I encontered the following warning:
WARNING: drivers/infiniband/sw/rxe/rxe_task.c:249 at rxe_sched_task+0x1c8/0x238 [rdma_rxe], CPU#0: swapper/0/0
...
libsha1 [last unloaded: ip6_udp_tunnel]
CPU: 0 UID: 0 PID: 0 Comm: swapper/0 Tainted: G C 6.19.0-rc5-64k-v8+ #37 PREEMPT
Tainted: [C]=CRAP
Hardware name: Raspberry Pi 4 Model B Rev 1.2
Call trace:
rxe_sched_task+0x1c8/0x238 [rdma_rxe] (P)
retransmit_timer+0x130/0x188 [rdma_rxe]
call_timer_fn+0x68/0x4d0
__run_timers+0x630/0x888
...
WARNING: drivers/infiniband/sw/rxe/rxe_task.c:38 at rxe_sched_task+0x1c0/0x238 [rdma_rxe], CPU#0: swapper/0/0
...
WARNING: drivers/infiniband/sw/rxe/rxe_task.c:111 at do_work+0x488/0x5c8 [rdma_rxe], CPU#3: kworker/u17:4/93400
...
refcount_t: underflow; use-after-free.
WARNING: lib/refcount.c:28 at refcount_warn_saturate+0x138/0x1a0, CPU#3: kworker/u17:4/93400
The issue is caused by a race condition between retransmit_timer() and
rxe_destroy_qp, leading to the Queue Pair's (QP) reference count dropping
to zero during timer handler execution.
It seems this warning is harmless because rxe_qp_do_cleanup() will flush
all pending timers and requests.
Example of flow causing the issue:
CPU0 CPU1
retransmit_timer() {
spin_lock_irqsave
rxe_destroy_qp()
__rxe_cleanup()
__rxe_put() // qp->ref_count decrease to 0
rxe_qp_do_cleanup() {
if (qp->valid) {
rxe_sched_task() {
WARN_ON(rxe_read(task->qp) <= 0);
}
}
spin_unlock_irqrestore
}
spin_lock_irqsave
qp->valid = 0
spin_unlock_irqrestore
}
Ensure the QP's reference count is maintained and its validity is checked
within the timer callbacks by adding calls to rxe_get(qp) and corresponding
rxe_put(qp) after use.
Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
Fixes: d946716325 ("RDMA/rxe: Rewrite rxe_task.c")
Link: https://patch.msgid.link/20260120074437.623018-1-lizhijian@fujitsu.com
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The UVERBS_HANDLER(MLX5_IB_METHOD_GET_DATA_DIRECT_SYSFS_PATH) function
allocates memory for the device path using kobject_get_path(). If the
length of the device path exceeds the output buffer length, the function
returns -ENOSPC but does not free the allocated memory, resulting in a
memory leak.
Add a kfree() call to the error path to ensure the allocated memory is
properly freed.
Compile tested only. Issue found using a prototype static analysis tool
and code review.
Fixes: ec7ad65309 ("RDMA/mlx5: Introduce GET_DATA_DIRECT_SYSFS_PATH ioctl")
Signed-off-by: Zilin Guan <zilin@seu.edu.cn>
Link: https://patch.msgid.link/20260126074801.627898-1-zilin@seu.edu.cn
Signed-off-by: Leon Romanovsky <leon@kernel.org>
ib_uverbs_post_send() uses cmd.wqe_size from userspace without any
validation before passing it to kmalloc() and using the allocated
buffer as struct ib_uverbs_send_wr.
If a user provides a small wqe_size value (e.g., 1), kmalloc() will
succeed, but subsequent accesses to user_wr->opcode, user_wr->num_sge,
and other fields will read beyond the allocated buffer, resulting in
an out-of-bounds read from kernel heap memory. This could potentially
leak sensitive kernel information to userspace.
Additionally, providing an excessively large wqe_size can trigger a
WARNING in the memory allocation path, as reported by syzkaller.
This is inconsistent with ib_uverbs_unmarshall_recv() which properly
validates that wqe_size >= sizeof(struct ib_uverbs_recv_wr) before
proceeding.
Add the same validation for ib_uverbs_post_send() to ensure wqe_size
is at least sizeof(struct ib_uverbs_send_wr).
