Skip to contentSkip to conversation

io_uring/cmd: don't skip completion for a non-armed multishot uring_cmd

RFC PATCH In review Vasileios Almpanis 2 participants 3 messages Aug 11 – Aug 20 io_uring

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 io_uring/cmd: don't skip completion for a non-armed multishot uring_cmd

io_uring_cmd() treats any uring_cmd carrying IORING_URING_CMD_MULTISHOT that returns >= 0 as "multishot armed, completion deferred" and returns IOU_ISSUE_SKIP_COMPLETE, expecting the provider to complete the request later. But the flag is user-controlled and validated only against buffer select, not against provider capability. A ->uring_cmd() handler that does not implement multishot and returns a normal >= 0 result then has its request skipped and never completed, leaking the io_kiocb and its io_async_cmd:

  BUG: memory leak
  unreferenced object (size 248):
    kmem_cache_alloc_bulk_noprof+0x272/0x3f0
    __io_alloc_req_refill+0x4a/0x150
    io_submit_sqes.cold+0x16e/0x20b
    __do_sys_io_uring_enter+0x56d/0xd60

syzbot hit this via ublk UBLK_U_CMD_ADD_DEV, but it is kernel-wide: the same leak reproduces with SOCKET_URING_OP_SIOCINQ on any socket fd, which returns the queued byte count and never inspects cmd->flags. No in-tree provider of multishot actually returns >= 0. Both io_cmd_poll_multishot() and ublk_handle_batch_fetch_cmd return -EIOCBQUEUED,

Only skip completion when the command is really multishot, i.e. REQ_F_APOLL_MULTISHOT is set. Otherwise fall through and complete the request normally with its result.

Fixes 620a50c92700 ("io_uring: uring_cmd: add multishot support")
Reported-by [email protected]

---
Questions / Notes:
- I put this in the core io_uring_cmd() rather than in provider specific code because the leak is independent of the provider. From my research so far no ->uring_cmd() rejects IORING_URING_CMD_MULTISHOT, they just ignore it. The flag is only validated against buffer-select in io_uring_cmd_prep(). So any handler that returns a plain >= 0 result with the user-set flag leaks.
- Instead of checking if it has been really armed should we just drop the >= 0 check? The in-tree code that supports multishot returns -EIOCBQUEUED from what I have seen so far.

---
 io_uring/uring_cmd.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
io_uring/uring_cmd.c +2 -1
index c14c22cff49e..a2899a852879 100644
@@ -269,7 +269,8 @@ int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags)
 	}
 
 	ret = file->f_op->uring_cmd(ioucmd, issue_flags);
-	if (ioucmd->flags & IORING_URING_CMD_MULTISHOT) {
+	if ((ioucmd->flags & IORING_URING_CMD_MULTISHOT) &&
+	    (req->flags & REQ_F_APOLL_MULTISHOT)) {
 		if (ret >= 0)
 			return IOU_ISSUE_SKIP_COMPLETE;
 	}

---
base-commit: d58772d8520c7ef247c4b95c9bd76d3a25da9ff5
change-id: 20260811-io_uring-169337c74617

Best regards,

Signature · 1 line
Vasileios Almpanis <[email protected]>
#1 lore raw

Hi everyone,

Just a gentle ping on this patch.

The issue is syzbot-reproduced and the proposed fix has been run with #syz test. The patch addresses the leak in the core |io_uring_cmd()|path rather than relying on individual providers to handle |IORING_URING_CMD_MULTISHOT|.

Could someone take a look when you get a chance?

Thanks,
Vasileios

Quoted 69 lines · Vasileios Almpanis

io_uring_cmd() treats any uring_cmd carrying IORING_URING_CMD_MULTISHOT that returns >= 0 as "multishot armed, completion deferred" and returns IOU_ISSUE_SKIP_COMPLETE, expecting the provider to complete the request later. But the flag is user-controlled and validated only against buffer select, not against provider capability. A ->uring_cmd() handler that does not implement multishot and returns a normal >= 0 result then has its request skipped and never completed, leaking the io_kiocb and its io_async_cmd:

   BUG: memory leak
   unreferenced object (size 248):
     kmem_cache_alloc_bulk_noprof+0x272/0x3f0
     __io_alloc_req_refill+0x4a/0x150
     io_submit_sqes.cold+0x16e/0x20b
     __do_sys_io_uring_enter+0x56d/0xd60

syzbot hit this via ublk UBLK_U_CMD_ADD_DEV, but it is kernel-wide: the same leak reproduces with SOCKET_URING_OP_SIOCINQ on any socket fd, which returns the queued byte count and never inspects cmd->flags. No in-tree provider of multishot actually returns >= 0. Both io_cmd_poll_multishot() and ublk_handle_batch_fetch_cmd return -EIOCBQUEUED,

Only skip completion when the command is really multishot, i.e. REQ_F_APOLL_MULTISHOT is set. Otherwise fall through and complete the request normally with its result.

