RDMA/mlx5: Fix unsafe xarray access in implicit ODP handling
Subsystems (2)
PATCH 1/1 RDMA/mlx5: Fix unsafe xarray access in implicit ODP handling
From: Or Har-Toov <[email protected]>
__xa_store() and __xa_erase() were used without holding the proper lock, which led to a lockdep warning due to unsafe RCPU usage. This patch replaces them with xa_store() and xa_erase(), which perform the necessary locking internally.
=============================
WARNING: suspicious RCPU usage 6.14.0-rc7_for_upstream_debug_2025_03_18_15_01 #1 Not tainted
-----------------------------./include/linux/xarray.h:1211 suspicious rcu_dereference_protected() usage!
other info that might help us debug this:
rcu_scheduler_active = 2, debug_locks = 1
3 locks held by kworker/u136:0/219:
at: process_one_work+0xbe4/0x15f0
process_one_work+0x75c/0x15f0
pagefault_mr+0x9a5/0x1390 [mlx5_ib]stack backtrace: CPU: 14 UID: 0 PID: 219 Comm: kworker/u136:0 Not tainted 6.14.0-rc7_for_upstream_debug_2025_03_18_15_01 #1 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.0-0-gd239552ce722-prebuilt.qemu.org 04/01/2014 Workqueue: mlx5_ib_page_fault mlx5_ib_eqe_pf_action [mlx5_ib] Call Trace: dump_stack_lvl+0xa8/0xc0 lockdep_rcu_suspicious+0x1e6/0x260 xas_create+0xb8a/0xee0 xas_store+0x73/0x14c0 __xa_store+0x13c/0x220 ? xa_store_range+0x390/0x390 ? spin_bug+0x1d0/0x1d0 pagefault_mr+0xcb5/0x1390 [mlx5_ib] ? _raw_spin_unlock+0x1f/0x30 mlx5_ib_eqe_pf_action+0x3be/0x2620 [mlx5_ib] ? lockdep_hardirqs_on_prepare+0x400/0x400 ? mlx5_ib_invalidate_range+0xcb0/0xcb0 [mlx5_ib] process_one_work+0x7db/0x15f0 ? pwq_dec_nr_in_flight+0xda0/0xda0 ? assign_work+0x168/0x240 worker_thread+0x57d/0xcd0 ? rescuer_thread+0xc40/0xc40 kthread+0x3b3/0x800 ? kthread_is_per_cpu+0xb0/0xb0 ? lock_downgrade+0x680/0x680 ? do_raw_spin_lock+0x12d/0x270 ? spin_bug+0x1d0/0x1d0 ? finish_task_switch.isra.0+0x284/0x9e0 ? lockdep_hardirqs_on_prepare+0x284/0x400 ? kthread_is_per_cpu+0xb0/0xb0 ret_from_fork+0x2d/0x70 ? kthread_is_per_cpu+0xb0/0xb0 ret_from_fork_asm+0x11/0x20
--- drivers/infiniband/hw/mlx5/odp.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/infiniband/hw/mlx5/odp.c +4 -4
@@ -259,8 +259,8 @@ static void destroy_unused_implicit_child_mr(struct mlx5_ib_mr *mr) }if (MLX5_CAP_ODP (mr_to_mdev (mr)->mdev, mem_page_fault)) -__xa_erase (&mr_to_mdev (mr)->odp_mkeys, -mlx5_base_mkey (mr->mmkey.key)); +xa_erase (&mr_to_mdev (mr)->odp_mkeys, +mlx5_base_mkey (mr->mmkey.key));xa_unlock (&imr->implicit_children);/* Freeing a MR is a sleeping operation, so bounce to a work queue */
@@ -532,8 +532,8 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr, }if (MLX5_CAP_ODP (dev->mdev, mem_page_fault)) { - ret =__xa_store (&dev->odp_mkeys,mlx5_base_mkey (mr->mmkey.key), - &mr->mmkey,GFP_KERNEL ); + ret =xa_store (&dev->odp_mkeys,mlx5_base_mkey (mr->mmkey.key), + &mr->mmkey,GFP_KERNEL );if (xa_is_err (ret)) { ret =ERR_PTR (xa_err (ret));__xa_erase (&imr->implicit_children, idx);
Signature · 1 line
2.49.0
Quoted 22 lines · Leon Romanovsky
From: Or Har-Toov <[email protected]>
__xa_store()and__xa_erase()were used without holding the proper lock, which led to a lockdep warning due to unsafe RCPU usage. This patch replaces them withxa_store()andxa_erase(), which perform the necessary locking internally.=============================
WARNING: suspicious RCPU usage 6.14.0-rc7_for_upstream_debug_2025_03_18_15_01 #1 Not tainted
-----------------------------./include/linux/xarray.h:1211suspiciousrcu_dereference_protected()usage!other info that might help us debug this:
rcu_scheduler_active= 2,debug_locks= 1
3 locks held by kworker/u136:0/219:at: process_one_work+0xbe4/0x15f0 process_one_work+0x75c/0x15f0 pagefault_mr+0x9a5/0x1390 [mlx5_ib]
Quoted 7 lines
stack backtrace: CPU: 14 UID: 0 PID: 219 Comm: kworker/u136:0 Not tainted 6.14.0-rc7_for_upstream_debug_2025_03_18_15_01 #1 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.0-0-gd239552ce722-prebuilt.qemu.org 04/01/2014 Workqueue: mlx5_ib_page_fault mlx5_ib_eqe_pf_action [mlx5_ib] Call Trace:
Don't work wrap oops messages.. Ignore the checkpatch thing that tells you to do that.
Applied to for-rc
Thanks,
Jason