X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/b7266188b87f3620ec3f9f717e57194a7dd989fe..d9a64523371fa019c4575bb400cbbc3a50ac9903:/osfmk/vm/vm_fault.c diff --git a/osfmk/vm/vm_fault.c b/osfmk/vm/vm_fault.c index b339dbd7d..abbe202ff 100644 --- a/osfmk/vm/vm_fault.c +++ b/osfmk/vm/vm_fault.c @@ -2,7 +2,7 @@ * Copyright (c) 2000-2009 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ - * + * * 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 @@ -11,10 +11,10 @@ * 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. - * + * * 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, @@ -22,34 +22,34 @@ * 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_OSREFERENCE_LICENSE_HEADER_END@ */ /* * @OSF_COPYRIGHT@ */ -/* +/* * Mach Operating System * Copyright (c) 1991,1990,1989,1988,1987 Carnegie Mellon University * All Rights Reserved. - * + * * Permission to use, copy, modify and distribute this software and its * documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. - * + * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND FOR * ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. - * + * * Carnegie Mellon requests users of this software to return to - * + * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 - * + * * any improvements or extensions that they make and grant Carnegie Mellon * the rights to redistribute these changes. */ @@ -64,7 +64,6 @@ #include #include -#include #include #include @@ -88,9 +87,10 @@ #include #include #include +#include -#include - +#include +#include #include #include #include @@ -102,21 +102,26 @@ #include #include #include /* Needed by some vm_page.h macros */ +#include + +#include +#include +#include -#include +#include #define VM_FAULT_CLASSIFY 0 #define TRACEFAULTPAGE 0 /* (TEST/DEBUG) */ -int vm_object_pagein_throttle = 16; +unsigned int vm_object_pagein_throttle = 16; /* - * We apply a hard throttle to the demand zero rate of tasks that we believe are running out of control which + * We apply a hard throttle to the demand zero rate of tasks that we believe are running out of control which * kicks in when swap space runs out. 64-bit programs have massive address spaces and can leak enormous amounts * of memory if they're buggy and can run the system completely out of swap space. If this happens, we * impose a hard throttle on them to prevent them from taking the last bit of memory left. This helps - * keep the UI active so that the user has a chance to kill the offending task before the system + * keep the UI active so that the user has a chance to kill the offending task before the system * completely hangs. * * The hard throttle is only applied when the system is nearly completely out of swap space and is only applied @@ -125,45 +130,56 @@ int vm_object_pagein_throttle = 16; * delay of HARD_THROTTLE_DELAY microseconds before being allowed to try the page fault again. */ -boolean_t thread_is_io_throttled(void); +extern void throttle_lowpri_io(int); + +extern struct vnode *vnode_pager_lookup_vnode(memory_object_t); uint64_t vm_hard_throttle_threshold; -extern unsigned int dp_pages_free, dp_pages_reserve; -#define NEED_TO_HARD_THROTTLE_THIS_TASK() (((dp_pages_free + dp_pages_reserve < 2000) && \ - (get_task_resident_size(current_task()) > vm_hard_throttle_threshold) && \ - (current_task() != kernel_task) && IP_VALID(memory_manager_default)) || \ - (vm_page_free_count < vm_page_throttle_limit && thread_is_io_throttled() && \ - (get_task_resident_size(current_task()) > vm_hard_throttle_threshold))) +#define NEED_TO_HARD_THROTTLE_THIS_TASK() (vm_wants_task_throttled(current_task()) || \ + ((vm_page_free_count < vm_page_throttle_limit || \ + HARD_THROTTLE_LIMIT_REACHED()) && \ + proc_get_effective_thread_policy(current_thread(), TASK_POLICY_IO) >= THROTTLE_LEVEL_THROTTLED)) -#define HARD_THROTTLE_DELAY 10000 /* 10000 us == 10 ms */ +#define HARD_THROTTLE_DELAY 10000 /* 10000 us == 10 ms */ +#define SOFT_THROTTLE_DELAY 200 /* 200 us == .2 ms */ -extern int cs_debug; +#define VM_PAGE_CREATION_THROTTLE_PERIOD_SECS 6 +#define VM_PAGE_CREATION_THROTTLE_RATE_PER_SEC 20000 -#if MACH_KDB -extern struct db_watchpoint *db_watchpoint_list; -#endif /* MACH_KDB */ boolean_t current_thread_aborted(void); /* Forward declarations of internal routines. */ -extern kern_return_t vm_fault_wire_fast( +static kern_return_t vm_fault_wire_fast( vm_map_t map, vm_map_offset_t va, + vm_prot_t prot, + vm_tag_t wire_tag, vm_map_entry_t entry, pmap_t pmap, - vm_map_offset_t pmap_addr); - -extern void vm_fault_continue(void); - -extern void vm_fault_copy_cleanup( + vm_map_offset_t pmap_addr, + ppnum_t *physpage_p); + +static kern_return_t vm_fault_internal( + vm_map_t map, + vm_map_offset_t vaddr, + vm_prot_t caller_prot, + boolean_t change_wiring, + vm_tag_t wire_tag, + int interruptible, + pmap_t pmap, + vm_map_offset_t pmap_addr, + ppnum_t *physpage_p); + +static void vm_fault_copy_cleanup( vm_page_t page, vm_page_t top_page); -extern void vm_fault_copy_dst_cleanup( +static void vm_fault_copy_dst_cleanup( vm_page_t page); #if VM_FAULT_CLASSIFY @@ -174,17 +190,42 @@ extern void vm_fault_classify(vm_object_t object, extern void vm_fault_classify_init(void); #endif +unsigned long vm_pmap_enter_blocked = 0; +unsigned long vm_pmap_enter_retried = 0; unsigned long vm_cs_validates = 0; unsigned long vm_cs_revalidates = 0; unsigned long vm_cs_query_modified = 0; unsigned long vm_cs_validated_dirtied = 0; - -#if CONFIG_ENFORCE_SIGNED_CODE -int cs_enforcement_disable=0; -#else -static const int cs_enforcement_disable=1; -#endif +unsigned long vm_cs_bitmap_validated = 0; +#if PMAP_CS +uint64_t vm_cs_defer_to_pmap_cs = 0; +uint64_t vm_cs_defer_to_pmap_cs_not = 0; +#endif /* PMAP_CS */ + +void vm_pre_fault(vm_map_offset_t); + +extern char *kdp_compressor_decompressed_page; +extern addr64_t kdp_compressor_decompressed_page_paddr; +extern ppnum_t kdp_compressor_decompressed_page_ppnum; + +struct vmrtfr { + int vmrtfr_maxi; + int vmrtfr_curi; + int64_t vmrtf_total; + vm_rtfault_record_t *vm_rtf_records; +} vmrtfrs; +#define VMRTF_DEFAULT_BUFSIZE (4096) +#define VMRTF_NUM_RECORDS_DEFAULT (VMRTF_DEFAULT_BUFSIZE / sizeof(vm_rtfault_record_t)) +int vmrtf_num_records = VMRTF_NUM_RECORDS_DEFAULT; + +static void vm_rtfrecord_lock(void); +static void vm_rtfrecord_unlock(void); +static void vm_record_rtfault(thread_t, uint64_t, vm_map_offset_t, int); + +lck_spin_t vm_rtfr_slock; +extern lck_grp_t vm_page_lck_grp_bucket; +extern lck_attr_t vm_page_lck_attr; /* * Routine: vm_fault_init @@ -194,26 +235,53 @@ static const int cs_enforcement_disable=1; void vm_fault_init(void) { -#if !SECURE_KERNEL -#if CONFIG_ENFORCE_SIGNED_CODE - PE_parse_boot_argn("cs_enforcement_disable", &cs_enforcement_disable, - sizeof (cs_enforcement_disable)); -#endif - PE_parse_boot_argn("cs_debug", &cs_debug, sizeof (cs_debug)); -#endif - + int i, vm_compressor_temp; + boolean_t need_default_val = TRUE; /* * Choose a value for the hard throttle threshold based on the amount of ram. The threshold is * computed as a percentage of available memory, and the percentage used is scaled inversely with - * the amount of memory. The pertange runs between 10% and 35%. We use 35% for small memory systems + * the amount of memory. The percentage runs between 10% and 35%. We use 35% for small memory systems * and reduce the value down to 10% for very large memory configurations. This helps give us a * definition of a memory hog that makes more sense relative to the amount of ram in the machine. * The formula here simply uses the number of gigabytes of ram to adjust the percentage. */ vm_hard_throttle_threshold = sane_size * (35 - MIN((int)(sane_size / (1024*1024*1024)), 25)) / 100; + + /* + * Configure compressed pager behavior. A boot arg takes precedence over a device tree entry. + */ + + if (PE_parse_boot_argn("vm_compressor", &vm_compressor_temp, sizeof (vm_compressor_temp))) { + for ( i = 0; i < VM_PAGER_MAX_MODES; i++) { + if (vm_compressor_temp > 0 && + ((vm_compressor_temp & ( 1 << i)) == vm_compressor_temp)) { + need_default_val = FALSE; + vm_compressor_mode = vm_compressor_temp; + break; + } + } + if (need_default_val) + printf("Ignoring \"vm_compressor\" boot arg %d\n", vm_compressor_temp); + } + if (need_default_val) { + /* If no boot arg or incorrect boot arg, try device tree. */ + PE_get_default("kern.vm_compressor", &vm_compressor_mode, sizeof(vm_compressor_mode)); + } + printf("\"vm_compressor_mode\" is %d\n", vm_compressor_mode); } +void vm_rtfault_record_init(void) { + PE_parse_boot_argn("vm_rtfault_records", &vmrtf_num_records, sizeof(vmrtf_num_records)); + + assert(vmrtf_num_records >= 1); + vmrtf_num_records = MAX(vmrtf_num_records, 1); + size_t kallocsz = vmrtf_num_records * sizeof(vm_rtfault_record_t); + vmrtfrs.vm_rtf_records = kalloc(kallocsz); + bzero(vmrtfrs.vm_rtf_records, kallocsz); + vmrtfrs.vmrtfr_maxi = vmrtf_num_records - 1; + lck_spin_init(&vm_rtfr_slock, &vm_page_lck_grp_bucket, &vm_page_lck_attr); +} /* * Routine: vm_fault_cleanup * Purpose: @@ -230,14 +298,14 @@ vm_fault_init(void) */ void vm_fault_cleanup( - register vm_object_t object, - register vm_page_t top_page) + vm_object_t object, + vm_page_t top_page) { vm_object_paging_end(object); - vm_object_unlock(object); + vm_object_unlock(object); if (top_page != VM_PAGE_NULL) { - object = top_page->object; + object = VM_PAGE_OBJECT(top_page); vm_object_lock(object); VM_PAGE_FREE(top_page); @@ -246,30 +314,12 @@ vm_fault_cleanup( } } -#if MACH_CLUSTER_STATS -#define MAXCLUSTERPAGES 16 -struct { - unsigned long pages_in_cluster; - unsigned long pages_at_higher_offsets; - unsigned long pages_at_lower_offsets; -} cluster_stats_in[MAXCLUSTERPAGES]; -#define CLUSTER_STAT(clause) clause -#define CLUSTER_STAT_HIGHER(x) \ - ((cluster_stats_in[(x)].pages_at_higher_offsets)++) -#define CLUSTER_STAT_LOWER(x) \ - ((cluster_stats_in[(x)].pages_at_lower_offsets)++) -#define CLUSTER_STAT_CLUSTER(x) \ - ((cluster_stats_in[(x)].pages_in_cluster)++) -#else /* MACH_CLUSTER_STATS */ -#define CLUSTER_STAT(clause) -#endif /* MACH_CLUSTER_STATS */ - #define ALIGNED(x) (((x) & (PAGE_SIZE_64 - 1)) == 0) boolean_t vm_page_deactivate_behind = TRUE; -/* - * default sizes given VM_BEHAVIOR_DEFAULT reference behavior +/* + * default sizes given VM_BEHAVIOR_DEFAULT reference behavior */ #define VM_DEFAULT_DEACTIVATE_BEHIND_WINDOW 128 #define VM_DEFAULT_DEACTIVATE_BEHIND_CLUSTER 16 /* don't make this too big... */ @@ -494,9 +544,21 @@ vm_fault_deactivate_behind( for (n = 0; n < max_pages_in_run; n++) { m = vm_page_lookup(object, offset + run_offset + (n * pg_offset)); - if (m && !m->busy && !m->no_cache && !m->throttled && !m->fictitious && !m->absent) { + if (m && !m->vmp_laundry && !m->vmp_busy && !m->vmp_no_cache && (m->vmp_q_state != VM_PAGE_ON_THROTTLED_Q) && !m->vmp_fictitious && !m->vmp_absent) { page_run[pages_in_run++] = m; - pmap_clear_reference(m->phys_page); + + /* + * by not passing in a pmap_flush_context we will forgo any TLB flushing, local or otherwise... + * + * a TLB flush isn't really needed here since at worst we'll miss the reference bit being + * updated in the PTE if a remote processor still has this mapping cached in its TLB when the + * new reference happens. If no futher references happen on the page after that remote TLB flushes + * we'll see a clean, non-referenced page when it eventually gets pulled out of the inactive queue + * by pageout_scan, which is just fine since the last reference would have happened quite far + * in the past (TLB caches don't hang around for very long), and of course could just as easily + * have happened before we did the deactivate_behind. + */ + pmap_clear_refmod_options(VM_PAGE_GET_PHYS_PAGE(m), VM_MEM_REFERENCED, PMAP_OPTIONS_NOFLUSH, (void *)NULL); } } if (pages_in_run) { @@ -521,53 +583,92 @@ vm_fault_deactivate_behind( } -static boolean_t -vm_page_throttled(void) +#if (DEVELOPMENT || DEBUG) +uint32_t vm_page_creation_throttled_hard = 0; +uint32_t vm_page_creation_throttled_soft = 0; +uint64_t vm_page_creation_throttle_avoided = 0; +#endif /* DEVELOPMENT || DEBUG */ + +static int +vm_page_throttled(boolean_t page_kept) { clock_sec_t elapsed_sec; clock_sec_t tv_sec; clock_usec_t tv_usec; - + thread_t thread = current_thread(); - + if (thread->options & TH_OPT_VMPRIV) - return (FALSE); + return (0); - thread->t_page_creation_count++; + if (thread->t_page_creation_throttled) { + thread->t_page_creation_throttled = 0; - if (NEED_TO_HARD_THROTTLE_THIS_TASK()) - return (TRUE); + if (page_kept == FALSE) + goto no_throttle; + } + if (NEED_TO_HARD_THROTTLE_THIS_TASK()) { +#if (DEVELOPMENT || DEBUG) + thread->t_page_creation_throttled_hard++; + OSAddAtomic(1, &vm_page_creation_throttled_hard); +#endif /* DEVELOPMENT || DEBUG */ + return (HARD_THROTTLE_DELAY); + } - if (vm_page_free_count < vm_page_throttle_limit && - thread->t_page_creation_count > vm_page_creation_throttle) { + if ((vm_page_free_count < vm_page_throttle_limit || (VM_CONFIG_COMPRESSOR_IS_PRESENT && SWAPPER_NEEDS_TO_UNTHROTTLE())) && + thread->t_page_creation_count > (VM_PAGE_CREATION_THROTTLE_PERIOD_SECS * VM_PAGE_CREATION_THROTTLE_RATE_PER_SEC)) { + if (vm_page_free_wanted == 0 && vm_page_free_wanted_privileged == 0) { +#if (DEVELOPMENT || DEBUG) + OSAddAtomic64(1, &vm_page_creation_throttle_avoided); +#endif + goto no_throttle; + } clock_get_system_microtime(&tv_sec, &tv_usec); elapsed_sec = tv_sec - thread->t_page_creation_time; - if (elapsed_sec <= 6 || (thread->t_page_creation_count / elapsed_sec) >= (vm_page_creation_throttle / 6)) { + if (elapsed_sec <= VM_PAGE_CREATION_THROTTLE_PERIOD_SECS || + (thread->t_page_creation_count / elapsed_sec) >= VM_PAGE_CREATION_THROTTLE_RATE_PER_SEC) { - if (elapsed_sec >= 60) { + if (elapsed_sec >= (3 * VM_PAGE_CREATION_THROTTLE_PERIOD_SECS)) { /* * we'll reset our stats to give a well behaved app * that was unlucky enough to accumulate a bunch of pages * over a long period of time a chance to get out of * the throttled state... we reset the counter and timestamp * so that if it stays under the rate limit for the next second - * it will be back in our good graces... if it exceeds it, it + * it will be back in our good graces... if it exceeds it, it * will remain in the throttled state */ thread->t_page_creation_time = tv_sec; - thread->t_page_creation_count = (vm_page_creation_throttle / 6) * 5; + thread->t_page_creation_count = VM_PAGE_CREATION_THROTTLE_RATE_PER_SEC * (VM_PAGE_CREATION_THROTTLE_PERIOD_SECS - 1); } - ++vm_page_throttle_count; + VM_PAGEOUT_DEBUG(vm_page_throttle_count, 1); - return (TRUE); + thread->t_page_creation_throttled = 1; + + if (VM_CONFIG_COMPRESSOR_IS_PRESENT && HARD_THROTTLE_LIMIT_REACHED()) { +#if (DEVELOPMENT || DEBUG) + thread->t_page_creation_throttled_hard++; + OSAddAtomic(1, &vm_page_creation_throttled_hard); +#endif /* DEVELOPMENT || DEBUG */ + return (HARD_THROTTLE_DELAY); + } else { +#if (DEVELOPMENT || DEBUG) + thread->t_page_creation_throttled_soft++; + OSAddAtomic(1, &vm_page_creation_throttled_soft); +#endif /* DEVELOPMENT || DEBUG */ + return (SOFT_THROTTLE_DELAY); + } } thread->t_page_creation_time = tv_sec; thread->t_page_creation_count = 0; } - return (FALSE); +no_throttle: + thread->t_page_creation_count++; + + return (0); } @@ -577,11 +678,13 @@ vm_page_throttled(void) * cleanup is based on being called from vm_fault_page * * object must be locked - * object == m->object + * object == m->vmp_object */ static vm_fault_return_t -vm_fault_check(vm_object_t object, vm_page_t m, vm_page_t first_m, boolean_t interruptible_state) +vm_fault_check(vm_object_t object, vm_page_t m, vm_page_t first_m, wait_interrupt_t interruptible_state, boolean_t page_throttle) { + int throttle_delay; + if (object->shadow_severed || VM_OBJECT_PURGEABLE_FAULT_ERROR(object)) { /* @@ -599,47 +702,28 @@ vm_fault_check(vm_object_t object, vm_page_t m, vm_page_t first_m, boolean_t int return (VM_FAULT_MEMORY_ERROR); } - if (vm_backing_store_low) { - /* - * are we protecting the system from - * backing store exhaustion. If so - * sleep unless we are privileged. - */ - if (!(current_task()->priv_flags & VM_BACKING_STORE_PRIV)) { - + if (page_throttle == TRUE) { + if ((throttle_delay = vm_page_throttled(FALSE))) { + /* + * we're throttling zero-fills... + * treat this as if we couldn't grab a page + */ if (m != VM_PAGE_NULL) - VM_PAGE_FREE(m); + VM_PAGE_FREE(m); vm_fault_cleanup(object, first_m); - assert_wait((event_t)&vm_backing_store_low, THREAD_UNINT); + VM_DEBUG_EVENT(vmf_check_zfdelay, VMF_CHECK_ZFDELAY, DBG_FUNC_NONE, throttle_delay, 0, 0, 0); - thread_block(THREAD_CONTINUE_NULL); - thread_interrupt_level(interruptible_state); - - return (VM_FAULT_RETRY); - } - } - if (vm_page_throttled()) { - /* - * we're throttling zero-fills... - * treat this as if we couldn't grab a page - */ - if (m != VM_PAGE_NULL) - VM_PAGE_FREE(m); - vm_fault_cleanup(object, first_m); - - if (NEED_TO_HARD_THROTTLE_THIS_TASK()) { - delay(HARD_THROTTLE_DELAY); + delay(throttle_delay); if (current_thread_aborted()) { thread_interrupt_level(interruptible_state); return VM_FAULT_INTERRUPTED; } - } - - thread_interrupt_level(interruptible_state); + thread_interrupt_level(interruptible_state); - return (VM_FAULT_MEMORY_SHORTAGE); + return (VM_FAULT_MEMORY_SHORTAGE); + } } return (VM_FAULT_SUCCESS); } @@ -649,13 +733,16 @@ vm_fault_check(vm_object_t object, vm_page_t m, vm_page_t first_m, boolean_t int * do the work to zero fill a page and * inject it into the correct paging queue * - * m->object must be locked + * m->vmp_object must be locked * page queue lock must NOT be held */ static int vm_fault_zero_page(vm_page_t m, boolean_t no_zero_fill) { int my_fault = DBG_ZERO_FILL_FAULT; + vm_object_t object; + + object = VM_PAGE_OBJECT(m); /* * This is is a zero-fill page fault... @@ -669,42 +756,51 @@ vm_fault_zero_page(vm_page_t m, boolean_t no_zero_fill) * execution. i.e. it is the responsibility * of higher layers to call for an instruction * sync after changing the contents and before - * sending a program into this area. We + * sending a program into this area. We * choose this approach for performance */ - m->pmapped = TRUE; + m->vmp_pmapped = TRUE; + + m->vmp_cs_validated = FALSE; + m->vmp_cs_tainted = FALSE; + m->vmp_cs_nx = FALSE; - m->cs_validated = FALSE; - m->cs_tainted = FALSE; + if (no_zero_fill == TRUE) { + my_fault = DBG_NZF_PAGE_FAULT; - if (no_zero_fill == TRUE) - my_fault = DBG_NZF_PAGE_FAULT; - else { + if (m->vmp_absent && m->vmp_busy) + return (my_fault); + } else { vm_page_zero_fill(m); VM_STAT_INCR(zero_fill_count); DTRACE_VM2(zfod, int, 1, (uint64_t *), NULL); } - assert(!m->laundry); - assert(m->object != kernel_object); - //assert(m->pageq.next == NULL && m->pageq.prev == NULL); - - if (!IP_VALID(memory_manager_default) && - (m->object->purgable == VM_PURGABLE_DENY || - m->object->purgable == VM_PURGABLE_NONVOLATILE || - m->object->purgable == VM_PURGABLE_VOLATILE )) { + assert(!m->vmp_laundry); + assert(object != kernel_object); + //assert(m->vmp_pageq.next == 0 && m->vmp_pageq.prev == 0); + + if (!VM_DYNAMIC_PAGING_ENABLED() && + (object->purgable == VM_PURGABLE_DENY || + object->purgable == VM_PURGABLE_NONVOLATILE || + object->purgable == VM_PURGABLE_VOLATILE )) { + vm_page_lockspin_queues(); - queue_enter(&vm_page_queue_throttled, m, vm_page_t, pageq); - m->throttled = TRUE; - vm_page_throttled_count++; + if (!VM_DYNAMIC_PAGING_ENABLED()) { + assert(!VM_PAGE_WIRED(m)); - vm_page_unlock_queues(); - } else { - if (current_thread()->t_page_creation_count > vm_page_creation_throttle) { - m->zero_fill = TRUE; - VM_ZF_COUNT_INCR(); + /* + * can't be on the pageout queue since we don't + * have a pager to try and clean to + */ + vm_page_queues_remove(m, TRUE); + vm_page_check_pageable_safe(m); + vm_page_queue_enter(&vm_page_queue_throttled, m, vm_page_t, vmp_pageq); + m->vmp_q_state = VM_PAGE_ON_THROTTLED_Q; + vm_page_throttled_count++; } + vm_page_unlock_queues(); } return (my_fault); } @@ -720,7 +816,7 @@ vm_fault_zero_page(vm_page_t m, boolean_t no_zero_fill) * The required permissions for the page is given * in "fault_type". Desired permissions are included * in "protection". - * fault_info