Date: Thu, 23 Nov 2023 08:13:26 -0800 From: Rick Macklem <rick.macklem@gmail.com> To: Mike Karels <mike@karels.net> Cc: FreeBSD CURRENT <freebsd-current@freebsd.org>, Alexander Motin <mav@freebsd.org>, Martin Matuska <mm@freebsd.org>, Garrett Wollman <wollman@bimajority.org> Subject: Re: NFS exports of ZFS snapshots broken Message-ID: <CAM5tNy7b_uWsvRaWOGd-CD=Sr4yZwVmepnqfNR_EUXVm_OjN2g@mail.gmail.com> In-Reply-To: <7C7920E0-B0FC-4255-AD5C-ECBDDC18CF43@karels.net> References: <25943.60056.880614.452966@hergotha.csail.mit.edu> <DDC3BBBB-561A-4779-B076-83EC054859C2@karels.net> <CAM5tNy6gcThf0nrcs4iC8r6J8cMFOHdUNcUjfJVDSpTW_tqXKQ@mail.gmail.com> <ACAFC3CD-0024-4076-8C69-CAA3B1550B41@karels.net> <AB4690A9-A8F4-490B-82AA-358CD8118DF0@karels.net> <CAM5tNy5LGWF424i9v_w2u%2B41wtzmW9kn6RZWwo6Pi0Y4wj92UA@mail.gmail.com> <CAM5tNy6K6psrtem%2Bc3ta0pM-7fhy9LYci0DEeR0O-EF%2B-iGo6Q@mail.gmail.com> <91988E23-ED50-4379-AA5F-4B069E08D80F@karels.net> <CAM5tNy68tq24C_BA0xNYM-xUbCzpAve0jY-gsr42ixQLHR5gHg@mail.gmail.com> <AD48F344-DDD1-476F-B353-B0F5C05C9808@karels.net> <CAM5tNy6jqH3nRrtAhTaOpAR1RCXKxygSKHtDMKeQStDnkC1skg@mail.gmail.com> <7C7920E0-B0FC-4255-AD5C-ECBDDC18CF43@karels.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sat, Nov 18, 2023 at 8:10=E2=80=AFPM Mike Karels <mike@karels.net> wrote=
:
>
> On 18 Nov 2023, at 21:19, Rick Macklem wrote:
>
> > On Sat, Nov 18, 2023 at 4:43=E2=80=AFPM Mike Karels <mike@karels.net> w=
rote:
> >>
> >> On 18 Nov 2023, at 17:23, Rick Macklem wrote:
> >>
> >>> On Sat, Nov 18, 2023 at 2:27=E2=80=AFPM Mike Karels <mike@karels.net>=
wrote:
> >>>>
> >>>> On 18 Nov 2023, at 15:58, Rick Macklem wrote:
> >>>>
> >>>>> On Sat, Nov 18, 2023 at 8:09=E2=80=AFAM Rick Macklem <rick.macklem@=
gmail.com> wrote:
> >>>>>>
> >>>>>> On Fri, Nov 17, 2023 at 8:19=E2=80=AFPM Mike Karels <mike@karels.n=
et> wrote:
> >>>>>>>
> >>>>>>> On 17 Nov 2023, at 22:14, Mike Karels wrote:
> >>>>>>>
> >>>>>>>> On 17 Nov 2023, at 21:24, Rick Macklem wrote:
> >>>>>>>>
> >>>>>>>>> Most of the changes in stable/13 that are not in releng/13.2
> >>>>>>>>> are the "make it work in a jail" stuff. Unfortunately, they are
> >>>>>>>>> a large # of changes (mostly trivial edits adding vnet macros),
> >>>>>>>>> but it also includes export check changes.
> >>>>>>>>>
> >>>>>>>>> I have attached a trivial patch that I think disables the expor=
t
> >>>>>>>>> checks for jails. If either of you can try it and see if it fix=
es
> >>>>>>>>> the problem, that would be great.
