On Fri 12-01-18 13:41:33, Sergey Senozhatsky wrote:
> On (01/08/18 19:22), Sergey Senozhatsky wrote:
> [..]
> > > Your changelog is rather modest on the information.
> >
> > fair point!
> >
> > > Could you be more specific on how the problem actually happens how
> > > likely it is?
> >
> > ok. so what we have is
> >
> > slow_path / swap-out page
> > __zram_bvec_write(page)
> > compressed_page = zcomp_compress(page)
> > zs_malloc(compressed_page)
> > // no available zspage found, need to allocate new
> > alloc_zspage()
> > {
> > for (i = 0; i < class->pages_per_zspage; i++)
> > page = alloc_page(gfp);
> > if (!page)
> > return NULL
> > }
> >
> > return -ENOMEM
> > ...
> > printk("Write-error on swap-device...");
> >
> >
> > zspage-s can consist of up to ->pages_per_zspage normal pages.
> > if alloc_page() fails then we can't allocate the entire zspage,
> > so we can't store the swapped out page, so it remains in ram
> > and we don't make any progress. so we try to swap another page
> > and may be do the whole zs_malloc()->alloc_zspage() again, may
> > be not. depending on how bad the OOM situation is there can be
> > few or many "Write-error on swap-device" errors.
> >
> > > And again, I do not think the throttling is an appropriate counter
> > > measure. We do want to print those messages when a critical situation
> > > happens. If we have a fallback then simply do not print at all.
> >
> > sure, but with the ratelimited printk we still print those messages.
> > we just don't print it for every single page we failed to write
> > to the device. the existing error messages can (*sometimes*) be noisy
> > and not very informative - "Write-error on swap-device (%u:%u:%llu)\n";
> > it's not like 1000 of those tell more than 1 or 10.
>
> Michal, does that make sense? with the updated/reworked commit
> message will the patch be good enough?
I am sorry but I didn't get to look into this yet. I still _believe_
that the ratelimit is just papering over a real problem here. So I would
prefer if the real fix was done instead. Maybe that is not that easy
easy, I haven't checked. Maybe I just do not understand the issue here.
--
Michal Hocko
SUSE Labs