is passed along to determine pagein cluster + * fault_info is passed along to determine pagein cluster * limits... it contains the expected reference pattern, * cluster size if available, etc... * @@ -752,13 +848,14 @@ vm_fault_zero_page(vm_page_t m, boolean_t no_zero_fill) * The "result_page" is also left busy. It is not removed * from the pageout queues. * Special Case: - * A return value of VM_FAULT_SUCCESS_NO_PAGE means that the + * A return value of VM_FAULT_SUCCESS_NO_PAGE means that the * fault succeeded but there's no VM page (i.e. the VM object * does not actually hold VM pages, but device memory or * large pages). The object is still locked and we still hold a * paging_in_progress reference. */ unsigned int vm_fault_page_blocked_access = 0; +unsigned int vm_fault_page_forced_retry = 0; vm_fault_return_t vm_fault_page( @@ -767,10 +864,11 @@ vm_fault_page( vm_object_offset_t first_offset, /* Offset into object */ vm_prot_t fault_type, /* What access is requested */ boolean_t must_be_resident,/* Must page be resident? */ + boolean_t caller_lookup, /* caller looked up page */ /* Modifies in place: */ vm_prot_t *protection, /* Protection for mapping */ - /* Returns: */ vm_page_t *result_page, /* Page found, if successful */ + /* Returns: */ vm_page_t *top_page, /* Page in top object, if * not result_page. */ int *type_of_fault, /* if non-null, fill in with type of fault @@ -778,13 +876,9 @@ vm_fault_page( /* More arguments: */ kern_return_t *error_code, /* code if page is in error */ boolean_t no_zero_fill, /* don't zero fill absent pages */ -#if MACH_PAGEMAP - boolean_t data_supply, /* treat as data_supply if + boolean_t data_supply, /* treat as data_supply if * it is a write fault and a full * page is provided */ -#else - __unused boolean_t data_supply, -#endif vm_object_fault_info_t fault_info) { vm_page_t m; @@ -794,54 +888,40 @@ vm_fault_page( vm_object_t next_object; vm_object_t copy_object; boolean_t look_for_page; + boolean_t force_fault_retry = FALSE; vm_prot_t access_required = fault_type; vm_prot_t wants_copy_flag; - CLUSTER_STAT(int pages_at_higher_offsets;) - CLUSTER_STAT(int pages_at_lower_offsets;) kern_return_t wait_result; - boolean_t interruptible_state; + wait_interrupt_t interruptible_state; + boolean_t data_already_requested = FALSE; + vm_behavior_t orig_behavior; + vm_size_t orig_cluster_size; vm_fault_return_t error; int my_fault; uint32_t try_failed_count; int interruptible; /* how may fault be interrupted? */ + int external_state = VM_EXTERNAL_STATE_UNKNOWN; memory_object_t pager; vm_fault_return_t retval; + int grab_options; /* - * MACH page map - an optional optimization where a bit map is maintained - * by the VM subsystem for internal objects to indicate which pages of - * the object currently reside on backing store. This existence map - * duplicates information maintained by the vnode pager. It is - * created at the time of the first pageout against the object, i.e. - * at the same time pager for the object is created. The optimization - * is designed to eliminate pager interaction overhead, if it is - * 'known' that the page does not exist on backing store. - * - * MUST_ASK_PAGER() evaluates to TRUE if the page specified by object/offset is - * either marked as paged out in the existence map for the object or no - * existence map exists for the object. MUST_ASK_PAGER() is one of the - * criteria in the decision to invoke the pager. It is also used as one - * of the criteria to terminate the scan for adjacent pages in a clustered - * pagein operation. Note that MUST_ASK_PAGER() always evaluates to TRUE for - * permanent objects. Note also that if the pager for an internal object - * has not been created, the pager is not invoked regardless of the value - * of MUST_ASK_PAGER() and that clustered pagein scans are only done on an object - * for which a pager has been created. + * MUST_ASK_PAGER() evaluates to TRUE if the page specified by object/offset is + * marked as paged out in the compressor pager or the pager doesn't exist. + * Note also that if the pager for an internal object + * has not been created, the pager is not invoked regardless of the value + * of MUST_ASK_PAGER(). * * PAGED_OUT() evaluates to TRUE if the page specified by the object/offset - * is marked as paged out in the existence map for the object. PAGED_OUT() + * is marked as paged out in the compressor pager. * PAGED_OUT() is used to determine if a page has already been pushed * into a copy object in order to avoid a redundant page out operation. */ -#if MACH_PAGEMAP -#define MUST_ASK_PAGER(o, f) (vm_external_state_get((o)->existence_map, (f)) \ - != VM_EXTERNAL_STATE_ABSENT) -#define PAGED_OUT(o, f) (vm_external_state_get((o)->existence_map, (f)) \ - == VM_EXTERNAL_STATE_EXISTS) -#else -#define MUST_ASK_PAGER(o, f) (TRUE) -#define PAGED_OUT(o, f) (FALSE) -#endif +#define MUST_ASK_PAGER(o, f, s) \ + ((s = VM_COMPRESSOR_PAGER_STATE_GET((o), (f))) != VM_EXTERNAL_STATE_ABSENT) + +#define PAGED_OUT(o, f) \ + (VM_COMPRESSOR_PAGER_STATE_GET((o), (f)) == VM_EXTERNAL_STATE_EXISTS) /* * Recovery actions @@ -849,40 +929,25 @@ vm_fault_page( #define RELEASE_PAGE(m) \ MACRO_BEGIN \ PAGE_WAKEUP_DONE(m); \ - if (!m->active && !m->inactive && !m->throttled) { \ - vm_page_lockspin_queues(); \ - if (!m->active && !m->inactive && !m->throttled) \ - vm_page_activate(m); \ - vm_page_unlock_queues(); \ - } \ + if ( !VM_PAGE_PAGEABLE(m)) { \ + vm_page_lockspin_queues(); \ + if ( !VM_PAGE_PAGEABLE(m)) { \ + if (VM_CONFIG_COMPRESSOR_IS_ACTIVE) \ + vm_page_deactivate(m); \ + else \ + vm_page_activate(m); \ + } \ + vm_page_unlock_queues(); \ + } \ MACRO_END #if TRACEFAULTPAGE dbgTrace(0xBEEF0002, (unsigned int) first_object, (unsigned int) first_offset); /* (TEST/DEBUG) */ #endif - -#if MACH_KDB - /* - * If there are watchpoints set, then - * we don't want to give away write permission - * on a read fault. Make the task write fault, - * so that the watchpoint code notices the access. - */ - if (db_watchpoint_list) { - /* - * If we aren't asking for write permission, - * then don't give it away. We're using write - * faults to set the dirty bit. - */ - if (!(fault_type & VM_PROT_WRITE)) - *protection &= ~VM_PROT_WRITE; - } -#endif /* MACH_KDB */ - interruptible = fault_info->interruptible; interruptible_state = thread_interrupt_level(interruptible); - + /* * INVARIANTS (through entire routine): * @@ -928,6 +993,14 @@ vm_fault_page( #if TRACEFAULTPAGE dbgTrace(0xBEEF0003, (unsigned int) 0, (unsigned int) 0); /* (TEST/DEBUG) */ #endif + + grab_options = 0; +#if CONFIG_SECLUDED_MEMORY + if (object->can_grab_secluded) { + grab_options |= VM_PAGE_GRAB_SECLUDED; + } +#endif /* CONFIG_SECLUDED_MEMORY */ + if (!object->alive) { /* * object is no longer valid @@ -945,6 +1018,7 @@ vm_fault_page( * must be a "large page" object. We do not deal * with VM pages for this object. */ + caller_lookup = FALSE; m = VM_PAGE_NULL; goto phys_contig_object; } @@ -956,6 +1030,7 @@ vm_fault_page( * a "activity_in_progress" reference and wait for * access to be unblocked. */ + caller_lookup = FALSE; /* no longer valid after sleep */ vm_object_activity_begin(object); vm_object_paging_end(object); while (object->blocked_access) { @@ -971,30 +1046,38 @@ vm_fault_page( /* * See whether the page at 'offset' is resident */ - m = vm_page_lookup(object, offset); + if (caller_lookup == TRUE) { + /* + * The caller has already looked up the page + * and gave us the result in "result_page". + * We can use this for the first lookup but + * it loses its validity as soon as we unlock + * the object. + */ + m = *result_page; + caller_lookup = FALSE; /* no longer valid after that */ + } else { + m = vm_page_lookup(object, offset); + } #if TRACEFAULTPAGE dbgTrace(0xBEEF0004, (unsigned int) m, (unsigned int) object); /* (TEST/DEBUG) */ #endif if (m != VM_PAGE_NULL) { - if (m->busy) { + if (m->vmp_busy) { /* * The page is being brought in, * wait for it and then retry. - * - * A possible optimization: if the page - * is known to be resident, we can ignore - * pages that are absent (regardless of - * whether they're busy). */ #if TRACEFAULTPAGE dbgTrace(0xBEEF0005, (unsigned int) m, (unsigned int) 0); /* (TEST/DEBUG) */ #endif wait_result = PAGE_SLEEP(object, m, interruptible); + XPR(XPR_VM_FAULT, "vm_f_page: block busy obj 0x%X, offset 0x%X, page 0x%X\n", - object, offset, - m, 0, 0); + object, offset, + m, 0, 0); counter(c_vm_fault_page_block_busy_kernel++); if (wait_result != THREAD_AWAKENED) { @@ -1002,14 +1085,19 @@ vm_fault_page( thread_interrupt_level(interruptible_state); if (wait_result == THREAD_RESTART) - return (VM_FAULT_RETRY); + return (VM_FAULT_RETRY); else return (VM_FAULT_INTERRUPTED); } continue; } + if (m->vmp_laundry) { + m->vmp_free_when_done = FALSE; - if (m->phys_page == vm_page_guard_addr) { + if (!m->vmp_cleaning) + vm_pageout_steal_laundry(m, FALSE); + } + if (VM_PAGE_GET_PHYS_PAGE(m) == vm_page_guard_addr) { /* * Guard page: off limits ! */ @@ -1020,12 +1108,13 @@ vm_fault_page( * be just to wire or unwire it. * Let's pretend it succeeded... */ - m->busy = TRUE; + m->vmp_busy = TRUE; *result_page = m; assert(first_m == VM_PAGE_NULL); *top_page = first_m; if (type_of_fault) *type_of_fault = DBG_GUARD_FAULT; + thread_interrupt_level(interruptible_state); return VM_FAULT_SUCCESS; } else { /* @@ -1038,7 +1127,7 @@ vm_fault_page( } } - if (m->error) { + if (m->vmp_error) { /* * The page is in error, give up now. */ @@ -1054,7 +1143,7 @@ vm_fault_page( return (VM_FAULT_MEMORY_ERROR); } - if (m->restart) { + if (m->vmp_restart) { /* * The pager wants us to restart * at the top of the chain, @@ -1071,7 +1160,7 @@ vm_fault_page( return (VM_FAULT_RETRY); } - if (m->absent) { + if (m->vmp_absent) { /* * The page isn't busy, but is absent, * therefore it's deemed "unavailable". @@ -1097,11 +1186,11 @@ vm_fault_page( /* * check for any conditions that prevent * us from creating a new zero-fill page - * vm_fault_check will do all of the + * vm_fault_check will do all of the * fault cleanup in the case of an error condition * including resetting the thread_interrupt_level */ - error = vm_fault_check(object, m, first_m, interruptible_state); + error = vm_fault_check(object, m, first_m, interruptible_state, (type_of_fault == NULL) ? TRUE : FALSE); if (error != VM_FAULT_SUCCESS) return (error); @@ -1125,7 +1214,7 @@ vm_fault_page( vm_object_unlock(object); /* - * grab the original page we + * grab the original page we * 'soldered' in place and * retake lock on 'first_object' */ @@ -1141,9 +1230,11 @@ vm_fault_page( * we're going to use the absent page we just found * so convert it to a 'busy' page */ - m->absent = FALSE; - m->busy = TRUE; + m->vmp_absent = FALSE; + m->vmp_busy = TRUE; } + if (fault_info->mark_zf_absent && no_zero_fill == TRUE) + m->vmp_absent = TRUE; /* * zero-fill the page and put it on * the correct paging queue @@ -1159,29 +1250,29 @@ vm_fault_page( VM_PAGE_FREE(m); } else { first_m = m; - m->absent = FALSE; - m->busy = TRUE; + m->vmp_absent = FALSE; + m->vmp_busy = TRUE; vm_page_lockspin_queues(); - VM_PAGE_QUEUES_REMOVE(m); + vm_page_queues_remove(m, FALSE); vm_page_unlock_queues(); } XPR(XPR_VM_FAULT, "vm_f_page: unavail obj 0x%X, off 0x%X, next_obj 0x%X, newoff 0x%X\n", object, offset, next_object, - offset+object->shadow_offset,0); + offset+object->vo_shadow_offset,0); - offset += object->shadow_offset; - fault_info->lo_offset += object->shadow_offset; - fault_info->hi_offset += object->shadow_offset; + offset += object->vo_shadow_offset; + fault_info->lo_offset += object->vo_shadow_offset; + fault_info->hi_offset += object->vo_shadow_offset; access_required = VM_PROT_READ; vm_object_lock(next_object); vm_object_unlock(object); object = next_object; vm_object_paging_begin(object); - + /* * reset to default type of fault */ @@ -1190,7 +1281,7 @@ vm_fault_page( continue; } } - if ((m->cleaning) + if ((m->vmp_cleaning) && ((object != first_object) || (object->copy != VM_OBJECT_NULL)) && (fault_type & VM_PROT_WRITE)) { /* @@ -1216,14 +1307,14 @@ vm_fault_page( vm_object_reference_locked(object); vm_fault_cleanup(object, first_m); - + counter(c_vm_fault_page_block_backoff_kernel++); vm_object_lock(object); assert(object->ref_count > 0); m = vm_page_lookup(object, offset); - if (m != VM_PAGE_NULL && m->cleaning) { + if (m != VM_PAGE_NULL && m->vmp_cleaning) { PAGE_ASSERT_WAIT(m, interruptible); vm_object_unlock(object); @@ -1240,14 +1331,14 @@ vm_fault_page( return (VM_FAULT_RETRY); } } - if (type_of_fault == NULL && m->speculative && + if (type_of_fault == NULL && (m->vmp_q_state == VM_PAGE_ON_SPECULATIVE_Q) && !(fault_info != NULL && fault_info->stealth)) { /* * If we were passed a non-NULL pointer for * "type_of_fault", than we came from * vm_fault... we'll let it deal with * this condition, since it - * needs to see m->speculative to correctly + * needs to see m->vmp_speculative to correctly * account the pageins, otherwise... * take it off the speculative queue, we'll * let the caller of vm_fault_page deal @@ -1258,35 +1349,13 @@ vm_fault_page( * the page in the speculative queue. */ vm_page_lockspin_queues(); - VM_PAGE_QUEUES_REMOVE(m); + if (m->vmp_q_state == VM_PAGE_ON_SPECULATIVE_Q) + vm_page_queues_remove(m, FALSE); vm_page_unlock_queues(); } + assert(object == VM_PAGE_OBJECT(m)); - if (m->encrypted) { - /* - * ENCRYPTED SWAP: - * the user needs access to a page that we - * encrypted before paging it out. - * Decrypt the page now. - * Keep it busy to prevent anyone from - * accessing it during the decryption. - */ - m->busy = TRUE; - vm_page_decrypt(m, 0); - assert(object == m->object); - assert(m->busy); - PAGE_WAKEUP_DONE(m); - - /* - * Retry from the top, in case - * something changed while we were - * decrypting. - */ - continue; - } - ASSERT_PAGE_DECRYPTED(m); - - if (m->object->code_signed) { + if (object->code_signed) { /* * CODE SIGNING: * We just paged in a page from a signed @@ -1311,13 +1380,13 @@ vm_fault_page( XPR(XPR_VM_FAULT, "vm_f_page: found page obj 0x%X, offset 0x%X, page 0x%X\n", object, offset, m, 0, 0); - assert(!m->busy); - assert(!m->absent); + assert(!m->vmp_busy); + assert(!m->vmp_absent); - m->busy = TRUE; + m->vmp_busy = TRUE; break; } - + /* * we get here when there is no page present in the object at @@ -1326,16 +1395,41 @@ vm_fault_page( * this object can provide the data or we're the top object... * object is locked; m == NULL */ - look_for_page = (object->pager_created && (MUST_ASK_PAGER(object, offset) == TRUE) && !data_supply); - + + if (must_be_resident) { + if (fault_type == VM_PROT_NONE && + object == kernel_object) { + /* + * We've been called from vm_fault_unwire() + * while removing a map entry that was allocated + * with KMA_KOBJECT and KMA_VAONLY. This page + * is not present and there's nothing more to + * do here (nothing to unwire). + */ + vm_fault_cleanup(object, first_m); + thread_interrupt_level(interruptible_state); + + return VM_FAULT_MEMORY_ERROR; + } + + goto dont_look_for_page; + } + + /* Don't expect to fault pages into the kernel object. */ + assert(object != kernel_object); + + data_supply = FALSE; + + look_for_page = (object->pager_created && (MUST_ASK_PAGER(object, offset, external_state) == TRUE) && !data_supply); + #if TRACEFAULTPAGE dbgTrace(0xBEEF000C, (unsigned int) look_for_page, (unsigned int) object); /* (TEST/DEBUG) */ #endif - if ((look_for_page || (object == first_object)) && !must_be_resident && !object->phys_contiguous) { + if (!look_for_page && object == first_object && !object->phys_contiguous) { /* - * Allocate a new page for this object/offset pair + * Allocate a new page for this object/offset pair as a placeholder */ - m = vm_page_grab(); + m = vm_page_grab_options(grab_options); #if TRACEFAULTPAGE dbgTrace(0xBEEF000D, (unsigned int) m, (unsigned int) object); /* (TEST/DEBUG) */ #endif @@ -1346,10 +1440,16 @@ vm_fault_page( return (VM_FAULT_MEMORY_SHORTAGE); } - vm_page_insert(m, object, offset); + + if (fault_info && fault_info->batch_pmap_op == TRUE) { + vm_page_insert_internal(m, object, offset, VM_KERN_MEMORY_NONE, FALSE, TRUE, TRUE, FALSE, NULL); + } else { + vm_page_insert(m, object, offset); + } } - if (look_for_page && !must_be_resident) { + if (look_for_page) { kern_return_t rc; + int my_fault_type; /* * If the memory manager is not ready, we @@ -1416,8 +1516,8 @@ vm_fault_page( vm_object_lock(object); assert(object->ref_count > 0); - if (object->paging_in_progress > vm_object_pagein_throttle) { - vm_object_assert_wait(object, VM_OBJECT_EVENT_PAGING_IN_PROGRESS, interruptible); + if (object->paging_in_progress >= vm_object_pagein_throttle) { + vm_object_assert_wait(object, VM_OBJECT_EVENT_PAGING_ONLY_IN_PROGRESS, interruptible); vm_object_unlock(object); wait_result = thread_block(THREAD_CONTINUE_NULL); @@ -1432,13 +1532,140 @@ vm_fault_page( return (VM_FAULT_RETRY); } } + if (object->internal) { + int compressed_count_delta; + + assert(VM_CONFIG_COMPRESSOR_IS_PRESENT); + + if (m == VM_PAGE_NULL) { + /* + * Allocate a new page for this object/offset pair as a placeholder + */ + m = vm_page_grab_options(grab_options); +#if TRACEFAULTPAGE + dbgTrace(0xBEEF000D, (unsigned int) m, (unsigned int) object); /* (TEST/DEBUG) */ +#endif + if (m == VM_PAGE_NULL) { + + vm_fault_cleanup(object, first_m); + thread_interrupt_level(interruptible_state); + + return (VM_FAULT_MEMORY_SHORTAGE); + } + + m->vmp_absent = TRUE; + if (fault_info && fault_info->batch_pmap_op == TRUE) { + vm_page_insert_internal(m, object, offset, VM_KERN_MEMORY_NONE, FALSE, TRUE, TRUE, FALSE, NULL); + } else { + vm_page_insert(m, object, offset); + } + } + assert(m->vmp_busy); + + m->vmp_absent = TRUE; + pager = object->pager; + + assert(object->paging_in_progress > 0); + vm_object_unlock(object); + + rc = vm_compressor_pager_get( + pager, + offset + object->paging_offset, + VM_PAGE_GET_PHYS_PAGE(m), + &my_fault_type, + 0, + &compressed_count_delta); + + if (type_of_fault == NULL) { + int throttle_delay; + + /* + * we weren't called from vm_fault, so we + * need to apply page creation throttling + * do it before we re-acquire any locks + */ + if (my_fault_type == DBG_COMPRESSOR_FAULT) { + if ((throttle_delay = vm_page_throttled(TRUE))) { + VM_DEBUG_EVENT(vmf_compressordelay, VMF_COMPRESSORDELAY, DBG_FUNC_NONE, throttle_delay, 0, 1, 0); + delay(throttle_delay); + } + } + } + vm_object_lock(object); + assert(object->paging_in_progress > 0); + + vm_compressor_pager_count( + pager, + compressed_count_delta, + FALSE, /* shared_lock */ + object); + + switch (rc) { + case KERN_SUCCESS: + m->vmp_absent = FALSE; + m->vmp_dirty = TRUE; + if ((object->wimg_bits & + VM_WIMG_MASK) != + VM_WIMG_USE_DEFAULT) { + /* + * If the page is not cacheable, + * we can't let its contents + * linger in the data cache + * after the decompression. + */ + pmap_sync_page_attributes_phys( + VM_PAGE_GET_PHYS_PAGE(m)); + } else { + m->vmp_written_by_kernel = TRUE; + } + + /* + * If the object is purgeable, its + * owner's purgeable ledgers have been + * updated in vm_page_insert() but the + * page was also accounted for in a + * "compressed purgeable" ledger, so + * update that now. + */ + if (((object->purgable != + VM_PURGABLE_DENY) || + object->vo_ledger_tag) && + (object->vo_owner != + NULL)) { + /* + * One less compressed + * purgeable/tagged page. + */ + vm_object_owner_compressed_update( + object, + -1); + } + + break; + case KERN_MEMORY_FAILURE: + m->vmp_unusual = TRUE; + m->vmp_error = TRUE; + m->vmp_absent = FALSE; + break; + case KERN_MEMORY_ERROR: + assert(m->vmp_absent); + break; + default: + panic("vm_fault_page(): unexpected " + "error %d from " + "vm_compressor_pager_get()\n", + rc); + } + PAGE_WAKEUP_DONE(m); + + rc = KERN_SUCCESS; + goto data_requested; + } + my_fault_type = DBG_PAGEIN_FAULT; + if (m != VM_PAGE_NULL) { - /* - * Indicate that the page is waiting for data - * from the memory manager. - */ - m->list_req_pending = TRUE; - m->absent = TRUE; + VM_PAGE_FREE(m); + m = VM_PAGE_NULL; } #if TRACEFAULTPAGE @@ -1447,7 +1674,7 @@ vm_fault_page( /* * It's possible someone called vm_object_destroy while we weren't - * holding the object lock. If that has happened, then