[lkml]   [2018]   [Aug]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 1/2] zram: remove BD_CAP_SYNCHRONOUS_IO with writeback feature
On Fri, Aug 03, 2018 at 01:13:02PM +0900, Sergey Senozhatsky wrote:
> Hi Minchan,
> On (08/03/18 12:00), Minchan Kim wrote:
> > > "Device is so fast that asynchronous IO would be inefficient."
> > >
> > > Which is not the reason why BDI_CAP_SYNCHRONOUS_IO is used by ZRAM.
> > > Probably, the comment needs to be updated as well.
> >
> > I couldn't catch your point. Could you clarify a little bit more?
> > What do you want to correct for the comment?
> >
> > > "efficiency" [looking at the comments], but in ZRAM's case the whole
> > > reason to use SYNC IO is a race condition and user-after-free that
> > > follows.
> >
> > Actually, it's not whole reason. As I wrote down, without it, swap_readpage
> > waits the IO completion for a long time by blk_poll so it causes system
> > sluggish problem when device is slow(e.g., zram with backing device).
> Sure, this is problem #1. But slow swap device probably doesn't do any
> irreversible harm to the system. Unlike use-after-free, which does. Thus
> use-after-free is a problem #0. BDI_CAP_SYNCHRONOUS_IO comment doesn't
> mention problem #0; it talks about problem #1 only. So, nothing serious,
> just wanted to point that out.
> So we probably can make ZRAM always ASYNC when WB is enabled.
> Or... maybe we can make swap out to be SYNC and perform WB in background.
> In __zram_bvec_write() we can always write compressed object to zmalloc,
> even the huge ones.
> Things to note:
> a) even when WB is enabled we still allocate huge classes
> b) even when WB is enabled we still may use those huge classes (consider
> a case when backing devices is full)
> So huge classes are still there and we still use them. So let's use
> them?
> For a huge object, after we stored it into zsmalloc, we can schedule a WB
> work, which would:
> a) write that particular object (page) to the backing device
> b) mark entry as WB entry
> c) remove object from zsmalloc, unlock necessary locks
> So swap in should either see object in zsmalloc or on backing device.
> How does this sound?
> And reading from a backing device can always be SYNC. Can it?
> Am I missing something?

AFAIK, onging writeback page couldn't freed so it was not writeabck problem.

What I'm tryig to fix is read part.
If we use swapcache, it shouldn't be a problem either because swapcache
has a reference count and we should wait PG_lock release before the freeing
from the swapcache so there is no race condition.

However, by the skip swapcache logic, we don't have a refcount any longer.

I think we can hold a new refcount in zram driver itself. With that, we
could get both benefits from writeback feature and skip swapcache.

However, I decided, at this moment, going this simple way for
stable-material to solve #0 and #1 problems at the same time.


> -ss

 \ /
  Last update: 2018-08-03 06:52    [W:0.103 / U:5.280 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site