Re: IO errors after "block: remove bio_get_nr_vecs()"

[Kent Overstreet <kent.overstreet@gmail.com>]

On Tue, Dec 22, 2015 at 10:59:09AM +0500, Artem S. Tashkinov wrote:
> On 2015-12-22 10:55, Kent Overstreet wrote:
> >On Tue, Dec 22, 2015 at 10:52:37AM +0500, Artem S. Tashkinov wrote:
> >>On 2015-12-22 10:38, Kent Overstreet wrote:
> >>>On Tue, Dec 22, 2015 at 05:26:12AM +0000, Junichi Nomura wrote:
> >>>>On 12/22/15 12:59, Kent Overstreet wrote:
> >>>>> reproduced it with 32 bit pae:
> >>>>>
> >>>>>> 1. Exclude memory above 4G line with boot param "max_addr=4G".
> >>>>>
> >>>>> doesn't work - max_addr=1G doesn't work either
> >>>>>
> >>>>>> 2. Disable highmem with "highmem=0".
> >>>>>
> >>>>> works!
> >>>>>
> >>>>>> 3. Try booting 64bit kernel.
> >>>>>
> >>>>> works
> >>>>
> >>>>blk_queue_bio() does split then bounce, which makes the segment
> >>>>counting based on pages before bouncing and could go wrong.
> >>>>
> >>>>What do you think of a patch like this?
> >>>
> >>>Artem, can you give this patch a try?
> >>
> >>
> >>This patch ostensibly fixes the issue - at least I cannot immediately
> >>reproduce it. You can count me in as "Tested-by: Artem S. Tashkinov"
> >
> >Let's all contemplate the fact that blk_segment_map_sg() _overrunning the
> >end of
> >the provided sglist_ was this much of a clusterfuck to debug.
> 
> From the look of it this fix has nothing to do with PAE, so then why only
> PAE users like me were affected by the original
> (b54ffb73cadcdcff9cc1ae0e11f502407e3e2e4c) patch?

The amusing thing is that I doubt PAE actually requires bouncing - addressing
limits come from the device, not the cpu.

But evidently in PAE mode, the block layer is in fact bouncing bios. Probably
from some default setting in the queue limits that no one ever looks at.

The whole queue limits design is an atrocity, it leads to exactly this kind of
crap where no one can predict the actual behaviour of any given setup.