[PATCH] mm: make fault_around_bytes configurable

[Vinayak Menon]

Mapping pages around fault is found to cause performance degradation
in certain use cases. The test performed here is launch of 10 apps
one by one, doing something with the app each time, and then repeating
the same sequence once more, on an ARM 64-bit Android device with 2GB
of RAM. The time taken to launch the apps is found to be better when
fault around feature is disabled by setting fault_around_bytes to page
size (4096 in this case).

The tests were done on 3.18 kernel. 4 extra vmstat counters were added
for debugging. pgpgoutclean accounts the clean pages reclaimed via
__delete_from_page_cache. pageref_activate, pageref_activate_vm_exec,
and pageref_keep accounts the mapped file pages activated and retained
by page_check_references.

=== Without swap ===
3.18 3.18-fault_around_bytes=4096
-----------------------------------------------------------------------
workingset_refault 691100 664339
workingset_activate 210379 179139
pgpgin 4676096 4492780
pgpgout 163967 96711
pgpgoutclean 1090664 990659
pgalloc_dma 3463111 3328299
pgfree 3502365 3363866
pgactivate 568134 238570
pgdeactivate 752260 392138
pageref_activate 315078 121705
pageref_activate_vm_exec 162940 55815
pageref_keep 141354 51011
pgmajfault 24863 23633
pgrefill_dma 1116370 544042
pgscan_kswapd_dma 1735186 1234622
pgsteal_kswapd_dma 1121769 1005725
pgscan_direct_dma 12966 1090
pgsteal_direct_dma 6209 967
slabs_scanned 1539849 977351
pageoutrun 1260 1333
allocstall 47 7

=== With swap ===
3.18 3.18-fault_around_bytes=4096
-----------------------------------------------------------------------
workingset_refault 597687 878109
workingset_activate 167169 254037
pgpgin 4035424 5157348
pgpgout 162151 85231
pgpgoutclean 928587 1225029
pswpin 46033 17100
pswpout 237952 127686
pgalloc_dma 3305034 3542614
pgfree 3354989 3592132
pgactivate 626468 355275
pgdeactivate 990205 771902
pageref_activate 294780 157106
pageref_activate_vm_exec 141722 63469
pageref_keep 121931 63028
pgmajfault 67818 45643
pgrefill_dma 1324023 977192
pgscan_kswapd_dma 1825267 1720322
pgsteal_kswapd_dma 1181882 1365500
pgscan_direct_dma 41957 9622
pgsteal_direct_dma 25136 6759
slabs_scanned 689575 542705
pageoutrun 1234 1538
allocstall 110 26

Looks like with fault_around, there is more pressure on reclaim because
of the presence of more mapped pages, resulting in more IO activity,
more faults, more swapping, and allocstalls.

Make fault_around_bytes configurable so that it can be tuned to avoid
performance degradation.

Signed-off-by: Vinayak Menon <vinm...@codeaurora.org>
---
mm/Kconfig | 10 ++++++++++
mm/memory.c | 2 +-
2 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/mm/Kconfig b/mm/Kconfig
index f644106..e3476fd 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -681,6 +681,16 @@ config ZONE_DEVICE

If FS_DAX is enabled, then say Y.

+config FAULT_AROUND_BYTES
+ int
+ range 4096 65536
+ default 65536
+ help
+ The number of bytes to be mapped around the fault. The default
+ value of 64 kilobytes effectively disables faultaround on
+ architectures with page size >= 64k, considering the fact that
+ the feature is less relevant when page size is bigger than 4k.
+
config FRAME_VECTOR
bool

diff --git a/mm/memory.c b/mm/memory.c
index 758b0b4..be06714 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -2939,7 +2939,7 @@ void do_set_pte(struct vm_area_struct *vma, unsigned long address,
}

static unsigned long fault_around_bytes __read_mostly =
- rounddown_pow_of_two(65536);
+ rounddown_pow_of_two(CONFIG_FAULT_AROUND_BYTES);

