nfs-root: destructive call to __detach_mounts /dev

nfs-root: destructive call to __detach_mounts /dev

[Sven Geggus]

Hello,

I have a couple of machines running Debian GNU/Linux 8 using an NFS-ro-mounted
root filesystem.

The systems seem to get bitten by VFS changes in some (unfortunaly
somewhat difficult to reproduce) circumstances.

The effect is, that /dev (dectmpfs) gets unmounted for some strange reason.

I added the following debug code to fs/namespace.c

printk(KERN_DEBUG "%s: %s\n", __func__, dentry->d_name.name);
dump_stack();

And this is what I get if one of those events happen:

__detach_mounts: dev
CPU: 1 PID: 5551 Comm: modtrack Not tainted 4.1.3-debug-00287-g0fe8050-dirty #1
Hardware name: Dell Inc. Precision WorkStation T3400  /0TP412, BIOS A03 01/31/2008
 ffff880127183c18 ffff880127183bc8 ffffffff815605f8 ffff88012bc4fcb0
 ffff880126c16198 ffff880127183bf8 ffffffff81106cf7 000000000000000f
 ffff880126c16198 ffff880127183c18 ffff880127183d00 ffff880127183c48
Call Trace:
 [<ffffffff815605f8>] dump_stack+0x4c/0x6e
 [<ffffffff81106cf7>] __detach_mounts+0x2b/0x12c
 [<ffffffff810ffbb6>] d_invalidate+0x9a/0xc8
 [<ffffffff810f6b13>] lookup_fast+0x1f5/0x26f
 [<ffffffff810f6dba>] do_last.isra.43+0xd6/0x9fb
 [<ffffffff810f916e>] path_openat+0x1d1/0x53e
 [<ffffffff810f9f41>] ? user_path_at_empty+0x63/0x93
 [<ffffffff810f9fe6>] do_filp_open+0x35/0x85
 [<ffffffff811f1893>] ? find_next_zero_bit+0x17/0x1d
 [<ffffffff81104225>] ? __alloc_fd+0xdd/0xef
 [<ffffffff810ec966>] do_sys_open+0x146/0x1d5
 [<ffffffff810eca1f>] SyS_openat+0xf/0x11
 [<ffffffff81565a17>] system_call_fastpath+0x12/0x6a

Regards

Sven

P.S.: Kernel is vanilla 4.1.3 with aufs patches, but aufs is not related in
this problem.

-- 
Der "normale Bürger" ist nicht an der TU Dresden und schreibt auch
nicht mit mutt. (Ulli Kuhnle in de.comp.os.unix.discussion)

/me is giggls@ircnet, http://sven.gegg.us/ on the Web

Re: nfs-root: destructive call to __detach_mounts /dev

[Eric W. Biederman]

I have added the linux-nfs list to hopefully add a wider interested audience.

Sven Geggus <lists@fuchsschwanzdomain.de> writes:

> Hello,
>
> I have a couple of machines running Debian GNU/Linux 8 using an NFS-ro-mounted
> root filesystem.
>
> The systems seem to get bitten by VFS changes in some (unfortunaly
> somewhat difficult to reproduce) circumstances.
>
> The effect is, that /dev (dectmpfs) gets unmounted for some strange reason.
>
> I added the following debug code to fs/namespace.c
>
> printk(KERN_DEBUG "%s: %s\n", __func__, dentry->d_name.name);
> dump_stack();
>
> And this is what I get if one of those events happen:

The only way to get to d_invalidate in lookup_fast is if d_revalidate
fails.

Which yields two functions to look at:
nfs_lookup_revalidate and nfs4_lookup_revalidate

If what is being revalidated is a mount point nfs4_lookup_revalidate
calls nfs_lookup_revalidate.  So nfs_lookup_revalidate is the only
interesting function.

The last round of this was with readdir having problems and invalidating
the dcache.  So now apparently we are down to something weird happening
in revalidate.

