virtiofsd: Fix data corruption with O_APPEND wirte in writeback mode
diff mbox series

Message ID 20191023122523.1816-1-misono.tomohiro@jp.fujitsu.com
State New
Headers show
Series
  • virtiofsd: Fix data corruption with O_APPEND wirte in writeback mode
Related show

Commit Message

Misono Tomohiro Oct. 23, 2019, 12:25 p.m. UTC
When writeback mode is enabled (-o writeback), O_APPEND handling is
done in kernel. Therefore virtiofsd clears O_APPEND flag when open.
Otherwise O_APPEND flag takes precedence over pwrite() and write
data may corrupt.

Currently clearing O_APPEND flag is done in lo_open(), but we also
need the same operation in lo_create(). So, factor out the flag
update operation in lo_open() to update_open_flags() and call it
in both lo_open() and lo_create().

This fixes the failure of xfstest generic/069 in writeback mode
(which tests O_APPEND write data integrity).

Signed-off-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
---
 contrib/virtiofsd/passthrough_ll.c | 56 +++++++++++++++---------------
 1 file changed, 28 insertions(+), 28 deletions(-)

Comments

Vivek Goyal Oct. 23, 2019, 8:07 p.m. UTC | #1
On Wed, Oct 23, 2019 at 09:25:23PM +0900, Misono Tomohiro wrote:
> When writeback mode is enabled (-o writeback), O_APPEND handling is
> done in kernel. Therefore virtiofsd clears O_APPEND flag when open.
> Otherwise O_APPEND flag takes precedence over pwrite() and write
> data may corrupt.
> 
> Currently clearing O_APPEND flag is done in lo_open(), but we also
> need the same operation in lo_create().

> So, factor out the flag
> update operation in lo_open() to update_open_flags() and call it
> in both lo_open() and lo_create().
> 
> This fixes the failure of xfstest generic/069 in writeback mode
> (which tests O_APPEND write data integrity).
> 

Hi,

Consolidating updation of flags both for lo_create() and lo_open() makes
sense to me. I will test it tomorrow.

Thanks
Vivek

> Signed-off-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
> ---
>  contrib/virtiofsd/passthrough_ll.c | 56 +++++++++++++++---------------
>  1 file changed, 28 insertions(+), 28 deletions(-)
> 
> diff --git a/contrib/virtiofsd/passthrough_ll.c b/contrib/virtiofsd/passthrough_ll.c
> index e8892c3c32..79fb78ecce 100644
> --- a/contrib/virtiofsd/passthrough_ll.c
> +++ b/contrib/virtiofsd/passthrough_ll.c
> @@ -1733,6 +1733,32 @@ static void lo_releasedir(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info
>  	fuse_reply_err(req, 0);
>  }
>  
> +static void update_open_flags(int writeback, struct fuse_file_info *fi)
> +{
> +	/* With writeback cache, kernel may send read requests even
> +	   when userspace opened write-only */
> +	if (writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> +		fi->flags &= ~O_ACCMODE;
> +		fi->flags |= O_RDWR;
> +	}
> +
> +	/* With writeback cache, O_APPEND is handled by the kernel.
> +	   This breaks atomicity (since the file may change in the
> +	   underlying filesystem, so that the kernel's idea of the
> +	   end of the file isn't accurate anymore). In this example,
> +	   we just accept that. A more rigorous filesystem may want
> +	   to return an error here */
> +	if (writeback && (fi->flags & O_APPEND))
> +		fi->flags &= ~O_APPEND;
> +
> +	/*
> +	 * O_DIRECT in guest should not necessarily mean bypassing page
> +	 * cache on host as well. If somebody needs that behavior, it
> +	 * probably should be a configuration knob in daemon.
> +	 */
> +	fi->flags &= ~O_DIRECT;
> +}
> +
>  static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
>  		      mode_t mode, struct fuse_file_info *fi)
>  {
> @@ -1760,12 +1786,7 @@ static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
>  	if (err)
>  		goto out;
>  
> -	/*
> -	 * O_DIRECT in guest should not necessarily mean bypassing page
> -	 * cache on host as well. If somebody needs that behavior, it
> -	 * probably should be a configuration knob in daemon.
> -	 */
> -	fi->flags &= ~O_DIRECT;
> +	update_open_flags(lo->writeback, fi);
>  
>  	fd = openat(parent_inode->fd, name,
>  		    (fi->flags | O_CREAT) & ~O_NOFOLLOW, mode);
> @@ -1966,28 +1987,7 @@ static void lo_open(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info *fi)
>  
>  	fuse_log(FUSE_LOG_DEBUG, "lo_open(ino=%" PRIu64 ", flags=%d)\n", ino, fi->flags);
>  
> -	/* With writeback cache, kernel may send read requests even
> -	   when userspace opened write-only */
> -	if (lo->writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> -		fi->flags &= ~O_ACCMODE;
> -		fi->flags |= O_RDWR;
> -	}
> -
> -	/* With writeback cache, O_APPEND is handled by the kernel.
> -	   This breaks atomicity (since the file may change in the
> -	   underlying filesystem, so that the kernel's idea of the
> -	   end of the file isn't accurate anymore). In this example,
> -	   we just accept that. A more rigorous filesystem may want
> -	   to return an error here */
> -	if (lo->writeback && (fi->flags & O_APPEND))
> -		fi->flags &= ~O_APPEND;
> -
> -	/*
> -	 * O_DIRECT in guest should not necessarily mean bypassing page
> -	 * cache on host as well. If somebody needs that behavior, it
> -	 * probably should be a configuration knob in daemon.
> -	 */
> -	fi->flags &= ~O_DIRECT;
> +	update_open_flags(lo->writeback, fi);
>  
>  	sprintf(buf, "%i", lo_fd(req, ino));
>  	fd = openat(lo->proc_self_fd, buf, fi->flags & ~O_NOFOLLOW);
> -- 
> 2.21.0
> 
> _______________________________________________
> Virtio-fs mailing list
> Virtio-fs@redhat.com
> https://www.redhat.com/mailman/listinfo/virtio-fs
Vivek Goyal Oct. 24, 2019, 3:02 p.m. UTC | #2
On Wed, Oct 23, 2019 at 04:07:52PM -0400, Vivek Goyal wrote:
> On Wed, Oct 23, 2019 at 09:25:23PM +0900, Misono Tomohiro wrote:
> > When writeback mode is enabled (-o writeback), O_APPEND handling is
> > done in kernel. Therefore virtiofsd clears O_APPEND flag when open.
> > Otherwise O_APPEND flag takes precedence over pwrite() and write
> > data may corrupt.
> > 
> > Currently clearing O_APPEND flag is done in lo_open(), but we also
> > need the same operation in lo_create().
> 
> > So, factor out the flag
> > update operation in lo_open() to update_open_flags() and call it
> > in both lo_open() and lo_create().
> > 
> > This fixes the failure of xfstest generic/069 in writeback mode
> > (which tests O_APPEND write data integrity).
> > 

