mm: swap: use fixed-size readahead during swapoff

mm/swap_state.c +2 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*

Andrea Righi <[email protected]> writes:

[snip]

Quoted 25 lines

mm/swap_state.c +9 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*
+	 * If current task is using too much memory or swapoff is running
+	 * simply use the max readahead size. Since we likely want to load a
+	 * lot of pages back into memory, using a fixed-size max readhaead can
+	 * give better performance in this case.
+	 */
+	if (oom_task_origin(current))
+		return max_pages;
 

mm/swap_state.c static unsigned long swapin_nr_pages(unsigned long offset) · line 521

Quoted 2 lines

mm/swap_state.c +0 -0

 	hits = atomic_xchg(&swapin_readahead_hits, 0);
 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,

Thinks this again. If my understanding were correct, the accessing pattern during swapoff is sequential, why swap readahead doesn't work? If so, can you root cause that firstly?

Quoted 29 lines · Ying

Andrea Righi <[email protected]> writes:

[snip]

Quoted 25 lines

mm/swap_state.c +9 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*
+	 * If current task is using too much memory or swapoff is running
+	 * simply use the max readahead size. Since we likely want to load a
+	 * lot of pages back into memory, using a fixed-size max readhaead can
+	 * give better performance in this case.
+	 */
+	if (oom_task_origin(current))
+		return max_pages;
 

Quoted 6 lines

mm/swap_state.c static unsigned long swapin_nr_pages(unsigned long offset) · line 521

Quoted 2 lines

mm/swap_state.c +0 -0

 	hits = atomic_xchg(&swapin_readahead_hits, 0);
 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,

Thinks this again. If my understanding were correct, the accessing pattern during swapoff is sequential, why swap readahead doesn't work? If so, can you root cause that firstly?

Theoretically if the pattern is sequential the current heuristic should already select a big readahead size, but apparently it's not doing that.

I'll repeat my tests tracing the readahead size during swapoff to see exactly what's going on here.

Quoted 40 lines · Andrea Righi

Quoted 29 lines · Ying

Andrea Righi <[email protected]> writes:

[snip]

Quoted 25 lines

mm/swap_state.c +9 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*
+	 * If current task is using too much memory or swapoff is running
+	 * simply use the max readahead size. Since we likely want to load a
+	 * lot of pages back into memory, using a fixed-size max readhaead can
+	 * give better performance in this case.
+	 */
+	if (oom_task_origin(current))
+		return max_pages;
 

Quoted 6 lines

mm/swap_state.c static unsigned long swapin_nr_pages(unsigned long offset) · line 521

Quoted 2 lines

mm/swap_state.c +0 -0

 	hits = atomic_xchg(&swapin_readahead_hits, 0);
 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,

Thinks this again. If my understanding were correct, the accessing pattern during swapoff is sequential, why swap readahead doesn't work? If so, can you root cause that firstly?

Theoretically if the pattern is sequential the current heuristic should already select a big readahead size, but apparently it's not doing that.

Quoted 2 lines

I'll repeat my tests tracing the readahead size during swapoff to see exactly what's going on here.

I haven't verify it. It may be helpful to call lookup_swap_cache() before swapin_readahead() in unuse_pte_range(). The theory behind it is to update the swap readahead statistics via lookup_swap_cache().

Quoted 42 lines · Ying

Quoted 40 lines · Andrea Righi

Quoted 29 lines · Ying

Andrea Righi <[email protected]> writes:

[snip]

Quoted 25 lines

mm/swap_state.c +9 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*
+	 * If current task is using too much memory or swapoff is running
+	 * simply use the max readahead size. Since we likely want to load a
+	 * lot of pages back into memory, using a fixed-size max readhaead can
+	 * give better performance in this case.
+	 */
+	if (oom_task_origin(current))
+		return max_pages;
 

Quoted 6 lines

mm/swap_state.c static unsigned long swapin_nr_pages(unsigned long offset) · line 521

Quoted 2 lines

mm/swap_state.c +0 -0

 	hits = atomic_xchg(&swapin_readahead_hits, 0);
 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,

Thinks this again. If my understanding were correct, the accessing pattern during swapoff is sequential, why swap readahead doesn't work? If so, can you root cause that firstly?

Theoretically if the pattern is sequential the current heuristic should already select a big readahead size, but apparently it's not doing that.

Quoted 6 lines

Quoted 2 lines

I'll repeat my tests tracing the readahead size during swapoff to see exactly what's going on here.

I haven't verify it. It may be helpful to call lookup_swap_cache() before swapin_readahead() in unuse_pte_range(). The theory behind it is to update the swap readahead statistics via lookup_swap_cache().

I did more tests trying to collect some useful information.

In particular I've been focusing at tracing the distribution of the values returned by swapin_nr_pages() in different scenarios.

