Re: dm bufio: Reduce dm_bufio_lock contention

From: Mikulas Patocka
Date: Mon Jun 25 2018 - 09:53:41 EST


y

On Mon, 25 Jun 2018, Michal Hocko wrote:

> On Fri 22-06-18 14:57:10, Mikulas Patocka wrote:
> >
> >
> > On Fri, 22 Jun 2018, Michal Hocko wrote:
> >
> > > On Fri 22-06-18 08:52:09, Mikulas Patocka wrote:
> > > >
> > > >
> > > > On Fri, 22 Jun 2018, Michal Hocko wrote:
> > > >
> > > > > On Fri 22-06-18 11:01:51, Michal Hocko wrote:
> > > > > > On Thu 21-06-18 21:17:24, Mikulas Patocka wrote:
> > > > > [...]
> > > > > > > What about this patch? If __GFP_NORETRY and __GFP_FS is not set (i.e. the
> > > > > > > request comes from a block device driver or a filesystem), we should not
> > > > > > > sleep.
> > > > > >
> > > > > > Why? How are you going to audit all the callers that the behavior makes
> > > > > > sense and moreover how are you going to ensure that future usage will
> > > > > > still make sense. The more subtle side effects gfp flags have the harder
> > > > > > they are to maintain.
> > > > >
> > > > > So just as an excercise. Try to explain the above semantic to users. We
> > > > > currently have the following.
> > > > >
> > > > > * __GFP_NORETRY: The VM implementation will try only very lightweight
> > > > > * memory direct reclaim to get some memory under memory pressure (thus
> > > > > * it can sleep). It will avoid disruptive actions like OOM killer. The
> > > > > * caller must handle the failure which is quite likely to happen under
> > > > > * heavy memory pressure. The flag is suitable when failure can easily be
> > > > > * handled at small cost, such as reduced throughput
> > > > >
> > > > > * __GFP_FS can call down to the low-level FS. Clearing the flag avoids the
> > > > > * allocator recursing into the filesystem which might already be holding
> > > > > * locks.
> > > > >
> > > > > So how are you going to explain gfp & (__GFP_NORETRY | ~__GFP_FS)? What
> > > > > is the actual semantic without explaining the whole reclaim or force
> > > > > users to look into the code to understand that? What about GFP_NOIO |
> > > > > __GFP_NORETRY? What does it mean to that "should not sleep". Do all
> > > > > shrinkers have to follow that as well?
> > > >
> > > > My reasoning was that there is broken code that uses __GFP_NORETRY and
> > > > assumes that it can't fail - so conditioning the change on !__GFP_FS would
> > > > minimize the diruption to the broken code.
> > > >
> > > > Anyway - if you want to test only on __GFP_NORETRY (and fix those 16
> > > > broken cases that assume that __GFP_NORETRY can't fail), I'm OK with that.
> > >
> > > As I've already said, this is a subtle change which is really hard to
> > > reason about. Throttling on congestion has its meaning and reason. Look
> > > at why we are doing that in the first place. You cannot simply say this
> >
> > So - explain why is throttling needed. You support throttling, I don't, so
> > you have to explain it :)
> >
> > > is ok based on your specific usecase. We do have means to achieve that.
> > > It is explicit and thus it will be applied only where it makes sense.
> > > You keep repeating that implicit behavior change for everybody is
> > > better.
> >
> > I don't want to change it for everybody. I want to change it for block
> > device drivers. I don't care what you do with non-block drivers.
>
> Well, it is usually onus of the patch submitter to justify any change.
> But let me be nice on you, for once. This throttling is triggered only
> if we all the pages we have encountered during the reclaim attempt are
> dirty and that means that we are rushing through the LRU list quicker
> than flushers are able to clean. If we didn't throttle we could hit
> stronger reclaim priorities (aka scan more to reclaim memory) and
> reclaim more pages as a result.

And the throttling in dm-bufio prevents kswapd from making forward
progress, causing this situation...

> > I'm sure you'll come up with another creative excuse why GFP_NORETRY
> > allocations need incur deliberate 100ms delays in block device drivers.
>
> ... is not really productive. I've tried to explain why I am not _sure_ what
> possible side effects such a change might have and your hand waving
> didn't really convince me. MD is not the only user of the page
> allocator...

But you are just doing that now - you're just coming up with another great
excuse why block device drivers need to sleep 100ms. The system stops to a
crawl when block device requests take 100ms and you - instead of fixing it
- are just arguing how is it needed.

> > > I guess we will not agree on that part. I consider it a hack
> > > rather than a systematic solution. I can easily imagine that we just
> > > find out other call sites that would cause over reclaim or similar
> >
> > If a __GFP_NORETRY allocation does overreclaim - it could be fixed by
> > returning NULL instead of doing overreclaim. The specification says that
> > callers must handle failure of __GFP_NORETRY allocations.
> >
> > So yes - if you think that just skipping throttling on __GFP_NORETRY could
> > cause excessive CPU consumption trying to reclaim unreclaimable pages or
> > something like that - then you can add more points where the __GFP_NORETRY
> > is failed with NULL to avoid the excessive CPU consumption.
>
> Which is exactly something I do not want to do. Spread __GFP_NORETRY all
> over the reclaim code. Especially for something we already have means
> for...

And so what do you want to do to prevent block drivers from sleeping?

Mikulas