#ifdef CONFIG_DEBUG_FS
static int fault_around_bytes_get(void *data, u64 *val)
--
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a
member of the Code Aurora Forum, hosted by The Linux Foundation

[Andrew Morton]

On Mon, 18 Apr 2016 20:47:16 +0530 Vinayak Menon <vinm...@codeaurora.org> wrote:

> Mapping pages around fault is found to cause performance degradation
> in certain use cases. The test performed here is launch of 10 apps
> one by one, doing something with the app each time, and then repeating
> the same sequence once more, on an ARM 64-bit Android device with 2GB
> of RAM. The time taken to launch the apps is found to be better when
> fault around feature is disabled by setting fault_around_bytes to page
> size (4096 in this case).

Well that's one workload, and a somewhat strange one. What is the
effect on other workloads (of which there are a lot!).

A few of those things did get a bit worse?

Do you have any data on actual wall-time changes? How much faster do
things become with the patch? If it is "0.1%" then I'd say "umm, no".

> Make fault_around_bytes configurable so that it can be tuned to avoid
> performance degradation.

It sounds like we need to be smarter about auto-tuning this thing.
Maybe the refault code could be taught to provide the feedback path but
that sounds hard.

Still. I do think it would be better to make this configurable at
runtime. Move the existing debugfs tunable into /proc/sys/vm (and
document it!). I do dislkie adding even more tunables but this one
does make sense. People will want to run their workloads with various
values until they find the peak throughput, and requiring a kernel
rebuild for that is a huge pain.

[Vinayak Menon]

On 04/22/2016 05:31 AM, Andrew Morton wrote:
> On Mon, 18 Apr 2016 20:47:16 +0530 Vinayak Menon <vinm...@codeaurora.org> wrote:
>
>> Mapping pages around fault is found to cause performance degradation
>> in certain use cases. The test performed here is launch of 10 apps
>> one by one, doing something with the app each time, and then repeating
>> the same sequence once more, on an ARM 64-bit Android device with 2GB
>> of RAM. The time taken to launch the apps is found to be better when
>> fault around feature is disabled by setting fault_around_bytes to page
>> size (4096 in this case).
>
> Well that's one workload, and a somewhat strange one. What is the
> effect on other workloads (of which there are a lot!).
>

This workload emulates the way a user would use his mobile device,
opening an application, using it for some time, switching to next, and
then coming back to the same application later. Another stat which shows
significant degradation on Android with fault_around is device boot up
time. I have not tried any other workload other than these.

I think some numbers (like workingset, pgpgin, pgpgoutclean etc) looks
better with fault_around because, increased number of mapped pages is
resulting in less number of file pages being reclaimed
(pageref_activate, pageref_activate_vm_exec, pageref_keep above), but
increased swapping. Latency numbers are far bad with fault_around_bytes
+ swap, possibly because of increased swapping, decrease in kswapd
efficiency and increase in allocstalls.
So the problem looks to be that unwanted pages are mapped around the
fault and page_check_references is unaware of this.

>
> Do you have any data on actual wall-time changes? How much faster do
> things become with the patch? If it is "0.1%" then I'd say "umm, no".
>

=== Without swap ====
3.18 3.18-fault_around_bytes=4096
Avg launch latency 1695ms 1300ms (23.3%)
Max launch latency 5097ms 3135ms (38.49%)

>> Make fault_around_bytes configurable so that it can be tuned to avoid
>> performance degradation.
>
> It sounds like we need to be smarter about auto-tuning this thing.
> Maybe the refault code could be taught to provide the feedback path but
> that sounds hard.
>
> Still. I do think it would be better to make this configurable at
> runtime. Move the existing debugfs tunable into /proc/sys/vm (and
> document it!). I do dislkie adding even more tunables but this one
> does make sense. People will want to run their workloads with various
> values until they find the peak throughput, and requiring a kernel
> rebuild for that is a huge pain.
>

I can send a v2 to do this runtime via /proc/sys/vm.

> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majo...@kvack.org. For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"do...@kvack.org"> em...@kvack.org </a>

[Kirill A. Shutemov]

Hm. It makes me think we should make ptes setup by faultaround old.

Although, it would defeat (to some extend) purpose of faultaround on
architectures without HW accessed bit :-/

Could you check if the patch below changes the situation?
It would require some more work to not mark the pte we've got fault for old.

diff --git a/include/linux/mm.h b/include/linux/mm.h
index a55e5be0894f..1066fabf17c3 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -584,7 +584,7 @@ static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)

}

void do_set_pte(struct vm_area_struct *vma, unsigned long address,

- struct page *page, pte_t *pte, bool write, bool anon);
+ struct page *page, pte_t *pte, bool write, bool anon, bool old);
#endif

/*
diff --git a/mm/filemap.c b/mm/filemap.c
index f2479af09da9..47ba88fd7192 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -2189,7 +2189,7 @@ repeat:
if (file->f_ra.mmap_miss > 0)
file->f_ra.mmap_miss--;
addr = address + (page->index - vmf->pgoff) * PAGE_SIZE;
- do_set_pte(vma, addr, page, pte, false, false);
+ do_set_pte(vma, addr, page, pte, false, false, true);
unlock_page(page);
goto next;
unlock:
diff --git a/mm/memory.c b/mm/memory.c
index 93897f23cc11..fa3ac184eafd 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -2836,7 +2836,7 @@ static int __do_fault(struct vm_area_struct *vma, unsigned long address,
* vm_ops->map_pages.
*/

void do_set_pte(struct vm_area_struct *vma, unsigned long address,

- struct page *page, pte_t *pte, bool write, bool anon)
+ struct page *page, pte_t *pte, bool write, bool anon, bool old)
{
pte_t entry;

@@ -2844,6 +2844,8 @@ void do_set_pte(struct vm_area_struct *vma, unsigned long address,
entry = mk_pte(page, vma->vm_page_prot);
if (write)
entry = maybe_mkwrite(pte_mkdirty(entry), vma);
+ if (old)
+ entry = pte_mkold(entry);
if (anon) {
inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
page_add_new_anon_rmap(page, vma, address, false);
@@ -2998,7 +3000,7 @@ static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
put_page(fault_page);
return ret;
}
- do_set_pte(vma, address, fault_page, pte, false, false);
+ do_set_pte(vma, address, fault_page, pte, false, false, false);
unlock_page(fault_page);
unlock_out:
pte_unmap_unlock(pte, ptl);
@@ -3050,7 +3052,7 @@ static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
}
goto uncharge_out;
}
- do_set_pte(vma, address, new_page, pte, true, true);
+ do_set_pte(vma, address, new_page, pte, true, true, false);
mem_cgroup_commit_charge(new_page, memcg, false, false);
lru_cache_add_active_or_unevictable(new_page, vma);
pte_unmap_unlock(pte, ptl);
@@ -3107,7 +3109,7 @@ static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
put_page(fault_page);
return ret;
}
- do_set_pte(vma, address, fault_page, pte, true, false);
+ do_set_pte(vma, address, fault_page, pte, true, false, false);
pte_unmap_unlock(pte, ptl);

if (set_page_dirty(fault_page))
--
Kirill A. Shutemov

[Minchan Kim]

On Fri, Apr 22, 2016 at 02:15:08PM +0530, Vinayak Menon wrote:

The page_check_references makes difference only when pte has marked access_bit.

enum page_references page_check_references(struct page *page)
{
referenced_ptes = page_referenced(page);
if (referenced_ptes) {
...
return PAGEREF_ACTIVATE
}
}

But map_pages doesn't mark ahead pages as pte_mkyoung. IOW, ptes are already
pte_mkold. So, I think page_check_reference shouldn't make any difference.

