The cited commit fixed a crash when cma_netevent_callback was called for
a cma_id while work on that id from a previous call had not yet started.
The work item was re-initialized in the second call, which corrupted the
work item currently in the work queue.
However, it left a problem when queue_work fails (because the item is
still pending in the work queue from a previous call). In this case,
cma_id_put (which is called in the work handler) is therefore not
called. This results in a userspace process hang (zombie process).
Fix this by calling cma_id_put() if queue_work fails.
Fixes: 45f5dcdd04 ("RDMA/cma: Fix workqueue crash in cma_netevent_work_handler")
Link: https://patch.msgid.link/r/4f3640b501e48d0166f312a64fdadf72b059bd04.1747827103.git.leon@kernel.org
Signed-off-by: Jack Morgenstein <jackm@nvidia.com>
Signed-off-by: Feng Liu <feliu@nvidia.com>
Reviewed-by: Vlad Dumitrescu <vdumitrescu@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Reviewed-by: Sharath Srinivasan <sharath.srinivasan@oracle.com>
Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The commit 59c68ac31e ("iw_cm: free cm_id resources on the last
deref") simplified cm_id resource management by freeing cm_id once all
references to the cm_id were removed. The references are removed either
upon completion of iw_cm event handlers or when the application destroys
the cm_id. This commit introduced the use-after-free condition where
cm_id_private object could still be in use by event handler works during
the destruction of cm_id. The commit aee2424246 ("RDMA/iwcm: Fix a
use-after-free related to destroying CM IDs") addressed this use-after-
free by flushing all pending works at the cm_id destruction.
However, still another use-after-free possibility remained. It happens
with the work objects allocated for each cm_id_priv within
alloc_work_entries() during cm_id creation, and subsequently freed in
dealloc_work_entries() once all references to the cm_id are removed.
If the cm_id's last reference is decremented in the event handler work,
the work object for the work itself gets removed, and causes the use-
after-free BUG below:
BUG: KASAN: slab-use-after-free in __pwq_activate_work+0x1ff/0x250
Read of size 8 at addr ffff88811f9cf800 by task kworker/u16:1/147091
CPU: 2 UID: 0 PID: 147091 Comm: kworker/u16:1 Not tainted 6.15.0-rc2+ #27 PREEMPT(voluntary)
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.3-3.fc41 04/01/2014
Workqueue: 0x0 (iw_cm_wq)
Call Trace:
<TASK>
dump_stack_lvl+0x6a/0x90
print_report+0x174/0x554
? __virt_addr_valid+0x208/0x430
? __pwq_activate_work+0x1ff/0x250
kasan_report+0xae/0x170
? __pwq_activate_work+0x1ff/0x250
__pwq_activate_work+0x1ff/0x250
pwq_dec_nr_in_flight+0x8c5/0xfb0
process_one_work+0xc11/0x1460
? __pfx_process_one_work+0x10/0x10
? assign_work+0x16c/0x240
worker_thread+0x5ef/0xfd0
? __pfx_worker_thread+0x10/0x10
kthread+0x3b0/0x770
? __pfx_kthread+0x10/0x10
? rcu_is_watching+0x11/0xb0
? _raw_spin_unlock_irq+0x24/0x50
? rcu_is_watching+0x11/0xb0
? __pfx_kthread+0x10/0x10
ret_from_fork+0x30/0x70
? __pfx_kthread+0x10/0x10
ret_from_fork_asm+0x1a/0x30
</TASK>
Allocated by task 147416:
kasan_save_stack+0x2c/0x50
kasan_save_track+0x10/0x30
__kasan_kmalloc+0xa6/0xb0
alloc_work_entries+0xa9/0x260 [iw_cm]
iw_cm_connect+0x23/0x4a0 [iw_cm]
rdma_connect_locked+0xbfd/0x1920 [rdma_cm]
nvme_rdma_cm_handler+0x8e5/0x1b60 [nvme_rdma]
cma_cm_event_handler+0xae/0x320 [rdma_cm]
cma_work_handler+0x106/0x1b0 [rdma_cm]
process_one_work+0x84f/0x1460
worker_thread+0x5ef/0xfd0
kthread+0x3b0/0x770
ret_from_fork+0x30/0x70
ret_from_fork_asm+0x1a/0x30
Freed by task 147091:
kasan_save_stack+0x2c/0x50
kasan_save_track+0x10/0x30
kasan_save_free_info+0x37/0x60
__kasan_slab_free+0x4b/0x70
kfree+0x13a/0x4b0
dealloc_work_entries+0x125/0x1f0 [iw_cm]
iwcm_deref_id+0x6f/0xa0 [iw_cm]
cm_work_handler+0x136/0x1ba0 [iw_cm]
process_one_work+0x84f/0x1460
worker_thread+0x5ef/0xfd0
kthread+0x3b0/0x770
ret_from_fork+0x30/0x70
ret_from_fork_asm+0x1a/0x30
Last potentially related work creation:
kasan_save_stack+0x2c/0x50
kasan_record_aux_stack+0xa3/0xb0
__queue_work+0x2ff/0x1390
queue_work_on+0x67/0xc0
cm_event_handler+0x46a/0x820 [iw_cm]
siw_cm_upcall+0x330/0x650 [siw]
siw_cm_work_handler+0x6b9/0x2b20 [siw]
process_one_work+0x84f/0x1460
worker_thread+0x5ef/0xfd0
kthread+0x3b0/0x770
ret_from_fork+0x30/0x70
ret_from_fork_asm+0x1a/0x30
This BUG is reproducible by repeating the blktests test case nvme/061
for the rdma transport and the siw driver.