Hi Misono,

Have you tried running pjdfstests. Looks like with the patch applied I
see following tests failing which were passing without the patch. Can
you please have a look. I ran daemon with options "-o cache=always -o
writeback"

/root/git/pjdfstest/tests/chmod/00.t          (Wstat: 0 Tests: 119 Failed: 1)
  Failed test:  64
/root/git/pjdfstest/tests/chown/00.t          (Wstat: 0 Tests: 1323 Failed: 1)
  Failed test:  946
  TODO passed:   1107, 1112, 1116, 1122, 1127, 1131, 1137
                1142, 1146, 1152, 1157, 1161, 1167, 1172
                1176, 1182, 1187
/root/git/pjdfstest/tests/link/00.t           (Wstat: 0 Tests: 202 Failed: 1)
  Failed test:  134
/root/git/pjdfstest/tests/utimensat/01.t      (Wstat: 0 Tests: 7 Failed: 1)
  Failed test:  4
Files=232, Tests=8789, 375 wallclock secs ( 4.00 usr  2.65 sys + 51.48 cusr 262.19 csys = 320.32 CPU)
Thanks
Vivek

> 
> Hi,
> 
> Consolidating updation of flags both for lo_create() and lo_open() makes
> sense to me. I will test it tomorrow.
> 
> Thanks
> Vivek
> 
> > Signed-off-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
> > ---
> >  contrib/virtiofsd/passthrough_ll.c | 56 +++++++++++++++---------------
> >  1 file changed, 28 insertions(+), 28 deletions(-)
> > 
> > diff --git a/contrib/virtiofsd/passthrough_ll.c b/contrib/virtiofsd/passthrough_ll.c
> > index e8892c3c32..79fb78ecce 100644
> > --- a/contrib/virtiofsd/passthrough_ll.c
> > +++ b/contrib/virtiofsd/passthrough_ll.c
> > @@ -1733,6 +1733,32 @@ static void lo_releasedir(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info
> >  	fuse_reply_err(req, 0);
> >  }
> >  
> > +static void update_open_flags(int writeback, struct fuse_file_info *fi)
> > +{
> > +	/* With writeback cache, kernel may send read requests even
> > +	   when userspace opened write-only */
> > +	if (writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> > +		fi->flags &= ~O_ACCMODE;
> > +		fi->flags |= O_RDWR;
> > +	}
> > +
> > +	/* With writeback cache, O_APPEND is handled by the kernel.
> > +	   This breaks atomicity (since the file may change in the
> > +	   underlying filesystem, so that the kernel's idea of the
> > +	   end of the file isn't accurate anymore). In this example,
> > +	   we just accept that. A more rigorous filesystem may want
> > +	   to return an error here */
> > +	if (writeback && (fi->flags & O_APPEND))
> > +		fi->flags &= ~O_APPEND;
> > +
> > +	/*
> > +	 * O_DIRECT in guest should not necessarily mean bypassing page
> > +	 * cache on host as well. If somebody needs that behavior, it
> > +	 * probably should be a configuration knob in daemon.
> > +	 */
> > +	fi->flags &= ~O_DIRECT;
> > +}
> > +
> >  static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
> >  		      mode_t mode, struct fuse_file_info *fi)
> >  {
> > @@ -1760,12 +1786,7 @@ static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
> >  	if (err)
> >  		goto out;
> >  
> > -	/*
> > -	 * O_DIRECT in guest should not necessarily mean bypassing page
> > -	 * cache on host as well. If somebody needs that behavior, it
> > -	 * probably should be a configuration knob in daemon.
> > -	 */
> > -	fi->flags &= ~O_DIRECT;
> > +	update_open_flags(lo->writeback, fi);
> >  
> >  	fd = openat(parent_inode->fd, name,
> >  		    (fi->flags | O_CREAT) & ~O_NOFOLLOW, mode);
> > @@ -1966,28 +1987,7 @@ static void lo_open(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info *fi)
> >  
> >  	fuse_log(FUSE_LOG_DEBUG, "lo_open(ino=%" PRIu64 ", flags=%d)\n", ino, fi->flags);
> >  
> > -	/* With writeback cache, kernel may send read requests even
> > -	   when userspace opened write-only */
> > -	if (lo->writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> > -		fi->flags &= ~O_ACCMODE;
> > -		fi->flags |= O_RDWR;
> > -	}
> > -
> > -	/* With writeback cache, O_APPEND is handled by the kernel.
> > -	   This breaks atomicity (since the file may change in the
> > -	   underlying filesystem, so that the kernel's idea of the
> > -	   end of the file isn't accurate anymore). In this example,
> > -	   we just accept that. A more rigorous filesystem may want
> > -	   to return an error here */
> > -	if (lo->writeback && (fi->flags & O_APPEND))
> > -		fi->flags &= ~O_APPEND;
> > -
> > -	/*
> > -	 * O_DIRECT in guest should not necessarily mean bypassing page
> > -	 * cache on host as well. If somebody needs that behavior, it
> > -	 * probably should be a configuration knob in daemon.
> > -	 */
> > -	fi->flags &= ~O_DIRECT;
> > +	update_open_flags(lo->writeback, fi);
> >  
> >  	sprintf(buf, "%i", lo_fd(req, ino));
> >  	fd = openat(lo->proc_self_fd, buf, fi->flags & ~O_NOFOLLOW);
> > -- 
> > 2.21.0
> > 
> > _______________________________________________
> > Virtio-fs mailing list
> > Virtio-fs@redhat.com
> > https://www.redhat.com/mailman/listinfo/virtio-fs
> 
> _______________________________________________
> Virtio-fs mailing list
> Virtio-fs@redhat.com
> https://www.redhat.com/mailman/listinfo/virtio-fs
misono.tomohiro@fujitsu.com Oct. 25, 2019, 10:02 a.m. UTC | #3
> On Wed, Oct 23, 2019 at 04:07:52PM -0400, Vivek Goyal wrote:
> > On Wed, Oct 23, 2019 at 09:25:23PM +0900, Misono Tomohiro wrote:
> > > When writeback mode is enabled (-o writeback), O_APPEND handling is
> > > done in kernel. Therefore virtiofsd clears O_APPEND flag when open.
> > > Otherwise O_APPEND flag takes precedence over pwrite() and write
> > > data may corrupt.
> > >
> > > Currently clearing O_APPEND flag is done in lo_open(), but we also
> > > need the same operation in lo_create().
> >
> > > So, factor out the flag
> > > update operation in lo_open() to update_open_flags() and call it in
> > > both lo_open() and lo_create().
> > >
> > > This fixes the failure of xfstest generic/069 in writeback mode
> > > (which tests O_APPEND write data integrity).
> > >
> 
> Hi Misono,
> 
> Have you tried running pjdfstests. Looks like with the patch applied I see following tests failing which were passing without
> the patch. Can you please have a look. I ran daemon with options "-o cache=always -o writeback"

