Message-ID: <SJ0PR11MB5678EA6E0A41B0C83CB7F8F3C9692@SJ0PR11MB5678.namprd11.prod.outlook.com>
Date: Wed, 25 Sep 2024 01:49:03 +0000
From: "Sridhar, Kanchana P" <kanchana.p.sridhar@...el.com>
To: Yosry Ahmed <yosryahmed@...gle.com>
CC: "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>, "hannes@...xchg.org"
	<hannes@...xchg.org>, "nphamcs@...il.com" <nphamcs@...il.com>,
	"chengming.zhou@...ux.dev" <chengming.zhou@...ux.dev>,
	"usamaarif642@...il.com" <usamaarif642@...il.com>, "shakeel.butt@...ux.dev"
	<shakeel.butt@...ux.dev>, "ryan.roberts@....com" <ryan.roberts@....com>,
	"Huang, Ying" <ying.huang@...el.com>, "21cnbao@...il.com"
	<21cnbao@...il.com>, "akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
	"Zou, Nanhai" <nanhai.zou@...el.com>, "Feghali, Wajdi K"
	<wajdi.k.feghali@...el.com>, "Gopal, Vinodh" <vinodh.gopal@...el.com>,
	"Sridhar, Kanchana P" <kanchana.p.sridhar@...el.com>
Subject: RE: [PATCH v7 5/8] mm: zswap: Compress and store a specific page in a
 folio.

> -----Original Message-----
> From: Yosry Ahmed <yosryahmed@...gle.com>
> Sent: Tuesday, September 24, 2024 5:47 PM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@...el.com>
> Cc: linux-kernel@...r.kernel.org; linux-mm@...ck.org;
> hannes@...xchg.org; nphamcs@...il.com; chengming.zhou@...ux.dev;
> usamaarif642@...il.com; shakeel.butt@...ux.dev; ryan.roberts@....com;
> Huang, Ying <ying.huang@...el.com>; 21cnbao@...il.com; akpm@...ux-
> foundation.org; Zou, Nanhai <nanhai.zou@...el.com>; Feghali, Wajdi K
> <wajdi.k.feghali@...el.com>; Gopal, Vinodh <vinodh.gopal@...el.com>
> Subject: Re: [PATCH v7 5/8] mm: zswap: Compress and store a specific page
> in a folio.
> 
> [..]
> > >
> > > > +{
> > > > +       swp_entry_t swp = folio->swap;
> > > > +       int type = swp_type(swp);
> > > > +       pgoff_t offset = swp_offset(swp) + index;
> > > > +       struct page *page = folio_page(folio, index);
> > > > +       struct xarray *tree = swap_zswap_tree(swp);
> > > > +       struct zswap_entry *entry;
> > > > +
> > > > +       if (objcg)
> > > > +               obj_cgroup_get(objcg);
> > > > +
> > > > +       if (zswap_check_limits())
> > > > +               goto reject;
> > > > +
> > > > +       /* allocate entry */
> > > > +       entry = zswap_entry_cache_alloc(GFP_KERNEL, folio_nid(folio));
> > > > +       if (!entry) {
> > > > +               zswap_reject_kmemcache_fail++;
> > > > +               goto reject;
> > > > +       }
> > > > +
> > > > +       /* if entry is successfully added, it keeps the reference */
> > > > +       if (!zswap_pool_get(pool))
> > > > +               goto freepage;
> > >
> > > I think we can batch this for all pages in zswap_store(), maybe first
> > > add zswap_pool_get_many().
> > >
> > > I am also wondering if it would be better to batch the limit checking
> > > and allocating the entries, to front load any failures before we start
> > > compression. Not sure if that's overall better though.
> > >
> > > To batch allocate entries we will have to also allocate an array to
> > > hold them. To batch the limit checking we will have to either allow
> > > going further over limit for mTHPs, or check if there is enough
> > > clearance to allow for compressing all the pages. Using the
> > > uncompressed size will lead to false negatives though, so maybe we can
> > > start tracking the average compression ratio for better limit
> > > checking.
> > >
> > > Nhat, Johannes, any thoughts here? I need someone to tell me if I am
> > > overthinking this :)
> >
> > These are all good points. I suppose I was thinking along the same lines
> > of what Nhat mentioned in an earlier comment. I was trying the
> > incremental zswap_pool_get() and limit checks and shrinker invocations
> > in case of (all) error conditions to allow different concurrent stores to make
> > progress, without favoring only one process's mTHP store. I was thinking
> > this would have minimal impact on the process(es) that see the zswap
> > limit being exceeded, and that this would be better than preemptively
> > checking for the entire mTHP and failing (this could also complicate things
> > where no one makes progress because multiple processes run the batch
> > checks and fail, when realistically one/many could have triggered
> > the shrinker before erroring out, and at least one could have made
> > progress).
> 
> On the other hand, if we allow concurrent mTHP swapouts to do limit
> checks incrementally, they may all fail at the last page. While if
> they all do limit checks beforehand, one of them can proceed.