To avoid the use-after-free of cm_id_private work objects, ensure that
the last reference to the cm_id is decremented not in the event handler
works, but in the cm_id destruction context. For that purpose, move
iwcm_deref_id() call from destroy_cm_id() to the callers of
destroy_cm_id(). In iw_destroy_cm_id(), call iwcm_deref_id() after
flushing the pending works.
During the fix work, I noticed that iw_destroy_cm_id() is called from
cm_work_handler() and process_event() context. However, the comment of
iw_destroy_cm_id() notes that the function "cannot be called by the
event thread". Drop the false comment.
Closes: https://lore.kernel.org/linux-rdma/r5676e754sv35aq7cdsqrlnvyhiq5zktteaurl7vmfih35efko@z6lay7uypy3c/
Fixes: 59c68ac31e ("iw_cm: free cm_id resources on the last deref")
Cc: stable@vger.kernel.org
Signed-off-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com>
Link: https://patch.msgid.link/20250510101036.1756439-1-shinichiro.kawasaki@wdc.com
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Create separate functions for the implicit ODP initialization
which is different from the explicit ODP initialization.
Tested-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Reuse newly added DMA API to cache IOVA and only link/unlink pages
in fast path for UMEM ODP flow.
Tested-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
As a preparation to remove dma_list, store access mask in PFN pointer
and not in dma_addr_t.
Tested-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Call Trace:
__dump_stack lib/dump_stack.c:94 [inline]
dump_stack_lvl+0x116/0x1f0 lib/dump_stack.c:120
print_address_description mm/kasan/report.c:408 [inline]
print_report+0xc3/0x670 mm/kasan/report.c:521
kasan_report+0xe0/0x110 mm/kasan/report.c:634
strlen+0x93/0xa0 lib/string.c:420
__fortify_strlen include/linux/fortify-string.h:268 [inline]
get_kobj_path_length lib/kobject.c:118 [inline]
kobject_get_path+0x3f/0x2a0 lib/kobject.c:158
kobject_uevent_env+0x289/0x1870 lib/kobject_uevent.c:545
ib_register_device drivers/infiniband/core/device.c:1472 [inline]
ib_register_device+0x8cf/0xe00 drivers/infiniband/core/device.c:1393
rxe_register_device+0x275/0x320 drivers/infiniband/sw/rxe/rxe_verbs.c:1552
rxe_net_add+0x8e/0xe0 drivers/infiniband/sw/rxe/rxe_net.c:550
rxe_newlink+0x70/0x190 drivers/infiniband/sw/rxe/rxe.c:225
nldev_newlink+0x3a3/0x680 drivers/infiniband/core/nldev.c:1796
rdma_nl_rcv_msg+0x387/0x6e0 drivers/infiniband/core/netlink.c:195
rdma_nl_rcv_skb.constprop.0.isra.0+0x2e5/0x450
netlink_unicast_kernel net/netlink/af_netlink.c:1313 [inline]
netlink_unicast+0x53a/0x7f0 net/netlink/af_netlink.c:1339
netlink_sendmsg+0x8d1/0xdd0 net/netlink/af_netlink.c:1883
sock_sendmsg_nosec net/socket.c:712 [inline]
__sock_sendmsg net/socket.c:727 [inline]
____sys_sendmsg+0xa95/0xc70 net/socket.c:2566
___sys_sendmsg+0x134/0x1d0 net/socket.c:2620
__sys_sendmsg+0x16d/0x220 net/socket.c:2652
do_syscall_x64 arch/x86/entry/syscall_64.c:63 [inline]
do_syscall_64+0xcd/0x260 arch/x86/entry/syscall_64.c:94
entry_SYSCALL_64_after_hwframe+0x77/0x7f
This problem is similar to the problem that the
commit 1d6a9e7449 ("RDMA/core: Fix use-after-free when rename device name")
fixes.
The root cause is: the function ib_device_rename() renames the name with
lock. But in the function kobject_uevent(), this name is accessed without
lock protection at the same time.
The solution is to add the lock protection when this name is accessed in
the function kobject_uevent().
Fixes: 779e0bf476 ("RDMA/core: Do not indicate device ready when device enablement fails")
Link: https://patch.msgid.link/r/20250506151008.75701-1-yanjun.zhu@linux.dev
Reported-by: syzbot+e2ce9e275ecc70a30b72@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=e2ce9e275ecc70a30b72
Signed-off-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
struct rdma_cm_id has member "struct work_struct net_work"
that is reused for enqueuing cma_netevent_work_handler()s
onto cma_wq.
