• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2013 Google Inc. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions are
6  * met:
7  *
8  *     * Redistributions of source code must retain the above copyright
9  * notice, this list of conditions and the following disclaimer.
10  *     * Redistributions in binary form must reproduce the above
11  * copyright notice, this list of conditions and the following disclaimer
12  * in the documentation and/or other materials provided with the
13  * distribution.
14  *     * Neither the name of Google Inc. nor the names of its
15  * contributors may be used to endorse or promote products derived from
16  * this software without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 
31 #include "config.h"
32 #include "wtf/PartitionAlloc.h"
33 
34 #include <string.h>
35 
36 #ifndef NDEBUG
37 #include <stdio.h>
38 #endif
39 
40 // Two partition pages are used as guard / metadata page so make sure the super
41 // page size is bigger.
42 COMPILE_ASSERT(WTF::kPartitionPageSize * 4 <= WTF::kSuperPageSize, ok_super_page_size);
43 COMPILE_ASSERT(!(WTF::kSuperPageSize % WTF::kPartitionPageSize), ok_super_page_multiple);
44 // Four system pages gives us room to hack out a still-guard-paged piece
45 // of metadata in the middle of a guard partition page.
46 COMPILE_ASSERT(WTF::kSystemPageSize * 4 <= WTF::kPartitionPageSize, ok_partition_page_size);
47 COMPILE_ASSERT(!(WTF::kPartitionPageSize % WTF::kSystemPageSize), ok_partition_page_multiple);
48 COMPILE_ASSERT(sizeof(WTF::PartitionPage) <= WTF::kPageMetadataSize, PartitionPage_not_too_big);
49 COMPILE_ASSERT(sizeof(WTF::PartitionBucket) <= WTF::kPageMetadataSize, PartitionBucket_not_too_big);
50 COMPILE_ASSERT(sizeof(WTF::PartitionSuperPageExtentEntry) <= WTF::kPageMetadataSize, PartitionSuperPageExtentEntry_not_too_big);
51 COMPILE_ASSERT(WTF::kPageMetadataSize * WTF::kNumPartitionPagesPerSuperPage <= WTF::kSystemPageSize, page_metadata_fits_in_hole);
52 // Check that some of our zanier calculations worked out as expected.
53 COMPILE_ASSERT(WTF::kGenericSmallestBucket == 8, generic_smallest_bucket);
54 COMPILE_ASSERT(WTF::kGenericMaxBucketed == 983040, generic_max_bucketed);
55 
56 namespace WTF {
57 
58 int PartitionRootBase::gInitializedLock = 0;
59 bool PartitionRootBase::gInitialized = false;
60 PartitionPage PartitionRootBase::gSeedPage;
61 PartitionBucket PartitionRootBase::gPagedBucket;
62 
partitionBucketNumSystemPages(size_t size)63 static size_t partitionBucketNumSystemPages(size_t size)
64 {
65     // This works out reasonably for the current bucket sizes of the generic
66     // allocator, and the current values of partition page size and constants.
67     // Specifically, we have enough room to always pack the slots perfectly into
68     // some number of system pages. The only waste is the waste associated with
69     // unfaulted pages (i.e. wasted address space).
70     // TODO: we end up using a lot of system pages for very small sizes. For
71     // example, we'll use 12 system pages for slot size 24. The slot size is
72     // so small that the waste would be tiny with just 4, or 1, system pages.
73     // Later, we can investigate whether there are anti-fragmentation benefits
74     // to using fewer system pages.
75     double bestWasteRatio = 1.0f;
76     size_t bestPages = 0;
77     if (size > kMaxSystemPagesPerSlotSpan * kSystemPageSize) {
78         ASSERT(!(size % kSystemPageSize));
79         return size / kSystemPageSize;
80     }
81     ASSERT(size <= kMaxSystemPagesPerSlotSpan * kSystemPageSize);
82     for (size_t i = kNumSystemPagesPerPartitionPage - 1; i <= kMaxSystemPagesPerSlotSpan; ++i) {
83         size_t pageSize = kSystemPageSize * i;
84         size_t numSlots = pageSize / size;
85         size_t waste = pageSize - (numSlots * size);
86         // Leaving a page unfaulted is not free; the page will occupy an empty page table entry.
87         // Make a simple attempt to account for that.
88         size_t numRemainderPages = i & (kNumSystemPagesPerPartitionPage - 1);
89         size_t numUnfaultedPages = numRemainderPages ? (kNumSystemPagesPerPartitionPage - numRemainderPages) : 0;
90         waste += sizeof(void*) * numUnfaultedPages;
91         double wasteRatio = (double) waste / (double) pageSize;
92         if (wasteRatio < bestWasteRatio) {
93             bestWasteRatio = wasteRatio;
94             bestPages = i;
95         }
96     }
97     ASSERT(bestPages > 0);
98     return bestPages;
99 }
100 
parititonAllocBaseInit(PartitionRootBase * root)101 static void parititonAllocBaseInit(PartitionRootBase* root)
102 {
103     ASSERT(!root->initialized);
104 
105     spinLockLock(&PartitionRootBase::gInitializedLock);
106     if (!PartitionRootBase::gInitialized) {
107         PartitionRootBase::gInitialized = true;
108         // We mark the seed page as free to make sure it is skipped by our
109         // logic to find a new active page.
110         PartitionRootBase::gPagedBucket.activePagesHead = &PartitionRootGeneric::gSeedPage;
111     }
112     spinLockUnlock(&PartitionRootBase::gInitializedLock);
113 
114     root->initialized = true;
115     root->totalSizeOfCommittedPages = 0;
116     root->totalSizeOfSuperPages = 0;
117     root->nextSuperPage = 0;
118     root->nextPartitionPage = 0;
119     root->nextPartitionPageEnd = 0;
120     root->firstExtent = 0;
121     root->currentExtent = 0;
122 
123     memset(&root->globalEmptyPageRing, '\0', sizeof(root->globalEmptyPageRing));
124     root->globalEmptyPageRingIndex = 0;
125 
126     // This is a "magic" value so we can test if a root pointer is valid.
