Re: [PATCH] fuse: defer dentry_tree_work until fuse_dentry_tree_init()

From: Sang-Heon Jeon

Date: Wed Nov 11 2026 - 10:25:27 EST


2026년 11월 11일 (수) 오후 10:58, Luis Henriques <luis@xxxxxxxxxx>님이 작성:


>
> On Wed, Nov 11 2026, Sang-Heon Jeon wrote:
>
> > 2026년 11월 11일 (수) 오후 8:27, Sang-Heon Jeon <ekffu200098@xxxxxxxxx>님이 작성:
> >>
> >> Hi, Luis
> >>
> >> 2026년 11월 11일 (수) 오후 6:28, Luis Henriques <luis@xxxxxxxxxx>님이 작성:
> >> >
> >> > On Wed, Nov 11 2026, Sang-Heon Jeon wrote:
> >> >
> >> > > inval_wq_set() schedules dentry_tree_work when inval_wq is enabled.
> >> > >
> >> > > But inval_wq_set() can be called at boot or module load before
> >> > > fuse_dentry_tree_init() initializes the work, so __queue_delayed_work()
> >> > > warns and the work never starts.
> >> > >
> >> > > So defer scheduling the work until fuse_dentry_tree_init() sets it up.
> >> > >
> >> > > Fixes: ab84ad597386 ("fuse: new work queue to periodically invalidate expired dentries")
> >> > > Cc: stable@xxxxxxxxxxxxxxx
> >> > > Signed-off-by: Sang-Heon Jeon <ekffu200098@xxxxxxxxx>
> >> > > ---
> >> > > QEMU based test result
> >> > > - booted with "fuse.inval_wq=5"
> >> >
> >> > Thank you, Sang-Heon. I've sent a fix for this already[1], but sashiko
> >> > still has some race complains which I haven't looked into yet. But maybe
> >> > Miklos is happier with your patch, in which case you can add my
> >>
> >> Sorry, I missed that thread.. sashiko also complains about the same
> >> race condition between init and sysfs setup. But I'm not sure it is a
> >> realistic scenario. So I want to hear other voices as well.
> >
> > I also checked the other sashiko's complaints from your patch, the one
> > about race between module cleanup and sysfs. It looks easier to
> > reproduce, but still I'm not sure this is a realistic problem. If it
> > is worth fixing, I will sent v2 with serialization between module
> > init, cleanup and inval_wq_set()
>
> I agree that the first race doesn't look realistic, though it could be
> fixed by adding an extra spinlock (which could protect both inval_wq and
> dentry_tree_is_initialised).
>
> The second race would be easier to fix by setting
> dentry_tree_is_initialised back to 'false' in fuse_dentry_tree_cleanup().

But I think race can still exist with that.

CPU0
CPU1

inval_wq_set()

READ_ONCE(dentry_tree_is_initialised) // true

<sleep>
fuse_dentry_tree_cleanup()
WRITE_ONCE(dentry_tree_is_initialised, false);
cancel_delayed_work_sync(&dentry_tree_work);

<wakeup>

mod_delayed_work(&dentry_tree_work) // already
false but executed


fuse_dentry_tree_work() // later can cause
UAF?

IMO, if we need to fix the first and second race both, then I prefer
to use lock. But if we need to fix the second race only, then
disable_delayed_work_sync(sashiko's suggestion) looks better.

Please correct me if I misunderstand something.

> One thing I like with my version of the fix is that it replaces the usage
> of schedule_delayed_work() by mod_delayed_work(). But maybe that's out
> of the scope of this fix, and I should send that separately.

I do believe that is a great enhancement. But as you said it could be
applied separately :)