bail out + * holding the object lock. If that has happened, then bail out * here. */ @@ -1464,6 +1691,10 @@ vm_fault_page( * so we can release the object lock. */ + if (object->object_is_shared_cache) { + set_thread_rwlock_boost(); + } + vm_object_unlock(object); /* @@ -1487,6 +1718,45 @@ vm_fault_page( object, offset, m, access_required | wants_copy_flag, 0); + if (object->copy == first_object) { + /* + * if we issue the memory_object_data_request in + * this state, we are subject to a deadlock with + * the underlying filesystem if it is trying to + * shrink the file resulting in a push of pages + * into the copy object... that push will stall + * on the placeholder page, and if the pushing thread + * is holding a lock that is required on the pagein + * path (such as a truncate lock), we'll deadlock... + * to avoid this potential deadlock, we throw away + * our placeholder page before calling memory_object_data_request + * and force this thread to retry the vm_fault_page after + * we have issued the I/O. the second time through this path + * we will find the page already in the cache (presumably still + * busy waiting for the I/O to complete) and then complete + * the fault w/o having to go through memory_object_data_request again + */ + assert(first_m != VM_PAGE_NULL); + assert(VM_PAGE_OBJECT(first_m) == first_object); + + vm_object_lock(first_object); + VM_PAGE_FREE(first_m); + vm_object_paging_end(first_object); + vm_object_unlock(first_object); + + first_m = VM_PAGE_NULL; + force_fault_retry = TRUE; + + vm_fault_page_forced_retry++; + } + + if (data_already_requested == TRUE) { + orig_behavior = fault_info->behavior; + orig_cluster_size = fault_info->cluster_size; + + fault_info->behavior = VM_BEHAVIOR_RANDOM; + fault_info->cluster_size = PAGE_SIZE; + } /* * Call the memory manager to retrieve the data. */ @@ -1497,11 +1767,23 @@ vm_fault_page( access_required | wants_copy_flag, (memory_object_fault_info_t)fault_info); + if (data_already_requested == TRUE) { + fault_info->behavior = orig_behavior; + fault_info->cluster_size = orig_cluster_size; + } else + data_already_requested = TRUE; + + DTRACE_VM2(maj_fault, int, 1, (uint64_t *), NULL); #if TRACEFAULTPAGE dbgTrace(0xBEEF0013, (unsigned int) object, (unsigned int) rc); /* (TEST/DEBUG) */ #endif vm_object_lock(object); + if (object->object_is_shared_cache) { + clear_thread_rwlock_boost(); + } + + data_requested: if (rc != KERN_SUCCESS) { vm_fault_cleanup(object, first_m); @@ -1513,22 +1795,31 @@ vm_fault_page( } else { clock_sec_t tv_sec; clock_usec_t tv_usec; - - clock_get_system_microtime(&tv_sec, &tv_usec); - current_thread()->t_page_creation_time = tv_sec; - current_thread()->t_page_creation_count = 0; + + if (my_fault_type == DBG_PAGEIN_FAULT) { + clock_get_system_microtime(&tv_sec, &tv_usec); + current_thread()->t_page_creation_time = tv_sec; + current_thread()->t_page_creation_count = 0; + } } - if ((interruptible != THREAD_UNINT) && (current_thread()->sched_mode & TH_MODE_ABORT)) { + if ((interruptible != THREAD_UNINT) && (current_thread()->sched_flags & TH_SFLAG_ABORT)) { vm_fault_cleanup(object, first_m); thread_interrupt_level(interruptible_state); return (VM_FAULT_INTERRUPTED); } + if (force_fault_retry == TRUE) { + + vm_fault_cleanup(object, first_m); + thread_interrupt_level(interruptible_state); + + return (VM_FAULT_RETRY); + } if (m == VM_PAGE_NULL && object->phys_contiguous) { /* * No page here means that the object we - * initially looked up was "physically + * initially looked up was "physically * contiguous" (i.e. device memory). However, * with Virtual VRAM, the object might not * be backed by that device memory anymore, @@ -1547,7 +1838,7 @@ vm_fault_page( * if we make it through the state checks * above, than we'll count it as such */ - my_fault = DBG_PAGEIN_FAULT; + my_fault = my_fault_type; /* * Retry with same object/offset, since new data may @@ -1556,12 +1847,12 @@ vm_fault_page( */ continue; } - +dont_look_for_page: /* - * We get here if the object has no pager, or an existence map + * We get here if the object has no pager, or an existence map * exists and indicates the page isn't present on the pager * or we're unwiring a page. If a pager exists, but there - * is no existence map, then the m->absent case above handles + * is no existence map, then the m->vmp_absent case above handles * the ZF case when the pager can't provide the page */ #if TRACEFAULTPAGE @@ -1595,23 +1886,23 @@ vm_fault_page( vm_object_lock(object); } m = first_m; - assert(m->object == object); + assert(VM_PAGE_OBJECT(m) == object); first_m = VM_PAGE_NULL; /* * check for any conditions that prevent * us from creating a new zero-fill page - * vm_fault_check will do all of the + * vm_fault_check will do all of the * fault cleanup in the case of an error condition * including resetting the thread_interrupt_level */ - error = vm_fault_check(object, m, first_m, interruptible_state); + error = vm_fault_check(object, m, first_m, interruptible_state, (type_of_fault == NULL) ? TRUE : FALSE); if (error != VM_FAULT_SUCCESS) return (error); if (m == VM_PAGE_NULL) { - m = vm_page_grab(); + m = vm_page_grab_options(grab_options); if (m == VM_PAGE_NULL) { vm_fault_cleanup(object, VM_PAGE_NULL); @@ -1621,6 +1912,9 @@ vm_fault_page( } vm_page_insert(m, object, offset); } + if (fault_info->mark_zf_absent && no_zero_fill == TRUE) + m->vmp_absent = TRUE; + my_fault = vm_fault_zero_page(m, no_zero_fill); break; @@ -1633,9 +1927,9 @@ vm_fault_page( if ((object != first_object) || must_be_resident) vm_object_paging_end(object); - offset += object->shadow_offset; - fault_info->lo_offset += object->shadow_offset; - fault_info->hi_offset += object->shadow_offset; + offset += object->vo_shadow_offset; + fault_info->lo_offset += object->vo_shadow_offset; + fault_info->hi_offset += object->vo_shadow_offset; access_required = VM_PROT_READ; vm_object_lock(next_object); @@ -1668,19 +1962,12 @@ vm_fault_page( dbgTrace(0xBEEF0015, (unsigned int) object, (unsigned int) m); /* (TEST/DEBUG) */ #endif #if EXTRA_ASSERTIONS - assert(m->busy && !m->absent); + assert(m->vmp_busy && !m->vmp_absent); assert((first_m == VM_PAGE_NULL) || - (first_m->busy && !first_m->absent && - !first_m->active && !first_m->inactive)); + (first_m->vmp_busy && !first_m->vmp_absent && + !first_m->vmp_active && !first_m->vmp_inactive && !first_m->vmp_secluded)); #endif /* EXTRA_ASSERTIONS */ - /* - * ENCRYPTED SWAP: - * If we found a page, we must have decrypted it before we - * get here... - */ - ASSERT_PAGE_DECRYPTED(m); - XPR(XPR_VM_FAULT, "vm_f_page: FOUND obj 0x%X, off 0x%X, page 0x%X, 1_obj 0x%X, 1_m 0x%X\n", object, offset, m, @@ -1706,25 +1993,6 @@ vm_fault_page( */ assert(!must_be_resident); - /* - * are we protecting the system from - * backing store exhaustion. If so - * sleep unless we are privileged. - */ - if (vm_backing_store_low) { - if (!(current_task()->priv_flags & VM_BACKING_STORE_PRIV)) { - - RELEASE_PAGE(m); - vm_fault_cleanup(object, first_m); - - assert_wait((event_t)&vm_backing_store_low, THREAD_UNINT); - - thread_block(THREAD_CONTINUE_NULL); - thread_interrupt_level(interruptible_state); - - return (VM_FAULT_RETRY); - } - } /* * If we try to collapse first_object at this * point, we may deadlock when we try to get @@ -1744,7 +2012,7 @@ vm_fault_page( /* * Allocate a page for the copy */ - copy_m = vm_page_grab(); + copy_m = vm_page_grab_options(grab_options); if (copy_m == VM_PAGE_NULL) { RELEASE_PAGE(m); @@ -1773,15 +2041,30 @@ vm_fault_page( * access to this page, then we could * avoid the pmap_disconnect() call. */ - if (m->pmapped) - pmap_disconnect(m->phys_page); + if (m->vmp_pmapped) + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m)); - assert(!m->cleaning); + if (m->vmp_clustered) { + VM_PAGE_COUNT_AS_PAGEIN(m); + VM_PAGE_CONSUME_CLUSTERED(m); + } + assert(!m->vmp_cleaning); /* * We no longer need the old page or object. */ - PAGE_WAKEUP_DONE(m); + RELEASE_PAGE(m); + + /* + * This check helps with marking the object as having a sequential pattern + * Normally we'll miss doing this below because this fault is about COW to + * the first_object i.e. bring page in from disk, push to object above but + * don't update the file object's sequential pattern. + */ + if (object->internal == FALSE) { + vm_fault_is_sequential(object, offset, fault_info->behavior); + } + vm_object_paging_end(object); vm_object_unlock(object); @@ -1800,14 +2083,14 @@ vm_fault_page( */ VM_PAGE_FREE(first_m); first_m = VM_PAGE_NULL; - + /* * and replace it with the * page we just copied into */ - assert(copy_m->busy); + assert(copy_m->vmp_busy); vm_page_insert(copy_m, object, offset); - copy_m->dirty = TRUE; + SET_PAGE_DIRTY(copy_m, TRUE); m = copy_m; /* @@ -1815,8 +2098,8 @@ vm_fault_page( * way, let's try to collapse the top object. * But we have to play ugly games with * paging_in_progress to do that... - */ - vm_object_paging_end(object); + */ + vm_object_paging_end(object); vm_object_collapse(object, offset, TRUE); vm_object_paging_begin(object); @@ -1880,9 +2163,9 @@ vm_fault_page( /* * Does the page exist in the copy? */ - copy_offset = first_offset - copy_object->shadow_offset; + copy_offset = first_offset - copy_object->vo_shadow_offset; - if (copy_object->size <= copy_offset) + if (copy_object->vo_size <= copy_offset) /* * Copy object doesn't cover this page -- do nothing. */ @@ -1891,7 +2174,7 @@ vm_fault_page( /* * Page currently exists in the copy object */ - if (copy_m->busy) { + if (copy_m->vmp_busy) { /* * If the page is being brought * in, wait for it and then retry. @@ -1913,13 +2196,8 @@ vm_fault_page( copy_object->ref_count--; assert(copy_object->ref_count > 0); copy_m = vm_page_lookup(copy_object, copy_offset); - /* - * ENCRYPTED SWAP: - * it's OK if the "copy_m" page is encrypted, - * because we're not moving it nor handling its - * contents. - */ - if (copy_m != VM_PAGE_NULL && copy_m->busy) { + + if (copy_m != VM_PAGE_NULL && copy_m->vmp_busy) { PAGE_ASSERT_WAIT(copy_m, interruptible); vm_object_unlock(copy_object); @@ -1945,32 +2223,7 @@ vm_fault_page( * for example) or it hasn't been paged out. * (VM_EXTERNAL_STATE_UNKNOWN||VM_EXTERNAL_STATE_ABSENT) * We must copy the page to the copy object. - */ - - if (vm_backing_store_low) { - /* - * we are protecting the system from - * backing store exhaustion. If so - * sleep unless we are privileged. - */ - if (!(current_task()->priv_flags & VM_BACKING_STORE_PRIV)) { - assert_wait((event_t)&vm_backing_store_low, THREAD_UNINT); - - RELEASE_PAGE(m); - VM_OBJ_RES_DECR(copy_object); - vm_object_lock_assert_exclusive(copy_object); - copy_object->ref_count--; - assert(copy_object->ref_count > 0); - - vm_object_unlock(copy_object); - vm_fault_cleanup(object, first_m); - thread_block(THREAD_CONTINUE_NULL); - thread_interrupt_level(interruptible_state); - - return (VM_FAULT_RETRY); - } - } - /* + * * Allocate a page for the copy */ copy_m = vm_page_alloc(copy_object, copy_offset); @@ -1993,43 +2246,46 @@ vm_fault_page( * Must copy page into copy-object. */ vm_page_copy(m, copy_m); - + /* * If the old page was in use by any users * of the copy-object, it must be removed * from all pmaps. (We can't know which * pmaps use it.) */ - if (m->pmapped) - pmap_disconnect(m->phys_page); + if (m->vmp_pmapped) + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m)); + if (m->vmp_clustered) { + VM_PAGE_COUNT_AS_PAGEIN(m); + VM_PAGE_CONSUME_CLUSTERED(m); + } /* * If there's a pager, then immediately * page out this page, using the "initialize" * option. Else, we use the copy. */ - if ((!copy_object->pager_created) -#if MACH_PAGEMAP - || vm_external_state_get(copy_object->existence_map, copy_offset) == VM_EXTERNAL_STATE_ABSENT -#endif - ) { + if ((!copy_object->pager_ready) + || VM_COMPRESSOR_PAGER_STATE_GET(copy_object, copy_offset) == VM_EXTERNAL_STATE_ABSENT + ) { vm_page_lockspin_queues(); - assert(!m->cleaning); + assert(!m->vmp_cleaning); vm_page_activate(copy_m); vm_page_unlock_queues(); - copy_m->dirty = TRUE; + SET_PAGE_DIRTY(copy_m, TRUE); PAGE_WAKEUP_DONE(copy_m); - } - else { - assert(copy_m->busy == TRUE); - assert(!m->cleaning); + + } else { + + assert(copy_m->vmp_busy == TRUE); + assert(!m->vmp_cleaning); /* * dirty is protected by the object lock */ - copy_m->dirty = TRUE; + SET_PAGE_DIRTY(copy_m, TRUE); /* * The page is already ready for pageout: @@ -2067,6 +2323,7 @@ vm_fault_page( */ vm_object_lock(object); } + /* * Because we're pushing a page upward * in the object tree, we must restart @@ -2076,8 +2333,8 @@ vm_fault_page( * wait result]. Can't turn off the page's * busy bit because we're not done with it. */ - if (m->wanted) { - m->wanted = FALSE; + if (m->vmp_wanted) { + m->vmp_wanted = FALSE; thread_wakeup_with_result((event_t) m, THREAD_RESTART); } } @@ -2092,7 +2349,7 @@ vm_fault_page( copy_object->ref_count--; assert(copy_object->ref_count > 0); - VM_OBJ_RES_DECR(copy_object); + VM_OBJ_RES_DECR(copy_object); vm_object_unlock(copy_object); break; @@ -2107,21 +2364,18 @@ done: object, offset, m, first_m, 0); if (m != VM_PAGE_NULL) { + assert(VM_PAGE_OBJECT(m) == object); + retval = VM_FAULT_SUCCESS; + if (my_fault == DBG_PAGEIN_FAULT) { - VM_STAT_INCR(pageins); - DTRACE_VM2(pgin, int, 1, (uint64_t *), NULL); - DTRACE_VM2(maj_fault, int, 1, (uint64_t *), NULL); - current_task()->pageins++; + VM_PAGE_COUNT_AS_PAGEIN(m); - if (m->object->internal) { - DTRACE_VM2(anonpgin, int, 1, (uint64_t *), NULL); + if (object->internal) my_fault = DBG_PAGEIND_FAULT; - } else { - DTRACE_VM2(fspgin, int, 1, (uint64_t *), NULL); + else my_fault = DBG_PAGEINV_FAULT; - } /* * evaluate access pattern and update state @@ -2129,8 +2383,23 @@ done: * state being up to date */ vm_fault_is_sequential(object, offset, fault_info->behavior); + vm_fault_deactivate_behind(object, offset, fault_info->behavior); + } else if (type_of_fault == NULL && my_fault == DBG_CACHE_HIT_FAULT) { + /* + * we weren't called from vm_fault, so handle the + * accounting here for hits in the cache + */ + if (m->vmp_clustered) { + VM_PAGE_COUNT_AS_PAGEIN(m); + VM_PAGE_CONSUME_CLUSTERED(m); + } + vm_fault_is_sequential(object, offset, fault_info->behavior); vm_fault_deactivate_behind(object, offset, fault_info->behavior); + + } else if (my_fault == DBG_COMPRESSOR_FAULT || my_fault == DBG_COMPRESSOR_SWAPIN_FAULT) { + + VM_STAT_INCR(decompressions); } if (type_of_fault) *type_of_fault = my_fault; @@ -2167,22 +2436,25 @@ backoff: * 3. the page belongs to a code-signed object * 4. the page has not been validated yet or has been mapped for write. */ -#define VM_FAULT_NEED_CS_VALIDATION(pmap, page) \ +#define VM_FAULT_NEED_CS_VALIDATION(pmap, page, page_obj) \ ((pmap) != kernel_pmap /*1*/ && \ - !(page)->cs_tainted /*2*/ && \ - (page)->object->code_signed /*3*/ && \ - (!(page)->cs_validated || (page)->wpmapped /*4*/)) + !(page)->vmp_cs_tainted /*2*/ && \ + (page_obj)->code_signed /*3*/ && \ + (!(page)->vmp_cs_validated || (page)->vmp_wpmapped /*4*/)) /* * page queue lock must NOT be held - * m->object must be locked + * m->vmp_object must be locked * - * NOTE: m->object could be locked "shared" only if we are called + * NOTE: m->vmp_object could be locked "shared" only if we are called * from vm_fault() as part of a soft fault. If so, we must be * careful not to modify the VM object in any way that is not * legal under a shared lock... */ +extern int panic_on_cs_killed; +extern int proc_selfpid(void); +extern char *proc_name_address(void *p); unsigned long cs_enter_tainted_rejected = 0; unsigned long cs_enter_tainted_accepted = 0; kern_return_t @@ -2190,65 +2462,93 @@ vm_fault_enter(vm_page_t m, pmap_t pmap, vm_map_offset_t vaddr, vm_prot_t prot, + vm_prot_t caller_prot, boolean_t wired, boolean_t change_wiring, - boolean_t no_cache, + vm_tag_t wire_tag, + vm_object_fault_info_t fault_info, + boolean_t *need_retry, int *type_of_fault) { - unsigned int cache_attr; - kern_return_t kr; - boolean_t previously_pmapped = m->pmapped; + kern_return_t kr, pe_result; + boolean_t previously_pmapped = m->vmp_pmapped; boolean_t must_disconnect = 0; boolean_t map_is_switched, map_is_switch_protected; - - vm_object_lock_assert_held(m->object); -#if DEBUG - lck_mtx_assert(&vm_page_queue_lock, LCK_MTX_ASSERT_NOTOWNED); -#endif /* DEBUG */ - - if (m->phys_page == vm_page_guard_addr) { - assert(m->fictitious); + boolean_t cs_violation; + int cs_enforcement_enabled; + vm_prot_t fault_type; + vm_object_t object; + boolean_t no_cache = fault_info->no_cache; + boolean_t cs_bypass = fault_info->cs_bypass; + int pmap_options = fault_info->pmap_options; + + fault_type = change_wiring ? VM_PROT_NONE : caller_prot; + object = VM_PAGE_OBJECT(m); + + vm_object_lock_assert_held(object); + +#if KASAN + if (pmap == kernel_pmap) { + kasan_notify_address(vaddr, PAGE_SIZE); + } +#endif + + LCK_MTX_ASSERT(&vm_page_queue_lock, LCK_MTX_ASSERT_NOTOWNED); + + if (VM_PAGE_GET_PHYS_PAGE(m) == vm_page_guard_addr) { + assert(m->vmp_fictitious); return KERN_SUCCESS; } - cache_attr = ((unsigned int)m->object->wimg_bits) & VM_WIMG_MASK; + if (*type_of_fault == DBG_ZERO_FILL_FAULT) { + + vm_object_lock_assert_exclusive(object); - if (m->pmapped == FALSE) { + } else if ((fault_type & VM_PROT_WRITE) == 0 && + (!m->vmp_wpmapped +#if VM_OBJECT_ACCESS_TRACKING + || object->access_tracking +#endif /* VM_OBJECT_ACCESS_TRACKING */ + )) { /* - * This is the first time this page is being - * mapped in an address space (pmapped == FALSE). - * - * Part of that page may still be in the data cache - * and not flushed to memory. In case we end up - * accessing that page via the instruction cache, - * we need to ensure that the 2 caches are in sync. + * This is not a "write" fault, so we + * might not have taken the object lock + * exclusively and we might not be able + * to update the "wpmapped" bit in + * vm_fault_enter(). + * Let's just grant read access to + * the page for now and we'll + * soft-fault again if we need write + * access later... */ - pmap_sync_page_data_phys(m->phys_page); - if ((*type_of_fault == DBG_CACHE_HIT_FAULT) && m->clustered) { - /* - * found it in the cache, but this - * is the first fault-in of the page (m->pmapped == FALSE) - * so it must have come in as part of - * a cluster... account 1 pagein against it - */ - VM_STAT_INCR(pageins); - DTRACE_VM2(pgin, int, 1, (uint64_t *), NULL); + /* This had better not be a JIT page. */ + if (!pmap_has_prot_policy(prot)) { + prot &= ~VM_PROT_WRITE; + } else { + assert(cs_bypass); + } + } + if (m->vmp_pmapped == FALSE) { - if (m->object->internal) { - DTRACE_VM2(anonpgin, int, 1, (uint64_t *), NULL); - *type_of_fault = DBG_PAGEIND_FAULT; - } else { - DTRACE_VM2(fspgin, int, 1, (uint64_t *), NULL); - *type_of_fault = DBG_PAGEINV_FAULT; - } + if (m->vmp_clustered) { + if (*type_of_fault == DBG_CACHE_HIT_FAULT) { + /* + * found it in the cache, but this + * is the first fault-in of the page (m->vmp_pmapped == FALSE) + * so it must have come in as part of + * a cluster... account 1 pagein against it + */ + if (object->internal) + *type_of_fault = DBG_PAGEIND_FAULT; + else + *type_of_fault = DBG_PAGEINV_FAULT; - current_task()->pageins++; + VM_PAGE_COUNT_AS_PAGEIN(m); + } + VM_PAGE_CONSUME_CLUSTERED(m); } - VM_PAGE_CONSUME_CLUSTERED(m); - - } else if (cache_attr != VM_WIMG_DEFAULT) - pmap_sync_page_attributes_phys(m->phys_page); + } if (*type_of_fault != DBG_COW_FAULT) { DTRACE_VM2(as_fault, int, 1, (uint64_t *), NULL); @@ -2259,72 +2559,129 @@ vm_fault_enter(vm_page_t m, } /* Validate code signature if necessary. */ - if (VM_FAULT_NEED_CS_VALIDATION(pmap, m)) { - vm_object_lock_assert_exclusive(m->object); + if (!cs_bypass && + VM_FAULT_NEED_CS_VALIDATION(pmap, m, object)) { + vm_object_lock_assert_exclusive(object); - if (m->cs_validated) { + if (m->vmp_cs_validated) { vm_cs_revalidates++; } - /* VM map is locked, so 1 ref will remain on VM object - + /* VM map is locked, so 1 ref will remain on VM object - * so no harm if vm_page_validate_cs drops the object lock */ + +#if PMAP_CS + if (fault_info->pmap_cs_associated && + pmap_cs_enforced(pmap) && + !m->vmp_cs_validated && + !m->vmp_cs_tainted && + !m->vmp_cs_nx && + (prot & VM_PROT_EXECUTE) && + (caller_prot & VM_PROT_EXECUTE)) { + /* + * With pmap_cs, the pmap layer will validate the + * code signature for any executable pmap mapping. + * No need for us to validate this page too: + * in pmap_cs we trust... + */ + vm_cs_defer_to_pmap_cs++; + } else { + vm_cs_defer_to_pmap_cs_not++; + vm_page_validate_cs(m); + } +#else /* PMAP_CS */ vm_page_validate_cs(m); +#endif /* PMAP_CS */ } -#define page_immutable(m,prot) ((m)->cs_validated /*&& ((prot) & VM_PROT_EXECUTE)*/) +#define page_immutable(m,prot) ((m)->vmp_cs_validated /*&& ((prot) & VM_PROT_EXECUTE)*/) +#define page_nx(m) ((m)->vmp_cs_nx) map_is_switched = ((pmap != vm_map_pmap(current_task()->map)) && (pmap == vm_map_pmap(current_thread()->map))); map_is_switch_protected = current_thread()->map->switch_protect; - + /* If the map is switched, and is switch-protected, we must protect - * some pages from being write-faulted: immutable pages because by + * some pages from being write-faulted: immutable pages because by * definition they may not be written, and executable pages because that * would provide a way to inject unsigned code. * If the page is immutable, we can simply return. However, we can't * immediately determine whether a page is executable anywhere. But, * we can disconnect it everywhere and remove the executable protection - * from the current map. We do that below right before we do the + * from the current map. We do that below right before we do the * PMAP_ENTER. */ - if(!cs_enforcement_disable && map_is_switched && - map_is_switch_protected && page_immutable(m, prot) && + cs_enforcement_enabled = cs_process_enforcement(NULL); + + if(cs_enforcement_enabled && map_is_switched && + map_is_switch_protected && page_immutable(m, prot) && (prot & VM_PROT_WRITE)) { return KERN_CODESIGN_ERROR; } + if (cs_enforcement_enabled && page_nx(m) && (prot & VM_PROT_EXECUTE)) { + if (cs_debug) + printf("page marked to be NX, not letting it be mapped EXEC\n"); + return KERN_CODESIGN_ERROR; + } + /* A page could be tainted, or pose a risk of being tainted later. * Check whether the receiving process wants it, and make it feel * the consequences (that hapens in cs_invalid_page()). - * For CS Enforcement, two other conditions will - * cause that page to be tainted as well: + * For CS Enforcement, two other conditions will + * cause that page to be tainted as well: * - pmapping an unsigned page executable - this means unsigned code; * - writeable mapping of a validated page - the content of that page * can be changed without the kernel noticing, therefore unsigned * code can be created */ - if (m->cs_tainted || - ( !cs_enforcement_disable && - (/* The page is unsigned and wants to be executable */ - (!m->cs_validated && (prot & VM_PROT_EXECUTE)) || - /* The page should be immutable, but is in danger of being modified - * This is the case where we want policy from the code directory - - * is the page immutable or not? For now we have to assume that - * code pages will be immutable, data pages not. - * We'll assume a page is a code page if it has a code directory - * and we fault for execution. - * That is good enough since if we faulted the code page for - * writing in another map before, it is wpmapped; if we fault - * it for writing in this map later it will also be faulted for executing - * at the same time; and if we fault for writing in another map - * later, we will disconnect it from this pmap so we'll notice - * the change. - */ - (page_immutable(m, prot) && ((prot & VM_PROT_WRITE) || m->wpmapped)) - )) - ) - { + if (cs_bypass) { + /* code-signing is bypassed */ + cs_violation = FALSE; + } else if (m->vmp_cs_tainted) { + /* tainted page */ + cs_violation = TRUE; + } else if (!cs_enforcement_enabled) { + /* no further code-signing enforcement */ + cs_violation = FALSE; + } else if (page_immutable(m, prot) && + ((prot & VM_PROT_WRITE) || + m->vmp_wpmapped)) { + /* + * The page should be immutable, but is in danger of being + * modified. + * This is the case where we want policy from the code + * directory - is the page immutable or not? For now we have + * to assume that code pages will be immutable, data pages not. + * We'll assume a page is a code page if it has a code directory + * and we fault for execution. + * That is good enough since if we faulted the code page for + * writing in another map before, it is wpmapped; if we fault + * it for writing in this map later it will also be faulted for + * executing at the same time; and if we fault for writing in + * another map later, we will disconnect it from this pmap so + * we'll notice the change. + */ + cs_violation = TRUE; + } else if (!m->vmp_cs_validated && + (prot & VM_PROT_EXECUTE) +#if PMAP_CS + /* + * Executable pages will be validated by pmap_cs; + * in pmap_cs we trust... + * If pmap_cs is turned off, this is a code-signing + * violation. + */ + && ! (pmap_cs_enforced(pmap)) +#endif /* PMAP_CS */ + ) { + cs_violation = TRUE; + } else { + cs_violation = FALSE; + } + + if (cs_violation) { /* We will have a tainted page. Have to handle the special case * of a switched map now. If the map is not switched, standard * procedure applies - call cs_invalid_page(). @@ -2332,182 +2689,674 @@ vm_fault_enter(vm_page_t m, * There is no point in invalidating the switching process since * it will not be executing from the map. So we don't call * cs_invalid_page() in that case. */ - boolean_t reject_page; - if(map_is_switched) { + boolean_t reject_page, cs_killed; + if(map_is_switched) { assert(pmap==vm_map_pmap(current_thread()->map)); assert(!(prot & VM_PROT_WRITE) || (map_is_switch_protected == FALSE)); reject_page = FALSE; } else { - reject_page = cs_invalid_page((addr64_t) vaddr); + if (cs_debug > 5) + printf("vm_fault: signed: %s validate: %s tainted: %s wpmapped: %s prot: 0x%x\n", + object->code_signed ? "yes" : "no", + m->vmp_cs_validated ? "yes" : "no", + m->vmp_cs_tainted ? "yes" : "no", + m->vmp_wpmapped ? "yes" : "no", + (int)prot); + reject_page = cs_invalid_page((addr64_t) vaddr, &cs_killed); } - + if (reject_page) { - /* reject the tainted page: abort the page fault */ + /* reject the invalid page: abort the page fault */ + int pid; + const char *procname; + task_t task; + vm_object_t file_object, shadow; + vm_object_offset_t file_offset; + char *pathname, *filename; + vm_size_t pathname_len, filename_len; + boolean_t truncated_path; +#define __PATH_MAX 1024 + struct timespec mtime, cs_mtime; + int shadow_depth; + os_reason_t codesigning_exit_reason = OS_REASON_NULL; + kr = KERN_CODESIGN_ERROR; cs_enter_tainted_rejected++; + + /* get process name and pid */ + procname = "?"; + task = current_task(); + pid = proc_selfpid(); + if (task->bsd_info != NULL) + procname = proc_name_address(task->bsd_info); + + /* get file's VM object */ + file_object = object; + file_offset = m->vmp_offset; + for (shadow = file_object->shadow, + shadow_depth = 0; + shadow != VM_OBJECT_NULL; + shadow = file_object->shadow, + shadow_depth++) { + vm_object_lock_shared(shadow); + if (file_object != object) { + vm_object_unlock(file_object); + } + file_offset += file_object->vo_shadow_offset; + file_object = shadow; + } + + mtime.tv_sec = 0; + mtime.tv_nsec = 0; + cs_mtime.tv_sec = 0; + cs_mtime.tv_nsec = 0; + + /* get file's pathname and/or filename */ + pathname = NULL; + filename = NULL; + pathname_len = 0; + filename_len = 0; + truncated_path = FALSE; + /* no pager -> no file -> no pathname, use "" in that case */ + if (file_object->pager != NULL) { + pathname = (char *)kalloc(__PATH_MAX * 2); + if (pathname) { + pathname[0] = '\0'; + pathname_len = __PATH_MAX; + filename = pathname + pathname_len; + filename_len = __PATH_MAX; + } + vnode_pager_get_object_name(file_object->pager, + pathname, + pathname_len, + filename, + filename_len, + &truncated_path); + if (pathname) { + /* safety first... */ + pathname[__PATH_MAX-1] = '\0'; + filename[__PATH_MAX-1] = '\0'; + } + vnode_pager_get_object_mtime(file_object->pager, + &mtime, + &cs_mtime); + } + printf("CODE SIGNING: process %d[%s]: " + "rejecting invalid page at address 0x%llx " + "from offset 0x%llx in file \"%s%s%s\" " + "(cs_mtime:%lu.%ld %s mtime:%lu.%ld) " + "(signed:%d validated:%d tainted:%d nx:%d " + "wpmapped:%d dirty:%d depth:%d)\n", + pid, procname, (addr64_t) vaddr, + file_offset, + (pathname ? pathname : ""), + (truncated_path ? "/.../" : ""), + (truncated_path ? filename : ""), + cs_mtime.tv_sec, cs_mtime.tv_nsec, + ((cs_mtime.tv_sec == mtime.tv_sec && + cs_mtime.tv_nsec == mtime.tv_nsec) + ? "==" + : "!="), + mtime.tv_sec, mtime.tv_nsec, + object->code_signed, + m->vmp_cs_validated, + m->vmp_cs_tainted, + m->vmp_cs_nx, + m->vmp_wpmapped, + m->vmp_dirty, + shadow_depth); + + /* + * We currently only generate an exit reason if cs_invalid_page directly killed a process. If cs_invalid_page + * did not kill the process (more the case on desktop), vm_fault_enter will not satisfy the fault and whether the + * process dies is dependent on whether there is a signal handler registered for SIGSEGV and how that handler + * will deal with the segmentation fault. + */ + if (cs_killed) { + KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_PROC, BSD_PROC_EXITREASON_CREATE) | DBG_FUNC_NONE, + pid, OS_REASON_CODESIGNING, CODESIGNING_EXIT_REASON_INVALID_PAGE, 0, 0); + + codesigning_exit_reason = os_reason_create(OS_REASON_CODESIGNING, CODESIGNING_EXIT_REASON_INVALID_PAGE); + if (codesigning_exit_reason == NULL) { + printf("vm_fault_enter: failed to allocate codesigning exit reason\n"); + } else { + mach_vm_address_t data_addr = 0; + struct codesigning_exit_reason_info *ceri = NULL; + uint32_t reason_buffer_size_estimate = kcdata_estimate_required_buffer_size(1, sizeof(*ceri)); + + if (os_reason_alloc_buffer_noblock(codesigning_exit_reason, reason_buffer_size_estimate)) { + printf("vm_fault_enter: failed to allocate buffer for codesigning exit reason\n"); + } else { + if (KERN_SUCCESS == kcdata_get_memory_addr(&codesigning_exit_reason->osr_kcd_descriptor, + EXIT_REASON_CODESIGNING_INFO, sizeof(*ceri), &data_addr)) { + ceri = (struct codesigning_exit_reason_info *)data_addr; + static_assert(__PATH_MAX == sizeof(ceri->ceri_pathname)); + + ceri->ceri_virt_addr = vaddr; + ceri->ceri_file_offset = file_offset; + if (pathname) + strncpy((char *)&ceri->ceri_pathname, pathname, sizeof(ceri->ceri_pathname)); + else + ceri->ceri_pathname[0] = '\0'; + if (filename) + strncpy((char *)&ceri->ceri_filename, filename, sizeof(ceri->ceri_filename)); + else + ceri->ceri_filename[0] = '\0'; + ceri->ceri_path_truncated = (truncated_path); + ceri->ceri_codesig_modtime_secs = cs_mtime.tv_sec; + ceri->ceri_codesig_modtime_nsecs = cs_mtime.tv_nsec; + ceri->ceri_page_modtime_secs = mtime.tv_sec; + ceri->ceri_page_modtime_nsecs = mtime.tv_nsec; + ceri->ceri_object_codesigned = (object->code_signed); + ceri->ceri_page_codesig_validated = (m->vmp_cs_validated); + ceri->ceri_page_codesig_tainted = (m->vmp_cs_tainted); + ceri->ceri_page_codesig_nx = (m->vmp_cs_nx); + ceri->ceri_page_wpmapped = (m->vmp_wpmapped); + ceri->ceri_page_slid = 0; + ceri->ceri_page_dirty = (m->vmp_dirty); + ceri->ceri_page_shadow_depth = shadow_depth; + } else { +#if DEBUG || DEVELOPMENT + panic("vm_fault_enter: failed to allocate kcdata for codesigning exit reason"); +#else + printf("vm_fault_enter: failed to allocate kcdata for codesigning exit reason\n"); +#endif /* DEBUG || DEVELOPMENT */ + /* Free the buffer */ + os_reason_alloc_buffer_noblock(codesigning_exit_reason, 0); + } + } + } + + set_thread_exit_reason(current_thread(), codesigning_exit_reason, FALSE); + } + if (panic_on_cs_killed && + object->object_is_shared_cache) { + panic("CODE SIGNING: process %d[%s]: " + "rejecting invalid page at address 0x%llx " + "from offset 0x%llx in file \"%s%s%s\" " + "(cs_mtime:%lu.%ld %s mtime:%lu.%ld) " + "(signed:%d validated:%d tainted:%d nx:%d" + "wpmapped:%d dirty:%d depth:%d)\n", + pid, procname, (addr64_t) vaddr, + file_offset, + (pathname ? pathname : ""), + (truncated_path ? "/.../" : ""), + (truncated_path ? filename : ""), + cs_mtime.tv_sec, cs_mtime.tv_nsec, + ((cs_mtime.tv_sec == mtime.tv_sec && + cs_mtime.tv_nsec == mtime.tv_nsec) + ? "==" + : "!="), + mtime.tv_sec, mtime.tv_nsec, + object->code_signed, + m->vmp_cs_validated, + m->vmp_cs_tainted, + m->vmp_cs_nx, + m->vmp_wpmapped, + m->vmp_dirty, + shadow_depth); + } + + if (file_object != object) { + vm_object_unlock(file_object); + } + if (pathname_len != 0) { + kfree(pathname, __PATH_MAX * 2); + pathname = NULL; + filename = NULL; + } } else { - /* proceed with the tainted page */ + /* proceed with the invalid page */ kr = KERN_SUCCESS; - /* Page might have been tainted before or not; now it - * definitively is. If the page wasn't tainted, we must - * disconnect it from all pmaps later. */ - must_disconnect = !m->cs_tainted; - m->cs_tainted = TRUE; + if (!m->vmp_cs_validated && + !object->code_signed) { + /* + * This page has not been (fully) validated but + * does not belong to a code-signed object + * so it should not be forcefully considered + * as tainted. + * We're just concerned about it here because + * we've been asked to "execute" it but that + * does not mean that it should cause other + * accesses to fail. + * This happens when a debugger sets a + * breakpoint and we then execute code in + * that page. Marking the page as "tainted" + * would cause any inspection tool ("leaks", + * "vmmap", "CrashReporter", ...) to get killed + * due to code-signing violation on that page, + * even though they're just reading it and not + * executing from it. + */ + } else { + /* + * Page might have been tainted before or not; + * now it definitively is. If the page wasn't + * tainted, we must disconnect it from all + * pmaps later, to force existing mappings + * through that code path for re-consideration + * of the validity of that page. + */ + must_disconnect = !m->vmp_cs_tainted; + m->vmp_cs_tainted = TRUE; + } cs_enter_tainted_accepted++; } - if (cs_debug || kr != KERN_SUCCESS) { - printf("CODESIGNING: vm_fault_enter(0x%llx): " - "page %p obj %p off 0x%llx *** INVALID PAGE ***\n", - (long long)vaddr, m, m->object, m->offset); + if (kr != KERN_SUCCESS) { + if (cs_debug) { + printf("CODESIGNING: vm_fault_enter(0x%llx): " + "*** INVALID PAGE ***\n", + (long long)vaddr); + } +#if !SECURE_KERNEL + if (cs_enforcement_panic) { + panic("CODESIGNING: panicking on invalid page\n"); + } +#endif } - + } else { /* proceed with the valid page */ kr = KERN_SUCCESS; } - /* If we have a KERN_SUCCESS from the previous checks, we either have - * a good page, or a tainted page that has been accepted by the process. - * In both cases the page will be entered into the pmap. - * If the page is writeable, we need to disconnect it from other pmaps - * now so those processes can take note. - */ - if (kr == KERN_SUCCESS) { - /* - * NOTE: we may only hold the vm_object lock SHARED - * at this point, but the update of pmapped is ok - * since this is the ONLY bit updated behind the SHARED - * lock... however, we need to figure out how to do an atomic - * update on a bit field to make this less fragile... right - * now I don't know how to coerce 'C' to give me the offset info - * that's needed for an AtomicCompareAndSwap - */ - m->pmapped = TRUE; - if (prot & VM_PROT_WRITE) { - vm_object_lock_assert_exclusive(m->object); - m->wpmapped = TRUE; - if(must_disconnect) { - /* We can only get here - * because of the CSE logic */ - assert(cs_enforcement_disable == FALSE); - pmap_disconnect(m->phys_page); - /* If we are faulting for a write, we can clear - * the execute bit - that will ensure the page is - * checked again before being executable, which - * protects against a map switch. - * This only happens the first time the page - * gets tainted, so we won't get stuck here - * to make an already writeable page executable. */ - prot &= ~VM_PROT_EXECUTE; - } - } - PMAP_ENTER(pmap, vaddr, m, prot, cache_attr, wired); - } + boolean_t page_queues_locked = FALSE; +#define __VM_PAGE_LOCKSPIN_QUEUES_IF_NEEDED() \ +MACRO_BEGIN \ + if (! page_queues_locked) { \ + page_queues_locked = TRUE; \ + vm_page_lockspin_queues(); \ + } \ +MACRO_END +#define __VM_PAGE_UNLOCK_QUEUES_IF_NEEDED() \ +MACRO_BEGIN \ + if (page_queues_locked) { \ + page_queues_locked = FALSE; \ + vm_page_unlock_queues(); \ + } \ +MACRO_END /* * Hold queues lock to manipulate * the page queues. Change wiring * case is obvious. */ - if (change_wiring) { - vm_page_lockspin_queues(); + assert((m->vmp_q_state == VM_PAGE_USED_BY_COMPRESSOR) || object != compressor_object); + +#if CONFIG_BACKGROUND_QUEUE + vm_page_update_background_state(m); +#endif + if (m->vmp_q_state == VM_PAGE_USED_BY_COMPRESSOR) { + /* + * Compressor pages are neither wired + * nor pageable and should never change. + */ + assert(object == compressor_object); + } else if (change_wiring) { + __VM_PAGE_LOCKSPIN_QUEUES_IF_NEEDED(); if (wired) { if (kr == KERN_SUCCESS) { - vm_page_wire(m); + vm_page_wire(m, wire_tag, TRUE); } } else { - vm_page_unwire(m); + vm_page_unwire(m, TRUE); } - vm_page_unlock_queues(); + /* we keep the page queues lock, if we need it later */ } else { + if (object->internal == TRUE) { + /* + * don't allow anonymous pages on + * the speculative queues + */ + no_cache = FALSE; + } if (kr != KERN_SUCCESS) { - vm_page_lockspin_queues(); + __VM_PAGE_LOCKSPIN_QUEUES_IF_NEEDED(); vm_page_deactivate(m); - vm_page_unlock_queues(); - } else { - if (((!m->active && !m->inactive) || no_cache) && !VM_PAGE_WIRED(m) && !m->throttled) { + /* we keep the page queues lock, if we need it later */ + } else if (((m->vmp_q_state == VM_PAGE_NOT_ON_Q) || + (m->vmp_q_state == VM_PAGE_ON_SPECULATIVE_Q) || + (m->vmp_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) || + ((m->vmp_q_state != VM_PAGE_ON_THROTTLED_Q) && no_cache)) && + !VM_PAGE_WIRED(m)) { - if ( vm_page_local_q && !no_cache && (*type_of_fault == DBG_COW_FAULT || *type_of_fault == DBG_ZERO_FILL_FAULT) ) { - struct vpl *lq; - uint32_t lid; + if (vm_page_local_q && + (*type_of_fault == DBG_COW_FAULT || + *type_of_fault == DBG_ZERO_FILL_FAULT) ) { + struct vpl *lq; + uint32_t lid; + + assert(m->vmp_q_state == VM_PAGE_NOT_ON_Q); + + __VM_PAGE_UNLOCK_QUEUES_IF_NEEDED(); + vm_object_lock_assert_exclusive(object); + + /* + * we got a local queue to stuff this + * new page on... + * its safe to manipulate local and + * local_id at this point since we're + * behind an exclusive object lock and + * the page is not on any global queue. + * + * we'll use the current cpu number to + * select the queue note that we don't + * need to disable preemption... we're + * going to be behind the local queue's + * lock to do the real work + */ + lid = cpu_number(); + lq = &vm_page_local_q[lid].vpl_un.vpl; + + VPL_LOCK(&lq->vpl_lock); + + vm_page_check_pageable_safe(m); + vm_page_queue_enter(&lq->vpl_queue, m, + vm_page_t, vmp_pageq); + m->vmp_q_state = VM_PAGE_ON_ACTIVE_LOCAL_Q; + m->vmp_local_id = lid; + lq->vpl_count++; + + if (object->internal) + lq->vpl_internal_count++; + else + lq->vpl_external_count++; + + VPL_UNLOCK(&lq->vpl_lock); + + if (lq->vpl_count > vm_page_local_q_soft_limit) + { /* - * we got a local queue to stuff this new page on... - * its safe to manipulate local and local_id at this point - * since we're behind an exclusive object lock and the - * page is not on any global queue. - * - * we'll use the current cpu number to select the queue - * note that we don't need to disable preemption... we're - * going to behind the local queue's lock to do the real - * work + * we're beyond the soft limit + * for the local queue + * vm_page_reactivate_local will + * 'try' to take the global page + * queue lock... if it can't + * that's ok... we'll let the + * queue continue to grow up + * to the hard limit... at that + * point we'll wait for the + * lock... once we've got the + * lock, we'll transfer all of + * the pages from the local + * queue to the global active + * queue */ - lid = cpu_number(); + vm_page_reactivate_local(lid, FALSE, FALSE); + } + } else { - lq = &vm_page_local_q[lid].vpl_un.vpl; + __VM_PAGE_LOCKSPIN_QUEUES_IF_NEEDED(); - VPL_LOCK(&lq->vpl_lock); + /* + * test again now that we hold the + * page