• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright (c) 2013 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
4 
5 #include "third_party/base/allocator/partition_allocator/partition_alloc.h"
6 
7 #include <string.h>
8 
9 #include "third_party/base/allocator/partition_allocator/oom.h"
10 #include "third_party/base/allocator/partition_allocator/spin_lock.h"
11 #include "third_party/base/compiler_specific.h"
12 
13 // Two partition pages are used as guard / metadata page so make sure the super
14 // page size is bigger.
15 static_assert(pdfium::base::kPartitionPageSize * 4 <=
16                   pdfium::base::kSuperPageSize,
17               "ok super page size");
18 static_assert(!(pdfium::base::kSuperPageSize %
19                 pdfium::base::kPartitionPageSize),
20               "ok super page multiple");
21 // Four system pages gives us room to hack out a still-guard-paged piece
22 // of metadata in the middle of a guard partition page.
23 static_assert(pdfium::base::kSystemPageSize * 4 <=
24                   pdfium::base::kPartitionPageSize,
25               "ok partition page size");
26 static_assert(!(pdfium::base::kPartitionPageSize %
27                 pdfium::base::kSystemPageSize),
28               "ok partition page multiple");
29 static_assert(sizeof(pdfium::base::PartitionPage) <=
30                   pdfium::base::kPageMetadataSize,
31               "PartitionPage should not be too big");
32 static_assert(sizeof(pdfium::base::PartitionBucket) <=
33                   pdfium::base::kPageMetadataSize,
34               "PartitionBucket should not be too big");
35 static_assert(sizeof(pdfium::base::PartitionSuperPageExtentEntry) <=
36                   pdfium::base::kPageMetadataSize,
37               "PartitionSuperPageExtentEntry should not be too big");
38 static_assert(pdfium::base::kPageMetadataSize *
39                       pdfium::base::kNumPartitionPagesPerSuperPage <=
40                   pdfium::base::kSystemPageSize,
41               "page metadata fits in hole");
42 // Check that some of our zanier calculations worked out as expected.
43 static_assert(pdfium::base::kGenericSmallestBucket == 8,
44               "generic smallest bucket");
45 static_assert(pdfium::base::kGenericMaxBucketed == 983040,
46               "generic max bucketed");
47 static_assert(pdfium::base::kMaxSystemPagesPerSlotSpan < (1 << 8),
48               "System pages per slot span must be less than 128.");
49 
50 namespace pdfium {
51 namespace base {
52 
53 subtle::SpinLock PartitionRootBase::gInitializedLock;
54 bool PartitionRootBase::gInitialized = false;
55 PartitionPage PartitionRootBase::gSeedPage;
56 PartitionBucket PartitionRootBase::gPagedBucket;
57 void (*PartitionRootBase::gOomHandlingFunction)() = nullptr;
58 PartitionAllocHooks::AllocationHook* PartitionAllocHooks::allocation_hook_ =
59     nullptr;
60 PartitionAllocHooks::FreeHook* PartitionAllocHooks::free_hook_ = nullptr;
61 
PartitionBucketNumSystemPages(size_t size)62 static uint8_t PartitionBucketNumSystemPages(size_t size) {
63   // This works out reasonably for the current bucket sizes of the generic
64   // allocator, and the current values of partition page size and constants.
65   // Specifically, we have enough room to always pack the slots perfectly into
66   // some number of system pages. The only waste is the waste associated with
67   // unfaulted pages (i.e. wasted address space).
68   // TODO: we end up using a lot of system pages for very small sizes. For
69   // example, we'll use 12 system pages for slot size 24. The slot size is
70   // so small that the waste would be tiny with just 4, or 1, system pages.
71   // Later, we can investigate whether there are anti-fragmentation benefits
72   // to using fewer system pages.
73   double best_waste_ratio = 1.0f;
74   uint16_t best_pages = 0;
75   if (size > kMaxSystemPagesPerSlotSpan * kSystemPageSize) {
76     DCHECK(!(size % kSystemPageSize));
77     best_pages = static_cast<uint16_t>(size / kSystemPageSize);
78     CHECK(best_pages < (1 << 8));
79     return static_cast<uint8_t>(best_pages);
80   }
81   DCHECK(size <= kMaxSystemPagesPerSlotSpan * kSystemPageSize);
82   for (uint16_t i = kNumSystemPagesPerPartitionPage - 1;
83        i <= kMaxSystemPagesPerSlotSpan; ++i) {
84     size_t page_size = kSystemPageSize * i;
85     size_t num_slots = page_size / size;
86     size_t waste = page_size - (num_slots * size);
87     // Leaving a page unfaulted is not free; the page will occupy an empty page
88     // table entry.  Make a simple attempt to account for that.
89     size_t num_remainder_pages = i & (kNumSystemPagesPerPartitionPage - 1);
90     size_t num_unfaulted_pages =
91         num_remainder_pages
92             ? (kNumSystemPagesPerPartitionPage - num_remainder_pages)
93             : 0;
94     waste += sizeof(void*) * num_unfaulted_pages;
95     double waste_ratio = (double)waste / (double)page_size;
96     if (waste_ratio < best_waste_ratio) {
97       best_waste_ratio = waste_ratio;
98       best_pages = i;
99     }
100   }
101   DCHECK(best_pages > 0);
102   CHECK(best_pages <= kMaxSystemPagesPerSlotSpan);
103   return static_cast<uint8_t>(best_pages);
104 }
105 
PartitionAllocBaseInit(PartitionRootBase * root)106 static void PartitionAllocBaseInit(PartitionRootBase* root) {
107   DCHECK(!root->initialized);
108   {
109     subtle::SpinLock::Guard guard(PartitionRootBase::gInitializedLock);
110     if (!PartitionRootBase::gInitialized) {
111       PartitionRootBase::gInitialized = true;
112       // We mark the seed page as free to make sure it is skipped by our
113       // logic to find a new active page.
114       PartitionRootBase::gPagedBucket.active_pages_head =
115           &PartitionRootGeneric::gSeedPage;
116     }
117   }
118 
119   root->initialized = true;
120   root->total_size_of_committed_pages = 0;
121   root->total_size_of_super_pages = 0;
122   root->total_size_of_direct_mapped_pages = 0;
123   root->next_super_page = 0;
124   root->next_partition_page = 0;
125   root->next_partition_page_end = 0;
126   root->first_extent = 0;
127   root->current_extent = 0;
128   root->direct_map_list = 0;
129 
130   memset(&root->global_empty_page_ring, '\0',
131          sizeof(root->global_empty_page_ring));
132   root->global_empty_page_ring_index = 0;
133 
134   // This is a "magic" value so we can test if a root pointer is valid.
135   root->inverted_self = ~reinterpret_cast<uintptr_t>(root);
136 }
137 
PartitionBucketInitBase(PartitionBucket * bucket,PartitionRootBase * root)138 static void PartitionBucketInitBase(PartitionBucket* bucket,
139                                     PartitionRootBase* root) {
140   bucket->active_pages_head = &PartitionRootGeneric::gSeedPage;
141   bucket->empty_pages_head = 0;
142   bucket->decommitted_pages_head = 0;
143   bucket->num_full_pages = 0;
144   bucket->num_system_pages_per_slot_span =
145       PartitionBucketNumSystemPages(bucket->slot_size);
146 }
147 
PartitionAllocGlobalInit(void (* oom_handling_function)())148 void PartitionAllocGlobalInit(void (*oom_handling_function)()) {
149   DCHECK(oom_handling_function);
150   PartitionRootBase::gOomHandlingFunction = oom_handling_function;
151 }
152 
PartitionAllocInit(PartitionRoot * root,size_t num_buckets,size_t max_allocation)153 void PartitionAllocInit(PartitionRoot* root,
154                         size_t num_buckets,
155                         size_t max_allocation) {
156   PartitionAllocBaseInit(root);
157 
158   root->num_buckets = num_buckets;
159   root->max_allocation = max_allocation;
160   size_t i;
161   for (i = 0; i < root->num_buckets; ++i) {
162     PartitionBucket* bucket = &root->buckets()[i];
163     if (!i)
164       bucket->slot_size = kAllocationGranularity;
165     else
166       bucket->slot_size = i << kBucketShift;
167     PartitionBucketInitBase(bucket, root);
168   }
169 }
170 
PartitionAllocGenericInit(PartitionRootGeneric * root)171 void PartitionAllocGenericInit(PartitionRootGeneric* root) {
172   subtle::SpinLock::Guard guard(root->lock);
173 
174   PartitionAllocBaseInit(root);
175 
176   // Precalculate some shift and mask constants used in the hot path.
177   // Example: malloc(41) == 101001 binary.
178   // Order is 6 (1 << 6-1) == 32 is highest bit set.
179   // order_index is the next three MSB == 010 == 2.
180   // sub_order_index_mask is a mask for the remaining bits == 11 (masking to 01
181   // for
182   // the sub_order_index).
