From 36dff0d011cc1cf5a41e7e0b30057206ea9fd6f9 Mon Sep 17 00:00:00 2001 From: ChiaHungDuan Date: Wed, 2 Oct 2024 13:34:46 -0700 Subject: [PATCH] [scudo] Reduce unsuccessful attempts of page releasing (#110583) We introduce a new strategy to track how many bytes are not released because of the contraint of release interval. This will change the `TryReleaseThreshold` adaptively so that we avoid releasing the same pages multiple times (and wasting time on the case of no pages to release). On Android, the number of release attempts decreases 33% (572 to 382) and the worst case drops from 251 to 33. At the same time, it maintains almost the same RSS usage (with some improvements as well). Note that in this CL, this is only applied to non small blocks. We will bring the strategy to all the size classes later. --- compiler-rt/lib/scudo/standalone/primary64.h | 133 +++++++++++++------ 1 file changed, 94 insertions(+), 39 deletions(-) diff --git a/compiler-rt/lib/scudo/standalone/primary64.h b/compiler-rt/lib/scudo/standalone/primary64.h index a3876479ec8158..c02058219b5e13 100644 --- a/compiler-rt/lib/scudo/standalone/primary64.h +++ b/compiler-rt/lib/scudo/standalone/primary64.h @@ -532,6 +532,11 @@ template class SizeClassAllocator64 { uptr BytesInFreeListAtLastCheckpoint; uptr RangesReleased; uptr LastReleasedBytes; + // The minimum size of pushed blocks to trigger page release. + uptr TryReleaseThreshold; + // The number of bytes not triggering `releaseToOSMaybe()` because of + // the length of release interval. + uptr PendingPushedBytesDelta; u64 LastReleaseAtNs; }; @@ -560,8 +565,6 @@ template class SizeClassAllocator64 { u32 RandState GUARDED_BY(MMLock) = 0; BlocksInfo FreeListInfo GUARDED_BY(FLLock); PagesInfo MemMapInfo GUARDED_BY(MMLock); - // The minimum size of pushed blocks to trigger page release. - uptr TryReleaseThreshold GUARDED_BY(MMLock) = 0; ReleaseToOsInfo ReleaseInfo GUARDED_BY(MMLock) = {}; bool Exhausted GUARDED_BY(MMLock) = false; bool isPopulatingFreeList GUARDED_BY(FLLock) = false; @@ -610,9 +613,8 @@ template class SizeClassAllocator64 { return BlockSize < PageSize / 16U; } - ALWAYS_INLINE static bool isLargeBlock(uptr BlockSize) { - const uptr PageSize = getPageSizeCached(); - return BlockSize > PageSize; + ALWAYS_INLINE uptr getMinReleaseAttemptSize(uptr BlockSize) { + return roundUp(BlockSize, getPageSizeCached()); } ALWAYS_INLINE void initRegion(RegionInfo *Region, uptr ClassId, @@ -631,12 +633,16 @@ template class SizeClassAllocator64 { (getRandomModN(&Region->RandState, 16) + 1) * PageSize; } + const uptr BlockSize = getSizeByClassId(ClassId); // Releasing small blocks is expensive, set a higher threshold to avoid // frequent page releases. - if (isSmallBlock(getSizeByClassId(ClassId))) - Region->TryReleaseThreshold = PageSize * SmallerBlockReleasePageDelta; - else - Region->TryReleaseThreshold = PageSize; + if (isSmallBlock(BlockSize)) { + Region->ReleaseInfo.TryReleaseThreshold = + PageSize * SmallerBlockReleasePageDelta; + } else { + Region->ReleaseInfo.TryReleaseThreshold = + getMinReleaseAttemptSize(BlockSize); + } } void pushBatchClassBlocks(RegionInfo *Region, CompactPtrT *Array, u32 Size) @@ -1245,6 +1251,7 @@ template class SizeClassAllocator64 { uptr BytesInFreeList; const uptr AllocatedUserEnd = Region->MemMapInfo.AllocatedUser + Region->RegionBeg; + uptr RegionPushedBytesDelta = 0; SinglyLinkedList GroupsToRelease; { @@ -1267,6 +1274,12 @@ template class SizeClassAllocator64 { return 0; } + // Given that we will unlock the freelist for block operations, cache the + // value here so that when we are adapting the `TryReleaseThreshold` + // later, we are using the right