X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/d7e50217d7adf6e52786a38bcaa4cd698cb9a79e..0b4c1975fb5e4eccf1012a35081f7e7799b81046:/osfmk/vm/vm_object.c
diff --git a/osfmk/vm/vm_object.c b/osfmk/vm/vm_object.c
index a0f8bff17..c71627a4f 100644
--- a/osfmk/vm/vm_object.c
+++ b/osfmk/vm/vm_object.c
@@ -1,16 +1,19 @@
/*
- * Copyright (c) 2000-2001 Apple Computer, Inc. All rights reserved.
+ * Copyright (c) 2000-2007 Apple Inc. All rights reserved.
*
- * @APPLE_LICENSE_HEADER_START@
- *
- * Copyright (c) 1999-2003 Apple Computer, 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
- * compliance with the License. Please obtain a copy of the License at
- * http://www.opensource.apple.com/apsl/ and read it before using this
- * file.
+ * compliance with the License. The rights granted to you under the License
+ * may not be used to create, or enable the creation or redistribution of,
+ * unlawful or unlicensed copies of an Apple operating system, or to
+ * circumvent, violate, or enable the circumvention or violation of, any
+ * terms of an Apple operating system software license agreement.
+ *
+ * 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
@@ -20,7 +23,7 @@
* Please see the License for the specific language governing rights and
* limitations under the License.
*
- * @APPLE_LICENSE_HEADER_END@
+ * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
*/
/*
* @OSF_COPYRIGHT@
@@ -59,11 +62,7 @@
* Virtual memory object module.
*/
-#ifdef MACH_BSD
-/* remove as part of compoenent support merge */
-extern int vnode_pager_workaround;
-#endif
-
+#include <debug.h>
#include <mach_pagemap.h>
#include <task_swapper.h>
@@ -72,7 +71,11 @@ extern int vnode_pager_workaround;
#include <mach/memory_object_default.h>
#include <mach/memory_object_control_server.h>
#include <mach/vm_param.h>
+
+#include <ipc/ipc_types.h>
#include <ipc/ipc_port.h>
+
+#include <kern/kern_types.h>
#include <kern/assert.h>
#include <kern/lock.h>
#include <kern/queue.h>
@@ -81,15 +84,20 @@ extern int vnode_pager_workaround;
#include <kern/host.h>
#include <kern/host_statistics.h>
#include <kern/processor.h>
+#include <kern/misc_protos.h>
+
#include <vm/memory_object.h>
#include <vm/vm_fault.h>
#include <vm/vm_map.h>
#include <vm/vm_object.h>
#include <vm/vm_page.h>
#include <vm/vm_pageout.h>
-#include <kern/misc_protos.h>
-
+#include <vm/vm_protos.h>
+#include <vm/vm_purgeable_internal.h>
+#if CONFIG_EMBEDDED
+#include <sys/kern_memorystatus.h>
+#endif
/*
* Virtual memory objects maintain the actual data
@@ -165,25 +173,12 @@ extern int vnode_pager_workaround;
*/
/* Forward declarations for internal functions. */
-static void _vm_object_allocate(
- vm_object_size_t size,
- vm_object_t object);
-
static kern_return_t vm_object_terminate(
vm_object_t object);
extern void vm_object_remove(
vm_object_t object);
-static vm_object_t vm_object_cache_trim(
- boolean_t called_from_vm_object_deallocate);
-
-static void vm_object_deactivate_all_pages(
- vm_object_t object);
-
-static void vm_object_abort_activity(
- vm_object_t object);
-
static kern_return_t vm_object_copy_call(
vm_object_t src_object,
vm_object_offset_t src_offset,
@@ -199,7 +194,8 @@ static void vm_object_do_bypass(
vm_object_t backing_object);
static void vm_object_release_pager(
- memory_object_t pager);
+ memory_object_t pager,
+ boolean_t hashed);
static zone_t vm_object_zone; /* vm backing store zone */
@@ -208,7 +204,8 @@ static zone_t vm_object_zone; /* vm backing store zone */
* memory object (kernel_object) to avoid wasting data structures.
*/
static struct vm_object kernel_object_store;
-__private_extern__ vm_object_t kernel_object = &kernel_object_store;
+vm_object_t kernel_object;
+
/*
* The submap object is used as a placeholder for vm_map_submap
@@ -228,6 +225,11 @@ static struct vm_object vm_submap_object_store;
*/
static struct vm_object vm_object_template;
+unsigned int vm_page_purged_wired = 0;
+unsigned int vm_page_purged_busy = 0;
+unsigned int vm_page_purged_others = 0;
+
+#if VM_OBJECT_CACHE
/*
* Virtual memory objects that are not referenced by
* any address maps, but that are allowed to persist
@@ -256,23 +258,41 @@ static struct vm_object vm_object_template;
* from the reference mechanism, so that the lock need
* not be held to make simple references.
*/
+static vm_object_t vm_object_cache_trim(
+ boolean_t called_from_vm_object_deallocate);
+
static queue_head_t vm_object_cached_list;
static int vm_object_cached_count=0;
static int vm_object_cached_high; /* highest # cached objects */
static int vm_object_cached_max = 512; /* may be patched*/
-static decl_mutex_data(,vm_object_cached_lock_data)
+static lck_mtx_t vm_object_cached_lock_data;
+static lck_mtx_ext_t vm_object_cached_lock_data_ext;
#define vm_object_cache_lock() \
- mutex_lock(&vm_object_cached_lock_data)
-#define vm_object_cache_lock_try() \
- mutex_try(&vm_object_cached_lock_data)
+ lck_mtx_lock(&vm_object_cached_lock_data)
+#define vm_object_cache_lock_try() \
+ lck_mtx_try_lock(&vm_object_cached_lock_data)
+#define vm_object_cache_lock_spin() \
+ lck_mtx_lock_spin(&vm_object_cached_lock_data)
#define vm_object_cache_unlock() \
- mutex_unlock(&vm_object_cached_lock_data)
+ lck_mtx_unlock(&vm_object_cached_lock_data)
+
+#endif /* VM_OBJECT_CACHE */
+
+
+static void vm_object_deactivate_all_pages(
+ vm_object_t object);
+
#define VM_OBJECT_HASH_COUNT 1024
+#define VM_OBJECT_HASH_LOCK_COUNT 512
+
+static lck_mtx_t vm_object_hashed_lock_data[VM_OBJECT_HASH_LOCK_COUNT];
+static lck_mtx_ext_t vm_object_hashed_lock_data_ext[VM_OBJECT_HASH_LOCK_COUNT];
+
static queue_head_t vm_object_hashtable[VM_OBJECT_HASH_COUNT];
-static struct zone *vm_object_hash_zone;
+static struct zone *vm_object_hash_zone;
struct vm_object_hash_entry {
queue_chain_t hash_link; /* hash chain link */
@@ -285,38 +305,81 @@ struct vm_object_hash_entry {
typedef struct vm_object_hash_entry *vm_object_hash_entry_t;
#define VM_OBJECT_HASH_ENTRY_NULL ((vm_object_hash_entry_t) 0)
-#define VM_OBJECT_HASH_SHIFT 8
+#define VM_OBJECT_HASH_SHIFT 5
#define vm_object_hash(pager) \
- ((((unsigned)pager) >> VM_OBJECT_HASH_SHIFT) % VM_OBJECT_HASH_COUNT)
+ ((int)((((uintptr_t)pager) >> VM_OBJECT_HASH_SHIFT) % VM_OBJECT_HASH_COUNT))
+
+#define vm_object_lock_hash(pager) \
+ ((int)((((uintptr_t)pager) >> VM_OBJECT_HASH_SHIFT) % VM_OBJECT_HASH_LOCK_COUNT))
+
+void vm_object_hash_entry_free(
+ vm_object_hash_entry_t entry);
+
+static void vm_object_reap(vm_object_t object);
+static void vm_object_reap_async(vm_object_t object);
+static void vm_object_reaper_thread(void);
+
+static lck_mtx_t vm_object_reaper_lock_data;
+static lck_mtx_ext_t vm_object_reaper_lock_data_ext;
+
+static queue_head_t vm_object_reaper_queue; /* protected by vm_object_reaper_lock() */
+unsigned int vm_object_reap_count = 0;
+unsigned int vm_object_reap_count_async = 0;
+
+#define vm_object_reaper_lock() \
+ lck_mtx_lock(&vm_object_reaper_lock_data)
+#define vm_object_reaper_lock_spin() \
+ lck_mtx_lock_spin(&vm_object_reaper_lock_data)
+#define vm_object_reaper_unlock() \
+ lck_mtx_unlock(&vm_object_reaper_lock_data)
+
+
+
+static lck_mtx_t *
+vm_object_hash_lock_spin(
+ memory_object_t pager)
+{
+ int index;
+
+ index = vm_object_lock_hash(pager);
+
+ lck_mtx_lock_spin(&vm_object_hashed_lock_data[index]);
+
+ return (&vm_object_hashed_lock_data[index]);
+}
+
+static void
+vm_object_hash_unlock(lck_mtx_t *lck)
+{
+ lck_mtx_unlock(lck);
+}
+
/*
* vm_object_hash_lookup looks up a pager in the hashtable
* and returns the corresponding entry, with optional removal.
*/
-
static vm_object_hash_entry_t
vm_object_hash_lookup(
memory_object_t pager,
boolean_t remove_entry)
{
- register queue_t bucket;
- register vm_object_hash_entry_t entry;
+ queue_t bucket;
+ vm_object_hash_entry_t entry;
bucket = &vm_object_hashtable[vm_object_hash(pager)];
entry = (vm_object_hash_entry_t)queue_first(bucket);
while (!queue_end(bucket, (queue_entry_t)entry)) {
- if (entry->pager == pager && !remove_entry)
- return(entry);
- else if (entry->pager == pager) {
- queue_remove(bucket, entry,
- vm_object_hash_entry_t, hash_link);
+ if (entry->pager == pager) {
+ if (remove_entry) {
+ queue_remove(bucket, entry,
+ vm_object_hash_entry_t, hash_link);
+ }
return(entry);
}
-
entry = (vm_object_hash_entry_t)queue_next(&entry->hash_link);
}
-
return(VM_OBJECT_HASH_ENTRY_NULL);
}
@@ -327,13 +390,17 @@ vm_object_hash_lookup(
static void
vm_object_hash_insert(
- vm_object_hash_entry_t entry)
+ vm_object_hash_entry_t entry,
+ vm_object_t object)
{
- register queue_t bucket;
+ queue_t bucket;
bucket = &vm_object_hashtable[vm_object_hash(entry->pager)];
queue_enter(bucket, entry, vm_object_hash_entry_t, hash_link);
+
+ entry->object = object;
+ object->hashed = TRUE;
}
static vm_object_hash_entry_t
@@ -354,7 +421,7 @@ void
vm_object_hash_entry_free(
vm_object_hash_entry_t entry)
{
- zfree(vm_object_hash_zone, (vm_offset_t)entry);
+ zfree(vm_object_hash_zone, entry);
}
/*
@@ -363,21 +430,21 @@ vm_object_hash_entry_free(
* Returns a new object with the given size.
*/
-static void
+__private_extern__ void
_vm_object_allocate(
vm_object_size_t size,
vm_object_t object)
{
XPR(XPR_VM_OBJECT,
"vm_object_allocate, object 0x%X size 0x%X\n",
- (integer_t)object, size, 0,0,0);
+ object, size, 0,0,0);
*object = vm_object_template;
queue_init(&object->memq);
queue_init(&object->msr_q);
-#ifdef UBC_DEBUG
+#if UPL_DEBUG
queue_init(&object->uplq);
-#endif /* UBC_DEBUG */
+#endif /* UPL_DEBUG */
vm_object_lock_init(object);
object->size = size;
}
@@ -398,6 +465,12 @@ vm_object_allocate(
return object;
}
+
+lck_grp_t vm_object_lck_grp;
+lck_grp_attr_t vm_object_lck_grp_attr;
+lck_attr_t vm_object_lck_attr;
+lck_attr_t kernel_object_lck_attr;
+
/*
* vm_object_bootstrap:
*
@@ -406,51 +479,83 @@ vm_object_allocate(
__private_extern__ void
vm_object_bootstrap(void)
{
- register i;
+ register int i;
vm_object_zone = zinit((vm_size_t) sizeof(struct vm_object),
- round_page_32(512*1024),
- round_page_32(12*1024),
+ round_page(512*1024),
+ round_page(12*1024),
"vm objects");
+ zone_change(vm_object_zone, Z_NOENCRYPT, TRUE);
+
+ vm_object_init_lck_grp();
+#if VM_OBJECT_CACHE
queue_init(&vm_object_cached_list);
- mutex_init(&vm_object_cached_lock_data, ETAP_VM_OBJ_CACHE);
+
+ lck_mtx_init_ext(&vm_object_cached_lock_data,
+ &vm_object_cached_lock_data_ext,
+ &vm_object_lck_grp,
+ &vm_object_lck_attr);
+#endif
+ queue_init(&vm_object_reaper_queue);
+
+ for (i = 0; i < VM_OBJECT_HASH_LOCK_COUNT; i++) {
+ lck_mtx_init_ext(&vm_object_hashed_lock_data[i],
+ &vm_object_hashed_lock_data_ext[i],
+ &vm_object_lck_grp,
+ &vm_object_lck_attr);
+ }
+ lck_mtx_init_ext(&vm_object_reaper_lock_data,
+ &vm_object_reaper_lock_data_ext,
+ &vm_object_lck_grp,
+ &vm_object_lck_attr);
vm_object_hash_zone =
zinit((vm_size_t) sizeof (struct vm_object_hash_entry),
- round_page_32(512*1024),
- round_page_32(12*1024),
+ round_page(512*1024),
+ round_page(12*1024),
"vm object hash entries");
+ zone_change(vm_object_hash_zone, Z_NOENCRYPT, TRUE);
for (i = 0; i < VM_OBJECT_HASH_COUNT; i++)
queue_init(&vm_object_hashtable[i]);
+
/*
* Fill in a template object, for quick initialization
*/
/* memq; Lock; init after allocation */
+ vm_object_template.memq.prev = NULL;
+ vm_object_template.memq.next = NULL;
+#if 0
+ /*
+ * We can't call vm_object_lock_init() here because that will
+ * allocate some memory and VM is not fully initialized yet.
+ * The lock will be initialized for each allocated object in
+ * _vm_object_allocate(), so we don't need to initialize it in
+ * the vm_object_template.
+ */
+ vm_object_lock_init(&vm_object_template);
+#endif
vm_object_template.size = 0;
- vm_object_template.frozen_size = 0;
+ vm_object_template.memq_hint = VM_PAGE_NULL;
vm_object_template.ref_count = 1;
#if TASK_SWAPPER
vm_object_template.res_count = 1;
#endif /* TASK_SWAPPER */
vm_object_template.resident_page_count = 0;
+ vm_object_template.wired_page_count = 0;
+ vm_object_template.reusable_page_count = 0;
vm_object_template.copy = VM_OBJECT_NULL;
vm_object_template.shadow = VM_OBJECT_NULL;
vm_object_template.shadow_offset = (vm_object_offset_t) 0;
- vm_object_template.cow_hint = 0;
- vm_object_template.true_share = FALSE;
-
vm_object_template.pager = MEMORY_OBJECT_NULL;
vm_object_template.paging_offset = 0;
- vm_object_template.pager_request = PAGER_REQUEST_NULL;
- /* msr_q; init after allocation */
-
+ vm_object_template.pager_control = MEMORY_OBJECT_CONTROL_NULL;
vm_object_template.copy_strategy = MEMORY_OBJECT_COPY_SYMMETRIC;
- vm_object_template.absent_count = 0;
vm_object_template.paging_in_progress = 0;
+ vm_object_template.activity_in_progress = 0;
/* Begin bitfields */
vm_object_template.all_wanted = 0; /* all bits FALSE */
@@ -464,30 +569,60 @@ vm_object_bootstrap(void)
vm_object_template.private = FALSE;
vm_object_template.pageout = FALSE;
vm_object_template.alive = TRUE;
- vm_object_template.lock_in_progress = FALSE;
- vm_object_template.lock_restart = FALSE;
+ vm_object_template.purgable = VM_PURGABLE_DENY;
+ vm_object_template.shadowed = FALSE;
vm_object_template.silent_overwrite = FALSE;
vm_object_template.advisory_pageout = FALSE;
- vm_object_template.shadowed = FALSE;
+ vm_object_template.true_share = FALSE;
vm_object_template.terminating = FALSE;
+ vm_object_template.named = FALSE;
vm_object_template.shadow_severed = FALSE;
vm_object_template.phys_contiguous = FALSE;
vm_object_template.nophyscache = FALSE;
/* End bitfields */
- /* cache bitfields */
- vm_object_template.wimg_bits = VM_WIMG_DEFAULT;
-
- /* cached_list; init after allocation */
+ vm_object_template.cached_list.prev = NULL;
+ vm_object_template.cached_list.next = NULL;
+ vm_object_template.msr_q.prev = NULL;
+ vm_object_template.msr_q.next = NULL;
+
vm_object_template.last_alloc = (vm_object_offset_t) 0;
- vm_object_template.cluster_size = 0;
+ vm_object_template.sequential = (vm_object_offset_t) 0;
+ vm_object_template.pages_created = 0;
+ vm_object_template.pages_used = 0;
+
#if MACH_PAGEMAP
vm_object_template.existence_map = VM_EXTERNAL_NULL;
#endif /* MACH_PAGEMAP */
+ vm_object_template.cow_hint = ~(vm_offset_t)0;
#if MACH_ASSERT
vm_object_template.paging_object = VM_OBJECT_NULL;
#endif /* MACH_ASSERT */
+ /* cache bitfields */
+ vm_object_template.wimg_bits = VM_WIMG_DEFAULT;
+ vm_object_template.code_signed = FALSE;
+ vm_object_template.hashed = FALSE;
+ vm_object_template.transposed = FALSE;
+ vm_object_template.mapping_in_progress = FALSE;
+ vm_object_template.volatile_empty = FALSE;
+ vm_object_template.volatile_fault = FALSE;
+ vm_object_template.all_reusable = FALSE;
+ vm_object_template.blocked_access = FALSE;
+ vm_object_template.__object2_unused_bits = 0;
+#if UPL_DEBUG
+ vm_object_template.uplq.prev = NULL;
+ vm_object_template.uplq.next = NULL;
+#endif /* UPL_DEBUG */
+#ifdef VM_PIP_DEBUG
+ bzero(&vm_object_template.pip_holders,
+ sizeof (vm_object_template.pip_holders));
+#endif /* VM_PIP_DEBUG */
+
+ vm_object_template.objq.next=NULL;
+ vm_object_template.objq.prev=NULL;
+
+
/*
* Initialize the "kernel object"
*/
@@ -500,11 +635,11 @@ vm_object_bootstrap(void)
*/
#ifdef ppc
- _vm_object_allocate((vm_last_addr - VM_MIN_KERNEL_ADDRESS) + 1,
- kernel_object);
+ _vm_object_allocate(vm_last_addr + 1,
+ kernel_object);
#else
- _vm_object_allocate((VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) + 1,
- kernel_object);
+ _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1,
+ kernel_object);
#endif
kernel_object->copy_strategy = MEMORY_OBJECT_COPY_NONE;
@@ -515,11 +650,11 @@ vm_object_bootstrap(void)
vm_submap_object = &vm_submap_object_store;
#ifdef ppc
- _vm_object_allocate((vm_last_addr - VM_MIN_KERNEL_ADDRESS) + 1,
- vm_submap_object);
+ _vm_object_allocate(vm_last_addr + 1,
+ vm_submap_object);
#else
- _vm_object_allocate((VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) + 1,
- vm_submap_object);
+ _vm_object_allocate(VM_MAX_KERNEL_ADDRESS + 1,
+ vm_submap_object);
#endif
vm_submap_object->copy_strategy = MEMORY_OBJECT_COPY_NONE;
@@ -535,6 +670,23 @@ vm_object_bootstrap(void)
#endif /* MACH_PAGEMAP */
}
+void
+vm_object_reaper_init(void)
+{
+ kern_return_t kr;
+ thread_t thread;
+
+ kr = kernel_thread_start_priority(
+ (thread_continue_t) vm_object_reaper_thread,
+ NULL,
+ BASEPRI_PREEMPT - 1,
+ &thread);
+ if (kr != KERN_SUCCESS) {
+ panic("failed to launch vm_object_reaper_thread kr=0x%x", kr);
+ }
+ thread_deallocate(thread);
+}
+
__private_extern__ void
vm_object_init(void)
{
@@ -543,18 +695,26 @@ vm_object_init(void)
*/
}
-/* remove the typedef below when emergency work-around is taken out */
-typedef struct vnode_pager {
- memory_object_t pager;
- memory_object_t pager_handle; /* pager */
- memory_object_control_t control_handle; /* memory object's control handle */
- void *vnode_handle; /* vnode handle */
-} *vnode_pager_t;
+__private_extern__ void
+vm_object_init_lck_grp(void)
+{
+ /*
+ * initialze the vm_object lock world
+ */
+ lck_grp_attr_setdefault(&vm_object_lck_grp_attr);
+ lck_grp_init(&vm_object_lck_grp, "vm_object", &vm_object_lck_grp_attr);
+ lck_attr_setdefault(&vm_object_lck_attr);
+ lck_attr_setdefault(&kernel_object_lck_attr);
+ lck_attr_cleardebug(&kernel_object_lck_attr);
+}
+
+#if VM_OBJECT_CACHE
#define MIGHT_NOT_CACHE_SHADOWS 1
#if MIGHT_NOT_CACHE_SHADOWS
static int cache_shadows = TRUE;
#endif /* MIGHT_NOT_CACHE_SHADOWS */
+#endif
/*
* vm_object_deallocate:
@@ -567,26 +727,92 @@ static int cache_shadows = TRUE;
*
* No object may be locked.
*/
+unsigned long vm_object_deallocate_shared_successes = 0;
+unsigned long vm_object_deallocate_shared_failures = 0;
+unsigned long vm_object_deallocate_shared_swap_failures = 0;
__private_extern__ void
vm_object_deallocate(
register vm_object_t object)
{
- boolean_t retry_cache_trim = FALSE;
- vm_object_t shadow;
+#if VM_OBJECT_CACHE
+ boolean_t retry_cache_trim = FALSE;
+ uint32_t try_failed_count = 0;
+#endif
+ vm_object_t shadow = VM_OBJECT_NULL;
// if(object)dbgLog(object, object->ref_count, object->can_persist, 3); /* (TEST/DEBUG) */
// else dbgLog(object, 0, 0, 3); /* (TEST/DEBUG) */
+ if (object == VM_OBJECT_NULL)
+ return;
- while (object != VM_OBJECT_NULL) {
+ if (object == kernel_object) {
+ vm_object_lock_shared(object);
+
+ OSAddAtomic(-1, &object->ref_count);
+
+ if (object->ref_count == 0) {
+ panic("vm_object_deallocate: losing kernel_object\n");
+ }
+ vm_object_unlock(object);
+ return;
+ }
+
+ if (object->ref_count > 2 ||
+ (!object->named && object->ref_count > 1)) {
+ UInt32 original_ref_count;
+ volatile UInt32 *ref_count_p;
+ Boolean atomic_swap;
/*
- * The cache holds a reference (uncounted) to
- * the object; we must lock it before removing
- * the object.
+ * The object currently looks like it is not being
+ * kept alive solely by the reference we're about to release.
+ * Let's try and release our reference without taking
+ * all the locks we would need if we had to terminate the
+ * object (cache lock + exclusive object lock).
+ * Lock the object "shared" to make sure we don't race with
+ * anyone holding it "exclusive".
+ */
+ vm_object_lock_shared(object);
+ ref_count_p = (volatile UInt32 *) &object->ref_count;
+ original_ref_count = object->ref_count;
+ /*
+ * Test again as "ref_count" could have changed.
+ * "named" shouldn't change.
*/
+ if (original_ref_count > 2 ||
+ (!object->named && original_ref_count > 1)) {
+ atomic_swap = OSCompareAndSwap(
+ original_ref_count,
+ original_ref_count - 1,
+ (UInt32 *) &object->ref_count);
+ if (atomic_swap == FALSE) {
+ vm_object_deallocate_shared_swap_failures++;
+ }
+
+ } else {
+ atomic_swap = FALSE;
+ }
+ vm_object_unlock(object);
+
+ if (atomic_swap) {
+ /*
+ * ref_count was updated atomically !
+ */
+ vm_object_deallocate_shared_successes++;
+ return;
+ }
+
+ /*
+ * Someone else updated the ref_count at the same
+ * time and we lost the race. Fall back to the usual
+ * slow but safe path...
+ */
+ vm_object_deallocate_shared_failures++;
+ }
+
+ while (object != VM_OBJECT_NULL) {
- vm_object_cache_lock();
vm_object_lock(object);
assert(object->ref_count > 0);
@@ -603,15 +829,27 @@ vm_object_deallocate(
/* more mappers for this object */
if (pager != MEMORY_OBJECT_NULL) {
+ vm_object_mapping_wait(object, THREAD_UNINT);
+ vm_object_mapping_begin(object);
vm_object_unlock(object);
- vm_object_cache_unlock();
-
- memory_object_unmap(pager);
- vm_object_cache_lock();
+ memory_object_last_unmap(pager);
+
vm_object_lock(object);
- assert(object->ref_count > 0);
+ vm_object_mapping_end(object);
}
+ /*
+ * recheck the ref_count since we dropped the object lock
+ * to call 'memory_object_last_unmap'... it's possible
+ * additional references got taken and we only want
+ * to deactivate the pages if this 'named' object will only
+ * referenced by the backing pager once we drop our reference
+ * below
+ */
+ if (!object->terminating && object->ref_count == 2)
+ vm_object_deactivate_all_pages(object);
+
+ assert(object->ref_count > 0);
}
/*
@@ -627,15 +865,38 @@ vm_object_deallocate(
/* terminate again. */
if ((object->ref_count > 1) || object->terminating) {
+ vm_object_lock_assert_exclusive(object);
object->ref_count--;
vm_object_res_deallocate(object);
- vm_object_unlock(object);
- vm_object_cache_unlock();
+
+ if (object->ref_count == 1 &&
+ object->shadow != VM_OBJECT_NULL) {
+ /*
+ * There's only one reference left on this
+ * VM object. We can't tell if it's a valid
+ * one (from a mapping for example) or if this
+ * object is just part of a possibly stale and
+ * useless shadow chain.
+ * We would like to try and collapse it into
+ * its parent, but we don't have any pointers
+ * back to this parent object.
+ * But we can try and collapse this object with
+ * its own shadows, in case these are useless
+ * too...
+ * We can't bypass this object though, since we
+ * don't know if this last reference on it is
+ * meaningful or not.