To do so I made swapin_nr_pages() trace-able and I used the following bcc command to measure the distrubution of the returned values:

 # argdist-bpfcc -c -C 'r::swapin_nr_pages(unsigned long offset):unsigned long:$retval'

I've collected this metric in the following scenarios:
- 5.6 vanilla
- 5.6 + lookup_swap_cache() before swapin_readahead() in unuse_pte_range()
- 5.6 + atomic_inc(&swapin_readahead_hits) before swapin_readahead() in unuse_pte_range()
- 5.6 + swapin_readahead_hits=last_readahead_pages (in the atomic way) before swapin_readahead() in unuse_pte_range()

Each kernel has been tested both with swappiness=0 and swappiness=60. Results are pretty much identical changing the swappiness, so I'm just reporting the default case here (w/ swappiness=60).

Result
======

= swapoff performance (elapsed time) =

 vanilla                 22.09s
 lookup_swap_cache()     23.87s
 hits++                  16.10s

hits=last_ra_pages 8.81s

= swapin_nr_pages() $retval distribution =

5.6 vanilla:
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	36948      $retval = 8
	44151      $retval = 4
	49290      $retval = 1
	527771     $retval = 2

5.6 lookup_swap_cache() before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	13093      $retval = 1
	56703      $retval = 8
	123067     $retval = 2
	366118     $retval = 4

5.6 atomic_inc(&swapin_readahead_hits) before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	2589       $retval = 1
	8016       $retval = 2
	40021      $retval = 8
	566038     $retval = 4

5.6 swapin_readahead_hits=last_readahead_pages before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	785        $retval = 2
	1072       $retval = 1
	21844      $retval = 4
	644168     $retval = 8

In the vanilla case, the readahead heuristic seems to choose 2 pages most of the time. This is because we are not properly considering the hits (hits are always 0 in the swapoff code path) and, as you correctly pointed out, we can fix this by calling lookup_swap_cache() in unuse_pte_range() before calling swapin_readahead().

With this change the distribution of the readahead size moves more toward 4 pages, but we still have some 2s. That looks good, however it doesn't seem to speed up swapoff very much... maybe because calling lookup_swap_cache() introduces a small overhead? (still need to investigate about this theory).

In the next test I've tried to always increment hits by 1 before calling swapin_readahead() in unuse_pte_range(). This is basically cheating, because I'm faking the hit ratio, forcing the heuristic to use a larger readahead size; in fact, the readahead size moves even more toward 4 pages and swapoff performance are a little better now.

Pushing even more the "cheating" I can pretend that the previous readahead was all hits (swapin_readahead_hits=last_readahead_pages), so I'm forcing the heuristic to move toward the max size and keep using it. The result here is pretty much identical to my fixed-size patch, because swapin_nr_pages() returns the max readahead size pretty much all the time during swapoff (8 pages or, more in general, vm.page-cluster).

Personally I don't like very much forcing the heuristic in this way, it'd be nice if it would just work by accounting the proper hit ratio (so just by adding lookup_swap_cache() as you correctly suggested), but this solution doesn't seem to improve performance in reality. For this reason I still think we should consider the swapoff scenario like a special one and somehow bypass the readahead heuristic and always return the max readahead size.

Looking at the hits of the previous step in the swapoff case just doesn't work, because we may have some misses, but they will become hits very soon, since we are reading all the swapped out pages back into memory. This is why using the max readahead size gives better swapoff performance.

What do you think?

Thanks,
-Andrea

Quoted 156 lines · Andrea Righi

Quoted 42 lines · Ying

Quoted 40 lines · Andrea Righi

Quoted 29 lines · Ying

Andrea Righi <[email protected]> writes:

[snip]

Quoted 25 lines

mm/swap_state.c +9 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*
+	 * If current task is using too much memory or swapoff is running
+	 * simply use the max readahead size. Since we likely want to load a
+	 * lot of pages back into memory, using a fixed-size max readhaead can
+	 * give better performance in this case.
+	 */
+	if (oom_task_origin(current))
+		return max_pages;
 

Quoted 6 lines

mm/swap_state.c static unsigned long swapin_nr_pages(unsigned long offset) · line 521

Quoted 2 lines

mm/swap_state.c +0 -0

 	hits = atomic_xchg(&swapin_readahead_hits, 0);
 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,

Thinks this again. If my understanding were correct, the accessing pattern during swapoff is sequential, why swap readahead doesn't work? If so, can you root cause that firstly?

Theoretically if the pattern is sequential the current heuristic should already select a big readahead size, but apparently it's not doing that.

Quoted 6 lines

Quoted 2 lines

I'll repeat my tests tracing the readahead size during swapoff to see exactly what's going on here.

