Re: 5.3-rc1 regression with XFS log recovery

[Ming Lei <ming.lei@redhat.com>]

On Tue, Aug 20, 2019 at 07:53:20AM +0200, hch@lst.de wrote:
> On Tue, Aug 20, 2019 at 02:41:35PM +1000, Dave Chinner wrote:
> > > With the following debug patch.  Based on that I think I'll just
> > > formally submit the vmalloc switch as we're at -rc5, and then we
> > > can restart the unaligned slub allocation drama..
> > 
> > This still doesn't make sense to me, because the pmem and brd code
> > have no aligment limitations in their make_request code - they can
> > handle byte adressing and should not have any problem at all with
> > 8 byte aligned memory in bios.
> > 
> > Digging a little furhter, I note that both brd and pmem use
> > identical mechanisms to marshall data in and out of bios, so they
> > are likely to have the same issue.
> > 
> > So, brd_make_request() does:
> > 
> >         bio_for_each_segment(bvec, bio, iter) {
> >                 unsigned int len = bvec.bv_len;
> >                 int err;
> > 
> >                 err = brd_do_bvec(brd, bvec.bv_page, len, bvec.bv_offset,
> >                                   bio_op(bio), sector);
> >                 if (err)
> >                         goto io_error;
> >                 sector += len >> SECTOR_SHIFT;
> >         }
> > 
> > So, the code behind bio_for_each_segment() splits multi-page bvecs
> > into individual pages, which are passed to brd_do_bvec(). An
> > unaligned 4kB io traces out as:
> > 
> >  [  121.295550] p,o,l,s 00000000a77f0146,768,3328,0x7d0048
> >  [  121.297635] p,o,l,s 000000006ceca91e,0,768,0x7d004e
> > 
> > i.e. page		offset	len	sector
> > 00000000a77f0146	768	3328	0x7d0048
> > 000000006ceca91e	0	768	0x7d004e
> > 
> > You should be able to guess what the problems are from this.

The problem should be that offset of '768' is passed to bio_add_page().

It should be one slub buffer used for block IO, looks an old unsolved
problem.

> > 
> > Both pmem and brd are _sector_ based. We've done a partial sector
> > copy on the first bvec, then the second bvec has started the copy
> > from the wrong offset into the sector we've done a partial copy
> > from.
> > 
> > IOWs, no error is reported when the bvec buffer isn't sector
> > aligned, no error is reported when the length of data to copy was
> > not a multiple of sector size, and no error was reported when we
> > copied the same partial sector twice.
> 
> Yes.  I think bio_for_each_segment is buggy here, as it should not
> blindly split by pages.

bio_for_each_segment() just keeps the original interface as before
introducing multi-page bvec.


Thanks,
Ming