On Tue, 3 Nov 2015 17:06:24 -0500
"J. Bruce Fields" <bfields@xxxxxxxxxxxx> wrote:
> On Mon, Nov 02, 2015 at 11:44:27AM -0500, Andrew W Elble wrote:
> >
> > I've been busy patching in things to observe what's going on with my
> > delegation issues.
> >
> > I implemented a seq_file interface on the server to walk the
> > file_hashtbl, and some interesting things have popped up:
>
> Zowie.
>
> So you're waiting for the problem to reproduce itself, and then dumping
> this information?
>
> (And if I remember correctly the problem was that the server was
> attempting to recall delegations that the client claimed not to have any
> knowledge of?)
>
> > This is a nfs4_file that has two open stateids allocated to the same client:
> >
> > file fh: 82369f70 R:2 F:4 FP:ffff881f24286b58 I:ffff881f6f6a91d0 D:ffff881febf08b40 /home/<stuff>
> > Deleg: Linux NFSv4.1 <hosta> : 2292335602000000d89e0700 120317
> > Deleg: Linux NFSv4.1 <hostb> : 2292335607000000437f0700 148262
> > Stateid: Linux NFSv4.1 <hosta> : 2292335602000000d69e0700/1
> > Stateid: Linux NFSv4.1 <hosta> : 2292335602000000d79e0700/1
> >
> > This one has four:
> >
> > file fh: 6ecdd5f8 R:4 F:4 FP:ffff88083e2974a0 I:ffff881f6c9349d0 D:ffff883f4ad1e780 /home/<stuff>
> > Stateid: Linux NFSv4.1 <hosta> : 2292335602000000f1620800/1
> > Stateid: Linux NFSv4.1 <hosta> : 2292335602000000f0620800/1
> > Stateid: Linux NFSv4.1 <hosta> : 2292335602000000c64c0800/1
> > Stateid: Linux NFSv4.1 <hosta> : 2292335602000000c74c0800/1
>
> I believe stateid's are per client,openowner,file. So multiple
> stateid's per file should be OK if there are multiple openowners.
>
Technically, there is a per-client hash of openowners and then each
open stateid is per-openowner. But your statement is basically correct.
> > These are nfs4_files that have elevated fi_ref counts but are
> > essentially "empty":
> > (note: fi_had_conflict is seemingly always nonzero on entries like
> > these)
> >
> > file fh: ee23414c F:1 FP:ffff88087fc52318 I:0000000000000000
> > file fh: 5a8e8c86 F:2 FP:ffff881f5e872318 I:0000000000000000
>
> So if I understand correctly, there's neither a delegated file nor
> anything in the fi_fds[] array? Hm.
>
Yes, looks that way. The code sometimes will take "ephemeral"
references to a file after looking it up, for instance. That said, if
they're sticking around for a long time afterward then we may have a
fi_ref leak somewhere.
> I'll have to stare at the code to figure out what that means.
>
> fi_had_conflict means a delegation was recalled on this file at some
> point.
>
Right.
>
> >
> > ======================
> >
> > for clarity:
> >
> > +static int nfs4_file_hash_list_show(struct seq_file *seq, void *v)
> > +{
> > + char *name = __getname();
> > + char *p;
> > + int i;
> > + struct dentry *dentry;
> > + struct inode *inode = NULL;
> > + struct nfs4_delegation *dp;
> > + struct nfs4_ol_stateid *stp;
> > + if (!name)
> > + goto out;
> > +
> > + if (v != SEQ_START_TOKEN) {
> > + struct nfs4_file *fp = (struct nfs4_file *) v;
> > + spin_lock(&fp->fi_lock);
> > + seq_printf(seq, "file fh:
> > %08x",nfsd_fhandle_hash(&fp->fi_fhandle));
> > + for (i = 0; i < 3; i++) {
> > + if (fp->fi_fds[i]) {
> > + if (!inode)
> > + inode =
> > (fp->fi_fds[i])->f_inode;
> > + dentry =
> > (fp->fi_fds[i])->f_path.dentry;
> > + if (dentry)
> > + break;
> > + }
> > + }
> > + if (atomic_read(&fp->fi_access[O_RDONLY]) > 0)
> > + seq_printf(seq, " R:%d",
> > +
> > atomic_read(&fp->fi_access[O_RDONLY]));
> > + if (atomic_read(&fp->fi_access[O_WRONLY]) > 0)
> > + seq_printf(seq, " W:%d",
> > +
> > atomic_read(&fp->fi_access[O_WRONLY]));
> > + seq_printf(seq, " F:%d",
> > + atomic_read(&fp->fi_ref));
> > + seq_printf(seq, " FP:%pK", fp);
> > + if ((!dentry) && fp->fi_deleg_file)
> > + dentry = fp->fi_deleg_file->f_path.dentry;
> > + if ((!inode) && fp->fi_deleg_file)
> > + inode = (fp->fi_deleg_file)->f_inode;
> > + seq_printf(seq, " I:%pK", inode);
> > + if (dentry) {
> > + p = dentry_path_raw(dentry, name, PATH_MAX);
> > + seq_printf(seq, " D:%pK %s\n", dentry, p);
> > + } else {
> > + seq_printf(seq, "\n");
> > + }
> > + list_for_each_entry(dp, &fp->fi_delegations,
> > dl_perfile) {
> > + seq_printf(seq, " Deleg: %.*s : %*phN
> > %lu\n",
> > +
> > (int)dp->dl_recall.cb_clp->cl_name.len,
> > + dp->dl_recall.cb_clp->cl_name.data,
> > + 12,
> > + &dp->dl_stid.sc_stateid.si_opaque,
> > + get_seconds() - dp->dl_alloctime);
> > + }
> > + list_for_each_entry(stp, &fp->fi_stateids,
> > st_perfile) {
> > + seq_printf(seq, " Stateid: %.*s :
> > %*phN/%x\n",
> > +
> > (int)stp->st_stid.sc_client->cl_name.len,
> > +
> > stp->st_stid.sc_client->cl_name.data,
> > + 12,
> > + &stp->st_stid.sc_stateid.si_opaque,
> > +
> > stp->st_stid.sc_stateid.si_generation);
> > + }
> > + spin_unlock(&fp->fi_lock);
> > + }
> > + __putname(name);
> > +out:
> > + return 0;
> > +}
> >
> > Thanks,
> >
> > Andy
> >
> > --
> > Andrew W. Elble
> > aweits@xxxxxxxxxxxxxxxxxx
> > Infrastructure Engineer, Communications Technical Lead
> > Rochester Institute of Technology
> > PGP: BFAD 8461 4CCF DC95 DA2C B0EB 965B 082E 863E C912
--
Jeff Layton <jlayton@xxxxxxxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