127     root->invertedSelf = ~reinterpret_cast<uintptr_t>(root);
128 }
129 
partitionBucketInitBase(PartitionBucket * bucket,PartitionRootBase * root)130 static void partitionBucketInitBase(PartitionBucket* bucket, PartitionRootBase* root)
131 {
132     bucket->activePagesHead = &PartitionRootGeneric::gSeedPage;
133     bucket->freePagesHead = 0;
134     bucket->numFullPages = 0;
135     bucket->numSystemPagesPerSlotSpan = partitionBucketNumSystemPages(bucket->slotSize);
136 }
137 
partitionAllocInit(PartitionRoot * root,size_t numBuckets,size_t maxAllocation)138 void partitionAllocInit(PartitionRoot* root, size_t numBuckets, size_t maxAllocation)
139 {
140     parititonAllocBaseInit(root);
141 
142     root->numBuckets = numBuckets;
143     root->maxAllocation = maxAllocation;
144     size_t i;
145     for (i = 0; i < root->numBuckets; ++i) {
146         PartitionBucket* bucket = &root->buckets()[i];
147         if (!i)
148             bucket->slotSize = kAllocationGranularity;
149         else
150             bucket->slotSize = i << kBucketShift;
151         partitionBucketInitBase(bucket, root);
152     }
153 }
154 
partitionAllocGenericInit(PartitionRootGeneric * root)155 void partitionAllocGenericInit(PartitionRootGeneric* root)
156 {
157     parititonAllocBaseInit(root);
158 
159     root->lock = 0;
160 
161     // Precalculate some shift and mask constants used in the hot path.
162     // Example: malloc(41) == 101001 binary.
163     // Order is 6 (1 << 6-1)==32 is highest bit set.
164     // orderIndex is the next three MSB == 010 == 2.
165     // subOrderIndexMask is a mask for the remaining bits == 11 (masking to 01 for the subOrderIndex).
166     size_t order;
167     for (order = 0; order <= kBitsPerSizet; ++order) {
168         size_t orderIndexShift;
169         if (order < kGenericNumBucketsPerOrderBits + 1)
170             orderIndexShift = 0;
171         else
172             orderIndexShift = order - (kGenericNumBucketsPerOrderBits + 1);
173         root->orderIndexShifts[order] = orderIndexShift;
174         size_t subOrderIndexMask;
175         if (order == kBitsPerSizet) {
176             // This avoids invoking undefined behavior for an excessive shift.
177             subOrderIndexMask = static_cast<size_t>(-1) >> (kGenericNumBucketsPerOrderBits + 1);
178         } else {
179             subOrderIndexMask = ((1 << order) - 1) >> (kGenericNumBucketsPerOrderBits + 1);
180         }
181         root->orderSubIndexMasks[order] = subOrderIndexMask;
182     }
183 
184     // Set up the actual usable buckets first.
185     // Note that typical values (i.e. min allocation size of 8) will result in
186     // invalid buckets (size==9 etc. or more generally, size is not a multiple
187     // of the smallest allocation granularity).
188     // We avoid them in the bucket lookup map, but we tolerate them to keep the
189     // code simpler and the structures more generic.
190     size_t i, j;
191     size_t currentSize = kGenericSmallestBucket;
192     size_t currentIncrement = kGenericSmallestBucket >> kGenericNumBucketsPerOrderBits;
193     PartitionBucket* bucket = &root->buckets[0];
194     for (i = 0; i < kGenericNumBucketedOrders; ++i) {
195         for (j = 0; j < kGenericNumBucketsPerOrder; ++j) {
196             bucket->slotSize = currentSize;
197             partitionBucketInitBase(bucket, root);
198             // Disable invalid buckets so that touching them faults.
199             if (currentSize % kGenericSmallestBucket)
200                 bucket->activePagesHead = 0;
201             currentSize += currentIncrement;
202             ++bucket;
203         }
204         currentIncrement <<= 1;
205     }
206     ASSERT(currentSize == 1 << kGenericMaxBucketedOrder);
207     ASSERT(bucket == &root->buckets[0] + (kGenericNumBucketedOrders * kGenericNumBucketsPerOrder));
208 
209     // Then set up the fast size -> bucket lookup table.
210     bucket = &root->buckets[0];
211     PartitionBucket** bucketPtr = &root->bucketLookups[0];
212     for (order = 0; order <= kBitsPerSizet; ++order) {
213         for (j = 0; j < kGenericNumBucketsPerOrder; ++j) {
214             if (order < kGenericMinBucketedOrder) {
215                 // Use the bucket of finest granularity for malloc(0) etc.
216                 *bucketPtr++ = &root->buckets[0];
217             } else if (order > kGenericMaxBucketedOrder) {
218                 *bucketPtr++ = &PartitionRootGeneric::gPagedBucket;
219             } else {
220                 PartitionBucket* validBucket = bucket;
221                 // Skip over invalid buckets.
222                 while (validBucket->slotSize % kGenericSmallestBucket)
223                     validBucket++;
224                 *bucketPtr++ = validBucket;
225                 bucket++;
226             }
227         }
228     }
229     ASSERT(bucket == &root->buckets[0] + (kGenericNumBucketedOrders * kGenericNumBucketsPerOrder));
230     ASSERT(bucketPtr == &root->bucketLookups[0] + ((kBitsPerSizet + 1) * kGenericNumBucketsPerOrder));
231     // And there's one last bucket lookup that will be hit for e.g. malloc(-1),
232     // which tries to overflow to a non-existant order.
233     *bucketPtr = &PartitionRootGeneric::gPagedBucket;
234 }
235 
partitionAllocShutdownBucket(PartitionBucket * bucket)236 static bool partitionAllocShutdownBucket(PartitionBucket* bucket)
237 {
238     // Failure here indicates a memory leak.
239     bool noLeaks = !bucket->numFullPages;
240     PartitionPage* page = bucket->activePagesHead;
241     while (page) {
242         if (page->numAllocatedSlots)
243             noLeaks = false;
244         page = page->nextPage;
245     }
246 
247     return noLeaks;
248 }
249 
partitionAllocBaseShutdown(PartitionRootBase * root)250 static void partitionAllocBaseShutdown(PartitionRootBase* root)
251 {
252     ASSERT(root->initialized);
253     root->initialized = false;
254 
255     // Now that we've examined all partition pages in all buckets, it's safe
256     // to free all our super pages. We first collect the super page pointers
257     // on the stack because some of them are themselves store in super pages.
