Home | History | Annotate | Download | only in partition_allocator
      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 #ifndef BASE_ALLOCATOR_PARTITION_ALLOCATOR_PARTITION_ALLOC_H
      6 #define BASE_ALLOCATOR_PARTITION_ALLOCATOR_PARTITION_ALLOC_H
      7 
      8 // DESCRIPTION
      9 // partitionAlloc() / PartitionAllocGeneric() and PartitionFree() /
     10 // PartitionFreeGeneric() are approximately analagous to malloc() and free().
     11 //
     12 // The main difference is that a PartitionRoot / PartitionRootGeneric object
     13 // must be supplied to these functions, representing a specific "heap partition"
     14 // that will be used to satisfy the allocation. Different partitions are
     15 // guaranteed to exist in separate address spaces, including being separate from
     16 // the main system heap. If the contained objects are all freed, physical memory
     17 // is returned to the system but the address space remains reserved.
     18 // See PartitionAlloc.md for other security properties PartitionAlloc provides.
     19 //
     20 // THE ONLY LEGITIMATE WAY TO OBTAIN A PartitionRoot IS THROUGH THE
     21 // SizeSpecificPartitionAllocator / PartitionAllocatorGeneric classes. To
     22 // minimize the instruction count to the fullest extent possible, the
     23 // PartitionRoot is really just a header adjacent to other data areas provided
     24 // by the allocator class.
     25 //
     26 // The partitionAlloc() variant of the API has the following caveats:
     27 // - Allocations and frees against a single partition must be single threaded.
     28 // - Allocations must not exceed a max size, chosen at compile-time via a
     29 // templated parameter to PartitionAllocator.
     30 // - Allocation sizes must be aligned to the system pointer size.
     31 // - Allocations are bucketed exactly according to size.
     32 //
     33 // And for PartitionAllocGeneric():
     34 // - Multi-threaded use against a single partition is ok; locking is handled.
     35 // - Allocations of any arbitrary size can be handled (subject to a limit of
     36 // INT_MAX bytes for security reasons).
     37 // - Bucketing is by approximate size, for example an allocation of 4000 bytes
     38 // might be placed into a 4096-byte bucket. Bucket sizes are chosen to try and
     39 // keep worst-case waste to ~10%.
     40 //
     41 // The allocators are designed to be extremely fast, thanks to the following
     42 // properties and design:
     43 // - Just two single (reasonably predicatable) branches in the hot / fast path
     44 //   for both allocating and (significantly) freeing.
     45 // - A minimal number of operations in the hot / fast path, with the slow paths
     46 //   in separate functions, leading to the possibility of inlining.
     47 // - Each partition page (which is usually multiple physical pages) has a
     48 //   metadata structure which allows fast mapping of free() address to an
     49 //   underlying bucket.
     50 // - Supports a lock-free API for fast performance in single-threaded cases.
     51 // - The freelist for a given bucket is split across a number of partition
     52 //   pages, enabling various simple tricks to try and minimize fragmentation.
     53 // - Fine-grained bucket sizes leading to less waste and better packing.
     54 //
     55 // The following security properties could be investigated in the future:
     56 // - Per-object bucketing (instead of per-size) is mostly available at the API,
     57 // but not used yet.
     58 // - No randomness of freelist entries or bucket position.
     59 // - Better checking for wild pointers in free().
     60 // - Better freelist masking function to guarantee fault on 32-bit.
     61 
     62 #include <limits.h>
     63 #include <string.h>
     64 
     65 #include "third_party/base/allocator/partition_allocator/page_allocator.h"
     66 #include "third_party/base/allocator/partition_allocator/spin_lock.h"
     67 #include "third_party/base/bits.h"
     68 #include "third_party/base/compiler_specific.h"
     69 #include "third_party/base/logging.h"
     70 #include "third_party/base/sys_byteorder.h"
     71 #include "third_party/build/build_config.h"
     72 
     73 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
     74 #include <stdlib.h>
     75 #endif
     76 
     77 namespace pdfium {
     78 namespace base {
     79 
     80 // Allocation granularity of sizeof(void*) bytes.
     81 static const size_t kAllocationGranularity = sizeof(void*);
     82 static const size_t kAllocationGranularityMask = kAllocationGranularity - 1;
     83 static const size_t kBucketShift = (kAllocationGranularity == 8) ? 3 : 2;
     84 
     85 // Underlying partition storage pages are a power-of-two size. It is typical
     86 // for a partition page to be based on multiple system pages. Most references to
     87 // "page" refer to partition pages.
     88 // We also have the concept of "super pages" -- these are the underlying system
     89 // allocations we make. Super pages contain multiple partition pages inside them
     90 // and include space for a small amount of metadata per partition page.
     91 // Inside super pages, we store "slot spans". A slot span is a continguous range
     92 // of one or more partition pages that stores allocations of the same size.
     93 // Slot span sizes are adjusted depending on the allocation size, to make sure
     94 // the packing does not lead to unused (wasted) space at the end of the last
     95 // system page of the span. For our current max slot span size of 64k and other
     96 // constant values, we pack _all_ PartitionAllocGeneric() sizes perfectly up
     97 // against the end of a system page.
     98 #if defined(_MIPS_ARCH_LOONGSON)
     99 static const size_t kPartitionPageShift = 16;  // 64KB
    100 #else
    101 static const size_t kPartitionPageShift = 14;  // 16KB
    102 #endif
    103 static const size_t kPartitionPageSize = 1 << kPartitionPageShift;
    104 static const size_t kPartitionPageOffsetMask = kPartitionPageSize - 1;
    105 static const size_t kPartitionPageBaseMask = ~kPartitionPageOffsetMask;
    106 static const size_t kMaxPartitionPagesPerSlotSpan = 4;
    107 
    108 // To avoid fragmentation via never-used freelist entries, we hand out partition
    109 // freelist sections gradually, in units of the dominant system page size.