+ */
+ vm_object_collapse(object, 0, FALSE);
+ }
+ vm_object_unlock(object);
+#if VM_OBJECT_CACHE
if (retry_cache_trim &&
((object = vm_object_cache_trim(TRUE)) !=
VM_OBJECT_NULL)) {
continue;
}
+#endif
return;
}
@@ -650,11 +911,12 @@ vm_object_deallocate(
VM_OBJECT_EVENT_INITIALIZED,
THREAD_UNINT);
vm_object_unlock(object);
- vm_object_cache_unlock();
+
thread_block(THREAD_CONTINUE_NULL);
continue;
}
+#if VM_OBJECT_CACHE
/*
* If this object can persist, then enter it in
* the cache. Otherwise, terminate it.
@@ -670,10 +932,12 @@ vm_object_deallocate(
* Now it is safe to decrement reference count,
* and to return if reference count is > 0.
*/
+
+ vm_object_lock_assert_exclusive(object);
if (--object->ref_count > 0) {
vm_object_res_deallocate(object);
vm_object_unlock(object);
- vm_object_cache_unlock();
+
if (retry_cache_trim &&
((object = vm_object_cache_trim(TRUE)) !=
VM_OBJECT_NULL)) {
@@ -702,16 +966,38 @@ vm_object_deallocate(
VM_OBJ_RES_DECR(object);
XPR(XPR_VM_OBJECT,
"vm_o_deallocate: adding %x to cache, queue = (%x, %x)\n",
- (integer_t)object,
- (integer_t)vm_object_cached_list.next,
- (integer_t)vm_object_cached_list.prev,0,0);
+ object,
+ vm_object_cached_list.next,
+ vm_object_cached_list.prev,0,0);
+
+
+ vm_object_unlock(object);
+
+ try_failed_count = 0;
+ for (;;) {
+ vm_object_cache_lock();
+
+ /*
+ * if we try to take a regular lock here
+ * we risk deadlocking against someone
+ * holding a lock on this object while
+ * trying to vm_object_deallocate a different
+ * object
+ */
+ if (vm_object_lock_try(object))
+ break;
+ vm_object_cache_unlock();
+ try_failed_count++;
+ mutex_pause(try_failed_count); /* wait a bit */
+ }
vm_object_cached_count++;
if (vm_object_cached_count > vm_object_cached_high)
vm_object_cached_high = vm_object_cached_count;
queue_enter(&vm_object_cached_list, object,
vm_object_t, cached_list);
vm_object_cache_unlock();
+
vm_object_deactivate_all_pages(object);
vm_object_unlock(object);
@@ -741,16 +1027,17 @@ vm_object_deallocate(
return;
}
retry_cache_trim = TRUE;
-
- } else {
+ } else
+#endif /* VM_OBJECT_CACHE */
+ {
/*
* This object is not cachable; terminate it.
*/
XPR(XPR_VM_OBJECT,
- "vm_o_deallocate: !cacheable 0x%X res %d paging_ops %d thread 0x%lX ref %d\n",
- (integer_t)object, object->resident_page_count,
- object->paging_in_progress,
- (natural_t)current_thread(),object->ref_count);
+ "vm_o_deallocate: !cacheable 0x%X res %d paging_ops %d thread 0x%p ref %d\n",
+ object, object->resident_page_count,
+ object->paging_in_progress,
+ (void *)current_thread(),object->ref_count);
VM_OBJ_RES_DECR(object); /* XXX ? */
/*
@@ -762,24 +1049,31 @@ vm_object_deallocate(
* a normal reference.
*/
shadow = object->pageout?VM_OBJECT_NULL:object->shadow;
- if(vm_object_terminate(object) != KERN_SUCCESS) {
+
+ if (vm_object_terminate(object) != KERN_SUCCESS) {
return;
}
if (shadow != VM_OBJECT_NULL) {
object = shadow;
continue;
}
+#if VM_OBJECT_CACHE
if (retry_cache_trim &&
((object = vm_object_cache_trim(TRUE)) !=
VM_OBJECT_NULL)) {
continue;
}
+#endif
return;
}
}
+#if VM_OBJECT_CACHE
assert(! retry_cache_trim);
+#endif
}
+
+#if VM_OBJECT_CACHE
/*
* Check to see whether we really need to trim
* down the cache. If so, remove an object from
@@ -800,6 +1094,8 @@ vm_object_cache_trim(
* If we no longer need to trim the cache,
* then we are done.
*/
+ if (vm_object_cached_count <= vm_object_cached_max)
+ return VM_OBJECT_NULL;
vm_object_cache_lock();
if (vm_object_cached_count <= vm_object_cached_max) {
@@ -813,8 +1109,8 @@ vm_object_cache_trim(
*/
XPR(XPR_VM_OBJECT,
"vm_object_cache_trim: removing from front of cache (%x, %x)\n",
- (integer_t)vm_object_cached_list.next,
- (integer_t)vm_object_cached_list.prev, 0, 0, 0);
+ vm_object_cached_list.next,
+ vm_object_cached_list.prev, 0, 0, 0);
object = (vm_object_t) queue_first(&vm_object_cached_list);
if(object == (vm_object_t) &vm_object_cached_list) {
@@ -832,6 +1128,7 @@ vm_object_cache_trim(
cached_list);
vm_object_cached_count--;
+ vm_object_cache_unlock();
/*
* Since this object is in the cache, we know
* that it is initialized and has no references.
@@ -840,6 +1137,7 @@ vm_object_cache_trim(
assert(object->pager_initialized);
assert(object->ref_count == 0);
+ vm_object_lock_assert_exclusive(object);
object->ref_count++;
/*
@@ -851,8 +1149,10 @@ vm_object_cache_trim(
* (We are careful here to limit recursion.)
*/
shadow = object->pageout?VM_OBJECT_NULL:object->shadow;
+
if(vm_object_terminate(object) != KERN_SUCCESS)
continue;
+
if (shadow != VM_OBJECT_NULL) {
if (called_from_vm_object_deallocate) {
return shadow;
@@ -862,8 +1162,8 @@ vm_object_cache_trim(
}
}
}
+#endif
-boolean_t vm_object_terminate_remove_all = FALSE;
/*
* Routine: vm_object_terminate
@@ -884,19 +1184,15 @@ boolean_t vm_object_terminate_remove_all = FALSE;
*/
static kern_return_t
vm_object_terminate(
- register vm_object_t object)
+ vm_object_t object)
{
- memory_object_t pager;
- register vm_page_t p;
- vm_object_t shadow_object;
+ vm_object_t shadow_object;
XPR(XPR_VM_OBJECT, "vm_object_terminate, object 0x%X ref %d\n",
- (integer_t)object, object->ref_count, 0, 0, 0);
+ object, object->ref_count, 0, 0, 0);
- if (!object->pageout && (!object->temporary || object->can_persist)
- && (object->pager != NULL || object->shadow_severed)) {
- vm_object_cache_unlock();
- while (!queue_empty(&object->memq)) {
+ if (!object->pageout && (!object->temporary || object->can_persist) &&
+ (object->pager != NULL || object->shadow_severed)) {
/*
* Clear pager_trusted bit so that the pages get yanked
* out of the object instead of cleaned in place. This
@@ -904,71 +1200,15 @@ vm_object_terminate(
*/
object->pager_trusted = FALSE;
- p = (vm_page_t) queue_first(&object->memq);
-
- VM_PAGE_CHECK(p);
-
- if (p->busy || p->cleaning) {
- if(p->cleaning || p->absent) {
- vm_object_paging_wait(object, THREAD_UNINT);
- continue;
- } else {
- panic("vm_object_terminate.3 0x%x 0x%x", object, p);
- }
- }
-
- vm_page_lock_queues();
- VM_PAGE_QUEUES_REMOVE(p);
- vm_page_unlock_queues();
-
- if (p->absent || p->private) {
-
- /*
- * For private pages, VM_PAGE_FREE just
- * leaves the page structure around for
- * its owner to clean up. For absent
- * pages, the structure is returned to
- * the appropriate pool.
- */
-
- goto free_page;
- }
-
- if (p->fictitious)
- panic("vm_object_terminate.4 0x%x 0x%x", object, p);
-
- if (!p->dirty)
- p->dirty = pmap_is_modified(p->phys_page);
-
- if ((p->dirty || p->precious) && !p->error && object->alive) {
- p->busy = TRUE;
- vm_object_paging_begin(object);
- /* protect the object from re-use/caching while it */
- /* is unlocked */
- vm_object_unlock(object);
- vm_pageout_cluster(p); /* flush page */
- vm_object_lock(object);
- vm_object_paging_wait(object, THREAD_UNINT);
- XPR(XPR_VM_OBJECT,
- "vm_object_terminate restart, object 0x%X ref %d\n",
- (integer_t)object, object->ref_count, 0, 0, 0);
- } else {
- free_page:
- VM_PAGE_FREE(p);
- }
- }
- vm_object_unlock(object);
- vm_object_cache_lock();
- vm_object_lock(object);
+ vm_object_reap_pages(object, REAP_TERMINATE);
}
-
/*
* Make sure the object isn't already being terminated
*/
- if(object->terminating) {
- object->ref_count -= 1;
+ if (object->terminating) {
+ vm_object_lock_assert_exclusive(object);
+ object->ref_count--;
assert(object->ref_count > 0);
- vm_object_cache_unlock();
vm_object_unlock(object);
return KERN_FAILURE;
}
@@ -977,11 +1217,11 @@ vm_object_terminate(
* Did somebody get a reference to the object while we were
* cleaning it?
*/
- if(object->ref_count != 1) {
- object->ref_count -= 1;
+ if (object->ref_count != 1) {
+ vm_object_lock_assert_exclusive(object);
+ object->ref_count--;
assert(object->ref_count > 0);
vm_object_res_deallocate(object);
- vm_object_cache_unlock();
vm_object_unlock(object);
return KERN_FAILURE;
}
@@ -992,43 +1232,107 @@ vm_object_terminate(
object->terminating = TRUE;
object->alive = FALSE;
- vm_object_remove(object);
+ if (object->hashed) {
+ lck_mtx_t *lck;
+
+ lck = vm_object_hash_lock_spin(object->pager);
+ vm_object_remove(object);
+ vm_object_hash_unlock(lck);
+ }
/*
* Detach the object from its shadow if we are the shadow's
- * copy.
+ * copy. The reference we hold on the shadow must be dropped
+ * by our caller.
*/
if (((shadow_object = object->shadow) != VM_OBJECT_NULL) &&
!(object->pageout)) {
vm_object_lock(shadow_object);
- assert((shadow_object->copy == object) ||
- (shadow_object->copy == VM_OBJECT_NULL));
- shadow_object->copy = VM_OBJECT_NULL;
+ if (shadow_object->copy == object)
+ shadow_object->copy = VM_OBJECT_NULL;
vm_object_unlock(shadow_object);
}
+ if (object->paging_in_progress != 0 ||
+ object->activity_in_progress != 0) {
+ /*
+ * There are still some paging_in_progress references
+ * on this object, meaning that there are some paging
+ * or other I/O operations in progress for this VM object.
+ * Such operations take some paging_in_progress references
+ * up front to ensure that the object doesn't go away, but
+ * they may also need to acquire a reference on the VM object,
+ * to map it in kernel space, for example. That means that
+ * they may end up releasing the last reference on the VM
+ * object, triggering its termination, while still holding
+ * paging_in_progress references. Waiting for these
+ * pending paging_in_progress references to go away here would
+ * deadlock.
+ *
+ * To avoid deadlocking, we'll let the vm_object_reaper_thread
+ * complete the VM object termination if it still holds
+ * paging_in_progress references at this point.
+ *
+ * No new paging_in_progress should appear now that the
+ * VM object is "terminating" and not "alive".
+ */
+ vm_object_reap_async(object);
+ vm_object_unlock(object);
+ /*
+ * Return KERN_FAILURE to let the caller know that we
+ * haven't completed the termination and it can't drop this
+ * object's reference on its shadow object yet.
+ * The reaper thread will take care of that once it has
+ * completed this object's termination.
+ */
+ return KERN_FAILURE;
+ }
+ /*
+ * complete the VM object termination
+ */
+ vm_object_reap(object);
+ object = VM_OBJECT_NULL;
+
/*
- * The pageout daemon might be playing with our pages.
- * Now that the object is dead, it won't touch any more
- * pages, but some pages might already be on their way out.
- * Hence, we wait until the active paging activities have ceased
- * before we break the association with the pager itself.
+ * the object lock was released by vm_object_reap()
+ *
+ * KERN_SUCCESS means that this object has been terminated
+ * and no longer needs its shadow object but still holds a
+ * reference on it.
+ * The caller is responsible for dropping that reference.
+ * We can't call vm_object_deallocate() here because that
+ * would create a recursion.
*/
- while (object->paging_in_progress != 0) {
- vm_object_cache_unlock();
- vm_object_wait(object,
- VM_OBJECT_EVENT_PAGING_IN_PROGRESS,
- THREAD_UNINT);
- vm_object_cache_lock();
- vm_object_lock(object);
- }
+ return KERN_SUCCESS;
+}
+
+
+/*
+ * vm_object_reap():
+ *
+ * Complete the termination of a VM object after it's been marked
+ * as "terminating" and "!alive" by vm_object_terminate().
+ *
+ * The VM object must be locked by caller.
+ * The lock will be released on return and the VM object is no longer valid.
+ */
+void
+vm_object_reap(
+ vm_object_t object)
+{
+ memory_object_t pager;
+
+ vm_object_lock_assert_exclusive(object);
+ assert(object->paging_in_progress == 0);
+ assert(object->activity_in_progress == 0);
+
+ vm_object_reap_count++;
pager = object->pager;
object->pager = MEMORY_OBJECT_NULL;
if (pager != MEMORY_OBJECT_NULL)
- memory_object_control_disable(object->pager_request);
- vm_object_cache_unlock();
+ memory_object_control_disable(object->pager_control);
object->ref_count--;
#if TASK_SWAPPER
@@ -1038,29 +1342,37 @@ vm_object_terminate(
assert (object->ref_count == 0);
/*
- * Clean or free the pages, as appropriate.
- * It is possible for us to find busy/absent pages,
- * if some faults on this object were aborted.
+ * remove from purgeable queue if it's on
+ */
+ if (object->objq.next || object->objq.prev) {
+ purgeable_q_t queue = vm_purgeable_object_remove(object);
+ assert(queue);
+
+ /* Must take page lock for this - using it to protect token queue */
+ vm_page_lock_queues();
+ vm_purgeable_token_delete_first(queue);
+
+ assert(queue->debug_count_objects>=0);
+ vm_page_unlock_queues();
+ }
+
+ /*
+ * Clean or free the pages, as appropriate.
+ * It is possible for us to find busy/absent pages,
+ * if some faults on this object were aborted.
*/
if (object->pageout) {
- assert(shadow_object != VM_OBJECT_NULL);
- assert(shadow_object == object->shadow);
+ assert(object->shadow != VM_OBJECT_NULL);
vm_pageout_object_terminate(object);
- } else if ((object->temporary && !object->can_persist) ||
- (pager == MEMORY_OBJECT_NULL)) {
- while (!queue_empty(&object->memq)) {
- p = (vm_page_t) queue_first(&object->memq);
+ } else if (((object->temporary && !object->can_persist) || (pager == MEMORY_OBJECT_NULL))) {
- VM_PAGE_CHECK(p);
- VM_PAGE_FREE(p);
- }
- } else if (!queue_empty(&object->memq)) {
- panic("vm_object_terminate: queue just emptied isn't");
+ vm_object_reap_pages(object, REAP_REAP);
}
-
+ assert(queue_empty(&object->memq));
assert(object->paging_in_progress == 0);
+ assert(object->activity_in_progress == 0);
assert(object->ref_count == 0);
/*
@@ -1070,7 +1382,7 @@ vm_object_terminate(
*/
if (pager != MEMORY_OBJECT_NULL) {
vm_object_unlock(object);
- vm_object_release_pager(pager);
+ vm_object_release_pager(pager, object->hashed);
vm_object_lock(object);
}
@@ -1084,11 +1396,356 @@ vm_object_terminate(
vm_external_destroy(object->existence_map, object->size);
#endif /* MACH_PAGEMAP */
+ object->shadow = VM_OBJECT_NULL;
+
+ vm_object_lock_destroy(object);
/*
* Free the space for the object.
*/
- zfree(vm_object_zone, (vm_offset_t) object);
- return KERN_SUCCESS;
+ zfree(vm_object_zone, object);
+ object = VM_OBJECT_NULL;
+}
+
+
+
+#define V_O_R_MAX_BATCH 128
+
+
+#define VM_OBJ_REAP_FREELIST(_local_free_q, do_disconnect) \
+ MACRO_BEGIN \
+ if (_local_free_q) { \
+ if (do_disconnect) { \
+ vm_page_t m; \
+ for (m = _local_free_q; \
+ m != VM_PAGE_NULL; \
+ m = (vm_page_t) m->pageq.next) { \
+ if (m->pmapped) { \
+ pmap_disconnect(m->phys_page); \
+ } \
+ } \
+ } \
+ vm_page_free_list(_local_free_q, TRUE); \
+ _local_free_q = VM_PAGE_NULL; \
+ } \
+ MACRO_END
+
+
+void
+vm_object_reap_pages(
+ vm_object_t object,
+ int reap_type)
+{
+ vm_page_t p;
+ vm_page_t next;
+ vm_page_t local_free_q = VM_PAGE_NULL;
+ int loop_count;
+ boolean_t disconnect_on_release;
+
+ if (reap_type == REAP_DATA_FLUSH) {
+ /*
+ * We need to disconnect pages from all pmaps before
+ * releasing them to the free list
+ */
+ disconnect_on_release = TRUE;
+ } else {
+ /*
+ * Either the caller has already disconnected the pages
+ * from all pmaps, or we disconnect them here as we add
+ * them to out local list of pages to be released.
+ * No need to re-disconnect them when we release the pages
+ * to the free list.
+ */
+ disconnect_on_release = FALSE;
+ }
+
+restart_after_sleep:
+ if (queue_empty(&object->memq))
+ return;
+ loop_count = V_O_R_MAX_BATCH + 1;
+
+ vm_page_lockspin_queues();
+
+ next = (vm_page_t)queue_first(&object->memq);
+
+ while (!queue_end(&object->memq, (queue_entry_t)next)) {
+
+ p = next;
+ next = (vm_page_t)queue_next(&next->listq);
+
+ if (--loop_count == 0) {
+
+ vm_page_unlock_queues();
+
+ if (local_free_q) {
+ /*
+ * Free the pages we reclaimed so far
+ * and take a little break to avoid
+ * hogging the page queue lock too long
+ */
+ VM_OBJ_REAP_FREELIST(local_free_q,
+ disconnect_on_release);
+ } else
+ mutex_pause(0);
+
+ loop_count = V_O_R_MAX_BATCH + 1;
+
+ vm_page_lockspin_queues();
+ }
+ if (reap_type == REAP_DATA_FLUSH || reap_type == REAP_TERMINATE) {
+
+ if (reap_type == REAP_DATA_FLUSH &&
+ ((p->pageout == TRUE || p->cleaning == TRUE) && p->list_req_pending == TRUE)) {
+ p->list_req_pending = FALSE;
+ p->cleaning = FALSE;
+ /*
+ * need to drop the laundry count...
+ * we may also need to remove it
+ * from the I/O paging queue...
+ * vm_pageout_throttle_up handles both cases
+ *
+ * the laundry and pageout_queue flags are cleared...
+ */
+#if CONFIG_EMBEDDED
+ if (p->laundry)
+ vm_pageout_throttle_up(p);
+#else
+ vm_pageout_throttle_up(p);
+#endif
+ if (p->pageout == TRUE) {
+ /*
+ * toss the wire count we picked up
+ * when we initially set this page up
+ * to be cleaned and stolen...
+ */
+ vm_page_unwire(p, TRUE);
+ p->pageout = FALSE;
+ }
+ PAGE_WAKEUP(p);
+
+ } else if (p->busy || p->cleaning) {
+
+ vm_page_unlock_queues();
+ /*
+ * free the pages reclaimed so far
+ */
+ VM_OBJ_REAP_FREELIST(local_free_q,
+ disconnect_on_release);
+
+ PAGE_SLEEP(object, p, THREAD_UNINT);
+
+ goto restart_after_sleep;
+ }
+ }
+ switch (reap_type) {
+
+ case REAP_DATA_FLUSH:
+ if (VM_PAGE_WIRED(p)) {
+ /*
+ * this is an odd case... perhaps we should
+ * zero-fill this page since we're conceptually
+ * tossing its data at this point, but leaving
+ * it on the object to honor the 'wire' contract
+ */
+ continue;
+ }
+ break;
+
+ case REAP_PURGEABLE:
+ if (VM_PAGE_WIRED(p)) {
+ /* can't purge a wired page */
+ vm_page_purged_wired++;
+ continue;
+ }
+
+ if (p->busy) {
+ /*
+ * We can't reclaim a busy page but we can
+ * make it pageable (it's not wired) to make
+ * sure that it gets considered by
+ * vm_pageout_scan() later.
+ */
+ vm_page_deactivate(p);
+ vm_page_purged_busy++;
+ continue;
+ }
+
+ if (p->cleaning || p->laundry || p->list_req_pending) {
+ /*
+ * page is being acted upon,
+ * so don't mess with it
+ */
+ vm_page_purged_others++;
+ continue;
+ }
+ assert(p->object != kernel_object);
+
+ /*
+ * we can discard this page...
+ */
+ if (p->pmapped == TRUE) {
+ int refmod_state;
+ /*
+ * unmap the page
+ */
+ refmod_state = pmap_disconnect(p->phys_page);
+ if (refmod_state & VM_MEM_MODIFIED) {
+ p->dirty = TRUE;
+ }
+ }
+ if (p->dirty || p->precious) {
+ /*
+ * we saved the cost of cleaning this page !
+ */
+ vm_page_purged_count++;
+ }
+
+ break;
+
+ case REAP_TERMINATE:
+ if (p->absent || p->private) {
+ /*
+ * For private pages, VM_PAGE_FREE just
+ * leaves the page structure around for
+ * its owner to clean up. For absent
+ * pages, the structure is returned to
+ * the appropriate pool.
+ */
+ break;
+ }
+ if (p->fictitious) {
+ assert (p->phys_page == vm_page_guard_addr);
+ break;
+ }
+ if (!p->dirty && p->wpmapped)
+ p->dirty = pmap_is_modified(p->phys_page);
+
+ if ((p->dirty || p->precious) && !p->error && object->alive) {
+
+ p->busy = TRUE;
+
+ VM_PAGE_QUEUES_REMOVE(p);
+
+ vm_page_unlock_queues();
+ /*
+ * free the pages reclaimed so far
+ */
+ VM_OBJ_REAP_FREELIST(local_free_q,
+ disconnect_on_release);
+
+ /*
+ * flush page... page will be freed
+ * upon completion of I/O
+ */
+ vm_pageout_cluster(p);
+ vm_object_paging_wait(object, THREAD_UNINT);
+
+ goto restart_after_sleep;
+ }
+ break;
+
+ case REAP_REAP:
+ break;
+ }
+ vm_page_free_prepare_queues(p);
+ assert(p->pageq.next == NULL && p->pageq.prev == NULL);
+ /*
+ * Add this page to our list of reclaimed pages,
+ * to be freed later.
+ */
+ p->pageq.next = (queue_entry_t) local_free_q;
+ local_free_q = p;
+ }
+ vm_page_unlock_queues();
+
+ /*
+ * Free the remaining reclaimed pages
+ */
+ VM_OBJ_REAP_FREELIST(local_free_q,
+ disconnect_on_release);
+}
+
+
+void
+vm_object_reap_async(
+ vm_object_t object)
+{
+ vm_object_lock_assert_exclusive(object);
+
+ vm_object_reaper_lock_spin();
+
+ vm_object_reap_count_async++;
+
+ /* enqueue the VM object... */
+ queue_enter(&vm_object_reaper_queue, object,
+ vm_object_t, cached_list);
+
+ vm_object_reaper_unlock();
+
+ /* ... and wake up the reaper thread */
+ thread_wakeup((event_t) &vm_object_reaper_queue);
+}
+
+
+void
+vm_object_reaper_thread(void)
+{
+ vm_object_t object, shadow_object;
+
+ vm_object_reaper_lock_spin();
+
+ while (!queue_empty(&vm_object_reaper_queue)) {
+ queue_remove_first(&vm_object_reaper_queue,
+ object,
+ vm_object_t,
+ cached_list);
+
+ vm_object_reaper_unlock();
+ vm_object_lock(object);
+
+ assert(object->terminating);
+ assert(!object->alive);
+
+ /*
+ * The pageout daemon might be playing with our pages.
+ * Now that the object is dead, it won't touch any more
+ * pages, but some pages might already be on their way out.
+ * Hence, we wait until the active paging activities have
+ * ceased before we break the association with the pager
+ * itself.
+ */
+ while (object->paging_in_progress != 0 ||
+ object->activity_in_progress != 0) {
+ vm_object_wait(object,
+ VM_OBJECT_EVENT_PAGING_IN_PROGRESS,
+ THREAD_UNINT);
+ vm_object_lock(object);
+ }
+
+ shadow_object =
+ object->pageout ? VM_OBJECT_NULL : object->shadow;
+
+ vm_object_reap(object);
+ /* cache is unlocked and object is no longer valid */
+ object = VM_OBJECT_NULL;
+
+ if (shadow_object != VM_OBJECT_NULL) {
+ /*
+ * Drop the reference "object" was holding on
+ * its shadow object.
+ */
+ vm_object_deallocate(shadow_object);
+ shadow_object = VM_OBJECT_NULL;
+ }
+ vm_object_reaper_lock_spin();
+ }
+
+ /* wait for more work... */
+ assert_wait((event_t) &vm_object_reaper_queue, THREAD_UNINT);
+
+ vm_object_reaper_unlock();
+
+ thread_block((thread_continue_t) vm_object_reaper_thread);
+ /*NOTREACHED*/
}
/*
@@ -1102,16 +1759,18 @@ vm_object_pager_wakeup(
{
vm_object_hash_entry_t entry;
boolean_t waiting = FALSE;
+ lck_mtx_t *lck;
/*
* If anyone was waiting for the memory_object_terminate
* to be queued, wake them up now.
*/
- vm_object_cache_lock();
+ lck = vm_object_hash_lock_spin(pager);
entry = vm_object_hash_lookup(pager, TRUE);
if (entry != VM_OBJECT_HASH_ENTRY_NULL)
waiting = entry->waiting;
- vm_object_cache_unlock();
+ vm_object_hash_unlock(lck);
+
if (entry != VM_OBJECT_HASH_ENTRY_NULL) {
if (waiting)
thread_wakeup((event_t) pager);
@@ -1130,7 +1789,8 @@ vm_object_pager_wakeup(
*/
static void
vm_object_release_pager(
- memory_object_t pager)
+ memory_object_t pager,
+ boolean_t hashed)
{
/*
@@ -1139,72 +1799,17 @@ vm_object_release_pager(
(void) memory_object_terminate(pager);
- /*
- * Wakeup anyone waiting for this terminate
- */
- vm_object_pager_wakeup(pager);
-
- /*
- * Release reference to pager.
- */
- memory_object_deallocate(pager);
-}
-
-/*
- * Routine: vm_object_abort_activity [internal use only]
- * Purpose:
- * Abort paging requests pending on this object.
