mirror of
https://github.com/ClickHouse/ClickHouse.git
synced 2024-12-16 11:22:12 +00:00
e1898ca89a
Scenarios that use Arena::allocContinue may waste quadratically many memory and perform quadratically many copying, when the memory range size reaches Arena's linear allocation threshold. To alleviate this, make sure that the next memory chunk allocated by allocContinue is at least linear_growth_threshold bytes bigger than the previous one, so that we don't reallocate and copy that often.
312 lines
10 KiB
C++
312 lines
10 KiB
C++
#pragma once
|
|
|
|
#include <string.h>
|
|
#include <memory>
|
|
#include <vector>
|
|
#include <boost/noncopyable.hpp>
|
|
#include <common/likely.h>
|
|
#if __has_include(<sanitizer/asan_interface.h>)
|
|
# include <sanitizer/asan_interface.h>
|
|
#endif
|
|
#include <Core/Defines.h>
|
|
#include <Common/memcpySmall.h>
|
|
#include <Common/ProfileEvents.h>
|
|
#include <Common/Allocator.h>
|
|
|
|
|
|
namespace ProfileEvents
|
|
{
|
|
extern const Event ArenaAllocChunks;
|
|
extern const Event ArenaAllocBytes;
|
|
}
|
|
|
|
namespace DB
|
|
{
|
|
|
|
|
|
/** Memory pool to append something. For example, short strings.
|
|
* Usage scenario:
|
|
* - put lot of strings inside pool, keep their addresses;
|
|
* - addresses remain valid during lifetime of pool;
|
|
* - at destruction of pool, all memory is freed;
|
|
* - memory is allocated and freed by large chunks;
|
|
* - freeing parts of data is not possible (but look at ArenaWithFreeLists if you need);
|
|
*/
|
|
class Arena : private boost::noncopyable
|
|
{
|
|
private:
|
|
/// Padding allows to use 'memcpySmallAllowReadWriteOverflow15' instead of 'memcpy'.
|
|
static constexpr size_t pad_right = 15;
|
|
|
|
/// Contiguous chunk of memory and pointer to free space inside it. Member of single-linked list.
|
|
struct alignas(16) Chunk : private Allocator<false> /// empty base optimization
|
|
{
|
|
char * begin;
|
|
char * pos;
|
|
char * end; /// does not include padding.
|
|
|
|
Chunk * prev;
|
|
|
|
Chunk(size_t size_, Chunk * prev_)
|
|
{
|
|
ProfileEvents::increment(ProfileEvents::ArenaAllocChunks);
|
|
ProfileEvents::increment(ProfileEvents::ArenaAllocBytes, size_);
|
|
|
|
begin = reinterpret_cast<char *>(Allocator<false>::alloc(size_));
|
|
pos = begin;
|
|
end = begin + size_ - pad_right;
|
|
prev = prev_;
|
|
|
|
ASAN_POISON_MEMORY_REGION(begin, size_);
|
|
}
|
|
|
|
~Chunk()
|
|
{
|
|
/// We must unpoison the memory before returning to the allocator,
|
|
/// because the allocator might not have asan integration, and the
|
|
/// memory would stay poisoned forever. If the allocator supports
|
|
/// asan, it will correctly poison the memory by itself.
|
|
ASAN_UNPOISON_MEMORY_REGION(begin, size());
|
|
|
|
Allocator<false>::free(begin, size());
|
|
|
|
if (prev)
|
|
delete prev;
|
|
}
|
|
|
|
size_t size() const { return end + pad_right - begin; }
|
|
size_t remaining() const { return end - pos; }
|
|
};
|
|
|
|
size_t growth_factor;
|
|
size_t linear_growth_threshold;
|
|
|
|
/// Last contiguous chunk of memory.
|
|
Chunk * head;
|
|
size_t size_in_bytes;
|
|
|
|
static size_t roundUpToPageSize(size_t s)
|
|
{
|
|
return (s + 4096 - 1) / 4096 * 4096;
|
|
}
|
|
|
|
/// If chunks size is less than 'linear_growth_threshold', then use exponential growth, otherwise - linear growth
|
|
/// (to not allocate too much excessive memory).