    110 // What we're actually doing is avoiding filling the full partition page (16 KB)
    111 // with freelist pointers right away. Writing freelist pointers will fault and
    112 // dirty a private page, which is very wasteful if we never actually store
    113 // objects there.
    114 static const size_t kNumSystemPagesPerPartitionPage =
    115     kPartitionPageSize / kSystemPageSize;
    116 static const size_t kMaxSystemPagesPerSlotSpan =
    117     kNumSystemPagesPerPartitionPage * kMaxPartitionPagesPerSlotSpan;
    118 
    119 // We reserve virtual address space in 2MB chunks (aligned to 2MB as well).
    120 // These chunks are called "super pages". We do this so that we can store
    121 // metadata in the first few pages of each 2MB aligned section. This leads to
    122 // a very fast free(). We specifically choose 2MB because this virtual address
    123 // block represents a full but single PTE allocation on ARM, ia32 and x64.
    124 //
    125 // The layout of the super page is as follows. The sizes below are the same
    126 // for 32 bit and 64 bit.
    127 //
    128 //   | Guard page (4KB)    |
    129 //   | Metadata page (4KB) |
    130 //   | Guard pages (8KB)   |
    131 //   | Slot span           |
    132 //   | Slot span           |
    133 //   | ...                 |
    134 //   | Slot span           |
    135 //   | Guard page (4KB)    |
    136 //
    137 //   - Each slot span is a contiguous range of one or more PartitionPages.
    138 //   - The metadata page has the following format. Note that the PartitionPage
    139 //     that is not at the head of a slot span is "unused". In other words,
    140 //     the metadata for the slot span is stored only in the first PartitionPage
    141 //     of the slot span. Metadata accesses to other PartitionPages are
    142 //     redirected to the first PartitionPage.
    143 //
    144 //     | SuperPageExtentEntry (32B)                 |
    145 //     | PartitionPage of slot span 1 (32B, used)   |
    146 //     | PartitionPage of slot span 1 (32B, unused) |
    147 //     | PartitionPage of slot span 1 (32B, unused) |
    148 //     | PartitionPage of slot span 2 (32B, used)   |
    149 //     | PartitionPage of slot span 3 (32B, used)   |
    150 //     | ...                                        |
    151 //     | PartitionPage of slot span N (32B, unused) |
    152 //
    153 // A direct mapped page has a similar layout to fake it looking like a super
    154 // page:
    155 //
    156 //     | Guard page (4KB)     |
    157 //     | Metadata page (4KB)  |
    158 //     | Guard pages (8KB)    |
    159 //     | Direct mapped object |
    160 //     | Guard page (4KB)     |
    161 //
    162 //    - The metadata page has the following layout:
    163 //
    164 //     | SuperPageExtentEntry (32B)    |
    165 //     | PartitionPage (32B)           |
    166 //     | PartitionBucket (32B)         |
    167 //     | PartitionDirectMapExtent (8B) |
    168 static const size_t kSuperPageShift = 21;  // 2MB
    169 static const size_t kSuperPageSize = 1 << kSuperPageShift;
    170 static const size_t kSuperPageOffsetMask = kSuperPageSize - 1;
    171 static const size_t kSuperPageBaseMask = ~kSuperPageOffsetMask;
    172 static const size_t kNumPartitionPagesPerSuperPage =
    173     kSuperPageSize / kPartitionPageSize;
    174 
    175 static const size_t kPageMetadataShift = 5;  // 32 bytes per partition page.
    176 static const size_t kPageMetadataSize = 1 << kPageMetadataShift;
    177 
    178 // The following kGeneric* constants apply to the generic variants of the API.
    179 // The "order" of an allocation is closely related to the power-of-two size of
    180 // the allocation. More precisely, the order is the bit index of the
    181 // most-significant-bit in the allocation size, where the bit numbers starts
    182 // at index 1 for the least-significant-bit.
    183 // In terms of allocation sizes, order 0 covers 0, order 1 covers 1, order 2
    184 // covers 2->3, order 3 covers 4->7, order 4 covers 8->15.
    185 static const size_t kGenericMinBucketedOrder = 4;  // 8 bytes.
    186 static const size_t kGenericMaxBucketedOrder =
    187     20;  // Largest bucketed order is 1<<(20-1) (storing 512KB -> almost 1MB)
    188 static const size_t kGenericNumBucketedOrders =
    189     (kGenericMaxBucketedOrder - kGenericMinBucketedOrder) + 1;
    190 // Eight buckets per order (for the higher orders), e.g. order 8 is 128, 144,
    191 // 160, ..., 240:
    192 static const size_t kGenericNumBucketsPerOrderBits = 3;
    193 static const size_t kGenericNumBucketsPerOrder =
    194     1 << kGenericNumBucketsPerOrderBits;
    195 static const size_t kGenericNumBuckets =
    196     kGenericNumBucketedOrders * kGenericNumBucketsPerOrder;
    197 static const size_t kGenericSmallestBucket = 1
    198                                              << (kGenericMinBucketedOrder - 1);
    199 static const size_t kGenericMaxBucketSpacing =
    200     1 << ((kGenericMaxBucketedOrder - 1) - kGenericNumBucketsPerOrderBits);
    201 static const size_t kGenericMaxBucketed =
    202     (1 << (kGenericMaxBucketedOrder - 1)) +
    203     ((kGenericNumBucketsPerOrder - 1) * kGenericMaxBucketSpacing);
    204 static const size_t kGenericMinDirectMappedDownsize =
    205     kGenericMaxBucketed +
    206     1;  // Limit when downsizing a direct mapping using realloc().
