1 // Copyright 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.
5 #include "base/memory/discardable_memory_allocator_android.h"
12 #include "base/basictypes.h"
13 #include "base/containers/hash_tables.h"
14 #include "base/logging.h"
15 #include "base/memory/discardable_memory.h"
16 #include "base/memory/discardable_memory_android.h"
17 #include "base/memory/scoped_vector.h"
18 #include "base/synchronization/lock.h"
19 #include "base/threading/thread_checker.h"
21 // The allocator consists of three parts (classes):
22 // - DiscardableMemoryAllocator: entry point of all allocations (through its
23 // Allocate() method) that are dispatched to the AshmemRegion instances (which
25 // - AshmemRegion: manages allocations and destructions inside a single large
26 // (e.g. 32 MBytes) ashmem region.
27 // - DiscardableAshmemChunk: class implementing the DiscardableMemory interface
28 // whose instances are returned to the client. DiscardableAshmemChunk lets the
29 // client seamlessly operate on a subrange of the ashmem region managed by
35 // Only tolerate fragmentation in used chunks *caused by the client* (as opposed
36 // to the allocator when a free chunk is reused). The client can cause such
37 // fragmentation by e.g. requesting 4097 bytes. This size would be rounded up to
38 // 8192 by the allocator which would cause 4095 bytes of fragmentation (which is
39 // currently the maximum allowed). If the client requests 4096 bytes and a free
40 // chunk of 8192 bytes is available then the free chunk gets splitted into two
41 // pieces to minimize fragmentation (since 8192 - 4096 = 4096 which is greater
43 // TODO(pliard): tune this if splitting chunks too often leads to performance
45 const size_t kMaxChunkFragmentationBytes
= 4096 - 1;
47 const size_t kMinAshmemRegionSize
= 32 * 1024 * 1024;
49 // Returns 0 if the provided size is too high to be aligned.
50 size_t AlignToNextPage(size_t size
) {
51 const size_t kPageSize
= 4096;
52 DCHECK_EQ(static_cast<int>(kPageSize
), getpagesize());
53 if (size
> std::numeric_limits
<size_t>::max() - kPageSize
+ 1)
55 const size_t mask
= ~(kPageSize
- 1);
56 return (size
+ kPageSize
- 1) & mask
;
63 class DiscardableMemoryAllocator::DiscardableAshmemChunk
64 : public DiscardableMemory
{
66 // Note that |ashmem_region| must outlive |this|.
67 DiscardableAshmemChunk(AshmemRegion
* ashmem_region
,
72 : ashmem_region_(ashmem_region
),
80 // Implemented below AshmemRegion since this requires the full definition of
82 virtual ~DiscardableAshmemChunk();
85 virtual DiscardableMemoryLockStatus
Lock() OVERRIDE
{
88 return internal::LockAshmemRegion(fd_
, offset_
, size_
, address_
);
91 virtual void Unlock() OVERRIDE
{
94 internal::UnlockAshmemRegion(fd_
, offset_
, size_
, address_
);
97 virtual void* Memory() const OVERRIDE
{
102 AshmemRegion
* const ashmem_region_
;
104 void* const address_
;
105 const size_t offset_
;
109 DISALLOW_COPY_AND_ASSIGN(DiscardableAshmemChunk
);
112 class DiscardableMemoryAllocator::AshmemRegion
{
114 // Note that |allocator| must outlive |this|.
115 static scoped_ptr
<AshmemRegion
> Create(
117 const std::string
& name
,
118 DiscardableMemoryAllocator
* allocator
) {
119 DCHECK_EQ(size
, AlignToNextPage(size
));
122 if (!internal::CreateAshmemRegion(name
.c_str(), size
, &fd
, &base
))
123 return scoped_ptr
<AshmemRegion
>();
124 return make_scoped_ptr(new AshmemRegion(fd
, size
, base
, allocator
));
128 const bool result
= internal::CloseAshmemRegion(fd_
, size_
, base_
);
132 // Returns a new instance of DiscardableMemory whose size is greater or equal
133 // than |actual_size| (which is expected to be greater or equal than
134 // |client_requested_size|).
135 // Allocation works as follows:
136 // 1) Reuse a previously freed chunk and return it if it succeeded. See
137 // ReuseFreeChunk_Locked() below for more information.
138 // 2) If no free chunk could be reused and the region is not big enough for
139 // the requested size then NULL is returned.
140 // 3) If there is enough room in the ashmem region then a new chunk is
141 // returned. This new chunk starts at |offset_| which is the end of the
142 // previously highest chunk in the region.
