On Wed, 10 Sep 2014 07:59:43 -0400 Jeff Layton <jeff.layton@xxxxxxxxxxxxxxx>
wrote:
> On Wed, 10 Sep 2014 13:57:39 +1000
> NeilBrown <neilb@xxxxxxx> wrote:
>
> > On Tue, 9 Sep 2014 13:50:06 -0400 Jeff Layton <jeff.layton@xxxxxxxxxxxxxxx>
> > wrote:
> >
> > > On Tue, 9 Sep 2014 12:15:46 -0400
> > > Jeff Layton <jeff.layton@xxxxxxxxxxxxxxx> wrote:
> > >
> > > > On Tue, 9 Sep 2014 12:12:44 -0400
> > > > Jeff Layton <jlayton@xxxxxxxxxxxxxxx> wrote:
> > > >
> > > > > On Tue, 9 Sep 2014 08:42:11 -0700
> > > > > Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > > On Tue, Sep 09, 2014 at 10:59:18AM -0400, Jeff Layton wrote:
> > > > > > > > [ 5497.405585] [<ffffffff8135d0c1>] nfs4_do_open.constprop.84+0x681/0x960
> > > > > > > > [ 5497.405585] [<ffffffff8135d459>] nfs4_atomic_open+0x9/0x20
> > > > > > > > [ 5497.405585] [<ffffffff8136cc3d>] nfs4_file_open+0xcd/0x1b0
> > > > > > > > [ 5497.405585] [<ffffffff811b8092>] do_dentry_open.isra.13+0x1f2/0x320
> > > > > > > > [ 5497.405585] [<ffffffff811b82ad>] finish_open+0x1d/0x30
> > > > > > > > [ 5497.405585] [<ffffffff811c98e9>] path_openat+0xb9/0x670
> > > > > > > > [ 5497.405585] [<ffffffff811ca26e>] do_filp_open+0x3e/0xa0
> > > > > > > > [ 5497.405585] [<ffffffff811b95cc>] do_sys_open+0x13c/0x230
> > > > > > > > [ 5497.405585] [<ffffffff811b96dd>] SyS_open+0x1d/0x20
> > > > > > > > [ 5497.405585] [<ffffffff81d9f5a9>] system_call_fastpath+0x16/0x1b
> > > > > > > >
> > > > > > >
> > > > > > > Odd...
> > > > > > >
> > > > > > > It looks like you hit the BUG() in d_instantiate.
> > > > > > >
> > > > > > > I don't see any calls to d_instantiate in the current nfs_code (aside
> > > > > > > from the one in nfs_lookup, and I don't think that's getting called
> > > > > > > here). d_instantiate is called by d_add though, and that gets called
> > > > > > > from nfs_atomic_open. Is that what happened here?
> > > > > > >
> > > > > > > Maybe you can use gdb to figure out what line of code
> > > > > > > "nfs4_do_open.constprop.84+0x681" actually is?
> > > > > >
> > > > > > My gdb can't cope with these constprop expressions unfortunately.
> > > > > >
> > > > > > But when you remove the questionable symbols as I've done above it's
> > > > > > pretty clear that this must be the
> > > > > >
> > > > > > nfs4_atomic_open
> > > > > > -> nfs4_do_open
> > > > > > -> _nfs4_do_open
> > > > > > -> _nfs4_open_and_get_state
> > > > > > -> d_add
> > > > > > -> d_instantiate
> > > > > >
> > > > > > call chain. There is heavy inlining going on inside nfs4file.c, and
> > > > > > d_add now is a simple inline around d_instantiate and d_rehash.
> > > > >
> > > > > Ok. So I'm guessing that means that the current scheme of doing a
> > > > > d_drop/d_add is not valid. d_drop doesn't remove the dentry from the
> > > > > i_alias list.
> > > > >
> > > > > It looks like the first call there should just be doing a d_delete
> > > > > instead, since it's trying to turn the thing into a negative dentry.
> > > >
> > > > (cc'ing Neil B.)
> > > >
> > > > ...and I'd hazard a guess that 4fa2c54b5198d might be the culprit. You
> > > > might want to try backing that out and see if it's still reproducible.
> > > >
> > > > Neil, any thoughts?
> > > >
> > >
> > > In fact, maybe this patch would make sense?
> > >
> > > ------------------[snip]-------------------
> > >
> > > [PATCH] nfs: d_drop/d_add of positive dentry may be harmful
> > >
> > > Christoph reported the following oops, when running xfstests:
> > >
> > > generic/089 199s ...
> > > [ 5497.402293] ------------[ cut here ]------------
> > > [ 5497.403150] kernel BUG at ../fs/dcache.c:1620!