I haven't verify it. It may be helpful to call lookup_swap_cache() before swapin_readahead() in unuse_pte_range(). The theory behind it is to update the swap readahead statistics via lookup_swap_cache().

I did more tests trying to collect some useful information.

In particular I've been focusing at tracing the distribution of the values returned by swapin_nr_pages() in different scenarios.

To do so I made swapin_nr_pages() trace-able and I used the following bcc command to measure the distrubution of the returned values:

 # argdist-bpfcc -c -C 'r::swapin_nr_pages(unsigned long offset):unsigned long:$retval'

I've collected this metric in the following scenarios:
- 5.6 vanilla
- 5.6 + lookup_swap_cache() before swapin_readahead() in unuse_pte_range()
- 5.6 + atomic_inc(&swapin_readahead_hits) before swapin_readahead() in unuse_pte_range()
- 5.6 + swapin_readahead_hits=last_readahead_pages (in the atomic way) before swapin_readahead() in unuse_pte_range()

Each kernel has been tested both with swappiness=0 and swappiness=60. Results are pretty much identical changing the swappiness, so I'm just reporting the default case here (w/ swappiness=60).

Result
======

= swapoff performance (elapsed time) =

 vanilla                 22.09s
 lookup_swap_cache()     23.87s
 hits++                  16.10s

hits=last_ra_pages 8.81s

= swapin_nr_pages() $retval distribution =

5.6 vanilla:
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	36948      $retval = 8
	44151      $retval = 4
	49290      $retval = 1
	527771     $retval = 2

5.6 lookup_swap_cache() before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	13093      $retval = 1
	56703      $retval = 8
	123067     $retval = 2
	366118     $retval = 4

5.6 atomic_inc(&swapin_readahead_hits) before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	2589       $retval = 1
	8016       $retval = 2
	40021      $retval = 8
	566038     $retval = 4

5.6 swapin_readahead_hits=last_readahead_pages before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	785        $retval = 2
	1072       $retval = 1
	21844      $retval = 4
	644168     $retval = 8

In the vanilla case, the readahead heuristic seems to choose 2 pages most of the time. This is because we are not properly considering the hits (hits are always 0 in the swapoff code path) and, as you correctly pointed out, we can fix this by calling lookup_swap_cache() in unuse_pte_range() before calling swapin_readahead().

With this change the distribution of the readahead size moves more toward 4 pages, but we still have some 2s. That looks good, however it doesn't seem to speed up swapoff very much... maybe because calling lookup_swap_cache() introduces a small overhead? (still need to investigate about this theory).

In the next test I've tried to always increment hits by 1 before calling swapin_readahead() in unuse_pte_range(). This is basically cheating, because I'm faking the hit ratio, forcing the heuristic to use a larger readahead size; in fact, the readahead size moves even more toward 4 pages and swapoff performance are a little better now.

Pushing even more the "cheating" I can pretend that the previous readahead was all hits (swapin_readahead_hits=last_readahead_pages), so I'm forcing the heuristic to move toward the max size and keep using it. The result here is pretty much identical to my fixed-size patch, because swapin_nr_pages() returns the max readahead size pretty much all the time during swapoff (8 pages or, more in general, vm.page-cluster).

Personally I don't like very much forcing the heuristic in this way, it'd be nice if it would just work by accounting the proper hit ratio (so just by adding lookup_swap_cache() as you correctly suggested), but this solution doesn't seem to improve performance in reality. For this reason I still think we should consider the swapoff scenario like a special one and somehow bypass the readahead heuristic and always return the max readahead size.

Looking at the hits of the previous step in the swapoff case just doesn't work, because we may have some misses, but they will become hits very soon, since we are reading all the swapped out pages back into memory. This is why using the max readahead size gives better swapoff performance.

Quoted 3 lines

What do you think?

From your description, it appears that you are using cluster readahead

instead of vma readahead. Can you verify this via,

# cat /sys/kernel/mm/swap/vma_ra_enabled

And if it returns false, you can enable it via,

# echo 1 > /sys/kernel/mm/swap/vma_ra_enabled

Because now swapoff code swapin pages in the page table order instead of the swap entry order. But this will turn the sequential disk read to random disk read too. Let's see the performance results.

And please make sure that in unuse_pte_range(), after lookup_swap_cache() returns non-NULL page, it's unnecessary to call swapin_readahead().

mm/swapfile.c +3 -1

-		page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, &vmf);
+		page = lookup_swap_cache(entry, vma, addr);
+		if (!page)
+			page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, &vmf);
 		if (!page) {
 			if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
 				goto try_next;

Thanks!
-Andrea

Quoted 3 lines · Andrea Righi

 mm/swapfile.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

mm/swapfile.c static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd, · line 1946

Quoted 8 lines

mm/swapfile.c +0 -0

index 9fd47e6f7a86..cb9eb517178d 100644
@@ -1944,7 +1944,9 @@ static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 		vmf.pmd = pmd;
 		last_ra = atomic_read(&last_readahead_pages);
 		atomic_set(&swapin_readahead_hits, last_ra);

You need to remove the above 2 lines firstly.

