Re: kernel BUG scsi_dh_alua sleeping from invalid context && kernel WARNING do not call blocking ops when !TASK_RUNNING

[Bart Van Assche <bvanassche@acm.org>]

On 1/17/23 14:03, Martin Wilck wrote:
> On Tue, 2023-01-17 at 13:52 -0800, Bart Van Assche wrote:
>> On 1/17/23 13:48, Martin Wilck wrote:
>>> Yes, that was my suggestion. Just defer the scsi_device_put() call
>>> in
>>> alua_rtpg_queue() in the case where the actual RTPG handler is not
>>> queued. I won't have time for that before next week though.
>>
>> Hi Martin,
>>
>> Do you agree that the call trace shared by Steffen is not sufficient
>> to
>> conclude that this change is necessary?
> 
> Hmm, I suppose I missed your point... to re-iterate my thinking:
> 
>   1 alua_queue_rtpg() must take a ref to the sdev before queueing work,
>     whether or not the caller already has one
>   2 queue_delayed_work() can fail
>   3 if queue_delayed_work() fails, alua_queue_rtpg() must drop the ref
>     it just took
>   4 BUT (and this is what I guess I missed) this ref can't be the last
>     one dropped, because the caller of alua_rtpg_queue() must still hold
>     a reference. And scsi_device_put() only sleeps if the last ref is
>     dropped. Therefore the issue in Steffen's call stack should
>     indeed be fixed just by removing the might_sleep(). If all callers
>     callers of alua_rtpg_queue() must hold an sdev reference (I believe
>     they do), we can indeed remove the might_sleep() entirely.
> 
> Is this correct reasoning, and what you meant previously? If yes, I
> agree, and I apologize for not realizing it in the first place.
> But I think this is subtle enough to deserve a comment in the code.

Yes, that's what I'm thinking.

How about the patch below?

Thanks,

Bart.

[PATCH] scsi: device_handler: alua: Remove a might_sleep() annotation

The might_sleep() annotation in alua_rtpg_queue() is not correct since the
command completion code may call this function from atomic context.
Calling alua_rtpg_queue() from atomic context in the command completion
path is fine since request submitters must hold an sdev reference until
command execution has completed. This patch fixes the following kernel
warning:

BUG: sleeping function called from invalid context at drivers/scsi/device_handler/scsi_dh_alua.c:992
Call Trace:
  dump_stack_lvl+0xac/0x100
  __might_resched+0x284/0x2c8
  alua_rtpg_queue+0x3c/0x98 [scsi_dh_alua]
  alua_check+0x122/0x250 [scsi_dh_alua]
  alua_check_sense+0x172/0x228 [scsi_dh_alua]
  scsi_check_sense+0x8a/0x2e0
  scsi_decide_disposition+0x286/0x298
  scsi_complete+0x6a/0x108
  blk_complete_reqs+0x6e/0x88
  __do_softirq+0x13e/0x6b8
  __irq_exit_rcu+0x14a/0x170
  irq_exit_rcu+0x22/0x50
  do_ext_irq+0x10a/0x1d0

Reported-by: Steffen Maier <maier@linux.ibm.com>
Cc: Steffen Maier <maier@linux.ibm.com>
Cc: Martin Wilck <mwilck@suse.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
---
  drivers/scsi/device_handler/scsi_dh_alua.c | 5 +++--
  1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c b/drivers/scsi/device_handler/scsi_dh_alua.c
index 55a5073248f8..362fa631f39b 100644
--- a/drivers/scsi/device_handler/scsi_dh_alua.c
+++ b/drivers/scsi/device_handler/scsi_dh_alua.c
@@ -987,6 +987,9 @@ static void alua_rtpg_work(struct work_struct *work)
   *
   * Returns true if and only if alua_rtpg_work() will be called asynchronously.
   * That function is responsible for calling @qdata->fn().
+ *
+ * Context: may be called from atomic context (alua_check()) only if the caller
+ *	holds an sdev reference.
   */
  static bool alua_rtpg_queue(struct alua_port_group *pg,
  			    struct scsi_device *sdev,
@@ -995,8 +998,6 @@ static bool alua_rtpg_queue(struct alua_port_group *pg,
  	int start_queue = 0;
  	unsigned long flags;

-	might_sleep();
-
  	if (WARN_ON_ONCE(!pg) || scsi_device_get(sdev))
  		return false;