Fixes to tsan allocator64

Deallocate: drain blocks to central cache if cached too much.
Allocate: batch allocate fixed amount of blocks.

[dvyukov, 2012-08-23 16:17:05]

Please review.

[dvyukov, 2012-08-23 16:19:52]

On 2012/08/23 16:17:05, dvyukov wrote:
> Please review.

Number of cached blocks may be not optimal, as well as list traversal under a
mutex, but let's start from something simple.

[kcc1, 2012-08-24 08:17:44]

LGTM++ 
My fault that I left the FIXME unfixed.

http://codereview.appspot.com/6478053/diff/1/sanitizer_common/sanitizer_alloc...
File sanitizer_common/sanitizer_allocator64.h (right):

http://codereview.appspot.com/6478053/diff/1/sanitizer_common/sanitizer_alloc...
sanitizer_common/sanitizer_allocator64.h:151: for (uptr i = 0; i < count &&
!region->free_list.empty(); i++) {
region->free list has a size, so you don't need to empty() on every iteration. 
Also, if region->free_list has less than count chunks we want to preserve the
old behavior. 
This can be done in a separate CL if you wish

http://codereview.appspot.com/6478053/diff/1/sanitizer_common/tests/sanitizer...
File sanitizer_common/tests/sanitizer_allocator64_test.cc (right):

http://codereview.appspot.com/6478053/diff/1/sanitizer_common/tests/sanitizer...
sanitizer_common/tests/sanitizer_allocator64_test.cc:31: SCMap::MaxCached(i));
please don't remove the old debug code, just append the new one.

[dvyukov, 2012-08-24 15:54:26]

On Fri, Aug 24, 2012 at 12:17 PM, <kcc@google.com> wrote:

> LGTM++
> My fault that I left the FIXME unfixed.
>
>
> http://codereview.appspot.com/**6478053/diff/1/sanitizer_**
>
common/sanitizer_allocator64.h<http://codereview.appspot.com/6478053/diff/1/sanitizer_common/sanitizer_allocator64.h>
> File sanitizer_common/sanitizer_**allocator64.h (right):
>
> http://codereview.appspot.com/**6478053/diff/1/sanitizer_**
>
common/sanitizer_allocator64.**h#newcode151<http://codereview.appspot.com/6478053/diff/1/sanitizer_common/sanitizer_allocator64.h#newcode151>
> sanitizer_common/sanitizer_**allocator64.h:151: for (uptr i = 0; i < count
> && !region->free_list.empty(); i++) {
> region->free list has a size, so you don't need to empty() on every
> iteration.
> Also, if region->free_list has less than count chunks we want to
> preserve the old behavior.
> This can be done in a separate CL if you wish
>

Done in r162561.



>
> http://codereview.appspot.com/**6478053/diff/1/sanitizer_**
>
common/tests/sanitizer_**allocator64_test.cc<http://codereview.appspot.com/6478053/diff/1/sanitizer_common/tests/sanitizer_allocator64_test.cc>
> File sanitizer_common/tests/**sanitizer_allocator64_test.cc (right):
>
> http://codereview.appspot.com/**6478053/diff/1/sanitizer_**
>
common/tests/sanitizer_**allocator64_test.cc#newcode31<http://codereview.appspot.com/6478053/diff/1/sanitizer_common/tests/sanitizer_allocator64_test.cc#newcode31>
> sanitizer_common/tests/**sanitizer_allocator64_test.cc:**31:
> SCMap::MaxCached(i));
> please don't remove the old debug code, just append the new one.
>
>
http://codereview.appspot.com/**6478053/<http://codereview.appspot.com/6478053/>
>