143 scoped_ptr
<DiscardableMemory
> Allocate_Locked(size_t client_requested_size
,
144 size_t actual_size
) {
145 DCHECK_LE(client_requested_size
, actual_size
);
146 allocator_
->lock_
.AssertAcquired();
147 scoped_ptr
<DiscardableMemory
> memory
= ReuseFreeChunk_Locked(
148 client_requested_size
, actual_size
);
150 return memory
.Pass();
151 if (size_
- offset_
< actual_size
) {
152 // This region does not have enough space left to hold the requested size.
153 return scoped_ptr
<DiscardableMemory
>();
155 void* const address
= static_cast<char*>(base_
) + offset_
;
157 new DiscardableAshmemChunk(this, fd_
, address
, offset_
, actual_size
));
158 used_to_previous_chunk_map_
.insert(
159 std::make_pair(address
, highest_allocated_chunk_
));
160 highest_allocated_chunk_
= address
;
161 offset_
+= actual_size
;
162 DCHECK_LE(offset_
, size_
);
163 return memory
.Pass();
166 void OnChunkDeletion(void* chunk
, size_t size
) {
167 AutoLock
auto_lock(allocator_
->lock_
);
168 MergeAndAddFreeChunk_Locked(chunk
, size
);
169 // Note that |this| might be deleted beyond this point.
174 FreeChunk(void* previous_chunk
, void* start
, size_t size
)
175 : previous_chunk(previous_chunk
),
180 void* const previous_chunk
;
184 bool is_null() const { return !start
; }
186 bool operator<(const FreeChunk
& other
) const {
187 return size
< other
.size
;
191 // Note that |allocator| must outlive |this|.
195 DiscardableMemoryAllocator
* allocator
)
199 allocator_(allocator
),
200 highest_allocated_chunk_(NULL
),
203 DCHECK_GE(size
, kMinAshmemRegionSize
);
208 // Tries to reuse a previously freed chunk by doing a closest size match.
209 scoped_ptr
<DiscardableMemory
> ReuseFreeChunk_Locked(
210 size_t client_requested_size
,
211 size_t actual_size
) {
212 allocator_
->lock_
.AssertAcquired();
213 const FreeChunk reused_chunk
= RemoveFreeChunkFromIterator_Locked(
214 free_chunks_
.lower_bound(FreeChunk(NULL
, NULL
, actual_size
)));
215 if (reused_chunk
.is_null())
216 return scoped_ptr
<DiscardableMemory
>();
218 used_to_previous_chunk_map_
.insert(
219 std::make_pair(reused_chunk
.start
, reused_chunk
.previous_chunk
));
220 size_t reused_chunk_size
= reused_chunk
.size
;
221 // |client_requested_size| is used below rather than |actual_size| to
222 // reflect the amount of bytes that would not be usable by the client (i.e.
223 // wasted). Using |actual_size| instead would not allow us to detect
224 // fragmentation caused by the client if he did misaligned allocations.
225 DCHECK_GE(reused_chunk
.size
, client_requested_size
);
226 const size_t fragmentation_bytes
=
227 reused_chunk
.size
- client_requested_size
;
228 if (fragmentation_bytes
> kMaxChunkFragmentationBytes
) {
229 // Split the free chunk being recycled so that its unused tail doesn't get
230 // reused (i.e. locked) which would prevent it from being evicted under
232 reused_chunk_size
= actual_size
;
233 void* const new_chunk_start
=
234 static_cast<char*>(reused_chunk
.start
) + actual_size
;
235 DCHECK_GT(reused_chunk
.size
, actual_size
);
236 const size_t new_chunk_size
= reused_chunk
.size
- actual_size
;
237 // Note that merging is not needed here since there can't be contiguous
238 // free chunks at this point.
240 FreeChunk(reused_chunk
.start
, new_chunk_start
, new_chunk_size
));
242 const size_t offset
=
243 static_cast<char*>(reused_chunk
.start
) - static_cast<char*>(base_
);
244 internal::LockAshmemRegion(
245 fd_
, offset
, reused_chunk_size
, reused_chunk
.start
);
246 scoped_ptr
<DiscardableMemory
> memory(
247 new DiscardableAshmemChunk(this, fd_
, reused_chunk
.start
, offset
,
249 return memory
.Pass();
252 // Makes the chunk identified with the provided arguments free and possibly
253 // merges this chunk with the previous and next contiguous ones.
254 // If the provided chunk is the only one used (and going to be freed) in the
255 // region then the internal ashmem region is closed so that the underlying
256 // physical pages are immediately released.
257 // Note that free chunks are unlocked therefore they can be reclaimed by the
258 // kernel if needed (under memory pressure) but they are not immediately
259 // released unfortunately since madvise(MADV_REMOVE) and
260 // fallocate(FALLOC_FL_PUNCH_HOLE) don't seem to work on ashmem. This might
261 // change in versions of kernel >=3.5 though. The fact that free chunks are
262 // not immediately released is the reason why we are trying to minimize
263 // fragmentation in order not to cause "artificial" memory pressure.
