Randy Kobes wrote:
On Thu, 22 Mar 2007, Steve Hay wrote:
Randy Kobes wrote:
On Wed, 14 Mar 2007, Steve Hay wrote:
I tried your patch with the current svn version (revision 518242),
but I'm still seeing intermittent failures (usually in tests 15, 16
and/or 20) either when I run "nmake test" from the top-level,
[ ... ]
Perhaps just to narrow things down, could you try the
attached C file (a VC++ Makefile is also attached)?
[ ... ]
I ran each program with an argument of 500 for good measure. All four
programs successfully create and then delete a large number
(presumably 500, I didn't count 'em!) of apreq* temp files.
apr_temp and apr_temp_sh both deleted all those files about as quickly
as they were made (i.e. very quickly), while apr_temp_nc and
apr_temp_nc_sh both output loads of CLEANUP messages in the console
quickly but then seemed to hang for nearly a minute, during which time
loads of apreq* files were visible in the temp directory, and then
eventually exited and all the files disappeared.
On my system (Windows XP, VC++ 6, Apache/2.2.2), there's
essentially no difference in speed between the four
programs in cleaning things up, and like you, all temp
files are gone at the end. I'm afraid I'm stuck ...
Since for you apr_temp/apr_temp_sh (fast) and
apr_temp_nc/apr_temp_nc_sh (hangs a minute) pairwise
behave similarly, that suggests APR_SHARELOCK isn't
relevant (which is supported by it's apparent lack of
relevance to this problem in the Apache sources). APR_FILE_NOCLEANUP
does seem to make a (negative)
difference in speed, but perhaps
http://marc.info/?l=apr-dev&m=117448738223552&w=2
is relevant here. In any case, the above results suggests
that having APR_FILE_NOCLEANUP and APR_SHARELOCK both
undefined would be better, which makes sense, but
both apparently help in the Perl glue on Win32. Sigh ...
I'm wondering the following ... The failed upload
tests (15, 16, and/or 20), for me, involve uploading
the perl.exe binary. Is that the same for you? Might
there be a problem with uploading a file which is
a program that might be in use at the time? Does the
attached patches against
apreq/library/util.c
which removes the
flag |= APR_FILE_NOCLEANUP | APR_SHARELOCK
line, and
apreq/glue/perl/t/apreq/upload.t
which skips using perl.exe (and httpd.exe) for
the upload tests on Win32 work for you?
I've tried running both "nmake test" and just the upload.t test many
times over each and haven't yet seen either one fail with those two
changes in place :-)
If I just apply the util.c change (i.e. drop the NOCLEANUP and SHARELOCK
flags, but continue using perl.exe and httpd.exe for the upload tests)
then all the tests still seem to pass anyway, but I do then see some
stray "apreqXXXXXX" files left over in my temp directory from time to
time (mostly when running "nmake test" rather than just the upload.t
test), which didn't happen with the upload.t change applied as well.
That's really weird. I'm not sure I understand what's going on, but I
think you might have cracked it at last!
--