Other thing it can make the difference about reclaiming is that it can
make more pressure slab shrinking.

unsigned long shrink_page_list()
{
..
/* Double the slab pressure for mapped and swapcache pages */
if (page_mapped(page) || PageSwapCache(page))
sc->nr_scanned++;
..
}

But I'm not sure it can make such difference.
Could you explain why I am missing?

[Kirill A. Shutemov]

Actually, I've checked and mk_pte() produces young ptes for me. Not sure
why.

--
Kirill A. Shutemov

[Kirill A. Shutemov]

Ah. Okay, _PAGE_ACCESSED included into pgprot mask, which is reasonable to
have if you handle page fault for the address. But it should be adjusted
for faultaround.

--
Kirill A. Shutemov

[Minchan Kim]

Thanks for pointing out quickly!
Your suggestion does make sense to me.

[Minchan Kim]

So, faultaround should be disabled for non HW access bit architecture?
As you said, it would defeat faultaround benefit. As well, it adds reclaim
overhead because rmap should handle it to remove ptes and more pressure to slab.

[Kirill A. Shutemov]

Not necessarily. Need to be tested. For those architectures, after
faultaround, we would get faults to set accessed bit, which should be
cheaper than fault to pte_none().

> As you said, it would defeat faultaround benefit. As well, it adds reclaim
> overhead because rmap should handle it to remove ptes and more pressure to slab.

--
Kirill A. Shutemov

[Vinayak Menon]

Column at the end shows the values with the patch

3.18 3.18-fab=4096 3.18-Kirill's-fix

---------------------------------------------------------

workingset_refault 597687 878109 790207

workingset_activate 167169 254037 207912

pgpgin 4035424 5157348 4793116

pgpgout 162151 85231 85539

pgpgoutclean 928587 1225029 1129088

pswpin 46033 17100 8926

pswpout 237952 127686 103435

pgalloc_dma 3305034 3542614 3401000

pgfree 3354989 3592132 3457783

pgactivate 626468 355275 326716

pgdeactivate 990205 771902 697392

pageref_activate 294780 157106 138451

pageref_activate_vm_exec 141722 63469 64585

pageref_keep 121931 63028 65811

pgmajfault 67818 45643 34944

pgrefill_dma 1324023 977192 874497

pgscan_kswapd_dma 1825267 1720322 1577483

pgsteal_kswapd_dma 1181882 1365500 1243968

pgscan_direct_dma 41957 9622 9387

pgsteal_direct_dma 25136 6759 7108

slabs_scanned 689575 542705 618839

pageoutrun 1234 1538 1450

allocstall 110 26 13

Everything seems to have improved except slabs_scanned, possibly because
of this check which Minchan pointed out, that results in higher pressure on slabs.

if (page_mapped(page) || PageSwapCache(page))

sc->nr_scanned++;

I had added some traces to monitor the vmpressure values. Those also seems to
be high, possibly because of the same reason.

Should the pressure be doubled only if page is mapped and referenced ?

There is big improvement in avg latency, but still 5% higher than with fault_around
disabled. I will try to debug this further.

[Minchan Kim]

Hello,

Yes, pte_mkold is not perfect at the moment.

Anyway, above heuristic has been in there for a long time since I was born
maybe :) (I don't want to argue why it's there and whether it's right) So,
I'm really hesitant to change it that it might bite some workloads.
(But I don't mean I'm against it but just don't want to make it by myself
to avoid potential blame). IOW, Kirill's fault_around broke it too so it
could bite some workloads.

At least, as Vinayak mentioned, it would change vmpressure level so users of
vmpressure can be affected. AFAIK, some vendors in embedded side relies on
vmpressure to control memory management so it will hurt them.
As well, slab shrinking behavior was changed, too. Unfortunately, I don't
know any workload is dependent with it.