Sorry, I didn't run pjdfstests. I will run the test and check the results.

Thanks,
Misono

> 
> /root/git/pjdfstest/tests/chmod/00.t          (Wstat: 0 Tests: 119 Failed: 1)
>   Failed test:  64
> /root/git/pjdfstest/tests/chown/00.t          (Wstat: 0 Tests: 1323 Failed: 1)
>   Failed test:  946
>   TODO passed:   1107, 1112, 1116, 1122, 1127, 1131, 1137
>                 1142, 1146, 1152, 1157, 1161, 1167, 1172
>                 1176, 1182, 1187
> /root/git/pjdfstest/tests/link/00.t           (Wstat: 0 Tests: 202 Failed: 1)
>   Failed test:  134
> /root/git/pjdfstest/tests/utimensat/01.t      (Wstat: 0 Tests: 7 Failed: 1)
>   Failed test:  4
> Files=232, Tests=8789, 375 wallclock secs ( 4.00 usr  2.65 sys + 51.48 cusr 262.19 csys = 320.32 CPU) Thanks Vivek
> 
> >
> > Hi,
> >
> > Consolidating updation of flags both for lo_create() and lo_open()
> > makes sense to me. I will test it tomorrow.
> >
> > Thanks
> > Vivek
> >
> > > Signed-off-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
> > > ---
> > >  contrib/virtiofsd/passthrough_ll.c | 56
> > > +++++++++++++++---------------
> > >  1 file changed, 28 insertions(+), 28 deletions(-)
> > >
> > > diff --git a/contrib/virtiofsd/passthrough_ll.c
> > > b/contrib/virtiofsd/passthrough_ll.c
> > > index e8892c3c32..79fb78ecce 100644
> > > --- a/contrib/virtiofsd/passthrough_ll.c
> > > +++ b/contrib/virtiofsd/passthrough_ll.c
> > > @@ -1733,6 +1733,32 @@ static void lo_releasedir(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info
> > >  	fuse_reply_err(req, 0);
> > >  }
> > >
> > > +static void update_open_flags(int writeback, struct fuse_file_info
> > > +*fi) {
> > > +	/* With writeback cache, kernel may send read requests even
> > > +	   when userspace opened write-only */
> > > +	if (writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> > > +		fi->flags &= ~O_ACCMODE;
> > > +		fi->flags |= O_RDWR;
> > > +	}
> > > +
> > > +	/* With writeback cache, O_APPEND is handled by the kernel.
> > > +	   This breaks atomicity (since the file may change in the
> > > +	   underlying filesystem, so that the kernel's idea of the
> > > +	   end of the file isn't accurate anymore). In this example,
> > > +	   we just accept that. A more rigorous filesystem may want
> > > +	   to return an error here */
> > > +	if (writeback && (fi->flags & O_APPEND))
> > > +		fi->flags &= ~O_APPEND;
> > > +
> > > +	/*
> > > +	 * O_DIRECT in guest should not necessarily mean bypassing page
> > > +	 * cache on host as well. If somebody needs that behavior, it
> > > +	 * probably should be a configuration knob in daemon.
> > > +	 */
> > > +	fi->flags &= ~O_DIRECT;
> > > +}
> > > +
> > >  static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
> > >  		      mode_t mode, struct fuse_file_info *fi)  { @@ -1760,12
> > > +1786,7 @@ static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
> > >  	if (err)
> > >  		goto out;
> > >
> > > -	/*
> > > -	 * O_DIRECT in guest should not necessarily mean bypassing page
> > > -	 * cache on host as well. If somebody needs that behavior, it
> > > -	 * probably should be a configuration knob in daemon.
> > > -	 */
> > > -	fi->flags &= ~O_DIRECT;
> > > +	update_open_flags(lo->writeback, fi);
> > >
> > >  	fd = openat(parent_inode->fd, name,
> > >  		    (fi->flags | O_CREAT) & ~O_NOFOLLOW, mode); @@ -1966,28
> > > +1987,7 @@ static void lo_open(fuse_req_t req, fuse_ino_t ino,
> > > struct fuse_file_info *fi)
> > >
> > >  	fuse_log(FUSE_LOG_DEBUG, "lo_open(ino=%" PRIu64 ", flags=%d)\n",
> > > ino, fi->flags);
> > >
> > > -	/* With writeback cache, kernel may send read requests even
> > > -	   when userspace opened write-only */
> > > -	if (lo->writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> > > -		fi->flags &= ~O_ACCMODE;
> > > -		fi->flags |= O_RDWR;
> > > -	}
> > > -
> > > -	/* With writeback cache, O_APPEND is handled by the kernel.
> > > -	   This breaks atomicity (since the file may change in the
> > > -	   underlying filesystem, so that the kernel's idea of the
> > > -	   end of the file isn't accurate anymore). In this example,
> > > -	   we just accept that. A more rigorous filesystem may want
> > > -	   to return an error here */
> > > -	if (lo->writeback && (fi->flags & O_APPEND))
> > > -		fi->flags &= ~O_APPEND;
> > > -
> > > -	/*
> > > -	 * O_DIRECT in guest should not necessarily mean bypassing page
> > > -	 * cache on host as well. If somebody needs that behavior, it
> > > -	 * probably should be a configuration knob in daemon.
> > > -	 */
> > > -	fi->flags &= ~O_DIRECT;
> > > +	update_open_flags(lo->writeback, fi);
> > >
> > >  	sprintf(buf, "%i", lo_fd(req, ino));
> > >  	fd = openat(lo->proc_self_fd, buf, fi->flags & ~O_NOFOLLOW);
> > > --
> > > 2.21.0
> > >
> > > _______________________________________________
> > > Virtio-fs mailing list
> > > Virtio-fs@redhat.com
> > > https://www.redhat.com/mailman/listinfo/virtio-fs
> >
> > _______________________________________________
> > Virtio-fs mailing list
> > Virtio-fs@redhat.com
> > https://www.redhat.com/mailman/listinfo/virtio-fs
misono.tomohiro@fujitsu.com Oct. 29, 2019, 10:07 a.m. UTC | #4
> -----Original Message-----
> From: virtio-fs-bounces@redhat.com [mailto:virtio-fs-bounces@redhat.com] On Behalf Of misono.tomohiro@fujitsu.com
> Sent: Friday, October 25, 2019 7:02 PM
> To: 'Vivek Goyal' <vgoyal@redhat.com>
> Cc: virtio-fs@redhat.com; qemu-devel@nongnu.org
> Subject: Re: [Virtio-fs] [PATCH] virtiofsd: Fix data corruption with O_APPEND wirte in writeback mode
> 
> > On Wed, Oct 23, 2019 at 04:07:52PM -0400, Vivek Goyal wrote:
> > > On Wed, Oct 23, 2019 at 09:25:23PM +0900, Misono Tomohiro wrote:
> > > > When writeback mode is enabled (-o writeback), O_APPEND handling
> > > > is done in kernel. Therefore virtiofsd clears O_APPEND flag when open.
> > > > Otherwise O_APPEND flag takes precedence over pwrite() and write
> > > > data may corrupt.
> > > >
> > > > Currently clearing O_APPEND flag is done in lo_open(), but we also
> > > > need the same operation in lo_create().
> > >
> > > > So, factor out the flag
> > > > update operation in lo_open() to update_open_flags() and call it
> > > > in both lo_open() and lo_create().
> > > >
> > > > This fixes the failure of xfstest generic/069 in writeback mode
> > > > (which tests O_APPEND write data integrity).
> > > >
> >
> > Hi Misono,
> >
> > Have you tried running pjdfstests. Looks like with the patch applied I
> > see following tests failing which were passing without the patch. Can you please have a look. I ran daemon with options "-o
> cache=always -o writeback"