Below crash[1] can occur if more than one call to
cma_netevent_callback() occurs in quick succession,
which further enqueues cma_netevent_work_handler()s for the
same rdma_cm_id, overwriting any previously queued work-item(s)
that was just scheduled to run i.e. there is no guarantee
the queued work item may run between two successive calls
to cma_netevent_callback() and the 2nd INIT_WORK would overwrite
the 1st work item (for the same rdma_cm_id), despite grabbing
id_table_lock during enqueue.
Also drgn analysis [2] indicates the work item was likely overwritten.
Fix this by moving the INIT_WORK() to __rdma_create_id(),
so that it doesn't race with any existing queue_work() or
its worker thread.
[1] Trimmed crash stack:
=============================================
BUG: kernel NULL pointer dereference, address: 0000000000000008
kworker/u256:6 ... 6.12.0-0...
Workqueue: cma_netevent_work_handler [rdma_cm] (rdma_cm)
RIP: 0010:process_one_work+0xba/0x31a
Call Trace:
worker_thread+0x266/0x3a0
kthread+0xcf/0x100
ret_from_fork+0x31/0x50
ret_from_fork_asm+0x1a/0x30
=============================================
[2] drgn crash analysis:
>>> trace = prog.crashed_thread().stack_trace()
>>> trace
(0) crash_setup_regs (./arch/x86/include/asm/kexec.h:111:15)
(1) __crash_kexec (kernel/crash_core.c:122:4)
(2) panic (kernel/panic.c:399:3)
(3) oops_end (arch/x86/kernel/dumpstack.c:382:3)
...
(8) process_one_work (kernel/workqueue.c:3168:2)
(9) process_scheduled_works (kernel/workqueue.c:3310:3)
(10) worker_thread (kernel/workqueue.c:3391:4)
(11) kthread (kernel/kthread.c:389:9)
Line workqueue.c:3168 for this kernel version is in process_one_work():
3168 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
>>> trace[8]["work"]
*(struct work_struct *)0xffff92577d0a21d8 = {
.data = (atomic_long_t){
.counter = (s64)536870912, <=== Note
},
.entry = (struct list_head){
.next = (struct list_head *)0xffff924d075924c0,
.prev = (struct list_head *)0xffff924d075924c0,
},
.func = (work_func_t)cma_netevent_work_handler+0x0 = 0xffffffffc2cec280,
}
Suspicion is that pwq is NULL:
>>> trace[8]["pwq"]
(struct pool_workqueue *)<absent>
In process_one_work(), pwq is assigned from:
struct pool_workqueue *pwq = get_work_pwq(work);
and get_work_pwq() is:
static struct pool_workqueue *get_work_pwq(struct work_struct *work)
{
unsigned long data = atomic_long_read(&work->data);
if (data & WORK_STRUCT_PWQ)
return work_struct_pwq(data);
else
return NULL;
}
WORK_STRUCT_PWQ is 0x4:
>>> print(repr(prog['WORK_STRUCT_PWQ']))
Object(prog, 'enum work_flags', value=4)
But work->data is 536870912 which is 0x20000000.
So, get_work_pwq() returns NULL and we crash in process_one_work():
3168 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
=============================================
Fixes: 925d046e7e ("RDMA/core: Add a netevent notifier to cma")
Cc: stable@vger.kernel.org
Co-developed-by: Håkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: Sharath Srinivasan <sharath.srinivasan@oracle.com>
Reviewed-by: Patrisious Haddad <phaddad@nvidia.com>
Link: https://patch.msgid.link/bf0082f9-5b25-4593-92c6-d130aa8ba439@oracle.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
In workloads where there are many processes establishing connections using
RDMA CM in parallel (large scale MPI), there can be heavy contention for
mad_agent_lock in cm_alloc_msg.
This contention can occur while inside of a spin_lock_irq region, leading
to interrupts being disabled for extended durations on many
cores. Furthermore, it leads to the serialization of rdma_create_ah calls,
which has negative performance impacts for NICs which are capable of
processing multiple address handle creations in parallel.
The end result is the machine becoming unresponsive, hung task warnings,
netdev TX timeouts, etc.
Since the lock appears to be only for protection from cm_remove_one, it
can be changed to a rwlock to resolve these issues.
Reproducer:
Server:
for i in $(seq 1 512); do
ucmatose -c 32 -p $((i + 5000)) &
done
Client:
for i in $(seq 1 512); do
ucmatose -c 32 -p $((i + 5000)) -s 10.2.0.52 &
done
Fixes: 76039ac909 ("IB/cm: Protect cm_dev, cm_ports and mad_agent with kref and lock")
Link: https://patch.msgid.link/r/20250220175612.2763122-1-jmoroni@google.com
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Passing a non-literal format string to dev_set_name causes a warning:
drivers/infiniband/core/ucaps.c:173:33: error: format string is not a string literal (potentially insecure) [-Werror,-Wformat-security]
173 | ret = dev_set_name(&ucap->dev, ucap_names[type]);
| ^~~~~~~~~~~~~~~~
drivers/infiniband/core/ucaps.c:173:33: note: treat the string as an argument to avoid this
173 | ret = dev_set_name(&ucap->dev, ucap_names[type]);
| ^
| "%s",
Turn the name into the %s argument as suggested by gcc.
