Re: Strange hardlink behavior with CIFS

[Matt Mackall <mpm-VDJrAJ4Gl5ZBDgjK7y7TUQ@public.gmane.org>]

On Wed, 2010-10-27 at 16:32 -0400, Jeff Layton wrote:
> On Wed, 27 Oct 2010 12:31:26 -0500
> Matt Mackall <mpm-VDJrAJ4Gl5ZBDgjK7y7TUQ@public.gmane.org> wrote:
> 
> > On Tue, 2010-10-19 at 15:13 +0530, Suresh Jayaraman wrote:
> > > On 10/19/2010 12:34 AM, Matt Mackall wrote:
> > > > With Linux 2.6.32-35 and either Windows or Samba in nounix mode,
> > > > hardlink counts can mysteriously disappear:
> > > > 
> > > > - create hardlink pair foo,bar
> > > > - stat foo -> nlink = 2
> > > > - open foo
> > > > - stat foo -> nlink = 1
> > > > - close
> > > > - wait or sync
> > > > - ls -l -> nlink = 2
> > > > 
> > > > Original report is here:
> > > > 
> > > > http://mercurial.selenic.com/bts/issue1866
> > > > 
> > > 
> > > A quick test on 2.6.36-rc4 kernel reveals that the problem is no longer
> > > reproducible. Could you try a more recent kernel and see whether the
> > > problem is reproducible?
> > 
> > Test continues to fail with 2.6.36.
> > 
> > # uname -a
> > Linux calx 2.6.36 #71 SMP Tue Oct 26 02:23:34 CDT 2010 x86_64 GNU/Linux
> > # mount -o nounix //localhost/stuff cifs
> >            ^^^^^^
> > # cd cifs
> > # touch foo
> > # ln foo bar
> > # ./h
> > file: foo nlinks 2
> > file: foo nlinks 1
> > file: foo nlinks 1
> > 
> > # cat h.c
> > #include <sys/types.h>
> > #include <sys/stat.h>
> > #include <unistd.h>
> > #include <fcntl.h>
> > #include <stdio.h>
> > 
> > #define FPATH "foo"
> > 
> > int main(void)
> > {
> > 	int rc, fh;
> > 	struct stat st;
> > 
> > 	rc = lstat(FPATH, &st);
> > 	printf("file: %s nlinks %d\n", FPATH, st.st_nlink);
> > 
> > 	fh = open(FPATH, 0, O_RDONLY);
> > 
> > 	rc = lstat(FPATH, &st);
> > 	printf("file: %s nlinks %d\n", FPATH, st.st_nlink);
> > 
> > 	close(fh);
> > 
> > 	rc = lstat(FPATH, &st);
> > 	printf("file: %s nlinks %d\n", FPATH, st.st_nlink);
> > 
> > 	return 0;
> > }
> > 
> > 
> > 
> > 
> 
> Ok, I can reproduce this too. The problem is this nastiness in CIFSSMBOpen:
> 
>         if (rc) {
>                 cFYI(1, "Error in Open = %d", rc);
>         } else {
>                 *pOplock = pSMBr->OplockLevel; /* 1 byte no need to le_to_cpu */
>                 *netfid = pSMBr->Fid;   /* cifs fid stays in le */
>                 /* Let caller know file was created so we can set the mode. */
>                 /* Do we care about the CreateAction in any other cases? */
>                 if (cpu_to_le32(FILE_CREATE) == pSMBr->CreateAction)
>                         *pOplock |= CIFS_CREATE_ACTION;
>                 if (pfile_info) {
>                         memcpy((char *)pfile_info, (char *)&pSMBr->CreationTime,
>                                 36 /* CreationTime to Attributes */);
>                         /* the file_info buf is endian converted by caller */
>                         pfile_info->AllocationSize = pSMBr->AllocationSize;
>                         pfile_info->EndOfFile = pSMBr->EndOfFile;
>                         pfile_info->NumberOfLinks = cpu_to_le32(1);
>                         pfile_info->DeletePending = 0;
>                 }
>         }
> 
> In particular, the NumberOfLinks being set to 1 there. That probably
> just needs to go. I'll look over it when I have time.

I should have noted that the behavior is* different be .35 and .36 - the
link count reappears quite quickly after the test in .36 so a manual
test is unlikely to show the problem now. But back-to-back close/stat
still shows the problem.

-- 
Mathematics is the supreme nostalgia of our time.