I see these errors in both with and without this patch but not always.
Do you always see the failure?

I use:
  Kernel: Fuse's for-next branch
  Qemu: virtio-fs-dev branch
  backend: XFS (relatime)

These tests fail because a/c/m time is not updated as expected.
So it seems this is related to the failure of xfstest generic/003.
I will look into the problem more.

Thanks,
Misono

> 
> >
> > /root/git/pjdfstest/tests/chmod/00.t          (Wstat: 0 Tests: 119 Failed: 1)
> >   Failed test:  64
> > /root/git/pjdfstest/tests/chown/00.t          (Wstat: 0 Tests: 1323 Failed: 1)
> >   Failed test:  946
> >   TODO passed:   1107, 1112, 1116, 1122, 1127, 1131, 1137
> >                 1142, 1146, 1152, 1157, 1161, 1167, 1172
> >                 1176, 1182, 1187
> > /root/git/pjdfstest/tests/link/00.t           (Wstat: 0 Tests: 202 Failed: 1)
> >   Failed test:  134
> > /root/git/pjdfstest/tests/utimensat/01.t      (Wstat: 0 Tests: 7 Failed: 1)
> >   Failed test:  4
> > Files=232, Tests=8789, 375 wallclock secs ( 4.00 usr  2.65 sys + 51.48
> > cusr 262.19 csys = 320.32 CPU) Thanks Vivek
> >
> > >
> > > Hi,
> > >
> > > Consolidating updation of flags both for lo_create() and lo_open()
> > > makes sense to me. I will test it tomorrow.
> > >
> > > Thanks
> > > Vivek
> > >
> > > > Signed-off-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
> > > > ---
> > > >  contrib/virtiofsd/passthrough_ll.c | 56
> > > > +++++++++++++++---------------
> > > >  1 file changed, 28 insertions(+), 28 deletions(-)
> > > >
> > > > diff --git a/contrib/virtiofsd/passthrough_ll.c
> > > > b/contrib/virtiofsd/passthrough_ll.c
> > > > index e8892c3c32..79fb78ecce 100644
> > > > --- a/contrib/virtiofsd/passthrough_ll.c
> > > > +++ b/contrib/virtiofsd/passthrough_ll.c
> > > > @@ -1733,6 +1733,32 @@ static void lo_releasedir(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info
> > > >  	fuse_reply_err(req, 0);
> > > >  }
> > > >
> > > > +static void update_open_flags(int writeback, struct
> > > > +fuse_file_info
> > > > +*fi) {
> > > > +	/* With writeback cache, kernel may send read requests even
> > > > +	   when userspace opened write-only */
> > > > +	if (writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> > > > +		fi->flags &= ~O_ACCMODE;
> > > > +		fi->flags |= O_RDWR;
> > > > +	}
> > > > +
> > > > +	/* With writeback cache, O_APPEND is handled by the kernel.
> > > > +	   This breaks atomicity (since the file may change in the
> > > > +	   underlying filesystem, so that the kernel's idea of the
> > > > +	   end of the file isn't accurate anymore). In this example,
> > > > +	   we just accept that. A more rigorous filesystem may want
> > > > +	   to return an error here */
> > > > +	if (writeback && (fi->flags & O_APPEND))
> > > > +		fi->flags &= ~O_APPEND;
> > > > +
> > > > +	/*
> > > > +	 * O_DIRECT in guest should not necessarily mean bypassing page
> > > > +	 * cache on host as well. If somebody needs that behavior, it
> > > > +	 * probably should be a configuration knob in daemon.
> > > > +	 */
> > > > +	fi->flags &= ~O_DIRECT;
> > > > +}
> > > > +
> > > >  static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
> > > >  		      mode_t mode, struct fuse_file_info *fi)  { @@ -1760,12
> > > > +1786,7 @@ static void lo_create(fuse_req_t req, fuse_ino_t
> > > > +parent, const char *name,
> > > >  	if (err)
> > > >  		goto out;
> > > >
> > > > -	/*
> > > > -	 * O_DIRECT in guest should not necessarily mean bypassing page
> > > > -	 * cache on host as well. If somebody needs that behavior, it
> > > > -	 * probably should be a configuration knob in daemon.
> > > > -	 */
> > > > -	fi->flags &= ~O_DIRECT;
> > > > +	update_open_flags(lo->writeback, fi);
> > > >
> > > >  	fd = openat(parent_inode->fd, name,
> > > >  		    (fi->flags | O_CREAT) & ~O_NOFOLLOW, mode); @@ -1966,28
> > > > +1987,7 @@ static void lo_open(fuse_req_t req, fuse_ino_t ino,
