Re: e827091cb1 "block: merge: get the 1st and last bvec via helpers" broken

[Kent Overstreet]

On Sat, Mar 12, 2016 at 06:36:56PM +0800, Ming Lei wrote:
> Hi Kent,
> 
> On Sat, Mar 12, 2016 at 5:24 PM, Kent Overstreet
> <kent.overstreet@xxxxxxxxx> wrote:
> > On Sat, Mar 12, 2016 at 04:49:41PM +0800, Ming Lei wrote:
> >> On Sat, Mar 12, 2016 at 3:43 PM, Kent Overstreet
> >> <kent.overstreet@xxxxxxxxx> wrote:
> >> > I don't know exactly how it's broken, but with that patch segment counting is
> >> > broken - I'm seeing blk_rq_map_sg() overrun the end of the sgtable.
> >> >
> >> > I suggest reverting it for 4.5...
> >>
> >> Kent, could you share your test case? I'd like to figure out the root cause.
> >
> > xfstest 036 on bcachefs.
> 
> Given bcachefs isn't merged, could you provide one way to reproduce it
> with clean upstream kernel?
> 
> >
> >> BTW, I don't object to revert it given it is close to v4.5 release, but I am
> >> curious how it breaks segment couting.
> >
> > If you want to debug your version (personally I'd just revert to the simpler
> > one), I'd start by having your helper use both methods to calculate the last
> > biovec, and then assert that they're equal.
> >
> > Also make sure you're testing with a sub-page sized blocksize, if filesystem
> > blocksize == page size you're not going to be testing the interesting cases
> 
> I just run xfstests 036 over bcache and md, with block size 1024/2048, with
> xfs/ext4/btrfs, looks the segment counting issue can't be reproduced.
> 
> If the issue can only be reproduced with bcachefs, I suggest we don't revert
> it until the root cause is figured out.

Ming, if blk_rq_map_sg() is overrunning arrays and corrupting memory that's a
bug in your code - this is certainly a bug in the core block layer - and just
because you haven't been able to reproduce it yet does _not_ mean that no one
else will hit it.

I think I know why bcachefs hits it - that particular test is doing DIO writes
of sub page granularity, creating extents that are logically adjacent and
adjacent on disk (because they were written one right after the other), but that
don't get merged because I'm running my tests with data checksumming enabled.
Then, when we go to read that data - with a buffered read, so page granularity -
the read gets fragmented into multiple bios (because there's multiple extents),
where the two bios are adjacent (and pointing to the same page!) and adjacent on
disk - thus when they're issued you get sub page size segments from two different
bios that are able to be merged, which is otherwise a highly unusual situation.

I'm not about to write a test case for you though, it's your job to figure out
how to test your code.

Also note that it is entirely possible that the segment counting itself is
correct with your patch, and the bug is just that segments aren't getting merged
that the segment counting assumed would be. If that is the case then your patch
merely exposed the bug, but your patch still needs to be reverted in the
meantime.