| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <6bdec87c369b61b515a29a6d661b1e9473fba24c.camel@kernel.org>
Date: Fri, 21 Feb 2025 09:58:32 -0500
From: Jeff Layton <jlayton@...nel.org>
To: Benjamin Coddington <bcodding@...hat.com>
Cc: Chuck Lever <chuck.lever@...cle.com>, Li Lingfeng
<lilingfeng3@...wei.com>, neilb@...e.de, okorniev@...hat.com,
Dai.Ngo@...cle.com, tom@...pey.com, linux-nfs@...r.kernel.org,
linux-kernel@...r.kernel.org, yukuai1@...weicloud.com, houtao1@...wei.com,
yi.zhang@...wei.com, yangerkun@...wei.com, lilingfeng@...weicloud.com
Subject: Re: [PATCH] nfsd: decrease cl_cb_inflight if fail to queue cb_work
On Fri, 2025-02-21 at 09:46 -0500, Benjamin Coddington wrote:
> On 21 Feb 2025, at 9:37, Jeff Layton wrote:
>
> > On Fri, 2025-02-21 at 09:06 -0500, Benjamin Coddington wrote:
> > > On 18 Feb 2025, at 9:40, Jeff Layton wrote:
> > >
> > > > On Tue, 2025-02-18 at 09:31 -0500, Chuck Lever wrote:
> > > > > On 2/18/25 9:29 AM, Jeff Layton wrote:
> > > > > > On Tue, 2025-02-18 at 08:58 -0500, Jeff Layton wrote:
> > > > > > > On Tue, 2025-02-18 at 21:54 +0800, Li Lingfeng wrote:
> > > > > > > > In nfsd4_run_cb, cl_cb_inflight is increased before attempting to queue
> > > > > > > > cb_work to callback_wq. This count can be decreased in three situations:
> > > > > > > > 1) If queuing fails in nfsd4_run_cb, the count will be decremented
> > > > > > > > accordingly.
> > > > > > > > 2) After cb_work is running, the count is decreased in the exception
> > > > > > > > branch of nfsd4_run_cb_work via nfsd41_destroy_cb.
> > > > > > > > 3) The count is decreased in the release callback of rpc_task — either
> > > > > > > > directly calling nfsd41_cb_inflight_end in nfsd4_cb_probe_release, or
> > > > > > > > calling nfsd41_destroy_cb in .
> > > > > > > >
> > > > > > > > However, in nfsd4_cb_release, if the current cb_work needs to restart, the
> > > > > > > > count will not be decreased, with the expectation that it will be reduced
> > > > > > > > once cb_work is running.
> > > > > > > > If queuing fails here, then the count will leak, ultimately causing the
> > > > > > > > nfsd service to be unable to exit as shown below:
> > > > > > > > [root@..._test2 ~]# cat /proc/2271/stack
> > > > > > > > [<0>] nfsd4_shutdown_callback+0x22b/0x290
> > > > > > > > [<0>] __destroy_client+0x3cd/0x5c0
> > > > > > > > [<0>] nfs4_state_destroy_net+0xd2/0x330
> > > > > > > > [<0>] nfs4_state_shutdown_net+0x2ad/0x410
> > > > > > > > [<0>] nfsd_shutdown_net+0xb7/0x250
> > > > > > > > [<0>] nfsd_last_thread+0x15f/0x2a0
> > > > > > > > [<0>] nfsd_svc+0x388/0x3f0
> > > > > > > > [<0>] write_threads+0x17e/0x2b0
> > > > > > > > [<0>] nfsctl_transaction_write+0x91/0xf0
> > > > > > > > [<0>] vfs_write+0x1c4/0x750
> > > > > > > > [<0>] ksys_write+0xcb/0x170
> > > > > > > > [<0>] do_syscall_64+0x70/0x120
> > > > > > > > [<0>] entry_SYSCALL_64_after_hwframe+0x78/0xe2
> > > > > > > > [root@..._test2 ~]#
> > > > > > > >
> > > > > > > > Fix this by decreasing cl_cb_inflight if the restart fails.
