[svn.haxx.se] · SVN Dev · SVN Users · SVN Org · TSVN Dev · TSVN Users · Subclipse Dev · Subclipse Users · this month's index

Re: apr_file_flush and short writes

From: Stefan Sperling <stsp_at_elego.de>
Date: Tue, 22 Feb 2011 00:10:18 +0100

On Mon, Feb 21, 2011 at 02:55:50PM -0800, Blair Zajac wrote:
> On 2/9/11 10:30 AM, Blair Zajac wrote:
> >On 2/9/11 1:38 AM, John Szakmeister wrote:
> >>On Mon, Feb 7, 2011 at 4:26 PM, Blair Zajac<blair_at_orcaware.com> wrote:
> >>>[I sent this to dev_at_apr.apache.org but haven't received a response. Thread
> >>>here:
> >>>http://mail-archives.apache.org/mod_mbox/apr-dev/201102.mbox/%3CF7B1928D-D32F-48DD-B8D9-80B26906AE51@orcaware.com%3E
> >>>
> >>>. Given the importance of writing complete files for svn, could somebody
> >>>take a look and see if this is a valid issue].
> >>>
> >>>I was looking at apr_file_flush() and think I found a bug where if write()
> >>>doesn't do a full write, then the apr_file_t will destroy and buffered data
> >>>because it sets its internal buffer position back to 0.
> >>
> >>Yeah, that looks like a bug.
> >>
> >>>Here's a patch for this:
> >>>
> >>>Index: file_io/unix/readwrite.c
> >>>===================================================================
> >>>--- file_io/unix/readwrite.c (revision 1067340)
> >>>+++ file_io/unix/readwrite.c (working copy)
> >>>@@ -409,7 +409,11 @@
> >>>rv = errno;
> >>>} else {
> >>>thefile->filePtr += written;
> >>>- thefile->bufpos = 0;
> >>>+ if (written != thefile->bufpos)
> >>>+ memmove(thefile->buffer,
> >>>+ thefile->buffer + written,
> >>>+ thefile->bufpos - written);
> >>>+ thefile->bufpos -= written;
> >>>}
> >>>}
> >>>
> >>>Beyond this, there's no a mechanism to report that all the buffered data
> >>>didn't get into the file. Perhaps apr_file_flush() should loop until the
> >>>entire buffer is written or it gets a non-EINTR error?
> >>
> >>I think it you're right, it should loop around. Good catch!
> >
> >Thanks!
> >
> >Who here can commit to apr?
>
> Stefan Fritsch committed a looping solution in r1073142 (trunk),
> r1073145 (1.5.x), r1073146 (1.4.x).
>
> Blair

Is it possible that this is related to issue 3705?
http://subversion.tigris.org/issues/show_bug.cgi?id=3705
"fix root cause of known fixable FSFS corruption"
Received on 2011-02-22 00:11:01 CET

This is an archived mail posted to the Subversion Dev mailing list.