Re: [RFC 3/3] block: avoid deferral of blk_release_queue() work

[Nicolai Stange]

Nicolai Stange <nstange@xxxxxxx> writes:

> Bart Van Assche <bvanassche@xxxxxxx> writes:
>
>> The description of this patch mentions a single blk_release_queue() call
>> that happened in the past from a context from which sleeping is not
>> allowed and from which sleeping is allowed today. Have all other
>> blk_release_queue() / blk_put_queue() calls been verified to see whether
>> none of these happens from a context from which sleeping is not allowed?
>
> I've just done this today and found the following potentially
> problematic call paths to blk_put_queue().
>
> 1.) mem_cgroup_throttle_swaprate() takes a spinlock and
>     calls blkcg_schedule_throttle()->blk_put_queue().
>
>     Also note that AFAICS mem_cgroup_try_charge_delay() can be called
>     with GFP_ATOMIC.
>
> 2.) scsi_unblock_requests() gets called from a lot of drivers and
>     invoke blk_put_queue() through
>     scsi_unblock_requests() -> scsi_run_host_queues() ->
>     scsi_starved_list_run() -> blk_put_queue().
>
>     Most call sites are fine, the ones which are not are:
>     a.) pmcraid_complete_ioa_reset(). This gets assigned
>         to struct pmcraid_cmd's ->cmd_done and later invoked
>         under a spinlock.
>
>     b.) qla82xx_fw_dump() and qla8044_fw_dump().
>         These can potentially block w/o this patch already,
>         because both invoke qla2x00_wait_for_chip_reset().
>
> 	However, they can get called from IRQ context. For example,
>         qla82xx_intr_handler(), qla82xx_msix_default() and
>         qla82xx_poll() call qla2x00_async_event(), which calls
>         ->fw_dump().
>
> 	The aforementioned functions can also reach ->fw_dump() through
>         qla24xx_process_response_queue()->qlt_handle_abts_recv()->qlt_response_pkt_all_vps()
>         ->qlt_response_pkt()->qlt_handle_abts_completion()->qlt_chk_unresolv_exchg()
>         -> ->fw_dump().
>
> 	But I'd consider this a problem with the driver -- either
> 	->fw_dump() can sleep and must not be called from IRQ context
>         or they must not invoke qla2x00_wait_for_hba_ready().
>
>
> (I can share the full analysis, but it's lengthy and contains nothing
>  interesting except for what is listed above).
>
>
> One final note though: If I'm not mistaken, then the final
> blk_put_queue() can in principle block even today, simply by virtue of
> the kernfs operations invoked through
> kobject_put()->kobject_release()->kobject_cleanup()->kobject_del()
> ->sysfs_remove_dir()->kernfs_remove()->mutex_lock()?\

That's wrong, I missed kobject_del() invocation issued from
blk_unregister_queue(). Thus, blk_put_queue() in its current
implementation won't ever block.

Thanks,

Nicolai

-- 
SUSE Software Solutions Germany GmbH, Maxfeldstr. 5, 90409 NÃrnberg, Germany
(HRB 36809, AG NÃrnberg), GF: Felix ImendÃrffer