linux-stable.git/drivers/infiniband/core, branch v7.2.4 Linux kernel stable tree RDMA/uverbs: Guard legacy bundles without method_elm 2026-09-07T15:37:06+00:00 Yuhang Pan 242270054@hdu.edu.cn 2026-08-14T09:27:23+00:00 edae84e5461749ed9adf9fcf5c6e075381040ac3 commit 60a42d510113f46de47e86a84bf5758597644487 upstream. The legacy write() path dispatches through a uverbs_api_write_method, but the uverbs_attr_bundle passed to provider code does not have an ioctl method element. If malformed provider input causes the common uverbs validation code to emit an error message, uverbs_get_handler_fn() dereferences the uninitialized method_elm pointer. Initialize method_elm explicitly for legacy bundles and make uverbs_get_handler_fn() return NULL when no ioctl method is present. The legacy dispatcher continues to use its local write method, while the ioctl path continues to use the registered ioctl handler. Cc: stable@vger.kernel.org Fixes: 7122ff96068a ("RDMA/core: Do not read wild stack memory in uverbs_get_handler_fn()") Link: https://patch.msgid.link/r/AOYAQgCQK3IXqJLr1TB5Qao9.1.1787036796115.Hmail.242270054@hdu.edu.cn Signed-off-by: Yuhang Pan <242270054@hdu.edu.cn> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> 
commit 60a42d510113f46de47e86a84bf5758597644487 upstream.

The legacy write() path dispatches through a uverbs_api_write_method, but
the uverbs_attr_bundle passed to provider code does not have an ioctl
method element. If malformed provider input causes the common uverbs
validation code to emit an error message, uverbs_get_handler_fn()
dereferences the uninitialized method_elm pointer.

Initialize method_elm explicitly for legacy bundles and make
uverbs_get_handler_fn() return NULL when no ioctl method is present. The
legacy dispatcher continues to use its local write method, while the ioctl
path continues to use the registered ioctl handler.

Cc: stable@vger.kernel.org
Fixes: 7122ff96068a ("RDMA/core: Do not read wild stack memory in uverbs_get_handler_fn()")
Link: https://patch.msgid.link/r/AOYAQgCQK3IXqJLr1TB5Qao9.1.1787036796115.Hmail.242270054@hdu.edu.cn
Signed-off-by: Yuhang Pan <242270054@hdu.edu.cn>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
RDMA/uverbs: Add UVERBS_ATTR_UHW to UVERBS_METHOD_REG_MR 2026-09-07T15:37:06+00:00 Jason Gunthorpe jgg@nvidia.com 2026-07-02T16:47:10+00:00 9e71d0f4944ba5641e01db29b3256c963656d671 commit 011199f46f44a9fd93a9e5ab5d7fd1328d80e9bf upstream. The original commit missed that three drivers (mthca, irdma, siw) have UHW data associated with reg_mr that cannot be passed through the ioctl. They also assume that the udata cannot be NULL, so failing to pass a valid udata can trigger a NULL udata crash in those drivers. This never happens in real systems since in rdma-core ibv_cmd_reg_mr_ex() does not accept a udata and those three drivers don't use it, however a malicious userspace could trigger it. Cc: stable@vger.kernel.org Fixes: 5b2e45049dc0 ("IB/core: Add UVERBS_METHOD_REG_MR on the MR object") Reported-by: Jacob Moroni <jmoroni@google.com> Closes: https://lore.kernel.org/all/CAHYDg1TOGxRGZrS69d4Y--Shj_DZv0nJuM73iHUBwBM70g_t3Q@mail.gmail.com Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> 
commit 011199f46f44a9fd93a9e5ab5d7fd1328d80e9bf upstream.

The original commit missed that three drivers (mthca, irdma, siw) have UHW
data associated with reg_mr that cannot be passed through the ioctl. They
also assume that the udata cannot be NULL, so failing to pass a valid
udata can trigger a NULL udata crash in those drivers.

This never happens in real systems since in rdma-core ibv_cmd_reg_mr_ex()
does not accept a udata and those three drivers don't use it, however a
malicious userspace could trigger it.

