Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [PATCH v3] nvmet,rxe: defer ip datagram sending to tasklet
To:     Alexandru Moise <00moses.alexander00@gmail.com>,
        monis@mellanox.com, dledford@redhat.com, jgg@ziepe.ca,
        linux-rdma@vger.kernel.org, linux-kernel@vger.kernel.org
References: <20180507205858.GA11664@gmail.com>
From:   Yanjun Zhu <yanjun.zhu@oracle.com>
Organization: Oracle Corporation
Message-ID: <562873bf-075c-f148-c528-dddf188113c2@oracle.com>
Date:   Tue, 8 May 2018 11:09:18 +0800
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:52.0) Gecko/20100101
 Thunderbird/52.7.0
MIME-Version: 1.0
In-Reply-To: <20180507205858.GA11664@gmail.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 7bit
Content-Language: en-US
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


On 2018/5/8 4:58, Alexandru Moise wrote:
> This addresses 3 separate problems:
>
> 1. When using NVME over Fabrics we may end up sending IP
> packets in interrupt context, we should defer this work
> to a tasklet.
>
> [   50.939957] WARNING: CPU: 3 PID: 0 at kernel/softirq.c:161 __local_bh_enable_ip+0x1f/0xa0
> [   50.942602] CPU: 3 PID: 0 Comm: swapper/3 Kdump: loaded Tainted: G        W         4.17.0-rc3-ARCH+ #104
> [   50.945466] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.0-20171110_100015-anatol 04/01/2014
> [   50.948163] RIP: 0010:__local_bh_enable_ip+0x1f/0xa0
> [   50.949631] RSP: 0018:ffff88009c183900 EFLAGS: 00010006
> [   50.951029] RAX: 0000000080010403 RBX: 0000000000000200 RCX: 0000000000000001
> [   50.952636] RDX: 0000000000000000 RSI: 0000000000000200 RDI: ffffffff817e04ec
> [   50.954278] RBP: ffff88009c183910 R08: 0000000000000001 R09: 0000000000000614
> [   50.956000] R10: ffffea00021d5500 R11: 0000000000000001 R12: ffffffff817e04ec
> [   50.957779] R13: 0000000000000000 R14: ffff88009566f400 R15: ffff8800956c7000
> [   50.959402] FS:  0000000000000000(0000) GS:ffff88009c180000(0000) knlGS:0000000000000000
> [   50.961552] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [   50.963798] CR2: 000055c4ec0ccac0 CR3: 0000000002209001 CR4: 00000000000606e0
> [   50.966121] Call Trace:
> [   50.966845]  <IRQ>
> [   50.967497]  __dev_queue_xmit+0x62d/0x690
> [   50.968722]  dev_queue_xmit+0x10/0x20
> [   50.969894]  neigh_resolve_output+0x173/0x190
> [   50.971244]  ip_finish_output2+0x2b8/0x370
> [   50.972527]  ip_finish_output+0x1d2/0x220
> [   50.973785]  ? ip_finish_output+0x1d2/0x220
> [   50.975010]  ip_output+0xd4/0x100
> [   50.975903]  ip_local_out+0x3b/0x50
> [   50.976823]  rxe_send+0x74/0x120
> [   50.977702]  rxe_requester+0xe3b/0x10b0
> [   50.978881]  ? ip_local_deliver_finish+0xd1/0xe0
> [   50.980260]  rxe_do_task+0x85/0x100
> [   50.981386]  rxe_run_task+0x2f/0x40
> [   50.982470]  rxe_post_send+0x51a/0x550
> [   50.983591]  nvmet_rdma_queue_response+0x10a/0x170
> [   50.985024]  __nvmet_req_complete+0x95/0xa0
> [   50.986287]  nvmet_req_complete+0x15/0x60
> [   50.987469]  nvmet_bio_done+0x2d/0x40
> [   50.988564]  bio_endio+0x12c/0x140
> [   50.989654]  blk_update_request+0x185/0x2a0
> [   50.990947]  blk_mq_end_request+0x1e/0x80
> [   50.991997]  nvme_complete_rq+0x1cc/0x1e0
> [   50.993171]  nvme_pci_complete_rq+0x117/0x120
> [   50.994355]  __blk_mq_complete_request+0x15e/0x180
> [   50.995988]  blk_mq_complete_request+0x6f/0xa0
> [   50.997304]  nvme_process_cq+0xe0/0x1b0
> [   50.998494]  nvme_irq+0x28/0x50
> [   50.999572]  __handle_irq_event_percpu+0xa2/0x1c0
> [   51.000986]  handle_irq_event_percpu+0x32/0x80
> [   51.002356]  handle_irq_event+0x3c/0x60
> [   51.003463]  handle_edge_irq+0x1c9/0x200
> [   51.004473]  handle_irq+0x23/0x30
> [   51.005363]  do_IRQ+0x46/0xd0