Fixes: c3bea3d2dc ("RDMA/uverbs: Use the iterator for ib_uverbs_unmarshall_recv()")
Signed-off-by: Yi Liu <liuy22@mails.tsinghua.edu.cn>
Link: https://patch.msgid.link/20260122142900.2356276-2-liuy22@mails.tsinghua.edu.cn
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The hardware allows for an opaque CQ context field to be carried
over into CEQEs for the CQ. Previously, a pointer to the CQ was used
for this context. In the normal CQ destroy flow, the CEQ ring is
scrubbed to remove any preexisting CEQEs for the CQ that may not have
been processed yet so that the CQ structure is not dereferenced in the
CEQ ISR after the CQ has been freed.
However, in some cases, it is possible for a CEQE to be in flight in
HW even after the CQ destroy command completion is received, so it
could be missed during the scrub.
To protect against this, we can take advantage of the CQ table that
already exists and use the CQ ID for this context rather than a CQ
pointer.
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Link: https://patch.msgid.link/20260120212546.1893076-2-jmoroni@google.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The current implementation incorrectly handles memory regions (MRs) with
page sizes different from the system PAGE_SIZE. The core issue is that
rxe_set_page() is called with mr->page_size step increments, but the
page_list stores individual struct page pointers, each representing
PAGE_SIZE of memory.
ib_sg_to_page() has ensured that when i>=1 either
a) SG[i-1].dma_end and SG[i].dma_addr are contiguous
or
b) SG[i-1].dma_end and SG[i].dma_addr are mr->page_size aligned.
This leads to incorrect iova-to-va conversion in scenarios:
1) page_size < PAGE_SIZE (e.g., MR: 4K, system: 64K):
ibmr->iova = 0x181800
sg[0]: dma_addr=0x181800, len=0x800
sg[1]: dma_addr=0x173000, len=0x1000
Access iova = 0x181800 + 0x810 = 0x182010
Expected VA: 0x173010 (second SG, offset 0x10)
Before fix:
- index = (0x182010 >> 12) - (0x181800 >> 12) = 1
- page_offset = 0x182010 & 0xFFF = 0x10
- xarray[1] stores system page base 0x170000
- Resulting VA: 0x170000 + 0x10 = 0x170010 (wrong)
2) page_size > PAGE_SIZE (e.g., MR: 64K, system: 4K):
ibmr->iova = 0x18f800
sg[0]: dma_addr=0x18f800, len=0x800
sg[1]: dma_addr=0x170000, len=0x1000
Access iova = 0x18f800 + 0x810 = 0x190010
Expected VA: 0x170010 (second SG, offset 0x10)
Before fix:
- index = (0x190010 >> 16) - (0x18f800 >> 16) = 1
- page_offset = 0x190010 & 0xFFFF = 0x10
- xarray[1] stores system page for dma_addr 0x170000
- Resulting VA: system page of 0x170000 + 0x10 = 0x170010 (wrong)
Yi Zhang reported a kernel panic[1] years ago related to this defect.
Solution:
1. Replace xarray with pre-allocated rxe_mr_page array for sequential
indexing (all MR page indices are contiguous)
2. Each rxe_mr_page stores both struct page* and offset within the
system page
3. Handle MR page_size != PAGE_SIZE relationships:
- page_size > PAGE_SIZE: Split MR pages into multiple system pages
- page_size <= PAGE_SIZE: Store offset within system page
4. Add boundary checks and compatibility validation
This ensures correct iova-to-va conversion regardless of MR page size
and system PAGE_SIZE relationship, while improving performance through
array-based sequential access.
Tests on 4K and 64K PAGE_SIZE hosts:
- rdma-core/pytests
$ ./build/bin/run_tests.py --dev eth0_rxe
- blktest:
$ TIMEOUT=30 QUICK_RUN=1 USE_RXE=1 NVMET_TRTYPES=rdma ./check nvme srp rnbd
[1] https://lore.kernel.org/all/CAHj4cs9XRqE25jyVw9rj9YugffLn5+f=1znaBEnu1usLOciD+g@mail.gmail.com/T/
Fixes: 592627ccbd ("RDMA/rxe: Replace rxe_map and rxe_phys_buf by xarray")
Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
Link: https://patch.msgid.link/20260116032753.2574363-1-lizhijian@fujitsu.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
In rxe_map_mr_sg(), the `page_offset` member of the `rxe_mr` struct
was initialized based on `ibmr.iova`, which will be updated inside
ib_sg_to_pages() later.
