Re: [PATCH v2 08/16] scsi: fc: implement kref backed reference counting

From: Hannes Reinecke
Date: Thu Oct 13 2016 - 07:43:51 EST


On 10/12/2016 03:06 PM, Johannes Thumshirn wrote:
> Implement kref backed reference counting instead of rolling our own. This
> elimnates the need of the following fields in 'struct fc_bsg_job':
> * ref_cnt
> * state_flags
> * job_lock
> bringing us close to unification of 'struct fc_bsg_job' and 'struct bsg_job'.
>
> Signed-off-by: Johannes Thumshirn <jthumshirn@xxxxxxx>
> ---
> drivers/scsi/scsi_transport_fc.c | 38 +++++++++-----------------------------
> include/scsi/scsi_transport_fc.h | 4 +---
> 2 files changed, 10 insertions(+), 32 deletions(-)
>
> diff --git a/drivers/scsi/scsi_transport_fc.c b/drivers/scsi/scsi_transport_fc.c
> index 96b3a2e..b0e28af 100644
> --- a/drivers/scsi/scsi_transport_fc.c
> +++ b/drivers/scsi/scsi_transport_fc.c
> @@ -3560,16 +3560,9 @@ fc_vport_sched_delete(struct work_struct *work)
> * @job: fc_bsg_job that is to be torn down
> */
> static void
> -fc_destroy_bsgjob(struct fc_bsg_job *job)
> +fc_destroy_bsgjob(struct kref *kref)
> {
> - unsigned long flags;
> -
> - spin_lock_irqsave(&job->job_lock, flags);
> - if (job->ref_cnt) {
> - spin_unlock_irqrestore(&job->job_lock, flags);
> - return;
> - }
> - spin_unlock_irqrestore(&job->job_lock, flags);
> + struct fc_bsg_job *job = container_of(kref, struct fc_bsg_job, kref);
>
> put_device(job->dev); /* release reference for the request */
>
> @@ -3620,15 +3613,9 @@ EXPORT_SYMBOL_GPL(fc_bsg_jobdone);
> static void fc_bsg_softirq_done(struct request *rq)
> {
> struct fc_bsg_job *job = rq->special;
> - unsigned long flags;
> -
> - spin_lock_irqsave(&job->job_lock, flags);
> - job->state_flags |= FC_RQST_STATE_DONE;
> - job->ref_cnt--;
> - spin_unlock_irqrestore(&job->job_lock, flags);
>
> blk_end_request_all(rq, rq->errors);
> - fc_destroy_bsgjob(job);
> + kref_put(&job->kref, fc_destroy_bsgjob);
> }
>
> /**
Hmm. blk_end_request_all() (potentially) triggers a recursion into all
.end_io callbacks, which might end up doing god-knows-what.
With some delays in doing so
During that time we have no idea that bsg_softirq_done() is actually
running, and we might clash with eg. timeouts or somesuch.

Maybe it's an idea to move blk_end_request_all into the kref destroy
callback; that way we're guaranteed to call it only once and would avoid
this situation.

Cheers,

Hannes
--
Dr. Hannes Reinecke Teamlead Storage & Networking
hare@xxxxxxx +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: F. Imendörffer, J. Smithard, J. Guild, D. Upmanyu, G. Norton
HRB 21284 (AG Nürnberg)