I don't understand the what nfs_lookup_revalidate is doing particularly
well.  However it appears that if you enable the kconfig SUNRPC_DEBUG
option there will be more relevant information.

I am not certain what more to print out that SUNRPC_DEBUG won't, but I
have verified that SUNRPC_DEBUG will report when a failure path is hit
in nfs_lookup_revalidate and which failure path was hit (out_bad or
out_error).

Which should be sufficient to start narrowing down what is happening
even further.

I hope that helps a little bit in tracking down what is happening.

Eric

> __detach_mounts: dev
> CPU: 1 PID: 5551 Comm: modtrack Not tainted 4.1.3-debug-00287-g0fe8050-dirty #1
> Hardware name: Dell Inc. Precision WorkStation T3400  /0TP412, BIOS A03 01/31/2008
>  ffff880127183c18 ffff880127183bc8 ffffffff815605f8 ffff88012bc4fcb0
>  ffff880126c16198 ffff880127183bf8 ffffffff81106cf7 000000000000000f
>  ffff880126c16198 ffff880127183c18 ffff880127183d00 ffff880127183c48
> Call Trace:
>  [<ffffffff815605f8>] dump_stack+0x4c/0x6e
>  [<ffffffff81106cf7>] __detach_mounts+0x2b/0x12c
>  [<ffffffff810ffbb6>] d_invalidate+0x9a/0xc8
>  [<ffffffff810f6b13>] lookup_fast+0x1f5/0x26f
>  [<ffffffff810f6dba>] do_last.isra.43+0xd6/0x9fb
>  [<ffffffff810f916e>] path_openat+0x1d1/0x53e
>  [<ffffffff810f9f41>] ? user_path_at_empty+0x63/0x93
>  [<ffffffff810f9fe6>] do_filp_open+0x35/0x85
>  [<ffffffff811f1893>] ? find_next_zero_bit+0x17/0x1d
>  [<ffffffff81104225>] ? __alloc_fd+0xdd/0xef
>  [<ffffffff810ec966>] do_sys_open+0x146/0x1d5
>  [<ffffffff810eca1f>] SyS_openat+0xf/0x11
>  [<ffffffff81565a17>] system_call_fastpath+0x12/0x6a
>
> Regards
>
> Sven
>
> P.S.: Kernel is vanilla 4.1.3 with aufs patches, but aufs is not related in
> this problem.

Re: nfs-root: destructive call to __detach_mounts /dev

[Sven Geggus]

On 31-07-15 09:27 Eric W. Biederman wrote:

> I have added the linux-nfs list to hopefully add a wider interested
> audience.

... which made your mail get burried in my linux-nfs mailinglist folder :(
But I finaly found it.

> If what is being revalidated is a mount point nfs4_lookup_revalidate
> calls nfs_lookup_revalidate.  So nfs_lookup_revalidate is the only
> interesting function.

OK.

> I don't understand the what nfs_lookup_revalidate is doing particularly
> well.  

Neither do I.

Here is what I get from a broken machine (Kernel 4.1.5) using
"rpcdebug -m nfs -s lookupcache":

The mountpoint which got unmounted in this case is /proc not /dev, but the
stack-trace points to the same place.

