uring_cmd SQE corruptions

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

Quoted 36 lines · Caleb Sander Mateos

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

Quoted 5 lines

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any risk of overwriting SQEs. But I do think the code should copy for that case too, just to avoid it being a harder-to-use thing than it should be.

The two patches here look good, I'll go ahead with those. That'll give us a bit of time to figure out where this missing copy is.

Quoted 53 lines · Jens Axboe

Quoted 36 lines · Caleb Sander Mateos

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

Quoted 5 lines

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any risk of overwriting SQEs. But I do think the code should copy for that case too, just to avoid it being a harder-to-use thing than it should be.

Quoted 2 lines

The two patches here look good, I'll go ahead with those. That'll give us a bit of time to figure out where this missing copy is.

Can you try this on top of your 2 and see if you still hit anything odd?

io_uring/uring_cmd.c +0 -0

index bcfca18395c4..15a8a67f556e 100644

io_uring/uring_cmd.c +3 -2

+
 static int io_uring_cmd_prep_setup(struct io_kiocb *req,
                                   const struct io_uring_sqe *sqe)
 {
        struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
+       struct io_ring_ctx *ctx = req->ctx;
        struct io_uring_cmd_data *cache;
 
        cache = io_uring_alloc_async_data(&req->ctx->uring_cache, req);

@@ -190,7 +193,7 @@ static int io_uring_cmd_prep_setup(struct io_kiocb *req,
 
        ioucmd->sqe = sqe;
        /* defer memcpy until we need it */
-       if (unlikely(req->flags & REQ_F_FORCE_ASYNC))
+       if (unlikely(ctx->flags & IORING_SETUP_SQPOLL || req->flags & SQE_COPY_FLAGS))
                io_uring_cmd_cache_sqes(req);
        return 0;
 }
 
--

Jens Axboe

Quoted 68 lines · On Wed, Feb 12, 2025 at 1:02?PM Jens Axboe <[email protected]> wrote:

Quoted 53 lines · Jens Axboe

Quoted 36 lines · Caleb Sander Mateos

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

Quoted 5 lines

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any risk of overwriting SQEs. But I do think the code should copy for that case too, just to avoid it being a harder-to-use thing than it should be.

Quoted 2 lines

The two patches here look good, I'll go ahead with those. That'll give us a bit of time to figure out where this missing copy is.

Can you try this on top of your 2 and see if you still hit anything odd?

io_uring/uring_cmd.c +0 -0

index bcfca18395c4..15a8a67f556e 100644
@@ -177,10 +177,13 @@ static void io_uring_cmd_cache_sqes(struct io_kiocb *req)
        ioucmd->sqe = cache->sqes;
 }
 

io_uring/uring_cmd.c static void io_uring_cmd_cache_sqes(struct io_kiocb *req) · line 180

Quoted 1 line

io_uring/uring_cmd.c +1 -0

+#define SQE_COPY_FLAGS (REQ_F_FORCE_ASYNC|REQ_F_LINK|REQ_F_HARDLINK|REQ_F_IO_DRAIN)

I believe this still misses the last request in a linked chain, which won't have REQ_F_LINK/REQ_F_HARDLINK set?

IOSQE_IO_DRAIN also causes subsequent operations to be issued async; is REQ_F_IO_DRAIN set on those operations too?

Quoted 69 lines · Caleb Sander

Quoted 68 lines · On Wed, Feb 12, 2025 at 1:02?PM Jens Axboe <[email protected]> wrote:

Quoted 53 lines · Jens Axboe

Quoted 36 lines · Caleb Sander Mateos

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

Quoted 5 lines

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any risk of overwriting SQEs. But I do think the code should copy for that case too, just to avoid it being a harder-to-use thing than it should be.

Quoted 2 lines

The two patches here look good, I'll go ahead with those. That'll give us a bit of time to figure out where this missing copy is.