- * In/out conditions:
- * The object is locked on entry and exit.
- */
-static void
-vm_object_abort_activity(
- vm_object_t object)
-{
- register
- vm_page_t p;
- vm_page_t next;
-
- XPR(XPR_VM_OBJECT, "vm_object_abort_activity, object 0x%X\n",
- (integer_t)object, 0, 0, 0, 0);
-
- /*
- * Abort all activity that would be waiting
- * for a result on this memory object.
- *
- * We could also choose to destroy all pages
- * that we have in memory for this object, but
- * we don't.
- */
-
- p = (vm_page_t) queue_first(&object->memq);
- while (!queue_end(&object->memq, (queue_entry_t) p)) {
- next = (vm_page_t) queue_next(&p->listq);
-
+ if (hashed == TRUE) {
/*
- * If it's being paged in, destroy it.
- * If an unlock has been requested, start it again.
+ * Wakeup anyone waiting for this terminate
+ * and remove the entry from the hash
*/
-
- if (p->busy && p->absent) {
- VM_PAGE_FREE(p);
- }
- else {
- if (p->unlock_request != VM_PROT_NONE)
- p->unlock_request = VM_PROT_NONE;
- PAGE_WAKEUP(p);
- }
-
- p = next;
+ vm_object_pager_wakeup(pager);
}
-
/*
- * Wake up threads waiting for the memory object to
- * become ready.
+ * Release reference to pager.
*/
-
- object->pager_ready = TRUE;
- vm_object_wakeup(object, VM_OBJECT_EVENT_PAGER_READY);
+ memory_object_deallocate(pager);
}
/*
@@ -1217,7 +1822,7 @@ vm_object_abort_activity(
kern_return_t
vm_object_destroy(
vm_object_t object,
- kern_return_t reason)
+ __unused kern_return_t reason)
{
memory_object_t old_pager;
@@ -1233,22 +1838,24 @@ vm_object_destroy(
* the destroy call.]
*/
- vm_object_cache_lock();
vm_object_lock(object);
object->can_persist = FALSE;
object->named = FALSE;
object->alive = FALSE;
- /*
- * Rip out the pager from the vm_object now...
- */
-
- vm_object_remove(object);
+ if (object->hashed) {
+ lck_mtx_t *lck;
+ /*
+ * Rip out the pager from the vm_object now...
+ */
+ lck = vm_object_hash_lock_spin(object->pager);
+ vm_object_remove(object);
+ vm_object_hash_unlock(lck);
+ }
old_pager = object->pager;
object->pager = MEMORY_OBJECT_NULL;
if (old_pager != MEMORY_OBJECT_NULL)
- memory_object_control_disable(object->pager_request);
- vm_object_cache_unlock();
+ memory_object_control_disable(object->pager_control);
/*
* Wait for the existing paging activity (that got
@@ -1262,7 +1869,7 @@ vm_object_destroy(
* Terminate the object now.
*/
if (old_pager != MEMORY_OBJECT_NULL) {
- vm_object_release_pager(old_pager);
+ vm_object_release_pager(old_pager, object->hashed);
/*
* JMM - Release the caller's reference. This assumes the
@@ -1277,8 +1884,14 @@ vm_object_destroy(
return(KERN_SUCCESS);
}
+
+#define VM_OBJ_DEACT_ALL_STATS DEBUG
+#if VM_OBJ_DEACT_ALL_STATS
+uint32_t vm_object_deactivate_all_pages_batches = 0;
+uint32_t vm_object_deactivate_all_pages_pages = 0;
+#endif /* VM_OBJ_DEACT_ALL_STATS */
/*
- * vm_object_deactivate_pages
+ * vm_object_deactivate_all_pages
*
* Deactivate all pages in the specified object. (Keep its pages
* in memory even though it is no longer referenced.)
@@ -1290,99 +1903,716 @@ vm_object_deactivate_all_pages(
register vm_object_t object)
{
register vm_page_t p;
-
+ int loop_count;
+#if VM_OBJ_DEACT_ALL_STATS
+ int pages_count;
+#endif /* VM_OBJ_DEACT_ALL_STATS */
+#define V_O_D_A_P_MAX_BATCH 256
+
+ loop_count = V_O_D_A_P_MAX_BATCH;
+#if VM_OBJ_DEACT_ALL_STATS
+ pages_count = 0;
+#endif /* VM_OBJ_DEACT_ALL_STATS */
+ vm_page_lock_queues();
queue_iterate(&object->memq, p, vm_page_t, listq) {
- vm_page_lock_queues();
- if (!p->busy)
+ if (--loop_count == 0) {
+#if VM_OBJ_DEACT_ALL_STATS
+ hw_atomic_add(&vm_object_deactivate_all_pages_batches,
+ 1);
+ hw_atomic_add(&vm_object_deactivate_all_pages_pages,
+ pages_count);
+ pages_count = 0;
+#endif /* VM_OBJ_DEACT_ALL_STATS */
+ lck_mtx_yield(&vm_page_queue_lock);
+ loop_count = V_O_D_A_P_MAX_BATCH;
+ }
+ if (!p->busy && !p->throttled) {
+#if VM_OBJ_DEACT_ALL_STATS
+ pages_count++;
+#endif /* VM_OBJ_DEACT_ALL_STATS */
vm_page_deactivate(p);
- vm_page_unlock_queues();
+ }
+ }
+#if VM_OBJ_DEACT_ALL_STATS
+ if (pages_count) {
+ hw_atomic_add(&vm_object_deactivate_all_pages_batches, 1);
+ hw_atomic_add(&vm_object_deactivate_all_pages_pages,
+ pages_count);
+ pages_count = 0;
}
+#endif /* VM_OBJ_DEACT_ALL_STATS */
+ vm_page_unlock_queues();
}
-__private_extern__ void
-vm_object_deactivate_pages(
- vm_object_t object,
- vm_object_offset_t offset,
- vm_object_size_t size,
- boolean_t kill_page)
-{
- vm_object_t orig_object;
- int pages_moved = 0;
- int pages_found = 0;
- /*
- * entered with object lock held, acquire a paging reference to
- * prevent the memory_object and control ports from
- * being destroyed.
- */
- orig_object = object;
- for (;;) {
- register vm_page_t m;
- vm_object_offset_t toffset;
- vm_object_size_t tsize;
+/*
+ * when deallocating pages it is necessary to hold
+ * the vm_page_queue_lock (a hot global lock) for certain operations
+ * on the page... however, the majority of the work can be done
+ * while merely holding the object lock... to mitigate the time spent behind the
+ * global lock, go to a 2 pass algorithm... collect pages up to DELAYED_WORK_LIMIT
+ * while doing all of the work that doesn't require the vm_page_queue_lock...
+ * them call dw_do_work to acquire the vm_page_queue_lock and do the
+ * necessary work for each page... we will grab the busy bit on the page
+ * so that dw_do_work can drop the object lock if it can't immediately take the
+ * vm_page_queue_lock in order to compete for the locks in the same order that
+ * vm_pageout_scan takes them.
+ */
- vm_object_paging_begin(object);
- vm_page_lock_queues();
+#define DELAYED_WORK_LIMIT 32
- for (tsize = size, toffset = offset; tsize; tsize -= PAGE_SIZE, toffset += PAGE_SIZE) {
+#define DW_clear_reference 0x01
+#define DW_move_page 0x02
+#define DW_clear_busy 0x04
+#define DW_PAGE_WAKEUP 0x08
- if ((m = vm_page_lookup(object, toffset)) != VM_PAGE_NULL) {
- pages_found++;
+struct dw {
+ vm_page_t dw_m;
+ int dw_mask;
+};
- if ((m->wire_count == 0) && (!m->private) && (!m->gobbled) && (!m->busy)) {
+static void dw_do_work(vm_object_t object, struct dw *dwp, int dw_count);
- m->reference = FALSE;
- pmap_clear_reference(m->phys_page);
- if ((kill_page) && (object->internal)) {
- m->precious = FALSE;
- m->dirty = FALSE;
- pmap_clear_modify(m->phys_page);
- vm_external_state_clr(object->existence_map, offset);
- }
- VM_PAGE_QUEUES_REMOVE(m);
+static void
+dw_do_work(
+ vm_object_t object,
+ struct dw *dwp,
+ int dw_count)
+{
+ vm_page_t m;
+ int j;
- if(m->zero_fill) {
- queue_enter_first(
- &vm_page_queue_zf,
- m, vm_page_t, pageq);
- } else {
- queue_enter_first(
- &vm_page_queue_inactive,
- m, vm_page_t, pageq);
- }
+ /*
+ * pageout_scan takes the vm_page_lock_queues first
+ * then tries for the object lock... to avoid what
+ * is effectively a lock inversion, we'll go to the
+ * trouble of taking them in that same order... otherwise
+ * if this object contains the majority of the pages resident
+ * in the UBC (or a small set of large objects actively being
+ * worked on contain the majority of the pages), we could
+ * cause the pageout_scan thread to 'starve' in its attempt
+ * to find pages to move to the free queue, since it has to
+ * successfully acquire the object lock of any candidate page
+ * before it can steal/clean it.
+ */
+ if (!vm_page_trylockspin_queues()) {
+ vm_object_unlock(object);
- m->inactive = TRUE;
- if (!m->fictitious)
- vm_page_inactive_count++;
+ vm_page_lockspin_queues();
- pages_moved++;
- }
- }
+ for (j = 0; ; j++) {
+ if (!vm_object_lock_avoid(object) &&
+ _vm_object_lock_try(object))
+ break;
+ vm_page_unlock_queues();
+ mutex_pause(j);
+ vm_page_lockspin_queues();
}
- vm_page_unlock_queues();
- vm_object_paging_end(object);
+ }
+ for (j = 0; j < dw_count; j++, dwp++) {
- if (object->shadow) {
- vm_object_t tmp_object;
+ m = dwp->dw_m;
- kill_page = 0;
+ if (dwp->dw_mask & DW_clear_reference)
+ m->reference = FALSE;
- offset += object->shadow_offset;
+ if (dwp->dw_mask & DW_move_page) {
+ VM_PAGE_QUEUES_REMOVE(m);
- tmp_object = object->shadow;
- vm_object_lock(tmp_object);
+ assert(!m->laundry);
+ assert(m->object != kernel_object);
+ assert(m->pageq.next == NULL &&
+ m->pageq.prev == NULL);
+
+ if (m->zero_fill) {
+ queue_enter_first(&vm_page_queue_zf, m, vm_page_t, pageq);
+ vm_zf_queue_count++;
+ } else {
+ queue_enter_first(&vm_page_queue_inactive, m, vm_page_t, pageq);
+ }
+ m->inactive = TRUE;
- if (object != orig_object)
- vm_object_unlock(object);
- object = tmp_object;
- } else
- break;
- }
- if (object != orig_object)
+ if (!m->fictitious) {
+ vm_page_inactive_count++;
+ token_new_pagecount++;
+ } else {
+ assert(m->phys_page == vm_page_fictitious_addr);
+ }
+ }
+ if (dwp->dw_mask & DW_clear_busy)
+ dwp->dw_m->busy = FALSE;
+
+ if (dwp->dw_mask & DW_PAGE_WAKEUP)
+ PAGE_WAKEUP(dwp->dw_m);
+ }
+ vm_page_unlock_queues();
+
+#if CONFIG_EMBEDDED
+ {
+ int percent_avail;
+
+ /*
+ * Decide if we need to send a memory status notification.
+ */
+ percent_avail =
+ (vm_page_active_count + vm_page_inactive_count +
+ vm_page_speculative_count + vm_page_free_count +
+ (IP_VALID(memory_manager_default)?0:vm_page_purgeable_count) ) * 100 /
+ atop_64(max_mem);
+ if (percent_avail >= (kern_memorystatus_level + 5) ||
+ percent_avail <= (kern_memorystatus_level - 5)) {
+ kern_memorystatus_level = percent_avail;
+ thread_wakeup((event_t)&kern_memorystatus_wakeup);
+ }
+ }
+#endif
+}
+
+
+
+/*
+ * The "chunk" macros are used by routines below when looking for pages to deactivate. These
+ * exist because of the need to handle shadow chains. When deactivating pages, we only
+ * want to deactive the ones at the top most level in the object chain. In order to do
+ * this efficiently, the specified address range is divided up into "chunks" and we use
+ * a bit map to keep track of which pages have already been processed as we descend down
+ * the shadow chain. These chunk macros hide the details of the bit map implementation
+ * as much as we can.
+ *
+ * For convenience, we use a 64-bit data type as the bit map, and therefore a chunk is
+ * set to 64 pages. The bit map is indexed from the low-order end, so that the lowest
+ * order bit represents page 0 in the current range and highest order bit represents
+ * page 63.
+ *
+ * For further convenience, we also use negative logic for the page state in the bit map.
+ * The bit is set to 1 to indicate it has not yet been seen, and to 0 to indicate it has
+ * been processed. This way we can simply test the 64-bit long word to see if it's zero
+ * to easily tell if the whole range has been processed. Therefore, the bit map starts
+ * out with all the bits set. The macros below hide all these details from the caller.
+ */
+
+#define PAGES_IN_A_CHUNK 64 /* The number of pages in the chunk must */
+ /* be the same as the number of bits in */
+ /* the chunk_state_t type. We use 64 */
+ /* just for convenience. */
+
+#define CHUNK_SIZE (PAGES_IN_A_CHUNK * PAGE_SIZE_64) /* Size of a chunk in bytes */
+
+typedef uint64_t chunk_state_t;
+
+/*
+ * The bit map uses negative logic, so we start out with all 64 bits set to indicate
+ * that no pages have been processed yet. Also, if len is less than the full CHUNK_SIZE,
+ * then we mark pages beyond the len as having been "processed" so that we don't waste time
+ * looking at pages in that range. This can save us from unnecessarily chasing down the
+ * shadow chain.
+ */
+
+#define CHUNK_INIT(c, len) \
+ MACRO_BEGIN \
+ uint64_t p; \
+ \
+ (c) = 0xffffffffffffffffLL; \
+ \
+ for (p = (len) / PAGE_SIZE_64; p < PAGES_IN_A_CHUNK; p++) \
+ MARK_PAGE_HANDLED(c, p); \
+ MACRO_END
+
+/*
+ * Return true if all pages in the chunk have not yet been processed.
+ */
+
+#define CHUNK_NOT_COMPLETE(c) ((c) != 0)
+
+/*
+ * Return true if the page at offset 'p' in the bit map has already been handled
+ * while processing a higher level object in the shadow chain.
+ */
+
+#define PAGE_ALREADY_HANDLED(c, p) (((c) & (1LL << (p))) == 0)
+
+/*
+ * Mark the page at offset 'p' in the bit map as having been processed.
+ */
+
+#define MARK_PAGE_HANDLED(c, p) \
+MACRO_BEGIN \
+ (c) = (c) & ~(1LL << (p)); \
+MACRO_END
+
+
+/*
+ * Return true if the page at the given offset has been paged out. Object is
+ * locked upon entry and returned locked.
+ */
+
+static boolean_t
+page_is_paged_out(
+ vm_object_t object,
+ vm_object_offset_t offset)
+{
+ kern_return_t kr;
+ memory_object_t pager;
+
+ /*
+ * Check the existence map for the page if we have one, otherwise
+ * ask the pager about this page.
+ */
+
+#if MACH_PAGEMAP
+ if (object->existence_map) {
+ if (vm_external_state_get(object->existence_map, offset)
+ == VM_EXTERNAL_STATE_EXISTS) {
+ /*
+ * We found the page
+ */
+
+ return TRUE;
+ }
+ } else
+#endif
+ if (object->internal &&
+ object->alive &&
+ !object->terminating &&
+ object->pager_ready) {
+
+ /*
+ * We're already holding a "paging in progress" reference
+ * so the object can't disappear when we release the lock.
+ */
+
+ assert(object->paging_in_progress);
+ pager = object->pager;
+ vm_object_unlock(object);
+
+ kr = memory_object_data_request(
+ pager,
+ offset + object->paging_offset,
+ 0, /* just poke the pager */
+ VM_PROT_READ,
+ NULL);
+
+ vm_object_lock(object);
+
+ if (kr == KERN_SUCCESS) {
+
+ /*
+ * We found the page
+ */
+
+ return TRUE;
+ }
+ }
+
+ return FALSE;
+}
+
+
+/*
+ * Deactivate the pages in the specified object and range. If kill_page is set, also discard any
+ * page modified state from the pmap. Update the chunk_state as we go along. The caller must specify
+ * a size that is less than or equal to the CHUNK_SIZE.
+ */
+
+static void
+deactivate_pages_in_object(
+ vm_object_t object,
+ vm_object_offset_t offset,
+ vm_object_size_t size,
+ boolean_t kill_page,
+ boolean_t reusable_page,
+#if !MACH_ASSERT
+ __unused
+#endif
+ boolean_t all_reusable,
+ chunk_state_t *chunk_state)
+{
+ vm_page_t m;
+ int p;
+ struct dw dw_array[DELAYED_WORK_LIMIT];
+ struct dw *dwp;
+ int dw_count;
+ unsigned int reusable = 0;
+
+
+ /*
+ * Examine each page in the chunk. The variable 'p' is the page number relative to the start of the
+ * chunk. Since this routine is called once for each level in the shadow chain, the chunk_state may
+ * have pages marked as having been processed already. We stop the loop early if we find we've handled
+ * all the pages in the chunk.
+ */
+
+ dwp = &dw_array[0];
+ dw_count = 0;
+
+ for(p = 0; size && CHUNK_NOT_COMPLETE(*chunk_state); p++, size -= PAGE_SIZE_64, offset += PAGE_SIZE_64) {
+
+ /*
+ * If this offset has already been found and handled in a higher level object, then don't
+ * do anything with it in the current shadow object.
+ */
+
+ if (PAGE_ALREADY_HANDLED(*chunk_state, p))
+ continue;
+
+ /*
+ * See if the page at this offset is around. First check to see if the page is resident,
+ * then if not, check the existence map or with the pager.
+ */
+
+ if ((m = vm_page_lookup(object, offset)) != VM_PAGE_NULL) {
+
+ /*
+ * We found a page we were looking for. Mark it as "handled" now in the chunk_state
+ * so that we won't bother looking for a page at this offset again if there are more
+ * shadow objects. Then deactivate the page.
+ */
+
+ MARK_PAGE_HANDLED(*chunk_state, p);
+
+ if (( !VM_PAGE_WIRED(m)) && (!m->private) && (!m->gobbled) && (!m->busy)) {
+ int clear_refmod;
+
+ assert(!m->laundry);
+
+ clear_refmod = VM_MEM_REFERENCED;
+ dwp->dw_mask = DW_clear_reference;
+
+ if ((kill_page) && (object->internal)) {
+ m->precious = FALSE;
+ m->dirty = FALSE;
+
+ clear_refmod |= VM_MEM_MODIFIED;
+ if (m->throttled) {
+ /*
+ * This page is now clean and
+ * reclaimable. Move it out
+ * of the throttled queue, so
+ * that vm_pageout_scan() can
+ * find it.
+ */
+ dwp->dw_mask |= DW_move_page;
+ }
+#if MACH_PAGEMAP
+ vm_external_state_clr(object->existence_map, offset);
+#endif /* MACH_PAGEMAP */
+
+ if (reusable_page && !m->reusable) {
+ assert(!all_reusable);
+ assert(!object->all_reusable);
+ m->reusable = TRUE;
+ object->reusable_page_count++;
+ assert(object->resident_page_count >= object->reusable_page_count);
+ reusable++;
+#if CONFIG_EMBEDDED
+ } else {
+ if (m->reusable) {
+ m->reusable = FALSE;
+ object->reusable_page_count--;
+ }
+#endif
+ }
+ }
+ pmap_clear_refmod(m->phys_page, clear_refmod);
+
+ if (!m->throttled && !(reusable_page || all_reusable))
+ dwp->dw_mask |= DW_move_page;
+ /*
+ * dw_do_work may need to drop the object lock
+ * if it does, we need the pages its looking at to
+ * be held stable via the busy bit.
+ */
+ m->busy = TRUE;
+ dwp->dw_mask |= (DW_clear_busy | DW_PAGE_WAKEUP);
+
+ dwp->dw_m = m;
+ dwp++;
+ dw_count++;
+
+ if (dw_count >= DELAYED_WORK_LIMIT) {
+ if (reusable) {
+ OSAddAtomic(reusable,
+ &vm_page_stats_reusable.reusable_count);
+ vm_page_stats_reusable.reusable += reusable;
+ reusable = 0;
+ }
+ dw_do_work(object, &dw_array[0], dw_count);
+
+ dwp = &dw_array[0];
+ dw_count = 0;
+ }
+ }
+
+ } else {
+
+ /*
+ * The page at this offset isn't memory resident, check to see if it's
+ * been paged out. If so, mark it as handled so we don't bother looking
+ * for it in the shadow chain.
+ */
+
+ if (page_is_paged_out(object, offset)) {
+ MARK_PAGE_HANDLED(*chunk_state, p);
+
+ /*
+ * If we're killing a non-resident page, then clear the page in the existence
+ * map so we don't bother paging it back in if it's touched again in the future.
+ */
+
+ if ((kill_page) && (object->internal)) {
+#if MACH_PAGEMAP
+ vm_external_state_clr(object->existence_map, offset);
+#endif /* MACH_PAGEMAP */
+ }
+ }
+ }
+ }
+
+ if (reusable) {
+ OSAddAtomic(reusable, &vm_page_stats_reusable.reusable_count);
+ vm_page_stats_reusable.reusable += reusable;
+ reusable = 0;
+ }
+
+ if (dw_count)
+ dw_do_work(object, &dw_array[0], dw_count);
+}
+
+
+/*
+ * Deactive a "chunk" of the given range of the object starting at offset. A "chunk"
+ * will always be less than or equal to the given size. The total range is divided up
+ * into chunks for efficiency and performance related to the locks and handling the shadow
+ * chain. This routine returns how much of the given "size" it actually processed. It's
+ * up to the caler to loop and keep calling this routine until the entire range they want
+ * to process has been done.
+ */
+
+static vm_object_size_t
+deactivate_a_chunk(
+ vm_object_t orig_object,
+ vm_object_offset_t offset,
+ vm_object_size_t size,
+ boolean_t kill_page,
+ boolean_t reusable_page,
+ boolean_t all_reusable)
+{
+ vm_object_t object;
+ vm_object_t tmp_object;
+ vm_object_size_t length;
+ chunk_state_t chunk_state;
+
+
+ /*
+ * Get set to do a chunk. We'll do up to CHUNK_SIZE, but no more than the
+ * remaining size the caller asked for.
+ */
+
+ length = MIN(size, CHUNK_SIZE);
+
+ /*
+ * The chunk_state keeps track of which pages we've already processed if there's
+ * a shadow chain on this object. At this point, we haven't done anything with this
+ * range of pages yet, so initialize the state to indicate no pages processed yet.
+ */
+
+ CHUNK_INIT(chunk_state, length);
+ object = orig_object;
+
+ /*
+ * Start at the top level object and iterate around the loop once for each object
+ * in the shadow chain. We stop processing early if we've already found all the pages
+ * in the range. Otherwise we stop when we run out of shadow objects.
+ */
+
+ while (object && CHUNK_NOT_COMPLETE(chunk_state)) {
+ vm_object_paging_begin(object);
+
+ deactivate_pages_in_object(object, offset, length, kill_page, reusable_page, all_reusable, &chunk_state);
+
+ vm_object_paging_end(object);
+
+ /*
+ * We've finished with this object, see if there's a shadow object. If
+ * there is, update the offset and lock the new object. We also turn off
+ * kill_page at this point since we only kill pages in the top most object.
+ */
+
+ tmp_object = object->shadow;
+
+ if (tmp_object) {
+ kill_page = FALSE;
+ reusable_page = FALSE;
+ all_reusable = FALSE;
+ offset += object->shadow_offset;
+ vm_object_lock(tmp_object);
+ }
+
+ if (object != orig_object)
+ vm_object_unlock(object);
+
+ object = tmp_object;
+ }
+
+ if (object && object != orig_object)
vm_object_unlock(object);
+
+ return length;
+}
+
+
+
+/*
+ * Move any resident pages in the specified range to the inactive queue. If kill_page is set,
+ * we also clear the modified status of the page and "forget" any changes that have been made
+ * to the page.
+ */
+
+__private_extern__ void
+vm_object_deactivate_pages(
+ vm_object_t object,
+ vm_object_offset_t offset,
+ vm_object_size_t size,
+ boolean_t kill_page,
+ boolean_t reusable_page)
+{
+ vm_object_size_t length;
+ boolean_t all_reusable;
+
+ /*
+ * We break the range up into chunks and do one chunk at a time. This is for
+ * efficiency and performance while handling the shadow chains and the locks.
+ * The deactivate_a_chunk() function returns how much of the range it processed.
+ * We keep calling this routine until the given size is exhausted.
