retry_allocator.cc

// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "paddle/fluid/memory/allocation/retry_allocator.h"

namespace paddle {
namespace memory {
namespace allocation {

RetryAllocation::~RetryAllocation() {
  auto allocator = retry_allocator_.lock();
  // Allocator is destroyed before allocation. Should not happened usually.
  if (UNLIKELY(allocator == nullptr)) return;
  allocator->FreeUnderlyingAllocation(std::move(underlying_allocation_));
}

bool RetryAllocator::IsAllocThreadSafe() const { return true; }

std::shared_ptr<Allocation> RetryAllocator::AllocateShared(
    size_t size, Allocator::Attr attr) {
  return std::shared_ptr<Allocation>(AllocateImpl(size, attr));
}

std::unique_ptr<Allocation> RetryAllocator::Allocate(size_t size,
                                                     Allocator::Attr attr) {
  return std::unique_ptr<Allocation>(AllocateImpl(size, attr));
}

Allocation* RetryAllocator::AllocateImpl(size_t size, Allocator::Attr attr) {
  auto alloc_func = [&, this]() {
    return new RetryAllocation(underlying_allocator_->Allocate(size, attr),
                               this->shared_from_this());
  };
  // In fact, we can unify the code of allocation success and failure
  // But it would add lock even when allocation success at the first time
  try {
    return alloc_func();
  } catch (BadAlloc& bad_alloc) {
    {
      // We can just write allocation retry inside the predicate function of
      // wait_until
      // But it needs to acquire the lock when executing predicate function
      // For better performance, we use loop here
      auto end_time = std::chrono::high_resolution_clock::now() + retry_time_;
      auto wait_until = [&, this] {
        std::unique_lock<std::mutex> lock(mutex_);
        return cv_.wait_until(lock, end_time);
      };
      while (wait_until() != std::cv_status::timeout) {
        try {
          return alloc_func();
        } catch (BadAlloc& ex) {
          bad_alloc = ex;
        } catch (...) {
          throw;
        }
      }

      throw;  // rethrow the original exception or throw the internal bad_alloc
    }
  } catch (...) {
    throw;
  }
}
void RetryAllocator::FreeUnderlyingAllocation(
    std::unique_ptr<Allocation>&& allocation) {
  underlying_allocator_->Free(allocation.get());
  {
    // notify all waited allocators, they can try to allocate memory after free.
    std::lock_guard<std::mutex> lock(mutex_);
    cv_.notify_all();
  }
}

}  // namespace allocation
}  // namespace memory
}  // namespace paddle