Can you try this on top of your 2 and see if you still hit anything odd?

io_uring/uring_cmd.c +0 -0

index bcfca18395c4..15a8a67f556e 100644
@@ -177,10 +177,13 @@ static void io_uring_cmd_cache_sqes(struct io_kiocb *req)
        ioucmd->sqe = cache->sqes;
 }
 

Quoted 4 lines

io_uring/uring_cmd.c static void io_uring_cmd_cache_sqes(struct io_kiocb *req) · line 180

Quoted 1 line

io_uring/uring_cmd.c +1 -0

+#define SQE_COPY_FLAGS (REQ_F_FORCE_ASYNC|REQ_F_LINK|REQ_F_HARDLINK|REQ_F_IO_DRAIN)

I believe this still misses the last request in a linked chain, which won't have REQ_F_LINK/REQ_F_HARDLINK set?

Yeah good point, I think we should just be looking at link->head instead to see if the request is a link, or part of a linked submission. That may overshoot a bit, but that should be fine - it'll be a false positive. Alternatively, we can still check link flags and compare with link->last instead...

But the whole thing still feels a bit iffy. The whole uring_cmd setup with an SQE that's sometimes the actual SQE, and sometimes a copy when needed, does not fill me with joy.

Quoted 2 lines

IOSQE_IO_DRAIN also causes subsequent operations to be issued async; is REQ_F_IO_DRAIN set on those operations too?

The first 8 flags are directly set in the io_kiocb at init time. So if IOSQE_IO_DRAIN is set, then REQ_F_IO_DRAIN will be set as they are one and the same.

io_uring/uring_cmd.c +6 -2

index bcfca18395c4..9e60b5bb5a60 100644
@@ -177,10 +177,14 @@ static void io_uring_cmd_cache_sqes(struct io_kiocb *req)
        ioucmd->sqe = cache->sqes;
 }
 
+#define SQE_COPY_FLAGS (REQ_F_FORCE_ASYNC|REQ_F_IO_DRAIN)
+
 static int io_uring_cmd_prep_setup(struct io_kiocb *req,
                                   const struct io_uring_sqe *sqe)
 {
        struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
+       struct io_ring_ctx *ctx = req->ctx;
+       struct io_submit_link *link = &ctx->submit_state.link;
        struct io_uring_cmd_data *cache;
 
        cache = io_uring_alloc_async_data(&req->ctx->uring_cache, req);

@@ -190,7 +194,8 @@ static int io_uring_cmd_prep_setup(struct io_kiocb *req,
 
        ioucmd->sqe = sqe;
        /* defer memcpy until we need it */
-       if (unlikely(req->flags & REQ_F_FORCE_ASYNC))
+       if (unlikely(ctx->flags & IORING_SETUP_SQPOLL ||
+                    req->flags & SQE_COPY_FLAGS || link->head))
                io_uring_cmd_cache_sqes(req);
        return 0;
 }
 
--

Jens Axboe

Quoted 76 lines · On Wed, Feb 12, 2025 at 2:34?PM Jens Axboe <[email protected]> wrote:

Quoted 69 lines · Caleb Sander

Quoted 68 lines · On Wed, Feb 12, 2025 at 1:02?PM Jens Axboe <[email protected]> wrote:

Quoted 53 lines · Jens Axboe

Quoted 36 lines · Caleb Sander Mateos

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

Quoted 5 lines

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any risk of overwriting SQEs. But I do think the code should copy for that case too, just to avoid it being a harder-to-use thing than it should be.

Quoted 2 lines

The two patches here look good, I'll go ahead with those. That'll give us a bit of time to figure out where this missing copy is.

Can you try this on top of your 2 and see if you still hit anything odd?

io_uring/uring_cmd.c +0 -0

index bcfca18395c4..15a8a67f556e 100644
@@ -177,10 +177,13 @@ static void io_uring_cmd_cache_sqes(struct io_kiocb *req)
        ioucmd->sqe = cache->sqes;
 }
 

Quoted 4 lines

io_uring/uring_cmd.c static void io_uring_cmd_cache_sqes(struct io_kiocb *req) · line 180

Quoted 1 line

io_uring/uring_cmd.c +1 -0

+#define SQE_COPY_FLAGS (REQ_F_FORCE_ASYNC|REQ_F_LINK|REQ_F_HARDLINK|REQ_F_IO_DRAIN)

I believe this still misses the last request in a linked chain, which won't have REQ_F_LINK/REQ_F_HARDLINK set?

