On 7/9/07, Blair Zajac <blair@orcaware.com> wrote:
> Ivan Zhakov wrote:
> > On 7/8/07, Ivan Zhakov <chemodax@gmail.com> wrote:
> >> On 7/8/07, D.J. Heap <djheap@gmail.com> wrote:
> >> > On 7/7/07, Ivan Zhakov <chemodax@gmail.com> wrote:
> >> > [snip]
> >> > > And of course primary question: Does it work without my patch?
> >> Because
> >> > > these failures can be related to new txn generation code,
> >> committed by
> >> > > Blair in r25430.
> >> > >
> >> >
> >> >
> >> > I'm building with trunk r25679 and after backing out r25650, the
> >> > fs-test.exe passes all tests again. So, either the patch is bringing
> >> > out some lurking problem, or it is affecting the fs stuff in some way
> >> > I can't seem to figure out.
> >> >
> >> It seems I found way to reproduce problem even without my patch. Many
> >> fs-test will fail on Windows if you just remove time part from
> >> generated transaction name in fs_fs. See patch:
> >>
> >> Index: subversion/libsvn_fs_fs/fs_fs.c
> >> ===================================================================
> >> --- subversion/libsvn_fs_fs/fs_fs.c (revision 25654)
> >> +++ subversion/libsvn_fs_fs/fs_fs.c (working copy)
> >> @@ -3195,7 +3195,7 @@
> >> *p = '-';
> >>
> >> process_id = getpid();
> >> - now = apr_time_now();
> >> + now = 0; /* It's for testing purposes only! should apr_time_now(). */
> >>
> >> unique_basename = apr_psprintf(pool, "%s-%05d-%" APR_TIME_T_FMT,
> >> hostname_str, process_id, now);
> >> Index: subversion/libsvn_subr/utf.c
> >> ===================================================================
> >> --- subversion/libsvn_subr/utf.c (revision 25654)
> >> +++ subversion/libsvn_subr/utf.c (working copy)
> >> @@ -253,7 +253,7 @@
> >> pool = apr_hash_pool_get(xlate_handle_hash);
> >>
> >> /* Try to create a handle. */
> >> -#if defined( WIN32)
> >> +#if 0
> >> apr_err = svn_subr__win32_xlate_open(&handle, topage, frompage, pool);
> >> #elif defined(AS400)
> >> apr_err = apr_xlate_open(&handle, (int)topage, (int)frompage, pool);
> >>
> >> So it seems the problem of my patch that it made running Subversion
> >> faster and we got two transactions in the same time :)
> >>
> >> Now I'm trying to figure out why this transaction generation code
> >> fails on Windows.
> >>
> > It seems I understand what is happen here.
> > Before r25430 we've been generated transaction named
> > "<rev>-<uniquifier>". After r25430 we tried to guarantee transaction
> > names to be unique and added hostname, pid and time to transaction
> > name. And for optimization we removed revision number from txn name.
> > So we got "<hostname>-<pid>-<time>-<uniquifier>" txn names.
> >
> > Looks good, but bad news that this form of txn names isn't guarantee
> > to be unique. At least on Windows, where apr_time_now() could return
> > the same time for different transactions.
>
> How could that happen? Isn't the resolution of the time in microseconds?
No, it isn't. On Windows apr_time_now() using GetSystemTime(). In
returned in microseconds, but precision is approximately 10
milliseconds.
http://lists.boost.org/boost-users/2004/05/6635.php
>
> > And we are fail because fsfs assumes that transaction names unique
> > between revisions! So it MUST include revision number.
> >
> > So we have to problems:
> > 1. transaction names still isn't guaranteed to be unique
> > 2. fsfs depends that transaction names unique between revisions, i.e.
> > include revision number.
>
> The real solution, which I need to implement, is to add to fsfs a serial
> transaction number, much like BDB does. This will make the transaction
> names much shorter, which is a good thing.
>
Agreed. Do you have plans to do this? I really don't like that Windows
build slave is failing now.
--
Ivan Zhakov
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Mon Jul 9 11:42:16 2007