+ */
+
+
+ all_reusable = FALSE;
+ if (reusable_page &&
+ object->size != 0 &&
+ object->size == size &&
+ object->reusable_page_count == 0) {
+ all_reusable = TRUE;
+ reusable_page = FALSE;
+ }
+
+#if CONFIG_EMBEDDED
+ if ((reusable_page || all_reusable) && object->all_reusable) {
+ /* This means MADV_FREE_REUSABLE has been called twice, which
+ * is probably illegal. */
+ return;
+ }
+#endif
+
+ while (size) {
+ length = deactivate_a_chunk(object, offset, size, kill_page, reusable_page, all_reusable);
+
+ size -= length;
+ offset += length;
+ }
+
+ if (all_reusable) {
+ if (!object->all_reusable) {
+ unsigned int reusable;
+
+ object->all_reusable = TRUE;
+ assert(object->reusable_page_count == 0);
+ /* update global stats */
+ reusable = object->resident_page_count;
+ OSAddAtomic(reusable,
+ &vm_page_stats_reusable.reusable_count);
+ vm_page_stats_reusable.reusable += reusable;
+ vm_page_stats_reusable.all_reusable_calls++;
+ }
+ } else if (reusable_page) {
+ vm_page_stats_reusable.partial_reusable_calls++;
+ }
+}
+
+void
+vm_object_reuse_pages(
+ vm_object_t object,
+ vm_object_offset_t start_offset,
+ vm_object_offset_t end_offset,
+ boolean_t allow_partial_reuse)
+{
+ vm_object_offset_t cur_offset;
+ vm_page_t m;
+ unsigned int reused, reusable;
+
+#define VM_OBJECT_REUSE_PAGE(object, m, reused) \
+ MACRO_BEGIN \
+ if ((m) != VM_PAGE_NULL && \
+ (m)->reusable) { \
+ assert((object)->reusable_page_count <= \
+ (object)->resident_page_count); \
+ assert((object)->reusable_page_count > 0); \
+ (object)->reusable_page_count--; \
+ (m)->reusable = FALSE; \
+ (reused)++; \
+ } \
+ MACRO_END
+
+ reused = 0;
+ reusable = 0;
+
+ vm_object_lock_assert_exclusive(object);
+
+ if (object->all_reusable) {
+ assert(object->reusable_page_count == 0);
+ object->all_reusable = FALSE;
+ if (end_offset - start_offset == object->size ||
+ !allow_partial_reuse) {
+ vm_page_stats_reusable.all_reuse_calls++;
+ reused = object->resident_page_count;
+ } else {
+ vm_page_stats_reusable.partial_reuse_calls++;
+ queue_iterate(&object->memq, m, vm_page_t, listq) {
+ if (m->offset < start_offset ||
+ m->offset >= end_offset) {
+ m->reusable = TRUE;
+ object->reusable_page_count++;
+ assert(object->resident_page_count >= object->reusable_page_count);
+ continue;
+ } else {
+ assert(!m->reusable);
+ reused++;
+ }
+ }
+ }
+ } else if (object->resident_page_count >
+ ((end_offset - start_offset) >> PAGE_SHIFT)) {
+ vm_page_stats_reusable.partial_reuse_calls++;
+ for (cur_offset = start_offset;
+ cur_offset < end_offset;
+ cur_offset += PAGE_SIZE_64) {
+ if (object->reusable_page_count == 0) {
+ break;
+ }
+ m = vm_page_lookup(object, cur_offset);
+ VM_OBJECT_REUSE_PAGE(object, m, reused);
+ }
+ } else {
+ vm_page_stats_reusable.partial_reuse_calls++;
+ queue_iterate(&object->memq, m, vm_page_t, listq) {
+ if (object->reusable_page_count == 0) {
+ break;
+ }
+ if (m->offset < start_offset ||
+ m->offset >= end_offset) {
+ continue;
+ }
+ VM_OBJECT_REUSE_PAGE(object, m, reused);
+ }
+ }
+
+ /* update global stats */
+ OSAddAtomic(reusable-reused, &vm_page_stats_reusable.reusable_count);
+ vm_page_stats_reusable.reused += reused;
+ vm_page_stats_reusable.reusable += reusable;
}
/*
@@ -1413,21 +2643,44 @@ __private_extern__ void
vm_object_pmap_protect(
register vm_object_t object,
register vm_object_offset_t offset,
- vm_size_t size,
+ vm_object_size_t size,
pmap_t pmap,
- vm_offset_t pmap_start,
+ vm_map_offset_t pmap_start,
vm_prot_t prot)
{
if (object == VM_OBJECT_NULL)
return;
- size = round_page_64(size);
- offset = trunc_page_64(offset);
+ size = vm_object_round_page(size);
+ offset = vm_object_trunc_page(offset);
vm_object_lock(object);
+ if (object->phys_contiguous) {
+ if (pmap != NULL) {
+ vm_object_unlock(object);
+ pmap_protect(pmap, pmap_start, pmap_start + size, prot);
+ } else {
+ vm_object_offset_t phys_start, phys_end, phys_addr;
+
+ phys_start = object->shadow_offset + offset;
+ phys_end = phys_start + size;
+ assert(phys_start <= phys_end);
+ assert(phys_end <= object->shadow_offset + object->size);
+ vm_object_unlock(object);
+
+ for (phys_addr = phys_start;
+ phys_addr < phys_end;
+ phys_addr += PAGE_SIZE_64) {
+ pmap_page_protect((ppnum_t) (phys_addr >> PAGE_SHIFT), prot);
+ }
+ }
+ return;
+ }
+
+ assert(object->internal);
+
while (TRUE) {
- if (object->resident_page_count > atop_32(size) / 2 &&
- pmap != PMAP_NULL) {
+ if (ptoa_64(object->resident_page_count) > size/2 && pmap != PMAP_NULL) {
vm_object_unlock(object);
pmap_protect(pmap, pmap_start, pmap_start + size, prot);
return;
@@ -1436,7 +2689,7 @@ vm_object_pmap_protect(
/* if we are doing large ranges with respect to resident */
/* page count then we should interate over pages otherwise */
/* inverse page look-up will be faster */
- if ((object->resident_page_count / 4) < atop_32(size)) {
+ if (ptoa_64(object->resident_page_count / 4) < size) {
vm_page_t p;
vm_object_offset_t end;
@@ -1446,11 +2699,10 @@ vm_object_pmap_protect(
queue_iterate(&object->memq, p, vm_page_t, listq) {
if (!p->fictitious &&
(offset <= p->offset) && (p->offset < end)) {
+ vm_map_offset_t start;
- vm_offset_t start = pmap_start +
- (vm_offset_t)(p->offset - offset);
-
- pmap_protect(pmap, start, start + PAGE_SIZE, prot);
+ start = pmap_start + p->offset - offset;
+ pmap_protect(pmap, start, start + PAGE_SIZE_64, prot);
}
}
} else {
@@ -1458,8 +2710,7 @@ vm_object_pmap_protect(
if (!p->fictitious &&
(offset <= p->offset) && (p->offset < end)) {
- pmap_page_protect(p->phys_page,
- prot & ~p->page_lock);
+ pmap_page_protect(p->phys_page, prot);
}
}
}
@@ -1472,20 +2723,23 @@ vm_object_pmap_protect(
if (pmap != PMAP_NULL) {
for(target_off = offset;
- target_off < end; target_off += PAGE_SIZE) {
- if(p = vm_page_lookup(object, target_off)) {
- vm_offset_t start = pmap_start +
- (vm_offset_t)(p->offset - offset);
+ target_off < end;
+ target_off += PAGE_SIZE) {
+ p = vm_page_lookup(object, target_off);
+ if (p != VM_PAGE_NULL) {
+ vm_object_offset_t start;
+ start = pmap_start +
+ (p->offset - offset);
pmap_protect(pmap, start,
- start + PAGE_SIZE, prot);
+ start + PAGE_SIZE, prot);
}
}
} else {
for(target_off = offset;
target_off < end; target_off += PAGE_SIZE) {
- if(p = vm_page_lookup(object, target_off)) {
- pmap_page_protect(p->phys_page,
- prot & ~p->page_lock);
+ p = vm_page_lookup(object, target_off);
+ if (p != VM_PAGE_NULL) {
+ pmap_page_protect(p->phys_page, prot);
}
}
}
@@ -1565,8 +2819,7 @@ vm_object_copy_slowly(
vm_object_t new_object;
vm_object_offset_t new_offset;
- vm_object_offset_t src_lo_offset = src_offset;
- vm_object_offset_t src_hi_offset = src_offset + size;
+ struct vm_object_fault_info fault_info;
XPR(XPR_VM_OBJECT, "v_o_c_slowly obj 0x%x off 0x%x size 0x%x\n",
src_object, src_offset, size, 0, 0);
@@ -1581,9 +2834,7 @@ vm_object_copy_slowly(
* Prevent destruction of the source object while we copy.
*/
- assert(src_object->ref_count > 0);
- src_object->ref_count++;
- VM_OBJ_RES_INCR(src_object);
+ vm_object_reference_locked(src_object);
vm_object_unlock(src_object);
/*
@@ -1600,6 +2851,15 @@ vm_object_copy_slowly(
assert(size == trunc_page_64(size)); /* Will the loop terminate? */
+ fault_info.interruptible = interruptible;
+ fault_info.behavior = VM_BEHAVIOR_SEQUENTIAL;
+ fault_info.user_tag = 0;
+ fault_info.lo_offset = src_offset;
+ fault_info.hi_offset = src_offset + size;
+ fault_info.no_cache = FALSE;
+ fault_info.stealth = TRUE;
+ fault_info.mark_zf_absent = FALSE;
+
for ( ;
size != 0 ;
src_offset += PAGE_SIZE_64,
@@ -1608,14 +2868,22 @@ vm_object_copy_slowly(
vm_page_t new_page;
vm_fault_return_t result;
+ vm_object_lock(new_object);
+
while ((new_page = vm_page_alloc(new_object, new_offset))
== VM_PAGE_NULL) {
+
+ vm_object_unlock(new_object);
+
if (!vm_page_wait(interruptible)) {
vm_object_deallocate(new_object);
+ vm_object_deallocate(src_object);
*_result_object = VM_OBJECT_NULL;
return(MACH_SEND_INTERRUPTED);
}
+ vm_object_lock(new_object);
}
+ vm_object_unlock(new_object);
do {
vm_prot_t prot = VM_PROT_READ;
@@ -1628,99 +2896,121 @@ vm_object_copy_slowly(
vm_object_lock(src_object);
vm_object_paging_begin(src_object);
+ if (size > (vm_size_t) -1) {
+ /* 32-bit overflow */
+ fault_info.cluster_size = (vm_size_t) (0 - PAGE_SIZE);
+ } else {
+ fault_info.cluster_size = (vm_size_t) size;
+ assert(fault_info.cluster_size == size);
+ }
+
XPR(XPR_VM_FAULT,"vm_object_copy_slowly -> vm_fault_page",0,0,0,0,0);
result = vm_fault_page(src_object, src_offset,
- VM_PROT_READ, FALSE, interruptible,
- src_lo_offset, src_hi_offset,
- VM_BEHAVIOR_SEQUENTIAL,
+ VM_PROT_READ, FALSE,
&prot, &_result_page, &top_page,
(int *)0,
- &error_code, FALSE, FALSE, NULL, 0);
+ &error_code, FALSE, FALSE, &fault_info);
switch(result) {
- case VM_FAULT_SUCCESS:
- result_page = _result_page;
+ case VM_FAULT_SUCCESS:
+ result_page = _result_page;
- /*
- * We don't need to hold the object
- * lock -- the busy page will be enough.
- * [We don't care about picking up any
- * new modifications.]
- *
- * Copy the page to the new object.
- *
- * POLICY DECISION:
- * If result_page is clean,
- * we could steal it instead
- * of copying.
- */
+ /*
+ * We don't need to hold the object
+ * lock -- the busy page will be enough.
+ * [We don't care about picking up any
+ * new modifications.]
+ *
+ * Copy the page to the new object.
+ *
+ * POLICY DECISION:
+ * If result_page is clean,
+ * we could steal it instead
+ * of copying.
+ */
- vm_object_unlock(result_page->object);
- vm_page_copy(result_page, new_page);
+ vm_object_unlock(result_page->object);
+ vm_page_copy(result_page, new_page);
- /*
- * Let go of both pages (make them
- * not busy, perform wakeup, activate).
- */
+ /*
+ * Let go of both pages (make them
+ * not busy, perform wakeup, activate).
+ */
+ vm_object_lock(new_object);
+ new_page->dirty = TRUE;
+ PAGE_WAKEUP_DONE(new_page);
+ vm_object_unlock(new_object);
+
+ vm_object_lock(result_page->object);
+ PAGE_WAKEUP_DONE(result_page);
+
+ vm_page_lockspin_queues();
+ if (!result_page->active &&
+ !result_page->inactive &&
+ !result_page->throttled)
+ vm_page_activate(result_page);
+ vm_page_activate(new_page);
+ vm_page_unlock_queues();
- new_page->busy = FALSE;
- new_page->dirty = TRUE;
- vm_object_lock(result_page->object);
- PAGE_WAKEUP_DONE(result_page);
+ /*
+ * Release paging references and
+ * top-level placeholder page, if any.
+ */
- vm_page_lock_queues();
- if (!result_page->active &&
- !result_page->inactive)
- vm_page_activate(result_page);
- vm_page_activate(new_page);
- vm_page_unlock_queues();
+ vm_fault_cleanup(result_page->object,
+ top_page);
- /*
- * Release paging references and
- * top-level placeholder page, if any.
- */
+ break;
+
+ case VM_FAULT_RETRY:
+ break;
- vm_fault_cleanup(result_page->object,
- top_page);
+ case VM_FAULT_FICTITIOUS_SHORTAGE:
+ vm_page_more_fictitious();
+ break;
+ case VM_FAULT_MEMORY_SHORTAGE:
+ if (vm_page_wait(interruptible))
break;
-
- case VM_FAULT_RETRY:
- break;
+ /* fall thru */
- case VM_FAULT_FICTITIOUS_SHORTAGE:
- vm_page_more_fictitious();
- break;
+ case VM_FAULT_INTERRUPTED:
+ vm_object_lock(new_object);
+ VM_PAGE_FREE(new_page);
+ vm_object_unlock(new_object);
+
+ vm_object_deallocate(new_object);
+ vm_object_deallocate(src_object);
+ *_result_object = VM_OBJECT_NULL;
+ return(MACH_SEND_INTERRUPTED);
- case VM_FAULT_MEMORY_SHORTAGE:
- if (vm_page_wait(interruptible))
- break;
- /* fall thru */
+ case VM_FAULT_SUCCESS_NO_VM_PAGE:
+ /* success but no VM page: fail */
+ vm_object_paging_end(src_object);
+ vm_object_unlock(src_object);
+ /*FALLTHROUGH*/
+ case VM_FAULT_MEMORY_ERROR:
+ /*
+ * A policy choice:
+ * (a) ignore pages that we can't
+ * copy
+ * (b) return the null object if
+ * any page fails [chosen]
+ */
- case VM_FAULT_INTERRUPTED:
- vm_page_free(new_page);
- vm_object_deallocate(new_object);
- vm_object_deallocate(src_object);
- *_result_object = VM_OBJECT_NULL;
- return(MACH_SEND_INTERRUPTED);
+ vm_object_lock(new_object);
+ VM_PAGE_FREE(new_page);
+ vm_object_unlock(new_object);
- case VM_FAULT_MEMORY_ERROR:
- /*
- * A policy choice:
- * (a) ignore pages that we can't
- * copy
- * (b) return the null object if
- * any page fails [chosen]
- */
+ vm_object_deallocate(new_object);
+ vm_object_deallocate(src_object);
+ *_result_object = VM_OBJECT_NULL;
+ return(error_code ? error_code:
+ KERN_MEMORY_ERROR);
- vm_page_lock_queues();
- vm_page_free(new_page);
- vm_page_unlock_queues();
- vm_object_deallocate(new_object);
- vm_object_deallocate(src_object);
- *_result_object = VM_OBJECT_NULL;
- return(error_code ? error_code:
- KERN_MEMORY_ERROR);
+ default:
+ panic("vm_object_copy_slowly: unexpected error"
+ " 0x%x from vm_fault_page()\n", result);
}
} while (result != VM_FAULT_SUCCESS);
}
@@ -1728,7 +3018,6 @@ vm_object_copy_slowly(
/*
* Lose the extra reference, and return our object.
*/
-
vm_object_deallocate(src_object);
*_result_object = new_object;
return(KERN_SUCCESS);
@@ -1755,8 +3044,8 @@ vm_object_copy_slowly(
__private_extern__ boolean_t
vm_object_copy_quickly(
vm_object_t *_object, /* INOUT */
- vm_object_offset_t offset, /* IN */
- vm_object_size_t size, /* IN */
+ __unused vm_object_offset_t offset, /* IN */
+ __unused vm_object_size_t size, /* IN */
boolean_t *_src_needs_copy, /* OUT */
boolean_t *_dst_needs_copy) /* OUT */
{
@@ -1784,9 +3073,7 @@ vm_object_copy_quickly(
* Leave object/offset unchanged.
*/
- assert(object->ref_count > 0);
- object->ref_count++;
- vm_object_res_reference(object);
+ vm_object_reference_locked(object);
object->shadowed = TRUE;
vm_object_unlock(object);
@@ -1844,6 +3131,7 @@ vm_object_copy_call(
kern_return_t kr;
vm_object_t copy;
boolean_t check_ready = FALSE;
+ uint32_t try_failed_count = 0;
/*
* If a copy is already in progress, wait and retry.
@@ -1903,7 +3191,10 @@ Retry:
copy = src_object->copy;
if (!vm_object_lock_try(copy)) {
vm_object_unlock(src_object);
- mutex_pause(); /* wait a bit */
+
+ try_failed_count++;
+ mutex_pause(try_failed_count); /* wait a bit */
+
vm_object_lock(src_object);
goto Retry;
}
@@ -1936,8 +3227,6 @@ static int copy_delayed_lock_collisions = 0;
static int copy_delayed_max_collisions = 0;
static int copy_delayed_lock_contention = 0;
static int copy_delayed_protect_iterate = 0;
-static int copy_delayed_protect_lookup = 0;
-static int copy_delayed_protect_lookup_wait = 0;
/*
* Routine: vm_object_copy_delayed [internal]
@@ -1957,13 +3246,15 @@ __private_extern__ vm_object_t
vm_object_copy_delayed(
vm_object_t src_object,
vm_object_offset_t src_offset,
- vm_object_size_t size)
+ vm_object_size_t size,
+ boolean_t src_object_shared)
{
vm_object_t new_copy = VM_OBJECT_NULL;
vm_object_t old_copy;
vm_page_t p;
vm_object_size_t copy_size = src_offset + size;
+
int collisions = 0;
/*
* The user-level memory manager wants to see all of the changes
@@ -2009,9 +3300,17 @@ vm_object_copy_delayed(
/*
* Wait for paging in progress.
*/
- if (!src_object->true_share)
+ if (!src_object->true_share &&
+ (src_object->paging_in_progress != 0 ||
+ src_object->activity_in_progress != 0)) {
+ if (src_object_shared == TRUE) {
+ vm_object_unlock(src_object);
+ vm_object_lock(src_object);
+ src_object_shared = FALSE;
+ goto Retry;
+ }
vm_object_paging_wait(src_object, THREAD_UNINT);
-
+ }
/*
* See whether we can reuse the result of a previous
* copy operation.
@@ -2019,22 +3318,34 @@ vm_object_copy_delayed(
old_copy = src_object->copy;
if (old_copy != VM_OBJECT_NULL) {
+ int lock_granted;
+
/*
* Try to get the locks (out of order)
*/
- if (!vm_object_lock_try(old_copy)) {
+ if (src_object_shared == TRUE)
+ lock_granted = vm_object_lock_try_shared(old_copy);
+ else
+ lock_granted = vm_object_lock_try(old_copy);
+
+ if (!lock_granted) {
vm_object_unlock(src_object);
- mutex_pause();
- /* Heisenberg Rules */
- copy_delayed_lock_collisions++;
if (collisions++ == 0)
copy_delayed_lock_contention++;
+ mutex_pause(collisions);
+
+ /* Heisenberg Rules */
+ copy_delayed_lock_collisions++;
if (collisions > copy_delayed_max_collisions)
copy_delayed_max_collisions = collisions;
- vm_object_lock(src_object);
+ if (src_object_shared == TRUE)
+ vm_object_lock_shared(src_object);
+ else
+ vm_object_lock(src_object);
+
goto Retry;
}
@@ -2054,12 +3365,15 @@ vm_object_copy_delayed(
* needed).
*/
- if (new_copy != VM_OBJECT_NULL) {
- vm_object_unlock(new_copy);
- vm_object_deallocate(new_copy);
- }
-
if (old_copy->size < copy_size) {
+ if (src_object_shared == TRUE) {
+ vm_object_unlock(old_copy);
+ vm_object_unlock(src_object);
+
+ vm_object_lock(src_object);
+ src_object_shared = FALSE;
+ goto Retry;
+ }
/*
* We can't perform a delayed copy if any of the
* pages in the extended range are wired (because
@@ -2068,29 +3382,44 @@ vm_object_copy_delayed(
* go ahead and protect them.
*/
copy_delayed_protect_iterate++;
+
queue_iterate(&src_object->memq, p, vm_page_t, listq) {
if (!p->fictitious &&
p->offset >= old_copy->size &&
p->offset < copy_size) {
- if (p->wire_count > 0) {
+ if (VM_PAGE_WIRED(p)) {
vm_object_unlock(old_copy);
vm_object_unlock(src_object);
+
+ if (new_copy != VM_OBJECT_NULL) {
+ vm_object_unlock(new_copy);
+ vm_object_deallocate(new_copy);
+ }
+
return VM_OBJECT_NULL;
} else {
pmap_page_protect(p->phys_page,
- (VM_PROT_ALL & ~VM_PROT_WRITE &
- ~p->page_lock));
+ (VM_PROT_ALL & ~VM_PROT_WRITE));
}
}
}
old_copy->size = copy_size;
}
-
- vm_object_reference_locked(old_copy);
+ if (src_object_shared == TRUE)
+ vm_object_reference_shared(old_copy);
+ else
+ vm_object_reference_locked(old_copy);
vm_object_unlock(old_copy);
vm_object_unlock(src_object);
+
+ if (new_copy != VM_OBJECT_NULL) {
+ vm_object_unlock(new_copy);
+ vm_object_deallocate(new_copy);
+ }
return(old_copy);
}
+
+
/*
* Adjust the size argument so that the newly-created
@@ -2106,6 +3435,8 @@ vm_object_copy_delayed(
new_copy = vm_object_allocate(copy_size);
vm_object_lock(src_object);
vm_object_lock(new_copy);
+
+ src_object_shared = FALSE;
goto Retry;
}
new_copy->size = copy_size;
@@ -2125,6 +3456,8 @@ vm_object_copy_delayed(
new_copy = vm_object_allocate(copy_size);
vm_object_lock(src_object);
vm_object_lock(new_copy);
+
+ src_object_shared = FALSE;
goto Retry;
}
@@ -2140,9 +3473,10 @@ vm_object_copy_delayed(
* wired, then go ahead and protect them.
*/
copy_delayed_protect_iterate++;
+
queue_iterate(&src_object->memq, p, vm_page_t, listq) {
if (!p->fictitious && p->offset < copy_size) {
- if (p->wire_count > 0) {
+ if (VM_PAGE_WIRED(p)) {
if (old_copy)
vm_object_unlock(old_copy);
vm_object_unlock(src_object);
@@ -2151,12 +3485,10 @@ vm_object_copy_delayed(
return VM_OBJECT_NULL;
} else {
pmap_page_protect(p->phys_page,
- (VM_PROT_ALL & ~VM_PROT_WRITE &
- ~p->page_lock));
+ (VM_PROT_ALL & ~VM_PROT_WRITE));
}
}
}
-
if (old_copy != VM_OBJECT_NULL) {
/*
* Make the old copy-object shadow the new one.
@@ -2164,9 +3496,13 @@ vm_object_copy_delayed(
* object.
*/
- src_object->ref_count--; /* remove ref. from old_copy */
+ /* remove ref. from old_copy */
+ vm_object_lock_assert_exclusive(src_object);
+ src_object->ref_count--;
assert(src_object->ref_count > 0);
+ vm_object_lock_assert_exclusive(old_copy);
old_copy->shadow = new_copy;
+ vm_object_lock_assert_exclusive(new_copy);
assert(new_copy->ref_count > 0);
new_copy->ref_count++; /* for old_copy->shadow ref. */
@@ -2183,21 +3519,22 @@ vm_object_copy_delayed(
/*
* Point the new copy at the existing object.
*/
+ vm_object_lock_assert_exclusive(new_copy);
new_copy->shadow = src_object;
new_copy->shadow_offset = 0;
new_copy->shadowed = TRUE; /* caller must set needs_copy */
- assert(src_object->ref_count > 0);
- src_object->ref_count++;
- VM_OBJ_RES_INCR(src_object);
+
+ vm_object_lock_assert_exclusive(src_object);
+ vm_object_reference_locked(src_object);
src_object->copy = new_copy;
vm_object_unlock(src_object);
vm_object_unlock(new_copy);
XPR(XPR_VM_OBJECT,
"vm_object_copy_delayed: used copy object %X for source %X\n",
- (integer_t)new_copy, (integer_t)src_object, 0, 0, 0);
+ new_copy, src_object, 0, 0, 0);
- return(new_copy);
+ return new_copy;
}
/*
@@ -2219,11 +3556,18 @@ vm_object_copy_strategically(
{
boolean_t result;
boolean_t interruptible = THREAD_ABORTSAFE; /* XXX */
+ boolean_t object_lock_shared = FALSE;
memory_object_copy_strategy_t copy_strategy;
assert(src_object != VM_OBJECT_NULL);
- vm_object_lock(src_object);
+ copy_strategy = src_object->copy_strategy;
+
+ if (copy_strategy == MEMORY_OBJECT_COPY_DELAY) {
+ vm_object_lock_shared(src_object);
+ object_lock_shared = TRUE;
+ } else
+ vm_object_lock(src_object);
/*
* The copy strategy is only valid if the memory manager
@@ -2233,6 +3577,12 @@ vm_object_copy_strategically(
while (!src_object->internal && !src_object->pager_ready) {
wait_result_t wait_result;
+ if (object_lock_shared == TRUE) {
+ vm_object_unlock(src_object);
+ vm_object_lock(src_object);
+ object_lock_shared = FALSE;
+ continue;
+ }
wait_result = vm_object_sleep( src_object,
VM_OBJECT_EVENT_PAGER_READY,
interruptible);
@@ -2245,8 +3595,6 @@ vm_object_copy_strategically(
}
}
- copy_strategy = src_object->copy_strategy;
-
/*
* Use the appropriate copy strategy.
*/
@@ -2254,7 +3602,7 @@ vm_object_copy_strategically(
switch (copy_strategy) {
case MEMORY_OBJECT_COPY_DELAY:
*dst_object = vm_object_copy_delayed(src_object,
- src_offset, size);
+ src_offset, size, object_lock_shared);
if (*dst_object != VM_OBJECT_NULL) {
*dst_offset = src_offset;
*dst_needs_copy = TRUE;
@@ -2283,7 +3631,7 @@ vm_object_copy_strategically(
break;
case MEMORY_OBJECT_COPY_SYMMETRIC:
- XPR(XPR_VM_OBJECT, "v_o_c_strategically obj 0x%x off 0x%x size 0x%x\n",(natural_t)src_object, src_offset, size, 0, 0);
+ XPR(XPR_VM_OBJECT, "v_o_c_strategically obj 0x%x off 0x%x size 0x%x\n", src_object, src_offset, size, 0, 0);
vm_object_unlock(src_object);
result = KERN_MEMORY_RESTART_COPY;
break;
@@ -2317,7 +3665,22 @@ vm_object_shadow(
register vm_object_t result;
source = *object;
+#if 0
+ /*
+ * XXX FBDP
+ * This assertion is valid but it gets triggered by Rosetta for example
+ * due to a combination of vm_remap() that changes a VM object's
+ * copy_strategy from SYMMETRIC to DELAY and vm_protect(VM_PROT_COPY)
+ * that then sets "needs_copy" on its map entry. This creates a
+ * mapping situation that VM should never see and doesn't know how to
+ * handle.
+ * It's not clear if this can create any real problem but we should
+ * look into fixing this, probably by having vm_protect(VM_PROT_COPY)
+ * do more than just set "needs_copy" to handle the copy-on-write...
+ * In the meantime, let's disable the assertion.
+ */
assert(source->copy_strategy == MEMORY_OBJECT_COPY_SYMMETRIC);
+#endif
/*
* Determine if we really need a shadow.
@@ -2422,9 +3785,9 @@ vm_object_shadow(
* [Furthermore, each routine must cope with the simultaneous
* or previous operations of the others.]
*
- * In addition to the lock on the object, the vm_object_cache_lock
+ * In addition to the lock on the object, the vm_object_hash_lock
* governs the associations. References gained through the
- * association require use of the cache lock.