Aug 14 11:49:37 banthonytwarog kernel: NFS: nfs_lookup_revalidate(/proc) is valid
Aug 14 11:49:37 banthonytwarog kernel: NFS: nfs_lookup_revalidate(/proc) is invalid
Aug 14 11:49:37 banthonytwarog kernel: NFSROOT __detach_mounts: proc
Aug 14 11:49:37 banthonytwarog kernel: CPU: 2 PID: 28350 Comm: modtrack Tainted: P        W  O    4.1.5-lomac3-00293-gfdd763a #6
Aug 14 11:49:37 banthonytwarog kernel: Hardware name: System manufacturer System Product Name/P8H67, BIOS 3506 03/02/2012
Aug 14 11:49:37 banthonytwarog kernel: ffff8800d9b93bb8 ffff8800d9b93b78 ffffffff81560488 00000000446c446c
Aug 14 11:49:37 banthonytwarog kernel: ffff88040c427d98 ffff8800d9b93b98 ffffffff81106d36 00000000000000a2
Aug 14 11:49:37 banthonytwarog kernel: ffff88040c427d98 ffff8800d9b93be8 ffffffff810ffc0c 00000000d9b93c08
Aug 14 11:49:37 banthonytwarog kernel: Call Trace:
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff81560488>] dump_stack+0x4c/0x6e
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff81106d36>] __detach_mounts+0x20/0xdf
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810ffc0c>] d_invalidate+0x9a/0xc8
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f6b60>] lookup_fast+0x1f5/0x26f
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f5a44>] ? __inode_permission+0x37/0x95
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f7c02>] link_path_walk+0x204/0x749
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f4db2>] ? terminate_walk+0x10/0x2e
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f75e7>] ? do_last.isra.43+0x8b6/0x9fb
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f846f>] path_init+0x328/0x337
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f919a>] path_openat+0x1b0/0x53e
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810fa073>] do_filp_open+0x75/0x85
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff8110427b>] ? __alloc_fd+0xdd/0xef
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810ec9b3>] do_sys_open+0x146/0x1d5
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810d1c5c>] ? vm_munmap+0x4b/0x59
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810eca5b>] SyS_open+0x19/0x1b
Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff81565897>] system_call_fastpath+0x12/0x6a

I suppose, that the first two lines are particularly interesting as we have
"is valid" and a fraction of a second later we have "is invalid" at the same
mountpoint.

To me this looks like a job for the NFS client maintainers now, right?

Sven

-- 
Exploits and holes are a now a necessary protection against large
corporate interests. (Alan Cox)

/me is giggls@ircnet, http://sven.gegg.us/ on the Web

Re: nfs-root: destructive call to __detach_mounts /dev

[Eric W. Biederman]

Sven Geggus <lists@fuchsschwanzdomain.de> writes:

> On 31-07-15 09:27 Eric W. Biederman wrote:
>
>> I have added the linux-nfs list to hopefully add a wider interested
>> audience.
>
> ... which made your mail get burried in my linux-nfs mailinglist folder :(
> But I finaly found it.
>
>> If what is being revalidated is a mount point nfs4_lookup_revalidate
>> calls nfs_lookup_revalidate.  So nfs_lookup_revalidate is the only
>> interesting function.
>
> OK.
>
>> I don't understand the what nfs_lookup_revalidate is doing particularly
>> well.  
>
> Neither do I.
>
> Here is what I get from a broken machine (Kernel 4.1.5) using
> "rpcdebug -m nfs -s lookupcache":
>
> The mountpoint which got unmounted in this case is /proc not /dev, but the
> stack-trace points to the same place.
>
> Aug 14 11:49:37 banthonytwarog kernel: NFS: nfs_lookup_revalidate(/proc) is valid
> Aug 14 11:49:37 banthonytwarog kernel: NFS: nfs_lookup_revalidate(/proc) is invalid
> Aug 14 11:49:37 banthonytwarog kernel: NFSROOT __detach_mounts: proc
> Aug 14 11:49:37 banthonytwarog kernel: CPU: 2 PID: 28350 Comm: modtrack Tainted: P        W  O    4.1.5-lomac3-00293-gfdd763a #6
> Aug 14 11:49:37 banthonytwarog kernel: Hardware name: System manufacturer System Product Name/P8H67, BIOS 3506 03/02/2012
> Aug 14 11:49:37 banthonytwarog kernel: ffff8800d9b93bb8 ffff8800d9b93b78 ffffffff81560488 00000000446c446c
> Aug 14 11:49:37 banthonytwarog kernel: ffff88040c427d98 ffff8800d9b93b98 ffffffff81106d36 00000000000000a2
> Aug 14 11:49:37 banthonytwarog kernel: ffff88040c427d98 ffff8800d9b93be8 ffffffff810ffc0c 00000000d9b93c08
> Aug 14 11:49:37 banthonytwarog kernel: Call Trace:
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff81560488>] dump_stack+0x4c/0x6e
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff81106d36>] __detach_mounts+0x20/0xdf
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810ffc0c>] d_invalidate+0x9a/0xc8
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f6b60>] lookup_fast+0x1f5/0x26f
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f5a44>] ? __inode_permission+0x37/0x95
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f7c02>] link_path_walk+0x204/0x749
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f4db2>] ? terminate_walk+0x10/0x2e
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f75e7>] ? do_last.isra.43+0x8b6/0x9fb
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f846f>] path_init+0x328/0x337
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810f919a>] path_openat+0x1b0/0x53e
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810fa073>] do_filp_open+0x75/0x85
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff8110427b>] ? __alloc_fd+0xdd/0xef
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810ec9b3>] do_sys_open+0x146/0x1d5
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810d1c5c>] ? vm_munmap+0x4b/0x59
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff810eca5b>] SyS_open+0x19/0x1b
> Aug 14 11:49:37 banthonytwarog kernel: [<ffffffff81565897>] system_call_fastpath+0x12/0x6a
>
> I suppose, that the first two lines are particularly interesting as we have
> "is valid" and a fraction of a second later we have "is invalid" at the same
> mountpoint.

