Re: warning in ext4_journal_start_sb on filesystem freeze

Re: warning in ext4_journal_start_sb on filesystem freeze

[Jan Kara]

On Sat 22-02-14 09:50:06, Matthew Rahtz wrote:
> Thanks for your help Jan,
> 
> A few months later, we've noticed the issue is actually still there.
> Using 3.11.0-17-generic on Ubuntu 12.04, we’re seeing this in the kernel
> logs:
> 
> [29243.606215] WARNING: CPU: 0 PID: 1785 at
> /build/buildd/linux-lts-saucy-3.11.0/fs/ext4/ext4_jbd2.c:48
> ext4_journal_check_start+0x83/0x90()
> 
> Having a look at the Ubuntu source package for that version, it
> definitely does include commit 03d95eb2f2578083a3f6286262e1cb5d88a00c02,
> and the line generating the warning is still:
> 
> WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
> 
> Are there any other obvious possibilities for what may be causing this?
> There seem to be some users of Oracle Linux experiencing similar problems
> at https://community.oracle.com/thread/2617418, which was apparently
> fixed in Oracle's kernel version '3.8.13-26.el6uek'. Any word on when
> this might be integrated into the official kernel?
> 
> Full call trace included below.
  Looking at the trace below, now the problem seems to be in the NFS server
code. NFS should get protection against the filesystem being frozen (or
remounted read-only for that matter) via mnt_want_write() before calling
into notify_change() (actually before calling fh_lock() because of lock
ordering).  Similarly to what we do e.g. in fchownat(). Bruce?

								Honza

> [29243.606212] ------------[ cut here ]------------
> [29243.606215] WARNING: CPU: 0 PID: 1785 at /build/buildd/linux-lts-saucy-3.11.0/fs/ext4/ext4_jbd2.c:48 ext4_journal_check_start+0x83/0x90()
> [29243.606216] Modules linked in: parport_pc ppdev nfsd nfs_acl auth_rpcgss nfs fscache lockd sunrpc ext2 cirrus ttm drm_kms_helper drm sysimgblt psmouse i2c_piix4 virtio_balloon sysfillrect mac_hid serio_raw syscopyarea virtio_console lp parport floppy
> [29243.606227] CPU: 0 PID: 1785 Comm: nfsd Tainted: G        W    3.11.0-17-generic #31~precise1-Ubuntu
> [29243.606228] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007
> [29243.606228]  0000000000000030 ffff8801162f3b08 ffffffff8173c72d 0000000000000007
> [29243.606230]  0000000000000000 ffff8801162f3b48 ffffffff8106540c 0000000000000000
> [29243.606232]  ffff880114892800 0000000000000007 0000000000000068 0000000000000000
> [29243.606235] Call Trace:
> [29243.606237]  [<ffffffff8173c72d>] dump_stack+0x46/0x58
> [29243.606239]  [<ffffffff8106540c>] warn_slowpath_common+0x8c/0xc0
> [29243.606241]  [<ffffffff8106545a>] warn_slowpath_null+0x1a/0x20
> [29243.606244]  [<ffffffff8127ebb3>] ext4_journal_check_start+0x83/0x90
> [29243.606246]  [<ffffffff8127ec35>] __ext4_journal_start_sb+0x45/0x100
> [29243.606249]  [<ffffffff81258a03>] ? ext4_dirty_inode+0x33/0x70
> [29243.606251]  [<ffffffff81258a03>] ext4_dirty_inode+0x33/0x70
> [29243.606254]  [<ffffffff811de348>] __mark_inode_dirty+0x48/0x350
> [29243.606256]  [<ffffffff81256b53>] ext4_setattr+0x1b3/0x5b0
> [29243.606259]  [<ffffffff811d0903>] notify_change+0x1d3/0x390
> [29243.606263]  [<ffffffffa01c7fe2>] nfsd_setattr+0x232/0x2a0 [nfsd]
> [29243.606267]  [<ffffffffa01d00f6>] nfsd3_proc_setattr+0x76/0xc0 [nfsd]
> [29243.606271]  [<ffffffffa01c0d85>] nfsd_dispatch+0xe5/0x230 [nfsd]
> [29243.606283]  [<ffffffffa0128465>] svc_process_common+0x345/0x680 [sunrpc]
> [29243.606289]  [<ffffffffa0128af3>] svc_process+0x103/0x160 [sunrpc]
> [29243.606293]  [<ffffffffa01c08df>] nfsd+0xbf/0x130 [nfsd]
> [29243.606297]  [<ffffffffa01c0820>] ? nfsd_destroy+0x80/0x80 [nfsd]
> [29243.606299]  [<ffffffff81089170>] kthread+0xc0/0xd0
> [29243.606302]  [<ffffffff810890b0>] ? flush_kthread_worker+0xb0/0xb0
> [29243.606304]  [<ffffffff8175122c>] ret_from_fork+0x7c/0xb0
> [29243.606307]  [<ffffffff810890b0>] ? flush_kthread_worker+0xb0/0xb0
> [29243.606308] ---[ end trace e9d4726f92c62d43 ]---
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR

Re: warning in ext4_journal_start_sb on filesystem freeze

[J. Bruce Fields]

On Mon, Feb 24, 2014 at 10:55:25AM +0100, Jan Kara wrote:
> On Sat 22-02-14 09:50:06, Matthew Rahtz wrote:
> > Thanks for your help Jan,
> > 
> > A few months later, we've noticed the issue is actually still there.
> > Using 3.11.0-17-generic on Ubuntu 12.04, we’re seeing this in the kernel
> > logs:
> > 
> > [29243.606215] WARNING: CPU: 0 PID: 1785 at
> > /build/buildd/linux-lts-saucy-3.11.0/fs/ext4/ext4_jbd2.c:48
> > ext4_journal_check_start+0x83/0x90()
> > 
> > Having a look at the Ubuntu source package for that version, it
> > definitely does include commit 03d95eb2f2578083a3f6286262e1cb5d88a00c02,
> > and the line generating the warning is still:
> > 
> > WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
> > 
> > Are there any other obvious possibilities for what may be causing this?
> > There seem to be some users of Oracle Linux experiencing similar problems
> > at https://community.oracle.com/thread/2617418, which was apparently
> > fixed in Oracle's kernel version '3.8.13-26.el6uek'. Any word on when
> > this might be integrated into the official kernel?
> > 
> > Full call trace included below.
>   Looking at the trace below, now the problem seems to be in the NFS server
> code. NFS should get protection against the filesystem being frozen (or
> remounted read-only for that matter) via mnt_want_write() before calling
> into notify_change() (actually before calling fh_lock() because of lock
> ordering).  Similarly to what we do e.g. in fchownat(). Bruce?

Like this?

But I wonder why this is just popping up now--as far as I can tell we've
had the bug since those write counts were introduced.

--b.

diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
index 6d7be3f..d573b61 100644
--- a/fs/nfsd/vfs.c
+++ b/fs/nfsd/vfs.c
@@ -445,12 +445,16 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
 		err = nfserr_notsync;
 		goto out_put_write_access;
 	}
+	host_err = fh_want_write(fhp);
+	if (host_err)
+		goto out_nfserr;
 
 	fh_lock(fhp);
 	host_err = notify_change(dentry, iap, NULL);
 	fh_unlock(fhp);
+	fh_drop_write(fhp);
+out_nfserr:
 	err = nfserrno(host_err);
-
 out_put_write_access:
 	if (size_change)
 		put_write_access(inode);

Re: warning in ext4_journal_start_sb on filesystem freeze

[Jan Kara]

