Re: [PATCH] Fix checkout of large files to network shares under Windows XP

Hi,

I've cc:'d the mailing lists and the participants of the earlier
discussion to share your findings and to see if someone else may have an
explanation for this behaviour, or a solution.

The patch in question is c8b296450e5148c576697ea4709072b7855aacd5 and
has made it into git versions 1.7.1.1 and 1.7.2 (and up).  It caps
writes at 31MB to fix a problem with network drives in certain versions
of Windows.

Am 14.10.2010 21:26, schrieb Chad Warner:

I am experiencing issues with the patch you suggested.  I am working
with a repository via its UNC path and it is failing on large files.
I first tried lowing to several different values such as 4MB and still
had problems.  I eventually got it to work with the following:

return write(fd, buf, min(count, 1024 * 27));

I didn't notice any real delays with having to call write that many
more times.  However, I really don't know how to go about fixing this
issue or validating that this really fixes the problem.

Reducing the write cap to 27KB unconditionally sounds pretty drastic to
me; it could hurt the local case.  I didn't measure it, though, so I
might be wrong.  Depending on that measurement we perhaps need to find a
different solution.

Also, 27KB is an odd number.  I take it that 28KB won't fix the issue
for you?

I wonder about the cause of this threshold.  Wild guess: network driver
bug or other network issue?

Which version of Windows do you use on the client?  Which OS runs on the
server?  Do you have any other information that might help in
reproducing the problem?

Thanks,
René

On Sat, Oct 16, 2010 at 9:23 PM, René Scharfe
[off-list ref] wrote:

Also, 27KB is an odd number.  I take it that 28KB won't fix the issue
for you?

I guess the beginning of the written buffer is not page aligned, and the
system locks pages. Thus depending on the buffer alignment, the written
maximum can vary up to the size of one page.

I wonder about the cause of this threshold.  Wild guess: network driver
bug or other network issue?

Memory pressure? Git may consume large amount of virtual memory
during some operations, but I am not sure how it is translated in the
number of physical memory consumed by the process as well as what
quotas could be on that system. It would be interesting to see what
GetProcessMemoryInfo() returns when write() fails.


Dmitry