    207 static const size_t kGenericMaxDirectMapped = INT_MAX - kSystemPageSize;
    208 static const size_t kBitsPerSizeT = sizeof(void*) * CHAR_BIT;
    209 
    210 // Constants for the memory reclaim logic.
    211 static const size_t kMaxFreeableSpans = 16;
    212 
    213 // If the total size in bytes of allocated but not committed pages exceeds this
    214 // value (probably it is a "out of virtual address space" crash),
    215 // a special crash stack trace is generated at |partitionOutOfMemory|.
    216 // This is to distinguish "out of virtual address space" from
    217 // "out of physical memory" in crash reports.
    218 static const size_t kReasonableSizeOfUnusedPages = 1024 * 1024 * 1024;  // 1GiB
    219 
    220 #if DCHECK_IS_ON()
    221 // These two byte values match tcmalloc.
    222 static const unsigned char kUninitializedByte = 0xAB;
    223 static const unsigned char kFreedByte = 0xCD;
    224 static const size_t kCookieSize =
    225     16;  // Handles alignment up to XMM instructions on Intel.
    226 static const unsigned char kCookieValue[kCookieSize] = {
    227     0xDE, 0xAD, 0xBE, 0xEF, 0xCA, 0xFE, 0xD0, 0x0D,
    228     0x13, 0x37, 0xF0, 0x05, 0xBA, 0x11, 0xAB, 0x1E};
    229 #endif
    230 
    231 struct PartitionBucket;
    232 struct PartitionRootBase;
    233 
    234 struct PartitionFreelistEntry {
    235   PartitionFreelistEntry* next;
    236 };
    237 
    238 // Some notes on page states. A page can be in one of four major states:
    239 // 1) Active.
    240 // 2) Full.
    241 // 3) Empty.
    242 // 4) Decommitted.
    243 // An active page has available free slots. A full page has no free slots. An
    244 // empty page has no free slots, and a decommitted page is an empty page that
    245 // had its backing memory released back to the system.
    246 // There are two linked lists tracking the pages. The "active page" list is an
    247 // approximation of a list of active pages. It is an approximation because
    248 // full, empty and decommitted pages may briefly be present in the list until
    249 // we next do a scan over it.
    250 // The "empty page" list is an accurate list of pages which are either empty
    251 // or decommitted.
    252 //
    253 // The significant page transitions are:
    254 // - free() will detect when a full page has a slot free()'d and immediately
    255 // return the page to the head of the active list.
    256 // - free() will detect when a page is fully emptied. It _may_ add it to the
    257 // empty list or it _may_ leave it on the active list until a future list scan.
    258 // - malloc() _may_ scan the active page list in order to fulfil the request.
    259 // If it does this, full, empty and decommitted pages encountered will be
    260 // booted out of the active list. If there are no suitable active pages found,
    261 // an empty or decommitted page (if one exists) will be pulled from the empty
    262 // list on to the active list.
    263 struct PartitionPage {
    264   PartitionFreelistEntry* freelist_head;
    265   PartitionPage* next_page;
    266   PartitionBucket* bucket;
    267   // Deliberately signed, 0 for empty or decommitted page, -n for full pages:
    268   int16_t num_allocated_slots;
    269   uint16_t num_unprovisioned_slots;
    270   uint16_t page_offset;
    271   int16_t empty_cache_index;  // -1 if not in the empty cache.
    272 };
    273 
    274 struct PartitionBucket {
    275   PartitionPage* active_pages_head;  // Accessed most in hot path => goes first.
    276   PartitionPage* empty_pages_head;
    277   PartitionPage* decommitted_pages_head;
    278   uint32_t slot_size;
    279   unsigned num_system_pages_per_slot_span : 8;
    280   unsigned num_full_pages : 24;
    281 };
    282 
    283 // An "extent" is a span of consecutive superpages. We link to the partition's
    284 // next extent (if there is one) at the very start of a superpage's metadata
    285 // area.
    286 struct PartitionSuperPageExtentEntry {
    287   PartitionRootBase* root;
    288   char* super_page_base;
    289   char* super_pages_end;
    290   PartitionSuperPageExtentEntry* next;
    291 };
    292 
    293 struct PartitionDirectMapExtent {
    294   PartitionDirectMapExtent* next_extent;
    295   PartitionDirectMapExtent* prev_extent;
    296   PartitionBucket* bucket;
    297   size_t map_size;  // Mapped size, not including guard pages and meta-data.
    298 };
    299 
    300 struct BASE_EXPORT PartitionRootBase {
    301   size_t total_size_of_committed_pages;
    302   size_t total_size_of_super_pages;
    303   size_t total_size_of_direct_mapped_pages;
    304   // Invariant: total_size_of_committed_pages <=
    305   //                total_size_of_super_pages +
    306   //                total_size_of_direct_mapped_pages.
    307   unsigned num_buckets;
    308   unsigned max_allocation;
    309   bool initialized;
    310   char* next_super_page;
    311   char* next_partition_page;
    312   char* next_partition_page_end;
    313   PartitionSuperPageExtentEntry* current_extent;
    314   PartitionSuperPageExtentEntry* first_extent;
    315   PartitionDirectMapExtent* direct_map_list;
    316   PartitionPage* global_empty_page_ring[kMaxFreeableSpans];
    317   int16_t global_empty_page_ring_index;
    318   uintptr_t inverted_self;
    319 
    320   static subtle::SpinLock gInitializedLock;
    321   static bool gInitialized;
    322   // gSeedPage is used as a sentinel to indicate that there is no page
    323   // in the active page list. We can use nullptr, but in that case we need
    324   // to add a null-check branch to the hot allocation path. We want to avoid
    325   // that.