Fixes: 61e5168281 ("RDMA/uverbs: Introduce UCAP (User CAPabilities) API")
Link: https://patch.msgid.link/r/20250314155721.264083-1-arnd@kernel.org
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The ib_post_receive_mads() function handles posting receive work
requests (WRs) to MAD QPs and is called in two cases:
1) When a MAD port is opened.
2) When a receive WQE is consumed upon receiving a new MAD.
Whereas, if MADs arrive during the port open phase, a race condition
might cause an extra WR to be posted, exceeding the QP’s capacity.
This leads to failures such as:
infiniband mlx5_0: ib_post_recv failed: -12
infiniband mlx5_0: Couldn't post receive WRs
infiniband mlx5_0: Couldn't start port
infiniband mlx5_0: Couldn't open port 1
Fix this by checking the current receive count before posting a new WR.
If the QP’s receive queue is full, do not post additional WRs.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Maher Sanalla <msanalla@nvidia.com>
Link: https://patch.msgid.link/c4984ba3c3a98a5711a558bccefcad789587ecf1.1741875592.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Whenever a new counter is created, save inside it the user requested
configuration for optional-counters binding, for manual configuration it
is requested directly by the user and for the automatic configuration it
depends on if the automatic binding was enabled with or without
optional-counters binding.
This argument will later be used by the driver to determine if to bind the
optional-counters as well or not when trying to bind this counter to a QP.
It indicates that when binding counters to a QP we also want the
currently enabled link optional-counters to be bound as well.
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/82f1c357606a16932979ef9a5910122675c74a3a.1741875070.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Change rdma_counter allocation to use rdma_zalloc_drv_obj() instead of,
explicitly allocating at core, in order to be contained inside driver
specific structures.
Adjust all drivers that use it to have their containing structure, and
add driver specific initialization operation.
This change is needed to allow upcoming patches to implement
optional-counters binding whereas inside each driver specific counter
struct his bound optional-counters will be maintained.
Signed-off-by: Patrisious Haddad <phaddad@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Link: https://patch.msgid.link/a5a484f421fc2e5595158e61a354fba43272b02d.1741875070.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Currently, the IB uverbs API calls uobj_get_uobj_read(), which in turn
uses the rdma_lookup_get_uobject() helper to retrieve user objects.
In case of failure, uobj_get_uobj_read() returns NULL, overriding the
error code from rdma_lookup_get_uobject(). The IB uverbs API then
translates this NULL to -EINVAL, masking the actual error and
complicating debugging. For example, applications calling ibv_modify_qp
that fails with EBUSY when retrieving the QP uobject will see the
overridden error code EINVAL instead, masking the actual error.
Furthermore, based on rdma-core commit:
"2a22f1ced5f3 ("Merge pull request #1568 from jakemoroni/master")"
Kernel's IB uverbs return values are either ignored and passed on as is
to application or overridden with other errnos in a few cases.
Thus, to improve error reporting and debuggability, propagate the
original error from rdma_lookup_get_uobject() instead of replacing it
with EINVAL.
Signed-off-by: Maher Sanalla <msanalla@nvidia.com>
Link: https://patch.msgid.link/64f9d3711b183984e939962c2f83383904f97dfb.1740577869.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Implement a new User CAPabilities (UCAP) API to provide fine-grained
control over specific firmware features.
This approach offers more granular capabilities than the existing Linux
capabilities, which may be too generic for certain FW features.
This mechanism represents each capability as a character device with
root read-write access. Root processes can grant users special
privileges by allowing access to these character devices (e.g., using
chown).
UCAP character devices are located in /dev/infiniband and the class path
is /sys/class/infiniband_ucaps.
Signed-off-by: Chiara Meiohas <cmeiohas@nvidia.com>
Link: https://patch.msgid.link/5a1379187cd21178e8554afc81a3c941f21af22f.1741261611.git.leon@kernel.org
Reviewed-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Commit 467f432a52 ("RDMA/core: Split port and device counter sysfs
attributes") accidentally almost exposed hw counters to non-init net
namespaces. It didn't expose them fully, as an attempt to read any of
those counters leads to a crash like this one:
[42021.807566] BUG: kernel NULL pointer dereference, address: 0000000000000028
[42021.814463] #PF: supervisor read access in kernel mode