> Cheers,
> --
> Luís
>
>
> >> > Reviewed-by: Luis Henriques <luis@xxxxxxxxxx>
> >>
> >> Thanks for your review :)
> >>
> >> > [1] https://lore.kernel.org/20261106155512.24801-1-luis%40igalia.com
> >> >
> >> > Cheers,
> >> > --
> >> > Luís
> >> >
> >> > > 1) AS-IS (before fix)
> >> > > [ 0.011073] Kernel command line: rootwait root=/dev/vda console=tty1 console=ttyS0 nokaslr fuse.inval_wq=5
> >> > > [ 0.011114] ------------[ cut here ]------------
> >> > > [ 0.011115] WARNING: kernel/workqueue.c:2704 at __queue_delayed_work+0x11a/0x140, CPU#0: swapper/0
> >> > > [ 0.011124] Modules linked in:
> >> > > [ 0.011128] CPU: 0 UID: 0 PID: 0 Comm: swapper Not tainted 7.4.0-rc1-00010-gab9787d7e66a #1 PREEMPT(undef)
> >> > > [ 0.011130] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-debian-1.17.0-1ubuntu1 04/01/2014
> >> > > [ 0.011131] RIP: 0010:__queue_delayed_work+0x11a/0x140
> >> > > [ 0.011133] Code: c4 08 48 89 df 5b e9 85 28 0a 00 90 0f 0b 90 4c 8b 42 08 48 8d
> >> > > 42 08 49 39 c0 0f 84 1c ff ff ff 90 0f 0b 90 e9 13 ff ff ff 90 <0f> 0b 90 e9 ec
> >> > > fe ff ff bff
> >> > > [ 0.011134] RSP: 0000:ffffffff82e03e00 EFLAGS: 00010003 ORIG_RAX: 0000000000000000
> >> > > [ 0.011135] RAX: 0000000000000000 RBX: 0000000000000040 RCX: 0000000000001388
> >> > > [ 0.011136] RDX: ffffffff83a3f320 RSI: 0000000000000000 RDI: 0000000000000040
> >> > > [ 0.011136] RBP: 0000000000000000 R08: 0000000000000082 R09: 0000000000000001
> >> > > [ 0.011136] R10: 000000000000000a R11: 000000000000000a R12: ffffffff82c09fd0
> >> > > [ 0.011137] R13: ffff88807ffcfb89 R14: 0000000000000000 R15: 0000000000001388
> >> > > [ 0.011140] FS: 0000000000000000(0000) GS:ffff8880fa26e000(0000) knlGS:0000000000000000
> >> > > [ 0.011141] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> >> > > [ 0.011141] CR2: ffff888003a67000 CR3: 0000000002e30000 CR4: 00000000000000b0
> >> > > [ 0.011142] Call Trace:
> >> > > [ 0.011143] <TASK>
> >> > > [ 0.011143] queue_delayed_work_on+0x6a/0x80
> >> > > [ 0.011145] inval_wq_set+0xc4/0xe0
> >> > > [ 0.011149] parse_args+0x299/0x3c0
> >> > > [ 0.011151] ? __pfx_unknown_bootoption+0x10/0x10
> >> > > [ 0.011154] start_kernel+0x3b8/0x7c0
> >> > > [ 0.011156] ? __pfx_unknown_bootoption+0x10/0x10
> >> > > [ 0.011157] x86_64_start_reservations+0x24/0x30
> >> > > [ 0.011160] x86_64_start_kernel+0xd1/0xe0
> >> > > [ 0.011162] common_startup_64+0x13e/0x158
> >> > > [ 0.011165] </TASK>
> >> > > [ 0.011166] ---[ end trace 0000000000000000 ]---
> >> > > [ 0.011173] ------------[ cut here ]------------
> >> > >
> >> > > 2) TO-BE (after fix)
> >> > > - no error while booting, and also fuse_dentry_tree_work works well.
> >> > >
> >> > > ---
> >> > > fs/fuse/dir.c | 10 ++++++++++
> >> > > 1 file changed, 10 insertions(+)
> >> > >
> >> > > diff --git a/fs/fuse/dir.c b/fs/fuse/dir.c
> >> > > index da23ccd8685c..6965a77a871f 100644
> >> > > --- a/fs/fuse/dir.c
> >> > > +++ b/fs/fuse/dir.c
> >> > > @@ -39,6 +39,7 @@ static struct delayed_work dentry_tree_work;
> >> > > /* Minimum invalidation work queue frequency */
> >> > > #define FUSE_DENTRY_INVAL_FREQ_MIN 5
> >> > >
> >> > > +static bool inval_wq_ready;
> >> > > unsigned __read_mostly inval_wq;
> >> > > static int inval_wq_set(const char *val, const struct kernel_param *kp)
> >> > > {
> >> > > @@ -62,6 +63,10 @@ static int inval_wq_set(const char *val, const struct kernel_param *kp)
> >> > >
> >> > > *((unsigned int *)kp->arg) = num;
> >> > >
> >> > > + /* Defer the work until fuse_dentry_tree_init() sets it up. */
> >> > > + if (!inval_wq_ready)
> >> > > + return 0;
> >> > > +
> >> > > if (num && !old)
> >> > > schedule_delayed_work(&dentry_tree_work,
> >> > > secs_to_jiffies(num));
> >> > > @@ -224,6 +229,11 @@ void fuse_dentry_tree_init(void)
> >> > > dentry_hash[i].tree = RB_ROOT;
> >> > > }
> >> > > INIT_DELAYED_WORK(&dentry_tree_work, fuse_dentry_tree_work);
> >> > > + inval_wq_ready = true;
> >> > > +
> >> > > + if (inval_wq)
> >> > > + schedule_delayed_work(&dentry_tree_work,
> >> > > + secs_to_jiffies(inval_wq));
> >> > > }
> >> > >
> >> > > void fuse_dentry_tree_cleanup(void)
> >> > >
> >> > > base-commit: ab9787d7e66ac7e6e94fc975b37c5b56fff98ca9
> >> > > --
> >> > > 2.53.0
> >> > >
> >> >
> >>
> >> Best regards,
> >> Sang-Heon Jeon

Best regards,
Sang-Heon Jeon