As other regression in my company product, we have snapshot a process
with workingset for later fast resume. For that, we have considered
pte-mapped pages as workingset for snapshot but snapshot start to include
non-workingset pages since fault-around is merged. It means snapshot
image size is increased so that we need more storage space and it starts
the thing slow down. I guess mincore(2) users will be affected.

Additional Note: There are lots of products with ARM which is non-HW access
bit system in embedded world although ARM start to support it recenlty and
sequential file access workload is not important compared to memory reclaim
So, fault_around's benefit could be higly limited compared to HW-access bit
architectures on server workload.

I want to ask again.
I guess we could disable fault_around by kernel parameter but does it
sound reasonable to enable fault_around by default for every arches
at the cost of above regression?

I'm not against for that. Just what I want is some fixes about the
regression should go to -stable.

>
> There is big improvement in avg latency, but still 5% higher than with fault_around
> disabled. I will try to debug this further.
>
>

[Minchan Kim]

I did quick test in my ARM machine.

512M file mmap sequential every word read

= vanilla fault_around=4096 =
minor fault: 131291
elapsed time(usec): 6686236

= vanilla fault_around=65536 =
minor fault: 12577
elapsed time(usec): 6586959

I tested 3 times and result seemed to be stable.
90% minor fault was reduced. It's huge win but as looking at elapsed time,
it's not huge win. Just about 1.5%.

= pte_mkold applied fault_around=4096 =
minor fault: 131291
elapsed time(usec): 6608358

= pte_mkold applied fault_around=65536 =
minor fault: 143609
elapsed time(usec): 6772520

I tested 3 times and result seemed to be stable.
minor fault was rather increased and elapsed time was slow with
fault_around.
Gain is really not clear.

[Minchan Kim]

Kirill,
You wanted to test non-HW access bit system and I did.
What's your opinion?

[Kirill A. Shutemov]

Sorry, for late response.

My patch is incomlete: we need to find a way to not mark pte as old if we
handle page fault for the address the pte represents.

Once this will be done, the number of page faults shouldn't be higher with
fault-around enabled even on machines without hardware accessed bit. This
will address performance regression with the patch on such machines.

I'll try to find time to update the patch soon.

--
Kirill A. Shutemov

[Minchan Kim]

I'm sure you can handle it but my point is there wouldn't be a big gain
although you can handle it in non-HW access bit system. Okay, let's be
more clear because I don't have every non-HW access bit architecture.
At least, current mobile workload in ARM which I have wouldn't be huge
benefit.
I will say one more.
I tested the workload on quad-core system and core speed is not so slow
compared to recent other mobile phone SoC. Even when I tested the benchmark
without pte_mkold, the benefit is within noise because storage is really
slow so major fault is dominant factor. So, I decide test storage from eMMC
to eSATA. And then finally, I manage to see the a little beneift with
fault_around without pte_mkold.

However, let's consider side-effect aspect from fault_around.

1. Increase slab shrinking compard to old
2. high level vmpressure compared to old

With considering that regressions on my system, it's really not worth to
try at the moment.
That's why I wanted to disable fault_around as default in non-HW access
bit system.

>
> Once this will be done, the number of page faults shouldn't be higher with
> fault-around enabled even on machines without hardware accessed bit. This
> will address performance regression with the patch on such machines.

Although you solves that, I guess the benefit would be marginal in
some architectures but we should solve above side-effects.

>
> I'll try to find time to update the patch soon.

I hope you can solve above those regressions as well.

[Kirill A. Shutemov]

Feel free to post such patch. I guess it's reasonable.

> > Once this will be done, the number of page faults shouldn't be higher with
> > fault-around enabled even on machines without hardware accessed bit. This
> > will address performance regression with the patch on such machines.
>
> Although you solves that, I guess the benefit would be marginal in
> some architectures but we should solve above side-effects.
>
> >
> > I'll try to find time to update the patch soon.
>
> I hope you can solve above those regressions as well.

The patch is posted. Please test.

--
Kirill A. Shutemov