> >>>>>>>>> (Note that this is only for testing, although it probably does =
not
> >>>>>>>>> matter unless you are running nfsd(8) in vnet jails.)
> >>>>>>>>
> >>>>>>>> Yes, I can see snapshots with the patch. This system is just a =
test
> >>>>>>>> system that doesn't normally run ZFS or NFS, so no problem messi=
ng
> >>>>>>>> with permissions. It's a bhyve VM, so I just added a small disk=
and
> >>>>>>>> enabled ZFS for testing.
> >>>>>>>
> >>>>>>> btw, you might try to get mm@ or maybe mav@ to help out from the =
ZFS
> >>>>>>> side. It must be doing something differently inside a snapshot t=
han
> >>>>>>> outside, maybe with file handles or something like that.
> >>>>>> Yes. I've added freebsd-current@ (although Garrett is not on it, h=
e is
> >>>>>> cc'd) and these guys specifically...
> >>>>>>
> >>>>>> So, here's what appears to be the problem...
> >>>>>> Commit 88175af (in main and stable/13, but not 13.2) added checks =
for
> >>>>>> nfsd(8) running in jails by filling in mnt_exjail with a reference=
to the cred
> >>>>>> used when the file system is exported.
> >>>>>> When mnt_exjail is found NULL, the current nfsd code assumes that =
there
> >>>>>> is no access allowed for the mount.
> >>>>>>
> >>>>>> My vague understanding is that when a ZFS snapshot is accessed, it=
is
> >>>>>> "pseudo-mounted" by zfsctl_snapdir_lookup() and I am guessing that
> >>>>>> mnt_exjail is NULL as a result.
> >>>>>> Since I do not know the ZFS code and don't even have an easy way t=
o
> >>>>>> test this (thankfully Mike can test easily), I do not know what to=
do from
> >>>>>> here?
> >>>>>>
> >>>>>> Is there a "struct mount" constructed for this pseudo mount
> >>>>>> (or it actually appears to be the lookup of ".." that fails, so it
> >>>>>> might be the parent of the snapshot subdir?)?
> >>>>>>
> >>>>>> One thought is that I can check to see if the mount pointer is in =
the
> >>>>>> mountlist (I don't think the snapshot's mount is in the mountlist)=
and
> >>>>>> avoid the jail test for this case. This would assume that snapsho=
ts are
> >>>>>> always within the file system(s) exported via that jail (which inc=
ludes
> >>>>>> the case of prison0, of course), so that they do not need a separa=
te
> >>>>>> jail check.
> >>>>>>
> >>>>>> If this doesn't work, there will need to be some sort of messing a=
bout
> >>>>>> in ZFS to set mnt_exjail for these.
> >>>>> Ok, so now onto the hard part...
> >>>>> Thanks to Mike and others, I did create a snapshot under .zfs and I=
can
> >>>>> see the problem. It is that mnt_exjail =3D=3D NULL.
> >>>>> Now, is there a way that this "struct mount" can be recognized as "=
special"
> >>>>> for snapshots, so I can avoid the mnt_exjail =3D=3D NULL test?
> >>>>> (I had hoped that "mp->mnt_list.tqe_prev" would be NULL, but that i=
s not
> >>>>> the case.)
> >>>>
> >>>> Dumb question, is the mount point (mp presumably) different between =
the
> >>>> snapshot and the main file system?
> >>> Not a dump question and the answer is rather interesting...
> >>> It is "sometimes" or "usually" according to my printf().
> >>> It seems that when you first "cd <snapshot-name"" you get a different=
mp
> >>> where mnt_exjail =3D=3D NULL.. Then when you look at directories with=
in the
> >>> snapshot, you get the mp of the file system that .zfs exists in, whic=
h does
> >>> have mnt_exjail set non-NULL.
> >>>
> >>> There is this snippet of code in zfsctl_snapdir_lookup():
> >>> /*
> >>> * Fix up the root vnode mounted on .zfs/snapshot/<snapname>.
