| Message ID | 20220416004104.4089743-1-roman.gushchin@linux.dev (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm: do not call add_nr_deferred() with zero deferred | expand |
On 16.04.22 02:41, Roman Gushchin wrote: > add_nr_deferred() is often called with next_deferred equal to 0. > For instance, it's happening under low memory pressure for any > shrinkers with a low number of cached objects. A corresponding trace > looks like: > <...>-619914 [005] .... 467456.345160: mm_shrink_slab_end: \ > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > unused scan count 0 new scan count 0 total_scan 0 \ > last shrinker return val 0 > > <...>-619914 [005] .... 467456.345371: mm_shrink_slab_end: \ > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > unused scan count 0 new scan count 0 total_scan 0 \ > last shrinker return val 0 > > <...>-619914 [005] .... 467456.345380: mm_shrink_slab_end: \ > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 unused \ > scan count 0 new scan count 0 total_scan 0 \ > last shrinker return val 0 > > This lead to unnecessary checks and atomic operations, which can be > avoided by checking next_deferred for not being zero before calling > add_nr_deferred(). In this case the mm_shrink_slab_end trace point > will get a potentially slightly outdated "new scan count" value, but > it's totally fine. Sufficient improvement to justify added complexity for anybody reading that code? > > Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev> > --- > mm/vmscan.c | 5 ++++- > 1 file changed, 4 insertions(+), 1 deletion(-) > > diff --git a/mm/vmscan.c b/mm/vmscan.c > index d4a7d2bd276d..19d3d4fa1aad 100644 > --- a/mm/vmscan.c > +++ b/mm/vmscan.c > @@ -808,7 +808,10 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, > * move the unused scan count back into the shrinker in a > * manner that handles concurrent updates. > */ > - new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); > + if (next_deferred) > + new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); > + else > + new_nr = nr; > > trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); > return freed; And if we still want to do this optimization, why not put it into add_nr_deferred()?
On Tue, Apr 19, 2022 at 02:56:06PM +0200, David Hildenbrand wrote: > On 16.04.22 02:41, Roman Gushchin wrote: > > add_nr_deferred() is often called with next_deferred equal to 0. > > For instance, it's happening under low memory pressure for any > > shrinkers with a low number of cached objects. A corresponding trace > > looks like: > > <...>-619914 [005] .... 467456.345160: mm_shrink_slab_end: \ > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > > unused scan count 0 new scan count 0 total_scan 0 \ > > last shrinker return val 0 > > > > <...>-619914 [005] .... 467456.345371: mm_shrink_slab_end: \ > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > > unused scan count 0 new scan count 0 total_scan 0 \ > > last shrinker return val 0 > > > > <...>-619914 [005] .... 467456.345380: mm_shrink_slab_end: \ > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 unused \ > > scan count 0 new scan count 0 total_scan 0 \ > > last shrinker return val 0 > > > > This lead to unnecessary checks and atomic operations, which can be > > avoided by checking next_deferred for not being zero before calling > > add_nr_deferred(). In this case the mm_shrink_slab_end trace point > > will get a potentially slightly outdated "new scan count" value, but > > it's totally fine. > > Sufficient improvement to justify added complexity for anybody reading > that code? I don't have any numbers and really doubt the difference is significant, however the added complexity is also small: one "if" statement. Anyway, if you feel strongly against this change, I'm fine with dropping it. > > > > > Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev> > > --- > > mm/vmscan.c | 5 ++++- > > 1 file changed, 4 insertions(+), 1 deletion(-) > > > > diff --git a/mm/vmscan.c b/mm/vmscan.c > > index d4a7d2bd276d..19d3d4fa1aad 100644 > > --- a/mm/vmscan.c > > +++ b/mm/vmscan.c > > @@ -808,7 +808,10 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, > > * move the unused scan count back into the shrinker in a > > * manner that handles concurrent updates. > > */ > > - new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); > > + if (next_deferred) > > + new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); > > + else > > + new_nr = nr; > > > > trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); > > return freed; > > And if we still want to do this optimization, why not put it into > add_nr_deferred()? Because of the semantics of add_nr_deferred(), which returns the deferred value. It's not used for anything except tracing, so maybe it's a place for another change. Thanks!