> > > > > > > >
> > > > > > > > Fixes: cba5f62b1830 ("nfsd: fix callback restarts")
> > > > > > > > Signed-off-by: Li Lingfeng <lilingfeng3@...wei.com>
> > > > > > > > ---
> > > > > > > > fs/nfsd/nfs4callback.c | 10 +++++++---
> > > > > > > > 1 file changed, 7 insertions(+), 3 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/fs/nfsd/nfs4callback.c b/fs/nfsd/nfs4callback.c
> > > > > > > > index 484077200c5d..8a7d24efdd08 100644
> > > > > > > > --- a/fs/nfsd/nfs4callback.c
> > > > > > > > +++ b/fs/nfsd/nfs4callback.c
> > > > > > > > @@ -1459,12 +1459,16 @@ static void nfsd4_cb_done(struct rpc_task *task, void *calldata)
> > > > > > > > static void nfsd4_cb_release(void *calldata)
> > > > > > > > {
> > > > > > > > struct nfsd4_callback *cb = calldata;
> > > > > > > > + struct nfs4_client *clp = cb->cb_clp;
> > > > > > > > + int queued;
> > > > > > > >
> > > > > > > > trace_nfsd_cb_rpc_release(cb->cb_clp);
> > > > > > > >
> > > > > > > > - if (cb->cb_need_restart)
> > > > > > > > - nfsd4_queue_cb(cb);
> > > > > > > > - else
> > > > > > > > + if (cb->cb_need_restart) {
> > > > > > > > + queued = nfsd4_queue_cb(cb);
> > > > > > > > + if (!queued)
> > > > > > > > + nfsd41_cb_inflight_end(clp);
> > > > > > > > + } else
> > > > > > > > nfsd41_destroy_cb(cb);
> > > > > > > >
> > > > > > > > }
> > > > > > >
> > > > > > > Good catch!
> > > > > > >
> > > > > > > Reviewed-by: Jeff Layton <jlayton@...nel.org>
> > > > > > >
> > > > > >
> > > > > > Actually, I think this is not quite right. It's a bit more subtle than
> > > > > > it first appears. The problem of course is that the callback workqueue
> > > > > > jobs run in a different task than the RPC workqueue jobs, so they can
> > > > > > race.
> > > > > >
> > > > > > cl_cb_inflight gets bumped when the callback is first queued, and only
> > > > > > gets released in nfsd41_destroy_cb(). If it fails to be queued, it's
> > > > > > because something else has queued the workqueue job in the meantime.
> > > > > >
> > > > > > There are two places that can occur: nfsd4_cb_release() and
> > > > > > nfsd4_run_cb(). Since this is occurring in nfsd4_cb_release(), the only
> > > > > > other option is that something raced in and queued it via
> > > > > > nfsd4_run_cb().
> > > > >
> > > > > What would be the "something" that raced in?
> > > > >
> > > >
> > > > I think we may be able to get there via multiple __break_lease() calls
> > > > on the same layout or delegation. That could mean multiple calls to the
> > > > ->lm_break operation on the same object.
> > >
> > > Sorry for the late response, but isn't ->lm_break() already serialized in
> > > __break_lease for the same file_lease? We don't call lm_break(fl) if
> > > lease_breaking(fl).
> > >
> >
> > lease_breaking() is only checked when want_write is false. IOW, if
> > you're breaking the lease for write, then lm_break is always called.
> >
> > Is that a bug or a feature? I'm not sure, but it's been that way since
> > ~2011.
>
> Yeah.. why?
>
> Thanks I missed that detail when I refreshed my memory of it just now.
> Seems like you'd want to avoid constantly calling lm_break for both cases,
> spamming the lock manager adds nothing. 2 cents.
>
Sorry, it doesn't get called every time. If want_write is called then
we _do_ check FL_UNLOCK_PENDING. IOW, you can get a downgrade attempt
first, and then a full unlock request later.
nfsd doesn't do downgrades. It recalls the object either way. So,
ignoring subsequent lm_break calls is the right thing to do, I think.
--
Jeff Layton <jlayton@...nel.org>
Powered by blists - more mailing lists