Cc: stable@vger.kernel.org
Fixes: 5b2e45049dc0 ("IB/core: Add UVERBS_METHOD_REG_MR on the MR object")
Reported-by: Jacob Moroni <jmoroni@google.com>
Closes: https://lore.kernel.org/all/CAHYDg1TOGxRGZrS69d4Y--Shj_DZv0nJuM73iHUBwBM70g_t3Q@mail.gmail.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
RDMA/ucma: Lock the handler in ucma_write_cm_event() 2026-09-07T15:37:06+00:00 Norbert Szetei norbert@doyensec.com 2026-07-27T08:06:12+00:00 0be1955040a2eceed0ecfc387fdc92305411d273 commit f4cc21c6a8e9d392871477f9fd98d68e5ad80272 upstream. ctx->file may only be changed under the handler lock and the xa_lock, which is what stops uevents being queued for a ctx while ucma_migrate_id() moves it to another file. The CM core takes that lock before invoking ucma_event_handler(), but the write() paths that queue uevents themselves do not. ucma_write_cm_event() re-reads ctx->file for each of its four dereferences, so ucma_migrate_id() can swap it mid-sequence: mutex_lock(&ctx->file->mut); /* file A */ list_add_tail(&uevent->list, &ctx->file->event_list); /* file B */ mutex_unlock(&ctx->file->mut); /* file B */ wake_up_interruptible(&ctx->file->poll_wait); /* file B */ The window is the mutex_lock() itself: the writer sleeps in it while the migration reassigns ctx->file. The list_add_tail() then runs on file B's event_list holding only file A's mutex: list_add corruption. prev->next should be next (ffff888101320f30), but was ffff88814a08c418. (prev=ffff88814a075c18). kernel BUG at lib/list_debug.c:32! Call Trace: ucma_write_cm_event+0x36e/0x5e0 and file A's mut is left held forever, wedging its next writer in D state. The uevent is also stranded on a list ucma_cleanup_ctx_events() will not walk, so it outlives its context. /dev/infiniband/rdma_cm is 0666 and no RDMA device is involved, so an unprivileged user reaches all of this. Take the handler lock, as ucma_cleanup_mc_events() does; ctx->cm_id is pinned by the ucma_get_ctx() reference. Fixes: a3c9d0fcd371 ("RDMA/ucma: Support write an event into a CM") Link: https://patch.msgid.link/r/60544A67-EFD6-4D5D-974C-D983445F1070@doyensec.com Cc: stable@vger.kernel.org Assisted-by: Claude:claude-opus-5 Signed-off-by: Norbert Szetei <norbert@doyensec.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> 
commit f4cc21c6a8e9d392871477f9fd98d68e5ad80272 upstream.

ctx->file may only be changed under the handler lock and the xa_lock, which
is what stops uevents being queued for a ctx while ucma_migrate_id() moves
it to another file.  The CM core takes that lock before invoking
ucma_event_handler(), but the write() paths that queue uevents themselves
do not.

ucma_write_cm_event() re-reads ctx->file for each of its four dereferences,
so ucma_migrate_id() can swap it mid-sequence:

	mutex_lock(&ctx->file->mut);			/* file A */
	list_add_tail(&uevent->list, &ctx->file->event_list);	/* file B */
	mutex_unlock(&ctx->file->mut);			/* file B */
	wake_up_interruptible(&ctx->file->poll_wait);	/* file B */

The window is the mutex_lock() itself: the writer sleeps in it while the
migration reassigns ctx->file.  The list_add_tail() then runs on file B's
event_list holding only file A's mutex:

  list_add corruption. prev->next should be next (ffff888101320f30),
    but was ffff88814a08c418. (prev=ffff88814a075c18).
  kernel BUG at lib/list_debug.c:32!
  Call Trace:
   ucma_write_cm_event+0x36e/0x5e0

and file A's mut is left held forever, wedging its next writer in D state.
The uevent is also stranded on a list ucma_cleanup_ctx_events() will not
walk, so it outlives its context.  /dev/infiniband/rdma_cm is 0666 and no
RDMA device is involved, so an unprivileged user reaches all of this.

Take the handler lock, as ucma_cleanup_mc_events() does; ctx->cm_id is
pinned by the ucma_get_ctx() reference.