On Mon 24-02-14 10:45:32, J. Bruce Fields wrote:
> On Mon, Feb 24, 2014 at 10:55:25AM +0100, Jan Kara wrote:
> > On Sat 22-02-14 09:50:06, Matthew Rahtz wrote:
> > > Thanks for your help Jan,
> > > 
> > > A few months later, we've noticed the issue is actually still there.
> > > Using 3.11.0-17-generic on Ubuntu 12.04, we’re seeing this in the kernel
> > > logs:
> > > 
> > > [29243.606215] WARNING: CPU: 0 PID: 1785 at
> > > /build/buildd/linux-lts-saucy-3.11.0/fs/ext4/ext4_jbd2.c:48
> > > ext4_journal_check_start+0x83/0x90()
> > > 
> > > Having a look at the Ubuntu source package for that version, it
> > > definitely does include commit 03d95eb2f2578083a3f6286262e1cb5d88a00c02,
> > > and the line generating the warning is still:
> > > 
> > > WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
> > > 
> > > Are there any other obvious possibilities for what may be causing this?
> > > There seem to be some users of Oracle Linux experiencing similar problems
> > > at https://community.oracle.com/thread/2617418, which was apparently
> > > fixed in Oracle's kernel version '3.8.13-26.el6uek'. Any word on when
> > > this might be integrated into the official kernel?
> > > 
> > > Full call trace included below.
> >   Looking at the trace below, now the problem seems to be in the NFS server
> > code. NFS should get protection against the filesystem being frozen (or
> > remounted read-only for that matter) via mnt_want_write() before calling
> > into notify_change() (actually before calling fh_lock() because of lock
> > ordering).  Similarly to what we do e.g. in fchownat(). Bruce?
> 
> Like this?
  Yup, that looks right.

> But I wonder why this is just popping up now--as far as I can tell we've
> had the bug since those write counts were introduced.
  Yeah, I'm wondering as well. NFS server on ext4 should have been
complaining for a long time.

								Honza

> diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
> index 6d7be3f..d573b61 100644
> --- a/fs/nfsd/vfs.c
> +++ b/fs/nfsd/vfs.c
> @@ -445,12 +445,16 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
>  		err = nfserr_notsync;
>  		goto out_put_write_access;
>  	}
> +	host_err = fh_want_write(fhp);
> +	if (host_err)
> +		goto out_nfserr;
>  
>  	fh_lock(fhp);
>  	host_err = notify_change(dentry, iap, NULL);
>  	fh_unlock(fhp);
> +	fh_drop_write(fhp);
> +out_nfserr:
>  	err = nfserrno(host_err);
> -
>  out_put_write_access:
>  	if (size_change)
>  		put_write_access(inode);
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR

Re: warning in ext4_journal_start_sb on filesystem freeze

[J. Bruce Fields]

On Tue, Feb 25, 2014 at 11:21:26AM +0100, Jan Kara wrote:
> On Mon 24-02-14 10:45:32, J. Bruce Fields wrote:
> > On Mon, Feb 24, 2014 at 10:55:25AM +0100, Jan Kara wrote:
> > > On Sat 22-02-14 09:50:06, Matthew Rahtz wrote:
> > > > Thanks for your help Jan,
> > > > 
> > > > A few months later, we've noticed the issue is actually still there.
> > > > Using 3.11.0-17-generic on Ubuntu 12.04, we’re seeing this in the kernel
> > > > logs:
> > > > 
> > > > [29243.606215] WARNING: CPU: 0 PID: 1785 at
> > > > /build/buildd/linux-lts-saucy-3.11.0/fs/ext4/ext4_jbd2.c:48
> > > > ext4_journal_check_start+0x83/0x90()
> > > > 
> > > > Having a look at the Ubuntu source package for that version, it
> > > > definitely does include commit 03d95eb2f2578083a3f6286262e1cb5d88a00c02,
> > > > and the line generating the warning is still:
> > > > 
> > > > WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
> > > > 
> > > > Are there any other obvious possibilities for what may be causing this?
> > > > There seem to be some users of Oracle Linux experiencing similar problems
> > > > at https://community.oracle.com/thread/2617418, which was apparently
> > > > fixed in Oracle's kernel version '3.8.13-26.el6uek'. Any word on when
> > > > this might be integrated into the official kernel?
> > > > 
> > > > Full call trace included below.
> > >   Looking at the trace below, now the problem seems to be in the NFS server
> > > code. NFS should get protection against the filesystem being frozen (or
> > > remounted read-only for that matter) via mnt_want_write() before calling
> > > into notify_change() (actually before calling fh_lock() because of lock
> > > ordering).  Similarly to what we do e.g. in fchownat(). Bruce?
> > 
> > Like this?
>   Yup, that looks right.

