Re: next bio iters break discard?

[Kent Overstreet <kmo@daterainc.com>]

On Thu, Jan 16, 2014 at 12:21:10PM -0800, Hugh Dickins wrote:
> On Tue, 14 Jan 2014, Kent Overstreet wrote:
> > 
> > Does the below patch look like what we want? I'm assuming that if
> 
> You don't fill me with confidence ;)
> 
> > multiple WRITE_SAME bios are merged, since they're all writing the same
> > data we can consider the entire request to be a single segment.
> > 
> > commit 1755e7ffc5745591d37b8956ce2512f4052a104a
> > Author: Kent Overstreet <kmo@daterainc.com>
> > Date:   Tue Jan 14 14:22:01 2014 -0800
> > 
> >     block: Explicitly handle discard/write same when counting segments
> > 
> > diff --git a/block/blk-merge.c b/block/blk-merge.c
> > index 8f8adaa..7d977f8 100644
> > --- a/block/blk-merge.c
> > +++ b/block/blk-merge.c
> > @@ -21,6 +21,12 @@ static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
> >  	if (!bio)
> >  		return 0;
> >  
> > +	if (bio->bi_rw & REQ_DISCARD)
> > +		return 0;
> > +
> > +	if (bio->bi_rw & REQ_WRITE_SAME)
> > +		return 1;
> > +
> >  	fbio = bio;
> >  	cluster = blk_queue_cluster(q);
> >  	seg_size = 0;
> 
> For me this just shifts the crash,
> from __blk_recalc_rq_segments() to blk_rq_map_sg():
> 
> blk_rq_map_sg
> scsi_init_sgtable
> scsi_init_io
> scsi_setup_blk_pc_cmnd
> sd_prep_fn
> blk_peek_request
> scsi_request_fn
> __blk_run_queue
> blk_run_queue
> scsi_run_queue
> scsi_next_command
> scsi_io_completion
> scsi_finish_command
> scsi_softirq_done
> blk_done_softirq
> __do_softirq
> irq_exit
> do_IRQ
> common_interrupt
> <EOI>
> cpuidle_idle_call
> arch_cpu_idle
> cpu_startup_entry
> start_secondary
> 
> It's GPF'ing on struct scatter_list *sg 0x800000001473e064 in
> 
> static inline void sg_assign_page(struct scatterlist *sg, struct page *page)
> {
> 	unsigned long page_link = sg->page_link & 0x3;
> 
> It appears to be in the static inline __blk_segment_map_sg(),
> and that GPF'ing address is what it just got from sg_next().
> 
> Sorry, this isn't the kind of dump you'll be used to, but it's the
> best I can do at the moment, and I've just had to reboot the machine.
> 
> O, tried again and it hit the BUG_ON(count > sdb->table.nents)
> on line 1048 of drivers/scsi/scsi_lib.c:
> 
> scsi_init_sgtable
> <IRQ> scsi_init_io
> scsi_setup_blk_pc_cmnd
> sd_setup_discard_cmnd
> sd_prep_fn
> blk_peek_request
> etc. as before
> 
> I'll have to leave the machine shortly - I'm rather hoping
> you can do your own discard testing to see such crashes.

My simple hdparm/scsi_debug based test isn't hitting it - any suggestions on how
to reproduce it?