Fixes: a3c9d0fcd371 ("RDMA/ucma: Support write an event into a CM")
Link: https://patch.msgid.link/r/60544A67-EFD6-4D5D-974C-D983445F1070@doyensec.com
Cc: stable@vger.kernel.org
Assisted-by: Claude:claude-opus-5
Signed-off-by: Norbert Szetei <norbert@doyensec.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
RDMA/ucma: Lock the handler in ucma_set_ib_path() 2026-09-07T15:37:06+00:00 Norbert Szetei norbert@doyensec.com 2026-07-27T08:08:36+00:00 43a7d8ae9f645dc00235cfa4bf57673c20b769f8 commit ecbe7d36dc2de07e5dfbb4a8ff5b315ab43de820 upstream. ucma_set_ib_path() calls ucma_event_handler() straight from the write() path, without the handler lock that keeps ctx->file stable while a uevent is queued. The handler re-reads ctx->file for every dereference: mutex_lock(&ctx->file->mut); /* file A */ list_add_tail(&uevent->list, &ctx->file->event_list); /* file B */ mutex_unlock(&ctx->file->mut); /* file B */ wake_up_interruptible(&ctx->file->poll_wait); /* file B */ A concurrent ucma_migrate_id() reassigns ctx->file while the SET_OPTION caller sleeps in mutex_lock(), so the list_add_tail() lands on file B's event_list while only file A's mutex is held, racing every other user of that list: BUG: KASAN: slab-use-after-free in __list_add_valid_or_report+0x1aa/0x1c0 Read of size 8 at addr ffff888153c6a418 by task poc_corr/486 Call Trace: __list_add_valid_or_report+0x1aa/0x1c0 ucma_event_handler+0x1be/0xc00 ucma_set_ib_path+0x45e/0x710 ucma_set_option+0x32e/0x590 ucma_write+0x1f9/0x330 Allocated by task 505: ucma_write_cm_event+0x1a1/0x660 Freed by task 505: kfree+0x1da/0x4c0 ucma_get_event+0x5d5/0x7e0 The freed object is a ucma_event that another thread dequeued from file B's list under file B's mutex. File A's mut is left held on top of that, wedging its next writer in uninterruptible sleep. This path needs a bound and address-resolved cm_id, so it requires an RDMA device to be present. Take the handler lock around the call. Fixes: 09e328e47a69 ("RDMA/ucma: Fix the locking of ctx->file") Link: https://patch.msgid.link/r/2823D190-92D5-4714-8769-4FB643C64FF3@doyensec.com Cc: stable@vger.kernel.org Assisted-by: Claude:claude-opus-5 Signed-off-by: Norbert Szetei <norbert@doyensec.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> 
commit ecbe7d36dc2de07e5dfbb4a8ff5b315ab43de820 upstream.

ucma_set_ib_path() calls ucma_event_handler() straight from the write()
path, without the handler lock that keeps ctx->file stable while a uevent
is queued.  The handler re-reads ctx->file for every dereference:

	mutex_lock(&ctx->file->mut);			/* file A */
	list_add_tail(&uevent->list, &ctx->file->event_list);	/* file B */
	mutex_unlock(&ctx->file->mut);			/* file B */
	wake_up_interruptible(&ctx->file->poll_wait);	/* file B */

A concurrent ucma_migrate_id() reassigns ctx->file while the SET_OPTION
caller sleeps in mutex_lock(), so the list_add_tail() lands on file B's
event_list while only file A's mutex is held, racing every other user of
that list:

  BUG: KASAN: slab-use-after-free in __list_add_valid_or_report+0x1aa/0x1c0
  Read of size 8 at addr ffff888153c6a418 by task poc_corr/486
  Call Trace:
   __list_add_valid_or_report+0x1aa/0x1c0
   ucma_event_handler+0x1be/0xc00
   ucma_set_ib_path+0x45e/0x710
   ucma_set_option+0x32e/0x590
   ucma_write+0x1f9/0x330
  Allocated by task 505:
   ucma_write_cm_event+0x1a1/0x660
  Freed by task 505:
   kfree+0x1da/0x4c0
   ucma_get_event+0x5d5/0x7e0

The freed object is a ucma_event that another thread dequeued from file B's
list under file B's mutex.  File A's mut is left held on top of that,
wedging its next writer in uninterruptible sleep.

This path needs a bound and address-resolved cm_id, so it requires an RDMA
device to be present.

Take the handler lock around the call.

