On Wed, Sep 16, 2020 at 09:45:10AM +0100, Christoph Hellwig wrote:
> On Tue, Sep 15, 2020 at 05:12:42PM -0700, Darrick J. Wong wrote:
> > On Tue, Aug 25, 2020 at 10:49:17AM -0400, Brian Foster wrote:
> > > cc Ming
> > >
> > > On Tue, Aug 25, 2020 at 10:42:03AM +1000, Dave Chinner wrote:
> > > > On Mon, Aug 24, 2020 at 11:48:41AM -0400, Brian Foster wrote:
> > > > > On Mon, Aug 24, 2020 at 04:04:17PM +0100, Christoph Hellwig wrote:
> > > > > > On Mon, Aug 24, 2020 at 10:28:23AM -0400, Brian Foster wrote:
> > > > > > > Do I understand the current code (__bio_try_merge_page() ->
> > > > > > > page_is_mergeable()) correctly in that we're checking for physical page
> > > > > > > contiguity and not necessarily requiring a new bio_vec per physical
> > > > > > > page?
> > > > > >
> > > > > >
> > > > > > Yes.
> > > > > >
> > > > >
> > > > > Ok. I also realize now that this occurs on a kernel without commit
> > > > > 07173c3ec276 ("block: enable multipage bvecs"). That is probably a
> > > > > contributing factor, but it's not clear to me whether it's feasible to
> > > > > backport whatever supporting infrastructure is required for that
> > > > > mechanism to work (I suspect not).
> > > > >
> > > > > > > With regard to Dave's earlier point around seeing excessively sized bio
> > > > > > > chains.. If I set up a large memory box with high dirty mem ratios and
> > > > > > > do contiguous buffered overwrites over a 32GB range followed by fsync, I
> > > > > > > can see upwards of 1GB per bio and thus chains on the order of 32+ bios
> > > > > > > for the entire write. If I play games with how the buffered overwrite is
> > > > > > > submitted (i.e., in reverse) however, then I can occasionally reproduce
> > > > > > > a ~32GB chain of ~32k bios, which I think is what leads to problems in
> > > > > > > I/O completion on some systems. Granted, I don't reproduce soft lockup
> > > > > > > issues on my system with that behavior, so perhaps there's more to that
> > > > > > > particular issue.
> > > > > > >
> > > > > > > Regardless, it seems reasonable to me to at least have a conservative
> > > > > > > limit on the length of an ioend bio chain. Would anybody object to
> > > > > > > iomap_ioend growing a chain counter and perhaps forcing into a new ioend
> > > > > > > if we chain something like more than 1k bios at once?
> > > > > >
> > > > > > So what exactly is the problem of processing a long chain in the
> > > > > > workqueue vs multiple small chains? Maybe we need a cond_resched()
> > > > > > here and there, but I don't see how we'd substantially change behavior.
> > > > > >
> > > > >
> > > > > The immediate problem is a watchdog lockup detection in bio completion:
> > > > >
> > > > > NMI watchdog: Watchdog detected hard LOCKUP on cpu 25
> > > > >
> > > > > This effectively lands at the following segment of iomap_finish_ioend():
> > > > >
> > > > > ...
> > > > > /* walk each page on bio, ending page IO on them */
> > > > > bio_for_each_segment_all(bv, bio, iter_all)
> > > > > iomap_finish_page_writeback(inode, bv->bv_page, error);
> > > > >
> > > > > I suppose we could add a cond_resched(), but is that safe directly
> > > > > inside of a ->bi_end_io() handler? Another option could be to dump large
> > > > > chains into the completion workqueue, but we may still need to track the
> > > > > length to do that. Thoughts?
> > > >
> > > > We have ioend completion merging that will run the compeltion once
> > > > for all the pending ioend completions on that inode. IOWs, we do not
> > > > need to build huge chains at submission time to batch up completions
> > > > efficiently. However, huge bio chains at submission time do cause
> > > > issues with writeback fairness, pinning GBs of ram as unreclaimable
> > > > for seconds because they are queued for completion while we are
> > > > still submitting the bio chain and submission is being throttled by
> > > > the block layer writeback throttle, etc. Not to mention the latency
> > > > of stable pages in a situation like this - a mmap() write fault
> > > > could stall for many seconds waiting for a huge bio chain to finish
> > > > submission and run completion processing even when the IO for the
> > > > given page we faulted on was completed before the page fault
> > > > occurred...
> > > >
> > > > Hence I think we really do need to cap the length of the bio
> > > > chains here so that we start completing and ending page writeback on
> > > > large writeback ranges long before the writeback code finishes
> > > > submitting the range it was asked to write back.
> > > >
> > >
> > > Ming pointed out separately that limiting the bio chain itself might not
> > > be enough because with multipage bvecs, we can effectively capture the
> > > same number of pages in much fewer bios. Given that, what do you think
> > > about something like the patch below to limit ioend size? This
> > > effectively limits the number of pages per ioend regardless of whether
> > > in-core state results in a small chain of dense bios or a large chain of
> > > smaller bios, without requiring any new explicit page count tracking.
> > >
> > > Brian
> >
> > Dave was asking on IRC if I was going to pull this patch in. I'm unsure
> > of its status (other than it hasn't been sent as a proper [PATCH]) so I
> > wonder, is this necessary, and if so, can it be cleaned up and
> > submitted?
>
I was waiting on some feedback from a few different angles before
posting a proper patch..
> Maybe it is lost somewhere, but what is the point of this patch?
> What does the magic number try to represent?
>
Dave described the main purpose earlier in this thread [1]. The initial
motivation is that we've had downstream reports of soft lockup problems
in writeback bio completion down in the bio -> bvec loop of
iomap_finish_ioend() that has to finish writeback on each individual
page of insanely large bios and/or chains. We've also had an upstream
reports of a similar problem on linux-xfs [2].
The magic number itself was just pulled out of a hat. I picked it
because it seemed conservative enough to still allow large contiguous
bios (1GB w/ 4k pages) while hopefully preventing I/O completion
problems, but was hoping for some feedback on that bit if the general
approach was acceptable. I was also waiting for some feedback on either
of the two users who reported the problem but I don't think I've heard
back on that yet...
Brian
[1] https://lore.kernel.org/linux-fsdevel/20200821215358.GG7941@xxxxxxxxxxxxxxxxxxx/
[2] https://lore.kernel.org/linux-xfs/alpine.LRH.2.02.2008311513150.7870@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