> [   51.006182]  common_interrupt+0xf/0xf
> [   51.007129]  </IRQ>
>
> 2. Work must always be offloaded to tasklet when using NVMEoF in order to
> solve lock ordering between neigh->ha_lock seqlock and the nvme queue lock:
>
> [   77.833783]  Possible interrupt unsafe locking scenario:
> [   77.833783]
> [   77.835831]        CPU0                    CPU1
> [   77.837129]        ----                    ----
> [   77.838313]   lock(&(&n->ha_lock)->seqcount);
> [   77.839550]                                local_irq_disable();
> [   77.841377]                                lock(&(&nvmeq->q_lock)->rlock);
> [   77.843222]                                lock(&(&n->ha_lock)->seqcount);
> [   77.845178]   <Interrupt>
> [   77.846298]     lock(&(&nvmeq->q_lock)->rlock);
> [   77.847986]
> [   77.847986]  *** DEADLOCK ***
>
> 3. Same goes for the lock ordering between sch->q.lock and nvme queue lock:
>
> [   47.634271]  Possible interrupt unsafe locking scenario:
> [   47.634271]
> [   47.636452]        CPU0                    CPU1
> [   47.637861]        ----                    ----
> [   47.639285]   lock(&(&sch->q.lock)->rlock);
> [   47.640654]                                local_irq_disable();
> [   47.642451]                                lock(&(&nvmeq->q_lock)->rlock);
> [   47.644521]                                lock(&(&sch->q.lock)->rlock);
> [   47.646480]   <Interrupt>
> [   47.647263]     lock(&(&nvmeq->q_lock)->rlock);
> [   47.648492]
> [   47.648492]  *** DEADLOCK ***
>
> Using NVMEoF after this patch seems to finally be stable, without it,
> rxe eventually deadlocks the whole system and causes RCU stalls.
>
> Signed-off-by: Alexandru Moise <00moses.alexander00@gmail.com>
> ---
> v2->v3: Abandoned the idea of just offloading work to tasklet in irq
> context. NVMEoF needs the work to always be offloaded in order to
> function. I am unsure how large an impact this has on other rxe users.
>
>   drivers/infiniband/sw/rxe/rxe_verbs.c | 12 ++----------
>   1 file changed, 2 insertions(+), 10 deletions(-)
>
> diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.c b/drivers/infiniband/sw/rxe/rxe_verbs.c
> index 2cb52fd48cf1..564453060c54 100644
> --- a/drivers/infiniband/sw/rxe/rxe_verbs.c
> +++ b/drivers/infiniband/sw/rxe/rxe_verbs.c
> @@ -761,7 +761,6 @@ static int rxe_post_send_kernel(struct rxe_qp *qp, struct ib_send_wr *wr,
>   	unsigned int mask;
>   	unsigned int length = 0;
>   	int i;
> -	int must_sched;
>   
>   	while (wr) {
>   		mask = wr_opcode_mask(wr->opcode, qp);
> @@ -791,14 +790,7 @@ static int rxe_post_send_kernel(struct rxe_qp *qp, struct ib_send_wr *wr,
>   		wr = wr->next;
>   	}
>   
> -	/*
> -	 * Must sched in case of GSI QP because ib_send_mad() hold irq lock,
> -	 * and the requester call ip_local_out_sk() that takes spin_lock_bh.
> -	 */
> -	must_sched = (qp_type(qp) == IB_QPT_GSI) ||
> -			(queue_count(qp->sq.queue) > 1);
> -
> -	rxe_run_task(&qp->req.task, must_sched);
> +	rxe_run_task(&qp->req.task, 1);
>   	if (unlikely(qp->req.state == QP_STATE_ERROR))
>   		rxe_run_task(&qp->comp.task, 1);
>   
> @@ -822,7 +814,7 @@ static int rxe_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
>   
>   	if (qp->is_user) {
>   		/* Utilize process context to do protocol processing */
> -		rxe_run_task(&qp->req.task, 0);
> +		rxe_run_task(&qp->req.task, 1);
Hi,

 From your description, it seems that all the problems are caused by 
NVME kernel.

Am I correct?

It seems that this block handles process context. If we keep 
"rxe_run_task(&qp->req.task, 0);",

Do all the problems still occur?

Zhu Yanjun


>   		return 0;
>   	} else
>   		return rxe_post_send_kernel(qp, wr, bad_wr);