> >>> *
> >>> * This is where we lie about our v_vfsp in order to
> >>> * make .zfs/snapshot/<snapname> accessible over NFS
> >>> * without requiring manual mounts of <snapname>.
> >>> */
> >>> ASSERT3P(VTOZ(*vpp)->z_zfsvfs, !=3D, zfsvfs);
> >>> VTOZ(*vpp)->z_zfsvfs->z_parent =3D zfsvfs;
> >>>
> >>> /* Clear the root flag (set via VFS_ROOT) as well. */
> >>> (*vpp)->v_vflag &=3D ~VV_ROOT;
> >>> which seems to set the mp to that of the parent, but it
> >>> seems this does not happen for the initial lookup of
> >>> the <snapname>?
> >>>
> >>> I'll note that there is code before this in
> >>> zfsctl_snapdir_lookup() for handling cases
> >>> like "." and ".." that return without doing this.
> >>>
> >>> Now, why does this work without the mnt_exjail
> >>> check (as in 13.2)?
> >>> I am not quite sure, but there is this "cheat" in the
> >>> NFS server (it has been there for years, maybe decades):
> >>> /*
> >>> * Allow a Lookup, Getattr, GetFH, Secinfo on an
> >>> * non-exported directory if
> >>> * nfs_rootfhset. Do I need to allow any other Ops?
> >>> * (You can only have a non-exported vpnes if
> >>> * nfs_rootfhset is true. See nfsd_fhtovp())
> >>> * Allow AUTH_SYS to be used for file systems
> >>> * exported GSS only for certain Ops, to allow
> >>> * clients to do mounts more easily.
> >>> */
> >>> if (nfsv4_opflag[op].needscfh && vp) {
> >>> if (!NFSVNO_EXPORTED(&vpnes) &&
> >>> op !=3D NFSV4OP_LOOKUP &&
> >>> op !=3D NFSV4OP_GETATTR &&
> >>> op !=3D NFSV4OP_GETFH &&
> >>> op !=3D NFSV4OP_ACCESS &&
> >>> op !=3D NFSV4OP_READLINK &&
> >>> op !=3D NFSV4OP_SECINFO &&
> >>> op !=3D NFSV4OP_SECINFONONAME)
> >>> nd->nd_repstat =3D NFSERR_NOFILEHANDLE;
> >>> This allows certain operations to be done on
> >>> non-exported file systems and I think that is enough
> >>> to allow this to work when mnt_exjail is not checked.
> >>> (Note that NFSV4OP_LOOKUPP is not in the list,
> >>> which might explain why it is the one that fails for
> >>> Garrett. I don't think it can be added to this list
> >>> safely, since that would allow a client to move above
> >>> the exported file system into "uncharted territory".)
> >>>
> >>>> Just curious. Also, what is mnt_exjail
> >>>> normally set to for file systems not in a jail?
> >>> mnt_exjail is set to the credentials of the thread/process
> >>> that exported the file system (usually mountd(8)).
> >>> When not in a jail, cr_prison for these credentials
> >>> points to prison0.
> >>>
> >>> Btw, I checked and the "other mp that has mnt_exjail =3D=3D NULL
> >>> is in the mountlist, so the idea of checking "not in mountlist"
> >>> is a dead end.
> >>>
> >>> I am looking for something "unique" about this other mp,
> >>> but haven't found anything yet.
> >>> Alternately, it might be necessary to add code to
> >>> zfsctl_snapdir_lookup() to "cheat and change the mp"
> >>> in more cases, such as "." and ".." lookups?
> >>
> >> It seems to me that if ZFS is creating an additional mount structure,
> >> it should be responsible for setting it up correctly. That could
> >> involve a vfs-level routine to do some of the cloning. In any case,
> >> it seems to me that mnt_exjail should be set properly, e.g. by duping
> >> the one in the original mount structure. Probably ZFS is the only
> >> file system type that would need this added.