Ugh, actually, I didn't realize we can't do mnt_want_write recursively,
and there's a confusing mixture of callers that do and don't already
take it, so I'll have to do something a little more complicated.

Oh well.--b.

> 
> > But I wonder why this is just popping up now--as far as I can tell we've
> > had the bug since those write counts were introduced.
>   Yeah, I'm wondering as well. NFS server on ext4 should have been
> complaining for a long time.
> 
> 								Honza
> 
> > diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
> > index 6d7be3f..d573b61 100644
> > --- a/fs/nfsd/vfs.c
> > +++ b/fs/nfsd/vfs.c
> > @@ -445,12 +445,16 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
> >  		err = nfserr_notsync;
> >  		goto out_put_write_access;
> >  	}
> > +	host_err = fh_want_write(fhp);
> > +	if (host_err)
> > +		goto out_nfserr;
> >  
> >  	fh_lock(fhp);
> >  	host_err = notify_change(dentry, iap, NULL);
> >  	fh_unlock(fhp);
> > +	fh_drop_write(fhp);
> > +out_nfserr:
> >  	err = nfserrno(host_err);
> > -
> >  out_put_write_access:
> >  	if (size_change)
> >  		put_write_access(inode);
> -- 
> Jan Kara <jack@suse.cz>
> SUSE Labs, CR

Re: warning in ext4_journal_start_sb on filesystem freeze

[J. Bruce Fields]

On Tue, Mar 04, 2014 at 11:43:06AM -0500, J. Bruce Fields wrote:
> On Tue, Feb 25, 2014 at 11:21:26AM +0100, Jan Kara wrote:
> > On Mon 24-02-14 10:45:32, J. Bruce Fields wrote:
> > > On Mon, Feb 24, 2014 at 10:55:25AM +0100, Jan Kara wrote:
> > > > On Sat 22-02-14 09:50:06, Matthew Rahtz wrote:
> > > > > Thanks for your help Jan,
> > > > > 
> > > > > A few months later, we've noticed the issue is actually still there.
> > > > > Using 3.11.0-17-generic on Ubuntu 12.04, we’re seeing this in the kernel
> > > > > logs:
> > > > > 
> > > > > [29243.606215] WARNING: CPU: 0 PID: 1785 at
> > > > > /build/buildd/linux-lts-saucy-3.11.0/fs/ext4/ext4_jbd2.c:48
> > > > > ext4_journal_check_start+0x83/0x90()
> > > > > 
> > > > > Having a look at the Ubuntu source package for that version, it
> > > > > definitely does include commit 03d95eb2f2578083a3f6286262e1cb5d88a00c02,
> > > > > and the line generating the warning is still:
> > > > > 
> > > > > WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
> > > > > 
> > > > > Are there any other obvious possibilities for what may be causing this?
> > > > > There seem to be some users of Oracle Linux experiencing similar problems
> > > > > at https://community.oracle.com/thread/2617418, which was apparently
> > > > > fixed in Oracle's kernel version '3.8.13-26.el6uek'. Any word on when
> > > > > this might be integrated into the official kernel?
> > > > > 
> > > > > Full call trace included below.
> > > >   Looking at the trace below, now the problem seems to be in the NFS server
> > > > code. NFS should get protection against the filesystem being frozen (or
> > > > remounted read-only for that matter) via mnt_want_write() before calling
> > > > into notify_change() (actually before calling fh_lock() because of lock
> > > > ordering).  Similarly to what we do e.g. in fchownat(). Bruce?
> > > 
> > > Like this?
> >   Yup, that looks right.
> 
> Ugh, actually, I didn't realize we can't do mnt_want_write recursively,
> and there's a confusing mixture of callers that do and don't already
> take it, so I'll have to do something a little more complicated.

Actually it looks like there's an easy enough way to distinguish when we
need mnt_want_write and when we don't; hopefully the following does the
job.