258     char* superPages[kMaxPartitionSize / kSuperPageSize];
259     size_t numSuperPages = 0;
260     PartitionSuperPageExtentEntry* entry = root->firstExtent;
261     while (entry) {
262         char* superPage = entry->superPageBase;
263         while (superPage != entry->superPagesEnd) {
264             superPages[numSuperPages] = superPage;
265             numSuperPages++;
266             superPage += kSuperPageSize;
267         }
268         entry = entry->next;
269     }
270     ASSERT(numSuperPages == root->totalSizeOfSuperPages / kSuperPageSize);
271     for (size_t i = 0; i < numSuperPages; ++i)
272         freePages(superPages[i], kSuperPageSize);
273 }
274 
partitionAllocShutdown(PartitionRoot * root)275 bool partitionAllocShutdown(PartitionRoot* root)
276 {
277     bool noLeaks = true;
278     size_t i;
279     for (i = 0; i < root->numBuckets; ++i) {
280         PartitionBucket* bucket = &root->buckets()[i];
281         if (!partitionAllocShutdownBucket(bucket))
282             noLeaks = false;
283     }
284 
285     partitionAllocBaseShutdown(root);
286     return noLeaks;
287 }
288 
partitionAllocGenericShutdown(PartitionRootGeneric * root)289 bool partitionAllocGenericShutdown(PartitionRootGeneric* root)
290 {
291     bool noLeaks = true;
292     size_t i;
293     for (i = 0; i < kGenericNumBucketedOrders * kGenericNumBucketsPerOrder; ++i) {
294         PartitionBucket* bucket = &root->buckets[i];
295         if (!partitionAllocShutdownBucket(bucket))
296             noLeaks = false;
297     }
298     partitionAllocBaseShutdown(root);
299     return noLeaks;
300 }
301 
partitionOutOfMemory()302 static NEVER_INLINE void partitionOutOfMemory()
303 {
304     IMMEDIATE_CRASH();
305 }
306 
partitionFull()307 static NEVER_INLINE void partitionFull()
308 {
309     IMMEDIATE_CRASH();
310 }
311 
partitionDecommitSystemPages(PartitionRootBase * root,void * addr,size_t len)312 static ALWAYS_INLINE void partitionDecommitSystemPages(PartitionRootBase* root, void* addr, size_t len)
313 {
314     decommitSystemPages(addr, len);
315     ASSERT(root->totalSizeOfCommittedPages > len);
316     root->totalSizeOfCommittedPages -= len;
317 }
318 
partitionRecommitSystemPages(PartitionRootBase * root,void * addr,size_t len)319 static ALWAYS_INLINE void partitionRecommitSystemPages(PartitionRootBase* root, void* addr, size_t len)
320 {
321     recommitSystemPages(addr, len);
322     root->totalSizeOfCommittedPages += len;
323 }
324 
partitionAllocPartitionPages(PartitionRootBase * root,int flags,size_t numPartitionPages)325 static ALWAYS_INLINE void* partitionAllocPartitionPages(PartitionRootBase* root, int flags, size_t numPartitionPages)
326 {
327     ASSERT(!(reinterpret_cast<uintptr_t>(root->nextPartitionPage) % kPartitionPageSize));
328     ASSERT(!(reinterpret_cast<uintptr_t>(root->nextPartitionPageEnd) % kPartitionPageSize));
329     RELEASE_ASSERT(numPartitionPages <= kNumPartitionPagesPerSuperPage);
330     size_t totalSize = kPartitionPageSize * numPartitionPages;
331     root->totalSizeOfCommittedPages += totalSize;
332     size_t numPartitionPagesLeft = (root->nextPartitionPageEnd - root->nextPartitionPage) >> kPartitionPageShift;
333     if (LIKELY(numPartitionPagesLeft >= numPartitionPages)) {
334         // In this case, we can still hand out pages from the current super page
335         // allocation.
336         char* ret = root->nextPartitionPage;
337         root->nextPartitionPage += totalSize;
338         return ret;
339     }
340 
341     // Need a new super page.
342     root->totalSizeOfSuperPages += kSuperPageSize;
343     if (root->totalSizeOfSuperPages > kMaxPartitionSize)
344         partitionFull();
345     char* requestedAddress = root->nextSuperPage;
346     char* superPage = reinterpret_cast<char*>(allocPages(requestedAddress, kSuperPageSize, kSuperPageSize));
347     if (UNLIKELY(!superPage)) {
348         if (flags & PartitionAllocReturnNull)
349             return 0;
350         partitionOutOfMemory();
351     }
352     root->nextSuperPage = superPage + kSuperPageSize;
353     char* ret = superPage + kPartitionPageSize;
354     root->nextPartitionPage = ret + totalSize;
355     root->nextPartitionPageEnd = root->nextSuperPage - kPartitionPageSize;
356     // Make the first partition page in the super page a guard page, but leave a
357     // hole in the middle.
358     // This is where we put page metadata and also a tiny amount of extent
359     // metadata.
360     setSystemPagesInaccessible(superPage, kSystemPageSize);
361     setSystemPagesInaccessible(superPage + (kSystemPageSize * 2), kPartitionPageSize - (kSystemPageSize * 2));
362     // Also make the last partition page a guard page.
363     setSystemPagesInaccessible(superPage + (kSuperPageSize - kPartitionPageSize), kPartitionPageSize);
364 
365     // If we were after a specific address, but didn't get it, assume that
366     // the system chose a lousy address and re-randomize the next
367     // allocation.
368     if (requestedAddress && requestedAddress != superPage)
369         root->nextSuperPage = 0;
370 
371     // We allocated a new super page so update super page metadata.
372     // First check if this is a new extent or not.