Fixes: 09e328e47a69 ("RDMA/ucma: Fix the locking of ctx->file")
Link: https://patch.msgid.link/r/2823D190-92D5-4714-8769-4FB643C64FF3@doyensec.com
Cc: stable@vger.kernel.org
Assisted-by: Claude:claude-opus-5
Signed-off-by: Norbert Szetei <norbert@doyensec.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
RDMA/core: Fix memory leak in __ib_create_cq() on invalid cqe 2026-07-02T17:19:33+00:00 Chenguang Zhao zhaochenguang@kylinos.cn 2026-06-25T02:01:48+00:00 e939334ea7dd219f100f963dbb1cb43df520c20a Move the zero CQE validation before rdma_zalloc_drv_obj() to avoid leaking the CQ object when returning -EINVAL. Fixes: a2917582887a ("RDMA/core: Reject zero CQE count") Link: https://patch.msgid.link/r/20260625020148.224537-1-zhaochenguang@kylinos.cn Signed-off-by: Chenguang Zhao <zhaochenguang@kylinos.cn> Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> 
Move the zero CQE validation before rdma_zalloc_drv_obj() to avoid
leaking the CQ object when returning -EINVAL.

Fixes: a2917582887a ("RDMA/core: Reject zero CQE count")
Link: https://patch.msgid.link/r/20260625020148.224537-1-zhaochenguang@kylinos.cn
Signed-off-by: Chenguang Zhao <zhaochenguang@kylinos.cn>
Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
RDMA/cma: Fix hardware address comparison length in netevent callback 2026-07-02T11:03:46+00:00 Or Gerlitz ogerlitz@ddn.com 2026-06-17T11:21:05+00:00 18313833e2c6de222a4f6c072da759d0d5888528 The cited commit hardcoded the hardware address comparison len to ETH_ALEN. This breaks IPoIB, which uses 20-byte addresses. By truncating the memcmp, the CMA may incorrectly assume the target address is unchanged and fails to abort the stalled connection. Fix this by replacing ETH_ALEN with the dynamic neigh->dev->addr_len to correctly evaluate the full address regardless of the link layer. Fixes: 925d046e7e52 ("RDMA/core: Add a netevent notifier to cma") Signed-off-by: Or Gerlitz <ogerlitz@ddn.com> Link: https://patch.msgid.link/20260617-fix-cma-ipoib-v1-1-03f869344304@ddn.com Signed-off-by: Leon Romanovsky <leon@kernel.org> 
The cited commit hardcoded the hardware address comparison len to ETH_ALEN.

This breaks IPoIB, which uses 20-byte addresses. By truncating the
memcmp, the CMA may incorrectly assume the target address is
unchanged and fails to abort the stalled connection.

Fix this by replacing ETH_ALEN with the dynamic neigh->dev->addr_len
to correctly evaluate the full address regardless of the link layer.

Fixes: 925d046e7e52 ("RDMA/core: Add a netevent notifier to cma")
Signed-off-by: Or Gerlitz <ogerlitz@ddn.com>
Link: https://patch.msgid.link/20260617-fix-cma-ipoib-v1-1-03f869344304@ddn.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
IB/mad: Drop unmatched RMPP responses before reassembly 2026-06-29T12:08:59+00:00 Michael Bommarito michael.bommarito@gmail.com 2026-06-06T20:01:55+00:00 d2e52d610b9b09694261632340b801a421e0b0c5 Kernel-handled RMPP receive processing starts reassembly for active DATA responses before the response is matched to an outstanding send. The normal match happens later, after ib_process_rmpp_recv_wc() has either assembled a complete message or consumed the segment. That ordering lets an unsolicited response that routes to a kernel RMPP agent by the high TID bits allocate or extend RMPP receive state before the full TID and source address are checked against a real request. A reordered burst can therefore reach the receive-side insertion path even though the response would not match any send. For kernel-handled RMPP DATA responses, require the existing ib_find_send_mad() match before entering RMPP reassembly. The matcher already checks the full TID, management class and source address/GID against the agent wait, backlog and in-flight send lists. If there is no match, drop the response without creating RMPP state. This leaves the RMPP window behavior unchanged and only rejects responses that have no corresponding request. Fixes: fa619a77046b ("[PATCH] IB: Add RMPP implementation") Assisted-by: Codex:gpt-5-5-xhigh Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com> Link: https://patch.msgid.link/3170ff3bc389a930bb1641f2caa394a0b2241579.1780774907.git.michael.bommarito@gmail.com Signed-off-by: Leon Romanovsky <leon@kernel.org> 
Kernel-handled RMPP receive processing starts reassembly for active
DATA responses before the response is matched to an outstanding send.
The normal match happens later, after ib_process_rmpp_recv_wc() has
either assembled a complete message or consumed the segment.