On 19.04.22 18:42, Roman Gushchin wrote: > On Tue, Apr 19, 2022 at 02:56:06PM +0200, David Hildenbrand wrote: >> On 16.04.22 02:41, Roman Gushchin wrote: >>> add_nr_deferred() is often called with next_deferred equal to 0. >>> For instance, it's happening under low memory pressure for any >>> shrinkers with a low number of cached objects. A corresponding trace >>> looks like: >>> <...>-619914 [005] .... 467456.345160: mm_shrink_slab_end: \ >>> super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ >>> unused scan count 0 new scan count 0 total_scan 0 \ >>> last shrinker return val 0 >>> >>> <...>-619914 [005] .... 467456.345371: mm_shrink_slab_end: \ >>> super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ >>> unused scan count 0 new scan count 0 total_scan 0 \ >>> last shrinker return val 0 >>> >>> <...>-619914 [005] .... 467456.345380: mm_shrink_slab_end: \ >>> super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 unused \ >>> scan count 0 new scan count 0 total_scan 0 \ >>> last shrinker return val 0 >>> >>> This lead to unnecessary checks and atomic operations, which can be >>> avoided by checking next_deferred for not being zero before calling >>> add_nr_deferred(). In this case the mm_shrink_slab_end trace point >>> will get a potentially slightly outdated "new scan count" value, but >>> it's totally fine. >> >> Sufficient improvement to justify added complexity for anybody reading >> that code? > > I don't have any numbers and really doubt the difference is significant, > however the added complexity is also small: one "if" statement. > Anyway, if you feel strongly against this change, I'm fine with dropping it. > No strong opinion, naturally, more conditions make the code harder to read -- that's why I'm asking. >> >>> >>> Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev> >>> --- >>> mm/vmscan.c | 5 ++++- >>> 1 file changed, 4 insertions(+), 1 deletion(-) >>> >>> diff --git a/mm/vmscan.c b/mm/vmscan.c >>> index d4a7d2bd276d..19d3d4fa1aad 100644 >>> --- a/mm/vmscan.c >>> +++ b/mm/vmscan.c >>> @@ -808,7 +808,10 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, >>> * move the unused scan count back into the shrinker in a >>> * manner that handles concurrent updates. >>> */ >>> - new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); >>> + if (next_deferred) >>> + new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); >>> + else >>> + new_nr = nr; >>> >>> trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); >>> return freed; >> >> And if we still want to do this optimization, why not put it into >> add_nr_deferred()? > > Because of the semantics of add_nr_deferred(), which returns the deferred value. > It's not used for anything except tracing, so maybe it's a place for another > change. Skimming over the code I somehow missed that add_nr_deferred() doesn't have "nr" naturally available. LGTM Acked-by: David Hildenbrand <david@redhat.com>
On Tue, Apr 19, 2022 at 9:57 AM David Hildenbrand <david@redhat.com> wrote: > > On 19.04.22 18:42, Roman Gushchin wrote: > > On Tue, Apr 19, 2022 at 02:56:06PM +0200, David Hildenbrand wrote: > >> On 16.04.22 02:41, Roman Gushchin wrote: > >>> add_nr_deferred() is often called with next_deferred equal to 0. > >>> For instance, it's happening under low memory pressure for any > >>> shrinkers with a low number of cached objects. A corresponding trace > >>> looks like: > >>> <...>-619914 [005] .... 467456.345160: mm_shrink_slab_end: \ > >>> super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > >>> unused scan count 0 new scan count 0 total_scan 0 \ > >>> last shrinker return val 0 > >>> > >>> <...>-619914 [005] .... 467456.345371: mm_shrink_slab_end: \ > >>> super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > >>> unused scan count 0 new scan count 0 total_scan 0 \ > >>> last shrinker return val 0 > >>> > >>> <...>-619914 [005] .... 