[42021.819549] #PF: error_code(0x0000) - not-present page
[42021.824636] PGD 0 P4D 0
[42021.827145] Oops: 0000 [#1] SMP PTI
[42021.830598] CPU: 82 PID: 2843922 Comm: switchto-defaul Kdump: loaded Tainted: G S W I XXX
[42021.841697] Hardware name: XXX
[42021.849619] RIP: 0010:hw_stat_device_show+0x1e/0x40 [ib_core]
[42021.855362] Code: 90 90 90 90 90 90 90 90 90 90 90 90 f3 0f 1e fa 0f 1f 44 00 00 49 89 d0 4c 8b 5e 20 48 8b 8f b8 04 00 00 48 81 c7 f0 fa ff ff <48> 8b 41 28 48 29 ce 48 83 c6 d0 48 c1 ee 04 69 d6 ab aa aa aa 48
[42021.873931] RSP: 0018:ffff97fe90f03da0 EFLAGS: 00010287
[42021.879108] RAX: ffff9406988a8c60 RBX: ffff940e1072d438 RCX: 0000000000000000
[42021.886169] RDX: ffff94085f1aa000 RSI: ffff93c6cbbdbcb0 RDI: ffff940c7517aef0
[42021.893230] RBP: ffff97fe90f03e70 R08: ffff94085f1aa000 R09: 0000000000000000
[42021.900294] R10: ffff94085f1aa000 R11: ffffffffc0775680 R12: ffffffff87ca2530
[42021.907355] R13: ffff940651602840 R14: ffff93c6cbbdbcb0 R15: ffff94085f1aa000
[42021.914418] FS: 00007fda1a3b9700(0000) GS:ffff94453fb80000(0000) knlGS:0000000000000000
[42021.922423] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[42021.928130] CR2: 0000000000000028 CR3: 00000042dcfb8003 CR4: 00000000003726f0
[42021.935194] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[42021.942257] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[42021.949324] Call Trace:
[42021.951756] <TASK>
[42021.953842] [<ffffffff86c58674>] ? show_regs+0x64/0x70
[42021.959030] [<ffffffff86c58468>] ? __die+0x78/0xc0
[42021.963874] [<ffffffff86c9ef75>] ? page_fault_oops+0x2b5/0x3b0
[42021.969749] [<ffffffff87674b92>] ? exc_page_fault+0x1a2/0x3c0
[42021.975549] [<ffffffff87801326>] ? asm_exc_page_fault+0x26/0x30
[42021.981517] [<ffffffffc0775680>] ? __pfx_show_hw_stats+0x10/0x10 [ib_core]
[42021.988482] [<ffffffffc077564e>] ? hw_stat_device_show+0x1e/0x40 [ib_core]
[42021.995438] [<ffffffff86ac7f8e>] dev_attr_show+0x1e/0x50
[42022.000803] [<ffffffff86a3eeb1>] sysfs_kf_seq_show+0x81/0xe0
[42022.006508] [<ffffffff86a11134>] seq_read_iter+0xf4/0x410
[42022.011954] [<ffffffff869f4b2e>] vfs_read+0x16e/0x2f0
[42022.017058] [<ffffffff869f50ee>] ksys_read+0x6e/0xe0
[42022.022073] [<ffffffff8766f1ca>] do_syscall_64+0x6a/0xa0
[42022.027441] [<ffffffff8780013b>] entry_SYSCALL_64_after_hwframe+0x78/0xe2
The problem can be reproduced using the following steps:
ip netns add foo
ip netns exec foo bash
cat /sys/class/infiniband/mlx4_0/hw_counters/*
The panic occurs because of casting the device pointer into an
ib_device pointer using container_of() in hw_stat_device_show() is
wrong and leads to a memory corruption.
However the real problem is that hw counters should never been exposed
outside of the non-init net namespace.
Fix this by saving the index of the corresponding attribute group
(it might be 1 or 2 depending on the presence of driver-specific
attributes) and zeroing the pointer to hw_counters group for compat
devices during the initialization.
With this fix applied hw_counters are not available in a non-init
net namespace:
find /sys/class/infiniband/mlx4_0/ -name hw_counters
/sys/class/infiniband/mlx4_0/ports/1/hw_counters
/sys/class/infiniband/mlx4_0/ports/2/hw_counters
/sys/class/infiniband/mlx4_0/hw_counters
ip netns add foo
ip netns exec foo bash
find /sys/class/infiniband/mlx4_0/ -name hw_counters
Fixes: 467f432a52 ("RDMA/core: Split port and device counter sysfs attributes")
Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Leon Romanovsky <leon@kernel.org>
Cc: Maher Sanalla <msanalla@nvidia.com>
Cc: linux-rdma@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Link: https://patch.msgid.link/20250227165420.3430301-1-roman.gushchin@linux.dev
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
A single scatter-gather entry is limited by a 32 bits "length" field
that is practically 4GB - PAGE_SIZE. This means that even when the
memory is physically contiguous, we might need more than one entry to
represent it. Additionally when using dmabuf, the sg_table might be
originated outside the subsystem and optimized for other needs.
For instance an SGT of 16GB GPU continuous memory might look like this:
(a real life example)
dma_address 34401400000, length fffff000
dma_address 345013ff000, length fffff000
dma_address 346013fe000, length fffff000
dma_address 347013fd000, length fffff000
dma_address 348013fc000, length 4000
Since ib_umem_find_best_pgsz works within SG entries, in the above case
we will result with the worst possible 4KB page size.
Fix this by taking into consideration only the alignment of addresses of
real discontinuity points rather than treating SG entries as such, and
adjust the page iterator to correctly handle cross SG entry pages.
There is currently an assumption that drivers do not ask for pages
bigger than maximal DMA size supported by their devices.