That ordering lets an unsolicited response that routes to a kernel
RMPP agent by the high TID bits allocate or extend RMPP receive state
before the full TID and source address are checked against a real
request. A reordered burst can therefore reach the receive-side
insertion path even though the response would not match any send.

For kernel-handled RMPP DATA responses, require the existing
ib_find_send_mad() match before entering RMPP reassembly. The matcher
already checks the full TID, management class and source address/GID
against the agent wait, backlog and in-flight send lists. If there is
no match, drop the response without creating RMPP state.

This leaves the RMPP window behavior unchanged and only rejects
responses that have no corresponding request.

Fixes: fa619a77046b ("[PATCH] IB: Add RMPP implementation")
Assisted-by: Codex:gpt-5-5-xhigh
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
Link: https://patch.msgid.link/3170ff3bc389a930bb1641f2caa394a0b2241579.1780774907.git.michael.bommarito@gmail.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma 2026-06-18T15:16:21+00:00 Linus Torvalds torvalds@linux-foundation.org 2026-06-18T15:16:21+00:00 9e7e6633458362db72427b48effad8d759131c35 Pull rdma updates from Jason Gunthorpe: "Many AI driven bug fixes, and several big driver API cleanups - Driver bug fixes and minor cleanups in mlx5, hns, rxe, efa, siw, rtrs, mana, irdma, mlx4. Commonly error path flows, integer arithmetic overflows on unsafe data, out of bounds access, and use after free issues under races. - Second half of the new udata API for drivers focusing on uAPI response - bnxt_re supports more options for QP creation that will allow a dv path in rdma-core - Untangle the module dependencies so drivers don't link to ib_uverbs.ko as was originall intended - Provide a new way to handle umems with a consistent simplified uAPI and update several drivers to use it. This brings dmabuf support to more places and more drivers - Support for mlx5 rate limit and packet pacing for UD and UC - A batch of fixes for the new shared FRMR pools infrastructure" * tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (148 commits) RDMA/irdma: Replace waitqueue and flag with completion RDMA/hns: Fix memory leak of bonding resources RDMA/rtrs-srv: Bound RDMA-Write length to chunk size in rdma_write_sg docs: infiniband: correct name of option to enable the ib_uverbs module RDMA/bnxt_re: Reject GET_TOGGLE_MEM when toggle page was not allocated RDMA/bnxt_re: Fail DBR related page allocation UAPIs if the feature is disabled RDMA/bnxt_re: Avoid repeated requests to allocate WC pages RDMA/bnxt_re: Proper rollback if the ioremap fails RDMA/bnxt_re: Add a max slot check for SQ RDMA/bnxt_re: Avoid displaying the kernel pointer RDMA/bnxt_re: Free CQ toggle page after firmware teardown RDMA/bnxt_re: Free SRQ toggle page after firmware teardown RDMA/bnxt_re: Initialize dpi variable to zero ABI: sysfs-class-infiniband: minor cleanup RDMA/mlx5: Release the HW‑provided UAR index rather than the SW one RDMA/mlx5: Fix undefined shift of user RQ WQE size RDMA/mlx5: Remove raw RSS QP restrack tracking RDMA/mlx5: Remove DCT restrack tracking RDMA/mlx5: Drop FRMR pool handle on UMR revoke failure RDMA/core: Add ib_frmr_pool_drop for unrecoverable handles ... 
Pull rdma updates from Jason Gunthorpe:
 "Many AI driven bug fixes, and several big driver API cleanups

   - Driver bug fixes and minor cleanups in mlx5, hns, rxe, efa, siw,
     rtrs, mana, irdma, mlx4. Commonly error path flows, integer
     arithmetic overflows on unsafe data, out of bounds access, and use
     after free issues under races.