183   size_t order;
184   for (order = 0; order <= kBitsPerSizeT; ++order) {
185     size_t order_index_shift;
186     if (order < kGenericNumBucketsPerOrderBits + 1)
187       order_index_shift = 0;
188     else
189       order_index_shift = order - (kGenericNumBucketsPerOrderBits + 1);
190     root->order_index_shifts[order] = order_index_shift;
191     size_t sub_order_index_mask;
192     if (order == kBitsPerSizeT) {
193       // This avoids invoking undefined behavior for an excessive shift.
194       sub_order_index_mask =
195           static_cast<size_t>(-1) >> (kGenericNumBucketsPerOrderBits + 1);
196     } else {
197       sub_order_index_mask = ((static_cast<size_t>(1) << order) - 1) >>
198                              (kGenericNumBucketsPerOrderBits + 1);
199     }
200     root->order_sub_index_masks[order] = sub_order_index_mask;
201   }
202 
203   // Set up the actual usable buckets first.
204   // Note that typical values (i.e. min allocation size of 8) will result in
205   // pseudo buckets (size==9 etc. or more generally, size is not a multiple
206   // of the smallest allocation granularity).
207   // We avoid them in the bucket lookup map, but we tolerate them to keep the
208   // code simpler and the structures more generic.
209   size_t i, j;
210   size_t current_size = kGenericSmallestBucket;
211   size_t currentIncrement =
212       kGenericSmallestBucket >> kGenericNumBucketsPerOrderBits;
213   PartitionBucket* bucket = &root->buckets[0];
214   for (i = 0; i < kGenericNumBucketedOrders; ++i) {
215     for (j = 0; j < kGenericNumBucketsPerOrder; ++j) {
216       bucket->slot_size = current_size;
217       PartitionBucketInitBase(bucket, root);
218       // Disable psuedo buckets so that touching them faults.
219       if (current_size % kGenericSmallestBucket)
220         bucket->active_pages_head = 0;
221       current_size += currentIncrement;
222       ++bucket;
223     }
224     currentIncrement <<= 1;
225   }
226   DCHECK(current_size == 1 << kGenericMaxBucketedOrder);
227   DCHECK(bucket == &root->buckets[0] + kGenericNumBuckets);
228 
229   // Then set up the fast size -> bucket lookup table.
230   bucket = &root->buckets[0];
231   PartitionBucket** bucketPtr = &root->bucket_lookups[0];
232   for (order = 0; order <= kBitsPerSizeT; ++order) {
233     for (j = 0; j < kGenericNumBucketsPerOrder; ++j) {
234       if (order < kGenericMinBucketedOrder) {
235         // Use the bucket of the finest granularity for malloc(0) etc.
236         *bucketPtr++ = &root->buckets[0];
237       } else if (order > kGenericMaxBucketedOrder) {
238         *bucketPtr++ = &PartitionRootGeneric::gPagedBucket;
239       } else {
240         PartitionBucket* validBucket = bucket;
241         // Skip over invalid buckets.
242         while (validBucket->slot_size % kGenericSmallestBucket)
243           validBucket++;
244         *bucketPtr++ = validBucket;
245         bucket++;
246       }
247     }
248   }
249   DCHECK(bucket == &root->buckets[0] + kGenericNumBuckets);
250   DCHECK(bucketPtr ==
251          &root->bucket_lookups[0] +
252              ((kBitsPerSizeT + 1) * kGenericNumBucketsPerOrder));
253   // And there's one last bucket lookup that will be hit for e.g. malloc(-1),
254   // which tries to overflow to a non-existant order.
255   *bucketPtr = &PartitionRootGeneric::gPagedBucket;
256 }
257 
258 #if !defined(ARCH_CPU_64_BITS)
PartitionOutOfMemoryWithLotsOfUncommitedPages()259 static NOINLINE void PartitionOutOfMemoryWithLotsOfUncommitedPages() {
260   OOM_CRASH();
261 }
262 #endif
263 
PartitionOutOfMemory(const PartitionRootBase * root)264 static NOINLINE void PartitionOutOfMemory(const PartitionRootBase* root) {
265 #if !defined(ARCH_CPU_64_BITS)
266   // Check whether this OOM is due to a lot of super pages that are allocated
267   // but not committed, probably due to http://crbug.com/421387.
268   if (root->total_size_of_super_pages +
269           root->total_size_of_direct_mapped_pages -
270           root->total_size_of_committed_pages >
271       kReasonableSizeOfUnusedPages) {
272     PartitionOutOfMemoryWithLotsOfUncommitedPages();
273   }
274 #endif
275   if (PartitionRootBase::gOomHandlingFunction)
276     (*PartitionRootBase::gOomHandlingFunction)();
277   OOM_CRASH();
278 }
279 
PartitionExcessiveAllocationSize()280 static NOINLINE void PartitionExcessiveAllocationSize() {
281   OOM_CRASH();
282 }
283 
PartitionBucketFull()284 static NOINLINE void PartitionBucketFull() {
285   OOM_CRASH();
286 }
287 
288 // partitionPageStateIs*
289 // Note that it's only valid to call these functions on pages found on one of
290 // the page lists. Specifically, you can't call these functions on full pages
291 // that were detached from the active list.
292 static bool ALWAYS_INLINE
PartitionPageStateIsActive(const PartitionPage * page)293 PartitionPageStateIsActive(const PartitionPage* page) {
294   DCHECK(page != &PartitionRootGeneric::gSeedPage);
295   DCHECK(!page->page_offset);
296   return (page->num_allocated_slots > 0 &&
297           (page->freelist_head || page->num_unprovisioned_slots));
298 }
299 
PartitionPageStateIsFull(const PartitionPage * page)300 static bool ALWAYS_INLINE PartitionPageStateIsFull(const PartitionPage* page) {
301   DCHECK(page != &PartitionRootGeneric::gSeedPage);
302   DCHECK(!page->page_offset);
303   bool ret = (page->num_allocated_slots == PartitionBucketSlots(page->bucket));
304   if (ret) {
305     DCHECK(!page->freelist_head);
306     DCHECK(!page->num_unprovisioned_slots);
307   }
308   return ret;
309 }
310 
PartitionPageStateIsEmpty(const PartitionPage * page)311 static bool ALWAYS_INLINE PartitionPageStateIsEmpty(const PartitionPage* page) {
312   DCHECK(page != &PartitionRootGeneric::gSeedPage);
313   DCHECK(!page->page_offset);
314   return (!page->num_allocated_slots && page->freelist_head);
315 }
316 
317 static bool ALWAYS_INLINE
PartitionPageStateIsDecommitted(const PartitionPage * page)318 PartitionPageStateIsDecommitted(const PartitionPage* page) {
319   DCHECK(page != &PartitionRootGeneric::gSeedPage);
320   DCHECK(!page->page_offset);
321   bool ret = (!page->num_allocated_slots && !page->freelist_head);
322   if (ret) {
323     DCHECK(!page->num_unprovisioned_slots);
324     DCHECK(page->empty_cache_index == -1);
325   }
326   return ret;
327 }
328 
PartitionIncreaseCommittedPages(PartitionRootBase * root,size_t len)329 static void PartitionIncreaseCommittedPages(PartitionRootBase* root,
330                                             size_t len) {
331   root->total_size_of_committed_pages += len;
332   DCHECK(root->total_size_of_committed_pages <=
333          root->total_size_of_super_pages +
334              root->total_size_of_direct_mapped_pages);
335 }
336 
PartitionDecreaseCommittedPages(PartitionRootBase * root,size_t len)337 static void PartitionDecreaseCommittedPages(PartitionRootBase* root,
338                                             size_t len) {
339   root->total_size_of_committed_pages -= len;
340   DCHECK(root->total_size_of_committed_pages <=
341          root->total_size_of_super_pages +
342              root->total_size_of_direct_mapped_pages);
343 }
344 
PartitionDecommitSystemPages(PartitionRootBase * root,void * address,size_t length)345 static ALWAYS_INLINE void PartitionDecommitSystemPages(PartitionRootBase* root,
346                                                        void* address,
347                                                        size_t length) {
348   DecommitSystemPages(address, length);
349   PartitionDecreaseCommittedPages(root, length);
350 }
351 
PartitionRecommitSystemPages(PartitionRootBase * root,void * address,size_t length)352 static ALWAYS_INLINE void PartitionRecommitSystemPages(PartitionRootBase* root,
353                                                        void* address,
354                                                        size_t length) {
355   RecommitSystemPages(address, length);
356   PartitionIncreaseCommittedPages(root, length);
357 }
358 
PartitionAllocPartitionPages(PartitionRootBase * root,int flags,uint16_t num_partition_pages)359 static ALWAYS_INLINE void* PartitionAllocPartitionPages(
360     PartitionRootBase* root,
361     int flags,
362     uint16_t num_partition_pages) {
363   DCHECK(!(reinterpret_cast<uintptr_t>(root->next_partition_page) %
364            kPartitionPageSize));
365   DCHECK(!(reinterpret_cast<uintptr_t>(root->next_partition_page_end) %
366            kPartitionPageSize));
367   DCHECK(num_partition_pages <= kNumPartitionPagesPerSuperPage);
368   size_t total_size = kPartitionPageSize * num_partition_pages;
369   size_t num_partition_pages_left =
370       (root->next_partition_page_end - root->next_partition_page) >>
371       kPartitionPageShift;
372   if (LIKELY(num_partition_pages_left >= num_partition_pages)) {
373     // In this case, we can still hand out pages from the current super page
374     // allocation.