373     PartitionSuperPageExtentEntry* latestExtent = reinterpret_cast<PartitionSuperPageExtentEntry*>(partitionSuperPageToMetadataArea(superPage));
374     PartitionSuperPageExtentEntry* currentExtent = root->currentExtent;
375     bool isNewExtent = (superPage != requestedAddress);
376     if (UNLIKELY(isNewExtent)) {
377         latestExtent->next = 0;
378         if (UNLIKELY(!currentExtent)) {
379             root->firstExtent = latestExtent;
380         } else {
381             ASSERT(currentExtent->superPageBase);
382             currentExtent->next = latestExtent;
383         }
384         root->currentExtent = latestExtent;
385         currentExtent = latestExtent;
386         currentExtent->superPageBase = superPage;
387         currentExtent->superPagesEnd = superPage + kSuperPageSize;
388     } else {
389         // We allocated next to an existing extent so just nudge the size up a little.
390         currentExtent->superPagesEnd += kSuperPageSize;
391         ASSERT(ret >= currentExtent->superPageBase && ret < currentExtent->superPagesEnd);
392     }
393     // By storing the root in every extent metadata object, we have a fast way
394     // to go from a pointer within the partition to the root object.
395     latestExtent->root = root;
396 
397     return ret;
398 }
399 
partitionUnusePage(PartitionRootBase * root,PartitionPage * page)400 static ALWAYS_INLINE void partitionUnusePage(PartitionRootBase* root, PartitionPage* page)
401 {
402     ASSERT(page->bucket->numSystemPagesPerSlotSpan);
403     void* addr = partitionPageToPointer(page);
404     partitionDecommitSystemPages(root, addr, page->bucket->numSystemPagesPerSlotSpan * kSystemPageSize);
405 }
406 
partitionBucketSlots(const PartitionBucket * bucket)407 static ALWAYS_INLINE size_t partitionBucketSlots(const PartitionBucket* bucket)
408 {
409     return (bucket->numSystemPagesPerSlotSpan * kSystemPageSize) / bucket->slotSize;
410 }
411 
partitionBucketPartitionPages(const PartitionBucket * bucket)412 static ALWAYS_INLINE size_t partitionBucketPartitionPages(const PartitionBucket* bucket)
413 {
414     return (bucket->numSystemPagesPerSlotSpan + (kNumSystemPagesPerPartitionPage - 1)) / kNumSystemPagesPerPartitionPage;
415 }
416 
partitionPageReset(PartitionPage * page,PartitionBucket * bucket)417 static ALWAYS_INLINE void partitionPageReset(PartitionPage* page, PartitionBucket* bucket)
418 {
419     ASSERT(page != &PartitionRootGeneric::gSeedPage);
420     page->numAllocatedSlots = 0;
421     page->numUnprovisionedSlots = partitionBucketSlots(bucket);
422     ASSERT(page->numUnprovisionedSlots);
423     page->bucket = bucket;
424     page->nextPage = 0;
425     // NULLing the freelist is not strictly necessary but it makes an ASSERT in partitionPageFillFreelist simpler.
426     page->freelistHead = 0;
427     page->pageOffset = 0;
428     page->freeCacheIndex = -1;
429     size_t numPartitionPages = partitionBucketPartitionPages(bucket);
430     size_t i;
431     char* pageCharPtr = reinterpret_cast<char*>(page);
432     for (i = 1; i < numPartitionPages; ++i) {
433         pageCharPtr += kPageMetadataSize;
434         PartitionPage* secondaryPage = reinterpret_cast<PartitionPage*>(pageCharPtr);
435         secondaryPage->pageOffset = i;
436     }
437 }
438 
partitionPageAllocAndFillFreelist(PartitionPage * page)439 static ALWAYS_INLINE char* partitionPageAllocAndFillFreelist(PartitionPage* page)
440 {
441     ASSERT(page != &PartitionRootGeneric::gSeedPage);
442     size_t numSlots = page->numUnprovisionedSlots;
443     ASSERT(numSlots);
444     PartitionBucket* bucket = page->bucket;
445     // We should only get here when _every_ slot is either used or unprovisioned.
446     // (The third state is "on the freelist". If we have a non-empty freelist, we should not get here.)
447     ASSERT(numSlots + page->numAllocatedSlots == partitionBucketSlots(bucket));
448     // Similarly, make explicitly sure that the freelist is empty.
449     ASSERT(!page->freelistHead);
450     ASSERT(page->numAllocatedSlots >= 0);
451 
452     size_t size = bucket->slotSize;
453     char* base = reinterpret_cast<char*>(partitionPageToPointer(page));
454     char* returnObject = base + (size * page->numAllocatedSlots);
455     char* firstFreelistPointer = returnObject + size;
456     char* firstFreelistPointerExtent = firstFreelistPointer + sizeof(PartitionFreelistEntry*);
457     // Our goal is to fault as few system pages as possible. We calculate the
458     // page containing the "end" of the returned slot, and then allow freelist
459     // pointers to be written up to the end of that page.
460     char* subPageLimit = reinterpret_cast<char*>((reinterpret_cast<uintptr_t>(firstFreelistPointer) + kSystemPageOffsetMask) & kSystemPageBaseMask);
461     char* slotsLimit = returnObject + (size * page->numUnprovisionedSlots);
462     char* freelistLimit = subPageLimit;
463     if (UNLIKELY(slotsLimit < freelistLimit))
464         freelistLimit = slotsLimit;
465 
466     size_t numNewFreelistEntries = 0;
467     if (LIKELY(firstFreelistPointerExtent <= freelistLimit)) {
468         // Only consider used space in the slot span. If we consider wasted
469         // space, we may get an off-by-one when a freelist pointer fits in the
470         // wasted space, but a slot does not.
471         // We know we can fit at least one freelist pointer.
472         numNewFreelistEntries = 1;
473         // Any further entries require space for the whole slot span.
474         numNewFreelistEntries += (freelistLimit - firstFreelistPointerExtent) / size;
475     }
476 
477     // We always return an object slot -- that's the +1 below.
478     // We do not neccessarily create any new freelist entries, because we cross sub page boundaries frequently for large bucket sizes.