Quoted 5 lines

Yeah good point, I think we should just be looking at link->head instead to see if the request is a link, or part of a linked submission. That may overshoot a bit, but that should be fine - it'll be a false positive. Alternatively, we can still check link flags and compare with link->last instead...

Yeah, checking link.head sounds good to me. I don't think it should catch any extra requests. link.head will still be NULL when ->prep() is called on the first request of the chain, since it is set in io_submit_sqe() after io_init_req() (which calls ->prep()).

Quoted 10 lines

But the whole thing still feels a bit iffy. The whole uring_cmd setup with an SQE that's sometimes the actual SQE, and sometimes a copy when needed, does not fill me with joy.

Quoted 2 lines

IOSQE_IO_DRAIN also causes subsequent operations to be issued async; is REQ_F_IO_DRAIN set on those operations too?

The first 8 flags are directly set in the io_kiocb at init time. So if IOSQE_IO_DRAIN is set, then REQ_F_IO_DRAIN will be set as they are one and the same.

Sorry, I meant that a request marked IOSQE_IO_DRAIN/REQ_F_IO_DRAIN potentially causes both that request and any following requests to be submitted async. The first request waits for any outstanding requests to complete, and the following requests wait for the request marked IOSQE_IO_DRAIN to complete. I know REQ_F_IO_DRAIN = IOSQE_IO_DRAIN is already set on the first request, but will it also be set on any following requests that have to wait on that one?

Quoted 36 lines · Caleb Sander Mateos

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

Quoted 5 lines

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any risk of overwriting SQEs. But I do think the code should copy for that case too, just to avoid it being a harder-to-use thing than it should be.

The two patches here look good, I'll go ahead with those. That'll give us a bit of time to figure out where this missing copy is.

--
Jens Axboe

Quoted 44 lines · On Wed, Feb 12, 2025 at 12:55?PM Jens Axboe <[email protected]> wrote:

Quoted 36 lines · Caleb Sander Mateos

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

Quoted 5 lines

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

Quoted 9 lines

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any risk of overwriting SQEs. But I do think the code should copy for that case too, just to avoid it being a harder-to-use thing than it should be.

Yes, even with the EAGAIN case fixed, nvme_validate_passthru_nsid() is still catching the poisoned nsids. However, the log lines now come from the application thread rather than the iou-wrk thread. Indeed, we are not using SQPOLL. But we are using IOSQE_SQE_GROUP from Ming's SQE group patch set[1]. Like IOSQE_IO_LINK, subsequent operations in a group are issued only once the first completes. The first operation in the group is a UBLK_IO_PROVIDE_IO_BUF from Ming's other patch set[2], which should complete synchronously. The completion should be processed in __io_submit_flush_completions() -> io_free_batch_list() and queue the remaining grouped operations to be issued in task work. And all the pending task work should be processed

during io_uring_enter()'s return to userspace.
But some NVMe passthru operations must be going async because they are observing the poisoned values the application thread writes into the io_uring SQEs after io_uring_enter() returns. We don't yet have an explanation for how they end up being issued async; ftrace shows that in the typical case, all the NVMe passthru operations in the group are issued during task work before io_uring_enter() returns to userspace. Perhaps a pending signal could short-circuit the task work processing?

Quoted 66 lines · Caleb Sander

Quoted 44 lines · On Wed, Feb 12, 2025 at 12:55?PM Jens Axboe <[email protected]> wrote:

Quoted 36 lines · Caleb Sander Mateos

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

We confirmed this hypothesis by "poisoning" all SQEs up to the SQ head index in userspace upon return from io_uring_enter(). By overwriting the nvme_uring_cmd nsid field with a known garbage value, we were able to trigger the err message in nvme_validate_passthru_nsid(), which logged the garbage nsid value.

The issue is caused by commit 5eff57fa9f3a ("io_uring/uring_cmd: defer SQE copying until it's needed"). With this commit reverted, the poisoned values in the SQEs are no longer seen by nvme_uring_cmd_io().