375     char* ret = root->next_partition_page;
376     root->next_partition_page += total_size;
377     PartitionIncreaseCommittedPages(root, total_size);
378     return ret;
379   }
380 
381   // Need a new super page. We want to allocate super pages in a continguous
382   // address region as much as possible. This is important for not causing
383   // page table bloat and not fragmenting address spaces in 32 bit
384   // architectures.
385   char* requestedAddress = root->next_super_page;
386   char* super_page = reinterpret_cast<char*>(AllocPages(
387       requestedAddress, kSuperPageSize, kSuperPageSize, PageAccessible));
388   if (UNLIKELY(!super_page))
389     return 0;
390 
391   root->total_size_of_super_pages += kSuperPageSize;
392   PartitionIncreaseCommittedPages(root, total_size);
393 
394   root->next_super_page = super_page + kSuperPageSize;
395   char* ret = super_page + kPartitionPageSize;
396   root->next_partition_page = ret + total_size;
397   root->next_partition_page_end = root->next_super_page - kPartitionPageSize;
398   // Make the first partition page in the super page a guard page, but leave a
399   // hole in the middle.
400   // This is where we put page metadata and also a tiny amount of extent
401   // metadata.
402   SetSystemPagesInaccessible(super_page, kSystemPageSize);
403   SetSystemPagesInaccessible(super_page + (kSystemPageSize * 2),
404                              kPartitionPageSize - (kSystemPageSize * 2));
405   // Also make the last partition page a guard page.
406   SetSystemPagesInaccessible(super_page + (kSuperPageSize - kPartitionPageSize),
407                              kPartitionPageSize);
408 
409   // If we were after a specific address, but didn't get it, assume that
410   // the system chose a lousy address. Here most OS'es have a default
411   // algorithm that isn't randomized. For example, most Linux
412   // distributions will allocate the mapping directly before the last
413   // successful mapping, which is far from random. So we just get fresh
414   // randomness for the next mapping attempt.
415   if (requestedAddress && requestedAddress != super_page)
416     root->next_super_page = 0;
417 
418   // We allocated a new super page so update super page metadata.
419   // First check if this is a new extent or not.
420   PartitionSuperPageExtentEntry* latest_extent =
421       reinterpret_cast<PartitionSuperPageExtentEntry*>(
422           PartitionSuperPageToMetadataArea(super_page));
423   // By storing the root in every extent metadata object, we have a fast way
424   // to go from a pointer within the partition to the root object.
425   latest_extent->root = root;
426   // Most new extents will be part of a larger extent, and these three fields
427   // are unused, but we initialize them to 0 so that we get a clear signal
428   // in case they are accidentally used.
429   latest_extent->super_page_base = 0;
430   latest_extent->super_pages_end = 0;
431   latest_extent->next = 0;
432 
433   PartitionSuperPageExtentEntry* current_extent = root->current_extent;
434   bool isNewExtent = (super_page != requestedAddress);
435   if (UNLIKELY(isNewExtent)) {
436     if (UNLIKELY(!current_extent)) {
437       DCHECK(!root->first_extent);
438       root->first_extent = latest_extent;
439     } else {
440       DCHECK(current_extent->super_page_base);
441       current_extent->next = latest_extent;
442     }
443     root->current_extent = latest_extent;
444     latest_extent->super_page_base = super_page;
445     latest_extent->super_pages_end = super_page + kSuperPageSize;
446   } else {
447     // We allocated next to an existing extent so just nudge the size up a
448     // little.
449     DCHECK(current_extent->super_pages_end);
450     current_extent->super_pages_end += kSuperPageSize;
451     DCHECK(ret >= current_extent->super_page_base &&
452            ret < current_extent->super_pages_end);
453   }
454   return ret;
455 }
456 
457 static ALWAYS_INLINE uint16_t
PartitionBucketPartitionPages(const PartitionBucket * bucket)458 PartitionBucketPartitionPages(const PartitionBucket* bucket) {
459   return (bucket->num_system_pages_per_slot_span +
460           (kNumSystemPagesPerPartitionPage - 1)) /
461          kNumSystemPagesPerPartitionPage;
462 }
463 
PartitionPageReset(PartitionPage * page)464 static ALWAYS_INLINE void PartitionPageReset(PartitionPage* page) {
465   DCHECK(PartitionPageStateIsDecommitted(page));
466 
467   page->num_unprovisioned_slots = PartitionBucketSlots(page->bucket);
468   DCHECK(page->num_unprovisioned_slots);
469 
470   page->next_page = nullptr;
471 }
472 
PartitionPageSetup(PartitionPage * page,PartitionBucket * bucket)473 static ALWAYS_INLINE void PartitionPageSetup(PartitionPage* page,
474                                              PartitionBucket* bucket) {
475   // The bucket never changes. We set it up once.
476   page->bucket = bucket;
477   page->empty_cache_index = -1;
478 
479   PartitionPageReset(page);
480 
481   // If this page has just a single slot, do not set up page offsets for any
482   // page metadata other than the first one. This ensures that attempts to
483   // touch invalid page metadata fail.
484   if (page->num_unprovisioned_slots == 1)
485     return;
486 
487   uint16_t num_partition_pages = PartitionBucketPartitionPages(bucket);
488   char* page_char_ptr = reinterpret_cast<char*>(page);
489   for (uint16_t i = 1; i < num_partition_pages; ++i) {
490     page_char_ptr += kPageMetadataSize;
491     PartitionPage* secondary_page =
492         reinterpret_cast<PartitionPage*>(page_char_ptr);
493     secondary_page->page_offset = i;
494   }
495 }
496 
PartitionPageAllocAndFillFreelist(PartitionPage * page)497 static ALWAYS_INLINE char* PartitionPageAllocAndFillFreelist(
498     PartitionPage* page) {
499   DCHECK(page != &PartitionRootGeneric::gSeedPage);
500   uint16_t num_slots = page->num_unprovisioned_slots;
501   DCHECK(num_slots);
502   PartitionBucket* bucket = page->bucket;
503   // We should only get here when _every_ slot is either used or unprovisioned.
504   // (The third state is "on the freelist". If we have a non-empty freelist, we
505   // should not get here.)
506   DCHECK(num_slots + page->num_allocated_slots == PartitionBucketSlots(bucket));
507   // Similarly, make explicitly sure that the freelist is empty.
508   DCHECK(!page->freelist_head);
509   DCHECK(page->num_allocated_slots >= 0);
510 
511   size_t size = bucket->slot_size;
512   char* base = reinterpret_cast<char*>(PartitionPageToPointer(page));
513   char* return_object = base + (size * page->num_allocated_slots);
514   char* firstFreelistPointer = return_object + size;
515   char* firstFreelistPointerExtent =
516       firstFreelistPointer + sizeof(PartitionFreelistEntry*);
517   // Our goal is to fault as few system pages as possible. We calculate the
518   // page containing the "end" of the returned slot, and then allow freelist
519   // pointers to be written up to the end of that page.
520   char* sub_page_limit = reinterpret_cast<char*>(
521       RoundUpToSystemPage(reinterpret_cast<size_t>(firstFreelistPointer)));
522   char* slots_limit = return_object + (size * num_slots);
523   char* freelist_limit = sub_page_limit;
524   if (UNLIKELY(slots_limit < freelist_limit))
525     freelist_limit = slots_limit;
526 
527   uint16_t num_new_freelist_entries = 0;
528   if (LIKELY(firstFreelistPointerExtent <= freelist_limit)) {
529     // Only consider used space in the slot span. If we consider wasted
530     // space, we may get an off-by-one when a freelist pointer fits in the
531     // wasted space, but a slot does not.
532     // We know we can fit at least one freelist pointer.
533     num_new_freelist_entries = 1;
534     // Any further entries require space for the whole slot span.
535     num_new_freelist_entries += static_cast<uint16_t>(
536         (freelist_limit - firstFreelistPointerExtent) / size);
537   }
538 
539   // We always return an object slot -- that's the +1 below.
540   // We do not neccessarily create any new freelist entries, because we cross
541   // sub page boundaries frequently for large bucket sizes.