+ * association require use of the hash lock.
*
* Because the pager field may be cleared spontaneously, it
* cannot be used to determine whether a memory object has
@@ -2442,102 +3805,6 @@ vm_object_shadow(
* termination routines and vm_object_collapse.]
*/
-#if 0
-/*
- * Routine: vm_object_pager_dead
- *
- * Purpose:
- * A port is being destroy, and the IPC kobject code
- * can't tell if it represents a pager port or not.
- * So this function is called each time it sees a port
- * die.
- * THIS IS HORRIBLY INEFFICIENT. We should only call
- * this routine if we had requested a notification on
- * the port.
- */
-
-__private_extern__ void
-vm_object_pager_dead(
- ipc_port_t pager)
-{
- vm_object_t object;
- vm_object_hash_entry_t entry;
-
- /*
- * Perform essentially the same operations as in vm_object_lookup,
- * except that this time we look up based on the memory_object
- * port, not the control port.
- */
- vm_object_cache_lock();
- entry = vm_object_hash_lookup(pager, FALSE);
- if (entry == VM_OBJECT_HASH_ENTRY_NULL ||
- entry->object == VM_OBJECT_NULL) {
- vm_object_cache_unlock();
- return;
- }
-
- object = entry->object;
- entry->object = VM_OBJECT_NULL;
-
- vm_object_lock(object);
- if (object->ref_count == 0) {
- XPR(XPR_VM_OBJECT_CACHE,
- "vm_object_destroy: removing %x from cache, head (%x, %x)\n",
- (integer_t)object,
- (integer_t)vm_object_cached_list.next,
- (integer_t)vm_object_cached_list.prev, 0,0);
-
- queue_remove(&vm_object_cached_list, object,
- vm_object_t, cached_list);
- vm_object_cached_count--;
- }
- object->ref_count++;
- vm_object_res_reference(object);
-
- object->can_persist = FALSE;
-
- assert(object->pager == pager);
-
- /*
- * Remove the pager association.
- *
- * Note that the memory_object itself is dead, so
- * we don't bother with it.
- */
-
- object->pager = MEMORY_OBJECT_NULL;
-
- vm_object_unlock(object);
- vm_object_cache_unlock();
-
- vm_object_pager_wakeup(pager);
-
- /*
- * Release the pager reference. Note that there's no
- * point in trying the memory_object_terminate call
- * because the memory_object itself is dead. Also
- * release the memory_object_control reference, since
- * the pager didn't do that either.
- */
-
- memory_object_deallocate(pager);
- memory_object_control_deallocate(object->pager_request);
-
-
- /*
- * Restart pending page requests
- */
- vm_object_lock(object);
- vm_object_abort_activity(object);
- vm_object_unlock(object);
-
- /*
- * Lose the object reference.
- */
-
- vm_object_deallocate(object);
-}
-#endif
/*
* Routine: vm_object_enter
@@ -2558,6 +3825,8 @@ vm_object_enter(
vm_object_t new_object;
boolean_t must_init;
vm_object_hash_entry_t entry, new_entry;
+ uint32_t try_failed_count = 0;
+ lck_mtx_t *lck;
if (pager == MEMORY_OBJECT_NULL)
return(vm_object_allocate(size));
@@ -2569,100 +3838,107 @@ vm_object_enter(
/*
* Look for an object associated with this port.
*/
-
-restart:
- vm_object_cache_lock();
- for (;;) {
+Retry:
+ lck = vm_object_hash_lock_spin(pager);
+ do {
entry = vm_object_hash_lookup(pager, FALSE);
- /*
- * If a previous object is being terminated,
- * we must wait for the termination message
- * to be queued.
- *
- * We set kobject to a non-null value to let the
- * terminator know that someone is waiting.
- * Among the possibilities is that the port
- * could die while we're waiting. Must restart
- * instead of continuing the loop.
- */
-
- if (entry != VM_OBJECT_HASH_ENTRY_NULL) {
- if (entry->object != VM_OBJECT_NULL)
- break;
-
- entry->waiting = TRUE;
- assert_wait((event_t) pager, THREAD_UNINT);
- vm_object_cache_unlock();
- thread_block((void (*)(void))0);
- goto restart;
- }
-
- /*
- * We must unlock to create a new object;
- * if we do so, we must try the lookup again.
- */
-
- if (new_object == VM_OBJECT_NULL) {
- vm_object_cache_unlock();
- assert(new_entry == VM_OBJECT_HASH_ENTRY_NULL);
- new_entry = vm_object_hash_entry_alloc(pager);
- new_object = vm_object_allocate(size);
- vm_object_cache_lock();
- } else {
- /*
- * Lookup failed twice, and we have something
- * to insert; set the object.
- */
-
- if (entry == VM_OBJECT_HASH_ENTRY_NULL) {
- vm_object_hash_insert(new_entry);
+ if (entry == VM_OBJECT_HASH_ENTRY_NULL) {
+ if (new_object == VM_OBJECT_NULL) {
+ /*
+ * We must unlock to create a new object;
+ * if we do so, we must try the lookup again.
+ */
+ vm_object_hash_unlock(lck);
+ assert(new_entry == VM_OBJECT_HASH_ENTRY_NULL);
+ new_entry = vm_object_hash_entry_alloc(pager);
+ new_object = vm_object_allocate(size);
+ lck = vm_object_hash_lock_spin(pager);
+ } else {
+ /*
+ * Lookup failed twice, and we have something
+ * to insert; set the object.
+ */
+ vm_object_hash_insert(new_entry, new_object);
entry = new_entry;
new_entry = VM_OBJECT_HASH_ENTRY_NULL;
+ new_object = VM_OBJECT_NULL;
+ must_init = TRUE;
}
+ } else if (entry->object == VM_OBJECT_NULL) {
+ /*
+ * If a previous object is being terminated,
+ * we must wait for the termination message
+ * to be queued (and lookup the entry again).
+ */
+ entry->waiting = TRUE;
+ entry = VM_OBJECT_HASH_ENTRY_NULL;
+ assert_wait((event_t) pager, THREAD_UNINT);
+ vm_object_hash_unlock(lck);
- entry->object = new_object;
- new_object = VM_OBJECT_NULL;
- must_init = TRUE;
+ thread_block(THREAD_CONTINUE_NULL);
+ lck = vm_object_hash_lock_spin(pager);
}
- }
+ } while (entry == VM_OBJECT_HASH_ENTRY_NULL);
object = entry->object;
assert(object != VM_OBJECT_NULL);
if (!must_init) {
- vm_object_lock(object);
- assert(object->pager_created);
- assert(!internal || object->internal);
- if (named) {
- assert(!object->named);
- object->named = TRUE;
+ if ( !vm_object_lock_try(object)) {
+
+ vm_object_hash_unlock(lck);
+
+ try_failed_count++;
+ mutex_pause(try_failed_count); /* wait a bit */
+ goto Retry;
}
+ assert(!internal || object->internal);
+#if VM_OBJECT_CACHE
if (object->ref_count == 0) {
+ if ( !vm_object_cache_lock_try()) {
+
+ vm_object_hash_unlock(lck);
+ vm_object_unlock(object);
+
+ try_failed_count++;
+ mutex_pause(try_failed_count); /* wait a bit */
+ goto Retry;
+ }
XPR(XPR_VM_OBJECT_CACHE,
- "vm_object_enter: removing %x from cache, head (%x, %x)\n",
- (integer_t)object,
- (integer_t)vm_object_cached_list.next,
- (integer_t)vm_object_cached_list.prev, 0,0);
+ "vm_object_enter: removing %x from cache, head (%x, %x)\n",
+ object,
+ vm_object_cached_list.next,
+ vm_object_cached_list.prev, 0,0);
queue_remove(&vm_object_cached_list, object,
vm_object_t, cached_list);
vm_object_cached_count--;
+
+ vm_object_cache_unlock();
+ }
+#endif
+ if (named) {
+ assert(!object->named);
+ object->named = TRUE;
}
+ vm_object_lock_assert_exclusive(object);
object->ref_count++;
vm_object_res_reference(object);
+
+ vm_object_hash_unlock(lck);
vm_object_unlock(object);
- VM_STAT(hits++);
- }
- assert(object->ref_count > 0);
+ VM_STAT_INCR(hits);
+ } else
+ vm_object_hash_unlock(lck);
- VM_STAT(lookups++);
+ assert(object->ref_count > 0);
- vm_object_cache_unlock();
+ VM_STAT_INCR(lookups);
XPR(XPR_VM_OBJECT,
"vm_o_enter: pager 0x%x obj 0x%x must_init %d\n",
- (integer_t)pager, (integer_t)object, must_init, 0, 0);
+ pager, object, must_init, 0, 0);
/*
* If we raced to create a vm_object but lost, let's
@@ -2676,16 +3952,17 @@ restart:
vm_object_hash_entry_free(new_entry);
if (must_init) {
- pager_request_t pager_request;
+ memory_object_control_t control;
/*
* Allocate request port.
*/
- pager_request = memory_object_control_allocate(object);
- assert (pager_request != PAGER_REQUEST_NULL);
+ control = memory_object_control_allocate(object);
+ assert (control != MEMORY_OBJECT_CONTROL_NULL);
vm_object_lock(object);
+ assert(object != kernel_object);
/*
* Copy the reference we were given.
@@ -2700,7 +3977,7 @@ restart:
/* copy strategy invalid until set by memory manager */
object->copy_strategy = MEMORY_OBJECT_COPY_INVALID;
}
- object->pager_request = pager_request;
+ object->pager_control = control;
object->pager_ready = FALSE;
vm_object_unlock(object);
@@ -2710,7 +3987,7 @@ restart:
*/
(void) memory_object_init(pager,
- object->pager_request,
+ object->pager_control,
PAGE_SIZE);
vm_object_lock(object);
@@ -2745,7 +4022,7 @@ restart:
XPR(XPR_VM_OBJECT,
"vm_object_enter: vm_object %x, memory_object %x, internal %d\n",
- (integer_t)object, (integer_t)object->pager, internal, 0,0);
+ object, object->pager, internal, 0,0);
return(object);
}
@@ -2769,13 +4046,16 @@ vm_object_pager_create(
{
memory_object_t pager;
vm_object_hash_entry_t entry;
+ lck_mtx_t *lck;
#if MACH_PAGEMAP
vm_object_size_t size;
vm_external_map_t map;
#endif /* MACH_PAGEMAP */
XPR(XPR_VM_OBJECT, "vm_object_pager_create, object 0x%X\n",
- (integer_t)object, 0,0,0,0);
+ object, 0,0,0,0);
+
+ assert(object != kernel_object);
if (memory_manager_default_check() != KERN_SUCCESS)
return;
@@ -2820,6 +4100,11 @@ vm_object_pager_create(
vm_object_unlock(object);
#endif /* MACH_PAGEMAP */
+ if ((uint32_t) object->size != object->size) {
+ panic("vm_object_pager_create(): object size 0x%llx >= 4GB\n",
+ (uint64_t) object->size);
+ }
+
/*
* Create the [internal] pager, and associate it with this object.
*
@@ -2829,28 +4114,25 @@ vm_object_pager_create(
*/
{
memory_object_default_t dmm;
- vm_size_t cluster_size;
/* acquire a reference for the default memory manager */
- dmm = memory_manager_default_reference(&cluster_size);
- assert(cluster_size >= PAGE_SIZE);
+ dmm = memory_manager_default_reference();
- object->cluster_size = cluster_size; /* XXX ??? */
assert(object->temporary);
/* create our new memory object */
- (void) memory_object_create(dmm, object->size, &pager);
+ assert((vm_size_t) object->size == object->size);
+ (void) memory_object_create(dmm, (vm_size_t) object->size,
+ &pager);
memory_object_default_deallocate(dmm);
}
entry = vm_object_hash_entry_alloc(pager);
- vm_object_cache_lock();
- vm_object_hash_insert(entry);
-
- entry->object = object;
- vm_object_cache_unlock();
+ lck = vm_object_hash_lock_spin(pager);
+ vm_object_hash_insert(entry, object);
+ vm_object_hash_unlock(lck);
/*
* A reference was returned by
@@ -2887,7 +4169,6 @@ vm_object_remove(
vm_object_t object)
{
memory_object_t pager;
- pager_request_t pager_request;
if ((pager = object->pager) != MEMORY_OBJECT_NULL) {
vm_object_hash_entry_t entry;
@@ -2911,8 +4192,12 @@ static long object_bypasses = 0;
static boolean_t vm_object_collapse_allowed = TRUE;
static boolean_t vm_object_bypass_allowed = TRUE;
+#if MACH_PAGEMAP
static int vm_external_discarded;
static int vm_external_collapsed;
+#endif
+
+unsigned long vm_object_collapse_encrypted = 0;
/*
* Routine: vm_object_do_collapse
@@ -2934,6 +4219,9 @@ vm_object_do_collapse(
vm_object_offset_t new_offset, backing_offset;
vm_object_size_t size;
+ vm_object_lock_assert_exclusive(object);
+ vm_object_lock_assert_exclusive(backing_object);
+
backing_offset = object->shadow_offset;
size = object->size;
@@ -2951,7 +4239,7 @@ vm_object_do_collapse(
new_offset = (p->offset - backing_offset);
assert(!p->busy || p->absent);
-
+
/*
* If the parent has a page here, or if
* this page falls outside the parent,
@@ -2963,6 +4251,18 @@ vm_object_do_collapse(
if (p->offset < backing_offset || new_offset >= size) {
VM_PAGE_FREE(p);
} else {
+ /*
+ * ENCRYPTED SWAP:
+ * The encryption key includes the "pager" and the
+ * "paging_offset". These will not change during the
+ * object collapse, so we can just move an encrypted
+ * page from one object to the other in this case.
+ * We can't decrypt the page here, since we can't drop
+ * the object lock.
+ */
+ if (p->encrypted) {
+ vm_object_collapse_encrypted++;
+ }
pp = vm_page_lookup(object, new_offset);
if (pp == VM_PAGE_NULL) {
@@ -2971,7 +4271,7 @@ vm_object_do_collapse(
* Move the backing object's page up.
*/
- vm_page_rename(p, object, new_offset);
+ vm_page_rename(p, object, new_offset, TRUE);
#if MACH_PAGEMAP
} else if (pp->absent) {
@@ -2988,7 +4288,7 @@ vm_object_do_collapse(
*/
VM_PAGE_FREE(pp);
- vm_page_rename(p, object, new_offset);
+ vm_page_rename(p, object, new_offset, TRUE);
#endif /* MACH_PAGEMAP */
} else {
assert(! pp->absent);
@@ -3003,8 +4303,13 @@ vm_object_do_collapse(
}
}
- assert(object->pager == MEMORY_OBJECT_NULL ||
- backing_object->pager == MEMORY_OBJECT_NULL);
+#if !MACH_PAGEMAP
+ assert((!object->pager_created && (object->pager == MEMORY_OBJECT_NULL))
+ || (!backing_object->pager_created
+ && (backing_object->pager == MEMORY_OBJECT_NULL)));
+#else
+ assert(!object->pager_created && object->pager == MEMORY_OBJECT_NULL);
+#endif /* !MACH_PAGEMAP */
if (backing_object->pager != MEMORY_OBJECT_NULL) {
vm_object_hash_entry_t entry;
@@ -3017,27 +4322,33 @@ vm_object_do_collapse(
* unused portion.
*/
+ assert(!object->paging_in_progress);
+ assert(!object->activity_in_progress);
object->pager = backing_object->pager;
- entry = vm_object_hash_lookup(object->pager, FALSE);
- assert(entry != VM_OBJECT_HASH_ENTRY_NULL);
- entry->object = object;
+
+ if (backing_object->hashed) {
+ lck_mtx_t *lck;
+
+ lck = vm_object_hash_lock_spin(backing_object->pager);
+ entry = vm_object_hash_lookup(object->pager, FALSE);
+ assert(entry != VM_OBJECT_HASH_ENTRY_NULL);
+ entry->object = object;
+ vm_object_hash_unlock(lck);
+
+ object->hashed = TRUE;
+ }
object->pager_created = backing_object->pager_created;
- object->pager_request = backing_object->pager_request;
+ object->pager_control = backing_object->pager_control;
object->pager_ready = backing_object->pager_ready;
object->pager_initialized = backing_object->pager_initialized;
- object->cluster_size = backing_object->cluster_size;
object->paging_offset =
backing_object->paging_offset + backing_offset;
- if (object->pager_request != PAGER_REQUEST_NULL) {
- memory_object_control_collapse(object->pager_request,
+ if (object->pager_control != MEMORY_OBJECT_CONTROL_NULL) {
+ memory_object_control_collapse(object->pager_control,
object);
}
}
- vm_object_cache_unlock();
-
- object->paging_offset = backing_object->paging_offset + backing_offset;
-
#if MACH_PAGEMAP
/*
* If the shadow offset is 0, the use the existence map from
@@ -3070,10 +4381,17 @@ vm_object_do_collapse(
* moves from within backing_object to within object.
*/
+ assert(!object->phys_contiguous);
+ assert(!backing_object->phys_contiguous);
object->shadow = backing_object->shadow;
- object->shadow_offset += backing_object->shadow_offset;
+ if (object->shadow) {
+ object->shadow_offset += backing_object->shadow_offset;
+ } else {
+ /* no shadow, therefore no shadow offset... */
+ object->shadow_offset = 0;
+ }
assert((object->shadow == VM_OBJECT_NULL) ||
- (object->shadow->copy == VM_OBJECT_NULL));
+ (object->shadow->copy != backing_object));
/*
* Discard backing_object.
@@ -3085,15 +4403,18 @@ vm_object_do_collapse(
assert((backing_object->ref_count == 1) &&
(backing_object->resident_page_count == 0) &&
- (backing_object->paging_in_progress == 0));
+ (backing_object->paging_in_progress == 0) &&
+ (backing_object->activity_in_progress == 0));
backing_object->alive = FALSE;
vm_object_unlock(backing_object);
XPR(XPR_VM_OBJECT, "vm_object_collapse, collapsed 0x%X\n",
- (integer_t)backing_object, 0,0,0,0);
+ backing_object, 0,0,0,0);
- zfree(vm_object_zone, (vm_offset_t) backing_object);
+ vm_object_lock_destroy(backing_object);
+
+ zfree(vm_object_zone, backing_object);
object_collapses++;
}
@@ -3108,6 +4429,9 @@ vm_object_do_bypass(
* in the chain.
*/
+ vm_object_lock_assert_exclusive(object);
+ vm_object_lock_assert_exclusive(backing_object);
+
#if TASK_SWAPPER
/*
* Do object reference in-line to
@@ -3118,6 +4442,7 @@ vm_object_do_bypass(
*/
if (backing_object->shadow != VM_OBJECT_NULL) {
vm_object_lock(backing_object->shadow);
+ vm_object_lock_assert_exclusive(backing_object->shadow);
backing_object->shadow->ref_count++;
if (object->res_count != 0)
vm_object_res_reference(backing_object->shadow);
@@ -3127,8 +4452,15 @@ vm_object_do_bypass(
vm_object_reference(backing_object->shadow);
#endif /* TASK_SWAPPER */
+ assert(!object->phys_contiguous);
+ assert(!backing_object->phys_contiguous);
object->shadow = backing_object->shadow;
- object->shadow_offset += backing_object->shadow_offset;
+ if (object->shadow) {
+ object->shadow_offset += backing_object->shadow_offset;
+ } else {
+ /* no shadow, therefore no shadow offset... */
+ object->shadow_offset = 0;
+ }
/*
* Backing object might have had a copy pointer
@@ -3144,7 +4476,7 @@ vm_object_do_bypass(
* Since its ref_count was at least 2, it
* will not vanish; so we don't need to call
* vm_object_deallocate.
- * [FBDP: that doesn't seem to be true any more]
+ * [with a caveat for "named" objects]
*
* The res_count on the backing object is
* conditionally decremented. It's possible
@@ -3162,7 +4494,9 @@ vm_object_do_bypass(
* is temporary and cachable.
#endif
*/
- if (backing_object->ref_count > 1) {
+ if (backing_object->ref_count > 2 ||
+ (!backing_object->named && backing_object->ref_count > 1)) {
+ vm_object_lock_assert_exclusive(backing_object);
backing_object->ref_count--;
#if TASK_SWAPPER
if (object->res_count != 0)
@@ -3211,46 +4545,90 @@ vm_object_do_bypass(
* Requires that the object be locked and the page queues be unlocked.
*
*/
+static unsigned long vm_object_collapse_calls = 0;
+static unsigned long vm_object_collapse_objects = 0;
+static unsigned long vm_object_collapse_do_collapse = 0;
+static unsigned long vm_object_collapse_do_bypass = 0;
+static unsigned long vm_object_collapse_delays = 0;
__private_extern__ void
vm_object_collapse(
- register vm_object_t object)
+ register vm_object_t object,
+ register vm_object_offset_t hint_offset,
+ boolean_t can_bypass)
{
register vm_object_t backing_object;
- register vm_object_offset_t backing_offset;
- register vm_object_size_t size;
- register vm_object_offset_t new_offset;
- register vm_page_t p;
+ register unsigned int rcount;
+ register unsigned int size;
+ vm_object_t original_object;
+ int object_lock_type;
+ int backing_object_lock_type;
- vm_offset_t current_offset;
+ vm_object_collapse_calls++;
- if (! vm_object_collapse_allowed && ! vm_object_bypass_allowed) {
+ if (! vm_object_collapse_allowed &&
+ ! (can_bypass && vm_object_bypass_allowed)) {
return;
}
XPR(XPR_VM_OBJECT, "vm_object_collapse, obj 0x%X\n",
- (integer_t)object, 0,0,0,0);
+ object, 0,0,0,0);
- while (TRUE) {
- /*
- * Verify that the conditions are right for either
- * collapse or bypass:
- *
- * The object exists and no pages in it are currently
- * being paged out, and
+ if (object == VM_OBJECT_NULL)
+ return;
+
+ original_object = object;
+
+ /*
+ * The top object was locked "exclusive" by the caller.
+ * In the first pass, to determine if we can collapse the shadow chain,
+ * take a "shared" lock on the shadow objects. If we can collapse,
+ * we'll have to go down the chain again with exclusive locks.
+ */
+ object_lock_type = OBJECT_LOCK_EXCLUSIVE;
+ backing_object_lock_type = OBJECT_LOCK_SHARED;
+
+retry:
+ object = original_object;
+ vm_object_lock_assert_exclusive(object);
+
+ while (TRUE) {
+ vm_object_collapse_objects++;
+ /*
+ * Verify that the conditions are right for either
+ * collapse or bypass:
*/
- if (object == VM_OBJECT_NULL ||
- object->paging_in_progress != 0 ||
- object->absent_count != 0)
- return;
/*
* There is a backing object, and
*/
- if ((backing_object = object->shadow) == VM_OBJECT_NULL)
+ backing_object = object->shadow;
+ if (backing_object == VM_OBJECT_NULL) {
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
return;
-
- vm_object_lock(backing_object);
+ }
+ if (backing_object_lock_type == OBJECT_LOCK_SHARED) {
+ vm_object_lock_shared(backing_object);
+ } else {
+ vm_object_lock(backing_object);
+ }
+
+ /*
+ * No pages in the object are currently
+ * being paged out, and
+ */
+ if (object->paging_in_progress != 0 ||
+ object->activity_in_progress != 0) {
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
+ }
/*
* ...
@@ -3262,9 +4640,15 @@ vm_object_collapse(
*/
if (!backing_object->internal ||
- backing_object->paging_in_progress != 0) {
- vm_object_unlock(backing_object);
- return;
+ backing_object->paging_in_progress != 0 ||
+ backing_object->activity_in_progress != 0) {
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
}
/*
@@ -3278,9 +4662,14 @@ vm_object_collapse(
* parent object.
*/
if (backing_object->shadow != VM_OBJECT_NULL &&
- backing_object->shadow->copy != VM_OBJECT_NULL) {
- vm_object_unlock(backing_object);
- return;
+ backing_object->shadow->copy == backing_object) {
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
}
/*
@@ -3293,33 +4682,54 @@ vm_object_collapse(
* object, we may be able to collapse it into the
* parent.
*
- * The backing object must not have a pager
- * created for it, since collapsing an object
- * into a backing_object dumps new pages into
- * the backing_object that its pager doesn't
- * know about.
+ * If MACH_PAGEMAP is defined:
+ * The parent must not have a pager created for it,
+ * since collapsing a backing_object dumps new pages
+ * into the parent that its pager doesn't know about
+ * (and the collapse code can't merge the existence
+ * maps).
+ * Otherwise:
+ * As long as one of the objects is still not known
+ * to the pager, we can collapse them.
*/
-
if (backing_object->ref_count == 1 &&
- ! object->pager_created &&
- vm_object_collapse_allowed) {
-
- XPR(XPR_VM_OBJECT,
- "vm_object_collapse: %x to %x, pager %x, pager_request %x\n",
- (integer_t)backing_object, (integer_t)object,
- (integer_t)backing_object->pager,
- (integer_t)backing_object->pager_request, 0);
+ (!object->pager_created
+#if !MACH_PAGEMAP
+ || !backing_object->pager_created
+#endif /*!MACH_PAGEMAP */
+ ) && vm_object_collapse_allowed) {
/*
- * We need the cache lock for collapsing,
- * but we must not deadlock.
+ * We need the exclusive lock on the VM objects.
*/
-
- if (! vm_object_cache_lock_try()) {
+ if (backing_object_lock_type != OBJECT_LOCK_EXCLUSIVE) {
+ /*
+ * We have an object and its shadow locked
+ * "shared". We can't just upgrade the locks
+ * to "exclusive", as some other thread might
+ * also have these objects locked "shared" and
+ * attempt to upgrade one or the other to
+ * "exclusive". The upgrades would block
+ * forever waiting for the other "shared" locks
+ * to get released.
+ * So we have to release the locks and go
+ * down the shadow chain again (since it could
+ * have changed) with "exclusive" locking.
+ */
vm_object_unlock(backing_object);
- return;
+ if (object != original_object)
+ vm_object_unlock(object);
+ object_lock_type = OBJECT_LOCK_EXCLUSIVE;
+ backing_object_lock_type = OBJECT_LOCK_EXCLUSIVE;
+ goto retry;
}
+ XPR(XPR_VM_OBJECT,
+ "vm_object_collapse: %x to %x, pager %x, pager_control %x\n",
+ backing_object, object,
+ backing_object->pager,
+ backing_object->pager_control, 0);
+
/*
* Collapse the object with its backing
* object, and try again with the object's
@@ -3327,112 +4737,241 @@ vm_object_collapse(
*/
vm_object_do_collapse(object, backing_object);
+ vm_object_collapse_do_collapse++;
continue;
}
-
/*
* Collapsing the backing object was not possible
* or permitted, so let's try bypassing it.
*/
- if (! vm_object_bypass_allowed) {
- vm_object_unlock(backing_object);
- return;
+ if (! (can_bypass && vm_object_bypass_allowed)) {
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
}
/*
- * If the backing object has a pager but no pagemap,
- * then we cannot bypass it, because we don't know
- * what pages it has.
