]> www.infradead.org Git - users/jedix/linux-maple.git/commitdiff
crypto: hisilicon/qm - prevent soft lockup in receive loop
authorLongfang Liu <liulongfang@huawei.com>
Fri, 20 Oct 2023 09:35:58 +0000 (17:35 +0800)
committerHerbert Xu <herbert@gondor.apana.org.au>
Fri, 27 Oct 2023 10:04:27 +0000 (18:04 +0800)
In the scenario where the accelerator business is fully loaded.
When the workqueue receiving messages and performing callback
processing, there are a large number of messages that need to be
received, and there are continuously messages that have been
processed and need to be received.
This will cause the receive loop here to be locked for a long time.
This scenario will cause watchdog timeout problems on OS with kernel
preemption turned off.

The error logs:
watchdog: BUG: soft lockup - CPU#23 stuck for 23s! [kworker/u262:1:1407]
[ 1461.978428][   C23] Call trace:
[ 1461.981890][   C23]  complete+0x8c/0xf0
[ 1461.986031][   C23]  kcryptd_async_done+0x154/0x1f4 [dm_crypt]
[ 1461.992154][   C23]  sec_skcipher_callback+0x7c/0xf4 [hisi_sec2]
[ 1461.998446][   C23]  sec_req_cb+0x104/0x1f4 [hisi_sec2]
[ 1462.003950][   C23]  qm_poll_req_cb+0xcc/0x150 [hisi_qm]
[ 1462.009531][   C23]  qm_work_process+0x60/0xc0 [hisi_qm]
[ 1462.015101][   C23]  process_one_work+0x1c4/0x470
[ 1462.020052][   C23]  worker_thread+0x150/0x3c4
[ 1462.024735][   C23]  kthread+0x108/0x13c
[ 1462.028889][   C23]  ret_from_fork+0x10/0x18

Therefore, it is necessary to add an actively scheduled operation in the
while loop to prevent this problem.
After adding it, no matter whether the OS turns on or off the kernel
preemption function. Neither will cause watchdog timeout issues.

Signed-off-by: Longfang Liu <liulongfang@huawei.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
drivers/crypto/hisilicon/qm.c

index 6ee24313d851257b88fbd032d921f2454f18d5b8..18599f3634c3c1bc0adde2402cf8d12c119c0ae1 100644 (file)
@@ -888,6 +888,8 @@ static void qm_poll_req_cb(struct hisi_qp *qp)
                qm_db(qm, qp->qp_id, QM_DOORBELL_CMD_CQ,
                      qp->qp_status.cq_head, 0);
                atomic_dec(&qp->qp_status.used);
+
+               cond_resched();
        }
 
        /* set c_flag */