264 void MergeAndAddFreeChunk_Locked(void* chunk
, size_t size
) {
265 allocator_
->lock_
.AssertAcquired();
266 size_t new_free_chunk_size
= size
;
267 // Merge with the previous chunk.
268 void* first_free_chunk
= chunk
;
269 DCHECK(!used_to_previous_chunk_map_
.empty());
270 const hash_map
<void*, void*>::iterator previous_chunk_it
=
271 used_to_previous_chunk_map_
.find(chunk
);
272 DCHECK(previous_chunk_it
!= used_to_previous_chunk_map_
.end());
273 void* previous_chunk
= previous_chunk_it
->second
;
274 used_to_previous_chunk_map_
.erase(previous_chunk_it
);
275 if (previous_chunk
) {
276 const FreeChunk free_chunk
= RemoveFreeChunk_Locked(previous_chunk
);
277 if (!free_chunk
.is_null()) {
278 new_free_chunk_size
+= free_chunk
.size
;
279 first_free_chunk
= previous_chunk
;
280 // There should not be more contiguous previous free chunks.
281 DCHECK(!address_to_free_chunk_map_
.count(free_chunk
.previous_chunk
));
284 // Merge with the next chunk if free and present.
285 void* next_chunk
= static_cast<char*>(chunk
) + size
;
286 const FreeChunk next_free_chunk
= RemoveFreeChunk_Locked(next_chunk
);
287 if (!next_free_chunk
.is_null()) {
288 new_free_chunk_size
+= next_free_chunk
.size
;
290 DCHECK(!address_to_free_chunk_map_
.count(static_cast<char*>(next_chunk
) +
291 next_free_chunk
.size
));
293 const bool whole_ashmem_region_is_free
=
294 used_to_previous_chunk_map_
.empty();
295 if (!whole_ashmem_region_is_free
) {
297 FreeChunk(previous_chunk
, first_free_chunk
, new_free_chunk_size
));
300 // The whole ashmem region is free thus it can be deleted.
301 DCHECK_EQ(base_
, first_free_chunk
);
302 DCHECK(free_chunks_
.empty());
303 DCHECK(address_to_free_chunk_map_
.empty());
304 DCHECK(used_to_previous_chunk_map_
.empty());
305 allocator_
->DeleteAshmemRegion_Locked(this); // Deletes |this|.
308 void AddFreeChunk_Locked(const FreeChunk
& free_chunk
) {
309 allocator_
->lock_
.AssertAcquired();
310 const std::multiset
<FreeChunk
>::iterator it
= free_chunks_
.insert(
312 address_to_free_chunk_map_
.insert(std::make_pair(free_chunk
.start
, it
));
313 // Update the next used contiguous chunk, if any, since its previous chunk
314 // may have changed due to free chunks merging/splitting.
315 void* const next_used_contiguous_chunk
=
316 static_cast<char*>(free_chunk
.start
) + free_chunk
.size
;
317 hash_map
<void*, void*>::iterator previous_it
=
318 used_to_previous_chunk_map_
.find(next_used_contiguous_chunk
);
319 if (previous_it
!= used_to_previous_chunk_map_
.end())
320 previous_it
->second
= free_chunk
.start
;
323 // Finds and removes the free chunk, if any, whose start address is
324 // |chunk_start|. Returns a copy of the unlinked free chunk or a free chunk
325 // whose content is null if it was not found.
326 FreeChunk
RemoveFreeChunk_Locked(void* chunk_start
) {
327 allocator_
->lock_
.AssertAcquired();
329 void*, std::multiset
<FreeChunk
>::iterator
>::iterator it
=
330 address_to_free_chunk_map_
.find(chunk_start
);
331 if (it
== address_to_free_chunk_map_
.end())
332 return FreeChunk(NULL
, NULL
, 0U);
333 return RemoveFreeChunkFromIterator_Locked(it
->second
);
336 // Same as above but takes an iterator in.
337 FreeChunk
RemoveFreeChunkFromIterator_Locked(
338 std::multiset
<FreeChunk
>::iterator free_chunk_it
) {
339 allocator_
->lock_
.AssertAcquired();
340 if (free_chunk_it
== free_chunks_
.end())
341 return FreeChunk(NULL
, NULL
, 0U);
342 DCHECK(free_chunk_it
!= free_chunks_
.end());
343 const FreeChunk
free_chunk(*free_chunk_it
);
344 address_to_free_chunk_map_
.erase(free_chunk_it
->start
);
345 free_chunks_
.erase(free_chunk_it
);
352 DiscardableMemoryAllocator
* const allocator_
;
353 void* highest_allocated_chunk_
;
354 // Points to the end of |highest_allocated_chunk_|.