--b.

commit b0f5cd115e811a146a6e1a4dd1e7cb85808cca23
Author: J. Bruce Fields <bfields@redhat.com>
Date:   Mon Feb 24 14:59:47 2014 -0500

    nfsd: notify_change needs elevated write count
    
    Looks like this bug has been here since these write counts were
    introduced, not sure why it was just noticed now.
    
    Thanks also to Jan Kara for pointing out the problem.
    
    Reported-by: Matthew Rahtz <mrahtz@rapitasystems.com>
    Signed-off-by: J. Bruce Fields <bfields@redhat.com>

diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
index 6d7be3f..eea5ad1 100644
--- a/fs/nfsd/vfs.c
+++ b/fs/nfsd/vfs.c
@@ -404,6 +404,7 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
 	umode_t		ftype = 0;
 	__be32		err;
 	int		host_err;
+	bool		get_write_count;
 	int		size_change = 0;
 
 	if (iap->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_SIZE))
@@ -411,10 +412,18 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
 	if (iap->ia_valid & ATTR_SIZE)
 		ftype = S_IFREG;
 
+	/* Callers that do fh_verify should do the fh_want_write: */
+	get_write_count = !fhp->fh_dentry;
+
 	/* Get inode */
 	err = fh_verify(rqstp, fhp, ftype, accmode);
 	if (err)
 		goto out;
+	if (get_write_count) {
+		host_err = fh_want_write(fhp);
+		if (host_err)
+			return nfserrno(host_err);
+	}
 
 	dentry = fhp->fh_dentry;
 	inode = dentry->d_inode;

Re: warning in ext4_journal_start_sb on filesystem freeze

[Matthew Rahtz]

Brilliant :) Thank you for your work!

----- Original Message -----
From: "J. Bruce Fields" <bfields@fieldses.org>
To: "Jan Kara" <jack@suse.cz>
Cc: "Matthew Rahtz" <mrahtz@rapitasystems.com>, linux-ext4@vger.kernel.org, linux-nfs@vger.kernel.org
Sent: Tuesday, 4 March, 2014 7:04:42 PM
Subject: Re: warning in ext4_journal_start_sb on filesystem freeze

On Tue, Mar 04, 2014 at 11:43:06AM -0500, J. Bruce Fields wrote:
> On Tue, Feb 25, 2014 at 11:21:26AM +0100, Jan Kara wrote:
> > On Mon 24-02-14 10:45:32, J. Bruce Fields wrote:
> > > On Mon, Feb 24, 2014 at 10:55:25AM +0100, Jan Kara wrote:
> > > > On Sat 22-02-14 09:50:06, Matthew Rahtz wrote:
> > > > > Thanks for your help Jan,
> > > > > 
> > > > > A few months later, we've noticed the issue is actually still there.
> > > > > Using 3.11.0-17-generic on Ubuntu 12.04, we’re seeing this in the kernel
> > > > > logs:
> > > > > 
> > > > > [29243.606215] WARNING: CPU: 0 PID: 1785 at
> > > > > /build/buildd/linux-lts-saucy-3.11.0/fs/ext4/ext4_jbd2.c:48
> > > > > ext4_journal_check_start+0x83/0x90()
> > > > > 
> > > > > Having a look at the Ubuntu source package for that version, it
> > > > > definitely does include commit 03d95eb2f2578083a3f6286262e1cb5d88a00c02,
> > > > > and the line generating the warning is still:
> > > > > 
> > > > > WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
> > > > > 
> > > > > Are there any other obvious possibilities for what may be causing this?
> > > > > There seem to be some users of Oracle Linux experiencing similar problems
> > > > > at https://community.oracle.com/thread/2617418, which was apparently
> > > > > fixed in Oracle's kernel version '3.8.13-26.el6uek'. Any word on when
> > > > > this might be integrated into the official kernel?
> > > > > 
> > > > > Full call trace included below.
> > > >   Looking at the trace below, now the problem seems to be in the NFS server
> > > > code. NFS should get protection against the filesystem being frozen (or
> > > > remounted read-only for that matter) via mnt_want_write() before calling
> > > > into notify_change() (actually before calling fh_lock() because of lock
> > > > ordering).  Similarly to what we do e.g. in fchownat(). Bruce?
> > > 
> > > Like this?
> >   Yup, that looks right.
> 
> Ugh, actually, I didn't realize we can't do mnt_want_write recursively,
> and there's a confusing mixture of callers that do and don't already
> take it, so I'll have to do something a little more complicated.