    326   static PartitionPage gSeedPage;
    327   static PartitionBucket gPagedBucket;
    328   // gOomHandlingFunction is invoked when ParitionAlloc hits OutOfMemory.
    329   static void (*gOomHandlingFunction)();
    330 };
    331 
    332 // Never instantiate a PartitionRoot directly, instead use PartitionAlloc.
    333 struct PartitionRoot : public PartitionRootBase {
    334   // The PartitionAlloc templated class ensures the following is correct.
    335   ALWAYS_INLINE PartitionBucket* buckets() {
    336     return reinterpret_cast<PartitionBucket*>(this + 1);
    337   }
    338   ALWAYS_INLINE const PartitionBucket* buckets() const {
    339     return reinterpret_cast<const PartitionBucket*>(this + 1);
    340   }
    341 };
    342 
    343 // Never instantiate a PartitionRootGeneric directly, instead use
    344 // PartitionAllocatorGeneric.
    345 struct PartitionRootGeneric : public PartitionRootBase {
    346   subtle::SpinLock lock;
    347   // Some pre-computed constants.
    348   size_t order_index_shifts[kBitsPerSizeT + 1];
    349   size_t order_sub_index_masks[kBitsPerSizeT + 1];
    350   // The bucket lookup table lets us map a size_t to a bucket quickly.
    351   // The trailing +1 caters for the overflow case for very large allocation
    352   // sizes.  It is one flat array instead of a 2D array because in the 2D
    353   // world, we'd need to index array[blah][max+1] which risks undefined
    354   // behavior.
    355   PartitionBucket*
    356       bucket_lookups[((kBitsPerSizeT + 1) * kGenericNumBucketsPerOrder) + 1];
    357   PartitionBucket buckets[kGenericNumBuckets];
    358 };
    359 
    360 // Flags for PartitionAllocGenericFlags.
    361 enum PartitionAllocFlags {
    362   PartitionAllocReturnNull = 1 << 0,
    363 };
    364 
    365 // Struct used to retrieve total memory usage of a partition. Used by
    366 // PartitionStatsDumper implementation.
    367 struct PartitionMemoryStats {
    368   size_t total_mmapped_bytes;    // Total bytes mmaped from the system.
    369   size_t total_committed_bytes;  // Total size of commmitted pages.
    370   size_t total_resident_bytes;   // Total bytes provisioned by the partition.
    371   size_t total_active_bytes;     // Total active bytes in the partition.
    372   size_t total_decommittable_bytes;  // Total bytes that could be decommitted.
    373   size_t total_discardable_bytes;    // Total bytes that could be discarded.
    374 };
    375 
    376 // Struct used to retrieve memory statistics about a partition bucket. Used by
    377 // PartitionStatsDumper implementation.
    378 struct PartitionBucketMemoryStats {
    379   bool is_valid;       // Used to check if the stats is valid.
    380   bool is_direct_map;  // True if this is a direct mapping; size will not be
    381                        // unique.
    382   uint32_t bucket_slot_size;     // The size of the slot in bytes.
    383   uint32_t allocated_page_size;  // Total size the partition page allocated from
    384                                  // the system.
    385   uint32_t active_bytes;         // Total active bytes used in the bucket.
    386   uint32_t resident_bytes;       // Total bytes provisioned in the bucket.
    387   uint32_t decommittable_bytes;  // Total bytes that could be decommitted.
    388   uint32_t discardable_bytes;    // Total bytes that could be discarded.
    389   uint32_t num_full_pages;       // Number of pages with all slots allocated.
    390   uint32_t num_active_pages;     // Number of pages that have at least one
    391                                  // provisioned slot.
    392   uint32_t num_empty_pages;      // Number of pages that are empty
    393                                  // but not decommitted.
    394   uint32_t num_decommitted_pages;  // Number of pages that are empty
    395                                    // and decommitted.
    396 };
    397 
    398 // Interface that is passed to PartitionDumpStats and
    399 // PartitionDumpStatsGeneric for using the memory statistics.
    400 class BASE_EXPORT PartitionStatsDumper {
    401  public:
    402   // Called to dump total memory used by partition, once per partition.
    403   virtual void PartitionDumpTotals(const char* partition_name,
    404                                    const PartitionMemoryStats*) = 0;
    405 
    406   // Called to dump stats about buckets, for each bucket.
    407   virtual void PartitionsDumpBucketStats(const char* partition_name,
    408                                          const PartitionBucketMemoryStats*) = 0;
    409 };
    410 
    411 BASE_EXPORT void PartitionAllocGlobalInit(void (*oom_handling_function)());
    412 BASE_EXPORT void PartitionAllocInit(PartitionRoot*,
    413                                     size_t num_buckets,
    414                                     size_t max_allocation);
    415 BASE_EXPORT void PartitionAllocGenericInit(PartitionRootGeneric*);
    416 
    417 enum PartitionPurgeFlags {
    418   // Decommitting the ring list of empty pages is reasonably fast.
    419   PartitionPurgeDecommitEmptyPages = 1 << 0,
    420   // Discarding unused system pages is slower, because it involves walking all
    421   // freelists in all active partition pages of all buckets >= system page
    422   // size. It often frees a similar amount of memory to decommitting the empty
    423   // pages, though.