Quoted 5 lines · Ying

Quoted 3 lines · Andrea Righi

 mm/swapfile.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

Quoted 10 lines

mm/swapfile.c static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd, · line 1946

Quoted 8 lines

mm/swapfile.c +0 -0

index 9fd47e6f7a86..cb9eb517178d 100644
@@ -1944,7 +1944,9 @@ static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 		vmf.pmd = pmd;
 		last_ra = atomic_read(&last_readahead_pages);
 		atomic_set(&swapin_readahead_hits, last_ra);

You need to remove the above 2 lines firstly.

Meh... too much enthusiasm, and I definitely need more coffee this morning. Here's the right patch applied:

mm/swapfile.c +0 -0

index 5871a2aa86a5..8b38441b66fa 100644

mm/swapfile.c +0 -0

 		if (!page) {
 			if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
 				goto try_next;

And following the right results:

r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	1618       $retval = 1
	4960       $retval = 2
	41315      $retval = 4
	103521     $retval = 8

swapoff time: 12.19s

So, not as good as the fixed-size readahead, but it's definitely an improvement, considering that the swapoff time is ~22s without this patch applied.

I think this change can be a simple and reasonable compromise.

Thanks again and sorry for the noise,
-Andrea

Quoted 158 lines · Ying

Quoted 156 lines · Andrea Righi

Quoted 42 lines · Ying

Quoted 40 lines · Andrea Righi

Quoted 29 lines · Ying

Andrea Righi <[email protected]> writes:

[snip]

Quoted 25 lines

mm/swap_state.c +9 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*
+	 * If current task is using too much memory or swapoff is running
+	 * simply use the max readahead size. Since we likely want to load a
+	 * lot of pages back into memory, using a fixed-size max readhaead can
+	 * give better performance in this case.
+	 */
+	if (oom_task_origin(current))
+		return max_pages;
 

Quoted 6 lines

mm/swap_state.c static unsigned long swapin_nr_pages(unsigned long offset) · line 521

Quoted 2 lines

mm/swap_state.c +0 -0

 	hits = atomic_xchg(&swapin_readahead_hits, 0);
 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,

Thinks this again. If my understanding were correct, the accessing pattern during swapoff is sequential, why swap readahead doesn't work? If so, can you root cause that firstly?

Theoretically if the pattern is sequential the current heuristic should already select a big readahead size, but apparently it's not doing that.

Quoted 6 lines

Quoted 2 lines

I'll repeat my tests tracing the readahead size during swapoff to see exactly what's going on here.

I haven't verify it. It may be helpful to call lookup_swap_cache() before swapin_readahead() in unuse_pte_range(). The theory behind it is to update the swap readahead statistics via lookup_swap_cache().

I did more tests trying to collect some useful information.

In particular I've been focusing at tracing the distribution of the values returned by swapin_nr_pages() in different scenarios.

To do so I made swapin_nr_pages() trace-able and I used the following bcc command to measure the distrubution of the returned values:

 # argdist-bpfcc -c -C 'r::swapin_nr_pages(unsigned long offset):unsigned long:$retval'

I've collected this metric in the following scenarios:
- 5.6 vanilla
- 5.6 + lookup_swap_cache() before swapin_readahead() in unuse_pte_range()
- 5.6 + atomic_inc(&swapin_readahead_hits) before swapin_readahead() in unuse_pte_range()
- 5.6 + swapin_readahead_hits=last_readahead_pages (in the atomic way) before swapin_readahead() in unuse_pte_range()

Each kernel has been tested both with swappiness=0 and swappiness=60. Results are pretty much identical changing the swappiness, so I'm just reporting the default case here (w/ swappiness=60).

Result
======

= swapoff performance (elapsed time) =

 vanilla                 22.09s
 lookup_swap_cache()     23.87s
 hits++                  16.10s

hits=last_ra_pages 8.81s

= swapin_nr_pages() $retval distribution =

5.6 vanilla:
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	36948      $retval = 8
	44151      $retval = 4
	49290      $retval = 1
	527771     $retval = 2

5.6 lookup_swap_cache() before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	13093      $retval = 1
	56703      $retval = 8
	123067     $retval = 2
	366118     $retval = 4

5.6 atomic_inc(&swapin_readahead_hits) before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	2589       $retval = 1
	8016       $retval = 2
	40021      $retval = 8
	566038     $retval = 4

5.6 swapin_readahead_hits=last_readahead_pages before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	785        $retval = 2
	1072       $retval = 1
	21844      $retval = 4
	644168     $retval = 8

In the vanilla case, the readahead heuristic seems to choose 2 pages most of the time. This is because we are not properly considering the hits (hits are always 0 in the swapoff code path) and, as you correctly pointed out, we can fix this by calling lookup_swap_cache() in unuse_pte_range() before calling swapin_readahead().