479     ASSERT(numNewFreelistEntries + 1 <= numSlots);
480     numSlots -= (numNewFreelistEntries + 1);
481     page->numUnprovisionedSlots = numSlots;
482     page->numAllocatedSlots++;
483 
484     if (LIKELY(numNewFreelistEntries)) {
485         char* freelistPointer = firstFreelistPointer;
486         PartitionFreelistEntry* entry = reinterpret_cast<PartitionFreelistEntry*>(freelistPointer);
487         page->freelistHead = entry;
488         while (--numNewFreelistEntries) {
489             freelistPointer += size;
490             PartitionFreelistEntry* nextEntry = reinterpret_cast<PartitionFreelistEntry*>(freelistPointer);
491             entry->next = partitionFreelistMask(nextEntry);
492             entry = nextEntry;
493         }
494         entry->next = partitionFreelistMask(0);
495     } else {
496         page->freelistHead = 0;
497     }
498     return returnObject;
499 }
500 
501 // This helper function scans the active page list for a suitable new active
502 // page, starting at the passed in page.
503 // When it finds a suitable new active page (one that has free slots), it is
504 // set as the new active page and true is returned. If there is no suitable new
505 // active page, false is returned and the current active page is set to null.
506 // As potential pages are scanned, they are tidied up according to their state.
507 // Freed pages are swept on to the free page list and full pages are unlinked
508 // from any list.
partitionSetNewActivePage(PartitionPage * page)509 static ALWAYS_INLINE bool partitionSetNewActivePage(PartitionPage* page)
510 {
511     if (page == &PartitionRootBase::gSeedPage) {
512         ASSERT(!page->nextPage);
513         return false;
514     }
515 
516     PartitionPage* nextPage = 0;
517     PartitionBucket* bucket = page->bucket;
518 
519     for (; page; page = nextPage) {
520         nextPage = page->nextPage;
521         ASSERT(page->bucket == bucket);
522         ASSERT(page != bucket->freePagesHead);
523         ASSERT(!bucket->freePagesHead || page != bucket->freePagesHead->nextPage);
524 
525         // Page is usable if it has something on the freelist, or unprovisioned
526         // slots that can be turned into a freelist.
527         if (LIKELY(page->freelistHead != 0) || LIKELY(page->numUnprovisionedSlots)) {
528             bucket->activePagesHead = page;
529             return true;
530         }
531 
532         ASSERT(page->numAllocatedSlots >= 0);
533         if (LIKELY(page->numAllocatedSlots == 0)) {
534             ASSERT(page->freeCacheIndex == -1);
535             // We hit a free page, so shepherd it on to the free page list.
536             page->nextPage = bucket->freePagesHead;
537             bucket->freePagesHead = page;
538         } else {
539             // If we get here, we found a full page. Skip over it too, and also
540             // tag it as full (via a negative value). We need it tagged so that
541             // free'ing can tell, and move it back into the active page list.
542             ASSERT(page->numAllocatedSlots == static_cast<int>(partitionBucketSlots(bucket)));
543             page->numAllocatedSlots = -page->numAllocatedSlots;
544             ++bucket->numFullPages;
545             // numFullPages is a uint16_t for efficient packing so guard against
546             // overflow to be safe.
547             RELEASE_ASSERT(bucket->numFullPages);
548             // Not necessary but might help stop accidents.
549             page->nextPage = 0;
550         }
551     }
552 
553     bucket->activePagesHead = 0;
554     return false;
555 }
556 
557 struct PartitionDirectMapExtent {
558     size_t mapSize; // Mapped size, not including guard pages and meta-data.
559 };
560 
partitionPageToDirectMapExtent(PartitionPage * page)561 static ALWAYS_INLINE PartitionDirectMapExtent* partitionPageToDirectMapExtent(PartitionPage* page)
562 {
563     ASSERT(partitionBucketIsDirectMapped(page->bucket));
564     return reinterpret_cast<PartitionDirectMapExtent*>(reinterpret_cast<char*>(page) + 2 * kPageMetadataSize);
565 }
566 
partitionDirectMap(PartitionRootBase * root,int flags,size_t size)567 static ALWAYS_INLINE void* partitionDirectMap(PartitionRootBase* root, int flags, size_t size)
568 {
569     size = partitionDirectMapSize(size);
570 
571     // Because we need to fake looking like a super page, We need to allocate
572     // a bunch of system pages more than "size":
573     // - The first few system pages are the partition page in which the super
574     // page metadata is stored. We fault just one system page out of a partition
575     // page sized clump.
576     // - We add a trailing guard page.
577     size_t mapSize = size + kPartitionPageSize + kSystemPageSize;
578     // Round up to the allocation granularity.
579     mapSize += kPageAllocationGranularityOffsetMask;
580     mapSize &= kPageAllocationGranularityBaseMask;
581 
582     // TODO: we may want to let the operating system place these allocations
583     // where it pleases. On 32-bit, this might limit address space
584     // fragmentation and on 64-bit, this might have useful savings for TLB
585     // and page table overhead.
586     // TODO: if upsizing realloc()s are common on large sizes, we could
587     // consider over-allocating address space on 64-bit, "just in case".
588     // TODO: consider pre-populating page tables (e.g. MAP_POPULATE on Linux,
589     // MADV_WILLNEED on POSIX).
590     // TODO: these pages will be zero-filled. Consider internalizing an
591     // allocZeroed() API so we can avoid a memset() entirely in this case.
592     char* ptr = reinterpret_cast<char*>(allocPages(0, mapSize, kSuperPageSize));
593     if (!ptr) {
594         if (flags & PartitionAllocReturnNull)
595             return 0;
596         partitionOutOfMemory();
597     }
598     char* ret = ptr + kPartitionPageSize;
599     // TODO: due to all the guard paging, this arrangement creates 4 mappings.
600     // We could get it down to three by using read-only for the metadata page,
601     // or perhaps two by leaving out the trailing guard page on 64-bit.