    424   PartitionPurgeDiscardUnusedSystemPages = 1 << 1,
    425 };
    426 
    427 BASE_EXPORT void PartitionPurgeMemory(PartitionRoot*, int);
    428 BASE_EXPORT void PartitionPurgeMemoryGeneric(PartitionRootGeneric*, int);
    429 
    430 BASE_EXPORT NOINLINE void* PartitionAllocSlowPath(PartitionRootBase*,
    431                                                   int,
    432                                                   size_t,
    433                                                   PartitionBucket*);
    434 BASE_EXPORT NOINLINE void PartitionFreeSlowPath(PartitionPage*);
    435 BASE_EXPORT NOINLINE void* PartitionReallocGeneric(PartitionRootGeneric*,
    436                                                    void*,
    437                                                    size_t,
    438                                                    const char* type_name);
    439 
    440 BASE_EXPORT void PartitionDumpStats(PartitionRoot*,
    441                                     const char* partition_name,
    442                                     bool is_light_dump,
    443                                     PartitionStatsDumper*);
    444 BASE_EXPORT void PartitionDumpStatsGeneric(PartitionRootGeneric*,
    445                                            const char* partition_name,
    446                                            bool is_light_dump,
    447                                            PartitionStatsDumper*);
    448 
    449 class BASE_EXPORT PartitionAllocHooks {
    450  public:
    451   typedef void AllocationHook(void* address, size_t, const char* type_name);
    452   typedef void FreeHook(void* address);
    453 
    454   static void SetAllocationHook(AllocationHook* hook) {
    455     allocation_hook_ = hook;
    456   }
    457   static void SetFreeHook(FreeHook* hook) { free_hook_ = hook; }
    458 
    459   static void AllocationHookIfEnabled(void* address,
    460                                       size_t size,
    461                                       const char* type_name) {
    462     AllocationHook* hook = allocation_hook_;
    463     if (UNLIKELY(hook != nullptr))
    464       hook(address, size, type_name);
    465   }
    466 
    467   static void FreeHookIfEnabled(void* address) {
    468     FreeHook* hook = free_hook_;
    469     if (UNLIKELY(hook != nullptr))
    470       hook(address);
    471   }
    472 
    473   static void ReallocHookIfEnabled(void* old_address,
    474                                    void* new_address,
    475                                    size_t size,
    476                                    const char* type_name) {
    477     // Report a reallocation as a free followed by an allocation.
    478     AllocationHook* allocation_hook = allocation_hook_;
    479     FreeHook* free_hook = free_hook_;
    480     if (UNLIKELY(allocation_hook && free_hook)) {
    481       free_hook(old_address);
    482       allocation_hook(new_address, size, type_name);
    483     }
    484   }
    485 
    486  private:
    487   // Pointers to hook functions that PartitionAlloc will call on allocation and
    488   // free if the pointers are non-null.
    489   static AllocationHook* allocation_hook_;
    490   static FreeHook* free_hook_;
    491 };
    492 
    493 ALWAYS_INLINE PartitionFreelistEntry* PartitionFreelistMask(
    494     PartitionFreelistEntry* ptr) {
    495 // We use bswap on little endian as a fast mask for two reasons:
    496 // 1) If an object is freed and its vtable used where the attacker doesn't
    497 // get the chance to run allocations between the free and use, the vtable
    498 // dereference is likely to fault.
    499 // 2) If the attacker has a linear buffer overflow and elects to try and
    500 // corrupt a freelist pointer, partial pointer overwrite attacks are
    501 // thwarted.
    502 // For big endian, similar guarantees are arrived at with a negation.
    503 #if defined(ARCH_CPU_BIG_ENDIAN)
    504   uintptr_t masked = ~reinterpret_cast<uintptr_t>(ptr);
    505 #else
    506   uintptr_t masked = ByteSwapUintPtrT(reinterpret_cast<uintptr_t>(ptr));
    507 #endif
    508   return reinterpret_cast<PartitionFreelistEntry*>(masked);
    509 }
    510 
    511 ALWAYS_INLINE size_t PartitionCookieSizeAdjustAdd(size_t size) {
    512 #if DCHECK_IS_ON()
    513   // Add space for cookies, checking for integer overflow. TODO(palmer):
    514   // Investigate the performance and code size implications of using
    515   // CheckedNumeric throughout PA.
    516   DCHECK(size + (2 * kCookieSize) > size);
    517   size += 2 * kCookieSize;
    518 #endif
    519   return size;
    520 }
    521 
    522 ALWAYS_INLINE size_t PartitionCookieSizeAdjustSubtract(size_t size) {
    523 #if DCHECK_IS_ON()
    524   // Remove space for cookies.
    525   DCHECK(size >= 2 * kCookieSize);
    526   size -= 2 * kCookieSize;
    527 #endif
    528   return size;
    529 }
    530 
    531 ALWAYS_INLINE void* PartitionCookieFreePointerAdjust(void* ptr) {
    532 #if DCHECK_IS_ON()
    533   // The value given to the application is actually just after the cookie.
    534   ptr = static_cast<char*>(ptr) - kCookieSize;
    535 #endif
    536   return ptr;
    537 }
    538 
    539 ALWAYS_INLINE void PartitionCookieWriteValue(void* ptr) {
    540 #if DCHECK_IS_ON()
    541   auto* cookie_ptr = reinterpret_cast<unsigned char*>(ptr);
    542   for (size_t i = 0; i < kCookieSize; ++i, ++cookie_ptr)
    543     *cookie_ptr = kCookieValue[i];
    544 #endif
    545 }
    546 
    547 ALWAYS_INLINE void PartitionCookieCheckValue(void* ptr) {
    548 #if DCHECK_IS_ON()
    549   auto* cookie_ptr = reinterpret_cast<unsigned char*>(ptr);
    550   for (size_t i = 0; i < kCookieSize; ++i, ++cookie_ptr)
    551     DCHECK(*cookie_ptr == kCookieValue[i]);
    552 #endif
    553 }
    554 
    555 ALWAYS_INLINE char* PartitionSuperPageToMetadataArea(char* ptr) {
    556   auto pointer_as_uint = reinterpret_cast<uintptr_t>(ptr);
    557   DCHECK(!(pointer_as_uint & kSuperPageOffsetMask));
    558   // The metadata area is exactly one system page (the guard page) into the
    559   // super page.