+ * If the object doesn't have all its pages present,
+ * we have to make sure no pages in the backing object
+ * "show through" before bypassing it.
*/
- if (backing_object->pager_created
+ size = atop(object->size);
+ rcount = object->resident_page_count;
+ if (rcount != size) {
+ vm_object_offset_t offset;
+ vm_object_offset_t backing_offset;
+ unsigned int backing_rcount;
+ unsigned int lookups = 0;
+
+ /*
+ * If the backing object has a pager but no pagemap,
+ * then we cannot bypass it, because we don't know
+ * what pages it has.
+ */
+ if (backing_object->pager_created
#if MACH_PAGEMAP
- && (backing_object->existence_map == VM_EXTERNAL_NULL)
+ && (backing_object->existence_map == VM_EXTERNAL_NULL)
#endif /* MACH_PAGEMAP */
- ) {
- vm_object_unlock(backing_object);
- return;
- }
+ ) {
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
+ }
- /*
- * If the object has a pager but no pagemap,
- * then we cannot bypass it, because we don't know
- * what pages it has.
- */
- if (object->pager_created
+ /*
+ * If the object has a pager but no pagemap,
+ * then we cannot bypass it, because we don't know
+ * what pages it has.
+ */
+ if (object->pager_created
#if MACH_PAGEMAP
- && (object->existence_map == VM_EXTERNAL_NULL)
+ && (object->existence_map == VM_EXTERNAL_NULL)
#endif /* MACH_PAGEMAP */
- ) {
- vm_object_unlock(backing_object);
- return;
- }
+ ) {
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
+ }
- backing_offset = object->shadow_offset;
- size = object->size;
+ /*
+ * If all of the pages in the backing object are
+ * shadowed by the parent object, the parent
+ * object no longer has to shadow the backing
+ * object; it can shadow the next one in the
+ * chain.
+ *
+ * If the backing object has existence info,
+ * we must check examine its existence info
+ * as well.
+ *
+ */
- /*
- * If all of the pages in the backing object are
- * shadowed by the parent object, the parent
- * object no longer has to shadow the backing
- * object; it can shadow the next one in the
- * chain.
- *
- * If the backing object has existence info,
- * we must check examine its existence info
- * as well.
- *
- */
+ backing_offset = object->shadow_offset;
+ backing_rcount = backing_object->resident_page_count;
- if(object->cow_hint >= size)
- object->cow_hint = 0;
- current_offset = object->cow_hint;
- while(TRUE) {
- if (vm_page_lookup(object,
- (vm_object_offset_t)current_offset)
- != VM_PAGE_NULL) {
- current_offset+=PAGE_SIZE;
- } else if ((object->pager_created) &&
- (object->existence_map != NULL) &&
- (vm_external_state_get(object->existence_map,
- current_offset)
- != VM_EXTERNAL_STATE_ABSENT)) {
- current_offset+=PAGE_SIZE;
- } else if (vm_page_lookup(backing_object,
- (vm_object_offset_t)current_offset
- + backing_offset)!= VM_PAGE_NULL) {
- /* found a dependency */
- object->cow_hint = current_offset;
- vm_object_unlock(backing_object);
- return;
- } else if ((backing_object->pager_created) &&
- (backing_object->existence_map != NULL) &&
- (vm_external_state_get(
- backing_object->existence_map,
- current_offset + backing_offset)
- != VM_EXTERNAL_STATE_ABSENT)) {
- /* found a dependency */
- object->cow_hint = current_offset;
- vm_object_unlock(backing_object);
- return;
- } else {
- current_offset+=PAGE_SIZE;
+#if MACH_PAGEMAP
+#define EXISTS_IN_OBJECT(obj, off, rc) \
+ (vm_external_state_get((obj)->existence_map, \
+ (vm_offset_t)(off)) == VM_EXTERNAL_STATE_EXISTS || \
+ ((rc) && ++lookups && vm_page_lookup((obj), (off)) != VM_PAGE_NULL && (rc)--))
+#else
+#define EXISTS_IN_OBJECT(obj, off, rc) \
+ (((rc) && ++lookups && vm_page_lookup((obj), (off)) != VM_PAGE_NULL && (rc)--))
+#endif /* MACH_PAGEMAP */
+
+ /*
+ * Check the hint location first
+ * (since it is often the quickest way out of here).
+ */
+ if (object->cow_hint != ~(vm_offset_t)0)
+ hint_offset = (vm_object_offset_t)object->cow_hint;
+ else
+ hint_offset = (hint_offset > 8 * PAGE_SIZE_64) ?
+ (hint_offset - 8 * PAGE_SIZE_64) : 0;
+
+ if (EXISTS_IN_OBJECT(backing_object, hint_offset +
+ backing_offset, backing_rcount) &&
+ !EXISTS_IN_OBJECT(object, hint_offset, rcount)) {
+ /* dependency right at the hint */
+ object->cow_hint = (vm_offset_t) hint_offset; /* atomic */
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
}
- if(current_offset >= size) {
- /* wrap at end of object */
- current_offset = 0;
+
+ /*
+ * If the object's window onto the backing_object
+ * is large compared to the number of resident
+ * pages in the backing object, it makes sense to
+ * walk the backing_object's resident pages first.
+ *
+ * NOTE: Pages may be in both the existence map and
+ * resident. So, we can't permanently decrement
+ * the rcount here because the second loop may
+ * find the same pages in the backing object'
+ * existence map that we found here and we would
+ * double-decrement the rcount. We also may or
+ * may not have found the
+ */
+ if (backing_rcount &&
+#if MACH_PAGEMAP
+ size > ((backing_object->existence_map) ?
+ backing_rcount : (backing_rcount >> 1))
+#else
+ size > (backing_rcount >> 1)
+#endif /* MACH_PAGEMAP */
+ ) {
+ unsigned int rc = rcount;
+ vm_page_t p;
+
+ backing_rcount = backing_object->resident_page_count;
+ p = (vm_page_t)queue_first(&backing_object->memq);
+ do {
+ /* Until we get more than one lookup lock */
+ if (lookups > 256) {
+ vm_object_collapse_delays++;
+ lookups = 0;
+ mutex_pause(0);
+ }
+
+ offset = (p->offset - backing_offset);
+ if (offset < object->size &&
+ offset != hint_offset &&
+ !EXISTS_IN_OBJECT(object, offset, rc)) {
+ /* found a dependency */
+ object->cow_hint = (vm_offset_t) offset; /* atomic */
+
+ break;
+ }
+ p = (vm_page_t) queue_next(&p->listq);
+
+ } while (--backing_rcount);
+ if (backing_rcount != 0 ) {
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
+ }
}
- if(current_offset == object->cow_hint) {
- /* we are free of shadow influence */
- break;
+
+ /*
+ * Walk through the offsets looking for pages in the
+ * backing object that show through to the object.
+ */
+ if (backing_rcount
+#if MACH_PAGEMAP
+ || backing_object->existence_map
+#endif /* MACH_PAGEMAP */
+ ) {
+ offset = hint_offset;
+
+ while((offset =
+ (offset + PAGE_SIZE_64 < object->size) ?
+ (offset + PAGE_SIZE_64) : 0) != hint_offset) {
+
+ /* Until we get more than one lookup lock */
+ if (lookups > 256) {
+ vm_object_collapse_delays++;
+ lookups = 0;
+ mutex_pause(0);
+ }
+
+ if (EXISTS_IN_OBJECT(backing_object, offset +
+ backing_offset, backing_rcount) &&
+ !EXISTS_IN_OBJECT(object, offset, rcount)) {
+ /* found a dependency */
+ object->cow_hint = (vm_offset_t) offset; /* atomic */
+ break;
+ }
+ }
+ if (offset != hint_offset) {
+ /* try and collapse the rest of the shadow chain */
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
+ object = backing_object;
+ object_lock_type = backing_object_lock_type;
+ continue;
+ }
}
}
- /* reset the cow_hint for any objects deeper in the chain */
- object->cow_hint = 0;
+ /*
+ * We need "exclusive" locks on the 2 VM objects.
+ */
+ if (backing_object_lock_type != OBJECT_LOCK_EXCLUSIVE) {
+ vm_object_unlock(backing_object);
+ if (object != original_object)
+ vm_object_unlock(object);
+ object_lock_type = OBJECT_LOCK_EXCLUSIVE;
+ backing_object_lock_type = OBJECT_LOCK_EXCLUSIVE;
+ goto retry;
+ }
+ /* reset the offset hint for any objects deeper in the chain */
+ object->cow_hint = (vm_offset_t)0;
/*
* All interesting pages in the backing object
@@ -3441,6 +4980,7 @@ vm_object_collapse(
*/
vm_object_do_bypass(object, backing_object);
+ vm_object_collapse_do_bypass++;
/*
* Try again with this object's new backing object.
@@ -3448,6 +4988,10 @@ vm_object_collapse(
continue;
}
+
+ if (object != original_object) {
+ vm_object_unlock(object);
+ }
}
/*
@@ -3485,9 +5029,8 @@ vm_object_page_remove(
p = vm_page_lookup(object, start);
if (p != VM_PAGE_NULL) {
assert(!p->cleaning && !p->pageout);
- if (!p->fictitious)
- pmap_page_protect(p->phys_page,
- VM_PROT_NONE);
+ if (!p->fictitious && p->pmapped)
+ pmap_disconnect(p->phys_page);
VM_PAGE_FREE(p);
}
}
@@ -3499,9 +5042,8 @@ vm_object_page_remove(
next = (vm_page_t) queue_next(&p->listq);
if ((start <= p->offset) && (p->offset < end)) {
assert(!p->cleaning && !p->pageout);
- if (!p->fictitious)
- pmap_page_protect(p->phys_page,
- VM_PROT_NONE);
+ if (!p->fictitious && p->pmapped)
+ pmap_disconnect(p->phys_page);
VM_PAGE_FREE(p);
}
p = next;
@@ -3540,7 +5082,7 @@ vm_object_coalesce(
register vm_object_t prev_object,
vm_object_t next_object,
vm_object_offset_t prev_offset,
- vm_object_offset_t next_offset,
+ __unused vm_object_offset_t next_offset,
vm_object_size_t prev_size,
vm_object_size_t next_size)
{
@@ -3560,14 +5102,14 @@ vm_object_coalesce(
XPR(XPR_VM_OBJECT,
"vm_object_coalesce: 0x%X prev_off 0x%X prev_size 0x%X next_size 0x%X\n",
- (integer_t)prev_object, prev_offset, prev_size, next_size, 0);
+ prev_object, prev_offset, prev_size, next_size, 0);
vm_object_lock(prev_object);
/*
* Try to collapse the object first
*/
- vm_object_collapse(prev_object);
+ vm_object_collapse(prev_object, prev_offset, TRUE);
/*
* Can't coalesce if pages not mapped to
@@ -3576,6 +5118,7 @@ vm_object_coalesce(
* . paged out
* . shadows another object
* . has a copy elsewhere
+ * . is purgeable
* . paging references (pages might be in page-list)
*/
@@ -3584,7 +5127,9 @@ vm_object_coalesce(
(prev_object->shadow != VM_OBJECT_NULL) ||
(prev_object->copy != VM_OBJECT_NULL) ||
(prev_object->true_share != FALSE) ||
- (prev_object->paging_in_progress != 0)) {
+ (prev_object->purgable != VM_PURGABLE_DENY) ||
+ (prev_object->paging_in_progress != 0) ||
+ (prev_object->activity_in_progress != 0)) {
vm_object_unlock(prev_object);
return(FALSE);
}
@@ -3639,7 +5184,7 @@ vm_object_page_map(
vm_object_offset_t offset),
void *map_fn_data) /* private to map_fn */
{
- int num_pages;
+ int64_t num_pages;
int i;
vm_page_t m;
vm_page_t old_page;
@@ -3658,14 +5203,15 @@ vm_object_page_map(
if ((old_page = vm_page_lookup(object, offset))
!= VM_PAGE_NULL)
{
- vm_page_lock_queues();
- vm_page_free(old_page);
- vm_page_unlock_queues();
+ VM_PAGE_FREE(old_page);
}
- vm_page_init(m, addr);
- /* private normally requires lock_queues but since we */
- /* are initializing the page, its not necessary here */
+ assert((ppnum_t) addr == addr);
+ vm_page_init(m, (ppnum_t) addr, FALSE);
+ /*
+ * private normally requires lock_queues but since we
+ * are initializing the page, its not necessary here
+ */
m->private = TRUE; /* don`t free page */
m->wire_count = 1;
vm_page_insert(m, object, offset);
@@ -3708,8 +5254,9 @@ print_bitstring(
boolean_t
vm_object_cached(
- register vm_object_t object)
+ __unused register vm_object_t object)
{
+#if VM_OBJECT_CACHE
register vm_object_t o;
queue_iterate(&vm_object_cached_list, o, vm_object_t, cached_list) {
@@ -3717,6 +5264,7 @@ vm_object_cached(
return TRUE;
}
}
+#endif
return FALSE;
}
@@ -3726,23 +5274,23 @@ vm_object_cached(
*/
void
vm_external_print(
- vm_external_map_t map,
- vm_size_t size)
+ vm_external_map_t emap,
+ vm_object_size_t size)
{
- if (map == VM_EXTERNAL_NULL) {
+ if (emap == VM_EXTERNAL_NULL) {
printf("0 ");
} else {
- vm_size_t existence_size = stob(size);
- printf("{ size=%d, map=[", existence_size);
+ vm_object_size_t existence_size = stob(size);
+ printf("{ size=%lld, map=[", (uint64_t) existence_size);
if (existence_size > 0) {
- print_bitstring(map[0]);
+ print_bitstring(emap[0]);
}
if (existence_size > 1) {
- print_bitstring(map[1]);
+ print_bitstring(emap[1]);
}
if (existence_size > 2) {
printf("...");
- print_bitstring(map[existence_size-1]);
+ print_bitstring(emap[existence_size-1]);
}
printf("] }\n");
}
@@ -3754,8 +5302,6 @@ int
vm_follow_object(
vm_object_t object)
{
- extern db_indent;
-
int count = 0;
int orig_db_indent = db_indent;
@@ -3783,18 +5329,16 @@ vm_follow_object(
* vm_object_print: [ debug ]
*/
void
-vm_object_print(
- vm_object_t object,
- boolean_t have_addr,
- int arg_count,
- char *modif)
+vm_object_print(db_expr_t db_addr, __unused boolean_t have_addr,
+ __unused db_expr_t arg_count, __unused char *modif)
{
+ vm_object_t object;
register vm_page_t p;
- extern db_indent;
- char *s;
+ const char *s;
register int count;
+ object = (vm_object_t) (long) db_addr;
if (object == VM_OBJECT_NULL)
return;
@@ -3803,8 +5347,7 @@ vm_object_print(
db_indent += 2;
iprintf("size=0x%x", object->size);
- printf(", cluster=0x%x", object->cluster_size);
- printf(", frozen=0x%x", object->frozen_size);
+ printf(", memq_hint=%p", object->memq_hint);
printf(", ref_count=%d\n", object->ref_count);
iprintf("");
#if TASK_SWAPPER
@@ -3816,7 +5359,7 @@ vm_object_print(
if (object->shadow) {
register int i = 0;
vm_object_t shadow = object;
- while(shadow = shadow->shadow)
+ while((shadow = shadow->shadow))
i++;
printf(" (depth %d)", i);
}
@@ -3826,7 +5369,7 @@ vm_object_print(
iprintf("pager=0x%x", object->pager);
printf(", paging_offset=0x%x", object->paging_offset);
- printf(", pager_request=0x%x\n", object->pager_request);
+ printf(", pager_control=0x%x\n", object->pager_control);
iprintf("copy_strategy=%d[", object->copy_strategy);
switch (object->copy_strategy) {
@@ -3854,7 +5397,6 @@ vm_object_print(
printf("?");
}
printf("]");
- printf(", absent_count=%d\n", object->absent_count);
iprintf("all_wanted=0x%x<", object->all_wanted);
s = "";
@@ -3870,10 +5412,6 @@ vm_object_print(
printf("%spaging", s);
s = ",";
}
- if (vm_object_wanted(object, VM_OBJECT_EVENT_ABSENT_COUNT)) {
- printf("%sabsent", s);
- s = ",";
- }
if (vm_object_wanted(object, VM_OBJECT_EVENT_LOCK_IN_PROGRESS)) {
printf("%slock", s);
s = ",";
@@ -3892,6 +5430,7 @@ vm_object_print(
}
printf(">");
printf(", paging_in_progress=%d\n", object->paging_in_progress);
+ printf(", activity_in_progress=%d\n", object->activity_in_progress);
iprintf("%screated, %sinit, %sready, %spersist, %strusted, %spageout, %s, %s\n",
(object->pager_created ? "" : "!"),
@@ -3902,10 +5441,11 @@ vm_object_print(
(object->pageout ? "" : "!"),
(object->internal ? "internal" : "external"),
(object->temporary ? "temporary" : "permanent"));
- iprintf("%salive, %slock_in_progress, %slock_restart, %sshadowed, %scached, %sprivate\n",
+ iprintf("%salive, %spurgeable, %spurgeable_volatile, %spurgeable_empty, %sshadowed, %scached, %sprivate\n",
(object->alive ? "" : "!"),
- (object->lock_in_progress ? "" : "!"),
- (object->lock_restart ? "" : "!"),
+ ((object->purgable != VM_PURGABLE_DENY) ? "" : "!"),
+ ((object->purgable == VM_PURGABLE_VOLATILE) ? "" : "!"),
+ ((object->purgable == VM_PURGABLE_EMPTY) ? "" : "!"),
(object->shadowed ? "" : "!"),
(vm_object_cached(object) ? "" : "!"),
(object->private ? "" : "!"));
@@ -3936,7 +5476,7 @@ vm_object_print(
}
count++;
- printf("(off=0x%X,page=0x%X)", p->offset, (integer_t) p);
+ printf("(off=0x%llX,page=%p)", p->offset, p);
p = (vm_page_t) queue_next(&p->listq);
}
if (count != 0) {
@@ -3963,10 +5503,9 @@ vm_object_find(
task_t task;
vm_map_t map;
vm_map_entry_t entry;
- processor_set_t pset = &default_pset;
boolean_t found = FALSE;
- queue_iterate(&pset->tasks, task, task_t, pset_tasks) {
+ queue_iterate(&tasks, task, task_t, tasks) {
map = task->map;
for (entry = vm_map_first_entry(map);
entry && entry != vm_map_to_entry(map);
@@ -4036,38 +5575,58 @@ vm_object_populate_with_private(
m = vm_page_lookup(object, base_offset);
if(m != VM_PAGE_NULL) {
if(m->fictitious) {
- vm_page_lock_queues();
- m->fictitious = FALSE;
- m->private = TRUE;
- m->phys_page = base_page;
- if(!m->busy) {
- m->busy = TRUE;
- }
- if(!m->absent) {
- m->absent = TRUE;
- object->absent_count++;
+ if (m->phys_page != vm_page_guard_addr) {
+
+ vm_page_lockspin_queues();
+ m->private = TRUE;
+ vm_page_unlock_queues();
+
+ m->fictitious = FALSE;
+ m->phys_page = base_page;
+ if(!m->busy) {
+ m->busy = TRUE;
+ }
+ if(!m->absent) {
+ m->absent = TRUE;
+ }
+ m->list_req_pending = TRUE;
}
- m->list_req_pending = TRUE;
- vm_page_unlock_queues();
} else if (m->phys_page != base_page) {
- /* pmap call to clear old mapping */
- pmap_page_protect(m->phys_page,
- VM_PROT_NONE);
+ if (m->pmapped) {
+ /*
+ * pmap call to clear old mapping
+ */
+ pmap_disconnect(m->phys_page);
+ }
m->phys_page = base_page;
}
+
+ /*
+ * ENCRYPTED SWAP:
+ * We're not pointing to the same
+ * physical page any longer and the
+ * contents of the new one are not
+ * supposed to be encrypted.
+ * XXX What happens to the original
+ * physical page. Is it lost ?
+ */
+ m->encrypted = FALSE;
+
} else {
- while ((m = vm_page_grab_fictitious())
- == VM_PAGE_NULL)
+ while ((m = vm_page_grab_fictitious()) == VM_PAGE_NULL)
vm_page_more_fictitious();
- vm_page_lock_queues();
- m->fictitious = FALSE;
+
+ /*
+ * private normally requires lock_queues but since we
+ * are initializing the page, its not necessary here
+ */
m->private = TRUE;
+ m->fictitious = FALSE;
m->phys_page = base_page;
m->list_req_pending = TRUE;
m->absent = TRUE;
m->unusual = TRUE;
- object->absent_count++;
- vm_page_unlock_queues();
+
vm_page_insert(m, object, base_offset);
}
base_page++; /* Go to the next physical page */
@@ -4083,7 +5642,7 @@ vm_object_populate_with_private(
/* shadows on contiguous memory are not allowed */
/* we therefore can use the offset field */
- object->shadow_offset = (vm_object_offset_t)(phys_page << 12);
+ object->shadow_offset = (vm_object_offset_t)phys_page << PAGE_SHIFT;
object->size = size;
}
vm_object_unlock(object);
@@ -4094,7 +5653,7 @@ vm_object_populate_with_private(
* memory_object_free_from_cache:
*
* Walk the vm_object cache list, removing and freeing vm_objects
- * which are backed by the pager identified by the caller, (pager_id).
+ * which are backed by the pager identified by the caller, (pager_ops).
* Remove up to "count" objects, if there are that may available
* in the cache.
*
@@ -4104,13 +5663,12 @@ vm_object_populate_with_private(
__private_extern__ kern_return_t
memory_object_free_from_cache(
- host_t host,
- int *pager_id,
+ __unused host_t host,
+ __unused memory_object_pager_ops_t pager_ops,
int *count)
{
-
+#if VM_OBJECT_CACHE
int object_released = 0;
- int i;
register vm_object_t object = VM_OBJECT_NULL;
vm_object_t shadow;
@@ -4125,12 +5683,14 @@ memory_object_free_from_cache(
queue_iterate(&vm_object_cached_list, object,
vm_object_t, cached_list) {
- if (object->pager && (pager_id == object->pager->pager)) {
+ if (object->pager &&
+ (pager_ops == object->pager->mo_pager_ops)) {
vm_object_lock(object);
queue_remove(&vm_object_cached_list, object,
vm_object_t, cached_list);
vm_object_cached_count--;
+ vm_object_cache_unlock();
/*
* Since this object is in the cache, we know
* that it is initialized and has only a pager's
@@ -4140,6 +5700,7 @@ memory_object_free_from_cache(
assert(object->pager_initialized);
assert(object->ref_count == 0);
+ vm_object_lock_assert_exclusive(object);
object->ref_count++;
/*
@@ -4152,6 +5713,7 @@ memory_object_free_from_cache(
* (We are careful here to limit recursion.)
*/
shadow = object->pageout?VM_OBJECT_NULL:object->shadow;
+
if ((vm_object_terminate(object) == KERN_SUCCESS)
&& (shadow != VM_OBJECT_NULL)) {
vm_object_deallocate(shadow);
@@ -4164,6 +5726,9 @@ memory_object_free_from_cache(
}
vm_object_cache_unlock();
*count = object_released;
+#else
+ *count = 0;
+#endif
return KERN_SUCCESS;
}
@@ -4177,21 +5742,22 @@ memory_object_create_named(
{
vm_object_t object;
vm_object_hash_entry_t entry;
+ lck_mtx_t *lck;
*control = MEMORY_OBJECT_CONTROL_NULL;
if (pager == MEMORY_OBJECT_NULL)
return KERN_INVALID_ARGUMENT;
- vm_object_cache_lock();
+ lck = vm_object_hash_lock_spin(pager);
entry = vm_object_hash_lookup(pager, FALSE);
+
if ((entry != VM_OBJECT_HASH_ENTRY_NULL) &&
(entry->object != VM_OBJECT_NULL)) {
if (entry->object->named == TRUE)
panic("memory_object_create_named: caller already holds the right"); }
+ vm_object_hash_unlock(lck);
- vm_object_cache_unlock();
- if ((object = vm_object_enter(pager, size, FALSE, FALSE, TRUE))
- == VM_OBJECT_NULL) {
+ if ((object = vm_object_enter(pager, size, FALSE, FALSE, TRUE)) == VM_OBJECT_NULL) {
return(KERN_INVALID_OBJECT);
}
@@ -4204,7 +5770,7 @@ memory_object_create_named(
VM_OBJECT_EVENT_PAGER_READY,
THREAD_UNINT);
}
- *control = object->pager_request;
+ *control = object->pager_control;
vm_object_unlock(object);
}
return (KERN_SUCCESS);
@@ -4230,51 +5796,49 @@ memory_object_recover_named(
{
vm_object_t object;
- vm_object_cache_lock();
object = memory_object_control_to_vm_object(control);
if (object == VM_OBJECT_NULL) {
- vm_object_cache_unlock();
return (KERN_INVALID_ARGUMENT);
}
-
restart:
vm_object_lock(object);
if (object->terminating && wait_on_terminating) {
- vm_object_cache_unlock();
vm_object_wait(object,
VM_OBJECT_EVENT_PAGING_IN_PROGRESS,
THREAD_UNINT);
- vm_object_cache_lock();
goto restart;
}
if (!object->alive) {
- vm_object_cache_unlock();
vm_object_unlock(object);
return KERN_FAILURE;
}
if (object->named == TRUE) {
- vm_object_cache_unlock();
vm_object_unlock(object);
return KERN_SUCCESS;
}
-
- if((object->ref_count == 0) && (!object->terminating)){
+#if VM_OBJECT_CACHE
+ if ((object->ref_count == 0) && (!object->terminating)) {
+ if (!vm_object_cache_lock_try()) {
+ vm_object_unlock(object);
+ goto restart;
+ }
queue_remove(&vm_object_cached_list, object,
vm_object_t, cached_list);
- vm_object_cached_count--;
- XPR(XPR_VM_OBJECT_CACHE,
- "memory_object_recover_named: removing %X, head (%X, %X)\n",
- (integer_t)object,
- (integer_t)vm_object_cached_list.next,
- (integer_t)vm_object_cached_list.prev, 0,0);
+ vm_object_cached_count--;
+ XPR(XPR_VM_OBJECT_CACHE,
+ "memory_object_recover_named: removing %X, head (%X, %X)\n",
+ object,
+ vm_object_cached_list.next,
+ vm_object_cached_list.prev, 0,0);
+
+ vm_object_cache_unlock();
}
-
- vm_object_cache_unlock();
-
+#endif
object->named = TRUE;
+ vm_object_lock_assert_exclusive(object);
object->ref_count++;
vm_object_res_reference(object);
while (!object->pager_ready) {
@@ -4314,17 +5878,10 @@ vm_object_release_name(
while (object != VM_OBJECT_NULL) {
- /*
- * The cache holds a reference (uncounted) to
- * the object. We must locke it before removing
- * the object.