With this change the distribution of the readahead size moves more toward 4 pages, but we still have some 2s. That looks good, however it doesn't seem to speed up swapoff very much... maybe because calling lookup_swap_cache() introduces a small overhead? (still need to investigate about this theory).

In the next test I've tried to always increment hits by 1 before calling swapin_readahead() in unuse_pte_range(). This is basically cheating, because I'm faking the hit ratio, forcing the heuristic to use a larger readahead size; in fact, the readahead size moves even more toward 4 pages and swapoff performance are a little better now.

Pushing even more the "cheating" I can pretend that the previous readahead was all hits (swapin_readahead_hits=last_readahead_pages), so I'm forcing the heuristic to move toward the max size and keep using it. The result here is pretty much identical to my fixed-size patch, because swapin_nr_pages() returns the max readahead size pretty much all the time during swapoff (8 pages or, more in general, vm.page-cluster).

Personally I don't like very much forcing the heuristic in this way, it'd be nice if it would just work by accounting the proper hit ratio (so just by adding lookup_swap_cache() as you correctly suggested), but this solution doesn't seem to improve performance in reality. For this reason I still think we should consider the swapoff scenario like a special one and somehow bypass the readahead heuristic and always return the max readahead size.

Looking at the hits of the previous step in the swapoff case just doesn't work, because we may have some misses, but they will become hits very soon, since we are reading all the swapped out pages back into memory. This is why using the max readahead size gives better swapoff performance.

Quoted 6 lines

Quoted 3 lines

What do you think?

From your description, it appears that you are using cluster readahead

instead of vma readahead. Can you verify this via,

# cat /sys/kernel/mm/swap/vma_ra_enabled

# cat /sys/kernel/mm/swap/vma_ra_enabled
true

However, it's still using the cluster readahead because I'm using a swap file and nr_rotate_swap=1, so, according to the following, it's never using the vma readahead:

static inline bool swap_use_vma_readahead(void)

 {
         return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);

}

I'll investigate more on this, I think there's no reason to prevent the usage of vma readahead if the underlying device is non-rotational.

Quoted 173 lines · Andrea Righi

Quoted 158 lines · Ying

Quoted 156 lines · Andrea Righi

Quoted 42 lines · Ying

Quoted 40 lines · Andrea Righi

Quoted 29 lines · Ying

Andrea Righi <[email protected]> writes:

[snip]

Quoted 25 lines

mm/swap_state.c +9 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*
+	 * If current task is using too much memory or swapoff is running
+	 * simply use the max readahead size. Since we likely want to load a
+	 * lot of pages back into memory, using a fixed-size max readhaead can
+	 * give better performance in this case.
+	 */
+	if (oom_task_origin(current))
+		return max_pages;
 

Quoted 6 lines

mm/swap_state.c static unsigned long swapin_nr_pages(unsigned long offset) · line 521

Quoted 2 lines

mm/swap_state.c +0 -0

 	hits = atomic_xchg(&swapin_readahead_hits, 0);
 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,

Thinks this again. If my understanding were correct, the accessing pattern during swapoff is sequential, why swap readahead doesn't work? If so, can you root cause that firstly?

Theoretically if the pattern is sequential the current heuristic should already select a big readahead size, but apparently it's not doing that.

Quoted 6 lines

Quoted 2 lines

I'll repeat my tests tracing the readahead size during swapoff to see exactly what's going on here.

I haven't verify it. It may be helpful to call lookup_swap_cache() before swapin_readahead() in unuse_pte_range(). The theory behind it is to update the swap readahead statistics via lookup_swap_cache().

I did more tests trying to collect some useful information.

In particular I've been focusing at tracing the distribution of the values returned by swapin_nr_pages() in different scenarios.

To do so I made swapin_nr_pages() trace-able and I used the following bcc command to measure the distrubution of the returned values:

 # argdist-bpfcc -c -C 'r::swapin_nr_pages(unsigned long offset):unsigned long:$retval'

I've collected this metric in the following scenarios:
- 5.6 vanilla
- 5.6 + lookup_swap_cache() before swapin_readahead() in unuse_pte_range()
- 5.6 + atomic_inc(&swapin_readahead_hits) before swapin_readahead() in unuse_pte_range()
- 5.6 + swapin_readahead_hits=last_readahead_pages (in the atomic way) before swapin_readahead() in unuse_pte_range()

Each kernel has been tested both with swappiness=0 and swappiness=60. Results are pretty much identical changing the swappiness, so I'm just reporting the default case here (w/ swappiness=60).