|
|
size_t nextSize(size_t min_next_size) const
|
|
{
|
|
size_t size_after_grow = 0;
|
|
|
|
if (head->size() < linear_growth_threshold)
|
|
{
|
|
size_after_grow = std::max(min_next_size, head->size() * growth_factor);
|
|
}
|
|
else
|
|
{
|
|
// allocContinue() combined with linear growth results in quadratic
|
|
// behavior: we append the data by small amounts, and when it
|
|
// doesn't fit, we create a new chunk and copy all the previous data
|
|
// into it. The number of times we do this is directly proportional
|
|
// to the total size of data that is going to be serialized. To make
|
|
// the copying happen less often, round the next size up to the
|
|
// linear_growth_threshold.
|
|
size_after_grow = ((min_next_size + linear_growth_threshold - 1)
|
|
/ linear_growth_threshold) * linear_growth_threshold;
|
|
}
|
|
|
|
assert(size_after_grow >= min_next_size);
|
|
return roundUpToPageSize(size_after_grow);
|
|
}
|
|
|
|
/// Add next contiguous chunk of memory with size not less than specified.
|
|
void NO_INLINE addChunk(size_t min_size)
|
|
{
|
|
head = new Chunk(nextSize(min_size + pad_right), head);
|
|
size_in_bytes += head->size();
|
|
}
|
|
|
|
friend class ArenaAllocator;
|
|
template <size_t> friend class AlignedArenaAllocator;
|
|
|
|
public:
|
|
Arena(size_t initial_size_ = 4096, size_t growth_factor_ = 2, size_t linear_growth_threshold_ = 128 * 1024 * 1024)
|
|
: growth_factor(growth_factor_), linear_growth_threshold(linear_growth_threshold_),
|
|
head(new Chunk(initial_size_, nullptr)), size_in_bytes(head->size())
|
|
{
|
|
}
|
|
|
|
~Arena()
|
|
{
|
|
delete head;
|
|
}
|
|
|
|
/// Get piece of memory, without alignment.
|
|
char * alloc(size_t size)
|
|
{
|
|
if (unlikely(head->pos + size > head->end))
|
|
addChunk(size);
|
|
|
|
char * res = head->pos;
|
|
head->pos += size;
|
|
ASAN_UNPOISON_MEMORY_REGION(res, size + pad_right);
|
|
return res;
|
|
}
|
|
|
|
/// Get peice of memory with alignment
|
|
char * alignedAlloc(size_t size, size_t alignment)
|
|
{
|
|
do
|
|
{
|
|
void * head_pos = head->pos;
|
|
size_t space = head->end - head->pos;
|
|
|
|
auto res = static_cast<char *>(std::align(alignment, size, head_pos, space));
|
|
if (res)
|
|
{
|
|
head->pos = static_cast<char *>(head_pos);
|
|
head->pos += size;
|
|
ASAN_UNPOISON_MEMORY_REGION(res, size + pad_right);
|
|
return res;
|
|
}
|
|
|
|
addChunk(size + alignment);
|
|
} while (true);
|
|
}
|
|
|
|
template <typename T>
|
|
T * alloc()
|
|
{
|
|
return reinterpret_cast<T *>(alignedAlloc(sizeof(T), alignof(T)));
|
|
}
|
|
|
|
/** Rollback just performed allocation.
|
|
* Must pass size not more that was just allocated.
|
|
* Return the resulting head pointer, so that the caller can assert that
|
|
* the allocation it intended to roll back was indeed the last one.
|
|
*/
|
|
void * rollback(size_t size)
|
|
{
|
|
head->pos -= size;
|
|
ASAN_POISON_MEMORY_REGION(head->pos, size + pad_right);
|
|
return head->pos;
|
|
}
|
|
|
|
/** Begin or expand a contiguous range of memory.
|
|
* 'range_start' is the start of range. If nullptr, a new range is
|
|
* allocated.
|
|
* If there is no space in the current chunk to expand the range,
|
|
* the entire range is copied to a new, bigger memory chunk, and the value
|
|
* of 'range_start' is updated.
