On Mon, 27 Oct 2008 10:20:21 +0200 Boaz Harrosh <bharrosh@xxxxxxxxxxx> wrote: > FUJITA Tomonori wrote: > > On Sun, 26 Oct 2008 11:38:04 +0200 > > Boaz Harrosh <bharrosh@xxxxxxxxxxx> wrote: > > > >> FUJITA Tomonori wrote: > >>> CC'ed Jens, > >>> > >>> On Wed, 22 Oct 2008 19:27:35 -0700 > >>> Seokmann Ju <seokmann.ju@xxxxxxxxxx> wrote: > >>> > >>>> And it seems like that the panic is happening due to the fact that > >>>> blk_delete_timer() is not called upon having completion of the service. > >>>> In other words, the block layer calls blk_add_timer() prior to > >>>> dispatch the service but, it doesn't call blk_delete_timer() when it > >>>> returned. > >>> Yeah, we need to call blk_delete_timer somewhere. > >>> > >>> > >>>> Just for heck of it, I've tried out by adding blk_delete_timer() in > >>>> the ~/block/blk-exec.c:blk_end_sync_rq() and it seems fixes the problem. > >>> I think blk_end_sync_rq() is not the good place. From the perspective > >>> of bsg, we need to handle both blk_execute_rq_nowait and > >>> blk_execute_rq. > >>> > >>> > >>>> Seems like that there are APIs in the block layer that are call the > >>>> blk_delete_timer(), including, > >>>> - blk_end_io() > >>>> - __blk_end_request() > >>>> > >>>> Could you guide me what is right way to fix the problem? > >>> Exporting blk_delete_timer is one option, but it doesn't look very > >>> nice (since the block layer doesn't export any details about its timer > >>> infrastructure), I think. Modifying blk_end_io() to make it usable for > >>> requests via something like bsg might be better. > >>> > >>> Anyway, we need to ask Jens. > >>> > >>> Jens, fc people have working on fc pass through support via bsg, which > >>> hooks bsg's request queue on fc transport objects (We did the similar > >>> thing for sas transport). > >>> > >>> We want the timeout feature for fc pass through and I think that it's > >>> nice to use the block layer timeout feature for it. But the users of > >>> bsg request queue don't need (or call) APIs such as > >>> end_that_request_last to call blk_delete_timer internally. How should > >>> these users call blk_delete_timer? > >> TOMO Hi > >> If a command is queued by bsg to a scsi device, which is posible. Then > >> blk_end_request() is called by scsi-ml. So it does work. > > > > It doesn't work for bsg's scsi transport pass through stuff such as > > SMP (sas management protocol, we already support) and FC. Virtually, > > they don't use scsi-ml. > > > > Right, I know that, that's why I say. > > > > >> I think that all block-queue consumers should call one of > >> blk_end_request(), > > > > This is kinda what I suggested in the previous mail but as I wrote, > > some of them don't now. > > > > I think they should, specially if they're going to use the timer. > The way I see it they must. It's kind of a block layer API thing. > Someone calls blk_execute_xx then eventually someone needs to call > blk_end_request. As I said, with the current API, we don't need to do that. But it's kinda what I suggested in that mail. > You could call it from bsg but only temporary until > all are fixed. (because you will need an ugly check to see if request > was not already ended) Note that now nothing in mainline is broken with regard to this issue. SMP doesn't have the timeout feature. FC pass though is the first one that wants to do this. So no need to have any temporary fix. -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html