542   DCHECK(num_new_freelist_entries + 1 <= num_slots);
543   num_slots -= (num_new_freelist_entries + 1);
544   page->num_unprovisioned_slots = num_slots;
545   page->num_allocated_slots++;
546 
547   if (LIKELY(num_new_freelist_entries)) {
548     char* freelist_pointer = firstFreelistPointer;
549     PartitionFreelistEntry* entry =
550         reinterpret_cast<PartitionFreelistEntry*>(freelist_pointer);
551     page->freelist_head = entry;
552     while (--num_new_freelist_entries) {
553       freelist_pointer += size;
554       PartitionFreelistEntry* next_entry =
555           reinterpret_cast<PartitionFreelistEntry*>(freelist_pointer);
556       entry->next = PartitionFreelistMask(next_entry);
557       entry = next_entry;
558     }
559     entry->next = PartitionFreelistMask(0);
560   } else {
561     page->freelist_head = 0;
562   }
563   return return_object;
564 }
565 
566 // This helper function scans a bucket's active page list for a suitable new
567 // active page.
568 // When it finds a suitable new active page (one that has free slots and is not
569 // empty), it is set as the new active page. If there is no suitable new
570 // active page, the current active page is set to the seed page.
571 // As potential pages are scanned, they are tidied up according to their state.
572 // Empty pages are swept on to the empty page list, decommitted pages on to the
573 // decommitted page list and full pages are unlinked from any list.
PartitionSetNewActivePage(PartitionBucket * bucket)574 static bool PartitionSetNewActivePage(PartitionBucket* bucket) {
575   PartitionPage* page = bucket->active_pages_head;
576   if (page == &PartitionRootBase::gSeedPage)
577     return false;
578 
579   PartitionPage* next_page;
580 
581   for (; page; page = next_page) {
582     next_page = page->next_page;
583     DCHECK(page->bucket == bucket);
584     DCHECK(page != bucket->empty_pages_head);
585     DCHECK(page != bucket->decommitted_pages_head);
586 
587     // Deal with empty and decommitted pages.
588     if (LIKELY(PartitionPageStateIsActive(page))) {
589       // This page is usable because it has freelist entries, or has
590       // unprovisioned slots we can create freelist entries from.
591       bucket->active_pages_head = page;
592       return true;
593     }
594     if (LIKELY(PartitionPageStateIsEmpty(page))) {
595       page->next_page = bucket->empty_pages_head;
596       bucket->empty_pages_head = page;
597     } else if (LIKELY(PartitionPageStateIsDecommitted(page))) {
598       page->next_page = bucket->decommitted_pages_head;
599       bucket->decommitted_pages_head = page;
600     } else {
601       DCHECK(PartitionPageStateIsFull(page));
602       // If we get here, we found a full page. Skip over it too, and also
603       // tag it as full (via a negative value). We need it tagged so that
604       // free'ing can tell, and move it back into the active page list.
605       page->num_allocated_slots = -page->num_allocated_slots;
606       ++bucket->num_full_pages;
607       // num_full_pages is a uint16_t for efficient packing so guard against
608       // overflow to be safe.
609       if (UNLIKELY(!bucket->num_full_pages))
610         PartitionBucketFull();
611       // Not necessary but might help stop accidents.
612       page->next_page = 0;
613     }
614   }
615 
616   bucket->active_pages_head = &PartitionRootGeneric::gSeedPage;
617   return false;
618 }
619 
partitionPageToDirectMapExtent(PartitionPage * page)620 static ALWAYS_INLINE PartitionDirectMapExtent* partitionPageToDirectMapExtent(
621     PartitionPage* page) {
622   DCHECK(PartitionBucketIsDirectMapped(page->bucket));
623   return reinterpret_cast<PartitionDirectMapExtent*>(
624       reinterpret_cast<char*>(page) + 3 * kPageMetadataSize);
625 }
626 
PartitionPageSetRawSize(PartitionPage * page,size_t size)627 static ALWAYS_INLINE void PartitionPageSetRawSize(PartitionPage* page,
628                                                   size_t size) {
629   size_t* raw_size_ptr = PartitionPageGetRawSizePtr(page);
630   if (UNLIKELY(raw_size_ptr != nullptr))
631     *raw_size_ptr = size;
632 }
633 
PartitionDirectMap(PartitionRootBase * root,int flags,size_t raw_size)634 static ALWAYS_INLINE PartitionPage* PartitionDirectMap(PartitionRootBase* root,
635                                                        int flags,
636                                                        size_t raw_size) {
637   size_t size = PartitionDirectMapSize(raw_size);
638 
639   // Because we need to fake looking like a super page, we need to allocate
640   // a bunch of system pages more than "size":
641   // - The first few system pages are the partition page in which the super
642   // page metadata is stored. We fault just one system page out of a partition
643   // page sized clump.
644   // - We add a trailing guard page on 32-bit (on 64-bit we rely on the
645   // massive address space plus randomization instead).
646   size_t map_size = size + kPartitionPageSize;
647 #if !defined(ARCH_CPU_64_BITS)
648   map_size += kSystemPageSize;
649 #endif
650   // Round up to the allocation granularity.
651   map_size += kPageAllocationGranularityOffsetMask;
652   map_size &= kPageAllocationGranularityBaseMask;
653 
654   // TODO: these pages will be zero-filled. Consider internalizing an
655   // allocZeroed() API so we can avoid a memset() entirely in this case.
656   char* ptr = reinterpret_cast<char*>(
657       AllocPages(0, map_size, kSuperPageSize, PageAccessible));
658   if (UNLIKELY(!ptr))
659     return nullptr;
660 
661   size_t committed_page_size = size + kSystemPageSize;
662   root->total_size_of_direct_mapped_pages += committed_page_size;
663   PartitionIncreaseCommittedPages(root, committed_page_size);
664 
665   char* slot = ptr + kPartitionPageSize;
666   SetSystemPagesInaccessible(ptr + (kSystemPageSize * 2),
667                              kPartitionPageSize - (kSystemPageSize * 2));
668 #if !defined(ARCH_CPU_64_BITS)
669   SetSystemPagesInaccessible(ptr, kSystemPageSize);
670   SetSystemPagesInaccessible(slot + size, kSystemPageSize);
671 #endif
672 
673   PartitionSuperPageExtentEntry* extent =
674       reinterpret_cast<PartitionSuperPageExtentEntry*>(
675           PartitionSuperPageToMetadataArea(ptr));
676   extent->root = root;
677   // The new structures are all located inside a fresh system page so they
678   // will all be zeroed out. These DCHECKs are for documentation.
679   DCHECK(!extent->super_page_base);
680   DCHECK(!extent->super_pages_end);
681   DCHECK(!extent->next);
682   PartitionPage* page = PartitionPointerToPageNoAlignmentCheck(slot);
683   PartitionBucket* bucket = reinterpret_cast<PartitionBucket*>(
684       reinterpret_cast<char*>(page) + (kPageMetadataSize * 2));
685   DCHECK(!page->next_page);
686   DCHECK(!page->num_allocated_slots);
687   DCHECK(!page->num_unprovisioned_slots);
688   DCHECK(!page->page_offset);
689   DCHECK(!page->empty_cache_index);
690   page->bucket = bucket;
691   page->freelist_head = reinterpret_cast<PartitionFreelistEntry*>(slot);
692   PartitionFreelistEntry* next_entry =
693       reinterpret_cast<PartitionFreelistEntry*>(slot);
694   next_entry->next = PartitionFreelistMask(0);
695 
696   DCHECK(!bucket->active_pages_head);
697   DCHECK(!bucket->empty_pages_head);
698   DCHECK(!bucket->decommitted_pages_head);
699   DCHECK(!bucket->num_system_pages_per_slot_span);
700   DCHECK(!bucket->num_full_pages);
701   bucket->slot_size = size;
702 
703   PartitionDirectMapExtent* map_extent = partitionPageToDirectMapExtent(page);
704   map_extent->map_size = map_size - kPartitionPageSize - kSystemPageSize;
705   map_extent->bucket = bucket;
706 
707   // Maintain the doubly-linked list of all direct mappings.
708   map_extent->next_extent = root->direct_map_list;
709   if (map_extent->next_extent)
710     map_extent->next_extent->prev_extent = map_extent;
711   map_extent->prev_extent = nullptr;
712   root->direct_map_list = map_extent;
713 
714   return page;
715 }
716 
PartitionDirectUnmap(PartitionPage * page)717 static ALWAYS_INLINE void PartitionDirectUnmap(PartitionPage* page) {
718   PartitionRootBase* root = PartitionPageToRoot(page);
719   const PartitionDirectMapExtent* extent = partitionPageToDirectMapExtent(page);
720   size_t unmap_size = extent->map_size;
721 
722   // Maintain the doubly-linked list of all direct mappings.
723   if (extent->prev_extent) {
724     DCHECK(extent->prev_extent->next_extent == extent);
725     extent->prev_extent->next_extent = extent->next_extent;
726   } else {
727     root->direct_map_list = extent->next_extent;
728   }
729   if (extent->next_extent) {
730     DCHECK(extent->next_extent->prev_extent == extent);
731     extent->next_extent->prev_extent = extent->prev_extent;
732   }
733 
734   // Add on the size of the trailing guard page and preceeding partition
735   // page.