- *
- */
-
- vm_object_cache_lock();
vm_object_lock(object);
+
assert(object->alive);
- if(original_object)
+ if (original_object)
assert(object->named);
assert(object->ref_count > 0);
@@ -4339,7 +5896,6 @@ vm_object_release_name(
VM_OBJECT_EVENT_INITIALIZED,
THREAD_UNINT);
vm_object_unlock(object);
- vm_object_cache_unlock();
thread_block(THREAD_CONTINUE_NULL);
continue;
}
@@ -4348,22 +5904,19 @@ vm_object_release_name(
&& (flags & MEMORY_OBJECT_TERMINATE_IDLE))
|| (object->terminating)) {
vm_object_unlock(object);
- vm_object_cache_unlock();
return KERN_FAILURE;
} else {
if (flags & MEMORY_OBJECT_RELEASE_NO_OP) {
vm_object_unlock(object);
- vm_object_cache_unlock();
return KERN_SUCCESS;
}
}
if ((flags & MEMORY_OBJECT_RESPECT_CACHE) &&
(object->ref_count == 1)) {
- if(original_object)
+ if (original_object)
object->named = FALSE;
vm_object_unlock(object);
- vm_object_cache_unlock();
/* let vm_object_deallocate push this thing into */
/* the cache, if that it is where it is bound */
vm_object_deallocate(object);
@@ -4371,9 +5924,10 @@ vm_object_release_name(
}
VM_OBJ_RES_DECR(object);
shadow = object->pageout?VM_OBJECT_NULL:object->shadow;
- if(object->ref_count == 1) {
- if(vm_object_terminate(object) != KERN_SUCCESS) {
- if(original_object) {
+
+ if (object->ref_count == 1) {
+ if (vm_object_terminate(object) != KERN_SUCCESS) {
+ if (original_object) {
return KERN_FAILURE;
} else {
return KERN_SUCCESS;
@@ -4386,15 +5940,18 @@ vm_object_release_name(
}
return KERN_SUCCESS;
} else {
+ vm_object_lock_assert_exclusive(object);
object->ref_count--;
assert(object->ref_count > 0);
if(original_object)
object->named = FALSE;
vm_object_unlock(object);
- vm_object_cache_unlock();
return KERN_SUCCESS;
}
}
+ /*NOTREACHED*/
+ assert(0);
+ return KERN_FAILURE;
}
@@ -4407,12 +5964,13 @@ vm_object_lock_request(
int flags,
vm_prot_t prot)
{
- vm_object_offset_t original_offset = offset;
- boolean_t should_flush=flags & MEMORY_OBJECT_DATA_FLUSH;
+ __unused boolean_t should_flush;
+
+ should_flush = flags & MEMORY_OBJECT_DATA_FLUSH;
XPR(XPR_MEMORY_OBJECT,
"vm_o_lock_request, obj 0x%X off 0x%X size 0x%X flags %X prot %X\n",
- (integer_t)object, offset, size,
+ object, offset, size,
(((should_return&1)<<1)|should_flush), prot);
/*
@@ -4432,10 +5990,9 @@ vm_object_lock_request(
*/
vm_object_lock(object);
vm_object_paging_begin(object);
- offset -= object->paging_offset;
(void)vm_object_update(object,
- offset, size, should_return, flags, prot);
+ offset, size, NULL, NULL, should_return, flags, prot);
vm_object_paging_end(object);
vm_object_unlock(object);
@@ -4443,66 +6000,419 @@ vm_object_lock_request(
return (KERN_SUCCESS);
}
-
-
-#if TASK_SWAPPER
/*
- * vm_object_res_deallocate
- *
- * (recursively) decrement residence counts on vm objects and their shadows.
- * Called from vm_object_deallocate and when swapping out an object.
- *
- * The object is locked, and remains locked throughout the function,
- * even as we iterate down the shadow chain. Locks on intermediate objects
- * will be dropped, but not the original object.
+ * Empty a purgeable object by grabbing the physical pages assigned to it and
+ * putting them on the free queue without writing them to backing store, etc.
+ * When the pages are next touched they will be demand zero-fill pages. We
+ * skip pages which are busy, being paged in/out, wired, etc. We do _not_
+ * skip referenced/dirty pages, pages on the active queue, etc. We're more
+ * than happy to grab these since this is a purgeable object. We mark the
+ * object as "empty" after reaping its pages.
*
- * NOTE: this function used to use recursion, rather than iteration.
+ * On entry the object must be locked and it must be
+ * purgeable with no delayed copies pending.
*/
-
-__private_extern__ void
-vm_object_res_deallocate(
- vm_object_t object)
+void
+vm_object_purge(vm_object_t object)
{
- vm_object_t orig_object = object;
- /*
- * Object is locked so it can be called directly
- * from vm_object_deallocate. Original object is never
- * unlocked.
- */
- assert(object->res_count > 0);
- while (--object->res_count == 0) {
- assert(object->ref_count >= object->res_count);
- vm_object_deactivate_all_pages(object);
- /* iterate on shadow, if present */
- if (object->shadow != VM_OBJECT_NULL) {
- vm_object_t tmp_object = object->shadow;
- vm_object_lock(tmp_object);
- if (object != orig_object)
- vm_object_unlock(object);
- object = tmp_object;
- assert(object->res_count > 0);
- } else
- break;
+ vm_object_lock_assert_exclusive(object);
+
+ if (object->purgable == VM_PURGABLE_DENY)
+ return;
+
+ assert(object->copy == VM_OBJECT_NULL);
+ assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE);
+
+ if(object->purgable == VM_PURGABLE_VOLATILE) {
+ unsigned int delta;
+ assert(object->resident_page_count >=
+ object->wired_page_count);
+ delta = (object->resident_page_count -
+ object->wired_page_count);
+ if (delta != 0) {
+ assert(vm_page_purgeable_count >=
+ delta);
+ OSAddAtomic(-delta,
+ (SInt32 *)&vm_page_purgeable_count);
+ }
+ if (object->wired_page_count != 0) {
+ assert(vm_page_purgeable_wired_count >=
+ object->wired_page_count);
+ OSAddAtomic(-object->wired_page_count,
+ (SInt32 *)&vm_page_purgeable_wired_count);
+ }
}
- if (object != orig_object)
- vm_object_unlock(object);
+ object->purgable = VM_PURGABLE_EMPTY;
+
+ vm_object_reap_pages(object, REAP_PURGEABLE);
}
+
/*
- * vm_object_res_reference
+ * vm_object_purgeable_control() allows the caller to control and investigate the
+ * state of a purgeable object. A purgeable object is created via a call to
+ * vm_allocate() with VM_FLAGS_PURGABLE specified. A purgeable object will
+ * never be coalesced with any other object -- even other purgeable objects --
+ * and will thus always remain a distinct object. A purgeable object has
+ * special semantics when its reference count is exactly 1. If its reference
+ * count is greater than 1, then a purgeable object will behave like a normal
+ * object and attempts to use this interface will result in an error return
+ * of KERN_INVALID_ARGUMENT.
*
- * Internal function to increment residence count on a vm object
- * and its shadows. It is called only from vm_object_reference, and
- * when swapping in a vm object, via vm_map_swap.
+ * A purgeable object may be put into a "volatile" state which will make the
+ * object's pages elligable for being reclaimed without paging to backing
+ * store if the system runs low on memory. If the pages in a volatile
+ * purgeable object are reclaimed, the purgeable object is said to have been
+ * "emptied." When a purgeable object is emptied the system will reclaim as
+ * many pages from the object as it can in a convenient manner (pages already
+ * en route to backing store or busy for other reasons are left as is). When
+ * a purgeable object is made volatile, its pages will generally be reclaimed
+ * before other pages in the application's working set. This semantic is
+ * generally used by applications which can recreate the data in the object
+ * faster than it can be paged in. One such example might be media assets
+ * which can be reread from a much faster RAID volume.
*
- * The object is locked, and remains locked throughout the function,
- * even as we iterate down the shadow chain. Locks on intermediate objects
- * will be dropped, but not the original object.
+ * A purgeable object may be designated as "non-volatile" which means it will
+ * behave like all other objects in the system with pages being written to and
+ * read from backing store as needed to satisfy system memory needs. If the
+ * object was emptied before the object was made non-volatile, that fact will
+ * be returned as the old state of the purgeable object (see
+ * VM_PURGABLE_SET_STATE below). In this case, any pages of the object which
+ * were reclaimed as part of emptying the object will be refaulted in as
+ * zero-fill on demand. It is up to the application to note that an object
+ * was emptied and recreate the objects contents if necessary. When a
+ * purgeable object is made non-volatile, its pages will generally not be paged
+ * out to backing store in the immediate future. A purgeable object may also
+ * be manually emptied.
*
- * NOTE: this function used to use recursion, rather than iteration.
- */
-
-__private_extern__ void
+ * Finally, the current state (non-volatile, volatile, volatile & empty) of a
+ * volatile purgeable object may be queried at any time. This information may
+ * be used as a control input to let the application know when the system is
+ * experiencing memory pressure and is reclaiming memory.
+ *
+ * The specified address may be any address within the purgeable object. If
+ * the specified address does not represent any object in the target task's
+ * virtual address space, then KERN_INVALID_ADDRESS will be returned. If the
+ * object containing the specified address is not a purgeable object, then
+ * KERN_INVALID_ARGUMENT will be returned. Otherwise, KERN_SUCCESS will be
+ * returned.
+ *
+ * The control parameter may be any one of VM_PURGABLE_SET_STATE or
+ * VM_PURGABLE_GET_STATE. For VM_PURGABLE_SET_STATE, the in/out parameter
+ * state is used to set the new state of the purgeable object and return its
+ * old state. For VM_PURGABLE_GET_STATE, the current state of the purgeable
+ * object is returned in the parameter state.
+ *
+ * The in/out parameter state may be one of VM_PURGABLE_NONVOLATILE,
+ * VM_PURGABLE_VOLATILE or VM_PURGABLE_EMPTY. These, respectively, represent
+ * the non-volatile, volatile and volatile/empty states described above.
+ * Setting the state of a purgeable object to VM_PURGABLE_EMPTY will
+ * immediately reclaim as many pages in the object as can be conveniently
+ * collected (some may have already been written to backing store or be
+ * otherwise busy).
+ *
+ * The process of making a purgeable object non-volatile and determining its
+ * previous state is atomic. Thus, if a purgeable object is made
+ * VM_PURGABLE_NONVOLATILE and the old state is returned as
+ * VM_PURGABLE_VOLATILE, then the purgeable object's previous contents are
+ * completely intact and will remain so until the object is made volatile
+ * again. If the old state is returned as VM_PURGABLE_EMPTY then the object
+ * was reclaimed while it was in a volatile state and its previous contents
+ * have been lost.
+ */
+/*
+ * The object must be locked.
+ */
+kern_return_t
+vm_object_purgable_control(
+ vm_object_t object,
+ vm_purgable_t control,
+ int *state)
+{
+ int old_state;
+ int new_state;
+
+ if (object == VM_OBJECT_NULL) {
+ /*
+ * Object must already be present or it can't be purgeable.
+ */
+ return KERN_INVALID_ARGUMENT;
+ }
+
+ /*
+ * Get current state of the purgeable object.
+ */
+ old_state = object->purgable;
+ if (old_state == VM_PURGABLE_DENY)
+ return KERN_INVALID_ARGUMENT;
+
+ /* purgeable cant have delayed copies - now or in the future */
+ assert(object->copy == VM_OBJECT_NULL);
+ assert(object->copy_strategy == MEMORY_OBJECT_COPY_NONE);
+
+ /*
+ * Execute the desired operation.
+ */
+ if (control == VM_PURGABLE_GET_STATE) {
+ *state = old_state;
+ return KERN_SUCCESS;
+ }
+
+ if ((*state) & VM_PURGABLE_DEBUG_EMPTY) {
+ object->volatile_empty = TRUE;
+ }
+ if ((*state) & VM_PURGABLE_DEBUG_FAULT) {
+ object->volatile_fault = TRUE;
+ }
+
+ new_state = *state & VM_PURGABLE_STATE_MASK;
+ if (new_state == VM_PURGABLE_VOLATILE &&
+ object->volatile_empty) {
+ new_state = VM_PURGABLE_EMPTY;
+ }
+
+ switch (new_state) {
+ case VM_PURGABLE_DENY:
+ case VM_PURGABLE_NONVOLATILE:
+ object->purgable = new_state;
+
+ if (old_state == VM_PURGABLE_VOLATILE) {
+ unsigned int delta;
+
+ assert(object->resident_page_count >=
+ object->wired_page_count);
+ delta = (object->resident_page_count -
+ object->wired_page_count);
+
+ assert(vm_page_purgeable_count >= delta);
+
+ if (delta != 0) {
+ OSAddAtomic(-delta,
+ (SInt32 *)&vm_page_purgeable_count);
+ }
+ if (object->wired_page_count != 0) {
+ assert(vm_page_purgeable_wired_count >=
+ object->wired_page_count);
+ OSAddAtomic(-object->wired_page_count,
+ (SInt32 *)&vm_page_purgeable_wired_count);
+ }
+
+ vm_page_lock_queues();
+
+ assert(object->objq.next != NULL && object->objq.prev != NULL); /* object should be on a queue */
+ purgeable_q_t queue = vm_purgeable_object_remove(object);
+ assert(queue);
+
+ vm_purgeable_token_delete_first(queue);
+ assert(queue->debug_count_objects>=0);
+
+ vm_page_unlock_queues();
+ }
+ break;
+
+ case VM_PURGABLE_VOLATILE:
+ if (object->volatile_fault) {
+ vm_page_t p;
+ int refmod;
+
+ queue_iterate(&object->memq, p, vm_page_t, listq) {
+ if (p->busy ||
+ VM_PAGE_WIRED(p) ||
+ p->fictitious) {
+ continue;
+ }
+ refmod = pmap_disconnect(p->phys_page);
+ if ((refmod & VM_MEM_MODIFIED) &&
+ !p->dirty) {
+ p->dirty = TRUE;
+ }
+ }
+ }
+
+ if (old_state == VM_PURGABLE_EMPTY &&
+ object->resident_page_count == 0)
+ break;
+
+ purgeable_q_t queue;
+
+ /* find the correct queue */
+ if ((*state&VM_PURGABLE_ORDERING_MASK) == VM_PURGABLE_ORDERING_OBSOLETE)
+ queue = &purgeable_queues[PURGEABLE_Q_TYPE_OBSOLETE];
+ else {
+ if ((*state&VM_PURGABLE_BEHAVIOR_MASK) == VM_PURGABLE_BEHAVIOR_FIFO)
+ queue = &purgeable_queues[PURGEABLE_Q_TYPE_FIFO];
+ else
+ queue = &purgeable_queues[PURGEABLE_Q_TYPE_LIFO];
+ }
+
+ if (old_state == VM_PURGABLE_NONVOLATILE ||
+ old_state == VM_PURGABLE_EMPTY) {
+ unsigned int delta;
+
+ /* try to add token... this can fail */
+ vm_page_lock_queues();
+
+ kern_return_t result = vm_purgeable_token_add(queue);
+ if (result != KERN_SUCCESS) {
+ vm_page_unlock_queues();
+ return result;
+ }
+ vm_page_unlock_queues();
+
+ assert(object->resident_page_count >=
+ object->wired_page_count);
+ delta = (object->resident_page_count -
+ object->wired_page_count);
+
+ if (delta != 0) {
+ OSAddAtomic(delta,
+ &vm_page_purgeable_count);
+ }
+ if (object->wired_page_count != 0) {
+ OSAddAtomic(object->wired_page_count,
+ &vm_page_purgeable_wired_count);
+ }
+
+ object->purgable = new_state;
+
+ /* object should not be on a queue */
+ assert(object->objq.next == NULL && object->objq.prev == NULL);
+ }
+ else if (old_state == VM_PURGABLE_VOLATILE) {
+ /*
+ * if reassigning priorities / purgeable groups, we don't change the
+ * token queue. So moving priorities will not make pages stay around longer.
+ * Reasoning is that the algorithm gives most priority to the most important
+ * object. If a new token is added, the most important object' priority is boosted.
+ * This biases the system already for purgeable queues that move a lot.
+ * It doesn't seem more biasing is neccessary in this case, where no new object is added.
+ */
+ assert(object->objq.next != NULL && object->objq.prev != NULL); /* object should be on a queue */
+
+ purgeable_q_t old_queue=vm_purgeable_object_remove(object);
+ assert(old_queue);
+
+ if (old_queue != queue) {
+ kern_return_t result;
+
+ /* Changing queue. Have to move token. */
+ vm_page_lock_queues();
+ vm_purgeable_token_delete_first(old_queue);
+ result = vm_purgeable_token_add(queue);
+ vm_page_unlock_queues();
+
+ assert(result==KERN_SUCCESS); /* this should never fail since we just freed a token */
+ }
+ };
+ vm_purgeable_object_add(object, queue, (*state&VM_VOLATILE_GROUP_MASK)>>VM_VOLATILE_GROUP_SHIFT );
+
+ assert(queue->debug_count_objects>=0);
+
+ break;
+
+
+ case VM_PURGABLE_EMPTY:
+ if (object->volatile_fault) {
+ vm_page_t p;
+ int refmod;
+
+ queue_iterate(&object->memq, p, vm_page_t, listq) {
+ if (p->busy ||
+ VM_PAGE_WIRED(p) ||
+ p->fictitious) {
+ continue;
+ }
+ refmod = pmap_disconnect(p->phys_page);
+ if ((refmod & VM_MEM_MODIFIED) &&
+ !p->dirty) {
+ p->dirty = TRUE;
+ }
+ }
+ }
+
+ if (old_state != new_state) {
+ assert(old_state == VM_PURGABLE_NONVOLATILE ||
+ old_state == VM_PURGABLE_VOLATILE);
+ if (old_state == VM_PURGABLE_VOLATILE) {
+ purgeable_q_t old_queue;
+
+ /* object should be on a queue */
+ assert(object->objq.next != NULL &&
+ object->objq.prev != NULL);
+ old_queue = vm_purgeable_object_remove(object);
+ assert(old_queue);
+ vm_page_lock_queues();
+ vm_purgeable_token_delete_first(old_queue);
+ vm_page_unlock_queues();
+ }
+ (void) vm_object_purge(object);
+ }
+ break;
+
+ }
+ *state = old_state;
+
+ return KERN_SUCCESS;
+}
+
+#if TASK_SWAPPER
+/*
+ * vm_object_res_deallocate
+ *
+ * (recursively) decrement residence counts on vm objects and their shadows.
+ * Called from vm_object_deallocate and when swapping out an object.
+ *
+ * The object is locked, and remains locked throughout the function,
+ * even as we iterate down the shadow chain. Locks on intermediate objects
+ * will be dropped, but not the original object.
+ *
+ * NOTE: this function used to use recursion, rather than iteration.
+ */
+
+__private_extern__ void
+vm_object_res_deallocate(
+ vm_object_t object)
+{
+ vm_object_t orig_object = object;
+ /*
+ * Object is locked so it can be called directly
+ * from vm_object_deallocate. Original object is never
+ * unlocked.
+ */
+ assert(object->res_count > 0);
+ while (--object->res_count == 0) {
+ assert(object->ref_count >= object->res_count);
+ vm_object_deactivate_all_pages(object);
+ /* iterate on shadow, if present */
+ if (object->shadow != VM_OBJECT_NULL) {
+ vm_object_t tmp_object = object->shadow;
+ vm_object_lock(tmp_object);
+ if (object != orig_object)
+ vm_object_unlock(object);
+ object = tmp_object;
+ assert(object->res_count > 0);
+ } else
+ break;
+ }
+ if (object != orig_object)
+ vm_object_unlock(object);
+}
+
+/*
+ * vm_object_res_reference
+ *
+ * Internal function to increment residence count on a vm object
+ * and its shadows. It is called only from vm_object_reference, and
+ * when swapping in a vm object, via vm_map_swap.
+ *
+ * The object is locked, and remains locked throughout the function,
+ * even as we iterate down the shadow chain. Locks on intermediate objects
+ * will be dropped, but not the original object.
+ *
+ * NOTE: this function used to use recursion, rather than iteration.
+ */
+
+__private_extern__ void
vm_object_res_reference(
vm_object_t object)
{
@@ -4560,11 +6470,969 @@ vm_object_reference(
* This is also needed as number of vnodes can be dynamically scaled.
*/
kern_return_t
-adjust_vm_object_cache(vm_size_t oval, vm_size_t nval)
+adjust_vm_object_cache(
+ __unused vm_size_t oval,
+ __unused vm_size_t nval)
{
+#if VM_OBJECT_CACHE
vm_object_cached_max = nval;
vm_object_cache_trim(FALSE);
+#endif
return (KERN_SUCCESS);
}
#endif /* MACH_BSD */
+
+/*
+ * vm_object_transpose
+ *
+ * This routine takes two VM objects of the same size and exchanges
+ * their backing store.
+ * The objects should be "quiesced" via a UPL operation with UPL_SET_IO_WIRE
+ * and UPL_BLOCK_ACCESS if they are referenced anywhere.
+ *
+ * The VM objects must not be locked by caller.
+ */
+unsigned int vm_object_transpose_count = 0;
+kern_return_t
+vm_object_transpose(
+ vm_object_t object1,
+ vm_object_t object2,
+ vm_object_size_t transpose_size)
+{
+ vm_object_t tmp_object;
+ kern_return_t retval;
+ boolean_t object1_locked, object2_locked;
+ vm_page_t page;
+ vm_object_offset_t page_offset;
+ lck_mtx_t *hash_lck;
+ vm_object_hash_entry_t hash_entry;
+
+ tmp_object = VM_OBJECT_NULL;
+ object1_locked = FALSE; object2_locked = FALSE;
+
+ if (object1 == object2 ||
+ object1 == VM_OBJECT_NULL ||
+ object2 == VM_OBJECT_NULL) {
+ /*
+ * If the 2 VM objects are the same, there's
+ * no point in exchanging their backing store.
+ */
+ retval = KERN_INVALID_VALUE;
+ goto done;
+ }
+
+ /*
+ * Since we need to lock both objects at the same time,
+ * make sure we always lock them in the same order to
+ * avoid deadlocks.
+ */
+ if (object1 > object2) {
+ tmp_object = object1;
+ object1 = object2;
+ object2 = tmp_object;
+ }
+
+ /*
+ * Allocate a temporary VM object to hold object1's contents
+ * while we copy object2 to object1.
+ */
+ tmp_object = vm_object_allocate(transpose_size);
+ vm_object_lock(tmp_object);
+ tmp_object->can_persist = FALSE;
+
+
+ /*
+ * Grab control of the 1st VM object.
+ */
+ vm_object_lock(object1);
+ object1_locked = TRUE;
+ if (!object1->alive || object1->terminating ||
+ object1->copy || object1->shadow || object1->shadowed ||
+ object1->purgable != VM_PURGABLE_DENY) {
+ /*
+ * We don't deal with copy or shadow objects (yet).
+ */
+ retval = KERN_INVALID_VALUE;
+ goto done;
+ }
+ /*
+ * We're about to mess with the object's backing store and
+ * taking a "paging_in_progress" reference wouldn't be enough
+ * to prevent any paging activity on this object, so the caller should
+ * have "quiesced" the objects beforehand, via a UPL operation with
+ * UPL_SET_IO_WIRE (to make sure all the pages are there and wired)
+ * and UPL_BLOCK_ACCESS (to mark the pages "busy").
+ *
+ * Wait for any paging operation to complete (but only paging, not
+ * other kind of activities not linked to the pager). After we're
+ * statisfied that there's no more paging in progress, we keep the
+ * object locked, to guarantee that no one tries to access its pager.
+ */
+ vm_object_paging_only_wait(object1, THREAD_UNINT);
+
+ /*
+ * Same as above for the 2nd object...
+ */
+ vm_object_lock(object2);
+ object2_locked = TRUE;
+ if (! object2->alive || object2->terminating ||
+ object2->copy || object2->shadow || object2->shadowed ||
+ object2->purgable != VM_PURGABLE_DENY) {
+ retval = KERN_INVALID_VALUE;
+ goto done;
+ }
+ vm_object_paging_only_wait(object2, THREAD_UNINT);
+
+
+ if (object1->size != object2->size ||
+ object1->size != transpose_size) {
+ /*
+ * If the 2 objects don't have the same size, we can't
+ * exchange their backing stores or one would overflow.
+ * If their size doesn't match the caller's
+ * "transpose_size", we can't do it either because the
+ * transpose operation will affect the entire span of
+ * the objects.
+ */
+ retval = KERN_INVALID_VALUE;
+ goto done;
+ }
+
+
+ /*
+ * Transpose the lists of resident pages.
+ * This also updates the resident_page_count and the memq_hint.
+ */
+ if (object1->phys_contiguous || queue_empty(&object1->memq)) {
+ /*
+ * No pages in object1, just transfer pages
+ * from object2 to object1. No need to go through
+ * an intermediate object.
+ */
+ while (!queue_empty(&object2->memq)) {
+ page = (vm_page_t) queue_first(&object2->memq);
+ vm_page_rename(page, object1, page->offset, FALSE);
+ }
+ assert(queue_empty(&object2->memq));
+ } else if (object2->phys_contiguous || queue_empty(&object2->memq)) {
+ /*
+ * No pages in object2, just transfer pages
+ * from object1 to object2. No need to go through
+ * an intermediate object.