Actually it looks like there's an easy enough way to distinguish when we
need mnt_want_write and when we don't; hopefully the following does the
job.

--b.

commit b0f5cd115e811a146a6e1a4dd1e7cb85808cca23
Author: J. Bruce Fields <bfields@redhat.com>
Date:   Mon Feb 24 14:59:47 2014 -0500

    nfsd: notify_change needs elevated write count
    
    Looks like this bug has been here since these write counts were
    introduced, not sure why it was just noticed now.
    
    Thanks also to Jan Kara for pointing out the problem.
    
    Reported-by: Matthew Rahtz <mrahtz@rapitasystems.com>
    Signed-off-by: J. Bruce Fields <bfields@redhat.com>

diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
index 6d7be3f..eea5ad1 100644
--- a/fs/nfsd/vfs.c
+++ b/fs/nfsd/vfs.c
@@ -404,6 +404,7 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
 	umode_t		ftype = 0;
 	__be32		err;
 	int		host_err;
+	bool		get_write_count;
 	int		size_change = 0;
 
 	if (iap->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_SIZE))
@@ -411,10 +412,18 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
 	if (iap->ia_valid & ATTR_SIZE)
 		ftype = S_IFREG;
 
+	/* Callers that do fh_verify should do the fh_want_write: */
+	get_write_count = !fhp->fh_dentry;
+
 	/* Get inode */
 	err = fh_verify(rqstp, fhp, ftype, accmode);
 	if (err)
 		goto out;
+	if (get_write_count) {
+		host_err = fh_want_write(fhp);
+		if (host_err)
+			return nfserrno(host_err);
+	}
 
 	dentry = fhp->fh_dentry;
 	inode = dentry->d_inode;
Please Note: Rapita Systems has a new address and telephone number.
Telephone: +44 1904 413945
Address: Rapita Systems Ltd, Atlas House,
          Osbaldwick Link Road, YORK, YO10 3JB
          United Kingdom

Re: warning in ext4_journal_start_sb on filesystem freeze

[J. Bruce Fields]

On Sat, Mar 08, 2014 at 09:02:26AM +0000, Matthew Rahtz wrote:
> Brilliant :) Thank you for your work!

Just to make sure, have you been able to confirm yet that this
eliminates the warning you were seeing?

--b.