356 // Allows free chunks recycling (lookup, insertion and removal) in O(log N).
357 // Note that FreeChunk values are indexed by their size and also note that
358 // multiple free chunks can have the same size (which is why multiset<> is
359 // used instead of e.g. set<>).
360 std::multiset
<FreeChunk
> free_chunks_
;
361 // Used while merging free contiguous chunks to erase free chunks (from their
362 // start address) in constant time. Note that multiset<>::{insert,erase}()
363 // don't invalidate iterators (except the one for the element being removed
366 void*, std::multiset
<FreeChunk
>::iterator
> address_to_free_chunk_map_
;
367 // Maps the address of *used* chunks to the address of their previous
369 hash_map
<void*, void*> used_to_previous_chunk_map_
;
371 DISALLOW_COPY_AND_ASSIGN(AshmemRegion
);
374 DiscardableMemoryAllocator::DiscardableAshmemChunk::~DiscardableAshmemChunk() {
376 internal::UnlockAshmemRegion(fd_
, offset_
, size_
, address_
);
377 ashmem_region_
->OnChunkDeletion(address_
, size_
);
380 DiscardableMemoryAllocator::DiscardableMemoryAllocator(
381 const std::string
& name
,
382 size_t ashmem_region_size
)
385 std::max(kMinAshmemRegionSize
, AlignToNextPage(ashmem_region_size
))),
386 last_ashmem_region_size_(0) {
387 DCHECK_GE(ashmem_region_size_
, kMinAshmemRegionSize
);
390 DiscardableMemoryAllocator::~DiscardableMemoryAllocator() {
391 DCHECK(thread_checker_
.CalledOnValidThread());
392 DCHECK(ashmem_regions_
.empty());
395 scoped_ptr
<DiscardableMemory
> DiscardableMemoryAllocator::Allocate(
397 const size_t aligned_size
= AlignToNextPage(size
);
399 return scoped_ptr
<DiscardableMemory
>();
400 // TODO(pliard): make this function less naive by e.g. moving the free chunks
401 // multiset to the allocator itself in order to decrease even more
402 // fragmentation/speedup allocation. Note that there should not be more than a
403 // couple (=5) of AshmemRegion instances in practice though.
404 AutoLock
auto_lock(lock_
);
405 DCHECK_LE(ashmem_regions_
.size(), 5U);
406 for (ScopedVector
<AshmemRegion
>::iterator it
= ashmem_regions_
.begin();
407 it
!= ashmem_regions_
.end(); ++it
) {
408 scoped_ptr
<DiscardableMemory
> memory(
409 (*it
)->Allocate_Locked(size
, aligned_size
));
411 return memory
.Pass();
413 // The creation of the (large) ashmem region might fail if the address space
414 // is too fragmented. In case creation fails the allocator retries by
415 // repetitively dividing the size by 2.
416 const size_t min_region_size
= std::max(kMinAshmemRegionSize
, aligned_size
);
417 for (size_t region_size
= std::max(ashmem_region_size_
, aligned_size
);
418 region_size
>= min_region_size
;
419 region_size
= AlignToNextPage(region_size
/ 2)) {
420 scoped_ptr
<AshmemRegion
> new_region(
421 AshmemRegion::Create(region_size
, name_
.c_str(), this));
424 last_ashmem_region_size_
= region_size
;
425 ashmem_regions_
.push_back(new_region
.release());
426 return ashmem_regions_
.back()->Allocate_Locked(size
, aligned_size
);
428 // TODO(pliard): consider adding an histogram to see how often this happens.
429 return scoped_ptr
<DiscardableMemory
>();
432 size_t DiscardableMemoryAllocator::last_ashmem_region_size() const {
433 AutoLock
auto_lock(lock_
);
434 return last_ashmem_region_size_
;
437 void DiscardableMemoryAllocator::DeleteAshmemRegion_Locked(
438 AshmemRegion
* region
) {
439 lock_
.AssertAcquired();
440 // Note that there should not be more than a couple of ashmem region instances
441 // in |ashmem_regions_|.
442 DCHECK_LE(ashmem_regions_
.size(), 5U);
443 const ScopedVector
<AshmemRegion
>::iterator it
= std::find(
444 ashmem_regions_
.begin(), ashmem_regions_
.end(), region
);
445 DCHECK_NE(ashmem_regions_
.end(), it
);
446 std::swap(*it
, ashmem_regions_
.back());
447 ashmem_regions_
.pop_back();
450 } // namespace internal