736   unmap_size += kPartitionPageSize + kSystemPageSize;
737 
738   size_t uncommitted_page_size = page->bucket->slot_size + kSystemPageSize;
739   PartitionDecreaseCommittedPages(root, uncommitted_page_size);
740   DCHECK(root->total_size_of_direct_mapped_pages >= uncommitted_page_size);
741   root->total_size_of_direct_mapped_pages -= uncommitted_page_size;
742 
743   DCHECK(!(unmap_size & kPageAllocationGranularityOffsetMask));
744 
745   char* ptr = reinterpret_cast<char*>(PartitionPageToPointer(page));
746   // Account for the mapping starting a partition page before the actual
747   // allocation address.
748   ptr -= kPartitionPageSize;
749 
750   FreePages(ptr, unmap_size);
751 }
752 
PartitionAllocSlowPath(PartitionRootBase * root,int flags,size_t size,PartitionBucket * bucket)753 void* PartitionAllocSlowPath(PartitionRootBase* root,
754                              int flags,
755                              size_t size,
756                              PartitionBucket* bucket) {
757   // The slow path is called when the freelist is empty.
758   DCHECK(!bucket->active_pages_head->freelist_head);
759 
760   PartitionPage* new_page = nullptr;
761 
762   // For the PartitionAllocGeneric API, we have a bunch of buckets marked
763   // as special cases. We bounce them through to the slow path so that we
764   // can still have a blazing fast hot path due to lack of corner-case
765   // branches.
766   bool returnNull = flags & PartitionAllocReturnNull;
767   if (UNLIKELY(PartitionBucketIsDirectMapped(bucket))) {
768     DCHECK(size > kGenericMaxBucketed);
769     DCHECK(bucket == &PartitionRootBase::gPagedBucket);
770     DCHECK(bucket->active_pages_head == &PartitionRootGeneric::gSeedPage);
771     if (size > kGenericMaxDirectMapped) {
772       if (returnNull)
773         return nullptr;
774       PartitionExcessiveAllocationSize();
775     }
776     new_page = PartitionDirectMap(root, flags, size);
777   } else if (LIKELY(PartitionSetNewActivePage(bucket))) {
778     // First, did we find an active page in the active pages list?
779     new_page = bucket->active_pages_head;
780     DCHECK(PartitionPageStateIsActive(new_page));
781   } else if (LIKELY(bucket->empty_pages_head != nullptr) ||
782              LIKELY(bucket->decommitted_pages_head != nullptr)) {
783     // Second, look in our lists of empty and decommitted pages.
784     // Check empty pages first, which are preferred, but beware that an
785     // empty page might have been decommitted.
786     while (LIKELY((new_page = bucket->empty_pages_head) != nullptr)) {
787       DCHECK(new_page->bucket == bucket);
788       DCHECK(PartitionPageStateIsEmpty(new_page) ||
789              PartitionPageStateIsDecommitted(new_page));
790       bucket->empty_pages_head = new_page->next_page;
791       // Accept the empty page unless it got decommitted.
792       if (new_page->freelist_head) {
793         new_page->next_page = nullptr;
794         break;
795       }
796       DCHECK(PartitionPageStateIsDecommitted(new_page));
797       new_page->next_page = bucket->decommitted_pages_head;
798       bucket->decommitted_pages_head = new_page;
799     }
800     if (UNLIKELY(!new_page) &&
801         LIKELY(bucket->decommitted_pages_head != nullptr)) {
802       new_page = bucket->decommitted_pages_head;
803       DCHECK(new_page->bucket == bucket);
804       DCHECK(PartitionPageStateIsDecommitted(new_page));
805       bucket->decommitted_pages_head = new_page->next_page;
806       void* addr = PartitionPageToPointer(new_page);
807       PartitionRecommitSystemPages(root, addr,
808                                    PartitionBucketBytes(new_page->bucket));
809       PartitionPageReset(new_page);
810     }
811     DCHECK(new_page);
812   } else {
813     // Third. If we get here, we need a brand new page.
814     uint16_t num_partition_pages = PartitionBucketPartitionPages(bucket);
815     void* rawPages =
816         PartitionAllocPartitionPages(root, flags, num_partition_pages);
817     if (LIKELY(rawPages != nullptr)) {
818       new_page = PartitionPointerToPageNoAlignmentCheck(rawPages);
819       PartitionPageSetup(new_page, bucket);
820     }
821   }
822 
823   // Bail if we had a memory allocation failure.
824   if (UNLIKELY(!new_page)) {
825     DCHECK(bucket->active_pages_head == &PartitionRootGeneric::gSeedPage);
826     if (returnNull)
827       return nullptr;
828     PartitionOutOfMemory(root);
829   }
830 
831   bucket = new_page->bucket;
832   DCHECK(bucket != &PartitionRootBase::gPagedBucket);
833   bucket->active_pages_head = new_page;
834   PartitionPageSetRawSize(new_page, size);
835 
836   // If we found an active page with free slots, or an empty page, we have a
837   // usable freelist head.
838   if (LIKELY(new_page->freelist_head != nullptr)) {
839     PartitionFreelistEntry* entry = new_page->freelist_head;
840     PartitionFreelistEntry* new_head = PartitionFreelistMask(entry->next);
841     new_page->freelist_head = new_head;
842     new_page->num_allocated_slots++;
843     return entry;
844   }
845   // Otherwise, we need to build the freelist.
846   DCHECK(new_page->num_unprovisioned_slots);
847   return PartitionPageAllocAndFillFreelist(new_page);
848 }
849 
PartitionDecommitPage(PartitionRootBase * root,PartitionPage * page)850 static ALWAYS_INLINE void PartitionDecommitPage(PartitionRootBase* root,
851                                                 PartitionPage* page) {
852   DCHECK(PartitionPageStateIsEmpty(page));
853   DCHECK(!PartitionBucketIsDirectMapped(page->bucket));
854   void* addr = PartitionPageToPointer(page);
855   PartitionDecommitSystemPages(root, addr, PartitionBucketBytes(page->bucket));
856 
857   // We actually leave the decommitted page in the active list. We'll sweep
858   // it on to the decommitted page list when we next walk the active page
859   // list.
860   // Pulling this trick enables us to use a singly-linked page list for all
861   // cases, which is critical in keeping the page metadata structure down to
862   // 32 bytes in size.
863   page->freelist_head = 0;
864   page->num_unprovisioned_slots = 0;
865   DCHECK(PartitionPageStateIsDecommitted(page));
866 }
867 
PartitionDecommitPageIfPossible(PartitionRootBase * root,PartitionPage * page)868 static void PartitionDecommitPageIfPossible(PartitionRootBase* root,
869                                             PartitionPage* page) {
870   DCHECK(page->empty_cache_index >= 0);
871   DCHECK(static_cast<unsigned>(page->empty_cache_index) < kMaxFreeableSpans);
872   DCHECK(page == root->global_empty_page_ring[page->empty_cache_index]);
873   page->empty_cache_index = -1;
874   if (PartitionPageStateIsEmpty(page))
875     PartitionDecommitPage(root, page);
876 }
877 
PartitionRegisterEmptyPage(PartitionPage * page)878 static ALWAYS_INLINE void PartitionRegisterEmptyPage(PartitionPage* page) {
879   DCHECK(PartitionPageStateIsEmpty(page));
880   PartitionRootBase* root = PartitionPageToRoot(page);
881 
882   // If the page is already registered as empty, give it another life.
883   if (page->empty_cache_index != -1) {
884     DCHECK(page->empty_cache_index >= 0);
885     DCHECK(static_cast<unsigned>(page->empty_cache_index) < kMaxFreeableSpans);
886     DCHECK(root->global_empty_page_ring[page->empty_cache_index] == page);
887     root->global_empty_page_ring[page->empty_cache_index] = 0;
888   }
889 
890   int16_t current_index = root->global_empty_page_ring_index;
891   PartitionPage* pageToDecommit = root->global_empty_page_ring[current_index];
892   // The page might well have been re-activated, filled up, etc. before we get
893   // around to looking at it here.
894   if (pageToDecommit)
895     PartitionDecommitPageIfPossible(root, pageToDecommit);
896 
897   // We put the empty slot span on our global list of "pages that were once
898   // empty". thus providing it a bit of breathing room to get re-used before
899   // we really free it. This improves performance, particularly on Mac OS X
900   // which has subpar memory management performance.
901   root->global_empty_page_ring[current_index] = page;
902   page->empty_cache_index = current_index;
903   ++current_index;
904   if (current_index == kMaxFreeableSpans)
905     current_index = 0;
906   root->global_empty_page_ring_index = current_index;
907 }
908 
PartitionDecommitEmptyPages(PartitionRootBase * root)909 static void PartitionDecommitEmptyPages(PartitionRootBase* root) {
910   for (size_t i = 0; i < kMaxFreeableSpans; ++i) {
911     PartitionPage* page = root->global_empty_page_ring[i];
912     if (page)
913       PartitionDecommitPageIfPossible(root, page);
914     root->global_empty_page_ring[i] = nullptr;
915   }
916 }
917 
PartitionFreeSlowPath(PartitionPage * page)918 void PartitionFreeSlowPath(PartitionPage* page) {
919   PartitionBucket* bucket = page->bucket;
920   DCHECK(page != &PartitionRootGeneric::gSeedPage);
921   if (LIKELY(page->num_allocated_slots == 0)) {
922     // Page became fully unused.
