* [PATCH] nvmet: cache passthru reference in __nvmet_req_complete()
@ 2026-11-05 12:24 Nilay Shroff
2026-11-05 15:49 ` Keith Busch
0 siblings, 1 reply; 2+ messages in thread
From: Nilay Shroff @ 2026-11-05 12:24 UTC (permalink / raw)
To: kbusch, hch, sagi, chaitanyak, gjoyce
Cc: linux-nvme, shinichiro.kawasaki, Nilay Shroff
__nvmet_req_complete() releases the passthru reference held while
processing a passthru request. However, it currently does so after
calling ops->queue_response().
Once ops->queue_response() is called, the transport may complete and
recycle the request on another CPU. If the request is reused, its
req->p.ref_held field may be reset to false before
__nvmet_req_complete() checks it in nvmet_put_passthru_ref().
Consequently, the passthru reference may not be released, leaving
wait_for_completion(&passthru->disable_done) in
__nvmet_passthru_ctrl_disable() blocked indefinitely.
Fix this by caching the passthru pointer at the beginning of
__nvmet_req_complete(), while the request is still valid, and use
the cached pointer to release the reference after queuing the
response. This also changes nvmet_put_passthru_ref() to accept
the passthru pointer directly instead of accessing it through
the request.
Reported-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Closes: https://lore.kernel.org/all/auhpmq499RVE9YpC@shinmob/
Fixes: 5e12d81874a1 ("nvmet: fix use-after-free in passthru I/O hotpath")
Tested-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Signed-off-by: Nilay Shroff <nilay@linux.ibm.com>
---
drivers/nvme/target/core.c | 4 +++-
drivers/nvme/target/nvmet.h | 24 +++++++++++++++++-------
2 files changed, 20 insertions(+), 8 deletions(-)
diff --git a/drivers/nvme/target/core.c b/drivers/nvme/target/core.c
index eadab017c70b..b8bcb5bf6be6 100644
--- a/drivers/nvme/target/core.c
+++ b/drivers/nvme/target/core.c
@@ -806,6 +806,7 @@ static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
{
struct nvmet_ns *ns = req->ns;
struct nvmet_pr_per_ctrl_ref *pc_ref = req->pc_ref;
+ struct nvmet_passthru *passthru = nvmet_passthru_ref_held(req);
if (!req->sq->sqhd_disabled)
nvmet_update_sq_head(req);
@@ -824,7 +825,8 @@ static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
if (ns)
nvmet_put_namespace(ns);
- nvmet_put_passthru_ref(req);
+ if (passthru)
+ nvmet_put_passthru_ref(passthru);
}
void nvmet_req_complete(struct nvmet_req *req, u16 status)
diff --git a/drivers/nvme/target/nvmet.h b/drivers/nvme/target/nvmet.h
index 3cd58e813aa7..2c2460ef5e2b 100644
--- a/drivers/nvme/target/nvmet.h
+++ b/drivers/nvme/target/nvmet.h
@@ -832,6 +832,15 @@ static inline struct nvmet_passthru *nvmet_subsys_passthru(
return subsys->passthru;
}
+static inline struct nvmet_passthru *nvmet_passthru_ref_held(
+ struct nvmet_req *req)
+{
+ if (req->p.ref_held)
+ return nvmet_req_subsys(req)->passthru;
+
+ return NULL;
+}
+
static inline bool nvmet_get_passthru_ref(struct nvmet_req *req)
{
if (!percpu_ref_tryget_live(&nvmet_req_subsys(req)->passthru->ref))
@@ -841,13 +850,9 @@ static inline bool nvmet_get_passthru_ref(struct nvmet_req *req)
return true;
}
-static inline void nvmet_put_passthru_ref(struct nvmet_req *req)
+static inline void nvmet_put_passthru_ref(struct nvmet_passthru *passthru)
{
- if (!req->p.ref_held)
- return;
-
- req->p.ref_held = false;
- percpu_ref_put(&nvmet_req_subsys(req)->passthru->ref);
+ percpu_ref_put(&passthru->ref);
}
#else /* CONFIG_NVME_TARGET_PASSTHRU */
static inline void nvmet_passthru_subsys_free(struct nvmet_subsys *subsys)
@@ -873,11 +878,16 @@ static inline struct nvmet_passthru *nvmet_subsys_passthru(
{
return NULL;
}
+static inline struct nvmet_passthru *nvmet_passthru_ref_held(
+ struct nvmet_req *req)
+{
+ return NULL;
+}
static inline bool nvmet_get_passthru_ref(struct nvmet_req *req)
{
return false;
}
-static inline void nvmet_put_passthru_ref(struct nvmet_req *req)
+static inline void nvmet_put_passthru_ref(struct nvmet_passthru *passthru)
{
}
#endif /* CONFIG_NVME_TARGET_PASSTHRU */
base-commit: c64d9b2c1f4fd1d8fa52dab37b48428a99c4d58d
--
2.53.0
^ permalink raw reply related [flat|nested] 2+ messages in thread* Re: [PATCH] nvmet: cache passthru reference in __nvmet_req_complete()
2026-11-05 12:24 [PATCH] nvmet: cache passthru reference in __nvmet_req_complete() Nilay Shroff
@ 2026-11-05 15:49 ` Keith Busch
0 siblings, 0 replies; 2+ messages in thread
From: Keith Busch @ 2026-11-05 15:49 UTC (permalink / raw)
To: hch, sagi, chaitanyak, gjoyce, Nilay Shroff
Cc: linux-nvme, shinichiro.kawasaki
On Thu, 05 Nov 2026 17:54:55 +0530, Nilay Shroff wrote:
> __nvmet_req_complete() releases the passthru reference held while
> processing a passthru request. However, it currently does so after
> calling ops->queue_response().
>
> Once ops->queue_response() is called, the transport may complete and
> recycle the request on another CPU. If the request is reused, its
> req->p.ref_held field may be reset to false before
> __nvmet_req_complete() checks it in nvmet_put_passthru_ref().
> Consequently, the passthru reference may not be released, leaving
> wait_for_completion(&passthru->disable_done) in
> __nvmet_passthru_ctrl_disable() blocked indefinitely.
>
> [...]
Applied, thanks!
[1/1] nvmet: cache passthru reference in __nvmet_req_complete()
commit: 93a76fb240be5b7fc1c0d3df1706f7e6396e2577
Best regards,
--
Keith Busch <kbusch@kernel.org>
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-11-05 15:49 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-05 12:24 [PATCH] nvmet: cache passthru reference in __nvmet_req_complete() Nilay Shroff
2026-11-05 15:49 ` Keith Busch
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox