Re: [bug report] nvme passthrough: request failed when blocksize exceeding max_hw_sectors

[Kanchan Joshi <joshi.k@samsung.com>]

[-- Attachment #1: Type: text/plain, Size: 3610 bytes --]

On Mon, Jun 26, 2023 at 05:09:59PM +0800, Ferry Meng wrote:
>Hello:
>
>I'm testing the io-uring nvme passthrough via fio. But I have 
>encountered the following issue:
>When I specify 'blocksize' exceeding 128KB (actually the maximum size 
>per request can send 'max_sectors_kb'), the creation of request failed 
>and directly returned -22 (-EINVAL).
>
>For example:
>
># cat fio.job
>
>    [global]
>    ioengine=io_uring_cmd
>    thread=1
>    time_based
>    numjobs=1
>    iodepth=1
>    runtime=120
>    rw=randwrite
>    cmd_type=nvme
>    hipri=1
>
>    [randwrite]
>    bs=132k
>    filename=/dev/ng1n1
>
># fio fio.job
>randwrite: (g=0): rw=randwrite, bs=(R) 132KiB-132KiB, (W) 
>132KiB-132KiB, (T) 132KiB-132KiB, ioengine=io_uring_cmd, iodepth=1
>fio-3.34-10-g2fa0-dirty
>Starting 1 thread
>fio: io_u error on file /dev/ng1n1: Invalid argument: write 
>offset=231584956416, buflen=135168
>fio: pid=148989, err=22/file:io_u.c:1889, func=io_u error, 
>error=Invalid argument
>
>I tracked the position that returns the error val in kernel and dumped 
>calltrace.
>
>[   83.352715] nvme nvme1: 15/0/1 default/read/poll queues
>[   83.363273] nvme nvme1: Ignoring bogus Namespace Identifiers
>[   91.578457] CPU: 14 PID: 3993 Comm: fio Not tainted 
>6.4.0-rc7-00014-g692b7dc87ca6-dirty #2
>[   91.578462] Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS 
>2221b89 04/01/2014
>[   91.578463] Call Trace:
>[   91.578476]  <TASK>
>[   91.578478]  dump_stack_lvl+0x36/0x50
>[   91.578484]  ll_back_merge_fn+0x20d/0x320
>[   91.578490]  blk_rq_append_bio+0x6d/0xc0
>[   91.578492]  bio_map_user_iov+0x24a/0x3d0
>[   91.578494]  blk_rq_map_user_iov+0x292/0x680
>[   91.578496]  ? blk_mq_get_tag+0x249/0x280
>[   91.578500]  blk_rq_map_user+0x56/0x80
>[   91.578503]  nvme_map_user_request.isra.15+0x90/0x1e0 [nvme_core]
>[   91.578515]  nvme_uring_cmd_io+0x29d/0x2f0 [nvme_core]
>[   91.578522]  io_uring_cmd+0x89/0x110
>[   91.578526]  ? __pfx_io_uring_cmd+0x10/0x10
>[   91.578528]  io_issue_sqe+0x1e0/0x2d0
>[   91.578530]  io_submit_sqes+0x1e3/0x650
>[   91.578532]  __x64_sys_io_uring_enter+0x2da/0x450
>[   91.578534]  do_syscall_64+0x3b/0x90
>[   91.578537]  entry_SYSCALL_64_after_hwframe+0x72/0xdc
>
>Here in bio_map_user_iov()->blk_rq_append_bio(), I found the error val 
>-EINVAL:
>
>blk_rq_append_bio:
>    ...
>    if (!ll_back_merge_fn(rq, bio, nr_segs))
>        return -EINVAL;
>    rq->biotail->bi_next = bio;
>    ...
>
>And in ll_back_merge_fn(), returns 0 if merge can't happen. It checks 
>the request size:
>ll_back_merge_fn:
>    if (blk_rq_sectors(req) + bio_sectors(bio) >
>        blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
>            req_set_nomerge(req->q, req);
>            return 0;
>    }
>
>The ROOT cause is: In blk_rq_get_max_sectors, it returns 
>'max_hw_sectors' directly(in my device ,it's 256 sector, which means 
>128KB), causing the above inequality to hold true.
>blk_rq_get_max_sectors:
>    ...
>    if (blk_rq_is_passthrough(rq)){
>        return q->limits.max_hw_sectors;
>    }
>    ...
>
>I checked my disk's specs(cat 
>/sys/block/<mydisk>/queue/max_hw_sectors_kb 
>/sys/block/<mydisk>/queue/max_sectors_kb), both are 128KB.So I think 
>this arg causing the issue.
>
>I'm not sure if this is a designed restriction. Or should I have to 
>take care of it in application?

Right, passthrough interface does not abstract the device limits.
This needs to be handled in application.

[-- Attachment #2: Type: text/plain, Size: 0 bytes --]