602     setSystemPagesInaccessible(ptr, kSystemPageSize);
603     setSystemPagesInaccessible(ptr + (kSystemPageSize * 2), kPartitionPageSize - (kSystemPageSize * 2));
604     setSystemPagesInaccessible(ret + size, kSystemPageSize);
605 
606     PartitionSuperPageExtentEntry* extent = reinterpret_cast<PartitionSuperPageExtentEntry*>(partitionSuperPageToMetadataArea(ptr));
607     extent->root = root;
608     PartitionPage* page = partitionPointerToPageNoAlignmentCheck(ret);
609     PartitionBucket* bucket = reinterpret_cast<PartitionBucket*>(reinterpret_cast<char*>(page) + kPageMetadataSize);
610     page->freelistHead = 0;
611     page->nextPage = 0;
612     page->bucket = bucket;
613     page->numAllocatedSlots = 1;
614     page->numUnprovisionedSlots = 0;
615     page->pageOffset = 0;
616     page->freeCacheIndex = 0;
617 
618     bucket->activePagesHead = 0;
619     bucket->freePagesHead = 0;
620     bucket->slotSize = size;
621     bucket->numSystemPagesPerSlotSpan = 0;
622     bucket->numFullPages = 0;
623 
624     PartitionDirectMapExtent* mapExtent = partitionPageToDirectMapExtent(page);
625     mapExtent->mapSize = mapSize - kPartitionPageSize - kSystemPageSize;
626 
627     return ret;
628 }
629 
partitionDirectUnmap(PartitionPage * page)630 static ALWAYS_INLINE void partitionDirectUnmap(PartitionPage* page)
631 {
632     size_t unmapSize = partitionPageToDirectMapExtent(page)->mapSize;
633 
634     // Add on the size of the trailing guard page and preceeding partition
635     // page.
636     unmapSize += kPartitionPageSize + kSystemPageSize;
637 
638     ASSERT(!(unmapSize & kPageAllocationGranularityOffsetMask));
639 
640     char* ptr = reinterpret_cast<char*>(partitionPageToPointer(page));
641     // Account for the mapping starting a partition page before the actual
642     // allocation address.
643     ptr -= kPartitionPageSize;
644 
645     freePages(ptr, unmapSize);
646 }
647 
partitionAllocSlowPath(PartitionRootBase * root,int flags,size_t size,PartitionBucket * bucket)648 void* partitionAllocSlowPath(PartitionRootBase* root, int flags, size_t size, PartitionBucket* bucket)
649 {
650     // The slow path is called when the freelist is empty.
651     ASSERT(!bucket->activePagesHead->freelistHead);
652 
653     // For the partitionAllocGeneric API, we have a bunch of buckets marked
654     // as special cases. We bounce them through to the slow path so that we
655     // can still have a blazing fast hot path due to lack of corner-case
656     // branches.
657     bool returnNull = flags & PartitionAllocReturnNull;
658     if (UNLIKELY(partitionBucketIsDirectMapped(bucket))) {
659         ASSERT(size > kGenericMaxBucketed);
660         ASSERT(bucket == &PartitionRootBase::gPagedBucket);
661         if (size > kGenericMaxDirectMapped) {
662             if (returnNull)
663                 return 0;
664             RELEASE_ASSERT(false);
665         }
666         return partitionDirectMap(root, flags, size);
667     }
668 
669     // First, look for a usable page in the existing active pages list.
670     // Change active page, accepting the current page as a candidate.
671     if (LIKELY(partitionSetNewActivePage(bucket->activePagesHead))) {
672         PartitionPage* newPage = bucket->activePagesHead;
673         if (LIKELY(newPage->freelistHead != 0)) {
674             PartitionFreelistEntry* ret = newPage->freelistHead;
675             newPage->freelistHead = partitionFreelistMask(ret->next);
676             newPage->numAllocatedSlots++;
677             return ret;
678         }
679         ASSERT(newPage->numUnprovisionedSlots);
680         return partitionPageAllocAndFillFreelist(newPage);
681     }
682 
683     // Second, look in our list of freed but reserved pages.
684     PartitionPage* newPage = bucket->freePagesHead;
685     if (LIKELY(newPage != 0)) {
686         ASSERT(newPage != &PartitionRootGeneric::gSeedPage);
687         ASSERT(!newPage->freelistHead);
688         ASSERT(!newPage->numAllocatedSlots);
689         ASSERT(!newPage->numUnprovisionedSlots);
690         ASSERT(newPage->freeCacheIndex == -1);
691         bucket->freePagesHead = newPage->nextPage;
692         void* addr = partitionPageToPointer(newPage);
693         partitionRecommitSystemPages(root, addr, newPage->bucket->numSystemPagesPerSlotSpan * kSystemPageSize);
694     } else {
695         // Third. If we get here, we need a brand new page.
696         size_t numPartitionPages = partitionBucketPartitionPages(bucket);
697         void* rawNewPage = partitionAllocPartitionPages(root, flags, numPartitionPages);
698         if (UNLIKELY(!rawNewPage)) {
699             ASSERT(returnNull);
700             return 0;
701         }
702         // Skip the alignment check because it depends on page->bucket, which is not yet set.
703         newPage = partitionPointerToPageNoAlignmentCheck(rawNewPage);
704     }
705 
706     partitionPageReset(newPage, bucket);
707     bucket->activePagesHead = newPage;
708     return partitionPageAllocAndFillFreelist(newPage);
709 }
710 
partitionFreePage(PartitionRootBase * root,PartitionPage * page)711 static ALWAYS_INLINE void partitionFreePage(PartitionRootBase* root, PartitionPage* page)
712 {
713     ASSERT(page->freelistHead);
714     ASSERT(!page->numAllocatedSlots);
715     partitionUnusePage(root, page);
716     // We actually leave the freed page in the active list. We'll sweep it on
717     // to the free page list when we next walk the active page list. Pulling
718     // this trick enables us to use a singly-linked page list for all cases,
719     // which is critical in keeping the page metadata structure down to 32
720     // bytes in size.
721     page->freelistHead = 0;
722     page->numUnprovisionedSlots = 0;
723 }
724 
partitionRegisterEmptyPage(PartitionPage * page)725 static ALWAYS_INLINE void partitionRegisterEmptyPage(PartitionPage* page)
726 {
727     PartitionRootBase* root = partitionPageToRoot(page);
728 
729     // If the page is already registered as empty, give it another life.