queue lock + */ + if (!VM_PAGE_WIRED(m)) { + if (m->vmp_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) { + vm_page_queues_remove(m, FALSE); - queue_enter(&lq->vpl_queue, m, vm_page_t, pageq); - m->local = TRUE; - m->local_id = lid; - lq->vpl_count++; - - VPL_UNLOCK(&lq->vpl_lock); + VM_PAGEOUT_DEBUG(vm_pageout_cleaned_reactivated, 1); + VM_PAGEOUT_DEBUG(vm_pageout_cleaned_fault_reactivated, 1); + } - if (lq->vpl_count > vm_page_local_q_soft_limit) { + if ( !VM_PAGE_ACTIVE_OR_INACTIVE(m) || + no_cache) { /* - * we're beyond the soft limit for the local queue - * vm_page_reactivate_local will 'try' to take - * the global page queue lock... if it can't that's - * ok... we'll let the queue continue to grow up - * to the hard limit... at that point we'll wait - * for the lock... once we've got the lock, we'll - * transfer all of the pages from the local queue - * to the global active queue + * If this is a no_cache mapping + * and the page has never been + * mapped before or was + * previously a no_cache page, + * then we want to leave pages + * in the speculative state so + * that they can be readily + * recycled if free memory runs + * low. Otherwise the page is + * activated as normal. */ - vm_page_reactivate_local(lid, FALSE, FALSE); + + if (no_cache && + (!previously_pmapped || + m->vmp_no_cache)) { + m->vmp_no_cache = TRUE; + + if (m->vmp_q_state != VM_PAGE_ON_SPECULATIVE_Q) + vm_page_speculate(m, FALSE); + + } else if ( !VM_PAGE_ACTIVE_OR_INACTIVE(m)) { + vm_page_activate(m); + } } - return kr; } + /* we keep the page queues lock, if we need it later */ + } + } + } + /* we're done with the page queues lock, if we ever took it */ + __VM_PAGE_UNLOCK_QUEUES_IF_NEEDED(); - vm_page_lockspin_queues(); - /* - * test again now that we hold the page queue lock - */ - if (((!m->active && !m->inactive) || no_cache) && !VM_PAGE_WIRED(m)) { + /* If we have a KERN_SUCCESS from the previous checks, we either have + * a good page, or a tainted page that has been accepted by the process. + * In both cases the page will be entered into the pmap. + * If the page is writeable, we need to disconnect it from other pmaps + * now so those processes can take note. + */ + if (kr == KERN_SUCCESS) { + /* + * NOTE: we may only hold the vm_object lock SHARED + * at this point, so we need the phys_page lock to + * properly serialize updating the pmapped and + * xpmapped bits + */ + if ((prot & VM_PROT_EXECUTE) && !m->vmp_xpmapped) { + ppnum_t phys_page = VM_PAGE_GET_PHYS_PAGE(m); + + pmap_lock_phys_page(phys_page); + /* + * go ahead and take the opportunity + * to set 'pmapped' here so that we don't + * need to grab this lock a 2nd time + * just below + */ + m->vmp_pmapped = TRUE; + + if (!m->vmp_xpmapped) { + + m->vmp_xpmapped = TRUE; + + pmap_unlock_phys_page(phys_page); + + if (!object->internal) + OSAddAtomic(1, &vm_page_xpmapped_external_count); + +#if defined(__arm__) || defined(__arm64__) + pmap_sync_page_data_phys(phys_page); +#else + if (object->internal && + object->pager != NULL) { /* - * If this is a no_cache mapping and the page has never been - * mapped before or was previously a no_cache page, then we - * want to leave pages in the speculative state so that they - * can be readily recycled if free memory runs low. Otherwise - * the page is activated as normal. + * This page could have been + * uncompressed by the + * compressor pager and its + * contents might be only in + * the data cache. + * Since it's being mapped for + * "execute" for the fist time, + * make sure the icache is in + * sync. */ + assert(VM_CONFIG_COMPRESSOR_IS_PRESENT); + pmap_sync_page_data_phys(phys_page); + } +#endif + } else + pmap_unlock_phys_page(phys_page); + } else { + if (m->vmp_pmapped == FALSE) { + ppnum_t phys_page = VM_PAGE_GET_PHYS_PAGE(m); - if (no_cache && (!previously_pmapped || m->no_cache)) { - m->no_cache = TRUE; + pmap_lock_phys_page(phys_page); + m->vmp_pmapped = TRUE; + pmap_unlock_phys_page(phys_page); + } + } + + if (fault_type & VM_PROT_WRITE) { - if (m->active || m->inactive) - VM_PAGE_QUEUES_REMOVE(m); + if (m->vmp_wpmapped == FALSE) { + vm_object_lock_assert_exclusive(object); + if (!object->internal && object->pager) { + task_update_logical_writes(current_task(), PAGE_SIZE, TASK_WRITE_DEFERRED, vnode_pager_lookup_vnode(object->pager)); + } + m->vmp_wpmapped = TRUE; + } + if (must_disconnect) { + /* + * We can only get here + * because of the CSE logic + */ + assert(cs_enforcement_enabled); + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m)); + /* + * If we are faulting for a write, we can clear + * the execute bit - that will ensure the page is + * checked again before being executable, which + * protects against a map switch. + * This only happens the first time the page + * gets tainted, so we won't get stuck here + * to make an already writeable page executable. + */ + if (!cs_bypass){ + assert(!pmap_has_prot_policy(prot)); + prot &= ~VM_PROT_EXECUTE; + } + } + } + assert(VM_PAGE_OBJECT(m) == object); + +#if VM_OBJECT_ACCESS_TRACKING + if (object->access_tracking) { + DTRACE_VM2(access_tracking, vm_map_offset_t, vaddr, int, fault_type); + if (fault_type & VM_PROT_WRITE) { + object->access_tracking_writes++; + vm_object_access_tracking_writes++; + } else { + object->access_tracking_reads++; + vm_object_access_tracking_reads++; + } + } +#endif /* VM_OBJECT_ACCESS_TRACKING */ - if (!m->speculative) - vm_page_speculate(m, TRUE); +#if PMAP_CS + /* + * If CS enforcement is on, we don't ask for an executable page if the + * fault does not call for execution, because that can fail in + * situations where the caller only actually wanted read access. + * However, it may be better to instead retry without execute on + * failure, or pass a flag into pmap_enter to do the right thing. + */ + // TODO: maybe do something better than masking out VM_PROT_EXECUTE on non-execute faults + if (pmap_cs_enforced(pmap) && !(caller_prot & VM_PROT_EXECUTE)) { + prot &= ~VM_PROT_EXECUTE; + } +#endif - } else if (!m->active && !m->inactive) - vm_page_activate(m); + /* Prevent a deadlock by not + * holding the object lock if we need to wait for a page in + * pmap_enter() - */ + PMAP_ENTER_OPTIONS(pmap, vaddr, m, prot, fault_type, 0, + wired, + pmap_options | PMAP_OPTIONS_NOWAIT, + pe_result); +#if __x86_64__ + if (pe_result == KERN_INVALID_ARGUMENT && + pmap == PMAP_NULL && + wired) { + /* + * Wiring a page in a pmap-less VM map: + * VMware's "vmmon" kernel extension does this + * to grab pages. + * Let it proceed even though the PMAP_ENTER() failed. + */ + pe_result = KERN_SUCCESS; + } +#endif /* __x86_64__ */ - } + if(pe_result == KERN_RESOURCE_SHORTAGE) { - vm_page_unlock_queues(); + if (need_retry) { + /* + * this will be non-null in the case where we hold the lock + * on the top-object in this chain... we can't just drop + * the lock on the object we're inserting the page into + * and recall the PMAP_ENTER since we can still cause + * a deadlock if one of the critical paths tries to + * acquire the lock on the top-object and we're blocked + * in PMAP_ENTER waiting for memory... our only recourse + * is to deal with it at a higher level where we can + * drop both locks. + */ + *need_retry = TRUE; + vm_pmap_enter_retried++; + goto after_the_pmap_enter; } + /* The nonblocking version of pmap_enter did not succeed. + * and we don't need to drop other locks and retry + * at the level above us, so + * use the blocking version instead. Requires marking + * the page busy and unlocking the object */ + boolean_t was_busy = m->vmp_busy; + + vm_object_lock_assert_exclusive(object); + + m->vmp_busy = TRUE; + vm_object_unlock(object); + + PMAP_ENTER_OPTIONS(pmap, vaddr, m, prot, fault_type, + 0, wired, + pmap_options, pe_result); + + assert(VM_PAGE_OBJECT(m) == object); + + /* Take the object lock again. */ + vm_object_lock(object); + + /* If the page was busy, someone else will wake it up. + * Otherwise, we have to do it now. */ + assert(m->vmp_busy); + if(!was_busy) { + PAGE_WAKEUP_DONE(m); + } + vm_pmap_enter_blocked++; } + + kr = pe_result; } + +after_the_pmap_enter: return kr; } +void +vm_pre_fault(vm_map_offset_t vaddr) +{ + if (pmap_find_phys(current_map()->pmap, vaddr) == 0) { + + vm_fault(current_map(), /* map */ + vaddr, /* vaddr */ + VM_PROT_READ, /* fault_type */ + FALSE, /* change_wiring */ + VM_KERN_MEMORY_NONE, /* tag - not wiring */ + THREAD_UNINT, /* interruptible */ + NULL, /* caller_pmap */ + 0 /* caller_pmap_addr */); + } +} + /* * Routine: vm_fault @@ -2524,19 +3373,54 @@ vm_fault_enter(vm_page_t m, */ extern int _map_enter_debug; +extern uint64_t get_current_unique_pid(void); unsigned long vm_fault_collapse_total = 0; unsigned long vm_fault_collapse_skipped = 0; + +kern_return_t +vm_fault_external( + vm_map_t map, + vm_map_offset_t vaddr, + vm_prot_t fault_type, + boolean_t change_wiring, + int interruptible, + pmap_t caller_pmap, + vm_map_offset_t caller_pmap_addr) +{ + return vm_fault_internal(map, vaddr, fault_type, change_wiring, vm_tag_bt(), + interruptible, caller_pmap, caller_pmap_addr, + NULL); +} + kern_return_t vm_fault( vm_map_t map, vm_map_offset_t vaddr, vm_prot_t fault_type, boolean_t change_wiring, + vm_tag_t wire_tag, /* if wiring must pass tag != VM_KERN_MEMORY_NONE */ int interruptible, pmap_t caller_pmap, vm_map_offset_t caller_pmap_addr) +{ + return vm_fault_internal(map, vaddr, fault_type, change_wiring, wire_tag, + interruptible, caller_pmap, caller_pmap_addr, + NULL); +} + +kern_return_t +vm_fault_internal( + vm_map_t map, + vm_map_offset_t vaddr, + vm_prot_t caller_prot, + boolean_t change_wiring, + vm_tag_t wire_tag, /* if wiring must pass tag != VM_KERN_MEMORY_NONE */ + int interruptible, + pmap_t caller_pmap, + vm_map_offset_t caller_pmap_addr, + ppnum_t *physpage_p) { vm_map_version_t version; /* Map version for verificiation */ boolean_t wired; /* Should mapping be wired down? */ @@ -2551,42 +3435,80 @@ vm_fault( vm_page_t m; /* Fast access to result_page */ kern_return_t error_code; vm_object_t cur_object; + vm_object_t m_object = NULL; vm_object_offset_t cur_offset; vm_page_t cur_m; vm_object_t new_object; int type_of_fault; pmap_t pmap; - boolean_t interruptible_state; + wait_interrupt_t interruptible_state; vm_map_t real_map = map; vm_map_t original_map = map; + boolean_t object_locks_dropped = FALSE; + vm_prot_t fault_type; vm_prot_t original_fault_type; - struct vm_object_fault_info fault_info; + struct vm_object_fault_info fault_info = {}; boolean_t need_collapse = FALSE; + boolean_t need_retry = FALSE; + boolean_t *need_retry_ptr = NULL; int object_lock_type = 0; int cur_object_lock_type; vm_object_t top_object = VM_OBJECT_NULL; + vm_object_t written_on_object = VM_OBJECT_NULL; + memory_object_t written_on_pager = NULL; + vm_object_offset_t written_on_offset = 0; + int throttle_delay; + int compressed_count_delta; + int grab_options; + vm_map_offset_t trace_vaddr; + vm_map_offset_t trace_real_vaddr; +#if DEVELOPMENT || DEBUG + vm_map_offset_t real_vaddr; + + real_vaddr = vaddr; +#endif /* DEVELOPMENT || DEBUG */ + trace_real_vaddr = vaddr; + vaddr = vm_map_trunc_page(vaddr, PAGE_MASK); + + if (map == kernel_map) { + trace_vaddr = VM_KERNEL_ADDRHIDE(vaddr); + trace_real_vaddr = VM_KERNEL_ADDRHIDE(trace_real_vaddr); + } else { + trace_vaddr = vaddr; + } - - KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, 2)) | DBG_FUNC_START, - (int)((uint64_t)vaddr >> 32), - (int)vaddr, - 0, + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + (MACHDBG_CODE(DBG_MACH_VM, 2)) | DBG_FUNC_START, + ((uint64_t)trace_vaddr >> 32), + trace_vaddr, + (map == kernel_map), 0, 0); if (get_preemption_level() != 0) { - KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, 2)) | DBG_FUNC_END, - (int)((uint64_t)vaddr >> 32), - (int)vaddr, + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + (MACHDBG_CODE(DBG_MACH_VM, 2)) | DBG_FUNC_END, + ((uint64_t)trace_vaddr >> 32), + trace_vaddr, KERN_FAILURE, 0, 0); return (KERN_FAILURE); } - + + thread_t cthread = current_thread(); + boolean_t rtfault = (cthread->sched_mode == TH_MODE_REALTIME); + uint64_t fstart = 0; + + if (rtfault) { + fstart = mach_continuous_time(); + } + interruptible_state = thread_interrupt_level(interruptible); + fault_type = (change_wiring ? VM_PROT_NONE : caller_prot); + VM_STAT_INCR(faults); current_task()->faults++; original_fault_type = fault_type; @@ -2598,7 +3520,17 @@ vm_fault( cur_object_lock_type = OBJECT_LOCK_SHARED; + if ((map == kernel_map) && (caller_prot & VM_PROT_WRITE)) { + if (compressor_map) { + if ((vaddr >= vm_map_min(compressor_map)) && (vaddr < vm_map_max(compressor_map))) { + panic("Write fault on compressor map, va: %p type: %u bounds: %p->%p", (void *) vaddr, caller_prot, (void *) vm_map_min(compressor_map), (void *) vm_map_max(compressor_map)); + + } + } + } RetryFault: + assert(written_on_object == VM_OBJECT_NULL); + /* * assume we will hit a page in the cache * otherwise, explicitly override with @@ -2627,6 +3559,9 @@ RetryFault: pmap = real_map->pmap; fault_info.interruptible = interruptible; fault_info.stealth = FALSE; + fault_info.io_sync = FALSE; + fault_info.mark_zf_absent = FALSE; + fault_info.batch_pmap_op = FALSE; /* * If the page is wired, we must fault for the current protection @@ -2690,6 +3625,24 @@ RetryFault: * */ +#if defined(__arm64__) + /* + * Fail if reading an execute-only page in a + * pmap that enforces execute-only protection. + */ + if (fault_type == VM_PROT_READ && + (prot & VM_PROT_EXECUTE) && + !(prot & VM_PROT_READ) && + pmap_enforces_execute_only(pmap)) { + vm_object_unlock(object); + vm_map_unlock_read(map); + if (real_map != map) { + vm_map_unlock(real_map); + } + kr = KERN_PROTECTION_FAILURE; + goto done; + } +#endif /* * If this page is to be inserted in a copy delay object @@ -2703,6 +3656,13 @@ RetryFault: cur_object = object; cur_offset = offset; + grab_options = 0; +#if CONFIG_SECLUDED_MEMORY + if (object->can_grab_secluded) { + grab_options |= VM_PAGE_GRAB_SECLUDED; + } +#endif /* CONFIG_SECLUDED_MEMORY */ + while (TRUE) { if (!cur_object->pager_created && cur_object->phys_contiguous) /* superpage */ @@ -2717,9 +3677,12 @@ RetryFault: } m = vm_page_lookup(cur_object, cur_offset); + m_object = NULL; if (m != VM_PAGE_NULL) { - if (m->busy) { + m_object = cur_object; + + if (m->vmp_busy) { wait_result_t result; /* @@ -2727,7 +3690,6 @@ RetryFault: * have object that 'm' belongs to locked exclusively */ if (object != cur_object) { - vm_object_unlock(object); if (cur_object_lock_type == OBJECT_LOCK_SHARED) { @@ -2736,11 +3698,13 @@ RetryFault: if (vm_object_lock_upgrade(cur_object) == FALSE) { /* * couldn't upgrade so go do a full retry - * immediately since we've already dropped - * the top object lock associated with this page - * and the current one got dropped due to the - * failed upgrade... the state is no longer valid + * immediately since we can no longer be + * certain about cur_object (since we + * don't hold a reference on it)... + * first drop the top object lock */ + vm_object_unlock(object); + vm_map_unlock_read(map); if (real_map != map) vm_map_unlock(real_map); @@ -2767,6 +3731,32 @@ RetryFault: continue; } } + if ((m->vmp_q_state == VM_PAGE_ON_PAGEOUT_Q) && m_object->internal) { + /* + * m->vmp_busy == TRUE and the object is locked exclusively + * if m->pageout_queue == TRUE after we acquire the + * queues lock, we are guaranteed that it is stable on + * the pageout queue and therefore reclaimable + * + * NOTE: this is only true for the internal pageout queue + * in the compressor world + */ + assert(VM_CONFIG_COMPRESSOR_IS_PRESENT); + + vm_page_lock_queues(); + + if (m->vmp_q_state == VM_PAGE_ON_PAGEOUT_Q) { + vm_pageout_throttle_up(m); + vm_page_unlock_queues(); + + PAGE_WAKEUP_DONE(m); + goto reclaimed_from_pageout; + } + vm_page_unlock_queues(); + } + if (object != cur_object) + vm_object_unlock(object); + vm_map_unlock_read(map); if (real_map != map) vm_map_unlock(real_map); @@ -2786,70 +3776,28 @@ RetryFault: kr = KERN_ABORTED; goto done; } - if (m->phys_page == vm_page_guard_addr) { - /* - * Guard page: let the slow path deal with it - */ - break; - } - if (m->unusual && (m->error || m->restart || m->private || m->absent)) { - /* - * Unusual case... let the slow path deal with it - */ - break; - } - if (VM_OBJECT_PURGEABLE_FAULT_ERROR(m->object)) { - if (object != cur_object) - vm_object_unlock(object); - vm_map_unlock_read(map); - if (real_map != map) - vm_map_unlock(real_map); - vm_object_unlock(cur_object); - kr = KERN_MEMORY_ERROR; - goto done; - } - - if (m->encrypted) { - /* - * ENCRYPTED SWAP: - * We've soft-faulted (because it's not in the page - * table) on an encrypted page. - * Keep the page "busy" so that no one messes with - * it during the decryption. - * Release the extra locks we're holding, keep only - * the page's VM object lock. - * - * in order to set 'busy' on 'm', we must - * have object that 'm' belongs to locked exclusively - */ - if (object != cur_object) { - vm_object_unlock(object); - +reclaimed_from_pageout: + if (m->vmp_laundry) { + if (object != cur_object) { if (cur_object_lock_type == OBJECT_LOCK_SHARED) { + cur_object_lock_type = OBJECT_LOCK_EXCLUSIVE; - cur_object_lock_type = OBJECT_LOCK_EXCLUSIVE; + vm_object_unlock(object); + vm_object_unlock(cur_object); - if (vm_object_lock_upgrade(cur_object) == FALSE) { - /* - * couldn't upgrade so go do a full retry - * immediately since we've already dropped - * the top object lock associated with this page - * and the current one got dropped due to the - * failed upgrade... the state is no longer valid - */ - vm_map_unlock_read(map); - if (real_map != map) - vm_map_unlock(real_map); + vm_map_unlock_read(map); + if (real_map != map) + vm_map_unlock(real_map); - goto RetryFault; - } + goto RetryFault; } + } else if (object_lock_type == OBJECT_LOCK_SHARED) { - object_lock_type = OBJECT_LOCK_EXCLUSIVE; + object_lock_type = OBJECT_LOCK_EXCLUSIVE; if (vm_object_lock_upgrade(object) == FALSE) { - /* + /* * couldn't upgrade, so explictly take the lock * exclusively and go relookup the page since we * will have dropped the object lock and @@ -2858,32 +3806,41 @@ RetryFault: * no need for a full retry since we're * at the top level of the object chain */ - vm_object_lock(object); + vm_object_lock(object); continue; } } - m->busy = TRUE; + vm_pageout_steal_laundry(m, FALSE); + } + if (VM_PAGE_GET_PHYS_PAGE(m) == vm_page_guard_addr) { + /* + * Guard page: let the slow path deal with it + */ + break; + } + if (m->vmp_unusual && (m->vmp_error || m->vmp_restart || m->vmp_private || m->vmp_absent)) { + /* + * Unusual case... let the slow path deal with it + */ + break; + } + if (VM_OBJECT_PURGEABLE_FAULT_ERROR(m_object)) { + if (object != cur_object) + vm_object_unlock(object); vm_map_unlock_read(map); - if (real_map != map) - vm_map_unlock(real_map); - - vm_page_decrypt(m, 0); - - assert(m->busy); - PAGE_WAKEUP_DONE(m); - + if (real_map != map) + vm_map_unlock(real_map); vm_object_unlock(cur_object); - /* - * Retry from the top, in case anything - * changed while we were decrypting... - */ - goto RetryFault; + kr = KERN_MEMORY_ERROR; + goto done; } - ASSERT_PAGE_DECRYPTED(m); + assert(m_object == VM_PAGE_OBJECT(m)); - if (VM_FAULT_NEED_CS_VALIDATION(map->pmap, m)) { + if (VM_FAULT_NEED_CS_VALIDATION(map->pmap, m, m_object) || + (physpage_p != NULL && (prot & VM_PROT_WRITE))) { +upgrade_for_validation: /* * We might need to validate this page * against its code signature, so we @@ -2931,26 +3888,25 @@ RetryFault: */ if (object == cur_object && object->copy == VM_OBJECT_NULL) { - if ((fault_type & VM_PROT_WRITE) == 0) { - /* - * This is not a "write" fault, so we - * might not have taken the object lock - * exclusively and we might not be able - * to update the "wpmapped" bit in - * vm_fault_enter(). - * Let's just grant read access to - * the page for now and we'll - * soft-fault again if we need write - * access later... - */ - prot &= ~VM_PROT_WRITE; - } + goto FastPmapEnter; } if ((fault_type & VM_PROT_WRITE) == 0) { - - prot &= ~VM_PROT_WRITE; + if (!pmap_has_prot_policy(prot)) { + prot &= ~VM_PROT_WRITE; + } else { + /* + * For a protection that the pmap cares + * about, we must hand over the full + * set of protections (so that the pmap + * layer can apply any desired policy). + * This means that cs_bypass must be + * set, as this can force us to pass + * RWX. + */ + assert(fault_info.cs_bypass); + } if (object != cur_object) { /* @@ -2980,37 +3936,71 @@ RetryFault: object_lock_type = cur_object_lock_type; } FastPmapEnter: + assert(m_object == VM_PAGE_OBJECT(m)); + /* * prepare for the pmap_enter... * object and map are both locked * m contains valid data - * object == m->object + * object == m->vmp_object * cur_object == NULL or it's been unlocked * no paging references on either object or cur_object */ -#if MACH_KDB - if (db_watchpoint_list && (fault_type & VM_PROT_WRITE) == 0) - prot &= ~VM_PROT_WRITE; -#endif + if (top_object != VM_OBJECT_NULL || object_lock_type != OBJECT_LOCK_EXCLUSIVE) + need_retry_ptr = &need_retry; + else + need_retry_ptr = NULL; + if (caller_pmap) { kr = vm_fault_enter(m, caller_pmap, caller_pmap_addr, prot, + caller_prot, wired, change_wiring, - fault_info.no_cache, + wire_tag, + &fault_info, + need_retry_ptr, &type_of_fault); } else { kr = vm_fault_enter(m, pmap, vaddr, prot, + caller_prot, wired, change_wiring, - fault_info.no_cache, + wire_tag, + &fault_info, + need_retry_ptr, &type_of_fault); } +#if DEVELOPMENT || DEBUG + { + int event_code = 0; + + if (m_object->internal) + event_code = (MACHDBG_CODE(DBG_MACH_WORKINGSET, VM_REAL_FAULT_ADDR_INTERNAL)); + else if (m_object->object_is_shared_cache) + event_code = (MACHDBG_CODE(DBG_MACH_WORKINGSET, VM_REAL_FAULT_ADDR_SHAREDCACHE)); + else + event_code = (MACHDBG_CODE(DBG_MACH_WORKINGSET, VM_REAL_FAULT_ADDR_EXTERNAL)); + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, event_code, trace_real_vaddr, (fault_info.user_tag << 16) | (caller_prot << 8) | type_of_fault, m->vmp_offset, get_current_unique_pid(), 0); + + DTRACE_VM6(real_fault, vm_map_offset_t, real_vaddr, vm_map_offset_t, m->vmp_offset, int, event_code, int, caller_prot, int, type_of_fault, int, fault_info.user_tag); + } +#endif + if (kr == KERN_SUCCESS && + physpage_p != NULL) { + /* for vm_map_wire_and_extract() */ + *physpage_p = VM_PAGE_GET_PHYS_PAGE(m); + if (prot & VM_PROT_WRITE) { + vm_object_lock_assert_exclusive(m_object); + m->vmp_dirty = TRUE; + } + } if (top_object != VM_OBJECT_NULL) { /* @@ -3029,28 +4019,54 @@ FastPmapEnter: if (need_collapse == TRUE) vm_object_collapse(object, offset, TRUE); - if (type_of_fault == DBG_PAGEIND_FAULT || type_of_fault == DBG_PAGEINV_FAULT || type_of_fault == DBG_CACHE_HIT_FAULT) { + if (need_retry == FALSE && + (type_of_fault == DBG_PAGEIND_FAULT || type_of_fault == DBG_PAGEINV_FAULT || type_of_fault == DBG_CACHE_HIT_FAULT)) { /* * evaluate access pattern and update state * vm_fault_deactivate_behind depends on the * state being up to date */ - vm_fault_is_sequential(object, cur_offset, fault_info.behavior); + vm_fault_is_sequential(m_object, cur_offset, fault_info.behavior); - vm_fault_deactivate_behind(object, cur_offset, fault_info.behavior); + vm_fault_deactivate_behind(m_object, cur_offset, fault_info.behavior); } /* * That's it, clean up and return. */ - if (m->busy) + if (m->vmp_busy) PAGE_WAKEUP_DONE(m); + if (need_retry == FALSE && !m_object->internal && (fault_type & VM_PROT_WRITE)) { + + vm_object_paging_begin(m_object); + + assert(written_on_object == VM_OBJECT_NULL); + written_on_object = m_object; + written_on_pager = m_object->pager; + written_on_offset = m_object->paging_offset + m->vmp_offset; + } vm_object_unlock(object); vm_map_unlock_read(map); if (real_map != map) vm_map_unlock(real_map); + if (need_retry == TRUE) { + /* + * vm_fault_enter couldn't complete the PMAP_ENTER... + * at this point we don't hold any locks so it's safe + * to ask the pmap layer to expand the page table to + * accommodate this mapping... once expanded, we'll + * re-drive the fault which should result in vm_fault_enter + * being able to successfully enter the mapping this time around + */ + (void)pmap_enter_options( + pmap, vaddr, 0, 0, 0, 0, 0, + PMAP_OPTIONS_NOENTER, NULL); + + need_retry = FALSE; + goto RetryFault; + } goto done; } /* @@ -3058,30 +4074,6 @@ FastPmapEnter: */ assert(object_lock_type == OBJECT_LOCK_EXCLUSIVE); - if (vm_page_throttled()) { - /* - * drop all of our locks... - * wait until the free queue is - * pumped back up and then - * redrive the fault - */ - if (object != cur_object) - vm_object_unlock(cur_object); - vm_object_unlock(object); - vm_map_unlock_read(map); - if (real_map != map) - vm_map_unlock(real_map); - - if (NEED_TO_HARD_THROTTLE_THIS_TASK()) - delay(HARD_THROTTLE_DELAY); - - if (!current_thread_aborted() && vm_page_wait((change_wiring) ? - THREAD_UNINT : - THREAD_ABORTSAFE)) - goto RetryFault; - kr = KERN_ABORTED; - goto done; - } /* * If objects match, then * object->copy must not be NULL (else control @@ -3096,24 +4088,34 @@ FastPmapEnter: */ break; } + /* * This is now a shadow based copy on write * fault -- it requires a copy up the shadow * chain. - * + */ + assert(m_object == VM_PAGE_OBJECT(m)); + + if ((cur_object_lock_type == OBJECT_LOCK_SHARED) && + VM_FAULT_NEED_CS_VALIDATION(NULL, m, m_object)) { + goto upgrade_for_validation; + } + + /* * Allocate a page in the original top level * object. Give up if allocate fails. Also * need to remember current page, as it's the * source of the copy. * - * at this point we hold locks on both + * at this point we hold locks on both * object and cur_object... no need to take * paging refs or mark pages BUSY since * we don't drop either object lock until * the page has been copied and inserted */ cur_m = m; - m = vm_page_grab(); + m = vm_page_grab_options(grab_options); + m_object = NULL; if (m == VM_PAGE_NULL) { /* @@ -3130,14 +4132,20 @@ FastPmapEnter: */ vm_page_copy(cur_m, m); vm_page_insert(m, object, offset); - m->dirty = TRUE; + m_object = object; + SET_PAGE_DIRTY(m, FALSE); /* * Now cope with the source page and object */ - if (object->ref_count > 1 && cur_m->pmapped) - pmap_disconnect(cur_m->phys_page); + if (object->ref_count > 1 && cur_m->vmp_pmapped) + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(cur_m)); + if (cur_m->vmp_clustered) { + VM_PAGE_COUNT_AS_PAGEIN(cur_m); + VM_PAGE_CONSUME_CLUSTERED(cur_m); + vm_fault_is_sequential(cur_object, cur_offset, fault_info.behavior); + } need_collapse = TRUE; if (!cur_object->internal && @@ -3183,12 +4191,233 @@ FastPmapEnter: * No page at cur_object, cur_offset... m == NULL */ if (cur_object->pager_created) { - if (MUST_ASK_PAGER(cur_object, cur_offset) == TRUE) { + int compressor_external_state = VM_EXTERNAL_STATE_UNKNOWN; + + if (MUST_ASK_PAGER(cur_object, cur_offset, compressor_external_state) == TRUE) { + int my_fault_type; + int c_flags = C_DONT_BLOCK; + boolean_t insert_cur_object = FALSE; + /* * May have to talk to a pager... - * take the slow path. + * if so, take the slow path by + * doing a 'break' from the while (TRUE) loop + * + * external_state will only be set to VM_EXTERNAL_STATE_EXISTS + * if the compressor is active and the page exists there + */ + if (compressor_external_state != VM_EXTERNAL_STATE_EXISTS) + break; + + if (map == kernel_map || real_map == kernel_map) { + /* + * can't call into the compressor with the kernel_map + * lock held, since the compressor may try to operate + * on the kernel map in order to return an empty c_segment + */ + break; + } + if (object != cur_object) { + if (fault_type & VM_PROT_WRITE) + c_flags |= C_KEEP; + else + insert_cur_object = TRUE; + } + if (insert_cur_object == TRUE) { + + if (cur_object_lock_type == OBJECT_LOCK_SHARED) { + + cur_object_lock_type = OBJECT_LOCK_EXCLUSIVE; + + if (vm_object_lock_upgrade(cur_object) == FALSE) { + /* + * couldn't upgrade so go do a full retry + * immediately since we can no longer be + * certain about cur_object (since we + * don't hold a reference on it)... + * first drop the top object lock + */ + vm_object_unlock(object); + + vm_map_unlock_read(map); + if (real_map != map) + vm_map_unlock(real_map); + + goto RetryFault; + } + } + } else if (object_lock_type == OBJECT_LOCK_SHARED) { + + object_lock_type = OBJECT_LOCK_EXCLUSIVE; + + if (object != cur_object) { + /* + * we can't go for the upgrade on the top + * lock since the upgrade may block waiting + * for readers to drain... since we hold + * cur_object locked at this point, waiting + * for the readers to drain would represent + * a lock order inversion since the lock order + * for objects is the reference order in the + * shadown chain + */ + vm_object_unlock(object); + vm_object_unlock(cur_object); + + vm_map_unlock_read(map); + if (real_map != map) + vm_map_unlock(real_map); + + goto RetryFault; + } + if (vm_object_lock_upgrade(object) == FALSE) { + /* + * couldn't upgrade, so explictly take the lock + * exclusively and go relookup the page since we + * will have dropped the object lock and + * a different thread could have inserted + * a page at this offset + * no need for a full retry since we're + * at the top level of the object chain + */ + vm_object_lock(object); + + continue; + } + } + m = vm_page_grab_options(grab_options); + m_object = NULL; + + if (m == VM_PAGE_NULL) { + /* + * no free page currently available... + * must take the slow path + */ + break; + } + + /* + * The object is and remains locked + * so no need to take a + * "paging_in_progress" reference. + */ + boolean_t shared_lock; + if ((object == cur_object && + object_lock_type == OBJECT_LOCK_EXCLUSIVE) || + (object != cur_object && + cur_object_lock_type == OBJECT_LOCK_EXCLUSIVE)) { + shared_lock = FALSE; + } else { + shared_lock = TRUE; + } + + kr = vm_compressor_pager_get( + cur_object->pager, + (cur_offset + + cur_object->paging_offset), + VM_PAGE_GET_PHYS_PAGE(m), + &my_fault_type, + c_flags, + &compressed_count_delta); + + vm_compressor_pager_count( + cur_object->pager, + compressed_count_delta, + shared_lock, + cur_object); + + if (kr != KERN_SUCCESS) { + vm_page_release(m, FALSE); + m = VM_PAGE_NULL; + break; + } + m->vmp_dirty = TRUE; + + /* + * If the object is purgeable, its + * owner's purgeable ledgers will be + * updated in vm_page_insert() but the + * page was also accounted for in a + * "compressed purgeable" ledger, so + * update that now. */ - break; + if (object != cur_object && + !insert_cur_object) { + /* + * We're not going to insert + * the decompressed page into + * the object it came from. + * + * We're dealing with a + * copy-on-write fault on + * "object". + * We're going to decompress + * the page directly into the + * target "object" while + * keepin the compressed + * page for "cur_object", so + * no ledger update in that + * case. + */ + } else if (((cur_object->purgable == + VM_PURGABLE_DENY) && + (!cur_object->vo_ledger_tag)) || + (cur_object->vo_owner == + NULL)) { + /* + * "cur_object" is not purgeable + * and is not ledger-taged, or + * there's no owner for it, + * so no owner's ledgers to + * update. + */ + } else { + /* + * One less compressed + * purgeable/tagged page for + * cur_object's owner. + */ + vm_object_owner_compressed_update( + cur_object, + -1); + } + + if (insert_cur_object) { + vm_page_insert(m, cur_object, cur_offset); + m_object = cur_object; + } else { + vm_page_insert(m, object, offset); + m_object = object; + } + + if ((m_object->wimg_bits & VM_WIMG_MASK) != VM_WIMG_USE_DEFAULT) { + /* + * If the page is not cacheable, + * we can't let its contents + * linger in the data cache + * after the decompression. + */ + pmap_sync_page_attributes_phys(VM_PAGE_GET_PHYS_PAGE(m)); + } + + type_of_fault = my_fault_type; + + VM_STAT_INCR(decompressions); + + if (cur_object != object) { + if (insert_cur_object) { + top_object = object; + /* + * switch to the object that has the new page + */ + object = cur_object; + object_lock_type = cur_object_lock_type; + } else { + vm_object_unlock(cur_object); + cur_object = object; + } + } + goto FastPmapEnter; } /* * existence map present and indicates @@ -3201,8 +4430,10 @@ FastPmapEnter: * inserted into the original object. */ if (cur_object->shadow_severed || - VM_OBJECT_PURGEABLE_FAULT_ERROR(cur_object)) - { + VM_OBJECT_PURGEABLE_FAULT_ERROR(cur_object) || + cur_object == compressor_object || + cur_object == kernel_object || + cur_object == vm_submap_object) { if (object != cur_object) vm_object_unlock(cur_object); vm_object_unlock(object); @@ -3214,40 +4445,6 @@ FastPmapEnter: kr = KERN_MEMORY_ERROR; goto done; } - if (vm_page_throttled()) { - /* - * drop all of our locks... - * wait until the free queue is - * pumped back up and then - * redrive the fault - */ - if (object != cur_object) - vm_object_unlock(cur_object); - vm_object_unlock(object); - vm_map_unlock_read(map); - if (real_map != map) - vm_map_unlock(real_map); - - if (NEED_TO_HARD_THROTTLE_THIS_TASK()) - delay(HARD_THROTTLE_DELAY); - - if (!current_thread_aborted() && vm_page_wait((change_wiring) ? - THREAD_UNINT : - THREAD_ABORTSAFE)) - goto RetryFault; - kr = KERN_ABORTED; - goto done; - } - if (vm_backing_store_low) { - /* - * we are protecting the system from - * backing store exhaustion... - * must take the slow path if we're - * not privileged - */ - if (!(current_task()->priv_flags & VM_BACKING_STORE_PRIV)) - break; - } if (cur_object != object) { vm_object_unlock(cur_object); @@ -3272,6 +4469,7 @@ FastPmapEnter: } } m = vm_page_alloc(object, offset); + m_object = NULL; if (m == VM_PAGE_NULL) { /* @@ -3280,10 +4478,11 @@ FastPmapEnter: */ break; } + m_object = object; /* * Now zero fill page... - * the page is probably going to + * the page is probably going to * be written soon, so don't bother * to clear the modified bit * @@ -3297,7 +4496,7 @@ FastPmapEnter: /* * On to the next level in the shadow chain */ - cur_offset += cur_object->shadow_offset; + cur_offset += cur_object->vo_shadow_offset; new_object = cur_object->shadow; /* @@ -3346,6 +4545,28 @@ handle_copy_delay: if (real_map != map) vm_map_unlock(real_map); + if (__improbable(object == compressor_object || + object == kernel_object || + object == vm_submap_object)) { + /* + * These objects are explicitly managed and populated by the + * kernel. The virtual ranges backed by these objects should + * either have wired pages or "holes" that are not supposed to + * be accessed at all until they get explicitly populated. + * We should never have to resolve a fault on a mapping backed + * by one of these VM objects and providing a zero-filled page + * would be wrong here, so let's fail the fault and let the + * caller crash or recover. + */ + vm_object_unlock(object); + kr = KERN_MEMORY_ERROR; + goto done; + } + + assert(object != compressor_object); + assert(object != kernel_object); + assert(object != vm_submap_object); + /* * Make a reference to this object to * prevent its disposal while we are messing with @@ -3360,8 +4581,10 @@ handle_copy_delay: error_code = 0; + result_page = VM_PAGE_NULL; kr = vm_fault_page(object, offset, fault_type, (change_wiring && !wired), + FALSE, /* page not looked up */ &prot, &result_page, &top_page, &type_of_fault, &error_code, map->no_zero_fill, @@ -3377,7 +4600,7 @@ handle_copy_delay: * * the object is returned locked with a paging reference * - * if top_page != NULL, then it's BUSY and the + * if top_page != NULL, then it's BUSY and the * object it belongs to has a paging reference * but is returned unlocked */ @@ -3393,7 +4616,7 @@ handle_copy_delay: */ switch (kr) { case VM_FAULT_MEMORY_SHORTAGE: - if (vm_page_wait((change_wiring) ? + if (vm_page_wait((change_wiring) ? THREAD_UNINT : THREAD_ABORTSAFE)) goto RetryFault; @@ -3417,11 +4640,13 @@ handle_copy_delay: } } m = result_page; + m_object = NULL; if (m != VM_PAGE_NULL) { + m_object = VM_PAGE_OBJECT(m); assert((change_wiring && !wired) ? - (top_page == VM_PAGE_NULL) : - ((top_page == VM_PAGE_NULL) == (m->object == object))); + (top_page == VM_PAGE_NULL) : + ((top_page == VM_PAGE_NULL) == (m_object == object))); } /* @@ -3431,30 +4656,55 @@ handle_copy_delay: #define RELEASE_PAGE(m) \ MACRO_BEGIN \ PAGE_WAKEUP_DONE(m); \ - if (!m->active && !m->inactive && !m->throttled) { \ - vm_page_lockspin_queues(); \ - if (!m->active && !m->inactive && !m->throttled) \ - vm_page_activate(m); \ - vm_page_unlock_queues(); \ - } \ + if ( !VM_PAGE_PAGEABLE(m)) { \ + vm_page_lockspin_queues(); \ + if ( !VM_PAGE_PAGEABLE(m)) \ + vm_page_activate(m); \ + vm_page_unlock_queues(); \ + } \ MACRO_END + + object_locks_dropped = FALSE; /* * We must verify that the maps have not changed - * since our last lookup. + * since our last lookup. vm_map_verify() needs the + * map lock (shared) but we are holding object locks. + * So we do a try_lock() first and, if that fails, we + * drop the object locks and go in for the map lock again. */ - if (m != VM_PAGE_NULL) { - old_copy_object = m->object->copy; - vm_object_unlock(m->object); - } else { - old_copy_object = VM_OBJECT_NULL; - vm_object_unlock(object); + if (!vm_map_try_lock_read(original_map)) { + + if (m != VM_PAGE_NULL) { + old_copy_object = m_object->copy; + vm_object_unlock(m_object); + } else { + old_copy_object = VM_OBJECT_NULL; + vm_object_unlock(object); + } + + object_locks_dropped = TRUE; + + vm_map_lock_read(original_map); } - /* - * no object locks are held at this point - */ if ((map != original_map) || !vm_map_verify(map, &version)) { + + if (object_locks_dropped == FALSE) { + if (m != VM_PAGE_NULL) { + old_copy_object = m_object->copy; + vm_object_unlock(m_object); + } else { + old_copy_object = VM_OBJECT_NULL; + vm_object_unlock(object); + } + + object_locks_dropped = TRUE; + } + + /* + * no object locks are held at this point + */ vm_object_t retry_object; vm_object_offset_t retry_offset; vm_prot_t retry_prot; @@ -3469,7 +4719,6 @@ handle_copy_delay: * take another fault. */ map = original_map; - vm_map_lock_read(map); kr = vm_map_lookup_locked(&map, vaddr, fault_type & ~VM_PROT_WRITE, @@ -3484,17 +4733,19 @@ handle_copy_delay: vm_map_unlock_read(map); if (m != VM_PAGE_NULL) { + assert(VM_PAGE_OBJECT(m) == m_object); + /* * retake the lock so that * we can drop the paging reference * in vm_fault_cleanup and do the * PAGE_WAKEUP_DONE in RELEASE_PAGE */ - vm_object_lock(m->object); + vm_object_lock(m_object); RELEASE_PAGE(m); - vm_fault_cleanup(m->object, top_page); + vm_fault_cleanup(m_object, top_page); } else { /* * retake the lock so that @@ -3518,17 +4769,19 @@ handle_copy_delay: vm_map_unlock(real_map); if (m != VM_PAGE_NULL) { + assert(VM_PAGE_OBJECT(m) == m_object); + /* * retake the lock so that * we can drop the paging reference * in vm_fault_cleanup and do the * PAGE_WAKEUP_DONE in RELEASE_PAGE */ - vm_object_lock(m->object); + vm_object_lock(m_object); RELEASE_PAGE(m); - vm_fault_cleanup(m->object, top_page); + vm_fault_cleanup(m_object, top_page); } else { /* * retake the lock so that @@ -3547,20 +4800,31 @@ handle_copy_delay: * Check whether the protection has changed or the object * has been copied while we left the map unlocked. */ - prot &= retry_prot; + if (pmap_has_prot_policy(retry_prot)) { + /* If the pmap layer cares, pass the full set. */ + prot = retry_prot; + } else { + prot &= retry_prot; + } } - if (m != VM_PAGE_NULL) { - vm_object_lock(m->object); - if (m->object->copy != old_copy_object) { - /* - * The copy object changed while the top-level object - * was unlocked, so take away write permission. - */ - prot &= ~VM_PROT_WRITE; - } - } else - vm_object_lock(object); + if (object_locks_dropped == TRUE) { + if (m != VM_PAGE_NULL) { + vm_object_lock(m_object); + + if (m_object->copy != old_copy_object) { + /* + * The copy object changed while the top-level object + * was unlocked, so take away write permission. + */ + assert(!pmap_has_prot_policy(prot)); + prot &= ~VM_PROT_WRITE; + } + } else + vm_object_lock(object); + + object_locks_dropped = FALSE; + } /* * If we want to wire down this page, but no longer have @@ -3568,14 +4832,16 @@ handle_copy_delay: */ if (wired && (fault_type != (prot | VM_PROT_WRITE))) { - vm_map_verify_done(map, &version); + vm_map_unlock_read(map); if (real_map != map) vm_map_unlock(real_map); if (m != VM_PAGE_NULL) { + assert(VM_PAGE_OBJECT(m) == m_object); + RELEASE_PAGE(m); - vm_fault_cleanup(m->object, top_page); + vm_fault_cleanup(m_object, top_page); } else