Date: Fri, 4 Sep 2020 21:57:45 -0600
From: Alan Somers <asomers@freebsd.org>
To: Rick Macklem <rmacklem@uoguelph.ca>
Cc: FreeBSD Hackers <freebsd-hackers@freebsd.org>
Subject: Re: panic!("docallb") in nfsrv_docallback
Message-ID: <CAOtMX2hMbLxaXj3sgWg77knk8nJx3ff=G3j9HbeWaMzJT1TMTg@mail.gmail.com>
In-Reply-To: <YTBPR01MB396648BCDB31564AE81B59B4DD2A0@YTBPR01MB3966.CANPRD01.PROD.OUTLOOK.COM>
References: <CAOtMX2h3ZTVaZB8bdMg3QeGXCLtJLAuOCEQQm-WNTVaJz1HEDA@mail.gmail.com> <YTBPR01MB396648BCDB31564AE81B59B4DD2A0@YTBPR01MB3966.CANPRD01.PROD.OUTLOOK.COM>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, Sep 4, 2020 at 8:58 PM Rick Macklem <rmacklem@uoguelph.ca> wrote: > Alan Somers wrote: > >I just saw this panic on a 12-stable machine. Unfortunately, I don't have > >a core dump, just a stack trace. It was serving NFS v4.0, with > delegations > >enabled. The clients were all Debian, with Linux 3.16.0. > I will generically note that I believe the Linux NFS client developers > mostly > test NFSv4.1, 4.2, so if the clients support NFSv4.1, it might be worth > upgrading? > > Also, delegations aren't enabled by default for a couple of reasons. > 1 - For a long time, Linux only knew how to use read delegations and I felt > (and still feel) they are pretty useless. > 2 - They are complex to get right. > 3 - Although they should reduce the number of Open operations against the > server, I haven't observed dramatic performance improvements because > of them. > > >The proximal cause of the panic seems to be that the file had a write > >delegation issued to an unconfirmed client. Root cause is harder to > >determine. Did the kernel previously issue a delegation to an unconfirmed > >client? Or did the client somehow change to an unconfirmed state after > the > >delegation was issued, perhaps due to a race? > I think the first case is more likely. Since client confirmation happens > immediately > for NFSv4.1 (the ExchangeID and Createsession must occur before anything > else can happen), I wouldn;t be surprised if the Linux client tries to do > an Open > before the SetClientIDConfirm has completed for NFSv4.0. > > >It's hard to tell, but I don't see any checks for lc_flags & > >LCL_NEEDSCONFIRM in nfsrv_openctrl (which issues the delegations), so I'm > >guessing that that's the problem. > The server should definitely check for a confirmed ClientID during Open and > fail any Open attempt where that is not the case. > --> I'll take a look at the code. I wrote it about 20years ago, but I can > probably > figure out how it works.;-) > > > If so, then the event trace would look > >like this: > > > >1) Client Alice sends SETCLIENTID. The server creates a client state > >structure > > for her. > >_) Client Alice should've sent SETCLIENTID_CONFIRM, but doesn't. Bad > Alice! > >2) Client Alice sends OPEN for some file, and is issued a write > delegation. > > The server shouldn't have issued it, because Alice's client ID is > > unconfirmed. Bad server! > >3) Client Bob tries to do a GETATTR on that same file. > >4) In nfsrv_checkgetattr, the kernel finds a write delegation for that > file, > > owned by client Alice. > >5) The kernel tries to send a NFSV4OP_CBGETATTR callback to Alice, to see > >if the > > file's attributes have changed. > >6) But Alice's client ID is unconfirmed. Oh no! Panic! > > > >Does this sound plausible? Should there be a check for LCL_NEEDSCONFIRM > >somewhere around line 3166 in nfs_nfsdstate.c? Grateful for any help. > Yes, it does. I would have thought that I'd have checked for the > unconfirmed > ClientID, but maybe not. > > It is also possible that the client somehow did a SetClientID after the > Open > that issued the delegation, putting it back in "unconfirmed" state. > It that was the case, maybe the panic(), intended to catch corrupted data > structures, was overkill. > > >-Alan > > > >P.S.: stack trace > > > >kdb_backtrace > >vpanic > >panic > >nfsrv_docallback > >nfsrv_checkgetattr > nfsrv_checkgetattr() should probably check for the case of an unconfirmed > clientid and then return ignoring any delegations hanging off it instead > of attempting a callback. > --> This would handle the case where the client did a SetClientID after the > Open that acquired the delegation, leaving the ClientID unconfirmed. > - The two RPCs doing SetClientID and SetClientIDConfirm are normally > done only upon mounting or when the client thinks it has lost the > ClientID due to a lease expiry, but there is also the case where > it is > changing the callback address. (This could explain the SetClientID > happening after the Open that acquired the delegation.) > --> Hint. Can you now see why NFSv4.1 chose to do things differently? > > nfsrvd_getattr > nfsrvd_dorpc > nfssvc_program > svc_run_internal > svc_thread_start > fork_exit > fork_trampoline > > Thanks for reporting it. I'll take a look, rick > Wow, thanks! That's better feedback than I ever hoped for. About fixing the bug: my odds of reproducing it in the original setting are low, since it's a production server. And I doubt I could force it via normal traffic to a development server, either. It took many terabytes of traffic to hit the bug this once. I think my best shot would be to use LibNFS ( https://github.com/sahlberg/libnfs) to write a custom misbehaving NFS client. But that's daunting; the only NFSv4 examples are nearly 1,000 lines each. But, I could probably do it with enough time. Do you have any better ideas? -Alan
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAOtMX2hMbLxaXj3sgWg77knk8nJx3ff=G3j9HbeWaMzJT1TMTg>