private = clobbered_private;
+
+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);
+}
+
+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)
+{
+ return (lck_rw_try_lock_exclusive(&object->Lock));
+}
+
+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));
+}
+
+
+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);
+
+ queue_iterate(&object->memq, p, vm_page_t, listq) {
+
+ if (!p->fictitious)
+ pmap_set_cache_attributes(p->phys_page, 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
+
+kern_return_t vm_object_pack(
+ unsigned int *purgeable_count,
+ unsigned int *wired_count,
+ unsigned int *clean_count,
+ unsigned int *dirty_count,
+ unsigned int dirty_budget,
+ boolean_t *shared,
+ vm_object_t src_object,
+ struct default_freezer_handle *df_handle)
+{
+ kern_return_t kr = KERN_SUCCESS;
+
+ vm_object_lock(src_object);
+
+ *purgeable_count = *wired_count = *clean_count = *dirty_count = 0;
+ *shared = FALSE;
+
+ if (!src_object->alive || src_object->terminating){
+ kr = KERN_FAILURE;
+ goto done;
+ }
+
+ if (src_object->purgable == VM_PURGABLE_VOLATILE) {
+ *purgeable_count = src_object->resident_page_count;
+
+ /* If the default freezer handle is null, we're just walking the pages to discover how many can be hibernated */
+ if (df_handle != NULL) {
+ purgeable_q_t queue;
+ /* object should be on a queue */
+ assert(src_object->objq.next != NULL &&
+ src_object->objq.prev != NULL);
+
+ queue = vm_purgeable_object_remove(src_object);
+ assert(queue);
+ if (src_object->purgeable_when_ripe) {
+ vm_page_lock_queues();
+ vm_purgeable_token_delete_first(queue);
+ vm_page_unlock_queues();
+ }
+
+ vm_object_purge(src_object, 0);
+ assert(src_object->purgable == VM_PURGABLE_EMPTY);
+
+ /*
+ * This object was "volatile" so its pages must have
+ * already been accounted as "volatile": no change
+ * in accounting now that it's "empty".
+ */
+ }
+ goto done;
+ }
+
+ if (src_object->ref_count == 1) {
+ vm_object_pack_pages(wired_count, clean_count, dirty_count, dirty_budget, src_object, df_handle);
+ } else {
+ if (src_object->internal) {
+ *shared = TRUE;
+ }
+ }
+done:
+ vm_object_unlock(src_object);
+
+ return kr;
+}
+
+
+void
+vm_object_pack_pages(
+ unsigned int *wired_count,
+ unsigned int *clean_count,
+ unsigned int *dirty_count,
+ unsigned int dirty_budget,
+ vm_object_t src_object,
+ struct default_freezer_handle *df_handle)
+{
+ vm_page_t p, next;
+
+ next = (vm_page_t)queue_first(&src_object->memq);
+
+ while (!queue_end(&src_object->memq, (queue_entry_t)next)) {
+ p = next;
+ next = (vm_page_t)queue_next(&next->listq);
+
+ /* Finish up if we've hit our pageout limit */
+ if (dirty_budget && (dirty_budget == *dirty_count)) {
+ break;
+ }
+ assert(!p->laundry);
+
+ if (p->fictitious || p->busy )
+ continue;
+
+ if (p->absent || p->unusual || p->error)
+ continue;
+
+ if (VM_PAGE_WIRED(p)) {
+ (*wired_count)++;
+ continue;
+ }
+
+ if (df_handle == NULL) {
+ if (p->dirty || pmap_is_modified(p->phys_page)) {
+ (*dirty_count)++;
+ } else {
+ (*clean_count)++;
+ }
+ continue;
+ }
+
+ if (p->cleaning) {
+ p->pageout = TRUE;
+ continue;
+ }
+
+ if (p->pmapped == TRUE) {
+ int refmod_state;
+ refmod_state = pmap_disconnect(p->phys_page);
+ if (refmod_state & VM_MEM_MODIFIED) {
+ SET_PAGE_DIRTY(p, FALSE);
+ }
+ }
+
+ if (p->dirty) {
+ default_freezer_pack_page(p, df_handle);
+ (*dirty_count)++;
+ }
+ else {
+ VM_PAGE_FREE(p);
+ (*clean_count)++;
+ }
+ }
+}
+
+
+/*
+ * 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);
+
+ 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 (DEFAULT_FREEZER_COMPRESSED_PAGER_IS_SWAPBACKED) {
+ 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--) && !queue_empty(&object->memq)) {
+
+ p = (vm_page_t)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)) {
+ if (p->cleaning)
+ p->pageout = TRUE;
+
+ vm_page_unlock_queues();
+
+ KERNEL_DEBUG(0xe0430004 | DBG_FUNC_END, object, local_freed, 1, 0, 0);
+
+ queue_remove(&object->memq, p, vm_page_t, listq);
+ 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(p->phys_page, 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) {
+ p->pageout = FALSE;
+ vm_pageout_steal_laundry(p, TRUE);
+ }
+
+ vm_page_queues_remove(p);
+ 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.
+ */
+
+ queue_remove(&object->memq, p, vm_page_t, listq);
+ 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->pageq.next = (queue_entry_t)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);
+ }
+}
+
+kern_return_t
+vm_object_pagein(
+ vm_object_t object)
+{
+ memory_object_t pager;
+ kern_return_t kr;
+
+ vm_object_lock(object);
+
+ pager = object->pager;
+
+ if (!object->pager_ready || pager == MEMORY_OBJECT_NULL) {
+ vm_object_unlock(object);
+ return KERN_FAILURE;
+ }
+
+ vm_object_paging_wait(object, THREAD_UNINT);
+ vm_object_paging_begin(object);
+
+ object->blocked_access = TRUE;
+ vm_object_unlock(object);
+
+ kr = memory_object_data_reclaim(pager, TRUE);
+
+ vm_object_lock(object);
+
+ object->blocked_access = FALSE;
+ vm_object_paging_end(object);
+
+ vm_object_unlock(object);
+
+ return kr;
+}
+#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;
+
+ iq = &vm_pageout_queue_internal;
+
+ assert(object != VM_OBJECT_NULL );
+ assert(!DEFAULT_PAGER_IS_ACTIVE && !DEFAULT_FREEZER_IS_ACTIVE);
+
+ 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)queue_first(&object->memq);
+
+ while (!queue_end(&object->memq, (queue_entry_t)next)) {
+ p = next;
+ next = (vm_page_t)queue_next(&next->listq);
+
+ if (!(p->active || p->inactive || p->speculative) ||
+ p->encrypted_cleaning ||
+ p->cleaning ||
+ p->laundry ||
+ p->pageout ||
+ 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;
+
+ pmap_options = 0;
+ if (COMPRESSED_PAGER_IS_ACTIVE ||
+ DEFAULT_FREEZER_COMPRESSED_PAGER_IS_ACTIVE) {
+ /*
+ * 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(p->phys_page,
+ 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);
+ if (vm_pageout_cluster(p, TRUE, 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) && m->phys_page == 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, upl->upl_creator, m, 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, upl->upl_creator, 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
projects / apple / xnu.git / blobdiff