Result
======

= swapoff performance (elapsed time) =

 vanilla                 22.09s
 lookup_swap_cache()     23.87s
 hits++                  16.10s

hits=last_ra_pages 8.81s

= swapin_nr_pages() $retval distribution =

5.6 vanilla:
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	36948      $retval = 8
	44151      $retval = 4
	49290      $retval = 1
	527771     $retval = 2

5.6 lookup_swap_cache() before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	13093      $retval = 1
	56703      $retval = 8
	123067     $retval = 2
	366118     $retval = 4

5.6 atomic_inc(&swapin_readahead_hits) before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	2589       $retval = 1
	8016       $retval = 2
	40021      $retval = 8
	566038     $retval = 4

5.6 swapin_readahead_hits=last_readahead_pages before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	785        $retval = 2
	1072       $retval = 1
	21844      $retval = 4
	644168     $retval = 8

In the vanilla case, the readahead heuristic seems to choose 2 pages most of the time. This is because we are not properly considering the hits (hits are always 0 in the swapoff code path) and, as you correctly pointed out, we can fix this by calling lookup_swap_cache() in unuse_pte_range() before calling swapin_readahead().

With this change the distribution of the readahead size moves more toward 4 pages, but we still have some 2s. That looks good, however it doesn't seem to speed up swapoff very much... maybe because calling lookup_swap_cache() introduces a small overhead? (still need to investigate about this theory).

In the next test I've tried to always increment hits by 1 before calling swapin_readahead() in unuse_pte_range(). This is basically cheating, because I'm faking the hit ratio, forcing the heuristic to use a larger readahead size; in fact, the readahead size moves even more toward 4 pages and swapoff performance are a little better now.

Pushing even more the "cheating" I can pretend that the previous readahead was all hits (swapin_readahead_hits=last_readahead_pages), so I'm forcing the heuristic to move toward the max size and keep using it. The result here is pretty much identical to my fixed-size patch, because swapin_nr_pages() returns the max readahead size pretty much all the time during swapoff (8 pages or, more in general, vm.page-cluster).

Personally I don't like very much forcing the heuristic in this way, it'd be nice if it would just work by accounting the proper hit ratio (so just by adding lookup_swap_cache() as you correctly suggested), but this solution doesn't seem to improve performance in reality. For this reason I still think we should consider the swapoff scenario like a special one and somehow bypass the readahead heuristic and always return the max readahead size.

Looking at the hits of the previous step in the swapoff case just doesn't work, because we may have some misses, but they will become hits very soon, since we are reading all the swapped out pages back into memory. This is why using the max readahead size gives better swapoff performance.

Quoted 6 lines

Quoted 3 lines

What do you think?

From your description, it appears that you are using cluster readahead

instead of vma readahead. Can you verify this via,

# cat /sys/kernel/mm/swap/vma_ra_enabled

# cat /sys/kernel/mm/swap/vma_ra_enabled
true

However, it's still using the cluster readahead because I'm using a swap file and nr_rotate_swap=1, so, according to the following, it's never using the vma readahead:

Quoted 7 lines

static inline bool swap_use_vma_readahead(void)

 {
         return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);

}

I'll investigate more on this, I think there's no reason to prevent the usage of vma readahead if the underlying device is non-rotational.

Few more details about this.

Even if I have vma_ra_enabled (in /sys/kernel/mm/swap/vma_ra_enabled) it looks like my virtio_blk device is considered rotational by default:

$ cat /sys/block/vda/queue/rotational
1

Therefore the vma readahead is not used. If I change this to "rotational" then the vma readahead is used; this is confirmed by the fact that swapin_nr_pages isn't called anymore:

r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	13         $retval = 1
	18         $retval = 2
	23         $retval = 4
	29         $retval = 8

swapoff time: 16.44s

In terms of swapoff performance vma readahead works better than the vanilla cluster readahead (~21s), but the "improved" cluster readahead (with lookup_swap_cache() in unuse_pte_range()) still gives slightly better results (~12s).

