Received-SPF: pass (google.com: domain of linux-kernel+bounces-161321-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
Message-ID: <d200fc7c-c781-49f1-8277-bdb5d537b1f4@grimberg.me>
Date: Sun, 28 Apr 2024 11:10:43 +0300
Precedence: bulk
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [Bug Report] nvme connect deadlock in allocating tag
To: kwb <wangbing.kuang@shopee.com>, james.smart@broadcom.com,
 kbusch@kernel.org, axboe@fb.com, hch@lst.de, linux-nvme@lists.infradead.org,
 linux-kernel@vger.kernel.org
Cc: chunguang.xu@shopee.com
References: <20240428063147.88058-1-wangbing.kuang@shopee.com>
Content-Language: he-IL, en-US
From: Sagi Grimberg <sagi@grimberg.me>
In-Reply-To: <20240428063147.88058-1-wangbing.kuang@shopee.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit


On 28/04/2024 9:31, kwb wrote:
> Hi,
> We found nvme connect will dealock when it cannot alloc tag in admin queue. So we reproduce it and find a way to work around. The solution is to utilize reserve tag for connecting.
> Here is the deadlock environment:
> 1. the process [kworker/u129:1+nvme-wq] want to connect wait for geting tag, but tag is used up:
> [<0>] blk_mq_get_tag+0x11d/0x2d0
> [<0>] __blk_mq_alloc_request+0x92/0x180
> [<0>] blk_mq_alloc_request+0x7c/0xc0
> [<0>] nvme_alloc_request+0x28/0x100 [nvme_core]
> [<0>] __nvme_submit_sync_cmd+0x1ea/0x230 [nvme_core]
> [<0>] nvmf_reg_read64+0x62/0xa0 [nvme_fabrics]
> [<0>] nvme_enable_ctrl+0x25/0xb0 [nvme_core]
> [<0>] nvme_tcp_setup_ctrl+0x257/0x340 [nvme_tcp]
> [<0>] nvme_tcp_reconnect_ctrl_work+0x24/0x40 [nvme_tcp]
> [<0>] process_one_work+0x228/0x3d0
> [<0>] worker_thread+0x4d/0x3f0
> [<0>] kthread+0x127/0x150
> [<0>] ret_from_fork+0x1f/0x30
> 2. many processes (here is nvme list) is waiting for connecting:
> [<0>] blk_execute_rq+0x8d/0x110
> [<0>] nvme_execute_passthru_rq+0x60/0x1f0 [nvme_core]
> [<0>] nvme_submit_user_cmd+0x23e/0x400 [nvme_core]
> [<0>] nvme_user_cmd+0x163/0x1d0 [nvme_core]
> [<0>] nvme_ctrl_ioctl+0x2e/0x40 [nvme_core]
> [<0>] __nvme_ioctl+0x78/0xc0 [nvme_core]
> [<0>] nvme_ioctl+0x1e/0x20 [nvme_core]
> [<0>] blkdev_ioctl+0x126/0x260
> [<0>] block_ioctl+0x4a/0x60
> [<0>] __x64_sys_ioctl+0x91/0xc0
> [<0>] do_syscall_64+0x59/0xc0
> [<0>] entry_SYSCALL_64_after_hwframe+0x44/0xae
>
> Reproduce method is very eazy:
> 1. call many nvme list
> 2. make nvme io timeout to recover connection
> 3. trick is to make reconnect-delay much time, eg:30s
>
> The solution is the appending patch. it is tested and also consider keepalive and reset/showdown tag reserve.

The error_recovery work should unquiesce the admin_q, which should fail 
fast all pending admin commands,
so it is unclear to me how the connect process gets stuck.

What is step (2) - make nvme io timeout to recover the connection?

Is this reproducing with upstream nvme? or is this some distro kernel 
where this happens?
Do you have the below patch applied?
de105068fead ("nvme: fix reconnection fail due to reserved tag allocation")