> 
> ----- Original Message -----
> From: "J. Bruce Fields" <bfields@fieldses.org>
> To: "Jan Kara" <jack@suse.cz>
> Cc: "Matthew Rahtz" <mrahtz@rapitasystems.com>, linux-ext4@vger.kernel.org, linux-nfs@vger.kernel.org
> Sent: Tuesday, 4 March, 2014 7:04:42 PM
> Subject: Re: warning in ext4_journal_start_sb on filesystem freeze
> 
> On Tue, Mar 04, 2014 at 11:43:06AM -0500, J. Bruce Fields wrote:
> > On Tue, Feb 25, 2014 at 11:21:26AM +0100, Jan Kara wrote:
> > > On Mon 24-02-14 10:45:32, J. Bruce Fields wrote:
> > > > On Mon, Feb 24, 2014 at 10:55:25AM +0100, Jan Kara wrote:
> > > > > On Sat 22-02-14 09:50:06, Matthew Rahtz wrote:
> > > > > > Thanks for your help Jan,
> > > > > > 
> > > > > > A few months later, we've noticed the issue is actually still there.
> > > > > > Using 3.11.0-17-generic on Ubuntu 12.04, we’re seeing this in the kernel
> > > > > > logs:
> > > > > > 
> > > > > > [29243.606215] WARNING: CPU: 0 PID: 1785 at
> > > > > > /build/buildd/linux-lts-saucy-3.11.0/fs/ext4/ext4_jbd2.c:48
> > > > > > ext4_journal_check_start+0x83/0x90()
> > > > > > 
> > > > > > Having a look at the Ubuntu source package for that version, it
> > > > > > definitely does include commit 03d95eb2f2578083a3f6286262e1cb5d88a00c02,
> > > > > > and the line generating the warning is still:
> > > > > > 
> > > > > > WARN_ON(sb->s_writers.frozen == SB_FREEZE_COMPLETE);
> > > > > > 
> > > > > > Are there any other obvious possibilities for what may be causing this?
> > > > > > There seem to be some users of Oracle Linux experiencing similar problems
> > > > > > at https://community.oracle.com/thread/2617418, which was apparently
> > > > > > fixed in Oracle's kernel version '3.8.13-26.el6uek'. Any word on when
> > > > > > this might be integrated into the official kernel?
> > > > > > 
> > > > > > Full call trace included below.
> > > > >   Looking at the trace below, now the problem seems to be in the NFS server
> > > > > code. NFS should get protection against the filesystem being frozen (or
> > > > > remounted read-only for that matter) via mnt_want_write() before calling
> > > > > into notify_change() (actually before calling fh_lock() because of lock
> > > > > ordering).  Similarly to what we do e.g. in fchownat(). Bruce?
> > > > 
> > > > Like this?
> > >   Yup, that looks right.
> > 
> > Ugh, actually, I didn't realize we can't do mnt_want_write recursively,
> > and there's a confusing mixture of callers that do and don't already
> > take it, so I'll have to do something a little more complicated.
> 
> Actually it looks like there's an easy enough way to distinguish when we
> need mnt_want_write and when we don't; hopefully the following does the
> job.
> 
> --b.
> 
> commit b0f5cd115e811a146a6e1a4dd1e7cb85808cca23
> Author: J. Bruce Fields <bfields@redhat.com>
> Date:   Mon Feb 24 14:59:47 2014 -0500
> 
>     nfsd: notify_change needs elevated write count
>     
>     Looks like this bug has been here since these write counts were
>     introduced, not sure why it was just noticed now.
>     
>     Thanks also to Jan Kara for pointing out the problem.
>     
>     Reported-by: Matthew Rahtz <mrahtz@rapitasystems.com>
>     Signed-off-by: J. Bruce Fields <bfields@redhat.com>
> 
> diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
> index 6d7be3f..eea5ad1 100644
> --- a/fs/nfsd/vfs.c
> +++ b/fs/nfsd/vfs.c
> @@ -404,6 +404,7 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
>  	umode_t		ftype = 0;
>  	__be32		err;
>  	int		host_err;
> +	bool		get_write_count;
>  	int		size_change = 0;
>  
>  	if (iap->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_SIZE))
> @@ -411,10 +412,18 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
>  	if (iap->ia_valid & ATTR_SIZE)
>  		ftype = S_IFREG;
>  
> +	/* Callers that do fh_verify should do the fh_want_write: */
> +	get_write_count = !fhp->fh_dentry;
> +
>  	/* Get inode */
>  	err = fh_verify(rqstp, fhp, ftype, accmode);
>  	if (err)
>  		goto out;
> +	if (get_write_count) {
> +		host_err = fh_want_write(fhp);
> +		if (host_err)
> +			return nfserrno(host_err);
> +	}
>  
>  	dentry = fhp->fh_dentry;
>  	inode = dentry->d_inode;
> Please Note: Rapita Systems has a new address and telephone number.
> Telephone: +44 1904 413945
> Address: Rapita Systems Ltd, Atlas House,
>           Osbaldwick Link Road, YORK, YO10 3JB
>           United Kingdom

Re: warning in ext4_journal_start_sb on filesystem freeze

[Christoph Hellwig]