    560   return reinterpret_cast<char*>(pointer_as_uint + kSystemPageSize);
    561 }
    562 
    563 ALWAYS_INLINE PartitionPage* PartitionPointerToPageNoAlignmentCheck(void* ptr) {
    564   auto pointer_as_uint = reinterpret_cast<uintptr_t>(ptr);
    565   auto* super_page_ptr =
    566       reinterpret_cast<char*>(pointer_as_uint & kSuperPageBaseMask);
    567   uintptr_t partition_page_index =
    568       (pointer_as_uint & kSuperPageOffsetMask) >> kPartitionPageShift;
    569   // Index 0 is invalid because it is the metadata and guard area and
    570   // the last index is invalid because it is a guard page.
    571   DCHECK(partition_page_index);
    572   DCHECK(partition_page_index < kNumPartitionPagesPerSuperPage - 1);
    573   auto* page = reinterpret_cast<PartitionPage*>(
    574       PartitionSuperPageToMetadataArea(super_page_ptr) +
    575       (partition_page_index << kPageMetadataShift));
    576   // Partition pages in the same slot span can share the same page object.
    577   // Adjust for that.
    578   size_t delta = page->page_offset << kPageMetadataShift;
    579   page =
    580       reinterpret_cast<PartitionPage*>(reinterpret_cast<char*>(page) - delta);
    581   return page;
    582 }
    583 
    584 ALWAYS_INLINE void* PartitionPageToPointer(const PartitionPage* page) {
    585   auto pointer_as_uint = reinterpret_cast<uintptr_t>(page);
    586   uintptr_t super_page_offset = (pointer_as_uint & kSuperPageOffsetMask);
    587   DCHECK(super_page_offset > kSystemPageSize);
    588   DCHECK(super_page_offset < kSystemPageSize + (kNumPartitionPagesPerSuperPage *
    589                                                 kPageMetadataSize));
    590   uintptr_t partition_page_index =
    591       (super_page_offset - kSystemPageSize) >> kPageMetadataShift;
    592   // Index 0 is invalid because it is the metadata area and the last index is
    593   // invalid because it is a guard page.
    594   DCHECK(partition_page_index);
    595   DCHECK(partition_page_index < kNumPartitionPagesPerSuperPage - 1);
    596   uintptr_t super_page_base = (pointer_as_uint & kSuperPageBaseMask);
    597   auto* ret = reinterpret_cast<void*>(
    598       super_page_base + (partition_page_index << kPartitionPageShift));
    599   return ret;
    600 }
    601 
    602 ALWAYS_INLINE PartitionPage* PartitionPointerToPage(void* ptr) {
    603   PartitionPage* page = PartitionPointerToPageNoAlignmentCheck(ptr);
    604   // Checks that the pointer is a multiple of bucket size.
    605   DCHECK(!((reinterpret_cast<uintptr_t>(ptr) -
    606             reinterpret_cast<uintptr_t>(PartitionPageToPointer(page))) %
    607            page->bucket->slot_size));
    608   return page;
    609 }
    610 
    611 ALWAYS_INLINE bool PartitionBucketIsDirectMapped(
    612     const PartitionBucket* bucket) {
    613   return !bucket->num_system_pages_per_slot_span;
    614 }
    615 
    616 ALWAYS_INLINE size_t PartitionBucketBytes(const PartitionBucket* bucket) {
    617   return bucket->num_system_pages_per_slot_span * kSystemPageSize;
    618 }
    619 
    620 ALWAYS_INLINE uint16_t PartitionBucketSlots(const PartitionBucket* bucket) {
    621   return static_cast<uint16_t>(PartitionBucketBytes(bucket) /
    622                                bucket->slot_size);
    623 }
    624 
    625 ALWAYS_INLINE size_t* PartitionPageGetRawSizePtr(PartitionPage* page) {
    626   // For single-slot buckets which span more than one partition page, we
    627   // have some spare metadata space to store the raw allocation size. We
    628   // can use this to report better statistics.
    629   PartitionBucket* bucket = page->bucket;
    630   if (bucket->slot_size <= kMaxSystemPagesPerSlotSpan * kSystemPageSize)
    631     return nullptr;
    632 
    633   DCHECK((bucket->slot_size % kSystemPageSize) == 0);
    634   DCHECK(PartitionBucketIsDirectMapped(bucket) ||
    635          PartitionBucketSlots(bucket) == 1);
    636   page++;
    637   return reinterpret_cast<size_t*>(&page->freelist_head);
    638 }
    639 
    640 ALWAYS_INLINE size_t PartitionPageGetRawSize(PartitionPage* page) {
    641   size_t* raw_size_ptr = PartitionPageGetRawSizePtr(page);
    642   if (UNLIKELY(raw_size_ptr != nullptr))
    643     return *raw_size_ptr;
    644   return 0;
    645 }
    646 
    647 ALWAYS_INLINE PartitionRootBase* PartitionPageToRoot(PartitionPage* page) {
    648   auto* extent_entry = reinterpret_cast<PartitionSuperPageExtentEntry*>(
    649       reinterpret_cast<uintptr_t>(page) & kSystemPageBaseMask);
    650   return extent_entry->root;
    651 }
    652 
    653 ALWAYS_INLINE bool PartitionPointerIsValid(void* ptr) {
    654   PartitionPage* page = PartitionPointerToPage(ptr);
    655   PartitionRootBase* root = PartitionPageToRoot(page);
    656   return root->inverted_self == ~reinterpret_cast<uintptr_t>(root);
    657 }
    658 
    659 ALWAYS_INLINE void* PartitionBucketAlloc(PartitionRootBase* root,
    660                                          int flags,
    661                                          size_t size,
    662                                          PartitionBucket* bucket) {
    663   PartitionPage* page = bucket->active_pages_head;
    664   // Check that this page is neither full nor freed.
