Skip to contentSkip to conversation

fuse: clear stale intr_entry in fuse_remove_pending_req()

PATCH In review Baokun Li 3 participants 4 messages Jul 28 – Aug 14

Subsystems (2)
lore
Share & tools
lore

Baokun Li

author
#0 lore raw

PATCH 1/1 fuse: clear stale intr_entry in fuse_remove_pending_req()

Commit f8fce75fedf7 ("fuse: clear intr_entry in fuse_resend and fuse_remove_pending_req") removes stale interrupt entries in fuse_chan_resend() when requests are moved back to fiq->pending. However, that cleanup only covers interrupt entries that are already linked at scan time. It can race with a concurrent queue_interrupt() from the request holder:

CPU 0 (holder thread) CPU 1 (resend)
--------------------- --------------

req in processing (FR_SENT=1)

signal arrives
set_bit(FR_INTERRUPTED)
test_bit(FR_SENT) -> true
queue_interrupt():

    spins on fiq->lock ...
                                  fuse_chan_resend():
                                    set_bit(FR_PENDING)
                                    clear_bit(FR_SENT)
                                    spin_lock(&fiq->lock)
                                    cleanup scan:
                                      intr_entry not linked yet
                                      -> list_del_init is a no-op
                                    list_splice -> fiq->pending
                                    spin_unlock(&fiq->lock)
    ... acquires fiq->lock
    list_empty(&req->intr_entry) -> true
    FR_FINISHED not set
    -> intr_entry added to fiq->interrupts
       AFTER the cleanup already ran

fatal signal arrives
fuse_remove_pending_req():

    test_bit(FR_PENDING) -> true
    list_del(&req->list)
    __fuse_put_request
  fuse_put_request (refcount -> 0)
  -> req freed, intr_entry dangling
     on fiq->interrupts

fuse_dev_queue_interrupt() only checks list_empty() and FR_FINISHED before linking intr_entry -- it does not check FR_PENDING, so a request already spliced back to fiq->pending can still be added to fiq->interrupts. The lock contention itself produces the bad
ordering: while the resend holds fiq->lock to scan the queued requests, the holder spins in queue_interrupt() and links intr_entry right after the scan finishes.

The dangling entry then causes the same use-after-free that the above commit describes: fuse_read_interrupt() writes to the freed slab object via list_del_init() and leaks req->in.h.unique to userspace. Once the freed memory is reused, INIT_LIST_HEAD() turns the entry into a self-loop and list_empty(&fiq->interrupts) returns false forever, so the daemon reads the same phantom FUSE_INTERRUPT in an infinite loop and never consumes fiq->pending.

Close the race in fuse_remove_pending_req(), which is the common bail-out path for both the legacy and the io_uring transport: after the request is removed from the pending queue, also unlink intr_entry under fiq->lock before the reference is dropped. fiq->lock must be taken explicitly since the lock argument is the ring queue lock in the io_uring case, while fiq->interrupts is always protected by fiq->lock. This runs on the holder thread after any queue_interrupt() it issued, and no other path can re-link the entry once the request is off the queues (re-queueing an interrupt from FUSE_INTERRUPT's -EAGAIN reply requires finding the request in the processing queue first).

Fixes 760eac73f9f6 ("fuse: Introduce a new notification type for resend pending requests")
---
 fs/fuse/dev.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
fs/fuse/dev.c +14 -0
index 5763a7cd3b37..87891162985c 100644
@@ -678,6 +678,8 @@ static int queue_interrupt(struct fuse_req *req)
 
 bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock)
 {
+	struct fuse_iqueue *fiq = &req->chan->iq;
+
 	spin_lock(lock);
 	if (test_bit(FR_PENDING, &req->flags)) {
 		/*
@@ -686,6 +688,18 @@ bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock)
 		 */
 		list_del(&req->list);
 		spin_unlock(lock);
+
+		/*
+		 * Remove stale intr_entry queued by queue_interrupt() before
+		 * the request was requeued, which would otherwise dangle on
+		 * fiq->interrupts once the request is freed.
+		 */
+		if (test_bit(FR_INTERRUPTED, &req->flags)) {
+			spin_lock(&fiq->lock);
+			list_del_init(&req->intr_entry);
+			spin_unlock(&fiq->lock);
+		}
+
 		__fuse_put_request(req);
 		req->out.h.error = -EINTR;
 		return true;
Signature · 1 line
2.43.7
Reviewed-by on v1 1/1
#1 lore raw
Quoted 98 lines · Baokun Li

