src/sys/dev/mlx5/mlx5_ib, branch main

mlx5ib: use the eventfd_ctx API for DEVX event subscriptions

2026-07-14T19:25:48+00:00

The DEVX_SUBSCRIBE_EVENT redirect path resolved the user's eventfd with
fdget(), which on FreeBSD only finds LinuxKPI files.  rdma-core creates
the eventfd with the native FreeBSD eventfd(2), so the lookup failed and
subscription returned EBADF; the delivery side likewise assumed a
LinuxKPI-pollable file.

Use the LinuxKPI eventfd_ctx API instead: eventfd_ctx_fdget() resolves
the native eventfd, eventfd_signal() notifies it, and eventfd_ctx_put()
releases it.  DEVX async events can then be delivered through a redirect
eventfd.

Reviewed by: kib
Sponsored by: Nvidia networking
MFC after: 1 month

mlx5ib: initialize DEVX subscription state before the eventfd fdget()

2026-07-14T10:39:54+00:00

In the DEVX_SUBSCRIBE_EVENT handler the eventfd path can fail and
"goto err" before the subscription's xa keys and ev_file have been set;
they are still zeroed from kzalloc().  The cleanup then looks up a
level-1 xa entry with key 0, gets NULL, and faults dereferencing it.

Initialize the fields the cleanup path relies on right after the
subscription is allocated, before it is linked and before the fallible
fdget(), so a later failure unwinds cleanly.

Reviewed by: kib
Sponsored by: Nvidia networking
MFC after: 1 month

mlx5ib: advertise write-combining support for dynamic BlueFlame UARs

2026-07-07T11:25:15+00:00

Import Linux upstream commit 1f3db161881b7e21efb149e0ae8152b79a571a8f.

dev->wc_support was never set, so it was always false and the UAR ioctl
refused BlueFlame (write-combining) UAR allocations with EOPNOTSUPP.
That breaks QP creation in pure dynamic-UAR mode, where user space asks
for a BF doorbell UAR.

Reviewed by:	kib
Tested by:	Wafa Hamzah 
Sponsored by:	Nvidia networking
MFC after:	1 month

mlx5ib: encode dynamic UAR mmap offsets in the reserved command range

2026-07-07T11:25:06+00:00

The UAR ioctl handed user space a raw mmap offset, so the first dynamic
UAR landed at page offset 0.  mlx5_ib_mmap() decodes offset 0 as the
legacy regular-page command and routed the mapping through the old bfreg
path, which rejects dynamic-UAR contexts, so mmap() failed with EINVAL
and mlx5dv_devx_alloc_uar() returned NULL.

Follow the upstream scheme: reserve the mmap command range [9, 255] for
rdma_user_mmap entries and return command-encoded offsets, so the
dynamic-UAR mappings decode to the intended mlx5_ib_mmap() path.

Reviewed by:	kib
Tested by:	Wafa Hamzah 
Sponsored by:	Nvidia networking
MFC after:	1 month

mlx5ib: allocate IB queue counters as a shared resource

2026-07-07T11:24:57+00:00

A QP owned by a DEVX uid references the port's queue counter.  The
counter was allocated with uid 0, so RST2INIT_QP on a uid-owned QP
failed with "bad resource state".

Allocate and free the IB queue counters directly and, on devices that
support user contexts, stamp them with MLX5_SHARED_RESOURCE_UID so
uid-owned QPs can use them.

The code follows the Linux commit d2c8a1554c10d5e0443b1f97f480d7dacd55cf55
("IB/mlx5: Enable UAR to have DevX UID").

Reviewed by:	kib
Tested by:	Wafa Hamzah 
Sponsored by:	Nvidia networking
MFC after:	1 month

mlx5ib: implement the MLX5_IB_OBJECT_UAR ioctl object

2026-07-07T11:24:38+00:00

Import Linux upstream commit 342ee59de98a2ecdf15a46849a2534e7c808eb1f.

The dynamic UAR object was declared in the ABI headers but had no handler,
so the ioctl was rejected and dynamic-UAR contexts could not allocate a
doorbell UAR at all.