metric. + RegionPushedBytesDelta = + BytesInFreeList - Region->ReleaseInfo.BytesInFreeListAtLastCheckpoint; + // ==================================================================== // // 2. Determine which groups can release the pages. Use a heuristic to // gather groups that are candidates for doing a release. @@ -1310,12 +1323,45 @@ template class SizeClassAllocator64 { auto SkipRegion = [](UNUSED uptr RegionIndex) { return false; }; releaseFreeMemoryToOS(Context, Recorder, SkipRegion); if (Recorder.getReleasedRangesCount() > 0) { + // This is the case that we didn't hit the release threshold but it has + // been past a certain period of time. Thus we try to release some pages + // and if it does release some additional pages, it's hint that we are + // able to lower the threshold. Currently, this case happens when the + // `RegionPushedBytesDelta` is over half of the `TryReleaseThreshold`. As + // a result, we shrink the threshold to half accordingly. + // TODO(chiahungduan): Apply the same adjustment strategy to small blocks. + if (!isSmallBlock(BlockSize)) { + if (RegionPushedBytesDelta < Region->ReleaseInfo.TryReleaseThreshold && + Recorder.getReleasedBytes() > + Region->ReleaseInfo.LastReleasedBytes + + getMinReleaseAttemptSize(BlockSize)) { + Region->ReleaseInfo.TryReleaseThreshold = + Max(Region->ReleaseInfo.TryReleaseThreshold / 2, + getMinReleaseAttemptSize(BlockSize)); + } + } + Region->ReleaseInfo.BytesInFreeListAtLastCheckpoint = BytesInFreeList; Region->ReleaseInfo.RangesReleased += Recorder.getReleasedRangesCount(); Region->ReleaseInfo.LastReleasedBytes = Recorder.getReleasedBytes(); } Region->ReleaseInfo.LastReleaseAtNs = getMonotonicTimeFast(); + if (Region->ReleaseInfo.PendingPushedBytesDelta > 0) { + // Instead of increasing the threshold by the amount of + // `PendingPushedBytesDelta`, we only increase half of the amount so that + // it won't be a leap (which may lead to higher memory pressure) because + // of certain memory usage bursts which don't happen frequently. + Region->ReleaseInfo.TryReleaseThreshold += + Region->ReleaseInfo.PendingPushedBytesDelta / 2; + // This is another guard of avoiding the growth of threshold indefinitely. + // Note that we may consider to make this configurable if we have a better + // way to model this. + Region->ReleaseInfo.TryReleaseThreshold = Min( + Region->ReleaseInfo.TryReleaseThreshold, (1UL << GroupSizeLog) / 2); + Region->ReleaseInfo.PendingPushedBytesDelta = 0; + } + // ====================================================================== // // 5. Merge the `GroupsToRelease` back to the freelist. // ====================================================================== // @@ -1329,8 +1375,6 @@ template class SizeClassAllocator64 { REQUIRES(Region->MMLock, Region->FLLock) { DCHECK_GE(Region->FreeListInfo.PoppedBlocks, Region->FreeListInfo.PushedBlocks); - const uptr PageSize = getPageSizeCached(); - // Always update `BytesInFreeListAtLastCheckpoint` with the smallest value // so that we won't underestimate the releasable pages. For example, the // following is the region usage, @@ -1354,34 +1398,45 @@ template class SizeClassAllocator64 { const uptr RegionPushedBytesDelta = BytesInFreeList - Region->ReleaseInfo.BytesInFreeListAtLastCheckpoint; - if (RegionPushedBytesDelta < PageSize) - return false; - - // Releasing smaller blocks is expensive, so we want to make sure that a - // significant amount of bytes are free, and that there has been a good - // amount of batches pushed to the freelist before attempting to release. - if (isSmallBlock(BlockSize) && ReleaseType == ReleaseToOS::Normal) - if (RegionPushedBytesDelta < Region->TryReleaseThreshold) - return