Reviewed-by: Firas Jahjah <firasj@amazon.com>
Reviewed-by: Yonatan Nachum <ynachum@amazon.com>
Signed-off-by: Michael Margolin <mrgolin@amazon.com>
Link: https://patch.msgid.link/20250217141623.12428-1-mrgolin@amazon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Pull rdma updates from Jason Gunthorpe:
"Lighter that normal, but the now usual collection of driver fixes and
small improvements:
- Small fixes and minor improvements to cxgb4, bnxt_re, rxe, srp,
efa, cxgb4
- Update mlx4 to use the new umem APIs, avoiding direct use of
scatterlist
- Support ROCEv2 in erdma
- Remove various uncalled functions, constify bin_attribute
- Provide core infrastructure to catch netdev events and route them
to drivers, consolidating duplicated driver code
- Fix rare race condition crashes in mlx5 ODP flows"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (63 commits)
RDMA/mlx5: Fix implicit ODP use after free
RDMA/mlx5: Fix a race for an ODP MR which leads to CQE with error
RDMA/qib: Constify 'struct bin_attribute'
RDMA/hfi1: Constify 'struct bin_attribute'
RDMA/rxe: Fix the warning "__rxe_cleanup+0x12c/0x170 [rdma_rxe]"
RDMA/cxgb4: Notify rdma stack for IB_EVENT_QP_LAST_WQE_REACHED event
RDMA/bnxt_re: Allocate dev_attr information dynamically
RDMA/bnxt_re: Pass the context for ulp_irq_stop
RDMA/bnxt_re: Add support to handle DCB_CONFIG_CHANGE event
RDMA/bnxt_re: Query firmware defaults of CC params during probe
RDMA/bnxt_re: Add Async event handling support
bnxt_en: Add ULP call to notify async events
RDMA/mlx5: Fix indirect mkey ODP page count
MAINTAINERS: Update the bnxt_re maintainers
RDMA/hns: Clean up the legacy CONFIG_INFINIBAND_HNS
RDMA/rtrs: Add missing deinit() call
RDMA/efa: Align interrupt related fields to same type
RDMA/bnxt_re: Fix to drop reference to the mmap entry in case of error
RDMA/mlx5: Fix link status down event for MPV
RDMA/erdma: Support create_ah/destroy_ah in non-sleepable contexts
...
Pull rdma fixes from Jason Gunthorpe:
"A lot of fixes accumulated over the holiday break:
- Static tool fixes, value is already proven to be NULL, possible
integer overflow
- Many bnxt_re fixes:
- Crashes due to a mismatch in the maximum SGE list size
- Don't waste memory for user QPs by creating kernel-only
structures
- Fix compatability issues with older HW in some of the new HW
features recently introduced: RTS->RTS feature, work around 9096
- Do not allow destroy_qp to fail
- Validate QP MTU against device limits
- Add missing validation on madatory QP attributes for RTR->RTS
- Report port_num in query_qp as required by the spec
- Fix creation of QPs of the maximum queue size, and in the
variable mode
- Allow all QPs to be used on newer HW by limiting a work around
only to HW it affects
- Use the correct MSN table size for variable mode QPs
- Add missing locking in create_qp() accessing the qp_tbl
- Form WQE buffers correctly when some of the buffers are 0 hop
- Don't crash on QP destroy if the userspace doesn't setup the
dip_ctx
- Add the missing QP flush handler call on the DWQE path to avoid
hanging on error recovery
- Consistently use ENXIO for return codes if the devices is
fatally errored
- Try again to fix VLAN support on iwarp, previous fix was reverted
due to breaking other cards
- Correct error path return code for rdma netlink events
- Remove the seperate net_device pointer in siw and rxe which
syzkaller found a way to UAF
- Fix a UAF of a stack ib_sge in rtrs
- Fix a regression where old mlx5 devices and FW were wrongly
activing new device features and failing"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (28 commits)
RDMA/mlx5: Enable multiplane mode only when it is supported
RDMA/bnxt_re: Fix error recovery sequence
RDMA/rtrs: Ensure 'ib_sge list' is accessible
RDMA/rxe: Remove the direct link to net_device
RDMA/hns: Fix missing flush CQE for DWQE
RDMA/hns: Fix warning storm caused by invalid input in IO path
RDMA/hns: Fix accessing invalid dip_ctx during destroying QP
RDMA/hns: Fix mapping error of zero-hop WQE buffer
RDMA/bnxt_re: Fix the locking while accessing the QP table
RDMA/bnxt_re: Fix MSN table size for variable wqe mode
RDMA/bnxt_re: Add send queue size check for variable wqe
RDMA/bnxt_re: Disable use of reserved wqes
RDMA/bnxt_re: Fix max_qp_wrs reported
RDMA/siw: Remove direct link to net_device
RDMA/nldev: Set error code in rdma_nl_notify_event
RDMA/bnxt_re: Fix reporting hw_ver in query_device
RDMA/bnxt_re: Fix to export port num to ib_query_qp
RDMA/bnxt_re: Fix setting mandatory attributes for modify_qp
RDMA/bnxt_re: Add check for path mtu in modify_qp
RDMA/bnxt_re: Fix the check for 9060 condition
...
Currently the dispatching of link status events is implemented by
each RDMA driver independently, and most of them have very similar
patterns. Add support for this in ib_core so that we can get rid
of duplicate codes in each driver.
A new last_port_state is added in ib_port_cache to cache the port
state of the last link status events dispatching. The original
port_state in ib_port_cache is not used here because it will be
updated when ib_dispatch_event() is called, which means it may
be changed between two link status events, and may lead to a loss
of event dispatching.