> > > > struct fuse_file_info *fi)
> > > >
> > > >  	fuse_log(FUSE_LOG_DEBUG, "lo_open(ino=%" PRIu64 ", flags=%d)\n",
> > > > ino, fi->flags);
> > > >
> > > > -	/* With writeback cache, kernel may send read requests even
> > > > -	   when userspace opened write-only */
> > > > -	if (lo->writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> > > > -		fi->flags &= ~O_ACCMODE;
> > > > -		fi->flags |= O_RDWR;
> > > > -	}
> > > > -
> > > > -	/* With writeback cache, O_APPEND is handled by the kernel.
> > > > -	   This breaks atomicity (since the file may change in the
> > > > -	   underlying filesystem, so that the kernel's idea of the
> > > > -	   end of the file isn't accurate anymore). In this example,
> > > > -	   we just accept that. A more rigorous filesystem may want
> > > > -	   to return an error here */
> > > > -	if (lo->writeback && (fi->flags & O_APPEND))
> > > > -		fi->flags &= ~O_APPEND;
> > > > -
> > > > -	/*
> > > > -	 * O_DIRECT in guest should not necessarily mean bypassing page
> > > > -	 * cache on host as well. If somebody needs that behavior, it
> > > > -	 * probably should be a configuration knob in daemon.
> > > > -	 */
> > > > -	fi->flags &= ~O_DIRECT;
> > > > +	update_open_flags(lo->writeback, fi);
> > > >
> > > >  	sprintf(buf, "%i", lo_fd(req, ino));
> > > >  	fd = openat(lo->proc_self_fd, buf, fi->flags & ~O_NOFOLLOW);
> > > > --
> > > > 2.21.0
> > > >
> > > > _______________________________________________
> > > > Virtio-fs mailing list
> > > > Virtio-fs@redhat.com
> > > > https://www.redhat.com/mailman/listinfo/virtio-fs
> > >
> > > _______________________________________________
> > > Virtio-fs mailing list
> > > Virtio-fs@redhat.com
> > > https://www.redhat.com/mailman/listinfo/virtio-fs
> 
> 
> _______________________________________________
> Virtio-fs mailing list
> Virtio-fs@redhat.com
> https://www.redhat.com/mailman/listinfo/virtio-fs
misono.tomohiro@fujitsu.com Oct. 31, 2019, 9:39 a.m. UTC | #5
> > -----Original Message-----
> > From: virtio-fs-bounces@redhat.com
> > [mailto:virtio-fs-bounces@redhat.com] On Behalf Of
> > misono.tomohiro@fujitsu.com
> > Sent: Friday, October 25, 2019 7:02 PM
> > To: 'Vivek Goyal' <vgoyal@redhat.com>
> > Cc: virtio-fs@redhat.com; qemu-devel@nongnu.org
> > Subject: Re: [Virtio-fs] [PATCH] virtiofsd: Fix data corruption with
> > O_APPEND wirte in writeback mode
> >
> > > On Wed, Oct 23, 2019 at 04:07:52PM -0400, Vivek Goyal wrote:
> > > > On Wed, Oct 23, 2019 at 09:25:23PM +0900, Misono Tomohiro wrote:
> > > > > When writeback mode is enabled (-o writeback), O_APPEND handling
> > > > > is done in kernel. Therefore virtiofsd clears O_APPEND flag when open.
> > > > > Otherwise O_APPEND flag takes precedence over pwrite() and write
> > > > > data may corrupt.
> > > > >
> > > > > Currently clearing O_APPEND flag is done in lo_open(), but we
> > > > > also need the same operation in lo_create().
> > > >
> > > > > So, factor out the flag
> > > > > update operation in lo_open() to update_open_flags() and call it
> > > > > in both lo_open() and lo_create().
> > > > >
> > > > > This fixes the failure of xfstest generic/069 in writeback mode
> > > > > (which tests O_APPEND write data integrity).
> > > > >
> > >
> > > Hi Misono,
> > >
> > > Have you tried running pjdfstests. Looks like with the patch applied
> > > I see following tests failing which were passing without the patch.
> > > Can you please have a look. I ran daemon with options "-o
> > cache=always -o writeback"
> 
> I see these errors in both with and without this patch but not always.
> Do you always see the failure?
> 
> I use:
>   Kernel: Fuse's for-next branch
>   Qemu: virtio-fs-dev branch
>   backend: XFS (relatime)
> 
> These tests fail because a/c/m time is not updated as expected.
> So it seems this is related to the failure of xfstest generic/003.
> I will look into the problem more.
> 