923     if (UNLIKELY(PartitionBucketIsDirectMapped(bucket))) {
924       PartitionDirectUnmap(page);
925       return;
926     }
927     // If it's the current active page, change it. We bounce the page to
928     // the empty list as a force towards defragmentation.
929     if (LIKELY(page == bucket->active_pages_head))
930       (void)PartitionSetNewActivePage(bucket);
931     DCHECK(bucket->active_pages_head != page);
932 
933     PartitionPageSetRawSize(page, 0);
934     DCHECK(!PartitionPageGetRawSize(page));
935 
936     PartitionRegisterEmptyPage(page);
937   } else {
938     DCHECK(!PartitionBucketIsDirectMapped(bucket));
939     // Ensure that the page is full. That's the only valid case if we
940     // arrive here.
941     DCHECK(page->num_allocated_slots < 0);
942     // A transition of num_allocated_slots from 0 to -1 is not legal, and
943     // likely indicates a double-free.
944     CHECK(page->num_allocated_slots != -1);
945     page->num_allocated_slots = -page->num_allocated_slots - 2;
946     DCHECK(page->num_allocated_slots == PartitionBucketSlots(bucket) - 1);
947     // Fully used page became partially used. It must be put back on the
948     // non-full page list. Also make it the current page to increase the
949     // chances of it being filled up again. The old current page will be
950     // the next page.
951     DCHECK(!page->next_page);
952     if (LIKELY(bucket->active_pages_head != &PartitionRootGeneric::gSeedPage))
953       page->next_page = bucket->active_pages_head;
954     bucket->active_pages_head = page;
955     --bucket->num_full_pages;
956     // Special case: for a partition page with just a single slot, it may
957     // now be empty and we want to run it through the empty logic.
958     if (UNLIKELY(page->num_allocated_slots == 0))
959       PartitionFreeSlowPath(page);
960   }
961 }
962 
partitionReallocDirectMappedInPlace(PartitionRootGeneric * root,PartitionPage * page,size_t raw_size)963 bool partitionReallocDirectMappedInPlace(PartitionRootGeneric* root,
964                                          PartitionPage* page,
965                                          size_t raw_size) {
966   DCHECK(PartitionBucketIsDirectMapped(page->bucket));
967 
968   raw_size = PartitionCookieSizeAdjustAdd(raw_size);
969 
970   // Note that the new size might be a bucketed size; this function is called
971   // whenever we're reallocating a direct mapped allocation.
972   size_t new_size = PartitionDirectMapSize(raw_size);
973   if (new_size < kGenericMinDirectMappedDownsize)
974     return false;
975 
976   // bucket->slot_size is the current size of the allocation.
977   size_t current_size = page->bucket->slot_size;
978   if (new_size == current_size)
979     return true;
980 
981   char* char_ptr = static_cast<char*>(PartitionPageToPointer(page));
982 
983   if (new_size < current_size) {
984     size_t map_size = partitionPageToDirectMapExtent(page)->map_size;
985 
986     // Don't reallocate in-place if new size is less than 80 % of the full
987     // map size, to avoid holding on to too much unused address space.
988     if ((new_size / kSystemPageSize) * 5 < (map_size / kSystemPageSize) * 4)
989       return false;
990 
991     // Shrink by decommitting unneeded pages and making them inaccessible.
992     size_t decommitSize = current_size - new_size;
993     PartitionDecommitSystemPages(root, char_ptr + new_size, decommitSize);
994     SetSystemPagesInaccessible(char_ptr + new_size, decommitSize);
995   } else if (new_size <= partitionPageToDirectMapExtent(page)->map_size) {
996     // Grow within the actually allocated memory. Just need to make the
997     // pages accessible again.
998     size_t recommit_size = new_size - current_size;
999     bool ret = SetSystemPagesAccessible(char_ptr + current_size, recommit_size);
1000     CHECK(ret);
1001     PartitionRecommitSystemPages(root, char_ptr + current_size, recommit_size);
1002 
1003 #if DCHECK_IS_ON()
1004     memset(char_ptr + current_size, kUninitializedByte, recommit_size);
1005 #endif
1006   } else {
1007     // We can't perform the realloc in-place.
1008     // TODO: support this too when possible.
1009     return false;
1010   }
1011 
1012 #if DCHECK_IS_ON()
1013   // Write a new trailing cookie.
1014   PartitionCookieWriteValue(char_ptr + raw_size - kCookieSize);
1015 #endif
1016 
1017   PartitionPageSetRawSize(page, raw_size);
1018   DCHECK(PartitionPageGetRawSize(page) == raw_size);
1019 
1020   page->bucket->slot_size = new_size;
1021   return true;
1022 }
1023 
PartitionReallocGeneric(PartitionRootGeneric * root,void * ptr,size_t new_size,const char * type_name)1024 void* PartitionReallocGeneric(PartitionRootGeneric* root,
1025                               void* ptr,
1026                               size_t new_size,
1027                               const char* type_name) {
1028 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
1029   return realloc(ptr, new_size);
1030 #else
1031   if (UNLIKELY(!ptr))
1032     return PartitionAllocGeneric(root, new_size, type_name);
1033   if (UNLIKELY(!new_size)) {
1034     PartitionFreeGeneric(root, ptr);
1035     return 0;
1036   }
1037 
1038   if (new_size > kGenericMaxDirectMapped)
1039     PartitionExcessiveAllocationSize();
1040 
1041   DCHECK(PartitionPointerIsValid(PartitionCookieFreePointerAdjust(ptr)));
1042 
1043   PartitionPage* page =
1044       PartitionPointerToPage(PartitionCookieFreePointerAdjust(ptr));
1045 
1046   if (UNLIKELY(PartitionBucketIsDirectMapped(page->bucket))) {
1047     // We may be able to perform the realloc in place by changing the
1048     // accessibility of memory pages and, if reducing the size, decommitting
1049     // them.
1050     if (partitionReallocDirectMappedInPlace(root, page, new_size)) {
1051       PartitionAllocHooks::ReallocHookIfEnabled(ptr, ptr, new_size, type_name);
1052       return ptr;
1053     }
1054   }
1055 
1056   size_t actual_new_size = PartitionAllocActualSize(root, new_size);
1057   size_t actual_old_size = PartitionAllocGetSize(ptr);
1058 
1059   // TODO: note that tcmalloc will "ignore" a downsizing realloc() unless the
1060   // new size is a significant percentage smaller. We could do the same if we
1061   // determine it is a win.
1062   if (actual_new_size == actual_old_size) {
1063     // Trying to allocate a block of size new_size would give us a block of
1064     // the same size as the one we've already got, so re-use the allocation
1065     // after updating statistics (and cookies, if present).
1066     PartitionPageSetRawSize(page, PartitionCookieSizeAdjustAdd(new_size));
1067 #if DCHECK_IS_ON()
1068     // Write a new trailing cookie when it is possible to keep track of
1069     // |new_size| via the raw size pointer.
1070     if (PartitionPageGetRawSizePtr(page))
1071       PartitionCookieWriteValue(static_cast<char*>(ptr) + new_size);
1072 #endif
1073     return ptr;
1074   }
1075 
1076   // This realloc cannot be resized in-place. Sadness.
1077   void* ret = PartitionAllocGeneric(root, new_size, type_name);
1078   size_t copy_size = actual_old_size;
1079   if (new_size < copy_size)
1080     copy_size = new_size;
1081 
1082   memcpy(ret, ptr, copy_size);
1083   PartitionFreeGeneric(root, ptr);
1084   return ret;
1085 #endif
1086 }
1087 
PartitionPurgePage(PartitionPage * page,bool discard)1088 static size_t PartitionPurgePage(PartitionPage* page, bool discard) {
1089   const PartitionBucket* bucket = page->bucket;
1090   size_t slot_size = bucket->slot_size;
1091   if (slot_size < kSystemPageSize || !page->num_allocated_slots)
1092     return 0;
1093 
1094   size_t bucket_num_slots = PartitionBucketSlots(bucket);
1095   size_t discardable_bytes = 0;
1096 
1097   size_t raw_size = PartitionPageGetRawSize(const_cast<PartitionPage*>(page));
1098   if (raw_size) {
1099     uint32_t usedBytes = static_cast<uint32_t>(RoundUpToSystemPage(raw_size));
1100     discardable_bytes = bucket->slot_size - usedBytes;
1101     if (discardable_bytes && discard) {
1102       char* ptr = reinterpret_cast<char*>(PartitionPageToPointer(page));
1103       ptr += usedBytes;
1104       DiscardSystemPages(ptr, discardable_bytes);
1105     }
1106     return discardable_bytes;
1107   }
1108 
1109   const size_t max_slot_count =
1110       (kPartitionPageSize * kMaxPartitionPagesPerSlotSpan) / kSystemPageSize;
1111   DCHECK(bucket_num_slots <= max_slot_count);
1112   DCHECK(page->num_unprovisioned_slots < bucket_num_slots);
1113   size_t num_slots = bucket_num_slots - page->num_unprovisioned_slots;
1114   char slot_usage[max_slot_count];
1115   size_t last_slot = static_cast<size_t>(-1);
1116   memset(slot_usage, 1, num_slots);
1117   char* ptr = reinterpret_cast<char*>(PartitionPageToPointer(page));
1118   PartitionFreelistEntry* entry = page->freelist_head;
1119   // First, walk the freelist for this page and make a bitmap of which slots
1120   // are not in use.
