If bnxt FW behaves unexpectedly because of FW bug or unexpected behavior it
can send completions for old cookies that have already been handled by the
bnxt driver. If that old cookie was associated with an old calling context
the driver will try to access that caller memory again because the driver
never clean the is_waiter_alive flag after the caller successfully complete
waiting, and this access will cause the following kernel panic:
Call Trace:
<IRQ>
? __die+0x20/0x70
? page_fault_oops+0x75/0x170
? exc_page_fault+0xaa/0x140
? asm_exc_page_fault+0x22/0x30
? bnxt_qplib_process_qp_event.isra.0+0x20c/0x3a0 [bnxt_re]
? srso_return_thunk+0x5/0x5f
? __wake_up_common+0x78/0xa0
? srso_return_thunk+0x5/0x5f
bnxt_qplib_service_creq+0x18d/0x250 [bnxt_re]
tasklet_action_common+0xac/0x210
handle_softirqs+0xd3/0x2b0
__irq_exit_rcu+0x9b/0xc0
common_interrupt+0x7f/0xa0
</IRQ>
<TASK>
To avoid the above unexpected behavior clear the is_waiter_alive flag
every time the caller finishes waiting for a completion.
Fixes: 691eb7c6110f ("RDMA/bnxt_re: handle command completions after driver detect a timedout")
Signed-off-by: Mohammad Heib <mheib@xxxxxxxxxx>
---
drivers/infiniband/hw/bnxt_re/qplib_rcfw.c | 16 ++++++++--------
1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/drivers/infiniband/hw/bnxt_re/qplib_rcfw.c b/drivers/infiniband/hw/bnxt_re/qplib_rcfw.c
index f5713e3c39fb..eaf92029862b 100644
--- a/drivers/infiniband/hw/bnxt_re/qplib_rcfw.c
+++ b/drivers/infiniband/hw/bnxt_re/qplib_rcfw.c
@@ -511,15 +511,15 @@ static int __bnxt_qplib_rcfw_send_message(struct bnxt_qplib_rcfw *rcfw,
else
rc = __poll_for_resp(rcfw, cookie);
- if (rc) {
- spin_lock_irqsave(&rcfw->cmdq.hwq.lock, flags);
- crsqe = &rcfw->crsqe_tbl[cookie];
- crsqe->is_waiter_alive = false;
- if (rc == -ENODEV)
- set_bit(FIRMWARE_STALL_DETECTED, &rcfw->cmdq.flags);
- spin_unlock_irqrestore(&rcfw->cmdq.hwq.lock, flags);
+
+ spin_lock_irqsave(&rcfw->cmdq.hwq.lock, flags);
+ crsqe = &rcfw->crsqe_tbl[cookie];
+ crsqe->is_waiter_alive = false;
+ if (rc == -ENODEV)
+ set_bit(FIRMWARE_STALL_DETECTED, &rcfw->cmdq.flags);
+ spin_unlock_irqrestore(&rcfw->cmdq.hwq.lock, flags);
+ if (rc)
return -ETIMEDOUT;
- }
if (evnt->status) {
/* failed with status */
--
2.34.3