false; if (ReleaseType == ReleaseToOS::Normal) { - const s32 IntervalMs = atomic_load_relaxed(&ReleaseToOsIntervalMs); - if (IntervalMs < 0) + if (RegionPushedBytesDelta < Region->ReleaseInfo.TryReleaseThreshold / 2) + return false; + + const u64 IntervalNs = + static_cast(atomic_load_relaxed(&ReleaseToOsIntervalMs)) * + 1000000; + if (IntervalNs < 0) return false; - // The constant 8 here is selected from profiling some apps and the number - // of unreleased pages in the large size classes is around 16 pages or - // more. Choose half of it as a heuristic and which also avoids page - // release every time for every pushBlocks() attempt by large blocks. - const bool ByPassReleaseInterval = - isLargeBlock(BlockSize) && RegionPushedBytesDelta > 8 * PageSize; - if (!ByPassReleaseInterval) { - if (Region->ReleaseInfo.LastReleaseAtNs + - static_cast(IntervalMs) * 1000000 > - getMonotonicTimeFast()) { - // Memory was returned recently. + const u64 CurTimeNs = getMonotonicTimeFast(); + const u64 DiffSinceLastReleaseNs = + CurTimeNs - Region->ReleaseInfo.LastReleaseAtNs; + + // At here, `RegionPushedBytesDelta` is more than half of + // `TryReleaseThreshold`. If the last release happened 2 release interval + // before, we will still try to see if there's any chance to release some + // memory even it doesn't exceed the threshold. + if (RegionPushedBytesDelta < Region->ReleaseInfo.TryReleaseThreshold) { + // We want the threshold to have a shorter response time to the variant + // memory usage patterns. According to data collected during experiments + // (which were done with 1, 2, 4, 8 intervals), `2` strikes the better + // balance between the memory usage and number of page release attempts. + if (DiffSinceLastReleaseNs < 2 * IntervalNs) return false; - } + } else if (DiffSinceLastReleaseNs < IntervalNs) { + // In this case, we are over the threshold but we just did some page + // release in the same release interval. This is a hint that we may want + // a higher threshold so that we can release more memory at once. + // `TryReleaseThreshold` will be adjusted according to how many bytes + // are not released, i.e., the `PendingPushedBytesdelta` here. + // TODO(chiahungduan): Apply the same adjustment strategy to small + // blocks. + if (!isSmallBlock(BlockSize)) + Region->ReleaseInfo.PendingPushedBytesDelta = RegionPushedBytesDelta; + + // Memory was returned recently. + return false; } } // if (ReleaseType == ReleaseToOS::Normal) @@ -1397,10 +1452,10 @@ template class SizeClassAllocator64 { SinglyLinkedList GroupsToRelease; // We are examining each group and will take the minimum distance to the - // release threshold as the next Region::TryReleaseThreshold(). Note that if - // the size of free blocks has reached the release threshold, the distance - // to the next release will be PageSize * SmallerBlockReleasePageDelta. See - // the comment on `SmallerBlockReleasePageDelta` for more details. + // release threshold as the next `TryReleaseThreshold`. Note that if the + // size of free blocks has reached the release threshold, the distance to + // the next release will be PageSize * SmallerBlockReleasePageDelta. See the + // comment on `SmallerBlockReleasePageDelta` for more details. uptr MinDistToThreshold = GroupSize; for (BatchGroupT *BG = Region->FreeListInfo.BlockList.front(), @@ -1548,7 +1603,7 @@ template class SizeClassAllocator64 { // back to normal. if (MinDistToThreshold == GroupSize) MinDistToThreshold = PageSize * SmallerBlockReleasePageDelta; - Region->TryReleaseThreshold = MinDistToThreshold; + Region->ReleaseInfo.TryReleaseThreshold = MinDistToThreshold; } return GroupsToRelease;