> > I've created a patch that I think does this. It seemed to test ok for m=
y case.
> > It's in D42672 on reviews.freebsd.org.
> > I have also attached it here (this diff doesn't use -U999999, so it mig=
ht be
> > easier to apply).
>
> It works for me in 13-stable, and looks plausible.
The patch has now been fixed so that there is no race between vfs_exjail_cl=
one()
and a jail dying. I believe the current patch fixes the problem.
The patch can be found as an attachment to PR#275200 (that is a FreeBSD bug=
zilla
problem report, not an OpenZFS pull request). I will not close the PR until
the above has all happened.
This patch will be needed for systems running stable/13, stable/14 and 14.0=
.
The vfs_exjail_clone() part of the patch is now committed to main and will =
be
MFC'd in 3 days. The ZFS part of the patch is a pull request on OpenZFS.
I cannot say how long the ZFS part takes or if/when an errata to 14.0 will
happen.
Sorry about the breakage and thanks for reporting it (and persevering when
I said I knew nothing about ZFS, which is true). Thanks goes to Mike Karels=
,
who was able to determine that the problem was not in 13.2, but was in
stable/13,
which allowed me to conclude it was a "nfsd in jail" related breakage.
rick
>
> Mike
>
> > Hopefully this fixes the problem. Sorry for the breakage.
> >
> > rick
> >
> >>
> >> Mike
> >>
> >>> rick
> >>> ps: I added all the cc's back in because I want the
> >>> ZFS folk to hopefully chime in.
> >>>
> >>>>
> >>>> Mike
> >>>>
> >>>>> Do I need to search mountlist for it?
> >>>>>
> >>>>> rick
> >>>>> ps: The hack patch attached should fix the problem, but can only be
> >>>>> safely used if mountd/nfsd are not run in any jails.
> >>>>>
> >>>>>>
> >>>>>> I will try and get a test setup going here, which leads me to..
> >>>>>> how do I create a ZFS snapshot? (I do have a simple ZFS pool runni=
ng
> >>>>>> on a test machine, but I've never done a snapshot.)
> >>>>>>
> >>>>>> Although this problem is not in 13.2, it will have shipped in 14.0=
.
> >>>>>>
> >>>>>> Any help with be appreciated, rick
> >>>>>>
> >>>>>>>
> >>>>>>> Mike
> >>>>>>>>
> >>>>>>>>> rick
> >>>>>>>>>
> >>>>>>>>> On Fri, Nov 17, 2023 at 6:14=E2=80=AFPM Mike Karels <mike@karel=
s.net> wrote:
> >>>>>>>>>>
> >>>>>>>>>> CAUTION: This email originated from outside of the University =
of Guelph. Do not click links or open attachments unless you recognize the =
sender and know the content is safe. If in doubt, forward suspicious emails=
to IThelp@uoguelph.ca.
> >>>>>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> Rick, have you been following this thread on freebsd-stable? =
I have been able
> >>>>>>>>>> to reproduce this using a 13-stable server from Oct 7 and a 15=
-current system
> >>>>>>>>>> that is up to date using NFSv3. I did not reproduce with a 13=
.2 server. The
> >>>>>>>>>> client was running 13.2. Any ideas? A full bisect seems fair=
ly painful, but
> >>>>>>>>>> maybe you have an idea of points to try. Fortunately, these a=
re all test
> >>>>>>>>>> systems that I can reboot at will.
> >>>>>>>>>>
> >>>>>>>>>> Mike
> >>>>>>>>>>
> >>>>>>>>>> Forwarded message:
> >>>>>>>>>>
> >>>>>>>>>>> From: Garrett Wollman <wollman@bimajority.org>
> >>>>>>>>>>> To: Mike Karels <mike@karels.net>
> >>>>>>>>>>> Cc: freebsd-stable@freebsd.org
> >>>>>>>>>>> Subject: Re: NFS exports of ZFS snapshots broken
> >>>>>>>>>>> Date: Fri, 17 Nov 2023 17:35:04 -0500
> >>>>>>>>>>>
> >>>>>>>>>>> <<On Fri, 17 Nov 2023 15:57:42 -0600, Mike Karels <mike@karel=
s.net> said:
> >>>>>>>>>>>
> >>>>>>>>>>>> I have not run into this, so I tried it just now. I had no =
problem.