|
|
* If the optional 'start_alignment' is specified, the start of range is
|
|
* kept aligned to this value.
|
|
*
|
|
* NOTE This method is usable only for the last allocation made on this
|
|
* Arena. For earlier allocations, see 'realloc' method.
|
|
*/
|
|
char * allocContinue(size_t additional_bytes, char const *& range_start,
|
|
size_t start_alignment = 0)
|
|
{
|
|
if (!range_start)
|
|
{
|
|
// Start a new memory range.
|
|
char * result = start_alignment
|
|
? alignedAlloc(additional_bytes, start_alignment)
|
|
: alloc(additional_bytes);
|
|
|
|
range_start = result;
|
|
return result;
|
|
}
|
|
|
|
// Extend an existing memory range with 'additional_bytes'.
|
|
|
|
// This method only works for extending the last allocation. For lack of
|
|
// original size, check a weaker condition: that 'begin' is at least in
|
|
// the current Chunk.
|
|
assert(range_start >= head->begin && range_start < head->end);
|
|
|
|
if (head->pos + additional_bytes <= head->end)
|
|
{
|
|
// The new size fits into the last chunk, so just alloc the
|
|
// additional size. We can alloc without alignment here, because it
|
|
// only applies to the start of the range, and we don't change it.
|
|
return alloc(additional_bytes);
|
|
}
|
|
|
|
// New range doesn't fit into this chunk, will copy to a new one.
|
|
//
|
|
// Note: among other things, this method is used to provide a hack-ish
|
|
// implementation of realloc over Arenas in ArenaAllocators. It wastes a
|
|
// lot of memory -- quadratically so when we reach the linear allocation
|
|
// threshold. This deficiency is intentionally left as is, and should be
|
|
// solved not by complicating this method, but by rethinking the
|
|
// approach to memory management for aggregate function states, so that
|
|
// we can provide a proper realloc().
|
|
const size_t existing_bytes = head->pos - range_start;
|
|
const size_t new_bytes = existing_bytes + additional_bytes;
|
|
const char * old_range = range_start;
|
|
|
|
char * new_range = start_alignment
|
|
? alignedAlloc(new_bytes, start_alignment)
|
|
: alloc(new_bytes);
|
|
|
|
memcpy(new_range, old_range, existing_bytes);
|
|
|
|
range_start = new_range;
|
|
return new_range + existing_bytes;
|
|
}
|
|
|
|
/// NOTE Old memory region is wasted.
|
|
char * realloc(const char * old_data, size_t old_size, size_t new_size)
|
|
{
|
|
char * res = alloc(new_size);
|
|
if (old_data)
|
|
{
|
|
memcpy(res, old_data, old_size);
|
|
ASAN_POISON_MEMORY_REGION(old_data, old_size);
|
|
}
|
|
return res;
|
|
}
|
|
|
|
char * alignedRealloc(const char * old_data, size_t old_size, size_t new_size, size_t alignment)
|
|
{
|
|
char * res = alignedAlloc(new_size, alignment);
|
|
if (old_data)
|
|
{
|
|
memcpy(res, old_data, old_size);
|
|
ASAN_POISON_MEMORY_REGION(old_data, old_size);
|
|
}
|
|
return res;
|
|
}
|
|
|
|
/// Insert string without alignment.
|
|
const char * insert(const char * data, size_t size)
|
|
{
|
|
char * res = alloc(size);
|
|
memcpy(res, data, size);
|
|
return res;
|
|
}
|
|
|
|
const char * alignedInsert(const char * data, size_t size, size_t alignment)
|
|
{
|
|
char * res = alignedAlloc(size, alignment);
|
|
memcpy(res, data, size);
|
|
return res;
|
|
}
|
|
|
|
/// Size of chunks in bytes.
|
|
size_t size() const
|
|
{
|
|
return size_in_bytes;
|
|
}
|
|
|
|
size_t remainingSpaceInCurrentChunk() const
|
|
{
|
|
return head->remaining();
|
|
}
|
|
};
|
|
|
|
using ArenaPtr = std::shared_ptr<Arena>;
|
|
using Arenas = std::vector<ArenaPtr>;
|
|
|
|
|
|
}
|