Hi,

So I fugured out the problem. 
The failure occurs when guest clock is earlier than host clock in writeback mode.

The failure case checks:
 1. get current c/time of file
 2. sleep 1 second
 3. modify file's metadata/data
 4. check if c/mtime is updated

And the tests fail because:
 1. when file is opened, guest inode's time is initialized by host's inode information
 2. In writeback mode, guest clock is used for c/mtime update
 3. if guest clock is earlier than host clock, c/mtime could be updated
    earlier than current inode time. 

I can reproduce the problem reliably by deliberately setting guest clock earlier
than host clock. I'm not sure if virtiofs is to be blmaed in this case, but should
we not update c/mtime in guest if it is earlier then current c/mtime?

Anyway, I believe this O_APPEND fix patch is irrelevant to the problem,
could you please check it again?

Thanks,
Misono
Vivek Goyal Oct. 31, 2019, 3:47 p.m. UTC | #6
On Thu, Oct 31, 2019 at 09:39:23AM +0000, misono.tomohiro@fujitsu.com wrote:
[..]
> > > > Hi Misono,
> > > >
> > > > Have you tried running pjdfstests. Looks like with the patch applied
> > > > I see following tests failing which were passing without the patch.
> > > > Can you please have a look. I ran daemon with options "-o
> > > cache=always -o writeback"
> > 
> > I see these errors in both with and without this patch but not always.
> > Do you always see the failure?
> > 
> > I use:
> >   Kernel: Fuse's for-next branch
> >   Qemu: virtio-fs-dev branch
> >   backend: XFS (relatime)
> > 
> > These tests fail because a/c/m time is not updated as expected.
> > So it seems this is related to the failure of xfstest generic/003.
> > I will look into the problem more.
> > 
> 
> Hi,
> 
> So I fugured out the problem. 
> The failure occurs when guest clock is earlier than host clock in writeback mode.
> 
> The failure case checks:
>  1. get current c/time of file
>  2. sleep 1 second
>  3. modify file's metadata/data
>  4. check if c/mtime is updated
> 
> And the tests fail because:
>  1. when file is opened, guest inode's time is initialized by host's inode information
>  2. In writeback mode, guest clock is used for c/mtime update
>  3. if guest clock is earlier than host clock, c/mtime could be updated
>     earlier than current inode time. 
> 
> I can reproduce the problem reliably by deliberately setting guest clock earlier
> than host clock. I'm not sure if virtiofs is to be blmaed in this case, but should
> we not update c/mtime in guest if it is earlier then current c/mtime?
> 
> Anyway, I believe this O_APPEND fix patch is irrelevant to the problem,
> could you please check it again?

Hi Misono,

Thanks for looking into it. I agree that a/c/m time updates with -o writeback
is probably a different issue and not related to your patch. So I am fine
with your patch.

Thanks
Vivek
Vivek Goyal Oct. 31, 2019, 3:47 p.m. UTC | #7
On Wed, Oct 23, 2019 at 09:25:23PM +0900, Misono Tomohiro wrote:
> When writeback mode is enabled (-o writeback), O_APPEND handling is
> done in kernel. Therefore virtiofsd clears O_APPEND flag when open.
> Otherwise O_APPEND flag takes precedence over pwrite() and write
> data may corrupt.
> 
> Currently clearing O_APPEND flag is done in lo_open(), but we also
> need the same operation in lo_create(). So, factor out the flag
> update operation in lo_open() to update_open_flags() and call it
> in both lo_open() and lo_create().
> 
> This fixes the failure of xfstest generic/069 in writeback mode
> (which tests O_APPEND write data integrity).
> 
> Signed-off-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>