467456.345380: mm_shrink_slab_end: \ > >>> super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 unused \ > >>> scan count 0 new scan count 0 total_scan 0 \ > >>> last shrinker return val 0 > >>> > >>> This lead to unnecessary checks and atomic operations, which can be > >>> avoided by checking next_deferred for not being zero before calling > >>> add_nr_deferred(). In this case the mm_shrink_slab_end trace point > >>> will get a potentially slightly outdated "new scan count" value, but > >>> it's totally fine. > >> > >> Sufficient improvement to justify added complexity for anybody reading > >> that code? > > > > I don't have any numbers and really doubt the difference is significant, > > however the added complexity is also small: one "if" statement. > > Anyway, if you feel strongly against this change, I'm fine with dropping it. > > > > No strong opinion, naturally, more conditions make the code harder to > read -- that's why I'm asking. This is why that "if" was removed by commit 867508304685 ("mm: vmscan: use per memcg nr_deferred of shrinker") since it didn't bring in measurable performance improvement. TBH I'm not sure whether it is worth it with no measurable performance boost but harder to read code and potential outdated "new scan count". > > >> > >>> > >>> Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev> > >>> --- > >>> mm/vmscan.c | 5 ++++- > >>> 1 file changed, 4 insertions(+), 1 deletion(-) > >>> > >>> diff --git a/mm/vmscan.c b/mm/vmscan.c > >>> index d4a7d2bd276d..19d3d4fa1aad 100644 > >>> --- a/mm/vmscan.c > >>> +++ b/mm/vmscan.c > >>> @@ -808,7 +808,10 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, > >>> * move the unused scan count back into the shrinker in a > >>> * manner that handles concurrent updates. > >>> */ > >>> - new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); > >>> + if (next_deferred) > >>> + new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); > >>> + else > >>> + new_nr = nr; > >>> > >>> trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); > >>> return freed; > >> > >> And if we still want to do this optimization, why not put it into > >> add_nr_deferred()? > > > > Because of the semantics of add_nr_deferred(), which returns the deferred value. > > It's not used for anything except tracing, so maybe it's a place for another > > change. > > Skimming over the code I somehow missed that add_nr_deferred() doesn't > have "nr" naturally available. > > LGTM > > Acked-by: David Hildenbrand <david@redhat.com> > > > -- > Thanks, > > David / dhildenb >
On Tue, Apr 19, 2022 at 09:42:30AM -0700, Roman Gushchin wrote: > On Tue, Apr 19, 2022 at 02:56:06PM +0200, David Hildenbrand wrote: > > On 16.04.22 02:41, Roman Gushchin wrote: > > > add_nr_deferred() is often called with next_deferred equal to 0. > > > For instance, it's happening under low memory pressure for any > > > shrinkers with a low number of cached objects. A corresponding trace > > > looks like: > > > <...>-619914 [005] .... 467456.345160: mm_shrink_slab_end: \ > > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > > > unused scan count 0 new scan count 0 total_scan 0 \ > > > last shrinker return val 0 > > > > > > <...>-619914 [005] .... 467456.345371: mm_shrink_slab_end: \ > > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > > > unused scan count 0 new scan count 0 total_scan 0 \ > > > last shrinker return val 0 > > > > > > <...>-619914 [005] .... 467456.345380: mm_shrink_slab_end: \ > > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 unused \ > > > scan count 0 new scan count 0 total_scan 0 \ > > > last shrinker return val 0 > > > > > > This lead to unnecessary checks and atomic operations, which can be > > > avoided by checking next_deferred for not being zero before calling > > > add_nr_deferred(). In this case the mm_shrink_slab_end trace point > > > will get a potentially slightly outdated "new scan count" value, but > > > it's totally fine. > > > > Sufficient improvement to justify added complexity for anybody reading > > that code? > > I don't have any numbers and really doubt the difference is significant, Never been able to measure it myself. HwoeverI'd much prefer the tracepoint output stays accurate - I've had to post-process and/or graph the shrinker progress as reported by the start/end tracpoints to find problems in the algorithms in the past. That's why there is the additional complexity in the code to make sure the coutners are accurate in the first place. > however the added complexity is also small: one "if" statement. Yeah, complexity is not the problem here - it's that accuracy of the tracepoints has actually mattered to me in the past... Cheers, DAve.