On Tue, Mar 04, 2014 at 02:04:42PM -0500, J. Bruce Fields wrote:
> diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
> index 6d7be3f..eea5ad1 100644
> --- a/fs/nfsd/vfs.c
> +++ b/fs/nfsd/vfs.c
> @@ -404,6 +404,7 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
>  	umode_t		ftype = 0;
>  	__be32		err;
>  	int		host_err;
> +	bool		get_write_count;
>  	int		size_change = 0;
>  
>  	if (iap->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_SIZE))
> @@ -411,10 +412,18 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
>  	if (iap->ia_valid & ATTR_SIZE)
>  		ftype = S_IFREG;
>  
> +	/* Callers that do fh_verify should do the fh_want_write: */
> +	get_write_count = !fhp->fh_dentry;

Eww, this is nasty.  Given that there are only 6 callers of nfsd_setattr
in total, and only half of these might cause size changes I'd rather
deal with this properly, e.g. by taking both the fh_verify into the
callers.

Re: warning in ext4_journal_start_sb on filesystem freeze

[J. Bruce Fields]

On Mon, Mar 10, 2014 at 06:34:51AM -0700, Christoph Hellwig wrote:
> On Tue, Mar 04, 2014 at 02:04:42PM -0500, J. Bruce Fields wrote:
> > diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
> > index 6d7be3f..eea5ad1 100644
> > --- a/fs/nfsd/vfs.c
> > +++ b/fs/nfsd/vfs.c
> > @@ -404,6 +404,7 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
> >  	umode_t		ftype = 0;
> >  	__be32		err;
> >  	int		host_err;
> > +	bool		get_write_count;
> >  	int		size_change = 0;
> >  
> >  	if (iap->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_SIZE))
> > @@ -411,10 +412,18 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
> >  	if (iap->ia_valid & ATTR_SIZE)
> >  		ftype = S_IFREG;
> >  
> > +	/* Callers that do fh_verify should do the fh_want_write: */
> > +	get_write_count = !fhp->fh_dentry;
> 
> Eww, this is nasty.  Given that there are only 6 callers of nfsd_setattr
> in total, and only half of these might cause size changes I'd rather
> deal with this properly, e.g. by taking both the fh_verify into the
> callers.

Maybe so.

(Size is irrelevant, though, right?  Won't any setattr need an elevated
write count?)

--b.

Re: warning in ext4_journal_start_sb on filesystem freeze

[Christoph Hellwig]

On Mon, Mar 10, 2014 at 03:57:09PM -0400, J. Bruce Fields wrote:
> (Size is irrelevant, though, right?  Won't any setattr need an elevated
> write count?)

Indeed.  Not sure why I was thinking of truncate as a special case here.

Re: warning in ext4_journal_start_sb on filesystem freeze

[J. Bruce Fields]

On Mon, Mar 10, 2014 at 03:57:09PM -0400, J. Bruce Fields wrote:
> On Mon, Mar 10, 2014 at 06:34:51AM -0700, Christoph Hellwig wrote:
> > On Tue, Mar 04, 2014 at 02:04:42PM -0500, J. Bruce Fields wrote:
> > > diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
> > > index 6d7be3f..eea5ad1 100644
> > > --- a/fs/nfsd/vfs.c
> > > +++ b/fs/nfsd/vfs.c
> > > @@ -404,6 +404,7 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
> > >  	umode_t		ftype = 0;
> > >  	__be32		err;
> > >  	int		host_err;
> > > +	bool		get_write_count;
> > >  	int		size_change = 0;
> > >  
> > >  	if (iap->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_SIZE))
> > > @@ -411,10 +412,18 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp, struct iattr *iap,
> > >  	if (iap->ia_valid & ATTR_SIZE)
> > >  		ftype = S_IFREG;
> > >  
> > > +	/* Callers that do fh_verify should do the fh_want_write: */
> > > +	get_write_count = !fhp->fh_dentry;
> > 
> > Eww, this is nasty.  Given that there are only 6 callers of nfsd_setattr
> > in total, and only half of these might cause size changes I'd rather
> > deal with this properly, e.g. by taking both the fh_verify into the
> > callers.
> 
> Maybe so.

Gah, I found clearing out my invoice that a) I'd forgotten this, b) I'd
already committed and pushed out the patch.

And I'd rather leave the fix as is and the cleanup to be done later.

But it's not OK to just drop review like that and if you think it
warrants reverting or rebasing I can do that.

--b.