Prior to the commit, each uring_cmd SQE was unconditionally memcpy()ed to async_data at prep time. The commit moved this memcpy() to 2 cases when the request goes async:
- If REQ_F_FORCE_ASYNC is set to force the initial issue to go async
- If ->uring_cmd() returns -EAGAIN in the initial non-blocking issue

This patch set fixes a bug in the EAGAIN case where the uring_cmd's sqe pointer is not updated to point to async_data after the memcpy(), as it correctly is in the REQ_F_FORCE_ASYNC case.

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

Quoted 5 lines

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

Quoted 9 lines

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any risk of overwriting SQEs. But I do think the code should copy for that case too, just to avoid it being a harder-to-use thing than it should be.

Yes, even with the EAGAIN case fixed, nvme_validate_passthru_nsid() is still catching the poisoned nsids. However, the log lines now come from the application thread rather than the iou-wrk thread. Indeed, we are not using SQPOLL. But we are using IOSQE_SQE_GROUP from Ming's SQE group patch set[1]. Like IOSQE_IO_LINK, subsequent operations in a group are issued only once the first completes. The first operation in the group is a UBLK_IO_PROVIDE_IO_BUF from Ming's other patch set[2], which should complete synchronously. The completion should be processed in __io_submit_flush_completions() -> io_free_batch_list() and queue the remaining grouped operations to be issued in task work. And all the pending task work should be processed

Quoted 8 lines

during io_uring_enter()'s return to userspace.
But some NVMe passthru operations must be going async because they are observing the poisoned values the application thread writes into the io_uring SQEs after io_uring_enter() returns. We don't yet have an explanation for how they end up being issued async; ftrace shows that in the typical case, all the NVMe passthru operations in the group are issued during task work before io_uring_enter() returns to userspace. Perhaps a pending signal could short-circuit the task work processing?

I think it'd be a good idea to move testing to the newer patchset, as the sqe grouping was never really fully vetted or included. It's not impossible that it's a side effect of that, even though it very well cold be a generic issue that exists already without those patches. In any case, that'd move us all to a closer-to-upstream test base, without having older patches that aren't upstream in the mix.

How are the rings setup for what you're testing? I'm specifically thinking of the task_work related flags. Is it using DEFER_TASKRUN, or is it using the older signal style task_work?

--
Jens Axboe

Yes, we completely agree. We are working on incorporating Keith's patchset now. It looks like there is still an open question about whether userspace will need to enforce ordering between the requests (either using linked operations or waiting for completions before submitting the subsequent operations).

Quoted 6 lines · Caleb Sander Mateos

Yes, we completely agree. We are working on incorporating Keith's patchset now. It looks like there is still an open question about whether userspace will need to enforce ordering between the requests (either using linked operations or waiting for completions before submitting the subsequent operations).

In its current form, my series depends on you *not* using linked requests. I didn't think it would be a problem as it follows an existing pattern from the IORING_OP_FILES_UPDATE operation. That has to complete in its entirety before prepping any subsequent commands that reference the index, and using links would get the wrong results.

Quoted 9 lines · On Wed, Feb 12, 2025 at 3:21?PM Keith Busch <[email protected]> wrote:

Quoted 6 lines · Caleb Sander Mateos

Yes, we completely agree. We are working on incorporating Keith's patchset now. It looks like there is still an open question about whether userspace will need to enforce ordering between the requests (either using linked operations or waiting for completions before submitting the subsequent operations).

Quoted 5 lines

In its current form, my series depends on you *not* using linked requests. I didn't think it would be a problem as it follows an existing pattern from the IORING_OP_FILES_UPDATE operation. That has to complete in its entirety before prepping any subsequent commands that reference the index, and using links would get the wrong results.

As implementers of a ublk server, we would also prefer the current interface in your patch series! Having to explicitly order the requests would definitely make the interface more cumbersome and probably less performant. I was just saying that Ming and Pavel had raised some concerns about guaranteeing the order in which io_uring issues SQEs. IORING_OP_FILES_UPDATE is a good analogy. Do we have any examples of how applications use it? Are they waiting for a completion, linking it, or relying on io_uring to issue it synchronously?