Consequently, the value assigned to `page_offset` was incorrect. However,
since `page_offset` was never utilized throughout the code, it can be safely
removed to clean up the codebase and avoid future confusion.
Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
Link: https://patch.msgid.link/20260116032833.2574627-1-lizhijian@fujitsu.com
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
When querying speed information for a representor in switchdev mode,
the code previously used the first device in the eswitch, which may not
match the device that actually owns the representor. In setups such as
multi-port eswitch or LAG, this led to incorrect port attributes being
reported.
Fix this by retrieving the correct core device from the representor's
eswitch before querying its port attributes.
Fixes: 27f9e0ccb6 ("net/mlx5: Lag, Add single RDMA device in multiport mode")
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Link: https://patch.msgid.link/20260115-port-speed-query-fix-v2-1-3bde6a3c78e7@nvidia.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
During firmware reset in LAG mode, a race condition causes the driver
to hang indefinitely while waiting for UMR completion during device
unload. See [1].
In LAG mode the bond device is only registered on the master, so it
never sees sys_error events from the slave.
During firmware reset this causes UMR waits to hang forever on unload
as the slave is dead but the master hasn't entered error state yet, so
UMR posts succeed but completions never arrive.
Fix this by adding a sys_error notifier that gets registered before
MLX5_IB_STAGE_IB_REG and stays alive until after ib_unregister_device().
This ensures error events reach the bond device throughout teardown.
[1]
Call Trace:
__schedule+0x2bd/0x760
schedule+0x37/0xa0
schedule_preempt_disabled+0xa/0x10
__mutex_lock.isra.6+0x2b5/0x4a0
__mlx5_ib_dereg_mr+0x606/0x870 [mlx5_ib]
? __xa_erase+0x4a/0xa0
? _cond_resched+0x15/0x30
? wait_for_completion+0x31/0x100
ib_dereg_mr_user+0x48/0xc0 [ib_core]
? rdmacg_uncharge_hierarchy+0xa0/0x100
destroy_hw_idr_uobject+0x20/0x50 [ib_uverbs]
uverbs_destroy_uobject+0x37/0x150 [ib_uverbs]
__uverbs_cleanup_ufile+0xda/0x140 [ib_uverbs]
uverbs_destroy_ufile_hw+0x3a/0xf0 [ib_uverbs]
ib_uverbs_remove_one+0xc3/0x140 [ib_uverbs]
remove_client_context+0x8b/0xd0 [ib_core]
disable_device+0x8c/0x130 [ib_core]
__ib_unregister_device+0x10d/0x180 [ib_core]
ib_unregister_device+0x21/0x30 [ib_core]
__mlx5_ib_remove+0x1e4/0x1f0 [mlx5_ib]
auxiliary_bus_remove+0x1e/0x30
device_release_driver_internal+0x103/0x1f0
bus_remove_device+0xf7/0x170
device_del+0x181/0x410
mlx5_rescan_drivers_locked.part.10+0xa9/0x1d0 [mlx5_core]
mlx5_disable_lag+0x253/0x260 [mlx5_core]
mlx5_lag_disable_change+0x89/0xc0 [mlx5_core]
mlx5_eswitch_disable+0x67/0xa0 [mlx5_core]
mlx5_unload+0x15/0xd0 [mlx5_core]
mlx5_unload_one+0x71/0xc0 [mlx5_core]
mlx5_sync_reset_reload_work+0x83/0x100 [mlx5_core]
process_one_work+0x1a7/0x360
worker_thread+0x30/0x390
? create_worker+0x1a0/0x1a0
kthread+0x116/0x130
? kthread_flush_work_fn+0x10/0x10
ret_from_fork+0x22/0x40
Fixes: ede132a5cf ("RDMA/mlx5: Move events notifier registration to be after device registration")
Signed-off-by: Chiara Meiohas <cmeiohas@nvidia.com>
Signed-off-by: Maher Sanalla <msanalla@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Link: https://patch.msgid.link/20260113-umr-hand-lag-fix-v1-1-3dc476e00cd9@nvidia.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The commit e1168f0 ("RDMA/iwcm: Simplify cm_event_handler()")
changed the work submission logic to unconditionally call
queue_work() with the expectation that queue_work() would
have no effect if work was already pending. The problem is
that a free list of struct iwcm_work is used (for which
struct work_struct is embedded), so each call to queue_work()
is basically unique and therefore does indeed queue the work.
This causes a problem in the work handler which walks the work_list
until it's empty to process entries. This means that a single
run of the work handler could process item N+1 and release it
back to the free list while the actual workqueue entry is still
queued. It could then get reused (INIT_WORK...) and lead to
list corruption in the workqueue logic.
Fix this by just removing the work_list. The workqueue already
does this for us.
This fixes the following error that was observed when stress
testing with ucmatose on an Intel E830 in iWARP mode:
[ 151.465780] list_del corruption. next->prev should be ffff9f0915c69c08, but was ffff9f0a1116be08. (next=ffff9f0a15b11c08)
[ 151.466639] ------------[ cut here ]------------
[ 151.466986] kernel BUG at lib/list_debug.c:67!
[ 151.467349] Oops: invalid opcode: 0000 [#1] SMP NOPTI
[ 151.467753] CPU: 14 UID: 0 PID: 2306 Comm: kworker/u64:18 Not tainted 6.19.0-rc4+ #1 PREEMPT(voluntary)
[ 151.468466] Hardware name: QEMU Ubuntu 24.04 PC (i440FX + PIIX, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
[ 151.469192] Workqueue: 0x0 (iw_cm_wq)
[ 151.469478] RIP: 0010:__list_del_entry_valid_or_report+0xf0/0x100
[ 151.469942] Code: c7 58 5f 4c b2 e8 10 50 aa ff 0f 0b 48 89 ef e8 36 57 cb ff 48 8b 55 08 48 89 e9 48 89 de 48 c7 c7 a8 5f 4c b2 e8 f0 4f aa ff <0f> 0b 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 40 00 90 90 90 90 90 90
[ 151.471323] RSP: 0000:ffffb15644e7bd68 EFLAGS: 00010046
[ 151.471712] RAX: 000000000000006d RBX: ffff9f0915c69c08 RCX: 0000000000000027
[ 151.472243] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9f0a37d9c600
[ 151.472768] RBP: ffff9f0a15b11c08 R08: 0000000000000000 R09: c0000000ffff7fff
[ 151.473294] R10: 0000000000000001 R11: ffffb15644e7bba8 R12: ffff9f092339ee68
[ 151.473817] R13: ffff9f0900059c28 R14: ffff9f092339ee78 R15: 0000000000000000
[ 151.474344] FS: 0000000000000000(0000) GS:ffff9f0a847b5000(0000) knlGS:0000000000000000
[ 151.474934] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 151.475362] CR2: 0000559e233a9088 CR3: 000000020296b004 CR4: 0000000000770ef0
[ 151.475895] PKRU: 55555554
[ 151.476118] Call Trace:
[ 151.476331] <TASK>
[ 151.476497] move_linked_works+0x49/0xa0
[ 151.476792] __pwq_activate_work.isra.46+0x2f/0xa0
[ 151.477151] pwq_dec_nr_in_flight+0x1e0/0x2f0
[ 151.477479] process_scheduled_works+0x1c8/0x410
[ 151.477823] worker_thread+0x125/0x260
[ 151.478108] ? __pfx_worker_thread+0x10/0x10
[ 151.478430] kthread+0xfe/0x240
[ 151.478671] ? __pfx_kthread+0x10/0x10
[ 151.478955] ? __pfx_kthread+0x10/0x10
[ 151.479240] ret_from_fork+0x208/0x270
[ 151.479523] ? __pfx_kthread+0x10/0x10
[ 151.479806] ret_from_fork_asm+0x1a/0x30
[ 151.480103] </TASK>
Fixes: e1168f09b3 ("RDMA/iwcm: Simplify cm_event_handler()")
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Link: https://patch.msgid.link/20260112020006.1352438-1-jmoroni@google.com
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
In rxe_srq_from_init(), the queue pointer 'q' is assigned to
'srq->rq.queue' before copying the SRQ number to user space.