Commit f8fce75fedf7 ("fuse: clear intr_entry in fuse_resend and fuse_remove_pending_req") removes stale interrupt entries in fuse_chan_resend() when requests are moved back to fiq->pending. However, that cleanup only covers interrupt entries that are already linked at scan time. It can race with a concurrent queue_interrupt() from the request holder:

CPU 0 (holder thread) CPU 1 (resend)
--------------------- --------------

req in processing (FR_SENT=1)

signal arrives
set_bit(FR_INTERRUPTED)
test_bit(FR_SENT) -> true
queue_interrupt():

    spins on fiq->lock ...
                                  fuse_chan_resend():
                                    set_bit(FR_PENDING)
                                    clear_bit(FR_SENT)
                                    spin_lock(&fiq->lock)
                                    cleanup scan:
                                      intr_entry not linked yet
                                      -> list_del_init is a no-op
                                    list_splice -> fiq->pending
                                    spin_unlock(&fiq->lock)
    ... acquires fiq->lock
    list_empty(&req->intr_entry) -> true
    FR_FINISHED not set
    -> intr_entry added to fiq->interrupts
       AFTER the cleanup already ran

fatal signal arrives
fuse_remove_pending_req():

    test_bit(FR_PENDING) -> true
    list_del(&req->list)
    __fuse_put_request
  fuse_put_request (refcount -> 0)
  -> req freed, intr_entry dangling
     on fiq->interrupts

fuse_dev_queue_interrupt() only checks list_empty() and FR_FINISHED before linking intr_entry -- it does not check FR_PENDING, so a request already spliced back to fiq->pending can still be added to fiq->interrupts. The lock contention itself produces the bad
ordering: while the resend holds fiq->lock to scan the queued requests, the holder spins in queue_interrupt() and links intr_entry right after the scan finishes.

The dangling entry then causes the same use-after-free that the above commit describes: fuse_read_interrupt() writes to the freed slab object via list_del_init() and leaks req->in.h.unique to userspace. Once the freed memory is reused, INIT_LIST_HEAD() turns the entry into a self-loop and list_empty(&fiq->interrupts) returns false forever, so the daemon reads the same phantom FUSE_INTERRUPT in an infinite loop and never consumes fiq->pending.

Close the race in fuse_remove_pending_req(), which is the common bail-out path for both the legacy and the io_uring transport: after the request is removed from the pending queue, also unlink intr_entry under fiq->lock before the reference is dropped. fiq->lock must be taken explicitly since the lock argument is the ring queue lock in the io_uring case, while fiq->interrupts is always protected by fiq->lock. This runs on the holder thread after any queue_interrupt() it issued, and no other path can re-link the entry once the request is off the queues (re-queueing an interrupt from FUSE_INTERRUPT's -EAGAIN reply requires finding the request in the processing queue first).

Fixes 760eac73f9f6 ("fuse: Introduce a new notification type for resend pending requests")
---
 fs/fuse/dev.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
fs/fuse/dev.c +7 -0
index 5763a7cd3b37..87891162985c 100644
@@ -678,6 +678,8 @@ static int queue_interrupt(struct fuse_req *req)
 
 bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock)
 {
+	struct fuse_iqueue *fiq = &req->chan->iq;
+
 	spin_lock(lock);
 	if (test_bit(FR_PENDING, &req->flags)) {
 		/*
@@ -686,6 +688,18 @@ bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock)
 		 */
 		list_del(&req->list);
 		spin_unlock(lock);
+
+		/*
+		 * Remove stale intr_entry queued by queue_interrupt() before
+		 * the request was requeued, which would otherwise dangle on
+		 * fiq->interrupts once the request is freed.

fs/fuse/dev.c bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock) · line 705

Quoted 10 lines
fs/fuse/dev.c +7 -0
+		 */
+		if (test_bit(FR_INTERRUPTED, &req->flags)) {
+			spin_lock(&fiq->lock);
+			list_del_init(&req->intr_entry);
+			spin_unlock(&fiq->lock);
+		}
+
 		__fuse_put_request(req);
 		req->out.h.error = -EINTR;
 		return true;

LGTM.

