Re: [Bug 192571] zswap + zram enabled BUG

[Dan Streetman <ddstreet@ieee.org>]

On Wed, Jan 18, 2017 at 10:00 PM, Sergey Senozhatsky
<sergey.senozhatsky.work@gmail.com> wrote:
> On (01/18/17 20:36), Dan Streetman wrote:
>> On Wed, Jan 18, 2017 at 8:17 AM, Michal Hocko <mhocko@kernel.org> wrote:
>> > On Wed 18-01-17 10:39:48, Sergey Senozhatsky wrote:
>> >> Cc Dan
>> >>
>> >> On (01/17/17 12:22), Andrew Morton wrote:
>> >> > > https://bugzilla.kernel.org/show_bug.cgi?id=192571
>> >> > >
>> >> > > --- Comment #1 from Gluzskiy Alexandr <sss123next@list.ru> ---
>> >> > > [199961.576604] ------------[ cut here ]------------
>> >> > > [199961.577830] kernel BUG at mm/zswap.c:1108!
>> >>
>> >> zswap didn't manage to decompress the page:
>> >>
>> >> static int zswap_frontswap_load(unsigned type, pgoff_t offset,
>> >>                               struct page *page)
>> >> {
>> >> ...
>> >>       dst = kmap_atomic(page);
>> >>       tfm = *get_cpu_ptr(entry->pool->tfm);
>> >>       ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
>> >>       put_cpu_ptr(entry->pool->tfm);
>> >>       kunmap_atomic(dst);
>> >>       zpool_unmap_handle(entry->pool->zpool, entry->handle);
>> >>       BUG_ON(ret);
>> >>       ^^^^^^^^^^^
>> >
>> > Ugh, why do we even do that? This is not the way how to handle error
>> > situations. AFAIU propagating the error out wouldn't be a big deal
>> > because we would just fallback to regular swap, right?
>>
>> yeah this function definitely should never bug; it's just a callback
>> from the zpool to try to write a page back to the swapcache so the
>> zpool can free a page.  It's definitely ok for it to return an error.
>>
>
> good. Dan, Seth, care to send the patch?

damn, i misread the trace.  I need to stop reading bug emails late at night.

I just sent a patch to change the BUG calls in zswap_writeback_entry()
as those are totally recoverable, but the BUG here is from
zswap_frontswap_load(), and a failure there isn't recoverable.

Unfortunately, if we accepted the page, but now can't recover it, it's
gone and returning error from zswap_frontswap_load() will just cause
page_io.c to go read the swap disk; and what's on there is not the
page we're looking for, it's undefined as we didn't actually ever
write the swap page to disk (unless frontswap_writethrough is
enabled).

Ill have to look at this again in the morning.

> and one more thing... can you take a look at [1]?

yeah, i've been meaning to get to that as well, will send something tomorrow.

>
> [1] https://marc.info/?l=linux-mm&m=147031191906154&w=4
>
>         -ss
>

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>