If copy_to_user() fails, the function calls rxe_queue_cleanup()
to free the queue, but leaves the now-invalid pointer in
'srq->rq.queue'.
The caller of rxe_srq_from_init() (rxe_create_srq) eventually
calls rxe_srq_cleanup() upon receiving the error, which triggers
a second rxe_queue_cleanup() on the same memory, leading to a
double free.
The call trace looks like this:
kmem_cache_free+0x.../0x...
rxe_queue_cleanup+0x1a/0x30 [rdma_rxe]
rxe_srq_cleanup+0x42/0x60 [rdma_rxe]
rxe_elem_release+0x31/0x70 [rdma_rxe]
rxe_create_srq+0x12b/0x1a0 [rdma_rxe]
ib_create_srq_user+0x9a/0x150 [ib_core]
Fix this by moving 'srq->rq.queue = q' after copy_to_user.
Fixes: aae0484e15 ("IB/rxe: avoid srq memory leak")
Signed-off-by: Jiasheng Jiang <jiashengjiangcool@gmail.com>
Link: https://patch.msgid.link/20260112015412.29458-1-jiashengjiangcool@gmail.com
Reviewed-by: Zhu Yanjun <yanjun.Zhu@linux.dev>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Some ULPs, e.g. rpcrdma, rely on drain_qp() to ensure all outstanding
requests are completed before releasing related memory. If drain_qp()
fails, ULPs may release memory directly, and in-flight WRs may later be
flushed after the memory is freed, potentially leading to UAF.
drain_qp() failures can happen when HW enters an error state or is
reset. Add support to drain SQ and RQ in such cases by posting a
fake WR during reset, so the driver can process all remaining WRs in
sequence and generate corresponding completions.
Always invoke comp_handler() in drain process to ensure completions
are not lost under concurrency (e.g. concurrent post_send() and
reset, or QPs created during reset). If the CQ is already processed,
cancel any already scheduled comp_handler() to avoid concurrency
issues.
Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20260108113032.856306-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The AH CQP command wait loop executes in an atomic context and was
using a fixed 1 ms delay. Since many AH create commands can complete
much faster than 1 ms, use poll_timeout_us_atomic with a 1 us delay.
Also, use the timeout value indicated during the capability exchange
rather than a hard-coded value.
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Link: https://patch.msgid.link/20260105180550.2907858-1-jmoroni@google.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
A dma_wmb() is not necessary before a writel() because writel()
already has an even stronger store barrier. A dma_wmb() is only
required to order writes to consistent/DMA memory whereas the
barrier in writel() is specified to order writes to DMA memory as
well as MMIO.
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Link: https://patch.msgid.link/20260103172517.2088895-1-jmoroni@google.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
During several network incidents, a number of RTRS paths for a session
went through disconnect and reconnect phase. However, some of those did
not auto-reconnect successfully. Instead they failed with the following
logs,
On client,
kernel: rtrs_client L1991: <sess-name>: Connect rejected: status 28
(consumer defined), rtrs errno -104
kernel: rtrs_client L2698: <sess-name>: init_conns() failed: err=-104
path=gid:<gid1>@gid:<gid2> [mlx4_0:1]
On server, (log a)
kernel: ibtrs_server L1868: <>: Connection already exists: 0
When the misbehaving path was removed, and add_path was called to re-add
the path, the log on client side changed to, (log b)
kernel: rtrs_client L1991: <sess-name>: Connect rejected: status 28
(consumer defined), rtrs errno -17
There was no log on the server side for this, which is expected since
there is no logging in that path,
if (unlikely(__is_path_w_addr_exists(srv, &cm_id->route.addr))) {
err = -EEXIST;
goto err;
Because of the following check on server side,
if (unlikely(sess->state != IBTRS_SRV_CONNECTING)) {
ibtrs_err(s, "Session in wrong state: %s\n",
.. we know that the path in (log a) was in CONNECTING state.
The above state of the path persists for as long as we leave the session
be. This means that the path is in some zombie state, probably waiting
for the info_req packet to arrive, which never does.