-Andrea

Quoted 203 lines · Andrea Righi

Quoted 173 lines · Andrea Righi

Quoted 158 lines · Ying

Quoted 156 lines · Andrea Righi

Quoted 42 lines · Ying

Quoted 40 lines · Andrea Righi

Quoted 29 lines · Ying

Andrea Righi <[email protected]> writes:

[snip]

Quoted 25 lines

mm/swap_state.c +9 -0

index ebed37bbf7a3..c71abc8df304 100644
@@ -20,6 +20,7 @@
 #include <linux/migrate.h>
 #include <linux/vmalloc.h>
 #include <linux/swap_slots.h>
+#include <linux/oom.h>
 #include <linux/huge_mm.h>
 
 #include <asm/pgtable.h>

@@ -507,6 +508,14 @@ static unsigned long swapin_nr_pages(unsigned long offset)
 	max_pages = 1 << READ_ONCE(page_cluster);
 	if (max_pages <= 1)
 		return 1;
+	/*
+	 * If current task is using too much memory or swapoff is running
+	 * simply use the max readahead size. Since we likely want to load a
+	 * lot of pages back into memory, using a fixed-size max readhaead can
+	 * give better performance in this case.
+	 */
+	if (oom_task_origin(current))
+		return max_pages;
 

Quoted 6 lines

mm/swap_state.c static unsigned long swapin_nr_pages(unsigned long offset) · line 521

Quoted 2 lines

mm/swap_state.c +0 -0

 	hits = atomic_xchg(&swapin_readahead_hits, 0);
 	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,

Thinks this again. If my understanding were correct, the accessing pattern during swapoff is sequential, why swap readahead doesn't work? If so, can you root cause that firstly?

Theoretically if the pattern is sequential the current heuristic should already select a big readahead size, but apparently it's not doing that.

Quoted 6 lines

Quoted 2 lines

I'll repeat my tests tracing the readahead size during swapoff to see exactly what's going on here.

I haven't verify it. It may be helpful to call lookup_swap_cache() before swapin_readahead() in unuse_pte_range(). The theory behind it is to update the swap readahead statistics via lookup_swap_cache().

I did more tests trying to collect some useful information.

In particular I've been focusing at tracing the distribution of the values returned by swapin_nr_pages() in different scenarios.

To do so I made swapin_nr_pages() trace-able and I used the following bcc command to measure the distrubution of the returned values:

 # argdist-bpfcc -c -C 'r::swapin_nr_pages(unsigned long offset):unsigned long:$retval'

I've collected this metric in the following scenarios:
- 5.6 vanilla
- 5.6 + lookup_swap_cache() before swapin_readahead() in unuse_pte_range()
- 5.6 + atomic_inc(&swapin_readahead_hits) before swapin_readahead() in unuse_pte_range()
- 5.6 + swapin_readahead_hits=last_readahead_pages (in the atomic way) before swapin_readahead() in unuse_pte_range()

Each kernel has been tested both with swappiness=0 and swappiness=60. Results are pretty much identical changing the swappiness, so I'm just reporting the default case here (w/ swappiness=60).

Result
======

= swapoff performance (elapsed time) =

 vanilla                 22.09s
 lookup_swap_cache()     23.87s
 hits++                  16.10s

hits=last_ra_pages 8.81s

= swapin_nr_pages() $retval distribution =

5.6 vanilla:
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	36948      $retval = 8
	44151      $retval = 4
	49290      $retval = 1
	527771     $retval = 2

5.6 lookup_swap_cache() before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	13093      $retval = 1
	56703      $retval = 8
	123067     $retval = 2
	366118     $retval = 4

5.6 atomic_inc(&swapin_readahead_hits) before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	2589       $retval = 1
	8016       $retval = 2
	40021      $retval = 8
	566038     $retval = 4

5.6 swapin_readahead_hits=last_readahead_pages before swapin_readahead():
r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	785        $retval = 2
	1072       $retval = 1
	21844      $retval = 4
	644168     $retval = 8

In the vanilla case, the readahead heuristic seems to choose 2 pages most of the time. This is because we are not properly considering the hits (hits are always 0 in the swapoff code path) and, as you correctly pointed out, we can fix this by calling lookup_swap_cache() in unuse_pte_range() before calling swapin_readahead().

With this change the distribution of the readahead size moves more toward 4 pages, but we still have some 2s. That looks good, however it doesn't seem to speed up swapoff very much... maybe because calling lookup_swap_cache() introduces a small overhead? (still need to investigate about this theory).

In the next test I've tried to always increment hits by 1 before calling swapin_readahead() in unuse_pte_range(). This is basically cheating, because I'm faking the hit ratio, forcing the heuristic to use a larger readahead size; in fact, the readahead size moves even more toward 4 pages and swapoff performance are a little better now.

Pushing even more the "cheating" I can pretend that the previous readahead was all hits (swapin_readahead_hits=last_readahead_pages), so I'm forcing the heuristic to move toward the max size and keep using it. The result here is pretty much identical to my fixed-size patch, because swapin_nr_pages() returns the max readahead size pretty much all the time during swapoff (8 pages or, more in general, vm.page-cluster).

