G' day
It looks like XFS broken somewhere in 2.6.11-rc1,
sadly i can't sand "right" bugreport, some facts only.
Upgrade to 2.6.11-rc2 makes fcron non-working for me in case of
crontabs directory is placed on XFS partition.
When i try to install new crontab fcrontab die with error:
"could not write() buf to disk: Resource temporarily unavailable"
gluk@qa gluk $ crontab test
20:10:57 installing file /home/gluk/test for user gluk
20:10:57 could not write() buf to disk: Resource temporarily unavailable
20:10:57 Since fcrontab has not been able to save new.gluk's file, it will
keep the previous version (if any) of new.gluk.
20:10:57 Error while copying file. Aborting.
The same time it works with 2.6.10. Some trick like
mount -o bind from non-xfs (reiserfs in my case) partition helps too.
some googling shows that similar problem took plase for 2.6.11-rc1
and postfix:
http://www.webservertalk.com/message879262.html
--
Best regards.
Alexander Y. Fomichev <[email protected]>
Public PGP key: http://sysadminday.org.ru/gluk.asc
On Tue, Feb 08, 2005 at 08:51:36PM +0300, Alexander Y. Fomichev wrote:
> G' day
>
> It looks like XFS broken somewhere in 2.6.11-rc1,
> sadly i can't sand "right" bugreport, some facts only.
> Upgrade to 2.6.11-rc2 makes fcron non-working for me in case of
> crontabs directory is placed on XFS partition.
> When i try to install new crontab fcrontab die with error:
> "could not write() buf to disk: Resource temporarily unavailable"
Is that an O_SYNC write, do you know? Or a write to an inode
with the sync flag set?
> The same time it works with 2.6.10.
I'm chasing down a problem similar to this atm, so far looks like
something in the generic VM code below sync_page_range is giving
back EAGAIN, and that is getting passed back out to userspace by
XFS. Not sure where/why/how its been caused yet though ... I'll
let you know once I have a fix or have found the culprit change.
cheers.
--
Nathan
On Wednesday 09 February 2005 04:29, Nathan Scott wrote:
> On Tue, Feb 08, 2005 at 08:51:36PM +0300, Alexander Y. Fomichev wrote:
> > G' day
> >
> > It looks like XFS broken somewhere in 2.6.11-rc1,
> > sadly i can't sand "right" bugreport, some facts only.
> > Upgrade to 2.6.11-rc2 makes fcron non-working for me in case of
> > crontabs directory is placed on XFS partition.
> > When i try to install new crontab fcrontab die with error:
> > "could not write() buf to disk: Resource temporarily unavailable"
>
> Is that an O_SYNC write, do you know? Or a write to an inode
> with the sync flag set?
Yes, it is O_SYNC, as i can see from fcron sources, and, no, kernel
have been compiled without xattrs support (if i understand
your question correctly)
>
> > The same time it works with 2.6.10.
>
> I'm chasing down a problem similar to this atm, so far looks like
> something in the generic VM code below sync_page_range is giving
> back EAGAIN, and that is getting passed back out to userspace by
> XFS. Not sure where/why/how its been caused yet though ... I'll
> let you know once I have a fix or have found the culprit change.
>
> cheers.
Tnx for quick answer.
PS: i forgot to mention last time i tested 2.6.11-rc3-bk5 with the
same results.
--
Best regards.
Alexander Y. Fomichev <[email protected]>
Public PGP key: http://sysadminday.org.ru/gluk.asc
On Wed, Feb 09, 2005 at 05:44:54PM +0300, Alexander Y. Fomichev wrote:
> On Wednesday 09 February 2005 04:29, Nathan Scott wrote:
> > Is that an O_SYNC write, do you know? Or a write to an inode
> > with the sync flag set?
>
> Yes, it is O_SYNC, as i can see from fcron sources, and, no, kernel
OK, thanks.
> > I'm chasing down a problem similar to this atm, so far looks like
> > something in the generic VM code below sync_page_range is giving
> > back EAGAIN, and that is getting passed back out to userspace by
> > XFS. Not sure where/why/how its been caused yet though ... I'll
> > let you know once I have a fix or have found the culprit change.
Turns out it was actually XFS giving back this EAGAIN, indirectly -
and some of the generic VM routines have been tweaked recently to
propogate more sync write errors out to userspace. Try this patch,
it will fix your problem - we're still discussing if this is the
ideal fix, so something else may be merged in the end.
cheers.