The changes in this commits does 2 things.
1) Add logs at places where we see the errors happening. The logs would
shed more light at the state and lifetime of such zombie paths.
2) Close such zombie sessions, only if they are in CONNECTING state, and
after an inactivity period of 30 seconds.
i) The state check prevents closure of paths which are CONNECTED.
Also, from the above logs and code, we already know that the path could
only be on CONNECTING state, so we play safe and narrow our impact surface
area by closing only CONNECTING paths.
ii) The inactivity period is to allow requests for other cid to finish
processing, or for any stray packets to arrive/fail.
Signed-off-by: Md Haris Iqbal <haris.iqbal@ionos.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Signed-off-by: Grzegorz Prajsner <grzegorz.prajsner@ionos.com>
Link: https://patch.msgid.link/20260107161517.56357-7-haris.iqbal@ionos.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
This fixes the following error on the server side:
RTRS server session allocation failed: -EINVAL
caused by the caller of the `ib_dma_map_sg()`, which does not expect
less mapped entries, than requested, which is in the order of things
and can be easily reproduced on the machine with enabled IOMMU.
The fix is to treat any positive number of mapped sg entries as a
successful mapping and cache DMA addresses by traversing modified
SG table.
Fixes: 9cb8374804 ("RDMA/rtrs: server: main functionality")
Signed-off-by: Roman Penyaev <r.peniaev@gmail.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Signed-off-by: Grzegorz Prajsner <grzegorz.prajsner@ionos.com>
Link: https://patch.msgid.link/20260107161517.56357-2-haris.iqbal@ionos.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Implement the query_port_speed callback for mlx5 driver to support
querying effective port bandwidth.
For LAG configurations, query the aggregated speed from the LAG layer
or from the modified vport max_tx_speed.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Raise IB_EVENT_DEVICE_SPEED_CHANGE whenever the speed of one of the
device's ports changes. Usually all ports of the device changes
together.
This ensures user applications and upper-layer software are immediately
notified when bandwidth changes, improving traffic management in dynamic
environments. This is especially useful for vports which are part of a
LAG configuration, to know if the effective speed of the LAG was
changed.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Add new ibv_query_port_speed() verb to enable applications to query
the effective bandwidth of a port.
This verb is particularly useful when the speed is not a multiplication
of IB speed and width where width is 2^n.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Update sysfs rate_show() to rely on ib_port_attr_to_speed_info() for
converting IB port speed and width attributes to data rate and speed
string.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Introduce ib_port_attr_to_rate() to compute the data rate in 100 Mbps
units (deci-Gb/sec) from a port's active_speed and active_width
attributes. This generic helper removes duplicated speed-to-rate
calculations, which are used by sysfs and the upcoming new verb.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Add IB_EVENT_DEVICE_SPEED_CHANGE for notifying user applications on
device's ports speed changes.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Currently, mlx5 driver exposes only the parent function's speed to VFs,
providing no way to query the actual effective bandwidth in LAG and
MPESW configurations. This limitation prevents userspace and
upper-layer software from obtaining accurate bandwidth information,
which impacts traffic scheduling decisions.
This series addresses this by:
1. Adding mlx5 internal logic to calculate and propagate the effective
aggregated LAG speed to all attached vports. The vport speeds are
dynamically updated when LAG member link states change.
2. Extending RDMA core with a new ib_query_port_speed() verb and an
IB_EVENT_DEVICE_SPEED_CHANGE async event. These interfaces expose
the effective port speed to userspace, supporting speeds that are
not expressible as IB speed * width (where width is 2^n).
This series enables userspace applications to query the effective
port speed and receive notifications on speed changes in real-time.
In LAG configurations, each mlx5 port reports the aggregated bandwidth
of all active LAG members.
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Add mlx5_lag_query_bond_speed() to query the aggregated speed of
lag configurations with a bond device.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Add port change event handling logic for MPESW LAG mode, ensuring
VFs are updated when the speed of LAG physical ports changes.
This triggers a speed update workflow when relevant port state changes
occur, enabling consistent and accurate reporting of VF bandwidth.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Currently, vports report only their parent's uplink speed, which in LAG
setups does not reflect the true aggregated bandwidth. This makes it
hard for upper-layer software to optimize load balancing decisions
based on accurate bandwidth information.