Implement the alloc and destroy methods following the upstream driver:
grab a UAR stamped with the caller's DEVX uid, expose it to user space
through an rdma_user_mmap entry (write-combining or non-cached as
requested), and free it on destroy.

Reviewed by:	kib
Tested by:	Wafa Hamzah 
Sponsored by:	Nvidia networking
MFC after:	1 month

mlx5_ib: register DEVX objects in the uverbs ioctl parse tree

2026-07-07T11:24:29+00:00

Import Linux upstream commits a8b92ca1b0e5ce620e425e9d2f89ce44f1a82a82
and c59450c463695a016e823175bac421cff219935d.

The DEVX object and method definitions were already present, but nothing
pointed ib_device.driver_def at them.  ibcore therefore never merged
them into the uverbs uapi tree and every DEVX ioctl came back as
EPROTONOSUPPORT.

Reviewed by:	kib
Tested by:	Wafa Hamzah 
Sponsored by:	Nvidia networking
MFC after:	1 month

mlx5_ib: do not consume CMD/PAGE_REQUEST events in the DEVX notifier

2026-07-07T11:24:21+00:00

DEVX event notifier returned true for the command-completion and
page-request events. This is causing mlx5_eq_int() to skip the core EQ
handler, so the firmware command interface and the page supply stop
being serviced and the device wedges.

This commit also make notifier registration and dispatch safe against
the EQ interrupt running concurrently: publish the table pointer before
the callback and load it with acquire semantics. run the callback under
RCU, and drain it with synchronize_rcu() on teardown. Otherwise the
interrupt handler could observe a half-initialized notifier or race with
cleanup.

Reviewed by:	kib
Tested by:	Wafa Hamzah 
Sponsored by:	Nvidia networking
MFC after:	1 month

OFED: A few channges from Linux 5.0

2026-06-24T19:30:56+00:00

The main point of these changes is the last commit adding a couple of
helper functions used in ROCE drivers.

This contains changes from the following Linux commits:

a70c07397fd8 RDMA: Introduce and use GID attr helper to read RoCE L2 fields
8f9748602491 IB/cm: Reduce dependency on gid attribute ndev check
adb4a57a7a1d RDMA/cma: Use rdma_read_gid_attr_ndev_rcu to access netdev
b4fb4cc5ba83 RDMA/cma: Fix unbalanced cm_id reference count during address resolve
d5665a21250e RDMA/core: Add hash functions to calculate RoCEv2 flowlabel and UDP source port

Tested by:	Wafa Hamzah  (mlx5_ib)
Tested by:	John Baldwin  (iw_cxgbe)
Sponsored by:	Chelsio Communications

OFED: Various changes from Linux 4.19

2026-06-24T19:30:56+00:00

Many of these changes deal with shifting from fetcing SGID attributes
via an index in drivers to including references to SGID attributes in
other structures passed to driver callbacks which can be used directly
avoiding the need for fetching attributes via lookup functions.  I had
to fixup the ROCE GID management code directly as the version in
FreeBSD is very different from Linux.

I also tweaked the change from commit
475c8de7bf2dc86d9806709a04e9c8f524d4ec32 to use a goto to reduce diffs
with upstream to minimize conflicts when merging upstream changes.

This contains changes from the following Linux commits:

a1a4caeebac9 IB/core: Do not set the gid type when reserving default entries
1c36cf912ad1 IB/core: Store default GID property per-table instead of per-entry
1dfce2945771 IB: Replace ib_query_gid/ib_get_cached_gid with rdma_query_gid
83f6f8d29dd3 IB/core: Make rdma_find_gid_by_filter support all protocols
c3d71b69a75c IB/core: Provide rdma_ versions of the gid cache API
77e786fcbe2e IB/core: Replace ib_query_gid with rdma_get_gid_attr
bf399c2cadfa IB/core: Introduce GID attribute get, put and hold APIs
f4df9a7c34d8 RDMA: Use GID from the ib_gid_attr during the add_gid() callback
b150c3862d21 IB/core: Introduce GID entry reference counts
47ec38666210 RDMA: Convert drivers to use sgid_attr instead of sgid_index
d97099fe53ed IB{cm, core}: Introduce and use ah_attr copy, move, replace APIs
947c99ecfcb4 IB/core: Tidy ib_resolve_eth_dmac
8d9ec9addd6c IB/core: Add a sgid_attr pointer to struct rdma_ah_attr
89af969a6653 RDMA: Convert drivers to use the AH's sgid_attr in post_wr paths
1a1f460ff151 RDMA: Hold the sgid_attr inside the struct ib_ah/qp
7492052a186b IB/mlx4: Use GID attribute from ah attribute
59d40813328f IB/core: Free GID table entry during GID deletion
881456789248 RDMA/cma: Consider net namespace while leaving multicast group
4ed13a5f2d60 IB/cm: Keep track of the sgid_attr that created the cm id
aa74f4878d61 IB: Make init_ah_attr_grh_fields set sgid_attr
f685c19529f0 IB: Make ib_init_ah_from_mcmember set sgid_attr
b7403217656d IB: Make ib_init_ah_attr_from_wc set sgid_attr
ea8c2d8f6014 RDMA/core: Remove unused ib cache functions
a8872d53e9b7 IB/cm: Use sgid_attr from the AV
398391071f25 IB/cm: Replace members of sa_path_rec with 'struct sgid_attr *'
815d456ef21a IB/cm: Pass the sgid_attr through various events
4eefd62c17a9 include/rdma/opa_addr.h: Fix an endianness issue
f8c2d2280cf6 RDMA/core: Remove set-but-not-used variables
28e39894ed4f RDMA/core: Remove ib_find_cached_gid() and ib_find_cached_gid_by_port()
07e7056aff6c IB/core: Simplify check for RoCE route resolve
40ddacf2dda9 RDMA/umem: Don't hold mmap_sem for too long
1215cb7c88ec RDMA/umem: Refactor exit paths in ib_umem_get
c012691508f5 IB/cm: Remove cma_multicast->igmp_joined
4b4671a0f269 IB/IPoIB: Simplify ib_post_(send|recv|srq_recv)() calls
1fec77bf8fcd RDMA/core: Simplify ib_post_(send|recv|srq_recv)() calls
bb039a870c05 IB/core: Allow ULPs to specify NULL as the third ib_post_(send|recv|srq_recv)() argument
643d213a9a03 RDMA/cma: Do not ignore net namespace for unbound cm_id
d274e45ce1ed RDMA/cma: Consider netdevice for RoCE ports
cee104334c98 IB/core: Introduce and use sgid_attr in CM requests
7150c3d5544b RDMA/core: Remove {create,destroy}_ah from mandatory verbs
854633165164 RDMA/core: Prefix _ib to IB/RoCE specific functions
7582df826734 RDMA/core: Avoid holding lock while initializing fields on stack
79d684f02647 RDMA/core: Simplify gid type check in cma_acquire_dev()
ca3a8ace2b12 RDMA/core: Return bool instead of int
05e0b86c413d RDMA/cma: Get rid of 1 bit boolean
e7ff98aefc9e RDMA/cma: Constify path record, ib_cm_event, listen_id pointers
2df7dba855e1 RDMA/core: Constify dst_addr argument
219d2e9dfda9 RDMA/cma: Simplify rdma_resolve_addr() error flow
26e551c5aec5 RDMA: Fix return code check in rdma_set_cq_moderation
58796e67d5d5 IB/ucm: Initialize sgid request GID attribute pointer
dd81b2c8a333 IB/core: Change filter function return type from int to bool
93065ac753e4 mm, oom: distinguish blockable mode for mmu notifiers
0d23ba6034b9 RDMA/ucma: check fd type in ucma_migrate_id()

Part of cee104334c98 was previously merged into FreeBSD.  Namely, SGID
index fields had been added to the two param structures instead and
roce_get_net_dev_by_cm_event was implemented using SGID indices.

Tested by:	Wafa Hamzah  (mlx5_ib)
Tested by:	John Baldwin  (iw_cxgbe)
Sponsored by:	Chelsio Communications