    665   DCHECK(page->num_allocated_slots >= 0);
    666   void* ret = page->freelist_head;
    667   if (LIKELY(ret)) {
    668     // If these asserts fire, you probably corrupted memory.
    669     DCHECK(PartitionPointerIsValid(ret));
    670     // All large allocations must go through the slow path to correctly
    671     // update the size metadata.
    672     DCHECK(PartitionPageGetRawSize(page) == 0);
    673     PartitionFreelistEntry* new_head =
    674         PartitionFreelistMask(static_cast<PartitionFreelistEntry*>(ret)->next);
    675     page->freelist_head = new_head;
    676     page->num_allocated_slots++;
    677   } else {
    678     ret = PartitionAllocSlowPath(root, flags, size, bucket);
    679     DCHECK(!ret || PartitionPointerIsValid(ret));
    680   }
    681 #if DCHECK_IS_ON()
    682   if (!ret)
    683     return nullptr;
    684   // Fill the uninitialized pattern, and write the cookies.
    685   page = PartitionPointerToPage(ret);
    686   size_t slot_size = page->bucket->slot_size;
    687   size_t raw_size = PartitionPageGetRawSize(page);
    688   if (raw_size) {
    689     DCHECK(raw_size == size);
    690     slot_size = raw_size;
    691   }
    692   size_t no_cookie_size = PartitionCookieSizeAdjustSubtract(slot_size);
    693   auto* char_ret = static_cast<char*>(ret);
    694   // The value given to the application is actually just after the cookie.
    695   ret = char_ret + kCookieSize;
    696   memset(ret, kUninitializedByte, no_cookie_size);
    697   PartitionCookieWriteValue(char_ret);
    698   PartitionCookieWriteValue(char_ret + kCookieSize + no_cookie_size);
    699 #endif
    700   return ret;
    701 }
    702 
    703 ALWAYS_INLINE void* PartitionAlloc(PartitionRoot* root,
    704                                    size_t size,
    705                                    const char* type_name) {
    706 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
    707   void* result = malloc(size);
    708   CHECK(result);
    709   return result;
    710 #else
    711   size_t requested_size = size;
    712   size = PartitionCookieSizeAdjustAdd(size);
    713   DCHECK(root->initialized);
    714   size_t index = size >> kBucketShift;
    715   DCHECK(index < root->num_buckets);
    716   DCHECK(size == index << kBucketShift);
    717   PartitionBucket* bucket = &root->buckets()[index];
    718   void* result = PartitionBucketAlloc(root, 0, size, bucket);
    719   PartitionAllocHooks::AllocationHookIfEnabled(result, requested_size,
    720                                                type_name);
    721   return result;
    722 #endif  // defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
    723 }
    724 
    725 ALWAYS_INLINE void PartitionFreeWithPage(void* ptr, PartitionPage* page) {
    726 // If these asserts fire, you probably corrupted memory.
    727 #if DCHECK_IS_ON()
    728   size_t slot_size = page->bucket->slot_size;
    729   size_t raw_size = PartitionPageGetRawSize(page);
    730   if (raw_size)
    731     slot_size = raw_size;
    732   PartitionCookieCheckValue(ptr);
    733   PartitionCookieCheckValue(reinterpret_cast<char*>(ptr) + slot_size -
    734                             kCookieSize);
    735   memset(ptr, kFreedByte, slot_size);
    736 #endif
    737   DCHECK(page->num_allocated_slots);
    738   PartitionFreelistEntry* freelist_head = page->freelist_head;
    739   DCHECK(!freelist_head || PartitionPointerIsValid(freelist_head));
    740   CHECK(ptr != freelist_head);  // Catches an immediate double free.
    741   // Look for double free one level deeper in debug.
    742   DCHECK(!freelist_head || ptr != PartitionFreelistMask(freelist_head->next));
    743   auto* entry = static_cast<PartitionFreelistEntry*>(ptr);
    744   entry->next = PartitionFreelistMask(freelist_head);
    745   page->freelist_head = entry;
    746   --page->num_allocated_slots;
    747   if (UNLIKELY(page->num_allocated_slots <= 0)) {
    748     PartitionFreeSlowPath(page);
    749   } else {
    750     // All single-slot allocations must go through the slow path to
    751     // correctly update the size metadata.
    752     DCHECK(PartitionPageGetRawSize(page) == 0);
    753   }
    754 }
    755 
    756 ALWAYS_INLINE void PartitionFree(void* ptr) {
    757 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
    758   free(ptr);
    759 #else
    760   PartitionAllocHooks::FreeHookIfEnabled(ptr);
    761   ptr = PartitionCookieFreePointerAdjust(ptr);
    762   DCHECK(PartitionPointerIsValid(ptr));
    763   PartitionPage* page = PartitionPointerToPage(ptr);
    764   PartitionFreeWithPage(ptr, page);
    765 #endif
    766 }
    767 
    768 ALWAYS_INLINE PartitionBucket* PartitionGenericSizeToBucket(
    769     PartitionRootGeneric* root,
    770     size_t size) {
    771   size_t order = kBitsPerSizeT - bits::CountLeadingZeroBitsSizeT(size);
    772   // The order index is simply the next few bits after the most significant bit.
    773   size_t order_index = (size >> root->order_index_shifts[order]) &
    774                        (kGenericNumBucketsPerOrder - 1);
    775   // And if the remaining bits are non-zero we must bump the bucket up.