Signature · 2 lines
Thanks,
Jingbo
#2 lore raw
In reply to Baokun Li
Quoted 32 lines · Baokun Li

Commit f8fce75fedf7 ("fuse: clear intr_entry in fuse_resend and fuse_remove_pending_req") removes stale interrupt entries in fuse_chan_resend() when requests are moved back to fiq->pending. However, that cleanup only covers interrupt entries that are already linked at scan time. It can race with a concurrent queue_interrupt() from the request holder:

CPU 0 (holder thread) CPU 1 (resend)
--------------------- --------------

req in processing (FR_SENT=1)

signal arrives
set_bit(FR_INTERRUPTED)
test_bit(FR_SENT) -> true
queue_interrupt():

    spins on fiq->lock ...
                                  fuse_chan_resend():
                                    set_bit(FR_PENDING)
                                    clear_bit(FR_SENT)
                                    spin_lock(&fiq->lock)
                                    cleanup scan:
                                      intr_entry not linked yet
                                      -> list_del_init is a no-op
                                    list_splice -> fiq->pending
                                    spin_unlock(&fiq->lock)
    ... acquires fiq->lock
    list_empty(&req->intr_entry) -> true
    FR_FINISHED not set
    -> intr_entry added to fiq->interrupts
       AFTER the cleanup already ran
Quoted 9 lines

fatal signal arrives
fuse_remove_pending_req():

    test_bit(FR_PENDING) -> true
    list_del(&req->list)
    __fuse_put_request
  fuse_put_request (refcount -> 0)
  -> req freed, intr_entry dangling
     on fiq->interrupts

Good catch. We discovered this issue in our recent stability testing. The proposed fix looks reasonable, and we plan to validate it.

Quoted 16 lines

fuse_dev_queue_interrupt() only checks list_empty() and FR_FINISHED before linking intr_entry -- it does not check FR_PENDING, so a request already spliced back to fiq->pending can still be added to fiq->interrupts. The lock contention itself produces the bad
ordering: while the resend holds fiq->lock to scan the queued requests, the holder spins in queue_interrupt() and links intr_entry right after the scan finishes.

The dangling entry then causes the same use-after-free that the above commit describes: fuse_read_interrupt() writes to the freed slab object via list_del_init() and leaks req->in.h.unique to userspace. Once the freed memory is reused, INIT_LIST_HEAD() turns the entry into a self-loop and list_empty(&fiq->interrupts) returns false forever, so the daemon reads the same phantom FUSE_INTERRUPT in an infinite loop and never consumes fiq->pending.

Quoted 11 lines

Close the race in fuse_remove_pending_req(), which is the common bail-out path for both the legacy and the io_uring transport: after the request is removed from the pending queue, also unlink intr_entry under fiq->lock before the reference is dropped. fiq->lock must be taken explicitly since the lock argument is the ring queue lock in the io_uring case, while fiq->interrupts is always protected by fiq->lock. This runs on the holder thread after any queue_interrupt() it issued, and no other path can re-link the entry once the request is off the queues (re-queueing an interrupt from FUSE_INTERRUPT's -EAGAIN reply requires finding the request in the processing queue first).

The commit message is too verbose and looks AI-generated.

Signature · 2 lines
Best Regards,
Yi
Quoted 40 lines
Fixes 760eac73f9f6 ("fuse: Introduce a new notification type for resend pending requests")
---
 fs/fuse/dev.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
fs/fuse/dev.c +14 -0
index 5763a7cd3b37..87891162985c 100644
@@ -678,6 +678,8 @@ static int queue_interrupt(struct fuse_req *req)
 
 bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock)
 {
+	struct fuse_iqueue *fiq = &req->chan->iq;
+
 	spin_lock(lock);
 	if (test_bit(FR_PENDING, &req->flags)) {
 		/*
@@ -686,6 +688,18 @@ bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock)
 		 */
 		list_del(&req->list);
 		spin_unlock(lock);
+
+		/*
+		 * Remove stale intr_entry queued by queue_interrupt() before
+		 * the request was requeued, which would otherwise dangle on
+		 * fiq->interrupts once the request is freed.
+		 */
+		if (test_bit(FR_INTERRUPTED, &req->flags)) {
+			spin_lock(&fiq->lock);
+			list_del_init(&req->intr_entry);
+			spin_unlock(&fiq->lock);
+		}
+
 		__fuse_put_request(req);
 		req->out.h.error = -EINTR;
 		return true;
#3 lore raw
In reply to Baokun Li
Quoted 98 lines · Baokun Li

Commit f8fce75fedf7 ("fuse: clear intr_entry in fuse_resend and fuse_remove_pending_req") removes stale interrupt entries in fuse_chan_resend() when requests are moved back to fiq->pending. However, that cleanup only covers interrupt entries that are already linked at scan time. It can race with a concurrent queue_interrupt() from the request holder:

CPU 0 (holder thread) CPU 1 (resend)
--------------------- --------------

req in processing (FR_SENT=1)

signal arrives
set_bit(FR_INTERRUPTED)
test_bit(FR_SENT) -> true
queue_interrupt():

    spins on fiq->lock ...
                                  fuse_chan_resend():
                                    set_bit(FR_PENDING)
                                    clear_bit(FR_SENT)
                                    spin_lock(&fiq->lock)
                                    cleanup scan:
                                      intr_entry not linked yet
                                      -> list_del_init is a no-op
                                    list_splice -> fiq->pending
                                    spin_unlock(&fiq->lock)
    ... acquires fiq->lock
    list_empty(&req->intr_entry) -> true
    FR_FINISHED not set
    -> intr_entry added to fiq->interrupts
       AFTER the cleanup already ran

fatal signal arrives
fuse_remove_pending_req():

    test_bit(FR_PENDING) -> true
    list_del(&req->list)
    __fuse_put_request
  fuse_put_request (refcount -> 0)
  -> req freed, intr_entry dangling
     on fiq->interrupts

fuse_dev_queue_interrupt() only checks list_empty() and FR_FINISHED before linking intr_entry -- it does not check FR_PENDING, so a request already spliced back to fiq->pending can still be added to fiq->interrupts. The lock contention itself produces the bad
ordering: while the resend holds fiq->lock to scan the queued requests, the holder spins in queue_interrupt() and links intr_entry right after the scan finishes.

The dangling entry then causes the same use-after-free that the above commit describes: fuse_read_interrupt() writes to the freed slab object via list_del_init() and leaks req->in.h.unique to userspace. Once the freed memory is reused, INIT_LIST_HEAD() turns the entry into a self-loop and list_empty(&fiq->interrupts) returns false forever, so the daemon reads the same phantom FUSE_INTERRUPT in an infinite loop and never consumes fiq->pending.

Close the race in fuse_remove_pending_req(), which is the common bail-out path for both the legacy and the io_uring transport: after the request is removed from the pending queue, also unlink intr_entry under fiq->lock before the reference is dropped. fiq->lock must be taken explicitly since the lock argument is the ring queue lock in the io_uring case, while fiq->interrupts is always protected by fiq->lock. This runs on the holder thread after any queue_interrupt() it issued, and no other path can re-link the entry once the request is off the queues (re-queueing an interrupt from FUSE_INTERRUPT's -EAGAIN reply requires finding the request in the processing queue first).

Fixes 760eac73f9f6 ("fuse: Introduce a new notification type for resend pending requests")
---
 fs/fuse/dev.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
fs/fuse/dev.c +7 -0
index 5763a7cd3b37..87891162985c 100644
@@ -678,6 +678,8 @@ static int queue_interrupt(struct fuse_req *req)
 
 bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock)
 {
+	struct fuse_iqueue *fiq = &req->chan->iq;
+
 	spin_lock(lock);
 	if (test_bit(FR_PENDING, &req->flags)) {
 		/*
@@ -686,6 +688,18 @@ bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock)
 		 */
 		list_del(&req->list);
 		spin_unlock(lock);
+
+		/*
+		 * Remove stale intr_entry queued by queue_interrupt() before
+		 * the request was requeued, which would otherwise dangle on
+		 * fiq->interrupts once the request is freed.

fs/fuse/dev.c bool fuse_remove_pending_req(struct fuse_req *req, spinlock_t *lock) · line 705

Quoted 10 lines
fs/fuse/dev.c +7 -0
+		 */
+		if (test_bit(FR_INTERRUPTED, &req->flags)) {
+			spin_lock(&fiq->lock);
+			list_del_init(&req->intr_entry);
+			spin_unlock(&fiq->lock);
+		}
+
 		__fuse_put_request(req);
 		req->out.h.error = -EINTR;
 		return true;

Hi Baokun,

I just read Jun’s solution, and it seems his fix is more comprehensive. If you don’t mind, I hope you can take some time to read it.

https://lore.kernel.org/all/[email protected]/T/#mf9bb60e63dbd4ef0ab5cf826e97ad67037ccd3ff

To be frank, Jun doesn’t seem to have a deep understanding of the issue. His patchset appears to have been written by AI, and the problem description is a complete mess. However, his approach to solving the problem seems viable.

I’m not sure whether Jun is able to rewrite his patchset that Miklos would accept. If not, I’d be happy to help, since we encountered the same issue in our stability testing.

Signature · 2 lines
Best Regards,
Yi