Date: Sat, 18 Nov 2023 19:47:13 +0100 From: Ronald Klop <ronald@FreeBSD.org> To: FreeBSD CURRENT <freebsd-current@freebsd.org>, Rick Macklem <rmacklem@freebsd.org> Subject: Re: NFS exports of ZFS snapshots broken Message-ID: <805acb16-ee6f-4d09-b01b-39d9ae3f8c86@FreeBSD.org> In-Reply-To: <CAM5tNy5LGWF424i9v_w2u%2B41wtzmW9kn6RZWwo6Pi0Y4wj92UA@mail.gmail.com> 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>
next in thread | previous in thread | raw e-mail | index | archive | help
On 11/18/23 17:09, Rick Macklem wrote: > On Fri, Nov 17, 2023 at 8:19 PM Mike Karels <mike@karels.net> 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 export >>>> checks for jails. If either of you can try it and see if it fixes >>>> 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 messing >>> 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 than >> outside, maybe with file handles or something like that. > Yes. I've added freebsd-current@ (although Garrett is not on it, he 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 to > 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 snapshots are > always within the file system(s) exported via that jail (which includes > the case of prison0, of course), so that they do not need a separate > jail check. > > If this doesn't work, there will need to be some sort of messing about > in ZFS to set mnt_exjail for these. > > 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 running > on a test machine, but I've never done a snapshot.) # zfs list ... zroot/usr/local 4.59G 27.5G 2.76G /usr/local zroot/usr/ports 1.03G 27.5G 952M /usr/ports ... # zfs snapshot zroot/usr/local@myfirstsnapshot -- to view them # zfs list -t snapshot zroot/usr/local -- and to remove it: # zfs destroy zroot/usr/local@myfirstsnapshot -- more info # man zfs-snapshot If you get used to this you are going to love it. :-) Regards and happy hacking, Ronald. > > 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 PM Mike Karels <mike@karels.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 fairly painful, but >>>>> maybe you have an idea of points to try. Fortunately, these are 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@karels.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-2023-45" >>>>>> 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 error >>>>>> 25034 ls CALL close(0x4) >>>>>> 25034 ls RET close 0 >>>>>> 25034 ls CALL exit(0) >>>>>> >>>>>> Certainly a libc bug here that getdirentries(2) returning [EIO] >>>>>> 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 = 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 READDIR, 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?805acb16-ee6f-4d09-b01b-39d9ae3f8c86>