Quoted 16 lines · Caleb Sander Mateos

Quoted 9 lines · On Wed, Feb 12, 2025 at 3:21?PM Keith Busch <[email protected]> wrote:

Quoted 6 lines · Caleb Sander Mateos

Yes, we completely agree. We are working on incorporating Keith's patchset now. It looks like there is still an open question about whether userspace will need to enforce ordering between the requests (either using linked operations or waiting for completions before submitting the subsequent operations).

Quoted 5 lines

In its current form, my series depends on you *not* using linked requests. I didn't think it would be a problem as it follows an existing pattern from the IORING_OP_FILES_UPDATE operation. That has to complete in its entirety before prepping any subsequent commands that reference the index, and using links would get the wrong results.

Quoted 9 lines

As implementers of a ublk server, we would also prefer the current interface in your patch series! Having to explicitly order the requests would definitely make the interface more cumbersome and probably less performant. I was just saying that Ming and Pavel had raised some concerns about guaranteeing the order in which io_uring issues SQEs. IORING_OP_FILES_UPDATE is a good analogy. Do we have any examples of how applications use it? Are they waiting for a completion, linking it, or relying on io_uring to issue it synchronously?

Yes it's a good similar example - and I don't think it matters much how it's used. If you rely on its completion before making progress AND you set flags like IOSQE_ASYNC or LINK/DRAIN that will make it go async on purposes, then yes you'd need to similarly link dependents on it. If you don't set anything that forces it to go async, then it WILL complete inline - there's nothing in its implementation that would cause it needing to retry. Any failure would be fatal.

This is very much the same thing with the buf update / insertion, it'll behave in exactly the same way. You could argue "but you need to handle failures" and that is true. But if the failure case is that your consumer of the buffer fails with an import failure, then you can just as well handle that as you can the request getting failed with -ECANCELED because your dependent link failed.

Quoted 9 lines · Keith Busch

Quoted 6 lines · Caleb Sander Mateos

Yes, we completely agree. We are working on incorporating Keith's patchset now. It looks like there is still an open question about whether userspace will need to enforce ordering between the requests (either using linked operations or waiting for completions before submitting the subsequent operations).

Quoted 5 lines

In its current form, my series depends on you *not* using linked requests. I didn't think it would be a problem as it follows an existing pattern from the IORING_OP_FILES_UPDATE operation. That has to complete in its entirety before prepping any subsequent commands that reference the index, and using links would get the wrong results.

As implementers of a ublk server, we would also prefer the current interface in your patch series! Having to explicitly order the requests would definitely make the interface more cumbersome and probably less performant. I was just saying that Ming and Pavel had raised some concerns about guaranteeing the order in which io_uring issues SQEs. IORING_OP_FILES_UPDATE is a good analogy. Do we have any examples of how applications use it? Are they waiting for a completion, linking it, or relying on io_uring to issue it synchronously?

On 2/12/25 1:45 PM, Caleb Sander Mateos wrote:

Quoted 11 lines

However, uring_cmd's can be issued async in other cases not enumerated by 5eff57fa9f3a, also leading to SQE corruption. These include requests besides the first in a linked chain, which are only issued once prior requests complete. Requests waiting for a drain to complete would also be initially issued async.

While it's probably possible for io_uring_cmd_prep_setup() to check for each of these cases and avoid deferring the SQE memcpy(), we feel it might be safer to revert 5eff57fa9f3a to avoid the corruption risk. As discussed recently in regard to the ublk zero-copy patches[1], new async paths added in the future could break these delicate assumptions.

I don't think it's particularly delicate - did you manage to catch the case queueing a request for async execution where the sqe wasn't already copied? I did take a quick look after our out-of-band conversation, and the only missing bit I immediately spotted is using SQPOLL. But I don't think you're using that, right? And in any case, lifetime of SQEs with SQPOLL is the duration of the request anyway, so should not pose any

In our application issuing NVMe passthru commands, we have observed nvme_uring_cmd fields being corrupted between when userspace initializes the io_uring SQE and when nvme_uring_cmd_io() processes it.

We hypothesized that the uring_cmd's were executing asynchronously after the io_uring_enter() syscall returned, yet were still reading the SQE in the userspace-mapped SQ. Since io_uring_enter() had already incremented the SQ head index, userspace reused the SQ slot for a new SQE once the SQ wrapped around to it.

[...]