vm_fault_cleanup(object, top_page); @@ -3596,63 +4862,73 @@ handle_copy_delay: caller_pmap, caller_pmap_addr, prot, + caller_prot, wired, change_wiring, - fault_info.no_cache, + wire_tag, + &fault_info, + NULL, &type_of_fault); } else { kr = vm_fault_enter(m, pmap, vaddr, prot, + caller_prot, wired, change_wiring, - fault_info.no_cache, + wire_tag, + &fault_info, + NULL, &type_of_fault); } + assert(VM_PAGE_OBJECT(m) == m_object); + +#if DEVELOPMENT || DEBUG + { + int event_code = 0; + + if (m_object->internal) + event_code = (MACHDBG_CODE(DBG_MACH_WORKINGSET, VM_REAL_FAULT_ADDR_INTERNAL)); + else if (m_object->object_is_shared_cache) + event_code = (MACHDBG_CODE(DBG_MACH_WORKINGSET, VM_REAL_FAULT_ADDR_SHAREDCACHE)); + else + event_code = (MACHDBG_CODE(DBG_MACH_WORKINGSET, VM_REAL_FAULT_ADDR_EXTERNAL)); + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, event_code, trace_real_vaddr, (fault_info.user_tag << 16) | (caller_prot << 8) | type_of_fault, m->vmp_offset, get_current_unique_pid(), 0); + + DTRACE_VM6(real_fault, vm_map_offset_t, real_vaddr, vm_map_offset_t, m->vmp_offset, int, event_code, int, caller_prot, int, type_of_fault, int, fault_info.user_tag); + } +#endif if (kr != KERN_SUCCESS) { /* abort this page fault */ - vm_map_verify_done(map, &version); + vm_map_unlock_read(map); if (real_map != map) vm_map_unlock(real_map); PAGE_WAKEUP_DONE(m); - vm_fault_cleanup(m->object, top_page); + vm_fault_cleanup(m_object, top_page); vm_object_deallocate(object); goto done; } + if (physpage_p != NULL) { + /* for vm_map_wire_and_extract() */ + *physpage_p = VM_PAGE_GET_PHYS_PAGE(m); + if (prot & VM_PROT_WRITE) { + vm_object_lock_assert_exclusive(m_object); + m->vmp_dirty = TRUE; + } + } } else { vm_map_entry_t entry; vm_map_offset_t laddr; vm_map_offset_t ldelta, hdelta; - /* + /* * do a pmap block mapping from the physical address - * in the object + * in the object */ -#ifdef ppc - /* While we do not worry about execution protection in */ - /* general, certian pages may have instruction execution */ - /* disallowed. We will check here, and if not allowed */ - /* to execute, we return with a protection failure. */ - - if ((fault_type & VM_PROT_EXECUTE) && - (!pmap_eligible_for_execute((ppnum_t)(object->shadow_offset >> 12)))) { - - vm_map_verify_done(map, &version); - - if (real_map != map) - vm_map_unlock(real_map); - - vm_fault_cleanup(object, top_page); - vm_object_deallocate(object); - - kr = KERN_PROTECTION_FAILURE; - goto done; - } -#endif /* ppc */ - if (real_map != map) vm_map_unlock(real_map); @@ -3673,66 +4949,114 @@ handle_copy_delay: if (hdelta > (entry->vme_end - laddr)) hdelta = entry->vme_end - laddr; if (entry->is_sub_map) { - - laddr = (laddr - entry->vme_start) - + entry->offset; - vm_map_lock_read(entry->object.sub_map); + + laddr = ((laddr - entry->vme_start) + + VME_OFFSET(entry)); + vm_map_lock_read(VME_SUBMAP(entry)); if (map != real_map) vm_map_unlock_read(map); if (entry->use_pmap) { vm_map_unlock_read(real_map); - real_map = entry->object.sub_map; + real_map = VME_SUBMAP(entry); } - map = entry->object.sub_map; - + map = VME_SUBMAP(entry); + } else { break; } } - if (vm_map_lookup_entry(map, laddr, &entry) && - (entry->object.vm_object != NULL) && - (entry->object.vm_object == object)) { + if (vm_map_lookup_entry(map, laddr, &entry) && + (VME_OBJECT(entry) != NULL) && + (VME_OBJECT(entry) == object)) { + int superpage; + + if (!object->pager_created && + object->phys_contiguous && + VME_OFFSET(entry) == 0 && + (entry->vme_end - entry->vme_start == object->vo_size) && + VM_MAP_PAGE_ALIGNED(entry->vme_start, (object->vo_size-1))) { + superpage = VM_MEM_SUPERPAGE; + } else { + superpage = 0; + } + + if (superpage && physpage_p) { + /* for vm_map_wire_and_extract() */ + *physpage_p = (ppnum_t) + ((((vm_map_offset_t) + object->vo_shadow_offset) + + VME_OFFSET(entry) + + (laddr - entry->vme_start)) + >> PAGE_SHIFT); + } - int superpage = (!object->pager_created && object->phys_contiguous)? VM_MEM_SUPERPAGE : 0; if (caller_pmap) { /* * Set up a block mapped area */ - assert((uint32_t)((ldelta + hdelta) >> 12) == ((ldelta + hdelta) >> 12)); - pmap_map_block(caller_pmap, - (addr64_t)(caller_pmap_addr - ldelta), - (ppnum_t)((((vm_map_offset_t) (entry->object.vm_object->shadow_offset)) + - entry->offset + (laddr - entry->vme_start) - ldelta) >> 12), - (uint32_t)((ldelta + hdelta) >> 12), prot, - (VM_WIMG_MASK & (int)object->wimg_bits) | superpage, 0); - } else { + assert((uint32_t)((ldelta + hdelta) >> PAGE_SHIFT) == ((ldelta + hdelta) >> PAGE_SHIFT)); + kr = pmap_map_block(caller_pmap, + (addr64_t)(caller_pmap_addr - ldelta), + (ppnum_t)((((vm_map_offset_t) (VME_OBJECT(entry)->vo_shadow_offset)) + + VME_OFFSET(entry) + (laddr - entry->vme_start) - ldelta) >> PAGE_SHIFT), + (uint32_t)((ldelta + hdelta) >> PAGE_SHIFT), prot, + (VM_WIMG_MASK & (int)object->wimg_bits) | superpage, 0); + + if (kr != KERN_SUCCESS) { + goto cleanup; + } + } else { /* * Set up a block mapped area */ - assert((uint32_t)((ldelta + hdelta) >> 12) == ((ldelta + hdelta) >> 12)); - pmap_map_block(real_map->pmap, - (addr64_t)(vaddr - ldelta), - (ppnum_t)((((vm_map_offset_t)(entry->object.vm_object->shadow_offset)) + - entry->offset + (laddr - entry->vme_start) - ldelta) >> 12), - (uint32_t)((ldelta + hdelta) >> 12), prot, - (VM_WIMG_MASK & (int)object->wimg_bits) | superpage, 0); + assert((uint32_t)((ldelta + hdelta) >> PAGE_SHIFT) == ((ldelta + hdelta) >> PAGE_SHIFT)); + kr = pmap_map_block(real_map->pmap, + (addr64_t)(vaddr - ldelta), + (ppnum_t)((((vm_map_offset_t)(VME_OBJECT(entry)->vo_shadow_offset)) + + VME_OFFSET(entry) + (laddr - entry->vme_start) - ldelta) >> PAGE_SHIFT), + (uint32_t)((ldelta + hdelta) >> PAGE_SHIFT), prot, + (VM_WIMG_MASK & (int)object->wimg_bits) | superpage, 0); + + if (kr != KERN_SUCCESS) { + goto cleanup; + } } } } + /* + * Success + */ + kr = KERN_SUCCESS; + + /* + * TODO: could most of the done cases just use cleanup? + */ +cleanup: /* * Unlock everything, and return */ - vm_map_verify_done(map, &version); + vm_map_unlock_read(map); if (real_map != map) vm_map_unlock(real_map); if (m != VM_PAGE_NULL) { + assert(VM_PAGE_OBJECT(m) == m_object); + + if (!m_object->internal && (fault_type & VM_PROT_WRITE)) { + + vm_object_paging_begin(m_object); + + assert(written_on_object == VM_OBJECT_NULL); + written_on_object = m_object; + written_on_pager = m_object->pager; + written_on_offset = m_object->paging_offset + m->vmp_offset; + } PAGE_WAKEUP_DONE(m); - vm_fault_cleanup(m->object, top_page); + vm_fault_cleanup(m_object, top_page); } else vm_fault_cleanup(object, top_page); @@ -3740,13 +5064,51 @@ handle_copy_delay: #undef RELEASE_PAGE - kr = KERN_SUCCESS; done: thread_interrupt_level(interruptible_state); - KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, 2)) | DBG_FUNC_END, - (int)((uint64_t)vaddr >> 32), - (int)vaddr, + /* + * Only I/O throttle on faults which cause a pagein/swapin. + */ + if ((type_of_fault == DBG_PAGEIND_FAULT) || (type_of_fault == DBG_PAGEINV_FAULT) || (type_of_fault == DBG_COMPRESSOR_SWAPIN_FAULT)) { + throttle_lowpri_io(1); + } else { + if (kr == KERN_SUCCESS && type_of_fault != DBG_CACHE_HIT_FAULT && type_of_fault != DBG_GUARD_FAULT) { + + if ((throttle_delay = vm_page_throttled(TRUE))) { + + if (vm_debug_events) { + if (type_of_fault == DBG_COMPRESSOR_FAULT) + VM_DEBUG_EVENT(vmf_compressordelay, VMF_COMPRESSORDELAY, DBG_FUNC_NONE, throttle_delay, 0, 0, 0); + else if (type_of_fault == DBG_COW_FAULT) + VM_DEBUG_EVENT(vmf_cowdelay, VMF_COWDELAY, DBG_FUNC_NONE, throttle_delay, 0, 0, 0); + else + VM_DEBUG_EVENT(vmf_zfdelay, VMF_ZFDELAY, DBG_FUNC_NONE, throttle_delay, 0, 0, 0); + } + delay(throttle_delay); + } + } + } + + if (written_on_object) { + + vnode_pager_dirtied(written_on_pager, written_on_offset, written_on_offset + PAGE_SIZE_64); + + vm_object_lock(written_on_object); + vm_object_paging_end(written_on_object); + vm_object_unlock(written_on_object); + + written_on_object = VM_OBJECT_NULL; + } + + if (rtfault) { + vm_record_rtfault(cthread, fstart, trace_vaddr, type_of_fault); + } + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + (MACHDBG_CODE(DBG_MACH_VM, 2)) | DBG_FUNC_END, + ((uint64_t)trace_vaddr >> 32), + trace_vaddr, kr, type_of_fault, 0); @@ -3763,19 +5125,21 @@ kern_return_t vm_fault_wire( vm_map_t map, vm_map_entry_t entry, + vm_prot_t prot, + vm_tag_t wire_tag, pmap_t pmap, - vm_map_offset_t pmap_addr) + vm_map_offset_t pmap_addr, + ppnum_t *physpage_p) { - - register vm_map_offset_t va; - register vm_map_offset_t end_addr = entry->vme_end; - register kern_return_t rc; + vm_map_offset_t va; + vm_map_offset_t end_addr = entry->vme_end; + kern_return_t rc; assert(entry->in_transition); - if ((entry->object.vm_object != NULL) && - !entry->is_sub_map && - entry->object.vm_object->phys_contiguous) { + if ((VME_OBJECT(entry) != NULL) && + !entry->is_sub_map && + VME_OBJECT(entry)->phys_contiguous) { return KERN_SUCCESS; } @@ -3785,7 +5149,7 @@ vm_fault_wire( * page tables and such can be locked down as well. */ - pmap_pageable(pmap, pmap_addr, + pmap_pageable(pmap, pmap_addr, pmap_addr + (end_addr - entry->vme_start), FALSE); /* @@ -3794,14 +5158,18 @@ vm_fault_wire( */ for (va = entry->vme_start; va < end_addr; va += PAGE_SIZE) { - if ((rc = vm_fault_wire_fast( - map, va, entry, pmap, - pmap_addr + (va - entry->vme_start) - )) != KERN_SUCCESS) { - rc = vm_fault(map, va, VM_PROT_NONE, TRUE, - (pmap == kernel_pmap) ? - THREAD_UNINT : THREAD_ABORTSAFE, - pmap, pmap_addr + (va - entry->vme_start)); + rc = vm_fault_wire_fast(map, va, prot, wire_tag, entry, pmap, + pmap_addr + (va - entry->vme_start), + physpage_p); + if (rc != KERN_SUCCESS) { + rc = vm_fault_internal(map, va, prot, TRUE, wire_tag, + ((pmap == kernel_pmap) + ? THREAD_UNINT + : THREAD_ABORTSAFE), + pmap, + (pmap_addr + + (va - entry->vme_start)), + physpage_p); DTRACE_VM2(softlock, int, 1, (uint64_t *), NULL); } @@ -3810,7 +5178,7 @@ vm_fault_wire( /* unwire wired pages */ tmp_entry.vme_end = va; - vm_fault_unwire(map, + vm_fault_unwire(map, &tmp_entry, FALSE, pmap, pmap_addr); return rc; @@ -3832,13 +5200,13 @@ vm_fault_unwire( pmap_t pmap, vm_map_offset_t pmap_addr) { - register vm_map_offset_t va; - register vm_map_offset_t end_addr = entry->vme_end; + vm_map_offset_t va; + vm_map_offset_t end_addr = entry->vme_end; vm_object_t object; - struct vm_object_fault_info fault_info; + struct vm_object_fault_info fault_info = {}; + unsigned int unwired_pages; - object = (entry->is_sub_map) - ? VM_OBJECT_NULL : entry->object.vm_object; + object = (entry->is_sub_map) ? VM_OBJECT_NULL : VME_OBJECT(entry); /* * If it's marked phys_contiguous, then vm_fault_wire() didn't actually @@ -3851,12 +5219,18 @@ vm_fault_unwire( fault_info.interruptible = THREAD_UNINT; fault_info.behavior = entry->behavior; - fault_info.user_tag = entry->alias; - fault_info.lo_offset = entry->offset; - fault_info.hi_offset = (entry->vme_end - entry->vme_start) + entry->offset; + fault_info.user_tag = VME_ALIAS(entry); + if (entry->iokit_acct || + (!entry->is_sub_map && !entry->use_pmap)) { + fault_info.pmap_options |= PMAP_OPTIONS_ALT_ACCT; + } + fault_info.lo_offset = VME_OFFSET(entry); + fault_info.hi_offset = (entry->vme_end - entry->vme_start) + VME_OFFSET(entry); fault_info.no_cache = entry->no_cache; fault_info.stealth = TRUE; + unwired_pages = 0; + /* * Since the pages are wired down, we must be able to * get their mappings from the physical map system. @@ -3866,11 +5240,11 @@ vm_fault_unwire( if (object == VM_OBJECT_NULL) { if (pmap) { - pmap_change_wiring(pmap, + pmap_change_wiring(pmap, pmap_addr + (va - entry->vme_start), FALSE); } - (void) vm_fault(map, va, VM_PROT_NONE, - TRUE, THREAD_UNINT, pmap, pmap_addr); + (void) vm_fault(map, va, VM_PROT_NONE, + TRUE, VM_KERN_MEMORY_NONE, THREAD_UNINT, pmap, pmap_addr); } else { vm_prot_t prot; vm_page_t result_page; @@ -3878,13 +5252,12 @@ vm_fault_unwire( vm_object_t result_object; vm_fault_return_t result; - if (end_addr - va > (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) (end_addr - va); - assert(fault_info.cluster_size == end_addr - va); + /* cap cluster size at maximum UPL size */ + upl_size_t cluster_size; + if (os_sub_overflow(end_addr, va, &cluster_size)) { + cluster_size = 0 - (upl_size_t)PAGE_SIZE; } + fault_info.cluster_size = cluster_size; do { prot = VM_PROT_NONE; @@ -3894,13 +5267,16 @@ vm_fault_unwire( XPR(XPR_VM_FAULT, "vm_fault_unwire -> vm_fault_page\n", 0,0,0,0,0); + result_page = VM_PAGE_NULL; result = vm_fault_page( object, - entry->offset + (va - entry->vme_start), + (VME_OFFSET(entry) + + (va - entry->vme_start)), VM_PROT_NONE, TRUE, + FALSE, /* page not looked up */ &prot, &result_page, &top_page, (int *)0, - NULL, map->no_zero_fill, + NULL, map->no_zero_fill, FALSE, &fault_info); } while (result == VM_FAULT_RETRY); @@ -3910,35 +5286,52 @@ vm_fault_unwire( * move on to the next one in case the remaining pages are mapped from * different objects. During a forced unmount, the object is terminated * so the alive flag will be false if this happens. A forced unmount will - * will occur when an external disk is unplugged before the user does an + * will occur when an external disk is unplugged before the user does an * eject, so we don't want to panic in that situation. */ if (result == VM_FAULT_MEMORY_ERROR && !object->alive) continue; + if (result == VM_FAULT_MEMORY_ERROR && + object == kernel_object) { + /* + * This must have been allocated with + * KMA_KOBJECT and KMA_VAONLY and there's + * no physical page at this offset. + * We're done (no page to free). + */ + assert(deallocate); + continue; + } + if (result != VM_FAULT_SUCCESS) panic("vm_fault_unwire: failure"); - result_object = result_page->object; + result_object = VM_PAGE_OBJECT(result_page); - if ((pmap) && (result_page->phys_page != vm_page_guard_addr)) { - pmap_change_wiring(pmap, - pmap_addr + (va - entry->vme_start), FALSE); - } if (deallocate) { - assert(result_page->phys_page != + assert(VM_PAGE_GET_PHYS_PAGE(result_page) != vm_page_fictitious_addr); - pmap_disconnect(result_page->phys_page); + pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(result_page)); + if (VM_PAGE_WIRED(result_page)) { + unwired_pages++; + } VM_PAGE_FREE(result_page); } else { + if ((pmap) && (VM_PAGE_GET_PHYS_PAGE(result_page) != vm_page_guard_addr)) + pmap_change_wiring(pmap, + pmap_addr + (va - entry->vme_start), FALSE); + + if (VM_PAGE_WIRED(result_page)) { vm_page_lockspin_queues(); - vm_page_unwire(result_page); + vm_page_unwire(result_page, TRUE); vm_page_unlock_queues(); + unwired_pages++; } if(entry->zero_wired_pages) { - pmap_zero_page(result_page->phys_page); + pmap_zero_page(VM_PAGE_GET_PHYS_PAGE(result_page)); entry->zero_wired_pages = FALSE; } @@ -3954,9 +5347,12 @@ vm_fault_unwire( * such may be unwired themselves. */ - pmap_pageable(pmap, pmap_addr, + pmap_pageable(pmap, pmap_addr, pmap_addr + (end_addr - entry->vme_start), TRUE); + if (kernel_object == object) { + vm_tag_update_size(fault_info.user_tag, -ptoa_64(unwired_pages)); + } } /* @@ -3979,21 +5375,25 @@ vm_fault_unwire( * other than the common case will return KERN_FAILURE, and the caller * is expected to call vm_fault(). */ -kern_return_t +static kern_return_t vm_fault_wire_fast( __unused vm_map_t map, vm_map_offset_t va, + __unused vm_prot_t caller_prot, + vm_tag_t wire_tag, vm_map_entry_t entry, - pmap_t pmap, - vm_map_offset_t pmap_addr) + pmap_t pmap, + vm_map_offset_t pmap_addr, + ppnum_t *physpage_p) { vm_object_t object; vm_object_offset_t offset; - register vm_page_t m; + vm_page_t m; vm_prot_t prot; thread_t thread = current_thread(); int type_of_fault; kern_return_t kr; + struct vm_object_fault_info fault_info = {}; VM_STAT_INCR(faults); @@ -4008,7 +5408,7 @@ vm_fault_wire_fast( #define RELEASE_PAGE(m) { \ PAGE_WAKEUP_DONE(m); \ vm_page_lockspin_queues(); \ - vm_page_unwire(m); \ + vm_page_unwire(m, TRUE); \ vm_page_unlock_queues(); \ } @@ -4037,15 +5437,17 @@ vm_fault_wire_fast( /* * If this entry is not directly to a vm_object, bail out. */ - if (entry->is_sub_map) + if (entry->is_sub_map) { + assert(physpage_p == NULL); return(KERN_FAILURE); + } /* * Find the backing store object and offset into it. */ - object = entry->object.vm_object; - offset = (va - entry->vme_start) + entry->offset; + object = VME_OBJECT(entry); + offset = (va - entry->vme_start) + VME_OFFSET(entry); prot = entry->protection; /* @@ -4074,19 +5476,15 @@ vm_fault_wire_fast( /* * Look for page in top-level object. If it's not there or * there's something going on, give up. - * ENCRYPTED SWAP: use the slow fault path, since we'll need to - * decrypt the page before wiring it down. */ m = vm_page_lookup(object, offset); - if ((m == VM_PAGE_NULL) || (m->busy) || (m->encrypted) || - (m->unusual && ( m->error || m->restart || m->absent))) { + if ((m == VM_PAGE_NULL) || (m->vmp_busy) || + (m->vmp_unusual && ( m->vmp_error || m->vmp_restart || m->vmp_absent))) { GIVE_UP; } - ASSERT_PAGE_DECRYPTED(m); - - if (m->fictitious && - m->phys_page == vm_page_guard_addr) { + if (m->vmp_fictitious && + VM_PAGE_GET_PHYS_PAGE(m) == vm_page_guard_addr) { /* * Guard pages are fictitious pages and are never * entered into a pmap, so let's say it's been wired... @@ -4097,19 +5495,19 @@ vm_fault_wire_fast( /* * Wire the page down now. All bail outs beyond this - * point must unwire the page. + * point must unwire the page. */ vm_page_lockspin_queues(); - vm_page_wire(m); + vm_page_wire(m, wire_tag, TRUE); vm_page_unlock_queues(); /* * Mark page busy for other threads. */ - assert(!m->busy); - m->busy = TRUE; - assert(!m->absent); + assert(!m->vmp_busy); + m->vmp_busy = TRUE; + assert(!m->vmp_absent); /* * Give up if the page is being written and there's a copy object @@ -4119,6 +5517,13 @@ vm_fault_wire_fast( GIVE_UP; } + fault_info.user_tag = VME_ALIAS(entry); + fault_info.pmap_options = 0; + if (entry->iokit_acct || + (!entry->is_sub_map && !entry->use_pmap)) { + fault_info.pmap_options |= PMAP_OPTIONS_ALT_ACCT; + } + /* * Put this page into the physical map. */ @@ -4127,16 +5532,37 @@ vm_fault_wire_fast( pmap, pmap_addr, prot, - TRUE, - FALSE, - FALSE, + prot, + TRUE, /* wired */ + FALSE, /* change_wiring */ + wire_tag, + &fault_info, + NULL, &type_of_fault); + if (kr != KERN_SUCCESS) { + RELEASE_PAGE(m); + GIVE_UP; + } done: /* * Unlock everything, and return */ + if (physpage_p) { + /* for vm_map_wire_and_extract() */ + if (kr == KERN_SUCCESS) { + assert(object == VM_PAGE_OBJECT(m)); + *physpage_p = VM_PAGE_GET_PHYS_PAGE(m); + if (prot & VM_PROT_WRITE) { + vm_object_lock_assert_exclusive(object); + m->vmp_dirty = TRUE; + } + } else { + *physpage_p = 0; + } + } + PAGE_WAKEUP_DONE(m); UNLOCK_AND_DEALLOCATE; @@ -4150,37 +5576,38 @@ done: * Release a page used by vm_fault_copy. */ -void +static void vm_fault_copy_cleanup( vm_page_t page, vm_page_t top_page) { - vm_object_t object = page->object; + vm_object_t object = VM_PAGE_OBJECT(page); vm_object_lock(object); PAGE_WAKEUP_DONE(page); - if (!page->active && !page->inactive && !page->throttled) { + if ( !VM_PAGE_PAGEABLE(page)) { vm_page_lockspin_queues(); - if (!page->active && !page->inactive && !page->throttled) + if ( !VM_PAGE_PAGEABLE(page)) { vm_page_activate(page); + } vm_page_unlock_queues(); } vm_fault_cleanup(object, top_page); } -void +static void vm_fault_copy_dst_cleanup( vm_page_t page) { vm_object_t object; if (page != VM_PAGE_NULL) { - object = page->object; + object = VM_PAGE_OBJECT(page); vm_object_lock(object); vm_page_lockspin_queues(); - vm_page_unwire(page); + vm_page_unwire(page, TRUE); vm_page_unlock_queues(); - vm_object_paging_end(object); + vm_object_paging_end(object); vm_object_unlock(object); } } @@ -4224,7 +5651,7 @@ vm_fault_copy( int interruptible) { vm_page_t result_page; - + vm_page_t src_page; vm_page_t src_top_page; vm_prot_t src_prot; @@ -4235,12 +5662,13 @@ vm_fault_copy( vm_map_size_t amount_left; vm_object_t old_copy_object; + vm_object_t result_page_object = NULL; kern_return_t error = 0; vm_fault_return_t result; vm_map_size_t part_size; - struct vm_object_fault_info