    776   size_t sub_order_index = size & root->order_sub_index_masks[order];
    777   PartitionBucket* bucket =
    778       root->bucket_lookups[(order << kGenericNumBucketsPerOrderBits) +
    779                            order_index + !!sub_order_index];
    780   DCHECK(!bucket->slot_size || bucket->slot_size >= size);
    781   DCHECK(!(bucket->slot_size % kGenericSmallestBucket));
    782   return bucket;
    783 }
    784 
    785 ALWAYS_INLINE void* PartitionAllocGenericFlags(PartitionRootGeneric* root,
    786                                                int flags,
    787                                                size_t size,
    788                                                const char* type_name) {
    789 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
    790   void* result = malloc(size);
    791   CHECK(result || flags & PartitionAllocReturnNull);
    792   return result;
    793 #else
    794   DCHECK(root->initialized);
    795   size_t requested_size = size;
    796   size = PartitionCookieSizeAdjustAdd(size);
    797   PartitionBucket* bucket = PartitionGenericSizeToBucket(root, size);
    798   void* ret = nullptr;
    799   {
    800     subtle::SpinLock::Guard guard(root->lock);
    801     ret = PartitionBucketAlloc(root, flags, size, bucket);
    802   }
    803   PartitionAllocHooks::AllocationHookIfEnabled(ret, requested_size, type_name);
    804   return ret;
    805 #endif
    806 }
    807 
    808 ALWAYS_INLINE void* PartitionAllocGeneric(PartitionRootGeneric* root,
    809                                           size_t size,
    810                                           const char* type_name) {
    811   return PartitionAllocGenericFlags(root, 0, size, type_name);
    812 }
    813 
    814 ALWAYS_INLINE void PartitionFreeGeneric(PartitionRootGeneric* root, void* ptr) {
    815 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
    816   free(ptr);
    817 #else
    818   DCHECK(root->initialized);
    819 
    820   if (UNLIKELY(!ptr))
    821     return;
    822 
    823   PartitionAllocHooks::FreeHookIfEnabled(ptr);
    824   ptr = PartitionCookieFreePointerAdjust(ptr);
    825   DCHECK(PartitionPointerIsValid(ptr));
    826   PartitionPage* page = PartitionPointerToPage(ptr);
    827   {
    828     subtle::SpinLock::Guard guard(root->lock);
    829     PartitionFreeWithPage(ptr, page);
    830   }
    831 #endif
    832 }
    833 
    834 ALWAYS_INLINE size_t PartitionDirectMapSize(size_t size) {
    835   // Caller must check that the size is not above the kGenericMaxDirectMapped
    836   // limit before calling. This also guards against integer overflow in the
    837   // calculation here.
    838   DCHECK(size <= kGenericMaxDirectMapped);
    839   return (size + kSystemPageOffsetMask) & kSystemPageBaseMask;
    840 }
    841 
    842 ALWAYS_INLINE size_t PartitionAllocActualSize(PartitionRootGeneric* root,
    843                                               size_t size) {
    844 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
    845   return size;
    846 #else
    847   DCHECK(root->initialized);
    848   size = PartitionCookieSizeAdjustAdd(size);
    849   PartitionBucket* bucket = PartitionGenericSizeToBucket(root, size);
    850   if (LIKELY(!PartitionBucketIsDirectMapped(bucket))) {
    851     size = bucket->slot_size;
    852   } else if (size > kGenericMaxDirectMapped) {
    853     // Too large to allocate => return the size unchanged.
    854   } else {
    855     DCHECK(bucket == &PartitionRootBase::gPagedBucket);
    856     size = PartitionDirectMapSize(size);
    857   }
    858   return PartitionCookieSizeAdjustSubtract(size);
    859 #endif
    860 }
    861 
    862 ALWAYS_INLINE bool PartitionAllocSupportsGetSize() {
    863 #if defined(MEMORY_TOOL_REPLACES_ALLOCATOR)
    864   return false;
    865 #else
    866   return true;
    867 #endif
    868 }
    869 
    870 ALWAYS_INLINE size_t PartitionAllocGetSize(void* ptr) {
    871   // No need to lock here. Only |ptr| being freed by another thread could
    872   // cause trouble, and the caller is responsible for that not happening.
    873   DCHECK(PartitionAllocSupportsGetSize());
    874   ptr = PartitionCookieFreePointerAdjust(ptr);
    875   DCHECK(PartitionPointerIsValid(ptr));
    876   PartitionPage* page = PartitionPointerToPage(ptr);
    877   size_t size = page->bucket->slot_size;
    878   return PartitionCookieSizeAdjustSubtract(size);
    879 }
    880 
    881 // N (or more accurately, N - sizeof(void*)) represents the largest size in
    882 // bytes that will be handled by a SizeSpecificPartitionAllocator.
    883 // Attempts to partitionAlloc() more than this amount will fail.
    884 template <size_t N>
    885 class SizeSpecificPartitionAllocator {
    886  public:
    887   static const size_t kMaxAllocation = N - kAllocationGranularity;
    888   static const size_t kNumBuckets = N / kAllocationGranularity;
    889   void init() {
    890     PartitionAllocInit(&partition_root_, kNumBuckets, kMaxAllocation);
    891   }
    892   ALWAYS_INLINE PartitionRoot* root() { return &partition_root_; }
    893 
    894  private:
    895   PartitionRoot partition_root_;
    896   PartitionBucket actual_buckets_[kNumBuckets];
    897 };
    898 
    899 class PartitionAllocatorGeneric {
    900  public:
    901   void init() { PartitionAllocGenericInit(&partition_root_); }
    902   ALWAYS_INLINE PartitionRootGeneric* root() { return &partition_root_; }
    903 
    904  private:
    905   PartitionRootGeneric partition_root_;
    906 };
    907 
    908 }  // namespace base
    909 }  // namespace pdfium
    910 
    911 #endif  // BASE_ALLOCATOR_PARTITION_ALLOCATOR_PARTITION_ALLOC_H
    912