--
Nathan
Index: test/fs/xfs/linux-2.6/xfs_super.c
===================================================================
--- test.orig/fs/xfs/linux-2.6/xfs_super.c
+++ test/fs/xfs/linux-2.6/xfs_super.c
@@ -348,6 +348,12 @@
if (sync)
flags |= FLUSH_SYNC;
VOP_IFLUSH(vp, flags, error);
+ if (error == EAGAIN) {
+ if (sync)
+ VOP_IFLUSH(vp, flags | FLUSH_LOG, error);
+ else
+ error = 0;
+ }
}
return -error;
Index: test/fs/xfs/xfs_vnodeops.c
===================================================================
--- test.orig/fs/xfs/xfs_vnodeops.c
+++ test/fs/xfs/xfs_vnodeops.c
@@ -3681,27 +3681,27 @@
{
xfs_inode_t *ip;
xfs_mount_t *mp;
+ xfs_inode_log_item_t *iip;
int error = 0;
ip = XFS_BHVTOI(bdp);
mp = ip->i_mount;
+ iip = ip->i_itemp;
if (XFS_FORCED_SHUTDOWN(mp))
return XFS_ERROR(EIO);
- /* Bypass inodes which have already been cleaned by
+ /*
+ * Bypass inodes which have already been cleaned by
* the inode flush clustering code inside xfs_iflush
*/
if ((ip->i_update_core == 0) &&
- ((ip->i_itemp == NULL) ||
- !(ip->i_itemp->ili_format.ilf_fields & XFS_ILOG_ALL)))
+ ((iip == NULL) || !(iip->ili_format.ilf_fields & XFS_ILOG_ALL)))
return 0;
if (flags & FLUSH_LOG) {
- xfs_inode_log_item_t *iip = ip->i_itemp;
-
if (iip && iip->ili_last_lsn) {
- xlog_t *log = mp->m_log;
+ xlog_t *log = mp->m_log;
xfs_lsn_t sync_lsn;
int s, log_flags = XFS_LOG_FORCE;
@@ -3714,12 +3714,14 @@
if (flags & FLUSH_SYNC)
log_flags |= XFS_LOG_SYNC;
- return xfs_log_force(mp, iip->ili_last_lsn,
- log_flags);
+ error = xfs_log_force(mp, iip->ili_last_lsn, log_flags);
+ if (error)
+ return error;
}
}
- /* We make this non-blocking if the inode is contended,
+ /*
+ * We make this non-blocking if the inode is contended,
* return EAGAIN to indicate to the caller that they
* did not succeed. This prevents the flush path from
* blocking on inodes inside another operation right
@@ -3728,8 +3730,11 @@
if (flags & FLUSH_INODE) {
int flush_flags;
+ if (!(flags & FLUSH_LOG))
+ error = EAGAIN;
+
if (xfs_ipincount(ip))
- return EAGAIN;
+ return error;
if (flags & FLUSH_SYNC) {
xfs_ilock(ip, XFS_ILOCK_SHARED);
@@ -3737,10 +3742,10 @@
} else if (xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
if (xfs_ipincount(ip) || !xfs_iflock_nowait(ip)) {
xfs_iunlock(ip, XFS_ILOCK_SHARED);
- return EAGAIN;
+ return error;
}
} else {
- return EAGAIN;
+ return error;
}
if (flags & FLUSH_SYNC)
Index: test/fs/xfs/linux-2.6/xfs_lrw.c
===================================================================
--- test.orig/fs/xfs/linux-2.6/xfs_lrw.c
+++ test/fs/xfs/linux-2.6/xfs_lrw.c
@@ -962,9 +962,9 @@
xfs_trans_set_sync(tp);
error = xfs_trans_commit(tp, 0, NULL);
xfs_iunlock(xip, XFS_ILOCK_EXCL);
- if (error)
- goto out_unlock_internal;
}
+ if (error)
+ goto out_unlock_internal;
}
xfs_rwunlock(bdp, locktype);
On Thursday 10 February 2005 07:54, Nathan Scott wrote:
> On Wed, Feb 09, 2005 at 05:44:54PM +0300, Alexander Y. Fomichev wrote:
> > On Wednesday 09 February 2005 04:29, Nathan Scott wrote:
> > > Is that an O_SYNC write, do you know? Or a write to an inode
> > > with the sync flag set?
> >
> > Yes, it is O_SYNC, as i can see from fcron sources, and, no, kernel
>
> OK, thanks.
>
> > > I'm chasing down a problem similar to this atm, so far looks like
> > > something in the generic VM code below sync_page_range is giving
> > > back EAGAIN, and that is getting passed back out to userspace by
> > > XFS. Not sure where/why/how its been caused yet though ... I'll
> > > let you know once I have a fix or have found the culprit change.
>
> Turns out it was actually XFS giving back this EAGAIN, indirectly -
> and some of the generic VM routines have been tweaked recently to
> propogate more sync write errors out to userspace. Try this patch,
> it will fix your problem - we're still discussing if this is the
> ideal fix, so something else may be merged in the end.
>
> cheers.
Yes, it works. Thank you for quick patch.
--
Best regards.
Alexander Y. Fomichev <[email protected]>
Public PGP key: http://sysadminday.org.ru/gluk.asc
Hi Nathan
I can also confirm that this patch resolves an issue I am seeing
with re-aim-7 writing to xfs fs mounted on ramdisk, I was also
getting EAGAIN.
Thanks
Darren
On Thu, 10 Feb 2005, Nathan Scott wrote:
> On Wed, Feb 09, 2005 at 05:44:54PM +0300, Alexander Y. Fomichev wrote:
> > On Wednesday 09 February 2005 04:29, Nathan Scott wrote:
> > > Is that an O_SYNC write, do you know? Or a write to an inode
> > > with the sync flag set?
