On Fri, 2018-07-20 at 14:58 -0400, Olga Kornievskaia wrote:
> On Fri, Jul 20, 2018 at 2:46 PM, Trond Myklebust
> <trondmy@xxxxxxxxxxxxxxx> wrote:
> > On Fri, 2018-07-20 at 14:40 -0400, Olga Kornievskaia wrote:
> > > On Fri, Jul 20, 2018 at 2:17 PM, Olga Kornievskaia <aglo@xxxxxxxx
> > > u>
> > > wrote:
> > > > On Fri, Jul 20, 2018 at 2:05 PM, Trond Myklebust
> > > > <trondmy@xxxxxxxxxxxxxxx> wrote:
> > > > > On Fri, 2018-07-20 at 13:26 -0400, Olga Kornievskaia wrote:
> > > > > > Hi Trond,
> > > > > >
> > > > > > I would some help understanding attributes management.
> > > > > >
> > > > > > Right now, any time a directory inode that was marked with
> > > > > > INVALID_ACCESS (say to a change_attribute changed) ends up
> > > > > > triggering
> > > > > > sending a duplicate GETATTR. I don't think that's correct.
> > > > > >
> > > > > > In nfs_execute_ok() we check if
> > > > > > (nfs_check_cache_invalid(inode,
> > > > > > NFS_INO_INVALID_ACCESS)) and the call
> > > > > > __nfs_revalidate_inode()
> > > > > > which
> > > > > > will trigger GETATTR. I don't understand why after calling
> > > > > > this
> > > > > > function the INVALID_ACCESS doesn't get cleared? Because
> > > > > > that's
> > > > > > what
> > > > > > causes the double GETATTR to be sent.
> > > > > >
> > > > > > On the open path, the first time the getattr is sent is in
> > > > > >
> > > > > > Jul 19 15:39:52 ipa18 kernel: dump_stack+0x5a/0x73
> > > > > > Jul 19 15:39:52 ipa18 kernel: nfs4_proc_getattr+0x65/0x110
> > > > > > [nfsv4]
> > > > > > Jul 19 15:39:52 ipa18 kernel:
> > > > > > __nfs_revalidate_inode+0xe1/0x370
> > > > > > [nfs]
> > > > > > Jul 19 15:39:52 ipa18 kernel: nfs_permission+0x16b/0x1f0
> > > > > > [nfs]
> > > > > > Jul 19 15:39:52 ipa18 kernel: inode_permission+0xab/0x130
> > > > > > Jul 19 15:39:52 ipa18 kernel: link_path_walk+0x29d/0x520
> > > > > > Jul 19 15:39:52 ipa18 kernel: path_openat+0xf6/0x1230
> > > > > > Jul 19 15:39:52 ipa18 kernel: do_filp_open+0x91/0x100
> > > > > > Jul 19 15:39:52 ipa18 kernel: do_sys_open+0x126/0x210
> > > > > > Jul 19 15:39:52 ipa18 kernel: do_syscall_64+0x55/0x180
> > > > > > Jul 19 15:39:52 ipa18 kernel:
> > > > > > entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > > > > >
> > > > > > And then again during the open
> > > > > >
> > > > > > Jul 19 15:39:52 ipa18 kernel: dump_stack+0x5a/0x73
> > > > > > Jul 19 15:39:52 ipa18 kernel: nfs4_proc_getattr+0x65/0x110
> > > > > > [nfsv4]
> > > > > > Jul 19 15:39:52 ipa18 kernel:
> > > > > > __nfs_revalidate_inode+0xe1/0x370
> > > > > > [nfs]
> > > > > > Jul 19 15:39:52 ipa18 kernel: nfs_permission+0x16b/0x1f0
> > > > > > [nfs]
> > > > > > Jul 19 15:39:52 ipa18 kernel: inode_permission+0xab/0x130
> > > > > > Jul 19 15:39:52 ipa18 kernel: path_openat+0x942/0x1230
> > > > > > Jul 19 15:39:52 ipa18 kernel: do_filp_open+0x91/0x100
> > > > > > Jul 19 15:39:52 ipa18 kernel: do_sys_open+0x126/0x210
> > > > > > Jul 19 15:39:52 ipa18 kernel: do_syscall_64+0x55/0x180
> > > > > > Jul 19 15:39:52 ipa18 kernel:
> > > > > > entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > > > > >
> > > > > > Can't we remove INVALID_ACCESS after we revalidated the
> > > > > > inode?