Some drivers currently have some private stuff in their link status
events handler in addition to event dispatching, and cannot be
perfectly integrated into the ib_core handling process. For these
drivers, add a new ops report_port_event() so that they can keep
their current processing.
Finally, events of LAG devices are not supported yet in this patch
as currently there is no way to obtain ibdev from upper netdev in
ib_core. This can be a TODO work after the core have more support
for LAG.
Signed-off-by: Yuyu Li <liyuyu6@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
If traffic is over vlan, cma_validate_port() fails to match
net_device ifindex with bound_if_index and results in ENODEV error.
As iWARP gid table is static, it contains entry corresponding to
only one net device which is either real netdev or vlan netdev for
cases like siw attached to a vlan interface.
This patch fixes the issue by assigning bound_if_index with net
device index, if real net device obtained from bound if index matches
with net device retrieved from gid table
Fixes: f8ef1be816 ("RDMA/cma: Avoid GID lookups on iWARP devices")
Link: https://lore.kernel.org/all/ZzNgdrjo1kSCGbRz@chelsio.com/
Signed-off-by: Anumula Murali Mohan Reddy <anumula@chelsio.com>
Signed-off-by: Potnuri Bharat Teja <bharat@chelsio.com>
Link: https://patch.msgid.link/20241203140052.3985-1-anumula@chelsio.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
In the expression "cmd.wqe_size * cmd.wr_count", both variables are u32
values that come from the user so the multiplication can lead to integer
wrapping. Then we pass the result to uverbs_request_next_ptr() which also
could potentially wrap. The "cmd.sge_count * sizeof(struct ib_uverbs_sge)"
multiplication can also overflow on 32bit systems although it's fine on
64bit systems.
This patch does two things. First, I've re-arranged the condition in
uverbs_request_next_ptr() so that the use controlled variable "len" is on
one side of the comparison by itself without any math. Then I've modified
all the callers to use size_mul() for the multiplications.
Fixes: 67cdb40ca4 ("[IB] uverbs: Implement more commands")
Cc: stable@vger.kernel.org
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Link: https://patch.msgid.link/b8765ab3-c2da-4611-aae0-ddd6ba173d23@stanley.mountain
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Pull rdma updates from Jason Gunthorpe:
"Seveal fixes scattered across the drivers and a few new features:
- Minor updates and bug fixes to hfi1, efa, iopob, bnxt, hns
- Force disassociate the userspace FD when hns does an async reset
- bnxt new features for optimized modify QP to skip certain stayes,
CQ coalescing, better debug dumping
- mlx5 new data placement ordering feature
- Faster destruction of mlx5 devx HW objects
- Improvements to RDMA CM mad handling"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (51 commits)
RDMA/bnxt_re: Correct the sequence of device suspend
RDMA/bnxt_re: Use the default mode of congestion control
RDMA/bnxt_re: Support different traffic class
IB/cm: Rework sending DREQ when destroying a cm_id
IB/cm: Do not hold reference on cm_id unless needed
IB/cm: Explicitly mark if a response MAD is a retransmission
RDMA/mlx5: Move events notifier registration to be after device registration
RDMA/bnxt_re: Cache MSIx info to a local structure
RDMA/bnxt_re: Refurbish CQ to NQ hash calculation
RDMA/bnxt_re: Refactor NQ allocation
RDMA/bnxt_re: Fail probe early when not enough MSI-x vectors are reserved
RDMA/hns: Fix different dgids mapping to the same dip_idx
RDMA/bnxt_re: Add set_func_resources support for P5/P7 adapters
RDMA/bnxt_re: Enhance RoCE SRIOV resource configuration design
bnxt_en: Add support for RoCE sriov configuration
RDMA/hns: Fix NULL pointer derefernce in hns_roce_map_mr_sg()
RDMA/hns: Fix out-of-order issue of requester when setting FENCE
RDMA/nldev: Add IB device and net device rename events
RDMA/mlx5: Add implementation for ufile_hw_cleanup device operation
RDMA/core: Move ib_uverbs_file struct to uverbs_types.h
...
Pull 'struct fd' class updates from Al Viro:
"The bulk of struct fd memory safety stuff
Making sure that struct fd instances are destroyed in the same scope
where they'd been created, getting rid of reassignments and passing
them by reference, converting to CLASS(fd{,_pos,_raw}).
We are getting very close to having the memory safety of that stuff
trivial to verify"
* tag 'pull-fd' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (28 commits)
deal with the last remaing boolean uses of fd_file()
css_set_fork(): switch to CLASS(fd_raw, ...)
memcg_write_event_control(): switch to CLASS(fd)
assorted variants of irqfd setup: convert to CLASS(fd)
do_pollfd(): convert to CLASS(fd)
convert do_select()
convert vfs_dedupe_file_range().
convert cifs_ioctl_copychunk()
convert media_request_get_by_fd()
convert spu_run(2)
switch spufs_calls_{get,put}() to CLASS() use
convert cachestat(2)
convert do_preadv()/do_pwritev()
fdget(), more trivial conversions
fdget(), trivial conversions
privcmd_ioeventfd_assign(): don't open-code eventfd_ctx_fdget()
o2hb_region_dev_store(): avoid goto around fdget()/fdput()
introduce "fd_pos" class, convert fdget_pos() users to it.
fdget_raw() users: switch to CLASS(fd_raw)
convert vmsplice() to CLASS(fd)
...