> > > [ 5497.403974] invalid opcode: 0000 [#1] SMP
> > > [ 5497.404826] Modules linked in:
> > > [ 5497.405280] CPU: 1 PID: 14691 Comm: t_mtab Not tainted 3.17.0-rc3+ #264
> > > [ 5497.405585] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007
> > > [ 5497.405585] task: ffff88007ac801d0 ti: ffff8800670a4000 task.ti: ffff8800670a4000
> > > [ 5497.405585] RIP: 0010:[<ffffffff811d1345>] [<ffffffff811d1345>] d_instantiate+0x75/0x80
> > > [ 5497.405585] RSP: 0018:ffff8800670a7a68 EFLAGS: 00010286
> > > [ 5497.405585] RAX: 0000000000000001 RBX: ffff880066c399d8 RCX: ffff88007ac80990
> > > [ 5497.405585] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff880066c399d8
> > > [ 5497.405585] RBP: ffff8800670a7a88 R08: 0000000000000001 R09: 0000000000000000
> > > [ 5497.405585] R10: ffff880072f45000 R11: 000000000003fdf0 R12: ffff880066c399d8
> > > [ 5497.405585] R13: ffff88007a684800 R14: ffff88007acbc280 R15: ffff8800670de000
> > > [ 5497.405585] FS: 00007f6db6aae700(0000) GS:ffff88007fc80000(0000) knlGS:0000000000000000
> > > [ 5497.405585] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
> > > [ 5497.405585] CR2: 00007f6db5f56800 CR3: 000000007ac9e000 CR4: 00000000000006e0
> > > [ 5497.405585] Stack:
> > > [ 5497.405585] ffff8800670a7a88 ffff880066c399d8 ffff88007a684800 ffff88007a684800
> > > [ 5497.405585] ffff8800670a7b68 ffffffff8135d0c1 ffffffff00000004 ffff8800000000d0
> > > [ 5497.405585] ffff88007d400180 0000000000000246 ffff8800fffffffe ffff880072f45000
> > > [ 5497.405585] Call Trace:
> > > [ 5497.405585] [<ffffffff8135d0c1>] nfs4_do_open.constprop.84+0x681/0x960
> > > [ 5497.405585] [<ffffffff8135d459>] nfs4_atomic_open+0x9/0x20
> > > [ 5497.405585] [<ffffffff8136cc3d>] nfs4_file_open+0xcd/0x1b0
> > > [ 5497.405585] [<ffffffff81d9ee56>] ? _raw_spin_unlock+0x26/0x30
> > > [ 5497.405585] [<ffffffff8176b73d>] ? lockref_get+0x1d/0x30
> > > [ 5497.405585] [<ffffffff8136cb70>] ? nfs4_file_fsync+0xb0/0xb0
> > > [ 5497.405585] [<ffffffff811b8092>] do_dentry_open.isra.13+0x1f2/0x320
> > > [ 5497.405585] [<ffffffff81336f52>] ? nfs_permission+0x62/0x1d0
> > > [ 5497.405585] [<ffffffff811b82ad>] finish_open+0x1d/0x30
> > > [ 5497.405585] [<ffffffff811c91be>] do_last.isra.63+0x62e/0xca0
> > > [ 5497.405585] [<ffffffff811c5703>] ? inode_permission+0x13/0x50
> > > [ 5497.405585] [<ffffffff811c5bde>] ? link_path_walk+0x23e/0x850
> > > [ 5497.405585] [<ffffffff811c98e9>] path_openat+0xb9/0x670
> > > [ 5497.405585] [<ffffffff811b366b>] ? poison_obj+0x2b/0x40
> > > [ 5497.405585] [<ffffffff811ca26e>] do_filp_open+0x3e/0xa0
> > > [ 5497.405585] [<ffffffff811d7da1>] ? __alloc_fd+0xd1/0x120
> > > [ 5497.405585] [<ffffffff811b95cc>] do_sys_open+0x13c/0x230
> > > [ 5497.405585] [<ffffffff810f5bbd>] ? trace_hardirqs_on_caller+0x10d/0x1d0
> > > [ 5497.405585] [<ffffffff811b96dd>] SyS_open+0x1d/0x20
> > > [ 5497.405585] [<ffffffff81d9f5a9>] system_call_fastpath+0x16/0x1b
> > >
> > > The BUG() is due to the fact that the d_alias hlist is not empty when
> > > we called into d_instantiate. This is likely due to a situation where
> > > we did a successful open and instantiated the dentry and then later
> > > failed and ended up retrying. At that point, we try the open again and
> > > get back -ENOENT, and try to d_drop/d_add it.