> > > > > > This
> > > > > > removes the duplicated GETATTRs. Is this a valid way of
> > > > > > fixing
> > > > > > this
> > > > > > issue?
> > > > > >
> > > > > > diff --git a/fs/nfs/dir.c b/fs/nfs/dir.c
> > > > > > index 8f8e9e9..2b55a45 100644
> > > > > > --- a/fs/nfs/dir.c
> > > > > > +++ b/fs/nfs/dir.c
> > > > > > @@ -2504,6 +2504,8 @@ static int nfs_execute_ok(struct
> > > > > > inode
> > > > > > *inode,
> > > > > > int mask)
> > > > > > if (mask & MAY_NOT_BLOCK)
> > > > > > return -ECHILD;
> > > > > > ret = __nfs_revalidate_inode(server,
> > > > > > inode);
> > > > > > + if (!ret)
> > > > > > + NFS_I(inode)->cache_validity &=
> > > > > > ~NFS_INO_INVALID_ACCESS;
> > > > > > }
> > > > > > if (ret == 0 && !execute_ok(inode))
> > > > > > ret = -EACCES;
> > > > >
> > > > >
> > > > > I don't see how the above makes sense.
> > > > >
> > > > > Either the attribute revalidation confirmed that the change
> > > > > attr,
> > > > > mode
> > > > > + uid + gid are unchanged, in which case the call to
> > > > > nfs_refresh_inode() during revalidation will fail to set
> > > > > NFS_INO_INVALID_ACCESS, or it confirmed that at least one of
> > > > > them
> > > > > has
> > > > > changed, in which case we do want to revalidate the access
> > > > > cache.
> > > >
> > > > I'm confused as to against what are we checking then?
> > > >
> > > > We flagged a directory inode to have invalid attributes. So we
> > > > sent
> > > > a
> > > > GETATTR. I would think that after getting the result all our
> > > > attributes should be valid. Why would a client as the next
> > > > operation
> > > > send another GETATTR for the same inode?
> > > >
> > >
> > > I didn't answer your question but rather explained what I see
> > > client
> > > do.
> > >
> > > Let's me see if I can try again:
> > > in nfs_execute_ok() the check for the INVALID_CACHE is true so
> > > the
> > > code calls __nfs_revalidate_inode() which ends up sending a
> > > GETATTR.
> > > Could it be that what's happening is that GETATTRs reply for the
> > > change_attr is different than what we have stored. BUT that's
> > > obvious,
> > > we knew that to begin with since we are validating the
> > > attributes. So
> > > we update it and then we send another GETATTR now the GETATTR
> > > sends
> > > back the "same" change_attr and this time it matches what we have
> > > stored. Why is this 2nd GETATTR necessary? The attribute should
> > > be
> > > marked as valid after a getting a successful GETATTR.
> > >
> >
> > Oh... I see what you are saying. Yes, that check looks like it
> > should
> > be looking for NFS_INO_INVALID_OTHER.
>
> Whew. ok I'm glad something make sense. However, you lost me on
> "NFS_INO_INVALID_OTHER". What does that flag means?
> Are you talking about checking the check in nfs_execute_ok() to check
> for INVALID_OTHER instead of INVALID_ACCESS?
>
I'm saying that function should probably read as:
static int nfs_execute_ok(struct inode *inode, int mask)
{
struct nfs_server *server = NFS_SERVER(inode);
int ret = 0;
if (nfs_check_cache_invalid(inode, NFS_INO_INVALID_OTHER)) {
if (mask & MAY_NOT_BLOCK)
return -ECHILD;
ret = __nfs_revalidate_inode(server, inode);
}
if (ret == 0 && !execute_ok(inode))
ret = -EACCES;
return ret;
}
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx
��.n��������+%������w��{.n�����{��w��� jg��� �����ݢj����G���� ���j:+v���w�m������w�������h�����٥