+ */
+ while (!queue_empty(&object1->memq)) {
+ page = (vm_page_t) queue_first(&object1->memq);
+ vm_page_rename(page, object2, page->offset, FALSE);
+ }
+ assert(queue_empty(&object1->memq));
+ } else {
+ /* transfer object1's pages to tmp_object */
+ while (!queue_empty(&object1->memq)) {
+ page = (vm_page_t) queue_first(&object1->memq);
+ page_offset = page->offset;
+ vm_page_remove(page, TRUE);
+ page->offset = page_offset;
+ queue_enter(&tmp_object->memq, page, vm_page_t, listq);
+ }
+ assert(queue_empty(&object1->memq));
+ /* transfer object2's pages to object1 */
+ while (!queue_empty(&object2->memq)) {
+ page = (vm_page_t) queue_first(&object2->memq);
+ vm_page_rename(page, object1, page->offset, FALSE);
+ }
+ assert(queue_empty(&object2->memq));
+ /* transfer tmp_object's pages to object1 */
+ while (!queue_empty(&tmp_object->memq)) {
+ page = (vm_page_t) queue_first(&tmp_object->memq);
+ queue_remove(&tmp_object->memq, page,
+ vm_page_t, listq);
+ vm_page_insert(page, object2, page->offset);
+ }
+ assert(queue_empty(&tmp_object->memq));
+ }
+
+#define __TRANSPOSE_FIELD(field) \
+MACRO_BEGIN \
+ tmp_object->field = object1->field; \
+ object1->field = object2->field; \
+ object2->field = tmp_object->field; \
+MACRO_END
+
+ /* "Lock" refers to the object not its contents */
+ /* "size" should be identical */
+ assert(object1->size == object2->size);
+ /* "memq_hint" was updated above when transposing pages */
+ /* "ref_count" refers to the object not its contents */
+#if TASK_SWAPPER
+ /* "res_count" refers to the object not its contents */
+#endif
+ /* "resident_page_count" was updated above when transposing pages */
+ /* "wired_page_count" was updated above when transposing pages */
+ /* "reusable_page_count" was updated above when transposing pages */
+ /* there should be no "copy" */
+ assert(!object1->copy);
+ assert(!object2->copy);
+ /* there should be no "shadow" */
+ assert(!object1->shadow);
+ assert(!object2->shadow);
+ __TRANSPOSE_FIELD(shadow_offset); /* used by phys_contiguous objects */
+ __TRANSPOSE_FIELD(pager);
+ __TRANSPOSE_FIELD(paging_offset);
+ __TRANSPOSE_FIELD(pager_control);
+ /* update the memory_objects' pointers back to the VM objects */
+ if (object1->pager_control != MEMORY_OBJECT_CONTROL_NULL) {
+ memory_object_control_collapse(object1->pager_control,
+ object1);
+ }
+ if (object2->pager_control != MEMORY_OBJECT_CONTROL_NULL) {
+ memory_object_control_collapse(object2->pager_control,
+ object2);
+ }
+ __TRANSPOSE_FIELD(copy_strategy);
+ /* "paging_in_progress" refers to the object not its contents */
+ assert(!object1->paging_in_progress);
+ assert(!object2->paging_in_progress);
+ assert(object1->activity_in_progress);
+ assert(object2->activity_in_progress);
+ /* "all_wanted" refers to the object not its contents */
+ __TRANSPOSE_FIELD(pager_created);
+ __TRANSPOSE_FIELD(pager_initialized);
+ __TRANSPOSE_FIELD(pager_ready);
+ __TRANSPOSE_FIELD(pager_trusted);
+ __TRANSPOSE_FIELD(can_persist);
+ __TRANSPOSE_FIELD(internal);
+ __TRANSPOSE_FIELD(temporary);
+ __TRANSPOSE_FIELD(private);
+ __TRANSPOSE_FIELD(pageout);
+ /* "alive" should be set */
+ assert(object1->alive);
+ assert(object2->alive);
+ /* "purgeable" should be non-purgeable */
+ assert(object1->purgable == VM_PURGABLE_DENY);
+ assert(object2->purgable == VM_PURGABLE_DENY);
+ /* "shadowed" refers to the the object not its contents */
+ __TRANSPOSE_FIELD(silent_overwrite);
+ __TRANSPOSE_FIELD(advisory_pageout);
+ __TRANSPOSE_FIELD(true_share);
+ /* "terminating" should not be set */
+ assert(!object1->terminating);
+ assert(!object2->terminating);
+ __TRANSPOSE_FIELD(named);
+ /* "shadow_severed" refers to the object not its contents */
+ __TRANSPOSE_FIELD(phys_contiguous);
+ __TRANSPOSE_FIELD(nophyscache);
+ /* "cached_list.next" points to transposed object */
+ object1->cached_list.next = (queue_entry_t) object2;
+ object2->cached_list.next = (queue_entry_t) object1;
+ /* "cached_list.prev" should be NULL */
+ assert(object1->cached_list.prev == NULL);
+ assert(object2->cached_list.prev == NULL);
+ /* "msr_q" is linked to the object not its contents */
+ assert(queue_empty(&object1->msr_q));
+ assert(queue_empty(&object2->msr_q));
+ __TRANSPOSE_FIELD(last_alloc);
+ __TRANSPOSE_FIELD(sequential);
+ __TRANSPOSE_FIELD(pages_created);
+ __TRANSPOSE_FIELD(pages_used);
+#if MACH_PAGEMAP
+ __TRANSPOSE_FIELD(existence_map);
+#endif
+ __TRANSPOSE_FIELD(cow_hint);
+#if MACH_ASSERT
+ __TRANSPOSE_FIELD(paging_object);
+#endif
+ __TRANSPOSE_FIELD(wimg_bits);
+ __TRANSPOSE_FIELD(code_signed);
+ if (object1->hashed) {
+ hash_lck = vm_object_hash_lock_spin(object2->pager);
+ hash_entry = vm_object_hash_lookup(object2->pager, FALSE);
+ assert(hash_entry != VM_OBJECT_HASH_ENTRY_NULL);
+ hash_entry->object = object2;
+ vm_object_hash_unlock(hash_lck);
+ }
+ if (object2->hashed) {
+ hash_lck = vm_object_hash_lock_spin(object1->pager);
+ hash_entry = vm_object_hash_lookup(object1->pager, FALSE);
+ assert(hash_entry != VM_OBJECT_HASH_ENTRY_NULL);
+ hash_entry->object = object1;
+ vm_object_hash_unlock(hash_lck);
+ }
+ __TRANSPOSE_FIELD(hashed);
+ object1->transposed = TRUE;
+ object2->transposed = TRUE;
+ __TRANSPOSE_FIELD(mapping_in_progress);
+ __TRANSPOSE_FIELD(volatile_empty);
+ __TRANSPOSE_FIELD(volatile_fault);
+ __TRANSPOSE_FIELD(all_reusable);
+ assert(object1->blocked_access);
+ assert(object2->blocked_access);
+ assert(object1->__object2_unused_bits == 0);
+ assert(object2->__object2_unused_bits == 0);
+#if UPL_DEBUG
+ /* "uplq" refers to the object not its contents (see upl_transpose()) */
+#endif
+ assert(object1->objq.next == NULL);
+ assert(object1->objq.prev == NULL);
+ assert(object2->objq.next == NULL);
+ assert(object2->objq.prev == NULL);
+
+#undef __TRANSPOSE_FIELD
+
+ retval = KERN_SUCCESS;
+
+done:
+ /*
+ * Cleanup.
+ */
+ if (tmp_object != VM_OBJECT_NULL) {
+ vm_object_unlock(tmp_object);
+ /*
+ * Re-initialize the temporary object to avoid
+ * deallocating a real pager.
+ */
+ _vm_object_allocate(transpose_size, tmp_object);
+ vm_object_deallocate(tmp_object);
+ tmp_object = VM_OBJECT_NULL;
+ }
+
+ if (object1_locked) {
+ vm_object_unlock(object1);
+ object1_locked = FALSE;
+ }
+ if (object2_locked) {
+ vm_object_unlock(object2);
+ object2_locked = FALSE;
+ }
+
+ vm_object_transpose_count++;
+
+ return retval;
+}
+
+
+/*
+ * vm_object_cluster_size
+ *
+ * Determine how big a cluster we should issue an I/O for...
+ *
+ * Inputs: *start == offset of page needed
+ * *length == maximum cluster pager can handle
+ * Outputs: *start == beginning offset of cluster
+ * *length == length of cluster to try
+ *
+ * The original *start will be encompassed by the cluster
+ *
+ */
+extern int speculative_reads_disabled;
+#if CONFIG_EMBEDDED
+unsigned int preheat_pages_max = MAX_UPL_TRANSFER;
+unsigned int preheat_pages_min = 8;
+unsigned int preheat_pages_mult = 4;
+#else
+unsigned int preheat_pages_max = MAX_UPL_TRANSFER;
+unsigned int preheat_pages_min = 8;
+unsigned int preheat_pages_mult = 4;
+#endif
+
+uint32_t pre_heat_scaling[MAX_UPL_TRANSFER + 1];
+uint32_t pre_heat_cluster[MAX_UPL_TRANSFER + 1];
+
+
+__private_extern__ void
+vm_object_cluster_size(vm_object_t object, vm_object_offset_t *start,
+ vm_size_t *length, vm_object_fault_info_t fault_info, uint32_t *io_streaming)
+{
+ vm_size_t pre_heat_size;
+ vm_size_t tail_size;
+ vm_size_t head_size;
+ vm_size_t max_length;
+ vm_size_t cluster_size;
+ vm_object_offset_t object_size;
+ vm_object_offset_t orig_start;
+ vm_object_offset_t target_start;
+ vm_object_offset_t offset;
+ vm_behavior_t behavior;
+ boolean_t look_behind = TRUE;
+ boolean_t look_ahead = TRUE;
+ uint32_t throttle_limit;
+ int sequential_run;
+ int sequential_behavior = VM_BEHAVIOR_SEQUENTIAL;
+ unsigned int max_ph_size;
+ unsigned int min_ph_size;
+ unsigned int ph_mult;
+
+ assert( !(*length & PAGE_MASK));
+ assert( !(*start & PAGE_MASK_64));
+
+ if ( (ph_mult = preheat_pages_mult) < 1 )
+ ph_mult = 1;
+ if ( (min_ph_size = preheat_pages_min) < 1 )
+ min_ph_size = 1;
+ if ( (max_ph_size = preheat_pages_max) > MAX_UPL_TRANSFER )
+ max_ph_size = MAX_UPL_TRANSFER;
+
+ if ( (max_length = *length) > (max_ph_size * PAGE_SIZE) )
+ max_length = (max_ph_size * PAGE_SIZE);
+
+ /*
+ * we'll always return a cluster size of at least
+ * 1 page, since the original fault must always
+ * be processed
+ */
+ *length = PAGE_SIZE;
+ *io_streaming = 0;
+
+ if (speculative_reads_disabled || fault_info == NULL || max_length == 0) {
+ /*
+ * no cluster... just fault the page in
+ */
+ return;
+ }
+ orig_start = *start;
+ target_start = orig_start;
+ cluster_size = round_page(fault_info->cluster_size);
+ behavior = fault_info->behavior;
+
+ vm_object_lock(object);
+
+ if (object->internal)
+ object_size = object->size;
+ else if (object->pager != MEMORY_OBJECT_NULL)
+ vnode_pager_get_object_size(object->pager, &object_size);
+ else
+ goto out; /* pager is gone for this object, nothing more to do */
+
+ object_size = round_page_64(object_size);
+
+ if (orig_start >= object_size) {
+ /*
+ * fault occurred beyond the EOF...
+ * we need to punt w/o changing the
+ * starting offset
+ */
+ goto out;
+ }
+ if (object->pages_used > object->pages_created) {
+ /*
+ * must have wrapped our 32 bit counters
+ * so reset
+ */
+ object->pages_used = object->pages_created = 0;
+ }
+ if ((sequential_run = object->sequential)) {
+ if (sequential_run < 0) {
+ sequential_behavior = VM_BEHAVIOR_RSEQNTL;
+ sequential_run = 0 - sequential_run;
+ } else {
+ sequential_behavior = VM_BEHAVIOR_SEQUENTIAL;
+ }
+
+ }
+ switch(behavior) {
+
+ default:
+ behavior = VM_BEHAVIOR_DEFAULT;
+
+ case VM_BEHAVIOR_DEFAULT:
+ if (object->internal && fault_info->user_tag == VM_MEMORY_STACK)
+ goto out;
+
+ if (sequential_run >= (3 * PAGE_SIZE)) {
+ pre_heat_size = sequential_run + PAGE_SIZE;
+
+ if (sequential_behavior == VM_BEHAVIOR_SEQUENTIAL)
+ look_behind = FALSE;
+ else
+ look_ahead = FALSE;
+
+ *io_streaming = 1;
+ } else {
+
+ if (object->pages_created < 32 * ph_mult) {
+ /*
+ * prime the pump
+ */
+ pre_heat_size = PAGE_SIZE * 8 * ph_mult;
+ break;
+ }
+ /*
+ * Linear growth in PH size: The maximum size is max_length...
+ * this cacluation will result in a size that is neither a
+ * power of 2 nor a multiple of PAGE_SIZE... so round
+ * it up to the nearest PAGE_SIZE boundary
+ */
+ pre_heat_size = (ph_mult * (max_length * object->pages_used) / object->pages_created);
+
+ if (pre_heat_size < PAGE_SIZE * min_ph_size)
+ pre_heat_size = PAGE_SIZE * min_ph_size;
+ else
+ pre_heat_size = round_page(pre_heat_size);
+ }
+ break;
+
+ case VM_BEHAVIOR_RANDOM:
+ if ((pre_heat_size = cluster_size) <= PAGE_SIZE)
+ goto out;
+ break;
+
+ case VM_BEHAVIOR_SEQUENTIAL:
+ if ((pre_heat_size = cluster_size) == 0)
+ pre_heat_size = sequential_run + PAGE_SIZE;
+ look_behind = FALSE;
+ *io_streaming = 1;
+
+ break;
+
+ case VM_BEHAVIOR_RSEQNTL:
+ if ((pre_heat_size = cluster_size) == 0)
+ pre_heat_size = sequential_run + PAGE_SIZE;
+ look_ahead = FALSE;
+ *io_streaming = 1;
+
+ break;
+
+ }
+ throttle_limit = (uint32_t) max_length;
+ assert(throttle_limit == max_length);
+
+ if (vnode_pager_check_hard_throttle(object->pager, &throttle_limit, *io_streaming) == KERN_SUCCESS) {
+ if (max_length > throttle_limit)
+ max_length = throttle_limit;
+ }
+ if (pre_heat_size > max_length)
+ pre_heat_size = max_length;
+
+ if (behavior == VM_BEHAVIOR_DEFAULT) {
+ if (vm_page_free_count < vm_page_throttle_limit)
+ pre_heat_size = trunc_page(pre_heat_size / 8);
+ else if (vm_page_free_count < vm_page_free_target)
+ pre_heat_size = trunc_page(pre_heat_size / 2);
+
+ if (pre_heat_size <= PAGE_SIZE)
+ goto out;
+ }
+ if (look_ahead == TRUE) {
+ if (look_behind == TRUE) {
+ /*
+ * if we get here its due to a random access...
+ * so we want to center the original fault address
+ * within the cluster we will issue... make sure
+ * to calculate 'head_size' as a multiple of PAGE_SIZE...
+ * 'pre_heat_size' is a multiple of PAGE_SIZE but not
+ * necessarily an even number of pages so we need to truncate
+ * the result to a PAGE_SIZE boundary
+ */
+ head_size = trunc_page(pre_heat_size / 2);
+
+ if (target_start > head_size)
+ target_start -= head_size;
+ else
+ target_start = 0;
+
+ /*
+ * 'target_start' at this point represents the beginning offset
+ * of the cluster we are considering... 'orig_start' will be in
+ * the center of this cluster if we didn't have to clip the start
+ * due to running into the start of the file
+ */
+ }
+ if ((target_start + pre_heat_size) > object_size)
+ pre_heat_size = (vm_size_t)(round_page_64(object_size - target_start));
+ /*
+ * at this point caclulate the number of pages beyond the original fault
+ * address that we want to consider... this is guaranteed not to extend beyond
+ * the current EOF...
+ */
+ assert((vm_size_t)(orig_start - target_start) == (orig_start - target_start));
+ tail_size = pre_heat_size - (vm_size_t)(orig_start - target_start) - PAGE_SIZE;
+ } else {
+ if (pre_heat_size > target_start)
+ pre_heat_size = (vm_size_t) target_start; /* XXX: 32-bit vs 64-bit ? Joe ? */
+ tail_size = 0;
+ }
+ assert( !(target_start & PAGE_MASK_64));
+ assert( !(pre_heat_size & PAGE_MASK));
+
+ pre_heat_scaling[pre_heat_size / PAGE_SIZE]++;
+
+ if (pre_heat_size <= PAGE_SIZE)
+ goto out;
+
+ if (look_behind == TRUE) {
+ /*
+ * take a look at the pages before the original
+ * faulting offset... recalculate this in case
+ * we had to clip 'pre_heat_size' above to keep
+ * from running past the EOF.
+ */
+ head_size = pre_heat_size - tail_size - PAGE_SIZE;
+
+ for (offset = orig_start - PAGE_SIZE_64; head_size; offset -= PAGE_SIZE_64, head_size -= PAGE_SIZE) {
+ /*
+ * don't poke below the lowest offset
+ */
+ if (offset < fault_info->lo_offset)
+ break;
+ /*
+ * for external objects and internal objects w/o an existence map
+ * vm_externl_state_get will return VM_EXTERNAL_STATE_UNKNOWN
+ */
+#if MACH_PAGEMAP
+ if (vm_external_state_get(object->existence_map, offset) == VM_EXTERNAL_STATE_ABSENT) {
+ /*
+ * we know for a fact that the pager can't provide the page
+ * so don't include it or any pages beyond it in this cluster
+ */
+ break;
+ }
+#endif
+ if (vm_page_lookup(object, offset) != VM_PAGE_NULL) {
+ /*
+ * don't bridge resident pages
+ */
+ break;
+ }
+ *start = offset;
+ *length += PAGE_SIZE;
+ }
+ }
+ if (look_ahead == TRUE) {
+ for (offset = orig_start + PAGE_SIZE_64; tail_size; offset += PAGE_SIZE_64, tail_size -= PAGE_SIZE) {
+ /*
+ * don't poke above the highest offset
+ */
+ if (offset >= fault_info->hi_offset)
+ break;
+ assert(offset < object_size);
+
+ /*
+ * for external objects and internal objects w/o an existence map
+ * vm_externl_state_get will return VM_EXTERNAL_STATE_UNKNOWN
+ */
+#if MACH_PAGEMAP
+ if (vm_external_state_get(object->existence_map, offset) == VM_EXTERNAL_STATE_ABSENT) {
+ /*
+ * we know for a fact that the pager can't provide the page
+ * so don't include it or any pages beyond it in this cluster
+ */
+ break;
+ }
+#endif
+ if (vm_page_lookup(object, offset) != VM_PAGE_NULL) {
+ /*
+ * don't bridge resident pages
+ */
+ break;
+ }
+ *length += PAGE_SIZE;
+ }
+ }
+out:
+ if (*length > max_length)
+ *length = max_length;
+
+ pre_heat_cluster[*length / PAGE_SIZE]++;
+
+ vm_object_unlock(object);
+}
+
+
+/*
+ * Allow manipulation of individual page state. This is actually part of
+ * the UPL regimen but takes place on the VM object rather than on a UPL
+ */
+
+kern_return_t
+vm_object_page_op(
+ vm_object_t object,
+ vm_object_offset_t offset,
+ int ops,
+ ppnum_t *phys_entry,
+ int *flags)
+{
+ vm_page_t dst_page;
+
+ vm_object_lock(object);
+
+ if(ops & UPL_POP_PHYSICAL) {
+ if(object->phys_contiguous) {
+ if (phys_entry) {
+ *phys_entry = (ppnum_t)
+ (object->shadow_offset >> PAGE_SHIFT);
+ }
+ vm_object_unlock(object);
+ return KERN_SUCCESS;
+ } else {
+ vm_object_unlock(object);
+ return KERN_INVALID_OBJECT;
+ }
+ }
+ if(object->phys_contiguous) {
+ vm_object_unlock(object);
+ return KERN_INVALID_OBJECT;
+ }
+
+ while(TRUE) {
+ if((dst_page = vm_page_lookup(object,offset)) == VM_PAGE_NULL) {
+ vm_object_unlock(object);
+ return KERN_FAILURE;
+ }
+
+ /* Sync up on getting the busy bit */
+ if((dst_page->busy || dst_page->cleaning) &&
+ (((ops & UPL_POP_SET) &&
+ (ops & UPL_POP_BUSY)) || (ops & UPL_POP_DUMP))) {
+ /* someone else is playing with the page, we will */
+ /* have to wait */
+ PAGE_SLEEP(object, dst_page, THREAD_UNINT);
+ continue;
+ }
+
+ if (ops & UPL_POP_DUMP) {
+ if (dst_page->pmapped == TRUE)
+ pmap_disconnect(dst_page->phys_page);
+
+ VM_PAGE_FREE(dst_page);
+ break;
+ }
+
+ if (flags) {
+ *flags = 0;
+
+ /* Get the condition of flags before requested ops */
+ /* are undertaken */
+
+ if(dst_page->dirty) *flags |= UPL_POP_DIRTY;
+ if(dst_page->pageout) *flags |= UPL_POP_PAGEOUT;
+ if(dst_page->precious) *flags |= UPL_POP_PRECIOUS;
+ if(dst_page->absent) *flags |= UPL_POP_ABSENT;
+ if(dst_page->busy) *flags |= UPL_POP_BUSY;
+ }
+
+ /* The caller should have made a call either contingent with */
+ /* or prior to this call to set UPL_POP_BUSY */
+ if(ops & UPL_POP_SET) {
+ /* The protection granted with this assert will */
+ /* not be complete. If the caller violates the */
+ /* convention and attempts to change page state */
+ /* without first setting busy we may not see it */
+ /* because the page may already be busy. However */
+ /* if such violations occur we will assert sooner */
+ /* or later. */
+ assert(dst_page->busy || (ops & UPL_POP_BUSY));
+ if (ops & UPL_POP_DIRTY) dst_page->dirty = TRUE;
+ if (ops & UPL_POP_PAGEOUT) dst_page->pageout = TRUE;
+ if (ops & UPL_POP_PRECIOUS) dst_page->precious = TRUE;
+ if (ops & UPL_POP_ABSENT) dst_page->absent = TRUE;
+ if (ops & UPL_POP_BUSY) dst_page->busy = TRUE;
+ }
+
+ if(ops & UPL_POP_CLR) {
+ assert(dst_page->busy);
+ if (ops & UPL_POP_DIRTY) dst_page->dirty = FALSE;
+ if (ops & UPL_POP_PAGEOUT) dst_page->pageout = FALSE;
+ if (ops & UPL_POP_PRECIOUS) dst_page->precious = FALSE;
+ if (ops & UPL_POP_ABSENT) dst_page->absent = FALSE;
+ if (ops & UPL_POP_BUSY) {
+ dst_page->busy = FALSE;
+ PAGE_WAKEUP(dst_page);
+ }
+ }
+
+ if (dst_page->encrypted) {
+ /*
+ * ENCRYPTED SWAP:
+ * We need to decrypt this encrypted page before the
+ * caller can access its contents.
+ * But if the caller really wants to access the page's
+ * contents, they have to keep the page "busy".
+ * Otherwise, the page could get recycled or re-encrypted
+ * at any time.
+ */
+ if ((ops & UPL_POP_SET) && (ops & UPL_POP_BUSY) &&
+ dst_page->busy) {
+ /*
+ * The page is stable enough to be accessed by
+ * the caller, so make sure its contents are
+ * not encrypted.
+ */
+ vm_page_decrypt(dst_page, 0);
+ } else {
+ /*
+ * The page is not busy, so don't bother
+ * decrypting it, since anything could
+ * happen to it between now and when the
+ * caller wants to access it.
+ * We should not give the caller access
+ * to this page.
+ */
+ assert(!phys_entry);
+ }
+ }
+
+ if (phys_entry) {
+ /*
+ * The physical page number will remain valid
+ * only if the page is kept busy.
+ * ENCRYPTED SWAP: make sure we don't let the
+ * caller access an encrypted page.
+ */
+ assert(dst_page->busy);
+ assert(!dst_page->encrypted);
+ *phys_entry = dst_page->phys_page;
+ }
+
+ break;
+ }
+
+ vm_object_unlock(object);
+ return KERN_SUCCESS;
+
+}
+
+/*
+ * vm_object_range_op offers performance enhancement over
+ * vm_object_page_op for page_op functions which do not require page
+ * level state to be returned from the call. Page_op was created to provide
+ * a low-cost alternative to page manipulation via UPLs when only a single
+ * page was involved. The range_op call establishes the ability in the _op
+ * family of functions to work on multiple pages where the lack of page level
+ * state handling allows the caller to avoid the overhead of the upl structures.
+ */
+
+kern_return_t
+vm_object_range_op(
+ vm_object_t object,
+ vm_object_offset_t offset_beg,
+ vm_object_offset_t offset_end,
+ int ops,
+ uint32_t *range)
+{
+ vm_object_offset_t offset;
+ vm_page_t dst_page;
+
+ if (offset_end - offset_beg > (uint32_t) -1) {
+ /* range is too big and would overflow "*range" */
+ return KERN_INVALID_ARGUMENT;
+ }
+ if (object->resident_page_count == 0) {
+ if (range) {
+ if (ops & UPL_ROP_PRESENT) {
+ *range = 0;
+ } else {
+ *range = (uint32_t) (offset_end - offset_beg);
+ assert(*range == (offset_end - offset_beg));
+ }
+ }
+ return KERN_SUCCESS;
+ }
+ vm_object_lock(object);
+
+ if (object->phys_contiguous) {
+ vm_object_unlock(object);
+ return KERN_INVALID_OBJECT;
+ }
+
+ offset = offset_beg & ~PAGE_MASK_64;
+
+ while (offset < offset_end) {
+ dst_page = vm_page_lookup(object, offset);
+ if (dst_page != VM_PAGE_NULL) {
+ if (ops & UPL_ROP_DUMP) {
+ if (dst_page->busy || dst_page->cleaning) {
+ /*
+ * someone else is playing with the
+ * page, we will have to wait
+ */
+ PAGE_SLEEP(object, dst_page, THREAD_UNINT);
+ /*
+ * need to relook the page up since it's
+ * state may have changed while we slept
+ * it might even belong to a different object
+ * at this point
+ */
+ continue;
+ }
+ if (dst_page->pmapped == TRUE)
+ pmap_disconnect(dst_page->phys_page);
+
+ VM_PAGE_FREE(dst_page);
+
+ } else if ((ops & UPL_ROP_ABSENT) && !dst_page->absent)
+ break;
+ } else if (ops & UPL_ROP_PRESENT)
+ break;
+
+ offset += PAGE_SIZE;
+ }
+ vm_object_unlock(object);
+
+ if (range) {
+ if (offset > offset_end)
+ offset = offset_end;
+ if(offset > offset_beg) {
+ *range = (uint32_t) (offset - offset_beg);
+ assert(*range == (offset - offset_beg));
+ } else {
+ *range = 0;
+ }
+ }
+ return KERN_SUCCESS;
+}
+
+
+uint32_t scan_object_collision = 0;
+
+void
+vm_object_lock(vm_object_t object)
+{
+ if (object == vm_pageout_scan_wants_object) {
+ scan_object_collision++;
+ mutex_pause(2);
+ }
+ lck_rw_lock_exclusive(&object->Lock);
+}
+
+boolean_t
+vm_object_lock_avoid(vm_object_t object)
+{
+ if (object == vm_pageout_scan_wants_object) {
+ scan_object_collision++;
+ return TRUE;
+ }
+ return FALSE;
+}
+
+boolean_t
+_vm_object_lock_try(vm_object_t object)
+{
+ return (lck_rw_try_lock_exclusive(&object->Lock));
+}
+
+boolean_t
+vm_object_lock_try(vm_object_t object)
+{
+ // called from hibernate path so check before blocking
+ if (vm_object_lock_avoid(object) && ml_get_interrupts_enabled()) {
+ mutex_pause(2);
+ }
+ return _vm_object_lock_try(object);
+}
+void
+vm_object_lock_shared(vm_object_t object)
+{
+ if (vm_object_lock_avoid(object)) {
+ mutex_pause(2);
+ }
+ lck_rw_lock_shared(&object->Lock);
+}
+
+boolean_t
+vm_object_lock_try_shared(vm_object_t object)
+{
+ if (vm_object_lock_avoid(object)) {
+ mutex_pause(2);
+ }
+ return (lck_rw_try_lock_shared(&object->Lock));
+}