> > >
> > > The problem is that d_drop'ing a positive dentry is not sufficient to
> > > "clear" it for adding it back into the cache. That just makes it unfindable
> > > in the hash tables, but doesn't unhitch it from the inode.
> > >
> > > Switch to using the helper we already have for turning positive dentries
> > > into negative ones.
> > >
> > > Cc: Neil Brown <neilb@xxxxxxx>
> > > Reported-by: Christoph Hellwig <hch@xxxxxxxxxxxxx>
> > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxxxxxxx>
> > > ---
> > > fs/nfs/dir.c | 3 ++-
> > > fs/nfs/internal.h | 1 +
> > > fs/nfs/nfs4proc.c | 3 +--
> > > 3 files changed, 4 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/fs/nfs/dir.c b/fs/nfs/dir.c
> > > index 36d921f0c602..3938dba859c5 100644
> > > --- a/fs/nfs/dir.c
> > > +++ b/fs/nfs/dir.c
> > > @@ -1754,11 +1754,12 @@ out_err:
> > > }
> > > EXPORT_SYMBOL_GPL(nfs_mkdir);
> > >
> > > -static void nfs_dentry_handle_enoent(struct dentry *dentry)
> > > +void nfs_dentry_handle_enoent(struct dentry *dentry)
> > > {
> > > if (dentry->d_inode != NULL && !d_unhashed(dentry))
> > > d_delete(dentry);
> > > }
> > > +EXPORT_SYMBOL_GPL(nfs_dentry_handle_enoent);
> > >
> > > int nfs_rmdir(struct inode *dir, struct dentry *dentry)
> > > {
> > > diff --git a/fs/nfs/internal.h b/fs/nfs/internal.h
> > > index 9056622d2230..8d85a57ae499 100644
> > > --- a/fs/nfs/internal.h
> > > +++ b/fs/nfs/internal.h
> > > @@ -326,6 +326,7 @@ extern unsigned long nfs_access_cache_scan(struct shrinker *shrink,
> > > struct dentry *nfs_lookup(struct inode *, struct dentry *, unsigned int);
> > > int nfs_create(struct inode *, struct dentry *, umode_t, bool);
> > > int nfs_mkdir(struct inode *, struct dentry *, umode_t);
> > > +void nfs_dentry_handle_enoent(struct dentry *);
> > > int nfs_rmdir(struct inode *, struct dentry *);
> > > int nfs_unlink(struct inode *, struct dentry *);
> > > int nfs_symlink(struct inode *, struct dentry *, const char *);
> > > diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
> > > index 7dd8aca31c29..84ee3fb9e410 100644
> > > --- a/fs/nfs/nfs4proc.c
> > > +++ b/fs/nfs/nfs4proc.c
> > > @@ -2226,8 +2226,7 @@ static int _nfs4_open_and_get_state(struct nfs4_opendata *opendata,
> > > ret = _nfs4_proc_open(opendata);
> > > if (ret != 0) {
> > > if (ret == -ENOENT) {
> > > - d_drop(opendata->dentry);
> > > - d_add(opendata->dentry, NULL);
> > > + nfs_dentry_handle_enoent(opendata->dentry);
> > > nfs_set_verifier(opendata->dentry,
> > > nfs_save_change_attribute(opendata->dir->d_inode));
> > > }
> >
> >
> > The d_drop();d_add(); pattern is used a number of times in NFS, but what I
> > didn't notice before is that it is only used if ->d_inode is NULL.
> >
> > We *could* simply change the condition to
> > if (ret == -ENOENT && opendata->dentry->d_inode == NULL) {
> >
> > Calling nfs_dentry_handle_enoent() is possibly a good idea, but we do want to
> > still call d_add(). The whole point of the patch which introduce that code
> > was to ensure that negative entries were cached.
> > nfs_dentry_handle_enoent() is always enough to prepare for d_add(), so we
> > cannot just replace d_drop with that.
> >
> > Maybe:
> >
> > if (ret == -ENOENT) {
> > struct dentry *de = opendata->dentry;
> > d_drop(de->d_inode);
> > if (de->d_inode == NULL)
> > d_add(de, NULL);
> > nfs_set_verifier(de, .....);
> > }
> >
>
> Yeah using nfs_dentry_handle_enoent won't hash a brand new negative
> dentry. The above code though won't convert an existing positive dentry
> into a negative one. I guess though that's a rare enough situation that
> it's not worth optimizing for?
>
I agree. I've just posted a patch which I think is the best balance.
Thanks for your analysis and review!
NeilBrown
Attachment:
signature.asc
Description: PGP signature