> >
> > Yes, it is O_SYNC, as i can see from fcron sources, and, no, kernel
>
> OK, thanks.
>
> > > I'm chasing down a problem similar to this atm, so far looks like
> > > something in the generic VM code below sync_page_range is giving
> > > back EAGAIN, and that is getting passed back out to userspace by
> > > XFS. Not sure where/why/how its been caused yet though ... I'll
> > > let you know once I have a fix or have found the culprit change.
>
> Turns out it was actually XFS giving back this EAGAIN, indirectly -
> and some of the generic VM routines have been tweaked recently to
> propogate more sync write errors out to userspace. Try this patch,
> it will fix your problem - we're still discussing if this is the
> ideal fix, so something else may be merged in the end.
>
> cheers.
>
> --
> Nathan
>
>
> Index: test/fs/xfs/linux-2.6/xfs_super.c
> ===================================================================
> --- test.orig/fs/xfs/linux-2.6/xfs_super.c
> +++ test/fs/xfs/linux-2.6/xfs_super.c
> @@ -348,6 +348,12 @@
> if (sync)
> flags |= FLUSH_SYNC;
> VOP_IFLUSH(vp, flags, error);
> + if (error == EAGAIN) {
> + if (sync)
> + VOP_IFLUSH(vp, flags | FLUSH_LOG, error);
> + else
> + error = 0;
> + }
> }
>
> return -error;
> Index: test/fs/xfs/xfs_vnodeops.c
> ===================================================================
> --- test.orig/fs/xfs/xfs_vnodeops.c
> +++ test/fs/xfs/xfs_vnodeops.c
> @@ -3681,27 +3681,27 @@
> {
> xfs_inode_t *ip;
> xfs_mount_t *mp;
> + xfs_inode_log_item_t *iip;
> int error = 0;
>
> ip = XFS_BHVTOI(bdp);
> mp = ip->i_mount;
> + iip = ip->i_itemp;
>
> if (XFS_FORCED_SHUTDOWN(mp))
> return XFS_ERROR(EIO);
>
> - /* Bypass inodes which have already been cleaned by
> + /*
> + * Bypass inodes which have already been cleaned by
> * the inode flush clustering code inside xfs_iflush
> */
> if ((ip->i_update_core == 0) &&
> - ((ip->i_itemp == NULL) ||
> - !(ip->i_itemp->ili_format.ilf_fields & XFS_ILOG_ALL)))
> + ((iip == NULL) || !(iip->ili_format.ilf_fields & XFS_ILOG_ALL)))
> return 0;
>
> if (flags & FLUSH_LOG) {
> - xfs_inode_log_item_t *iip = ip->i_itemp;
> -
> if (iip && iip->ili_last_lsn) {
> - xlog_t *log = mp->m_log;
> + xlog_t *log = mp->m_log;
> xfs_lsn_t sync_lsn;
> int s, log_flags = XFS_LOG_FORCE;
>
> @@ -3714,12 +3714,14 @@
>
> if (flags & FLUSH_SYNC)
> log_flags |= XFS_LOG_SYNC;
> - return xfs_log_force(mp, iip->ili_last_lsn,
> - log_flags);
> + error = xfs_log_force(mp, iip->ili_last_lsn, log_flags);
> + if (error)
> + return error;
> }
> }
>
> - /* We make this non-blocking if the inode is contended,
> + /*
> + * We make this non-blocking if the inode is contended,
> * return EAGAIN to indicate to the caller that they
> * did not succeed. This prevents the flush path from
> * blocking on inodes inside another operation right
> @@ -3728,8 +3730,11 @@
> if (flags & FLUSH_INODE) {
> int flush_flags;
>
> + if (!(flags & FLUSH_LOG))
> + error = EAGAIN;
> +
> if (xfs_ipincount(ip))
> - return EAGAIN;
> + return error;
>
> if (flags & FLUSH_SYNC) {
> xfs_ilock(ip, XFS_ILOCK_SHARED);
> @@ -3737,10 +3742,10 @@
> } else if (xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
> if (xfs_ipincount(ip) || !xfs_iflock_nowait(ip)) {
> xfs_iunlock(ip, XFS_ILOCK_SHARED);
> - return EAGAIN;
> + return error;
> }
> } else {
> - return EAGAIN;
> + return error;
> }
>
> if (flags & FLUSH_SYNC)
> Index: test/fs/xfs/linux-2.6/xfs_lrw.c
> ===================================================================
> --- test.orig/fs/xfs/linux-2.6/xfs_lrw.c
> +++ test/fs/xfs/linux-2.6/xfs_lrw.c
> @@ -962,9 +962,9 @@
> xfs_trans_set_sync(tp);
> error = xfs_trans_commit(tp, 0, NULL);
> xfs_iunlock(xip, XFS_ILOCK_EXCL);
> - if (error)
> - goto out_unlock_internal;
> }
> + if (error)
> + goto out_unlock_internal;
> }
>
> xfs_rwunlock(bdp, locktype);
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
--------------------------------------------------
Darren Williams <dsw AT gelato.unsw.edu.au>
Gelato@UNSW <http://www.gelato.unsw.edu.au>
--------------------------------------------------