fault_info_src; - struct vm_object_fault_info fault_info_dst; + struct vm_object_fault_info fault_info_src = {}; + struct vm_object_fault_info fault_info_dst = {}; /* * In order not to confuse the clustered pageins, align @@ -4257,18 +5685,14 @@ vm_fault_copy( fault_info_src.interruptible = interruptible; fault_info_src.behavior = VM_BEHAVIOR_SEQUENTIAL; - fault_info_src.user_tag = 0; fault_info_src.lo_offset = vm_object_trunc_page(src_offset); fault_info_src.hi_offset = fault_info_src.lo_offset + amount_left; - fault_info_src.no_cache = FALSE; fault_info_src.stealth = TRUE; fault_info_dst.interruptible = interruptible; fault_info_dst.behavior = VM_BEHAVIOR_SEQUENTIAL; - fault_info_dst.user_tag = 0; fault_info_dst.lo_offset = vm_object_trunc_page(dst_offset); fault_info_dst.hi_offset = fault_info_dst.lo_offset + amount_left; - fault_info_dst.no_cache = FALSE; fault_info_dst.stealth = TRUE; do { /* while (amount_left > 0) */ @@ -4286,19 +5710,20 @@ vm_fault_copy( vm_object_lock(dst_object); vm_object_paging_begin(dst_object); - if (amount_left > (vm_size_t) -1) { - /* 32-bit overflow */ - fault_info_dst.cluster_size = (vm_size_t) (0 - PAGE_SIZE); - } else { - fault_info_dst.cluster_size = (vm_size_t) amount_left; - assert(fault_info_dst.cluster_size == amount_left); + /* cap cluster size at maximum UPL size */ + upl_size_t cluster_size; + if (os_convert_overflow(amount_left, &cluster_size)) { + cluster_size = 0 - (upl_size_t)PAGE_SIZE; } + fault_info_dst.cluster_size = cluster_size; XPR(XPR_VM_FAULT,"vm_fault_copy -> vm_fault_page\n",0,0,0,0,0); + dst_page = VM_PAGE_NULL; result = vm_fault_page(dst_object, vm_object_trunc_page(dst_offset), VM_PROT_WRITE|VM_PROT_READ, FALSE, + FALSE, /* page not looked up */ &dst_prot, &dst_page, &dst_top_page, (int *)0, &error, @@ -4331,7 +5756,8 @@ vm_fault_copy( } assert ((dst_prot & VM_PROT_WRITE) != VM_PROT_NONE); - old_copy_object = dst_page->object->copy; + assert(dst_object == VM_PAGE_OBJECT(dst_page)); + old_copy_object = dst_object->copy; /* * There exists the possiblity that the source and @@ -4340,16 +5766,16 @@ vm_fault_copy( * same, the call to vm_fault_page() for the * destination page will deadlock. To prevent this we * wire the page so we can drop busy without having - * the page daemon steal the page. We clean up the + * the page daemon steal the page. We clean up the * top page but keep the paging reference on the object * holding the dest page so it doesn't go away. */ vm_page_lockspin_queues(); - vm_page_wire(dst_page); + vm_page_wire(dst_page, VM_KERN_MEMORY_OSFMK, TRUE); vm_page_unlock_queues(); PAGE_WAKEUP_DONE(dst_page); - vm_object_unlock(dst_page->object); + vm_object_unlock(dst_object); if (dst_top_page != VM_PAGE_NULL) { vm_object_lock(dst_object); @@ -4378,22 +5804,22 @@ vm_fault_copy( src_prot = VM_PROT_READ; vm_object_paging_begin(src_object); - if (amount_left > (vm_size_t) -1) { - /* 32-bit overflow */ - fault_info_src.cluster_size = (vm_size_t) (0 - PAGE_SIZE); - } else { - fault_info_src.cluster_size = (vm_size_t) amount_left; - assert(fault_info_src.cluster_size == amount_left); + /* cap cluster size at maximum UPL size */ + if (os_convert_overflow(amount_left, &cluster_size)) { + cluster_size = 0 - (upl_size_t)PAGE_SIZE; } + fault_info_src.cluster_size = cluster_size; XPR(XPR_VM_FAULT, "vm_fault_copy(2) -> vm_fault_page\n", 0,0,0,0,0); + result_page = VM_PAGE_NULL; result = vm_fault_page( - src_object, + src_object, vm_object_trunc_page(src_offset), VM_PROT_READ, FALSE, - &src_prot, + FALSE, /* page not looked up */ + &src_prot, &result_page, &src_top_page, (int *)0, &error, FALSE, FALSE, &fault_info_src); @@ -4427,32 +5853,36 @@ vm_fault_copy( "vm_fault_page()\n", result); } - + result_page_object = VM_PAGE_OBJECT(result_page); assert((src_top_page == VM_PAGE_NULL) == - (result_page->object == src_object)); + (result_page_object == src_object)); } assert ((src_prot & VM_PROT_READ) != VM_PROT_NONE); - vm_object_unlock(result_page->object); + vm_object_unlock(result_page_object); } + vm_map_lock_read(dst_map); + if (!vm_map_verify(dst_map, dst_version)) { + vm_map_unlock_read(dst_map); if (result_page != VM_PAGE_NULL && src_page != dst_page) vm_fault_copy_cleanup(result_page, src_top_page); vm_fault_copy_dst_cleanup(dst_page); break; } + assert(dst_object == VM_PAGE_OBJECT(dst_page)); - vm_object_lock(dst_page->object); + vm_object_lock(dst_object); - if (dst_page->object->copy != old_copy_object) { - vm_object_unlock(dst_page->object); - vm_map_verify_done(dst_map, dst_version); + if (dst_object->copy != old_copy_object) { + vm_object_unlock(dst_object); + vm_map_unlock_read(dst_map); if (result_page != VM_PAGE_NULL && src_page != dst_page) vm_fault_copy_cleanup(result_page, src_top_page); vm_fault_copy_dst_cleanup(dst_page); break; } - vm_object_unlock(dst_page->object); + vm_object_unlock(dst_object); /* * Copy the page, and note that it is dirty @@ -4493,10 +5923,10 @@ vm_fault_copy( dst_page, (vm_offset_t) dst_po, (vm_size_t)part_size); - if(!dst_page->dirty){ + if(!dst_page->vmp_dirty){ vm_object_lock(dst_object); - dst_page->dirty = TRUE; - vm_object_unlock(dst_page->object); + SET_PAGE_DIRTY(dst_page, TRUE); + vm_object_unlock(dst_object); } } @@ -4506,11 +5936,14 @@ vm_fault_copy( if (result_page == VM_PAGE_NULL) vm_page_zero_fill(dst_page); else{ + vm_object_lock(result_page_object); vm_page_copy(result_page, dst_page); - if(!dst_page->dirty){ + vm_object_unlock(result_page_object); + + if(!dst_page->vmp_dirty){ vm_object_lock(dst_object); - dst_page->dirty = TRUE; - vm_object_unlock(dst_page->object); + SET_PAGE_DIRTY(dst_page, TRUE); + vm_object_unlock(dst_object); } } @@ -4520,7 +5953,7 @@ vm_fault_copy( * Unlock everything, and return */ - vm_map_verify_done(dst_map, dst_version); + vm_map_unlock_read(dst_map); if (result_page != VM_PAGE_NULL && src_page != dst_page) vm_fault_copy_cleanup(result_page, src_top_page); @@ -4534,7 +5967,7 @@ vm_fault_copy( RETURN(KERN_SUCCESS); #undef RETURN - /*NOTREACHED*/ + /*NOTREACHED*/ } #if VM_FAULT_CLASSIFY @@ -4567,15 +6000,15 @@ vm_fault_classify(vm_object_t object, while (TRUE) { m = vm_page_lookup(object, offset); - if (m != VM_PAGE_NULL) { - if (m->busy || m->error || m->restart || m->absent) { + if (m != VM_PAGE_NULL) { + if (m->vmp_busy || m->vmp_error || m->vmp_restart || m->vmp_absent) { type = VM_FAULT_TYPE_OTHER; break; } if (((fault_type & VM_PROT_WRITE) == 0) || ((level == 0) && object->copy == VM_OBJECT_NULL)) { type = VM_FAULT_TYPE_MAP_IN; - break; + break; } type = VM_FAULT_TYPE_COPY; break; @@ -4590,7 +6023,7 @@ vm_fault_classify(vm_object_t object, break; } - offset += object->shadow_offset; + offset += object->vo_shadow_offset; object = object->shadow; level++; continue; @@ -4622,29 +6055,134 @@ vm_fault_classify_init(void) } #endif /* VM_FAULT_CLASSIFY */ +vm_offset_t +kdp_lightweight_fault(vm_map_t map, vm_offset_t cur_target_addr) +{ + vm_map_entry_t entry; + vm_object_t object; + vm_offset_t object_offset; + vm_page_t m; + int compressor_external_state, compressed_count_delta; + int compressor_flags = (C_DONT_BLOCK | C_KEEP | C_KDP); + int my_fault_type = VM_PROT_READ; + kern_return_t kr; -extern int cs_validation; + if (not_in_kdp) { + panic("kdp_lightweight_fault called from outside of debugger context"); + } -void -vm_page_validate_cs_mapped( - vm_page_t page, - const void *kaddr) -{ - vm_object_t object; - vm_object_offset_t offset; - kern_return_t kr; - memory_object_t pager; - void *blobs; - boolean_t validated, tainted; + assert(map != VM_MAP_NULL); - assert(page->busy); - vm_object_lock_assert_exclusive(page->object); + assert((cur_target_addr & PAGE_MASK) == 0); + if ((cur_target_addr & PAGE_MASK) != 0) { + return 0; + } - if (!cs_validation) { - return; + if (kdp_lck_rw_lock_is_acquired_exclusive(&map->lock)) { + return 0; + } + + if (!vm_map_lookup_entry(map, cur_target_addr, &entry)) { + return 0; + } + + if (entry->is_sub_map) { + return 0; + } + + object = VME_OBJECT(entry); + if (object == VM_OBJECT_NULL) { + return 0; + } + + object_offset = cur_target_addr - entry->vme_start + VME_OFFSET(entry); + + while (TRUE) { + if (kdp_lck_rw_lock_is_acquired_exclusive(&object->Lock)) { + return 0; + } + + if (object->pager_created && (object->paging_in_progress || + object->activity_in_progress)) { + return 0; + } + + m = kdp_vm_page_lookup(object, object_offset); + + if (m != VM_PAGE_NULL) { + + if ((object->wimg_bits & VM_WIMG_MASK) != VM_WIMG_DEFAULT) { + return 0; + } + + if (m->vmp_laundry || m->vmp_busy || m->vmp_free_when_done || m->vmp_absent || m->vmp_error || m->vmp_cleaning || + m->vmp_overwriting || m->vmp_restart || m->vmp_unusual) { + return 0; + } + + assert(!m->vmp_private); + if (m->vmp_private) { + return 0; + } + + assert(!m->vmp_fictitious); + if (m->vmp_fictitious) { + return 0; + } + + assert(m->vmp_q_state != VM_PAGE_USED_BY_COMPRESSOR); + if (m->vmp_q_state == VM_PAGE_USED_BY_COMPRESSOR) { + return 0; + } + + return ptoa(VM_PAGE_GET_PHYS_PAGE(m)); + } + + compressor_external_state = VM_EXTERNAL_STATE_UNKNOWN; + + if (object->pager_created && MUST_ASK_PAGER(object, object_offset, compressor_external_state)) { + if (compressor_external_state == VM_EXTERNAL_STATE_EXISTS) { + kr = vm_compressor_pager_get(object->pager, (object_offset + object->paging_offset), + kdp_compressor_decompressed_page_ppnum, &my_fault_type, + compressor_flags, &compressed_count_delta); + if (kr == KERN_SUCCESS) { + return kdp_compressor_decompressed_page_paddr; + } else { + return 0; + } + } + } + + if (object->shadow == VM_OBJECT_NULL) { + return 0; + } + + object_offset += object->vo_shadow_offset; + object = object->shadow; } - if (page->wpmapped && !page->cs_tainted) { +} + +/* + * vm_page_validate_cs_fast(): + * Performs a few quick checks to determine if the page's code signature + * really needs to be fully validated. It could: + * 1. have been modified (i.e. automatically tainted), + * 2. have already been validated, + * 3. have already been found to be tainted, + * 4. no longer have a backing store. + * Returns FALSE if the page needs to be fully validated. + */ +static boolean_t +vm_page_validate_cs_fast( + vm_page_t page) +{ + vm_object_t object; + + object = VM_PAGE_OBJECT(page); + vm_object_lock_assert_held(object); + + if (page->vmp_wpmapped && !page->vmp_cs_tainted) { /* * This page was mapped for "write" access sometime in the * past and could still be modifiable in the future. @@ -4652,59 +6190,120 @@ vm_page_validate_cs_mapped( * [ If the page was already found to be "tainted", no * need to re-validate. ] */ - page->cs_validated = TRUE; - page->cs_tainted = TRUE; + vm_object_lock_assert_exclusive(object); + page->vmp_cs_validated = TRUE; + page->vmp_cs_tainted = TRUE; if (cs_debug) { - printf("CODESIGNING: vm_page_validate_cs: " + printf("CODESIGNING: %s: " "page %p obj %p off 0x%llx " "was modified\n", - page, page->object, page->offset); + __FUNCTION__, + page, object, page->vmp_offset); } vm_cs_validated_dirtied++; } - if (page->cs_validated) { - return; + if (page->vmp_cs_validated || page->vmp_cs_tainted) { + return TRUE; } + vm_object_lock_assert_exclusive(object); - vm_cs_validates++; +#if CHECK_CS_VALIDATION_BITMAP + kern_return_t kr; - object = page->object; - assert(object->code_signed); - offset = page->offset; + kr = vnode_pager_cs_check_validation_bitmap( + object->pager, + page->vmp_offset + object->paging_offset, + CS_BITMAP_CHECK); + if (kr == KERN_SUCCESS) { + page->vmp_cs_validated = TRUE; + page->vmp_cs_tainted = FALSE; + vm_cs_bitmap_validated++; + return TRUE; + } +#endif /* CHECK_CS_VALIDATION_BITMAP */ if (!object->alive || object->terminating || object->pager == NULL) { /* * The object is terminating and we don't have its pager * so we can't validate the data... */ - return; + return TRUE; } + + /* we need to really validate this page */ + vm_object_lock_assert_exclusive(object); + return FALSE; +} + +void +vm_page_validate_cs_mapped_slow( + vm_page_t page, + const void *kaddr) +{ + vm_object_t object; + memory_object_offset_t mo_offset; + memory_object_t pager; + struct vnode *vnode; + boolean_t validated; + unsigned tainted; + + assert(page->vmp_busy); + object = VM_PAGE_OBJECT(page); + vm_object_lock_assert_exclusive(object); + + vm_cs_validates++; + /* * Since we get here to validate a page that was brought in by * the pager, we know that this pager is all setup and ready * by now. */ + assert(object->code_signed); assert(!object->internal); assert(object->pager != NULL); assert(object->pager_ready); pager = object->pager; assert(object->paging_in_progress); - kr = vnode_pager_get_object_cs_blobs(pager, &blobs); - if (kr != KERN_SUCCESS) { - blobs = NULL; - } + vnode = vnode_pager_lookup_vnode(pager); + mo_offset = page->vmp_offset + object->paging_offset; /* verify the SHA1 hash for this page */ - validated = cs_validate_page(blobs, - offset + object->paging_offset, - (const void *)kaddr, - &tainted); - - page->cs_validated = validated; + tainted = 0; + validated = cs_validate_range(vnode, + pager, + mo_offset, + (const void *)((const char *)kaddr), + PAGE_SIZE_64, + &tainted); + + if (tainted & CS_VALIDATE_TAINTED) { + page->vmp_cs_tainted = TRUE; + } + if (tainted & CS_VALIDATE_NX) { + page->vmp_cs_nx = TRUE; + } if (validated) { - page->cs_tainted = tainted; + page->vmp_cs_validated = TRUE; + } + +#if CHECK_CS_VALIDATION_BITMAP + if (page->vmp_cs_validated && !page->vmp_cs_tainted) { + vnode_pager_cs_check_validation_bitmap(object->pager, + mo_offset, + CS_BITMAP_SET); + } +#endif /* CHECK_CS_VALIDATION_BITMAP */ +} + +void +vm_page_validate_cs_mapped( + vm_page_t page, + const void *kaddr) +{ + if (!vm_page_validate_cs_fast(page)) { + vm_page_validate_cs_mapped_slow(page, kaddr); } } @@ -4719,50 +6318,25 @@ vm_page_validate_cs( vm_offset_t kaddr; kern_return_t kr; boolean_t busy_page; + boolean_t need_unmap; - vm_object_lock_assert_held(page->object); - - if (!cs_validation) { - return; - } - - if (page->wpmapped && !page->cs_tainted) { - vm_object_lock_assert_exclusive(page->object); - - /* - * This page was mapped for "write" access sometime in the - * past and could still be modifiable in the future. - * Consider it tainted. - * [ If the page was already found to be "tainted", no - * need to re-validate. ] - */ - page->cs_validated = TRUE; - page->cs_tainted = TRUE; - if (cs_debug) { - printf("CODESIGNING: vm_page_validate_cs: " - "page %p obj %p off 0x%llx " - "was modified\n", - page, page->object, page->offset); - } - vm_cs_validated_dirtied++; - } + object = VM_PAGE_OBJECT(page); + vm_object_lock_assert_held(object); - if (page->cs_validated) { + if (vm_page_validate_cs_fast(page)) { return; } + vm_object_lock_assert_exclusive(object); - vm_object_lock_assert_exclusive(page->object); - - object = page->object; assert(object->code_signed); - offset = page->offset; + offset = page->vmp_offset; - busy_page = page->busy; + busy_page = page->vmp_busy; if (!busy_page) { /* keep page busy while we map (and unlock) the VM object */ - page->busy = TRUE; + page->vmp_busy = TRUE; } - + /* * Take a paging reference on the VM object * to protect it from collapse or bypass, @@ -4771,31 +6345,33 @@ vm_page_validate_cs( vm_object_paging_begin(object); /* map the page in the kernel address space */ - koffset = 0; ksize = PAGE_SIZE_64; - kr = vm_paging_map_object(&koffset, - page, + koffset = 0; + need_unmap = FALSE; + kr = vm_paging_map_object(page, object, offset, - &ksize, VM_PROT_READ, - FALSE); /* can't unlock object ! */ + FALSE, /* can't unlock object ! */ + &ksize, + &koffset, + &need_unmap); if (kr != KERN_SUCCESS) { - panic("vm_page_validate_cs: could not map page: 0x%x\n", kr); + panic("%s: could not map page: 0x%x\n", __FUNCTION__, kr); } kaddr = CAST_DOWN(vm_offset_t, koffset); /* validate the mapped page */ - vm_page_validate_cs_mapped(page, (const void *) kaddr); + vm_page_validate_cs_mapped_slow(page, (const void *) kaddr); - assert(page->busy); - assert(object == page->object); + assert(page->vmp_busy); + assert(object == VM_PAGE_OBJECT(page)); vm_object_lock_assert_exclusive(object); if (!busy_page) { PAGE_WAKEUP_DONE(page); } - if (koffset != 0) { + if (need_unmap) { /* unmap the map from the kernel address space */ vm_paging_unmap_object(object, koffset, koffset + ksize); koffset = 0; @@ -4804,3 +6380,166 @@ vm_page_validate_cs( } vm_object_paging_end(object); } + +void +vm_page_validate_cs_mapped_chunk( + vm_page_t page, + const void *kaddr, + vm_offset_t chunk_offset, + vm_size_t chunk_size, + boolean_t *validated_p, + unsigned *tainted_p) +{ + vm_object_t object; + vm_object_offset_t offset, offset_in_page; + memory_object_t pager; + struct vnode *vnode; + boolean_t validated; + unsigned tainted; + + *validated_p = FALSE; + *tainted_p = 0; + + assert(page->vmp_busy); + object = VM_PAGE_OBJECT(page); + vm_object_lock_assert_exclusive(object); + + assert(object->code_signed); + offset = page->vmp_offset; + + if (!object->alive || object->terminating || object->pager == NULL) { + /* + * The object is terminating and we don't have its pager + * so we can't validate the data... + */ + return; + } + /* + * Since we get here to validate a page that was brought in by + * the pager, we know that this pager is all setup and ready + * by now. + */ + assert(!object->internal); + assert(object->pager != NULL); + assert(object->pager_ready); + + pager = object->pager; + assert(object->paging_in_progress); + vnode = vnode_pager_lookup_vnode(pager); + + /* verify the signature for this chunk */ + offset_in_page = chunk_offset; + assert(offset_in_page < PAGE_SIZE); + + tainted = 0; + validated = cs_validate_range(vnode, + pager, + (object->paging_offset + + offset + + offset_in_page), + (const void *)((const char *)kaddr + + offset_in_page), + chunk_size, + &tainted); + if (validated) { + *validated_p = TRUE; + } + if (tainted) { + *tainted_p = tainted; + } +} + +static void vm_rtfrecord_lock(void) { + lck_spin_lock(&vm_rtfr_slock); +} + +static void vm_rtfrecord_unlock(void) { + lck_spin_unlock(&vm_rtfr_slock); +} + +unsigned int vmrtfaultinfo_bufsz(void) { + return (vmrtf_num_records * sizeof(vm_rtfault_record_t)); +} + +#include + +static void vm_record_rtfault(thread_t cthread, uint64_t fstart, vm_map_offset_t fault_vaddr, int type_of_fault) { + uint64_t fend = mach_continuous_time(); + + uint64_t cfpc = 0; + uint64_t ctid = cthread->thread_id; + uint64_t cupid = get_current_unique_pid(); + + uintptr_t bpc = 0; + uint32_t bfrs = 0; + bool u64 = false; + + /* Capture a single-frame backtrace; this extracts just the program + * counter at the point of the fault into "bpc", and should perform no + * further user stack traversals, thus avoiding copyin()s and further + * faults. + */ + int btr = backtrace_thread_user(cthread, &bpc, 1U, &bfrs, &u64); + + if ((btr == 0) && (bfrs > 0)) { + cfpc = bpc; + } + + assert((fstart != 0) && fend >= fstart); + vm_rtfrecord_lock(); + assert(vmrtfrs.vmrtfr_curi <= vmrtfrs.vmrtfr_maxi); + + vmrtfrs.vmrtf_total++; + vm_rtfault_record_t *cvmr = &vmrtfrs.vm_rtf_records[vmrtfrs.vmrtfr_curi++]; + + cvmr->rtfabstime = fstart; + cvmr->rtfduration = fend - fstart; + cvmr->rtfaddr = fault_vaddr; + cvmr->rtfpc = cfpc; + cvmr->rtftype = type_of_fault; + cvmr->rtfupid = cupid; + cvmr->rtftid = ctid; + + if (vmrtfrs.vmrtfr_curi > vmrtfrs.vmrtfr_maxi) { + vmrtfrs.vmrtfr_curi = 0; + } + + vm_rtfrecord_unlock(); +} + +int vmrtf_extract(uint64_t cupid, __unused boolean_t isroot, int vrecordsz, void *vrecords, int *vmrtfrv) { + vm_rtfault_record_t *cvmrd = vrecords; + size_t residue = vrecordsz; + int numextracted = 0; + boolean_t early_exit = FALSE; + + vm_rtfrecord_lock(); + + for (int vmfi = 0; vmfi <= vmrtfrs.vmrtfr_maxi; vmfi++) { + + if (residue < sizeof(vm_rtfault_record_t)) { + early_exit = TRUE; + break; + } + + if (vmrtfrs.vm_rtf_records[vmfi].rtfupid != cupid) { +#if DEVELOPMENT || DEBUG + if (isroot == FALSE) { + continue; + } +#else + continue; +#endif /* DEVDEBUG */ + } + + *cvmrd = vmrtfrs.vm_rtf_records[vmfi]; + cvmrd++; + residue -= sizeof(vm_rtfault_record_t); + numextracted++; + } + + vm_rtfrecord_unlock(); + + *vmrtfrv = numextracted; + return (early_exit); +}