Yes, this is possible too. Although, given the dynamic nature of the usage,
even with a check-before-store strategy for mTHP we could end up in a
similar situation as the optimistic approach in which we allowed progress
until there really was a reason to fail. 

> 
> I think we need to agree on a higher-level strategy for limit
> checking, both global and per-memcg. The per-memcg limit should be
> stricter though, so we may end up having different policies.

Sure, this makes sense. One possibility is we could allow zswap to
follow the "optimistic approach" used currently, while we manage
the limits checking at the memcg level? Something along the lines
of mem_cgroup_handle_over_high() that gets called every time after
a page-fault is handled; instead checks the cgroup's zswap usage and
triggers writeback? This seems like one way of not adding overhead to
the reclaim path (zswap will store mTHP until the limit checking causes
error and unwinding state), while triggering zswap-LRU based writeback
at a higher level to manage the limit.

> 
> >
> > Would appreciate your perspectives on how this should be handled,
> > and will implement a solution in v8 accordingly.
> >
> > Thanks,
> > Kanchana
> >
> > >
> > > > +
> > > > +       entry->pool = pool;
> > > > +
> > > > +       if (!zswap_compress(page, entry))
> > > > +               goto put_pool;
> > > > +
> > > > +       entry->swpentry = swp_entry(type, offset);
> > > > +       entry->objcg = objcg;
> > > > +       entry->referenced = true;
> > > > +
> > > > +       if (!zswap_store_entry(tree, entry))
> > > > +               goto store_failed;
> > > > +
> > > > +       if (objcg) {
> > > > +               obj_cgroup_charge_zswap(objcg, entry->length);
> > > > +               count_objcg_event(objcg, ZSWPOUT);
> > > > +       }
> > > > +
> > > > +       /*
> > > > +        * We finish initializing the entry while it's already in xarray.
> > > > +        * This is safe because:
> > > > +        *
> > > > +        * 1. Concurrent stores and invalidations are excluded by folio lock.
> > > > +        *
> > > > +        * 2. Writeback is excluded by the entry not being on the LRU yet.
> > > > +        *    The publishing order matters to prevent writeback from seeing
> > > > +        *    an incoherent entry.
> > > > +        */
> > > > +       if (entry->length) {
> > > > +               INIT_LIST_HEAD(&entry->lru);
> > > > +               zswap_lru_add(&zswap_list_lru, entry);
> > > > +       }
> > > > +
> > > > +       /* update stats */
> > > > +       atomic_inc(&zswap_stored_pages);
> > > > +       count_vm_event(ZSWPOUT);
> > >
> > > We should probably also batch updating the stats. It actually seems
> > > like now we don't handle rolling them back upon failure.
> >
> > Good point! I assume you are referring only to the "ZSWPOUT" vm event
> stats
> > updates and not the "zswap_stored_pages" (since latter is used in limit
> checking)?
> 
> I actually meant both. Do we rollback changes to zswap_stored_pages
> when some stores succeed and some of them fail?

Yes we do. zswap_tree_delete() calls zswap_entry_free() which will
decrement zswap_stored_pages. The only stat that is left in an incorrect
state in this case is the vmstat 'zswpout'.

> 
> I think it's more correct and efficient to update the atomic once
> after all the pages are successfully compressed and stored.

Actually this would need to co-relate with the limits checking strategy,
because the atomic is used there and needs to be as accurate as possible.

As far as the vmstat 'zswpout', the reason I left it as-is in my patchset
was to be more indicative of the actual zswpout compute events that
occurred (for things like getting the compressions count), regardless
of whether or not the overall mTHP store was successful. If this vmstat
needs to reflect only successful zswpout events (i.e., represent the zswap
usage), I can fix it by updating it once only if the mTHP is stored successfully.

Thanks,
Kanchana

Powered by blists - more mailing lists