Fixes 620a50c92700 ("io_uring: uring_cmd: add multishot support")
Reported-by [email protected]

---
Questions / Notes:
- I put this in the core io_uring_cmd() rather than in provider specific code because the leak is independent of the provider. From my research so far no ->uring_cmd() rejects IORING_URING_CMD_MULTISHOT, they just ignore it. The flag is only validated against buffer-select in io_uring_cmd_prep(). So any handler that returns a plain >= 0 result with the user-set flag leaks.
- Instead of checking if it has been really armed should we just drop the >= 0 check? The in-tree code that supports multishot returns -EIOCBQUEUED from what I have seen so far.

---
  io_uring/uring_cmd.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
io_uring/uring_cmd.c +2 -2
index c14c22cff49e..a2899a852879 100644
@@ -269,7 +269,8 @@ int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags)
  	}
 
  	ret = file->f_op->uring_cmd(ioucmd, issue_flags);
-	if (ioucmd->flags & IORING_URING_CMD_MULTISHOT) {
+	if ((ioucmd->flags & IORING_URING_CMD_MULTISHOT) &&
+	    (req->flags & REQ_F_APOLL_MULTISHOT)) {
  		if (ret >= 0)
  			return IOU_ISSUE_SKIP_COMPLETE;
  	}
 
---

base-commit: d58772d8520c7ef247c4b95c9bd76d3a25da9ff5
change-id: 20260811-io_uring-169337c74617

Best regards,
--
Vasileios Almpanis <[email protected]>

syzbot
#2 lore raw
unknown command "test."
Quoted 6 lines

Hi everyone,

Just a gentle ping on this patch.

The issue is syzbot-reproduced and the proposed fix has been run with #syz test.

unknown command "test."

Quoted 79 lines

The patch addresses the leak in the core |io_uring_cmd()|path rather than relying on individual providers to handle |IORING_URING_CMD_MULTISHOT|.

Could someone take a look when you get a chance?

Thanks,
Vasileios

Quoted 69 lines · Vasileios Almpanis

io_uring_cmd() treats any uring_cmd carrying IORING_URING_CMD_MULTISHOT that returns >= 0 as "multishot armed, completion deferred" and returns IOU_ISSUE_SKIP_COMPLETE, expecting the provider to complete the request later. But the flag is user-controlled and validated only against buffer select, not against provider capability. A ->uring_cmd() handler that does not implement multishot and returns a normal >= 0 result then has its request skipped and never completed, leaking the io_kiocb and its io_async_cmd:

   BUG: memory leak
   unreferenced object (size 248):
     kmem_cache_alloc_bulk_noprof+0x272/0x3f0
     __io_alloc_req_refill+0x4a/0x150
     io_submit_sqes.cold+0x16e/0x20b
     __do_sys_io_uring_enter+0x56d/0xd60

syzbot hit this via ublk UBLK_U_CMD_ADD_DEV, but it is kernel-wide: the same leak reproduces with SOCKET_URING_OP_SIOCINQ on any socket fd, which returns the queued byte count and never inspects cmd->flags. No in-tree provider of multishot actually returns >= 0. Both io_cmd_poll_multishot() and ublk_handle_batch_fetch_cmd return -EIOCBQUEUED,

Only skip completion when the command is really multishot, i.e. REQ_F_APOLL_MULTISHOT is set. Otherwise fall through and complete the request normally with its result.

Fixes 620a50c92700 ("io_uring: uring_cmd: add multishot support")
Reported-by [email protected]

---
Questions / Notes:
- I put this in the core io_uring_cmd() rather than in provider specific code because the leak is independent of the provider. From my research so far no ->uring_cmd() rejects IORING_URING_CMD_MULTISHOT, they just ignore it. The flag is only validated against buffer-select in io_uring_cmd_prep(). So any handler that returns a plain >= 0 result with the user-set flag leaks.
- Instead of checking if it has been really armed should we just drop the >= 0 check? The in-tree code that supports multishot returns -EIOCBQUEUED from what I have seen so far.

---
  io_uring/uring_cmd.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
io_uring/uring_cmd.c +2 -2
index c14c22cff49e..a2899a852879 100644
@@ -269,7 +269,8 @@ int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags)
  	}
 
  	ret = file->f_op->uring_cmd(ioucmd, issue_flags);
-	if (ioucmd->flags & IORING_URING_CMD_MULTISHOT) {
+	if ((ioucmd->flags & IORING_URING_CMD_MULTISHOT) &&
+	    (req->flags & REQ_F_APOLL_MULTISHOT)) {
  		if (ret >= 0)
  			return IOU_ISSUE_SKIP_COMPLETE;
  	}
 
---

base-commit: d58772d8520c7ef247c4b95c9bd76d3a25da9ff5
change-id: 20260811-io_uring-169337c74617

Best regards,
--
Vasileios Almpanis <[email protected]>