730     if (page->freeCacheIndex != -1) {
731         ASSERT(page->freeCacheIndex >= 0);
732         ASSERT(static_cast<unsigned>(page->freeCacheIndex) < kMaxFreeableSpans);
733         ASSERT(root->globalEmptyPageRing[page->freeCacheIndex] == page);
734         root->globalEmptyPageRing[page->freeCacheIndex] = 0;
735     }
736 
737     size_t currentIndex = root->globalEmptyPageRingIndex;
738     PartitionPage* pageToFree = root->globalEmptyPageRing[currentIndex];
739     // The page might well have been re-activated, filled up, etc. before we get
740     // around to looking at it here.
741     if (pageToFree) {
742         ASSERT(pageToFree != &PartitionRootBase::gSeedPage);
743         ASSERT(pageToFree->freeCacheIndex >= 0);
744         ASSERT(static_cast<unsigned>(pageToFree->freeCacheIndex) < kMaxFreeableSpans);
745         ASSERT(pageToFree == root->globalEmptyPageRing[pageToFree->freeCacheIndex]);
746         if (!pageToFree->numAllocatedSlots && pageToFree->freelistHead) {
747             // The page is still empty, and not freed, so _really_ free it.
748             partitionFreePage(root, pageToFree);
749         }
750         pageToFree->freeCacheIndex = -1;
751     }
752 
753     // We put the empty slot span on our global list of "pages that were once
754     // empty". thus providing it a bit of breathing room to get re-used before
755     // we really free it. This improves performance, particularly on Mac OS X
756     // which has subpar memory management performance.
757     root->globalEmptyPageRing[currentIndex] = page;
758     page->freeCacheIndex = currentIndex;
759     ++currentIndex;
760     if (currentIndex == kMaxFreeableSpans)
761         currentIndex = 0;
762     root->globalEmptyPageRingIndex = currentIndex;
763 }
764 
partitionFreeSlowPath(PartitionPage * page)765 void partitionFreeSlowPath(PartitionPage* page)
766 {
767     PartitionBucket* bucket = page->bucket;
768     ASSERT(page != &PartitionRootGeneric::gSeedPage);
769     ASSERT(bucket->activePagesHead != &PartitionRootGeneric::gSeedPage);
770     if (LIKELY(page->numAllocatedSlots == 0)) {
771         // Page became fully unused.
772         if (UNLIKELY(partitionBucketIsDirectMapped(bucket))) {
773             partitionDirectUnmap(page);
774             return;
775         }
776         // If it's the current active page, attempt to change it. We'd prefer to leave
777         // the page empty as a gentle force towards defragmentation.
778         if (LIKELY(page == bucket->activePagesHead) && page->nextPage) {
779             if (partitionSetNewActivePage(page->nextPage)) {
780                 ASSERT(bucket->activePagesHead != page);
781                 // Link the empty page back in after the new current page, to
782                 // avoid losing a reference to it.
783                 // TODO: consider walking the list to link the empty page after
784                 // all non-empty pages?
785                 PartitionPage* currentPage = bucket->activePagesHead;
786                 page->nextPage = currentPage->nextPage;
787                 currentPage->nextPage = page;
788             } else {
789                 bucket->activePagesHead = page;
790                 page->nextPage = 0;
791             }
792         }
793         partitionRegisterEmptyPage(page);
794     } else {
795         // Ensure that the page is full. That's the only valid case if we
796         // arrive here.
797         ASSERT(page->numAllocatedSlots < 0);
798         // A transition of numAllocatedSlots from 0 to -1 is not legal, and
799         // likely indicates a double-free.
800         RELEASE_ASSERT(page->numAllocatedSlots != -1);
801         page->numAllocatedSlots = -page->numAllocatedSlots - 2;
802         ASSERT(page->numAllocatedSlots == static_cast<int>(partitionBucketSlots(bucket) - 1));
803         // Fully used page became partially used. It must be put back on the
804         // non-full page list. Also make it the current page to increase the
805         // chances of it being filled up again. The old current page will be
806         // the next page.
807         page->nextPage = bucket->activePagesHead;
808         bucket->activePagesHead = page;
809         --bucket->numFullPages;
810         // Special case: for a partition page with just a single slot, it may
811         // now be empty and we want to run it through the empty logic.
812         if (UNLIKELY(page->numAllocatedSlots == 0))
813             partitionFreeSlowPath(page);
814     }
815 }
816 
partitionReallocDirectMappedInPlace(PartitionRootGeneric * root,PartitionPage * page,size_t newSize)817 bool partitionReallocDirectMappedInPlace(PartitionRootGeneric* root, PartitionPage* page, size_t newSize)
818 {
819     ASSERT(partitionBucketIsDirectMapped(page->bucket));
820 
821     newSize = partitionCookieSizeAdjustAdd(newSize);
822 
823     // Note that the new size might be a bucketed size; this function is called
824     // whenever we're reallocating a direct mapped allocation.
825     newSize = partitionDirectMapSize(newSize);
826     if (newSize < kGenericMinDirectMappedDownsize)
827         return false;
828 
829     // bucket->slotSize is the current size of the allocation.
830     size_t currentSize = page->bucket->slotSize;
831     if (newSize == currentSize)
832         return true;
833 
834     char* charPtr = static_cast<char*>(partitionPageToPointer(page));
835 
836     if (newSize < currentSize) {
837         size_t mapSize = partitionPageToDirectMapExtent(page)->mapSize;
838 
839         // Don't reallocate in-place if new size is less than 80 % of the full
840         // map size, to avoid holding on to too much unused address space.
841         if ((newSize / kSystemPageSize) * 5 < (mapSize / kSystemPageSize) * 4)
842             return false;
843 
844         // Shrink by decommitting unneeded pages and making them inaccessible.
845         size_t decommitSize = currentSize - newSize;
846         partitionDecommitSystemPages(root, charPtr + newSize, decommitSize);
847         setSystemPagesInaccessible(charPtr + newSize, decommitSize);
848     } else if (newSize <= partitionPageToDirectMapExtent(page)->mapSize) {
849         // Grow within the actually allocated memory. Just need to make the
850         // pages accessible again.
