On Fri, 2017-05-26 at 08:34 +0200, Christoph Hellwig wrote:

On Thu, May 25, 2017 at 11:43:16AM -0700, Bart Van Assche wrote:

quoted

Several block drivers need to initialize the driver-private data
after having called blk_get_request() and before .prep_rq_fn() is
called, e.g. when submitting a REQ_OP_SCSI_* request. Avoid that
that initialization code has to be repeated after every
blk_get_request() call by adding a new callback function to struct
request_queue.

=20
I still think we should do this at the end of blk_mq_alloc_request /
blk_old_get_request to exactly keep the old semantics and keep the
calls out of the hot path.

Hello Christoph,

I have tried to move that call into blk_mq_alloc_request() but that
resulted in a kernel oops during boot due to scsi_add_cmd_to_list()
dereferencing scsi_cmnd.device and due to that pointer being invalid.
I think that pointer was invalid because moving the initialize_rq_fn()
call into blk_mq_alloc_request() caused request initialization to be
skipped for the following code path:
submit_bio()
-> generic_make_request()
  -> .make_request_fn =3D=3D blk_mq_make_request()
 =A0  -> blk_mq_sched_get_request()
      -> __blk_mq_alloc_request()
 =A0 =A0 =A0  -> blk_mq_rq_ctx_init()

This is why I would like to keep the .initialize_rq_fn() call in
blk_mq_rq_ctx_init().

Bart.=