Rounds up VirtualAlloc calls on windows to dwAllocationGranularity to prevent fragmentation

Rounds up VirtualAlloc calls on windows to dwAllocationGranularity to prevent fragmentation

Actually on Windows VirtualAlloc'ated address might have bigger alignment than reported by SYSTEM_INFO.dwPageSize, see SYSTEM_INFO.dwAllocationGranularity.

As coalescing of spans in tcmalloc requires that those spans are physically adjacent, using smaller alignment inhibited coalescing. Additional benefit is we now
require ways less memory at least in some case.

Also adds assert()s for alignment values that would lead to pathological virtual address space waste.  No current callers pass in alignment values other than kPageSize and it's unlikely anyone will.

Original patch and discussion was at http://codereview.chromium.org/235003

Patch by: antonm@chromium.org

BUG=22701
TEST=Manually checked test page that did 10m XHR. Memory use goes from >2GB and a crash to 102MB. Membuster shows the change is roughly within noise and possibly a very slight regression ( <<5% ).

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=27162

[jamesr, 2009-09-24 22:33:15]

Follow up to http://codereview.chromium.org/235003 with assert()s for bad
alignment values.

[jar (doing other things), 2009-09-24 23:24:34]

http://codereview.chromium.org/222028/diff/1/2
File third_party/tcmalloc/port.cc (right):

http://codereview.chromium.org/222028/diff/1/2#newcode198
Line 198: // Safest is to make actual_size same as input-size.
The unexplained "safest" comment could be elaborated to:

Report the total number of bytes the OS delivered, rather than just |size| so
that spans can include the full amount, and not be artificially separated,
precluding coalescing.

http://codereview.chromium.org/222028/diff/1/2#newcode200
Line 200: // in actual_size and decommitting.
I don't understand this TODO comment.  Can you clarify?  Perhaps it should be
removed now that we have the assertions?

[jar (doing other things), 2009-09-24 23:26:03]

I should have added... with fixes to the comments: LGTM

[jamesr, 2009-09-25 01:42:41]

On 2009/09/24 23:26:03, jar wrote:
> I should have added... with fixes to the comments: LGTM

http://src.chromium.org/viewvc/chrome?view=rev&revision=27162

Thanks a lot to jar for the review and antonm for the patch.  I modified the
comments as you suggested, with some edits to make it a bit more readable.  I
think the TODO(antonm) is still valid since currently the interface doesn't
allow for the caller to deal with alignment issues generally.

[Anton Muhin, 2009-09-25 12:28:51]

just in case: LGTM

On Fri, Sep 25, 2009 at 5:42 AM,  <jamesr@chromium.org> wrote:
> On 2009/09/24 23:26:03, jar wrote:
>>
>> I should have added... with fixes to the comments: LGTM
>
> http://src.chromium.org/viewvc/chrome?view=3Drev&revision=3D27162
>
> Thanks a lot to jar for the review and antonm for the patch. =A0I modifie=
d the
> comments as you suggested, with some edits to make it a bit more readable=
.
> =A0I
> think the TODO(antonm) is still valid since currently the interface doesn=
't
> allow for the caller to deal with alignment issues generally.
>
> http://codereview.chromium.org/222028
>