Fix the issue by calculating the possible maximum speed of a LAG as
the sum of speeds of all active uplinks that are part of the LAG.
Propagate this effective max speed to vports associated with the LAG
whenever a relevant event occurs, such as physical port link state
changes or LAG creation/modification.
With this change, upper-layer components receive accurate bandwidth
information corresponding to the active members of the LAG and can
make better load balancing decisions.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Introduce the max_tx_speed field to the query and modify_vport_state
structures.
Add the esw_vport_state_max_tx_speed capability bit, indicating
the firmware support modifying the max_tx_speed field via the
MODIFY_VPORT_STATE command.
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Reviewed-by: Maher Sanalla <msanalla@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
During a reset, software-generated WCs cannot be reported via
interrupts. This may cause the ULP to miss some WCs.
To avoid this, add check in the CQ arm process: if a hardware reset
has occurred and there are still unreported soft-WCs, notify the ULP
to handle the remaining WCs, thereby preventing any loss of completions.
Fixes: 626903e935 ("RDMA/hns: Add support for reporting wc as software mode")
Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20260104064057.1582216-5-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
When sunrpc is used, if a reset triggered, our wq may lead the
following trace:
workqueue: WQ_MEM_RECLAIM xprtiod:xprt_rdma_connect_worker [rpcrdma]
is flushing !WQ_MEM_RECLAIM hns_roce_irq_workq:flush_work_handle
[hns_roce_hw_v2]
WARNING: CPU: 0 PID: 8250 at kernel/workqueue.c:2644 check_flush_dependency+0xe0/0x144
Call trace:
check_flush_dependency+0xe0/0x144
start_flush_work.constprop.0+0x1d0/0x2f0
__flush_work.isra.0+0x40/0xb0
flush_work+0x14/0x30
hns_roce_v2_destroy_qp+0xac/0x1e0 [hns_roce_hw_v2]
ib_destroy_qp_user+0x9c/0x2b4
rdma_destroy_qp+0x34/0xb0
rpcrdma_ep_destroy+0x28/0xcc [rpcrdma]
rpcrdma_ep_put+0x74/0xb4 [rpcrdma]
rpcrdma_xprt_disconnect+0x1d8/0x260 [rpcrdma]
xprt_rdma_connect_worker+0xc0/0x120 [rpcrdma]
process_one_work+0x1cc/0x4d0
worker_thread+0x154/0x414
kthread+0x104/0x144
ret_from_fork+0x10/0x18
Since QP destruction frees memory, this wq should have the WQ_MEM_RECLAIM.
Fixes: ffd541d457 ("RDMA/hns: Add the workqueue framework for flush cqe handler")
Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20260104064057.1582216-2-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Some HCAs (e.g: ConnectX4) do not trigger a IB_EVENT_GID_CHANGE on
subnet prefix update from SM (PortInfo).
Since the commit d58c23c925 ("IB/core: Only update PKEY and GID caches
on respective events"), the GID cache is updated exclusively on
IB_EVENT_GID_CHANGE. If this event is not emitted, the subnet prefix in the
IPoIB interface’s hardware address remains set to its default value
(0xfe80000000000000).
Then rdma_bind_addr() failed because it relies on hardware address to
find the port GID (subnet_prefix + port GUID).
This patch fixes this issue by updating the GID cache on
IB_EVENT_CLIENT_REREGISTER event (emitted on PortInfo::ClientReregister=1).
Fixes: d58c23c925 ("IB/core: Only update PKEY and GID caches on respective events")
Signed-off-by: Etienne AUJAMES <eaujames@ddn.com>
Link: https://patch.msgid.link/aVUfsO58QIDn5bGX@eaujamesFR0130
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
In limit_bank mode, QPs/CQs are restricted to using half of the banks.
HW concentrates resources on these banks, thereby improving performance
compared to the default mode.
Switch between limit_bank mode and default mode by setting the cap
flag in FW. Since the number of QPs and CQs will be halved, this mode
is suitable for scenarios where fewer QPs and CQs are required.
Signed-off-by: Lianfa Weng <wenglianfa@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20251230154911.3397584-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