Reviewed-by: Vivek Goyal <vgoyal@redhat.com>

Thanks
Vivek

> ---
>  contrib/virtiofsd/passthrough_ll.c | 56 +++++++++++++++---------------
>  1 file changed, 28 insertions(+), 28 deletions(-)
> 
> diff --git a/contrib/virtiofsd/passthrough_ll.c b/contrib/virtiofsd/passthrough_ll.c
> index e8892c3c32..79fb78ecce 100644
> --- a/contrib/virtiofsd/passthrough_ll.c
> +++ b/contrib/virtiofsd/passthrough_ll.c
> @@ -1733,6 +1733,32 @@ static void lo_releasedir(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info
>  	fuse_reply_err(req, 0);
>  }
>  
> +static void update_open_flags(int writeback, struct fuse_file_info *fi)
> +{
> +	/* With writeback cache, kernel may send read requests even
> +	   when userspace opened write-only */
> +	if (writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> +		fi->flags &= ~O_ACCMODE;
> +		fi->flags |= O_RDWR;
> +	}
> +
> +	/* With writeback cache, O_APPEND is handled by the kernel.
> +	   This breaks atomicity (since the file may change in the
> +	   underlying filesystem, so that the kernel's idea of the
> +	   end of the file isn't accurate anymore). In this example,
> +	   we just accept that. A more rigorous filesystem may want
> +	   to return an error here */
> +	if (writeback && (fi->flags & O_APPEND))
> +		fi->flags &= ~O_APPEND;
> +
> +	/*
> +	 * O_DIRECT in guest should not necessarily mean bypassing page
> +	 * cache on host as well. If somebody needs that behavior, it
> +	 * probably should be a configuration knob in daemon.
> +	 */
> +	fi->flags &= ~O_DIRECT;
> +}
> +
>  static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
>  		      mode_t mode, struct fuse_file_info *fi)
>  {
> @@ -1760,12 +1786,7 @@ static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
>  	if (err)
>  		goto out;
>  
> -	/*
> -	 * O_DIRECT in guest should not necessarily mean bypassing page
> -	 * cache on host as well. If somebody needs that behavior, it
> -	 * probably should be a configuration knob in daemon.
> -	 */
> -	fi->flags &= ~O_DIRECT;
> +	update_open_flags(lo->writeback, fi);
>  
>  	fd = openat(parent_inode->fd, name,
>  		    (fi->flags | O_CREAT) & ~O_NOFOLLOW, mode);
> @@ -1966,28 +1987,7 @@ static void lo_open(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info *fi)
>  
>  	fuse_log(FUSE_LOG_DEBUG, "lo_open(ino=%" PRIu64 ", flags=%d)\n", ino, fi->flags);
>  
> -	/* With writeback cache, kernel may send read requests even
> -	   when userspace opened write-only */
> -	if (lo->writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> -		fi->flags &= ~O_ACCMODE;
> -		fi->flags |= O_RDWR;
> -	}
> -
> -	/* With writeback cache, O_APPEND is handled by the kernel.
> -	   This breaks atomicity (since the file may change in the
> -	   underlying filesystem, so that the kernel's idea of the
> -	   end of the file isn't accurate anymore). In this example,
> -	   we just accept that. A more rigorous filesystem may want
> -	   to return an error here */
> -	if (lo->writeback && (fi->flags & O_APPEND))
> -		fi->flags &= ~O_APPEND;
> -
> -	/*
> -	 * O_DIRECT in guest should not necessarily mean bypassing page
> -	 * cache on host as well. If somebody needs that behavior, it
> -	 * probably should be a configuration knob in daemon.
> -	 */
> -	fi->flags &= ~O_DIRECT;
> +	update_open_flags(lo->writeback, fi);
>  
>  	sprintf(buf, "%i", lo_fd(req, ino));
>  	fd = openat(lo->proc_self_fd, buf, fi->flags & ~O_NOFOLLOW);
> -- 
> 2.21.0
> 
> _______________________________________________
> Virtio-fs mailing list
> Virtio-fs@redhat.com
> https://www.redhat.com/mailman/listinfo/virtio-fs
Dr. David Alan Gilbert Nov. 1, 2019, 8:08 a.m. UTC | #8
* Vivek Goyal (vgoyal@redhat.com) wrote:
> On Wed, Oct 23, 2019 at 09:25:23PM +0900, Misono Tomohiro wrote:
> > When writeback mode is enabled (-o writeback), O_APPEND handling is
> > done in kernel. Therefore virtiofsd clears O_APPEND flag when open.
> > Otherwise O_APPEND flag takes precedence over pwrite() and write
> > data may corrupt.
> > 
> > Currently clearing O_APPEND flag is done in lo_open(), but we also
> > need the same operation in lo_create(). So, factor out the flag
> > update operation in lo_open() to update_open_flags() and call it
> > in both lo_open() and lo_create().
> > 
> > This fixes the failure of xfstest generic/069 in writeback mode
> > (which tests O_APPEND write data integrity).
> > 
> > Signed-off-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
> 
> Reviewed-by: Vivek Goyal <vgoyal@redhat.com>

Thanks, merged.

