On Mon, Nov 06, 2017 at 07:45:23PM +0000, Bart Van Assche wrote:
> On Sat, 2017-11-04 at 08:19 -0600, Jens Axboe wrote:
> > On 11/03/2017 07:55 PM, Ming Lei wrote:
> > > It is very expensive to atomic_inc/atomic_dec the host wide counter of
> > > host->busy_count, and it should have been avoided via blk-mq's mechanism
> > > of getting driver tag, which uses the more efficient way of sbitmap queue.
> > >
> > > Also we don't check atomic_read(&sdev->device_busy) in
> > > scsi_mq_get_budget()
> > > and don't run queue if the counter becomes zero, so IO hang may be caused
> > > if all requests are completed just before the current SCSI device
> > > is added to shost->starved_list.
> >
> > This looks like an improvement. I have added it for 4.15.
> >
> > Bart, does this fix your hang?
>
> No, it doesn't. After I had reduced starget->can_queue in the SRP initiator I
> ran into the following hang while running the srp-test software:
>
> sysrq: SysRq : Show Blocked State
> task PC stack pid father
> systemd-udevd D 0 19882 467 0x80000106
> Call Trace:
> __schedule+0x2fa/0xbb0
> schedule+0x36/0x90
> io_schedule+0x16/0x40
> __lock_page+0x10a/0x140
> truncate_inode_pages_range+0x4ff/0x800
> truncate_inode_pages+0x15/0x20
> kill_bdev+0x35/0x40
> __blkdev_put+0x6d/0x1f0
> blkdev_put+0x4e/0x130
> blkdev_close+0x25/0x30
> __fput+0xed/0x1f0
> ____fput+0xe/0x10
> task_work_run+0x8b/0xc0
> do_exit+0x38d/0xc70
> do_group_exit+0x50/0xd0
> get_signal+0x2ad/0x8c0
> do_signal+0x28/0x680
> exit_to_usermode_loop+0x5a/0xa0
> do_syscall_64+0x12e/0x170
> entry_SYSCALL64_slow_path+0x25/0x25
>
> The SRP initiator driver was modified as follows for this test:
>
> diff --git a/drivers/infiniband/ulp/srp/ib_srp.c
> b/drivers/infiniband/ulp/srp/ib_srp.c
> index a6664467651e..9d24a871cc2e 100644
> --- a/drivers/infiniband/ulp/srp/ib_srp.c
> +++ b/drivers/infiniband/ulp/srp/ib_srp.c
>
> @@ -2835,6 +2839,13 @@ static int srp_reset_host(struct scsi_cmnd *scmnd)
> return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
> }
>
> +static int srp_target_alloc(struct scsi_target *starget)
> +{
> + starget->can_queue = 1;
> + return 0;
> +}
> +
> static int srp_slave_alloc(struct scsi_device *sdev)
> {
> struct Scsi_Host *shost = sdev->host;
> @@ -3039,6 +3050,7 @@ static struct scsi_host_template srp_template = {
> .module = THIS_MODULE,
> .name = "InfiniBand SRP initiator",
> .proc_name = DRV_NAME,
> + .target_alloc = srp_target_alloc,
> .slave_alloc = srp_slave_alloc,
> .slave_configure = srp_slave_configure,
> .info = srp_target_info,
Last time, you didn't mention the target patch for setting its
can_queue as 1, so I think you can't reproduce the issue on upstream
kernel without out-of-tree patch. Then looks it is another issue,
and we are making progress actually.
I just posted a one-line patch, which should address the small queue
depth issue, please let us know if it fixes your issue:
https://marc.info/?l=linux-block&m=151004881411480&w=2
--
Ming