> >>>>>>>>>>>> The server is 13.2, fully patched, the client is up-to-date =
-current,
> >>>>>>>>>>>> and the mount is v4.
> >>>>>>>>>>>
> >>>>>>>>>>> On my 13.2 client and 13-stable server, I see:
> >>>>>>>>>>>
> >>>>>>>>>>> 25034 ls CALL open(0x237d32f9a000,0x120004<O_RDONLY|O=
_NONBLOCK|O_DIRECTORY|O_CLOEXEC>)
> >>>>>>>>>>> 25034 ls NAMI "/mnt/tools/.zfs/snapshot/weekly-2023-4=
5"
> >>>>>>>>>>> 25034 ls RET open 4
> >>>>>>>>>>> 25034 ls CALL fcntl(0x4,F_ISUNIONSTACK,0x0)
> >>>>>>>>>>> 25034 ls RET fcntl 0
> >>>>>>>>>>> 25034 ls CALL getdirentries(0x4,0x237d32faa000,0x1000=
,0x237d32fa7028)
> >>>>>>>>>>> 25034 ls RET getdirentries -1 errno 5 Input/output e=
rror
> >>>>>>>>>>> 25034 ls CALL close(0x4)
> >>>>>>>>>>> 25034 ls RET close 0
> >>>>>>>>>>> 25034 ls CALL exit(0)
> >>>>>>>>>>>
> >>>>>>>>>>> Certainly a libc bug here that getdirentries(2) returning [EI=
O]
> >>>>>>>>>>> results in ls(1) returning EXIT_SUCCESS, but the [EIO] error =
is
> >>>>>>>>>>> consistent across both FreeBSD and Linux clients.
> >>>>>>>>>>>
> >>>>>>>>>>> Looking at this from the RPC side:
> >>>>>>>>>>>
> >>>>>>>>>>> (PUTFH, GETATTR, LOOKUP(snapshotname), GETFH, GETATTR)
> >>>>>>>>>>> [NFS4_OK for all ops]
> >>>>>>>>>>> (PUTFH, GETATTR)
> >>>>>>>>>>> [NFS4_OK, NFS4_OK]
> >>>>>>>>>>> (PUTFH, ACCESS(0x3f), GETATTR)
> >>>>>>>>>>> [NFS4_OK, NFS4_OK, rights =3D 0x03, NFS4_OK]
> >>>>>>>>>>> (PUTFH, GETATTR, LOOKUPP, GETFH, GETATTR)
> >>>>>>>>>>> [NFS4_OK, NFS4_OK, NFS4ERR_NOFILEHANDLE]
> >>>>>>>>>>>
> >>>>>>>>>>> and at this point the [EIO] is returned.
> >>>>>>>>>>>
> >>>>>>>>>>> It seems that clients always do a LOOKUPP before calling READ=
DIR, and
> >>>>>>>>>>> this is failing when the subject file handle is the snapshot.=
The
> >>>>>>>>>>> client is perfectly able to *traverse into* the snapshot: if =
I try to
> >>>>>>>>>>> list a subdirectory I know exists in the snapshot, the client=
is able to
> >>>>>>>>>>> LOOKUP(dirname) just fine, but LOOKUPP still fails with
> >>>>>>>>>>> NFS4ERR_NOFILEHANDLE *on the subndirectory*.
> >>>>>>>>>>>
> >>>>>>>>>>> -GAWollman
> >>>>>>>>>>
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAM5tNy7b_uWsvRaWOGd-CD=Sr4yZwVmepnqfNR_EUXVm_OjN2g>