   - Second half of the new udata API for drivers focusing on uAPI
     response

   - bnxt_re supports more options for QP creation that will allow a dv
     path in rdma-core

   - Untangle the module dependencies so drivers don't link to
     ib_uverbs.ko as was originall intended

   - Provide a new way to handle umems with a consistent simplified uAPI
     and update several drivers to use it. This brings dmabuf support to
     more places and more drivers

   - Support for mlx5 rate limit and packet pacing for UD and UC

   - A batch of fixes for the new shared FRMR pools infrastructure"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (148 commits)
  RDMA/irdma: Replace waitqueue and flag with completion
  RDMA/hns: Fix memory leak of bonding resources
  RDMA/rtrs-srv: Bound RDMA-Write length to chunk size in rdma_write_sg
  docs: infiniband: correct name of option to enable the ib_uverbs module
  RDMA/bnxt_re: Reject GET_TOGGLE_MEM when toggle page was not allocated
  RDMA/bnxt_re: Fail DBR related page allocation UAPIs if the feature is disabled
  RDMA/bnxt_re: Avoid repeated requests to allocate WC pages
  RDMA/bnxt_re: Proper rollback if the ioremap fails
  RDMA/bnxt_re: Add a max slot check for SQ
  RDMA/bnxt_re: Avoid displaying the kernel pointer
  RDMA/bnxt_re: Free CQ toggle page after firmware teardown
  RDMA/bnxt_re: Free SRQ toggle page after firmware teardown
  RDMA/bnxt_re: Initialize dpi variable to zero
  ABI: sysfs-class-infiniband: minor cleanup
  RDMA/mlx5: Release the HW‑provided UAR index rather than the SW one
  RDMA/mlx5: Fix undefined shift of user RQ WQE size
  RDMA/mlx5: Remove raw RSS QP restrack tracking
  RDMA/mlx5: Remove DCT restrack tracking
  RDMA/mlx5: Drop FRMR pool handle on UMR revoke failure
  RDMA/core: Add ib_frmr_pool_drop for unrecoverable handles
  ...
RDMA/core: Add ib_frmr_pool_drop for unrecoverable handles 2026-06-11T18:36:09+00:00 Michael Guralnik michaelgur@nvidia.com 2026-06-10T00:01:44+00:00 ddbc251be18fb82884ee6e9af634cc9f1171a4d6 A driver that has popped a handle from an FRMR pool can hit failures that leave the handle in a state where it can't safely be returned for reuse. The driver destroys the handle itself, but the pool has no way to learn about it, so the in_use counter drifts upward. Add ib_frmr_pool_drop to balance the pool's accounting in this case. Every pop is now balanced by exactly one push or drop. Fixes: 36680ef7bceb ("RDMA/mlx5: Switch from MR cache to FRMR pools") Link: https://patch.msgid.link/r/20260610000145.820592-9-michaelgur@nvidia.com Signed-off-by: Michael Guralnik <michaelgur@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> 
A driver that has popped a handle from an FRMR pool can hit failures
that leave the handle in a state where it can't safely be returned
for reuse. The driver destroys the handle itself, but the pool has
no way to learn about it, so the in_use counter drifts upward.

Add ib_frmr_pool_drop to balance the pool's accounting in this case.
Every pop is now balanced by exactly one push or drop.

Fixes: 36680ef7bceb ("RDMA/mlx5: Switch from MR cache to FRMR pools")
Link: https://patch.msgid.link/r/20260610000145.820592-9-michaelgur@nvidia.com
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
RDMA/core: Fix FRMR handle leak on push failure 2026-06-11T18:36:09+00:00 Michael Guralnik michaelgur@nvidia.com 2026-06-10T00:01:43+00:00 8c76126b866649d8e8acc09a06f2b03b6ff88900 Failure to push a handle to the pool, caused by ENOMEM on queue page allocation, will trigger missing in_use counter update, skewing pool state indefinitely. Fix that by moving the handling of handle destruction in such case into the FRMR code, ensuring the handle is either pushed to the pool or destroyed inside the same function. Adjust mlx5_ib call site accordingly. Fixes: ce5df0b891ed ("IB/core: Introduce FRMR pools") Link: https://patch.msgid.link/r/20260610000145.820592-8-michaelgur@nvidia.com Signed-off-by: Michael Guralnik <michaelgur@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> 
Failure to push a handle to the pool, caused by ENOMEM on queue page
allocation, will trigger missing in_use counter update, skewing pool
state indefinitely.
Fix that by moving the handling of handle destruction in such case
into the FRMR code, ensuring the handle is either pushed to the pool
or destroyed inside the same function.

Adjust mlx5_ib call site accordingly.

Fixes: ce5df0b891ed ("IB/core: Introduce FRMR pools")
Link: https://patch.msgid.link/r/20260610000145.820592-8-michaelgur@nvidia.com
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>