1121   while (entry) {
1122     size_t slotIndex = (reinterpret_cast<char*>(entry) - ptr) / slot_size;
1123     DCHECK(slotIndex < num_slots);
1124     slot_usage[slotIndex] = 0;
1125     entry = PartitionFreelistMask(entry->next);
1126     // If we have a slot where the masked freelist entry is 0, we can
1127     // actually discard that freelist entry because touching a discarded
1128     // page is guaranteed to return original content or 0.
1129     // (Note that this optimization won't fire on big endian machines
1130     // because the masking function is negation.)
1131     if (!PartitionFreelistMask(entry))
1132       last_slot = slotIndex;
1133   }
1134 
1135   // If the slot(s) at the end of the slot span are not in used, we can
1136   // truncate them entirely and rewrite the freelist.
1137   size_t truncated_slots = 0;
1138   while (!slot_usage[num_slots - 1]) {
1139     truncated_slots++;
1140     num_slots--;
1141     DCHECK(num_slots);
1142   }
1143   // First, do the work of calculating the discardable bytes. Don't actually
1144   // discard anything unless the discard flag was passed in.
1145   char* begin_ptr = nullptr;
1146   char* end_ptr = nullptr;
1147   size_t unprovisioned_bytes = 0;
1148   if (truncated_slots) {
1149     begin_ptr = ptr + (num_slots * slot_size);
1150     end_ptr = begin_ptr + (slot_size * truncated_slots);
1151     begin_ptr = reinterpret_cast<char*>(
1152         RoundUpToSystemPage(reinterpret_cast<size_t>(begin_ptr)));
1153     // We round the end pointer here up and not down because we're at the
1154     // end of a slot span, so we "own" all the way up the page boundary.
1155     end_ptr = reinterpret_cast<char*>(
1156         RoundUpToSystemPage(reinterpret_cast<size_t>(end_ptr)));
1157     DCHECK(end_ptr <= ptr + PartitionBucketBytes(bucket));
1158     if (begin_ptr < end_ptr) {
1159       unprovisioned_bytes = end_ptr - begin_ptr;
1160       discardable_bytes += unprovisioned_bytes;
1161     }
1162   }
1163   if (unprovisioned_bytes && discard) {
1164     DCHECK(truncated_slots > 0);
1165     size_t num_new_entries = 0;
1166     page->num_unprovisioned_slots += static_cast<uint16_t>(truncated_slots);
1167     // Rewrite the freelist.
1168     PartitionFreelistEntry** entry_ptr = &page->freelist_head;
1169     for (size_t slotIndex = 0; slotIndex < num_slots; ++slotIndex) {
1170       if (slot_usage[slotIndex])
1171         continue;
1172       PartitionFreelistEntry* entry = reinterpret_cast<PartitionFreelistEntry*>(
1173           ptr + (slot_size * slotIndex));
1174       *entry_ptr = PartitionFreelistMask(entry);
1175       entry_ptr = reinterpret_cast<PartitionFreelistEntry**>(entry);
1176       num_new_entries++;
1177     }
1178     // Terminate the freelist chain.
1179     *entry_ptr = nullptr;
1180     // The freelist head is stored unmasked.
1181     page->freelist_head = PartitionFreelistMask(page->freelist_head);
1182     DCHECK(num_new_entries == num_slots - page->num_allocated_slots);
1183     // Discard the memory.
1184     DiscardSystemPages(begin_ptr, unprovisioned_bytes);
1185   }
1186 
1187   // Next, walk the slots and for any not in use, consider where the system
1188   // page boundaries occur. We can release any system pages back to the
1189   // system as long as we don't interfere with a freelist pointer or an
1190   // adjacent slot.
1191   for (size_t i = 0; i < num_slots; ++i) {
1192     if (slot_usage[i])
1193       continue;
1194     // The first address we can safely discard is just after the freelist
1195     // pointer. There's one quirk: if the freelist pointer is actually a
1196     // null, we can discard that pointer value too.
1197     char* begin_ptr = ptr + (i * slot_size);
1198     char* end_ptr = begin_ptr + slot_size;
1199     if (i != last_slot)
1200       begin_ptr += sizeof(PartitionFreelistEntry);
1201     begin_ptr = reinterpret_cast<char*>(
1202         RoundUpToSystemPage(reinterpret_cast<size_t>(begin_ptr)));
1203     end_ptr = reinterpret_cast<char*>(
1204         RoundDownToSystemPage(reinterpret_cast<size_t>(end_ptr)));
1205     if (begin_ptr < end_ptr) {
1206       size_t partial_slot_bytes = end_ptr - begin_ptr;
1207       discardable_bytes += partial_slot_bytes;
1208       if (discard)
1209         DiscardSystemPages(begin_ptr, partial_slot_bytes);
1210     }
1211   }
1212   return discardable_bytes;
1213 }
1214 
PartitionPurgeBucket(PartitionBucket * bucket)1215 static void PartitionPurgeBucket(PartitionBucket* bucket) {
1216   if (bucket->active_pages_head != &PartitionRootGeneric::gSeedPage) {
1217     for (PartitionPage* page = bucket->active_pages_head; page;
1218          page = page->next_page) {
1219       DCHECK(page != &PartitionRootGeneric::gSeedPage);
1220       (void)PartitionPurgePage(page, true);
1221     }
1222   }
1223 }
1224 
PartitionPurgeMemory(PartitionRoot * root,int flags)1225 void PartitionPurgeMemory(PartitionRoot* root, int flags) {
1226   if (flags & PartitionPurgeDecommitEmptyPages)
1227     PartitionDecommitEmptyPages(root);
1228   // We don't currently do anything for PartitionPurgeDiscardUnusedSystemPages
1229   // here because that flag is only useful for allocations >= system page
1230   // size. We only have allocations that large inside generic partitions
1231   // at the moment.
1232 }
1233 
PartitionPurgeMemoryGeneric(PartitionRootGeneric * root,int flags)1234 void PartitionPurgeMemoryGeneric(PartitionRootGeneric* root, int flags) {
1235   subtle::SpinLock::Guard guard(root->lock);
1236   if (flags & PartitionPurgeDecommitEmptyPages)
1237     PartitionDecommitEmptyPages(root);
1238   if (flags & PartitionPurgeDiscardUnusedSystemPages) {
1239     for (size_t i = 0; i < kGenericNumBuckets; ++i) {
1240       PartitionBucket* bucket = &root->buckets[i];
1241       if (bucket->slot_size >= kSystemPageSize)
1242         PartitionPurgeBucket(bucket);
1243     }
1244   }
1245 }
1246 
PartitionDumpPageStats(PartitionBucketMemoryStats * stats_out,const PartitionPage * page)1247 static void PartitionDumpPageStats(PartitionBucketMemoryStats* stats_out,
1248                                    const PartitionPage* page) {
1249   uint16_t bucket_num_slots = PartitionBucketSlots(page->bucket);
1250 
1251   if (PartitionPageStateIsDecommitted(page)) {
1252     ++stats_out->num_decommitted_pages;
1253     return;
1254   }
1255 
1256   stats_out->discardable_bytes +=
1257       PartitionPurgePage(const_cast<PartitionPage*>(page), false);
1258 
1259   size_t raw_size = PartitionPageGetRawSize(const_cast<PartitionPage*>(page));
1260   if (raw_size)
1261     stats_out->active_bytes += static_cast<uint32_t>(raw_size);
1262   else
1263     stats_out->active_bytes +=
1264         (page->num_allocated_slots * stats_out->bucket_slot_size);
1265 
1266   size_t page_bytes_resident =
1267       RoundUpToSystemPage((bucket_num_slots - page->num_unprovisioned_slots) *
1268                           stats_out->bucket_slot_size);
1269   stats_out->resident_bytes += page_bytes_resident;
1270   if (PartitionPageStateIsEmpty(page)) {
1271     stats_out->decommittable_bytes += page_bytes_resident;
1272     ++stats_out->num_empty_pages;
1273   } else if (PartitionPageStateIsFull(page)) {
1274     ++stats_out->num_full_pages;
1275   } else {
1276     DCHECK(PartitionPageStateIsActive(page));
1277     ++stats_out->num_active_pages;
1278   }
1279 }
1280 
PartitionDumpBucketStats(PartitionBucketMemoryStats * stats_out,const PartitionBucket * bucket)1281 static void PartitionDumpBucketStats(PartitionBucketMemoryStats* stats_out,
1282                                      const PartitionBucket* bucket) {
1283   DCHECK(!PartitionBucketIsDirectMapped(bucket));
1284   stats_out->is_valid = false;
1285   // If the active page list is empty (== &PartitionRootGeneric::gSeedPage),
1286   // the bucket might still need to be reported if it has a list of empty,
1287   // decommitted or full pages.
1288   if (bucket->active_pages_head == &PartitionRootGeneric::gSeedPage &&
1289       !bucket->empty_pages_head && !bucket->decommitted_pages_head &&
1290       !bucket->num_full_pages)
1291     return;
1292 
1293   memset(stats_out, '\0', sizeof(*stats_out));
1294   stats_out->is_valid = true;
1295   stats_out->is_direct_map = false;
1296   stats_out->num_full_pages = static_cast<size_t>(bucket->num_full_pages);
1297   stats_out->bucket_slot_size = bucket->slot_size;
1298   uint16_t bucket_num_slots = PartitionBucketSlots(bucket);
1299   size_t bucket_useful_storage = stats_out->bucket_slot_size * bucket_num_slots;
1300   stats_out->allocated_page_size = PartitionBucketBytes(bucket);
1301   stats_out->active_bytes = bucket->num_full_pages * bucket_useful_storage;
1302   stats_out->resident_bytes =
1303       bucket->num_full_pages * stats_out->allocated_page_size;
1304 
1305   for (const PartitionPage* page = bucket->empty_pages_head; page;
1306        page = page->next_page) {
1307     DCHECK(PartitionPageStateIsEmpty(page) ||
1308            PartitionPageStateIsDecommitted(page));
1309     PartitionDumpPageStats(stats_out, page);
1310   }
1311   for (const PartitionPage* page = bucket->decommitted_pages_head; page;
1312        page = page->next_page) {
1313     DCHECK(PartitionPageStateIsDecommitted(page));
1314     PartitionDumpPageStats(stats_out, page);
1315   }
1316 
1317   if (bucket->active_pages_head != &PartitionRootGeneric::gSeedPage) {
1318     for (const PartitionPage* page = bucket->active_pages_head; page;
1319          page = page->next_page) {
1320       DCHECK(page != &PartitionRootGeneric::gSeedPage);
1321       PartitionDumpPageStats(stats_out, page);
1322     }
1323   }
1324 }
1325 
PartitionDumpStatsGeneric(PartitionRootGeneric * partition,const char * partition_name,bool is_light_dump,PartitionStatsDumper * dumper)1326 void PartitionDumpStatsGeneric(PartitionRootGeneric* partition,
1327                                const char* partition_name,
1328                                bool is_light_dump,
1329                                PartitionStatsDumper* dumper) {
1330   PartitionMemoryStats stats = {0};
1331   stats.total_mmapped_bytes = partition->total_size_of_super_pages +
1332                               partition->total_size_of_direct_mapped_pages;
1333   stats.total_committed_bytes = partition->total_size_of_committed_pages;
1334 
1335   size_t direct_mapped_allocations_total_size = 0;
1336 
1337   static const size_t kMaxReportableDirectMaps = 4096;
1338 
1339   // Allocate on the heap rather than on the stack to avoid stack overflow
1340   // skirmishes (on Windows, in particular).
1341   std::unique_ptr<uint32_t[]> direct_map_lengths = nullptr;
1342   if (!is_light_dump) {
1343     direct_map_lengths =
1344         std::unique_ptr<uint32_t[]>(new uint32_t[kMaxReportableDirectMaps]);
1345   }
1346 
1347   PartitionBucketMemoryStats bucket_stats[kGenericNumBuckets];
1348   size_t num_direct_mapped_allocations = 0;
1349   {
1350     subtle::SpinLock::Guard guard(partition->lock);
1351 
1352     for (size_t i = 0; i < kGenericNumBuckets; ++i) {
1353       const PartitionBucket* bucket = &partition->buckets[i];
1354       // Don't report the pseudo buckets that the generic allocator sets up in
1355       // order to preserve a fast size->bucket map (see
1356       // PartitionAllocGenericInit for details).
1357       if (!bucket->active_pages_head)
1358         bucket_stats[i].is_valid = false;
1359       else
1360         PartitionDumpBucketStats(&bucket_stats[i], bucket);
1361       if (bucket_stats[i].is_valid) {
1362         stats.total_resident_bytes += bucket_stats[i].resident_bytes;
1363         stats.total_active_bytes += bucket_stats[i].active_bytes;
1364         stats.total_decommittable_bytes += bucket_stats[i].decommittable_bytes;
1365         stats.total_discardable_bytes += bucket_stats[i].discardable_bytes;
1366       }
1367     }
1368 
1369     for (PartitionDirectMapExtent *extent = partition->direct_map_list;
1370          extent && num_direct_mapped_allocations < kMaxReportableDirectMaps;
1371          extent = extent->next_extent, ++num_direct_mapped_allocations) {
1372       DCHECK(!extent->next_extent ||
1373              extent->next_extent->prev_extent == extent);
1374       size_t slot_size = extent->bucket->slot_size;
1375       direct_mapped_allocations_total_size += slot_size;
1376       if (is_light_dump)
1377         continue;
1378       direct_map_lengths[num_direct_mapped_allocations] = slot_size;
1379     }
1380   }
1381 
1382   if (!is_light_dump) {
1383     // Call |PartitionsDumpBucketStats| after collecting stats because it can
1384     // try to allocate using |PartitionAllocGeneric| and it can't obtain the
1385     // lock.
1386     for (size_t i = 0; i < kGenericNumBuckets; ++i) {
1387       if (bucket_stats[i].is_valid)
1388         dumper->PartitionsDumpBucketStats(partition_name, &bucket_stats[i]);
1389     }
1390 
1391     for (size_t i = 0; i < num_direct_mapped_allocations; ++i) {
1392       uint32_t size = direct_map_lengths[i];
1393 
1394       PartitionBucketMemoryStats stats;
1395       memset(&stats, '\0', sizeof(stats));
1396       stats.is_valid = true;
1397       stats.is_direct_map = true;
1398       stats.num_full_pages = 1;
1399       stats.allocated_page_size = size;
1400       stats.bucket_slot_size = size;
1401       stats.active_bytes = size;
1402       stats.resident_bytes = size;
1403       dumper->PartitionsDumpBucketStats(partition_name, &stats);
1404     }
1405   }
1406 
1407   stats.total_resident_bytes += direct_mapped_allocations_total_size;
1408   stats.total_active_bytes += direct_mapped_allocations_total_size;
1409   dumper->PartitionDumpTotals(partition_name, &stats);
1410 }
1411 
PartitionDumpStats(PartitionRoot * partition,const char * partition_name,bool is_light_dump,PartitionStatsDumper * dumper)1412 void PartitionDumpStats(PartitionRoot* partition,
1413                         const char* partition_name,
1414                         bool is_light_dump,
1415                         PartitionStatsDumper* dumper) {
1416   static const size_t kMaxReportableBuckets = 4096 / sizeof(void*);
1417   PartitionBucketMemoryStats memory_stats[kMaxReportableBuckets];
1418   const size_t partitionNumBuckets = partition->num_buckets;
1419   DCHECK(partitionNumBuckets <= kMaxReportableBuckets);
1420 
1421   for (size_t i = 0; i < partitionNumBuckets; ++i)
1422     PartitionDumpBucketStats(&memory_stats[i], &partition->buckets()[i]);
1423 
1424   // PartitionsDumpBucketStats is called after collecting stats because it
1425   // can use PartitionAlloc to allocate and this can affect the statistics.
1426   PartitionMemoryStats stats = {0};
1427   stats.total_mmapped_bytes = partition->total_size_of_super_pages;
1428   stats.total_committed_bytes = partition->total_size_of_committed_pages;
1429   DCHECK(!partition->total_size_of_direct_mapped_pages);
1430   for (size_t i = 0; i < partitionNumBuckets; ++i) {
1431     if (memory_stats[i].is_valid) {
1432       stats.total_resident_bytes += memory_stats[i].resident_bytes;
1433       stats.total_active_bytes += memory_stats[i].active_bytes;
1434       stats.total_decommittable_bytes += memory_stats[i].decommittable_bytes;
1435       stats.total_discardable_bytes += memory_stats[i].discardable_bytes;
1436       if (!is_light_dump)
1437         dumper->PartitionsDumpBucketStats(partition_name, &memory_stats[i]);
1438     }
1439   }
1440   dumper->PartitionDumpTotals(partition_name, &stats);
1441 }
1442 
1443 }  // namespace base
1444 }  // namespace pdfium
1445