That does sound interesting.

> To me this looks like a job for the NFS client maintainers now, right?

My educated but unsupported guess would be that there is likely
something funny going on with attributes somewhere like there was with
nfs_prime_dcache.

At a quick look the failure possibilities are:
nfs_lookup_verify_inode failing,
NFS_STALE(inode) being true,
NFS_PROTO(dir)->lookup failing,
nfs_compare_fh failing,
nfs_refresh_inode failing,

I expect what needs to happen now is to drill down into
nfs_lookup_revalidate especially into the branches that can lead to
out_bad and add some print statments so that it becomes clear just
what conditions are causing nfs_lookup_invalidate to fail.

I don't have a clue what the issue would be but I would start with
something like the patch below.  That will help narrow it down even
further.  And there are still enough possibilities that I don't think
anyone has enough information yet to figure out what is going on from
first principles.

Eric

diff --git a/fs/nfs/dir.c b/fs/nfs/dir.c
index 547308a5ec6f..97c70c887b23 100644
--- a/fs/nfs/dir.c
+++ b/fs/nfs/dir.c
@@ -1167,7 +1167,7 @@ static int nfs_lookup_revalidate(struct dentry *dentry, unsigned int flags)
 		return -ECHILD;
 
 	if (NFS_STALE(inode))
-		goto out_bad;
+		goto out_bad1;
 
 	error = -ENOMEM;
 	fhandle = nfs_alloc_fhandle();
@@ -1183,11 +1183,11 @@ static int nfs_lookup_revalidate(struct dentry *dentry, unsigned int flags)
 	error = NFS_PROTO(dir)->lookup(dir, &dentry->d_name, fhandle, fattr, label);
 	trace_nfs_lookup_revalidate_exit(dir, dentry, flags, error);
 	if (error)
-		goto out_bad;
+		goto out_bad2;
 	if (nfs_compare_fh(NFS_FH(inode), fhandle))
-		goto out_bad;
+		goto out_bad3;
 	if ((error = nfs_refresh_inode(inode, fattr)) != 0)
-		goto out_bad;
+		goto out_bad4;
 
 	nfs_setsecurity(inode, fattr, label);
 