Personally I don't like very much forcing the heuristic in this way, it'd be nice if it would just work by accounting the proper hit ratio (so just by adding lookup_swap_cache() as you correctly suggested), but this solution doesn't seem to improve performance in reality. For this reason I still think we should consider the swapoff scenario like a special one and somehow bypass the readahead heuristic and always return the max readahead size.

Looking at the hits of the previous step in the swapoff case just doesn't work, because we may have some misses, but they will become hits very soon, since we are reading all the swapped out pages back into memory. This is why using the max readahead size gives better swapoff performance.

Quoted 6 lines

Quoted 3 lines

What do you think?

From your description, it appears that you are using cluster readahead

instead of vma readahead. Can you verify this via,

# cat /sys/kernel/mm/swap/vma_ra_enabled

# cat /sys/kernel/mm/swap/vma_ra_enabled
true

However, it's still using the cluster readahead because I'm using a swap file and nr_rotate_swap=1, so, according to the following, it's never using the vma readahead:

Quoted 7 lines

static inline bool swap_use_vma_readahead(void)

 {
         return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);

}

I'll investigate more on this, I think there's no reason to prevent the usage of vma readahead if the underlying device is non-rotational.

Few more details about this.

Even if I have vma_ra_enabled (in /sys/kernel/mm/swap/vma_ra_enabled) it looks like my virtio_blk device is considered rotational by default:

$ cat /sys/block/vda/queue/rotational
1

Therefore the vma readahead is not used. If I change this to "rotational" then the vma readahead is used; this is confirmed by the fact that swapin_nr_pages isn't called anymore:

r::swapin_nr_pages(unsigned long offset):unsigned long:$retval

	COUNT      EVENT
	13         $retval = 1
	18         $retval = 2
	23         $retval = 4
	29         $retval = 8

swapoff time: 16.44s

Quoted 4 lines

In terms of swapoff performance vma readahead works better than the vanilla cluster readahead (~21s), but the "improved" cluster readahead (with lookup_swap_cache() in unuse_pte_range()) still gives slightly better results (~12s).

Thanks for testing. Can you check the return value of __swapin_nr_pages() to verify whether vma readahead works as expected?

The global swap-in readahead policy takes in account the previous access patterns, using a scaling heuristic to determine the optimal readahead chunk dynamically.

This works pretty well in most cases, but like any heuristic there are specific cases when this approach is not ideal, for example the swapoff scenario.

During swapoff we just want to load back into memory all the swapped-out pages and for this specific use case a fixed-size readahead is more efficient.

The specific use case this patch is addressing is to improve swapoff performance when a VM has been hibernated, resumed and all memory needs to be forced back to RAM by disabling swap (see the test case below).

Quoted 8 lines · Andrea Righi

The global swap-in readahead policy takes in account the previous access patterns, using a scaling heuristic to determine the optimal readahead chunk dynamically.

This works pretty well in most cases, but like any heuristic there are specific cases when this approach is not ideal, for example the swapoff scenario.

Quoted 7 lines

During swapoff we just want to load back into memory all the swapped-out pages and for this specific use case a fixed-size readahead is more efficient.

The specific use case this patch is addressing is to improve swapoff performance when a VM has been hibernated, resumed and all memory needs to be forced back to RAM by disabling swap (see the test case below).

Why do you need to swapoff after resuming? The swap device isn't used except hibernation? I guess the process is,

1) add swap device to VM
2) hibernate
3) resume
4) swapoff

Some pages are swapped out in step 2? If os, can we just set /proc/sys/vm/swappiness to 0 to avoid swapping in step 2?

Quoted 18 lines · Ying

Quoted 8 lines · Andrea Righi

The global swap-in readahead policy takes in account the previous access patterns, using a scaling heuristic to determine the optimal readahead chunk dynamically.

This works pretty well in most cases, but like any heuristic there are specific cases when this approach is not ideal, for example the swapoff scenario.

Quoted 7 lines

During swapoff we just want to load back into memory all the swapped-out pages and for this specific use case a fixed-size readahead is more efficient.

The specific use case this patch is addressing is to improve swapoff performance when a VM has been hibernated, resumed and all memory needs to be forced back to RAM by disabling swap (see the test case below).

Quoted 7 lines

Why do you need to swapoff after resuming? The swap device isn't used except hibernation? I guess the process is,

1) add swap device to VM
2) hibernate
3) resume
4) swapoff

Correct, the swap device is used only for hibernation, when the system is resumed the swap is disabled (swapoff).

Quoted 1 line

Quoted 2 lines

Some pages are swapped out in step 2? If os, can we just set /proc/sys/vm/swappiness to 0 to avoid swapping in step 2?

Sorry, can you elaborate more on this? All anonymous pages are swapped out during step 2, it doesn't matter if we set swappiness to 0, they are swapped out anyway, because we need save them somewhere in order to hibernate, shutting down the system.

Thanks,
-Andrea