851         size_t recommitSize = newSize - currentSize;
852         setSystemPagesAccessible(charPtr + currentSize, recommitSize);
853         partitionRecommitSystemPages(root, charPtr + currentSize, recommitSize);
854 
855 #if ENABLE(ASSERT)
856         memset(charPtr + currentSize, kUninitializedByte, recommitSize);
857 #endif
858     } else {
859         // We can't perform the realloc in-place.
860         // TODO: support this too when possible.
861         return false;
862     }
863 
864 #if ENABLE(ASSERT)
865     // Write a new trailing cookie.
866     partitionCookieWriteValue(charPtr + newSize - kCookieSize);
867 #endif
868 
869     page->bucket->slotSize = newSize;
870     return true;
871 }
872 
partitionReallocGeneric(PartitionRootGeneric * root,void * ptr,size_t newSize)873 void* partitionReallocGeneric(PartitionRootGeneric* root, void* ptr, size_t newSize)
874 {
875 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
876     return realloc(ptr, newSize);
877 #else
878     if (UNLIKELY(!ptr))
879         return partitionAllocGeneric(root, newSize);
880     if (UNLIKELY(!newSize)) {
881         partitionFreeGeneric(root, ptr);
882         return 0;
883     }
884 
885     RELEASE_ASSERT(newSize <= kGenericMaxDirectMapped);
886 
887     ASSERT(partitionPointerIsValid(partitionCookieFreePointerAdjust(ptr)));
888 
889     PartitionPage* page = partitionPointerToPage(partitionCookieFreePointerAdjust(ptr));
890 
891     if (UNLIKELY(partitionBucketIsDirectMapped(page->bucket))) {
892         // We may be able to perform the realloc in place by changing the
893         // accessibility of memory pages and, if reducing the size, decommitting
894         // them.
895         if (partitionReallocDirectMappedInPlace(root, page, newSize))
896             return ptr;
897     }
898 
899     size_t actualNewSize = partitionAllocActualSize(root, newSize);
900     size_t actualOldSize = partitionAllocGetSize(ptr);
901 
902     // TODO: note that tcmalloc will "ignore" a downsizing realloc() unless the
903     // new size is a significant percentage smaller. We could do the same if we
904     // determine it is a win.
905     if (actualNewSize == actualOldSize) {
906         // Trying to allocate a block of size newSize would give us a block of
907         // the same size as the one we've already got, so no point in doing
908         // anything here.
909         return ptr;
910     }
911 
912     // This realloc cannot be resized in-place. Sadness.
913     void* ret = partitionAllocGeneric(root, newSize);
914     size_t copySize = actualOldSize;
915     if (newSize < copySize)
916         copySize = newSize;
917 
918     memcpy(ret, ptr, copySize);
919     partitionFreeGeneric(root, ptr);
920     return ret;
921 #endif
922 }
923 
924 #ifndef NDEBUG
925 
partitionDumpStats(const PartitionRoot & root)926 void partitionDumpStats(const PartitionRoot& root)
927 {
928     size_t i;
929     size_t totalLive = 0;
930     size_t totalResident = 0;
931     size_t totalFreeable = 0;
932     for (i = 0; i < root.numBuckets; ++i) {
933         const PartitionBucket& bucket = root.buckets()[i];
934         if (bucket.activePagesHead == &PartitionRootGeneric::gSeedPage && !bucket.freePagesHead && !bucket.numFullPages) {
935             // Empty bucket with no freelist or full pages. Skip reporting it.
936             continue;
937         }
938         size_t numFreePages = 0;
939         PartitionPage* freePages = bucket.freePagesHead;
940         while (freePages) {
941             ++numFreePages;
942             freePages = freePages->nextPage;
943         }
944         size_t bucketSlotSize = bucket.slotSize;
945         size_t bucketNumSlots = partitionBucketSlots(&bucket);
946         size_t bucketUsefulStorage = bucketSlotSize * bucketNumSlots;
947         size_t bucketPageSize = bucket.numSystemPagesPerSlotSpan * kSystemPageSize;
948         size_t bucketWaste = bucketPageSize - bucketUsefulStorage;
949         size_t numActiveBytes = bucket.numFullPages * bucketUsefulStorage;
950         size_t numResidentBytes = bucket.numFullPages * bucketPageSize;
951         size_t numFreeableBytes = 0;
952         size_t numActivePages = 0;
953         const PartitionPage* page = bucket.activePagesHead;
954         while (page) {
955             ASSERT(page != &PartitionRootGeneric::gSeedPage);
956             // A page may be on the active list but freed and not yet swept.
957             if (!page->freelistHead && !page->numUnprovisionedSlots && !page->numAllocatedSlots) {
958                 ++numFreePages;
959             } else {
960                 ++numActivePages;
961                 numActiveBytes += (page->numAllocatedSlots * bucketSlotSize);
962                 size_t pageBytesResident = (bucketNumSlots - page->numUnprovisionedSlots) * bucketSlotSize;
963                 // Round up to system page size.
964                 pageBytesResident = (pageBytesResident + kSystemPageOffsetMask) & kSystemPageBaseMask;
965                 numResidentBytes += pageBytesResident;
966                 if (!page->numAllocatedSlots)
967                     numFreeableBytes += pageBytesResident;
968             }
969             page = page->nextPage;
970         }
971         totalLive += numActiveBytes;
972         totalResident += numResidentBytes;
973         totalFreeable += numFreeableBytes;
974         printf("bucket size %zu (pageSize %zu waste %zu): %zu alloc/%zu commit/%zu freeable bytes, %zu/%zu/%zu full/active/free pages\n", bucketSlotSize, bucketPageSize, bucketWaste, numActiveBytes, numResidentBytes, numFreeableBytes, static_cast<size_t>(bucket.numFullPages), numActivePages, numFreePages);
975     }
976     printf("total live: %zu bytes\n", totalLive);
977     printf("total resident: %zu bytes\n", totalResident);
978     printf("total freeable: %zu bytes\n", totalFreeable);
979     fflush(stdout);
980 }
981 
982 #endif // !NDEBUG
983 
984 } // namespace WTF
985 
986