On Fri, Apr 22, 2022 at 11:19:05AM +1000, Dave Chinner wrote: > On Tue, Apr 19, 2022 at 09:42:30AM -0700, Roman Gushchin wrote: > > On Tue, Apr 19, 2022 at 02:56:06PM +0200, David Hildenbrand wrote: > > > On 16.04.22 02:41, Roman Gushchin wrote: > > > > add_nr_deferred() is often called with next_deferred equal to 0. > > > > For instance, it's happening under low memory pressure for any > > > > shrinkers with a low number of cached objects. A corresponding trace > > > > looks like: > > > > <...>-619914 [005] .... 467456.345160: mm_shrink_slab_end: \ > > > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > > > > unused scan count 0 new scan count 0 total_scan 0 \ > > > > last shrinker return val 0 > > > > > > > > <...>-619914 [005] .... 467456.345371: mm_shrink_slab_end: \ > > > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ > > > > unused scan count 0 new scan count 0 total_scan 0 \ > > > > last shrinker return val 0 > > > > > > > > <...>-619914 [005] .... 467456.345380: mm_shrink_slab_end: \ > > > > super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 unused \ > > > > scan count 0 new scan count 0 total_scan 0 \ > > > > last shrinker return val 0 > > > > > > > > This lead to unnecessary checks and atomic operations, which can be > > > > avoided by checking next_deferred for not being zero before calling > > > > add_nr_deferred(). In this case the mm_shrink_slab_end trace point > > > > will get a potentially slightly outdated "new scan count" value, but > > > > it's totally fine. > > > > > > Sufficient improvement to justify added complexity for anybody reading > > > that code? > > > > I don't have any numbers and really doubt the difference is significant, > > Never been able to measure it myself. > > HwoeverI'd much prefer the tracepoint output stays accurate - I've had to > post-process and/or graph the shrinker progress as reported by the > start/end tracpoints to find problems in the algorithms in the past. > That's why there is the additional complexity in the code to make > sure the coutners are accurate in the first place. Sure, no problems. Andrew, can you, please, drop this patch? Thanks!
diff --git a/mm/vmscan.c b/mm/vmscan.c index d4a7d2bd276d..19d3d4fa1aad 100644 --- a/mm/vmscan.c +++ b/mm/vmscan.c @@ -808,7 +808,10 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, * move the unused scan count back into the shrinker in a * manner that handles concurrent updates. */ - new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); + if (next_deferred) + new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); + else + new_nr = nr; trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); return freed;
add_nr_deferred() is often called with next_deferred equal to 0. For instance, it's happening under low memory pressure for any shrinkers with a low number of cached objects. A corresponding trace looks like: <...>-619914 [005] .... 467456.345160: mm_shrink_slab_end: \ super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ unused scan count 0 new scan count 0 total_scan 0 \ last shrinker return val 0 <...>-619914 [005] .... 467456.345371: mm_shrink_slab_end: \ super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 \ unused scan count 0 new scan count 0 total_scan 0 \ last shrinker return val 0 <...>-619914 [005] .... 467456.345380: mm_shrink_slab_end: \ super_cache_scan+0x0/0x1a0 0000000087027f06: nid: 1 unused \ scan count 0 new scan count 0 total_scan 0 \ last shrinker return val 0 This lead to unnecessary checks and atomic operations, which can be avoided by checking next_deferred for not being zero before calling add_nr_deferred(). In this case the mm_shrink_slab_end trace point will get a potentially slightly outdated "new scan count" value, but it's totally fine. Signed-off-by: Roman Gushchin <roman.gushchin@linux.dev> --- mm/vmscan.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)