@@ -1210,6 +1210,8 @@ out_set_verifier:
 			__func__, dentry);
 	return 1;
 out_zap_parent:
+	dfprintk(LOOKUPCACHE "NFS: %s(%pd2): nfs_lookup_verify_inode() failed\n",
+		 __func__, dentry);
 	nfs_zap_caches(dir);
  out_bad:
 	WARN_ON(flags & LOOKUP_RCU);
@@ -1233,6 +1235,22 @@ out_zap_parent:
 	dfprintk(LOOKUPCACHE, "NFS: %s(%pd2) is invalid\n",
 			__func__, dentry);
 	return 0;
+out_bad1:
+	dfprintk(LOOKUPCACHE, "NFS: %s(%pd2): NFS_STALE(inode)\n",
+		 __func__, dentry);
+	goto out_bad;
+out_bad2:
+	dfprintk(LOOKUPCACHE, "NFS: %s(%pd2): NFS_PROTO(dir)->lookup -> %u\n",
+		 __func__, dentry, error);
+	goto out_bad;
+out_bad3:
+	dfprintk(LOOKUPCACHE, "NFS: %s(%pd2): nfs_compare_fh() failed\n",
+		 __func__, dentry);
+	goto out_bad;
+out_bad4:
+	dfprintk(LOOKUPCACHE "NFS: %s(%pd2): nfs_refresh_inode() -> %u\n",
+		 __func__, dentry, error);
+	goto out_bad;
 out_error:
 	WARN_ON(flags & LOOKUP_RCU);
 	nfs_free_fattr(fattr);

Re: nfs-root: destructive call to __detach_mounts /dev

[Sven Geggus]

Eric W. Biederman wrote:

> I don't have a clue what the issue would be but I would start with
> something like the patch below.

OK using it I got the following...
This time with /dev on another machine (same nfs root-fs):

Aug 17 13:57:04 dhcp195 kernel: NFS: nfs_lookup_revalidate(/dev): nfs_lookup_verify_inode() failed
Aug 17 13:57:04 dhcp195 kernel: NFS: nfs_lookup_revalidate(/dev) is invalid
Aug 17 13:57:04 dhcp195 kernel: NFSROOT __detach_mounts: dev
Aug 17 13:57:04 dhcp195 kernel: CPU: 3 PID: 32644 Comm: modtrack Tainted: P O    4.1.6-lomac1-00296-gfeae456 #7
Aug 17 13:57:04 dhcp195 kernel: Hardware name: System manufacturer System Product Name/P7P55D, BIOS 1807    07/22/2010
Aug 17 13:57:04 dhcp195 kernel: ffff88021315bc18 ffff88021315bbd8 ffffffff81560638 000000000000b413
Aug 17 13:57:04 dhcp195 kernel: ffff880216460f18 ffff88021315bbf8 ffffffff81106e1f 000000000000001b
Aug 17 13:57:04 dhcp195 kernel: ffff880216460f18 ffff88021315bc48 ffffffff810ffcf5 0000000005dd05dd
Aug 17 13:57:04 dhcp195 kernel: Call Trace:
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff81560638>] dump_stack+0x4c/0x6e
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff81106e1f>] __detach_mounts+0x20/0xdf
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff810ffcf5>] d_invalidate+0x9a/0xc8
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff810f6c49>] lookup_fast+0x1f5/0x26f
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff810f6ef0>] do_last.isra.43+0xd6/0x9fb
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff810f92a4>] path_openat+0x1d1/0x53e
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff810fa077>] ? user_path_at_empty+0x63/0x93
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff810fa11c>] do_filp_open+0x35/0x85
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff811f1aa9>] ? find_next_zero_bit+0x17/0x1d
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff81104364>] ? __alloc_fd+0xdd/0xef
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff810eca9c>] do_sys_open+0x146/0x1d5
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff810ecb55>] SyS_openat+0xf/0x11
Aug 17 13:57:04 dhcp195 kernel: [<ffffffff81565a57>] system_call_fastpath+0x12/0x6a

So this looks like I would need to add some debug output to
"nfs_lookup_verify_inode" then right?

Sven

-- 
Why are there so many Unix-haters-handbooks and not even one
Microsoft-Windows-haters handbook?
Gurer vf ab arrq sbe n unaqobbx gb ungr Zvpebfbsg Jvaqbjf!
/me is giggls@ircnet, http://sven.gegg.us/ on the Web