A DREQ is sent in 2 situations:
1. When requested by the user.
This DREQ has to wait for a DREP, which will be routed to the user.
2. When the cm_id is destroyed.
This DREQ is generated by the CM to notify the peer that the
connection has been destroyed.
In the latter case, any DREP that is received will be discarded.
There's no need to hold a reference on the cm_id. Today, both
situations are covered by the same function: cm_send_dreq_locked().
When invoked in the cm_id destroy path, the cm_id reference would be
held until the DREQ completes, blocking the destruction. Because it
could take several seconds to minutes before the DREQ receives a DREP,
the destroy call posts a send for the DREQ then immediately cancels the
MAD. However, cancellation is not immediate in the MAD layer. There
could still be a delay before the MAD layer returns the DREQ to the CM.
Moreover, the only guarantee is that the DREQ will be sent at most once.
Introduce a separate flow for sending a DREQ when destroying the cm_id.
The new flow will not hold a reference on the cm_id, allowing it to be
cleaned up immediately. The cancellation trick is no longer needed.
The MAD layer will send the DREQ exactly once.
Signed-off-by: Sean Hefty <shefty@nvidia.com>
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Signed-off-by: Vlad Dumitrescu <vdumitrescu@nvidia.com>
Link: https://patch.msgid.link/a288a098b8e0550305755fd4a7937431699317f4.1731495873.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Typically, when the CM sends a MAD it bumps a reference count
on the associated cm_id. There are some exceptions, such
as when the MAD is a direct response to a receive MAD. For
example, the CM may generate an MRA in response to a duplicate
REQ. But, in general, if a MAD may be sent as a result of
the user invoking an API call (e.g. ib_send_cm_rep(),
ib_send_cm_rtu(), etc.), a reference is taken on the cm_id.
This reference is necessary if the MAD requires a response.
The reference allows routing a response MAD back to the
cm_id, or, if no response is received, allows updating the
cm_id state to reflect the failure.
For MADs which do not generate a response from the
target, however, there's no need to hold a reference on the cm_id.
Such MADs will not be retried by the MAD layer and their
completions do not change the state of the cm_id.
There are 2 internal calls used to allocate MADs which take
a reference on the cm_id: cm_alloc_msg() and cm_alloc_priv_msg().
The latter calls the former. It turns out that all other places
where cm_alloc_msg() is called are for MADs that do not generate
a response from the target: sending an RTU, DREP, REJ, MRA, or
SIDR REP. In all of these cases, there's no need to hold a
reference on the cm_id.
The benefit of dropping unneeded references is that it allows
destruction of the cm_id to proceed immediately. Currently,
the cm_destroy_id() call blocks as long as there's a reference
held on the cm_id. Worse, is that cm_destroy_id() will send
MADs, which it then needs to complete. Sending the MADs is
beneficial, as they notify the peer that a connection is
being destroyed. However, since the MADs hold a reference
on the cm_id, they block destruction and cannot be retried.
Move cm_id referencing from cm_alloc_msg() to cm_alloc_priv_msg().
The latter should hold a reference on the cm_id in all cases but
one, which will be handled in a separate patch. cm_alloc_priv_msg()
is used when sending a REQ, REP, DREQ, and SIDR REQ, all of which
require a response.
Also, merge common code into cm_alloc_priv_msg() and combine the
freeing of all messages which do not need a response.
Signed-off-by: Sean Hefty <shefty@nvidia.com>
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Signed-off-by: Vlad Dumitrescu <vdumitrescu@nvidia.com>
Link: https://patch.msgid.link/1f0f96acace72790ecf89087fc765dead960189e.1731495873.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
In several situations the CM may send a reply to a received MAD
without the reply being directly linked with a cm_id. For
example, it may send a REJ in response to a REQ which does not
match a listener. Or, it may send a DREP in response to a DREQ
if the cm_id has already been destroyed. This can happen if the
original DREP was lost and the DREQ was retried.
When such a response MAD completes, it updates a counter tracking
how many MADs were retried. However, not all response MADs issued
directly by the CM may be retries. The REJ mentioned in the example
above is such a case. To distinguish between responses which were
retries versus those that are not, the send_handler performs the
following check: is a retry if the response is not associated with
a cm_id and the response is not a REJ message.
Replace this indirect method of checking if a response is a retry
with an explicit check. Note that these retries are generated
directly by the CM, rather than retried by the MAD layer.
This change will be needed by later changes which would otherwise
break the indirect check.
Signed-off-by: Sean Hefty <shefty@nvidia.com>
Signed-off-by: Or Har-Toov <ohartoov@nvidia.com>
Signed-off-by: Vlad Dumitrescu <vdumitrescu@nvidia.com>
Link: https://patch.msgid.link/1ee6e2a68f8de1992b9da23aa1d7e3f9f25e0036.1731495873.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
