From mboxrd@z Thu Jan 1 00:00:00 1970 From: Johannes Weiner Subject: Re: [PATCH] mm/vmscan: check references from all memcgs for swapbacked memory Date: Wed, 5 Oct 2022 10:04:05 -0400 Message-ID: References: <20221004233446.787056-1-yosryahmed@google.com> Mime-Version: 1.0 Return-path: DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=cmpxchg-org.20210112.gappssmtp.com; s=20210112; h=in-reply-to:content-disposition:mime-version:references:message-id :subject:cc:to:from:date:from:to:cc:subject:date; bh=FXi0XoAQOBYrGesfkJtJr1D/6fSwUi+ciVYcxjp3MmE=; b=mJHIld8cSRSCYaAW4MMM2N5yWGRlbk2nS1p0K7//H9ccXyb44BcbTd5JtB+mm91SHc vj4bKWYfab3HIaX1/tARDmcGIzlgnXluW22totZD2rYJHXsO53Ld0KGtv4vDjlHyvPgc 3h+0Z0GJG709ND7TPCiWA8Jl0OKqaVCJXUu5rLUK7hIrff6saZStgRN01tRn/upSvqRY ZLTSDrP0oEm9eagXypT89rzH+C0lLbywPuE/sIzA157xwrOa2+dK8MGvO2cNgJfGKkGy MIqr4F3cSF0X2KflYeEwEBag+4GX52eT00MimlhMUH8EdhC7N2gXHX8jIa2e/AqbJ7rb zDmg== Content-Disposition: inline In-Reply-To: <20221004233446.787056-1-yosryahmed-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org> List-ID: Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Yosry Ahmed Cc: Andrew Morton , Michal Hocko , Roman Gushchin , Shakeel Butt , Muchun Song , Greg Thelen , David Rientjes , cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, linux-mm-Bw31MaZKKs3YtjvyW6yDsg@public.gmane.org Hi Yosry, On Tue, Oct 04, 2022 at 11:34:46PM +0000, Yosry Ahmed wrote: > During page/folio reclaim, we check folio is referenced using > folio_referenced() to avoid reclaiming folios that have been recently > accessed (hot memory). The ratinale is that this memory is likely to be > accessed soon, and hence reclaiming it will cause a refault. > > For memcg reclaim, we pass in sc->target_mem_cgroup to > folio_referenced(), which means we only check accesses to the folio > from processes in the subtree of the target memcg. This behavior was > originally introduced by commit bed7161a519a ("Memory controller: make > page_referenced() cgroup aware") a long time ago. Back then, refaulted > pages would get charged to the memcg of the process that was faulting them > in. It made sense to only consider accesses coming from processes in the > subtree of target_mem_cgroup. If a page was charged to memcg A but only > being accessed by a sibling memcg B, we would reclaim it if memcg A is > under pressure. memcg B can then fault it back in and get charged for it > appropriately. > > Today, this behavior still makes sense for file pages. However, unlike > file pages, when swapbacked pages are refaulted they are charged to the > memcg that was originally charged for them during swapout. Which > means that if a swapbacked page is charged to memcg A but only used by > memcg B, and we reclaim it when memcg A is under pressure, it would > simply be faulted back in and charged again to memcg A once memcg B > accesses it. In that sense, accesses from all memcgs matter equally when > considering if a swapbacked page/folio is a viable reclaim target. > > Add folio_referenced_memcg() which decides what memcg we should pass to > folio_referenced() based on the folio type, and includes an elaborate > comment about why we should do so. This should help reclaim make better > decision and reduce refaults when reclaiming swapbacked memory that is > used by multiple memcgs. Great observation, and I agree with this change. Just one nitpick: > Signed-off-by: Yosry Ahmed > --- > mm/vmscan.c | 38 ++++++++++++++++++++++++++++++++++---- > 1 file changed, 34 insertions(+), 4 deletions(-) > > diff --git a/mm/vmscan.c b/mm/vmscan.c > index c5a4bff11da6..f9fa0f9287e5 100644 > --- a/mm/vmscan.c > +++ b/mm/vmscan.c > @@ -1443,14 +1443,43 @@ enum folio_references { > FOLIOREF_ACTIVATE, > }; > > +/* What memcg should we pass to folio_referenced()? */ > +static struct mem_cgroup *folio_referenced_memcg(struct folio *folio, > + struct mem_cgroup *target_memcg) > +{ > + /* > + * We check references to folios to make sure we don't reclaim hot > + * folios that are likely to be refaulted soon. We pass a memcg to > + * folio_referenced() to only check references coming from processes in > + * that memcg's subtree. > + * > + * For file folios, we only consider references from processes in the > + * subtree of the target memcg. If a folio is charged to > + * memcg A but is only referenced by processes in memcg B, we reclaim it > + * if memcg A is under pressure. If it is later accessed by memcg B it > + * will be faulted back in and charged to memcg B. For memcg A, this is > + * called memory that should be reclaimed. > + * > + * On the other hand, when swapbacked folios are faulted in, they get > + * charged to the memcg that was originally charged for them at the time > + * of swapping out. This means that if a folio that is charged to > + * memcg A gets swapped out, it will get charged back to A when *any* > + * memcg accesses it. In that sense, we need to consider references from > + * *all* processes when considering whether to reclaim a swapbacked > + * folio. > + */ > + return folio_test_swapbacked(folio) ? NULL : target_memcg; > +} > + > static enum folio_references folio_check_references(struct folio *folio, > struct scan_control *sc) > { > int referenced_ptes, referenced_folio; > unsigned long vm_flags; > + struct mem_cgroup *memcg = folio_referenced_memcg(folio, > + sc->target_mem_cgroup); > > - referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup, > - &vm_flags); > + referenced_ptes = folio_referenced(folio, 1, memcg, &vm_flags); > referenced_folio = folio_test_clear_referenced(folio); > > /* > @@ -2581,6 +2610,7 @@ static void shrink_active_list(unsigned long nr_to_scan, > > while (!list_empty(&l_hold)) { > struct folio *folio; > + struct mem_cgroup *memcg; > > cond_resched(); > folio = lru_to_folio(&l_hold); > @@ -2600,8 +2630,8 @@ static void shrink_active_list(unsigned long nr_to_scan, > } > > /* Referenced or rmap lock contention: rotate */ > - if (folio_referenced(folio, 0, sc->target_mem_cgroup, > - &vm_flags) != 0) { > + memcg = folio_referenced_memcg(folio, sc->target_mem_cgroup); > + if (folio_referenced(folio, 0, memcg, &vm_flags) != 0) { Would you mind moving this to folio_referenced() directly? There is already a comment and branch in there that IMO would extend quite naturally to cover the new exception: /* * If we are reclaiming on behalf of a cgroup, skip * counting on behalf of references from different * cgroups */ if (memcg) { rwc.invalid_vma = invalid_folio_referenced_vma; } That would keep the decision-making and doc in one place. Thanks! Johannnes