On Tue, 6 Oct 2015, Mike Snitzer wrote: > On Tue, Oct 06 2015 at 9:28am -0400, > Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote: > > > > > > > On Mon, 5 Oct 2015, Mike Snitzer wrote: > > > > > Mikulas, > > > > > > Could it be that cond_resched() wasn't unplugging? As was > > > recently raised in this thread: https://lkml.org/lkml/2015/9/18/378 > > > Chris Mason's patch from that thread fixed this issue... I _think_ Linus > > > has since committed Chris' work but I haven't kept my finger on the > > > pulse of that issue. > > > > I think it doesn't matter (regarding correctness) if cond_reched unplugs > > on not. If it didn't unplug, the process will be scheduled later, and it > > will eventually reach the point where it unplugs. > > Couldn't the original deadlock you fixed (as described in your first > patch) manifest when a new process is scheduled? > > > > FYI, I've put rebased versions of your 2 patches in my wip branch, see: > > > http://git.kernel.org/cgit/linux/kernel/git/snitzer/linux.git/log/?h=wip > > > > > > I tweaked the 2nd patch that adds bio_list to plug so that > > > generic_make_request's checks for in_generic_make_request isn't racey > > > (your original patch could happen to have current-plug set but > > > in_generic_make_request not yet set). > > > > I don't recommend that second patch > > (http://git.kernel.org/cgit/linux/kernel/git/snitzer/linux.git/commit/?h=wip&id=5e740c2e45a767d8d6ef8ca36b0db705ef6259c4). > > The patch just complicates things without adding any value. It's also not > > correct because it plugs bios at places when bios aren't supposed to be > > plugged > > Avoiding another hook the the scheduler is a requirement (from Jens). > "without adding any value": it offers a different strategy for recording > bios to the bio_list by making it part of the plug. The plug doesn't > actually block bios like requests are plugged. Unfortunatelly, it does (when the second patch is applied). > What am I missing? Bios allocated with bio_kmalloc can't be unplugged because they lack the rescurer workqueue (they shouldn't be used by the stacking drivers, they could only be used by top-level mm code). If you plug those bios at incorrect places (i.e. between blk_start_plug and blk_finish_plug), you are introducing other deadlock possibility. > Avoiding another hook the the scheduler is a requirement (from Jens). So, you can add a flag that is set when either current->plug or current->bio_list is non-NULL and if the flag is set, run a function that unplugs both current->plug and current->bio_list. But anyway - if you look at struct task_struct, you see that "bio_list" and "plug" fields are next to each other - so they will likely be in the same cacheline. So testing current->bio_list in the schedule function has no performance overhead because the cacheline was already loaded when testing current->plug. Mikulas -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