Dave

> Thanks
> Vivek
> 
> > ---
> >  contrib/virtiofsd/passthrough_ll.c | 56 +++++++++++++++---------------
> >  1 file changed, 28 insertions(+), 28 deletions(-)
> > 
> > diff --git a/contrib/virtiofsd/passthrough_ll.c b/contrib/virtiofsd/passthrough_ll.c
> > index e8892c3c32..79fb78ecce 100644
> > --- a/contrib/virtiofsd/passthrough_ll.c
> > +++ b/contrib/virtiofsd/passthrough_ll.c
> > @@ -1733,6 +1733,32 @@ static void lo_releasedir(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info
> >  	fuse_reply_err(req, 0);
> >  }
> >  
> > +static void update_open_flags(int writeback, struct fuse_file_info *fi)
> > +{
> > +	/* With writeback cache, kernel may send read requests even
> > +	   when userspace opened write-only */
> > +	if (writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> > +		fi->flags &= ~O_ACCMODE;
> > +		fi->flags |= O_RDWR;
> > +	}
> > +
> > +	/* With writeback cache, O_APPEND is handled by the kernel.
> > +	   This breaks atomicity (since the file may change in the
> > +	   underlying filesystem, so that the kernel's idea of the
> > +	   end of the file isn't accurate anymore). In this example,
> > +	   we just accept that. A more rigorous filesystem may want
> > +	   to return an error here */
> > +	if (writeback && (fi->flags & O_APPEND))
> > +		fi->flags &= ~O_APPEND;
> > +
> > +	/*
> > +	 * O_DIRECT in guest should not necessarily mean bypassing page
> > +	 * cache on host as well. If somebody needs that behavior, it
> > +	 * probably should be a configuration knob in daemon.
> > +	 */
> > +	fi->flags &= ~O_DIRECT;
> > +}
> > +
> >  static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
> >  		      mode_t mode, struct fuse_file_info *fi)
> >  {
> > @@ -1760,12 +1786,7 @@ static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
> >  	if (err)
> >  		goto out;
> >  
> > -	/*
> > -	 * O_DIRECT in guest should not necessarily mean bypassing page
> > -	 * cache on host as well. If somebody needs that behavior, it
> > -	 * probably should be a configuration knob in daemon.
> > -	 */
> > -	fi->flags &= ~O_DIRECT;
> > +	update_open_flags(lo->writeback, fi);
> >  
> >  	fd = openat(parent_inode->fd, name,
> >  		    (fi->flags | O_CREAT) & ~O_NOFOLLOW, mode);
> > @@ -1966,28 +1987,7 @@ static void lo_open(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info *fi)
> >  
> >  	fuse_log(FUSE_LOG_DEBUG, "lo_open(ino=%" PRIu64 ", flags=%d)\n", ino, fi->flags);
> >  
> > -	/* With writeback cache, kernel may send read requests even
> > -	   when userspace opened write-only */
> > -	if (lo->writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
> > -		fi->flags &= ~O_ACCMODE;
> > -		fi->flags |= O_RDWR;
> > -	}
> > -
> > -	/* With writeback cache, O_APPEND is handled by the kernel.
> > -	   This breaks atomicity (since the file may change in the
> > -	   underlying filesystem, so that the kernel's idea of the
> > -	   end of the file isn't accurate anymore). In this example,
> > -	   we just accept that. A more rigorous filesystem may want
> > -	   to return an error here */
> > -	if (lo->writeback && (fi->flags & O_APPEND))
> > -		fi->flags &= ~O_APPEND;
> > -
> > -	/*
> > -	 * O_DIRECT in guest should not necessarily mean bypassing page
> > -	 * cache on host as well. If somebody needs that behavior, it
> > -	 * probably should be a configuration knob in daemon.
> > -	 */
> > -	fi->flags &= ~O_DIRECT;
> > +	update_open_flags(lo->writeback, fi);
> >  
> >  	sprintf(buf, "%i", lo_fd(req, ino));
> >  	fd = openat(lo->proc_self_fd, buf, fi->flags & ~O_NOFOLLOW);
> > -- 
> > 2.21.0
> > 
> > _______________________________________________
> > Virtio-fs mailing list
> > Virtio-fs@redhat.com
> > https://www.redhat.com/mailman/listinfo/virtio-fs
> 
> _______________________________________________
> Virtio-fs mailing list
> Virtio-fs@redhat.com
> https://www.redhat.com/mailman/listinfo/virtio-fs
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

Patch
diff mbox series

diff --git a/contrib/virtiofsd/passthrough_ll.c b/contrib/virtiofsd/passthrough_ll.c
index e8892c3c32..79fb78ecce 100644
--- a/contrib/virtiofsd/passthrough_ll.c
+++ b/contrib/virtiofsd/passthrough_ll.c
@@ -1733,6 +1733,32 @@  static void lo_releasedir(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info
 	fuse_reply_err(req, 0);
 }
 
+static void update_open_flags(int writeback, struct fuse_file_info *fi)
+{
+	/* With writeback cache, kernel may send read requests even
+	   when userspace opened write-only */
+	if (writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
+		fi->flags &= ~O_ACCMODE;
+		fi->flags |= O_RDWR;
+	}
+
+	/* With writeback cache, O_APPEND is handled by the kernel.
+	   This breaks atomicity (since the file may change in the
+	   underlying filesystem, so that the kernel's idea of the
+	   end of the file isn't accurate anymore). In this example,
+	   we just accept that. A more rigorous filesystem may want
+	   to return an error here */
+	if (writeback && (fi->flags & O_APPEND))
+		fi->flags &= ~O_APPEND;
+
+	/*
+	 * O_DIRECT in guest should not necessarily mean bypassing page
+	 * cache on host as well. If somebody needs that behavior, it
+	 * probably should be a configuration knob in daemon.
+	 */
+	fi->flags &= ~O_DIRECT;
+}
+
 static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
 		      mode_t mode, struct fuse_file_info *fi)
 {
@@ -1760,12 +1786,7 @@  static void lo_create(fuse_req_t req, fuse_ino_t parent, const char *name,
 	if (err)
 		goto out;
 
-	/*
-	 * O_DIRECT in guest should not necessarily mean bypassing page
-	 * cache on host as well. If somebody needs that behavior, it
-	 * probably should be a configuration knob in daemon.
-	 */
-	fi->flags &= ~O_DIRECT;
+	update_open_flags(lo->writeback, fi);
 
 	fd = openat(parent_inode->fd, name,
 		    (fi->flags | O_CREAT) & ~O_NOFOLLOW, mode);
@@ -1966,28 +1987,7 @@  static void lo_open(fuse_req_t req, fuse_ino_t ino, struct fuse_file_info *fi)
 
 	fuse_log(FUSE_LOG_DEBUG, "lo_open(ino=%" PRIu64 ", flags=%d)\n", ino, fi->flags);
 
-	/* With writeback cache, kernel may send read requests even
-	   when userspace opened write-only */
-	if (lo->writeback && (fi->flags & O_ACCMODE) == O_WRONLY) {
-		fi->flags &= ~O_ACCMODE;
-		fi->flags |= O_RDWR;
-	}
-
-	/* With writeback cache, O_APPEND is handled by the kernel.
-	   This breaks atomicity (since the file may change in the
-	   underlying filesystem, so that the kernel's idea of the
-	   end of the file isn't accurate anymore). In this example,
-	   we just accept that. A more rigorous filesystem may want
-	   to return an error here */
-	if (lo->writeback && (fi->flags & O_APPEND))
-		fi->flags &= ~O_APPEND;
-
-	/*
-	 * O_DIRECT in guest should not necessarily mean bypassing page
-	 * cache on host as well. If somebody needs that behavior, it
-	 * probably should be a configuration knob in daemon.
-	 */
-	fi->flags &= ~O_DIRECT;
+	update_open_flags(lo->writeback, fi);
 
 	sprintf(buf, "%i", lo_fd(req, ino));
 	fd = openat(lo->proc_self_fd, buf, fi->flags & ~O_NOFOLLOW);