Skip to contentSkip to conversation

Introduce MADV_COLD and MADV_PAGEOUT

PATCH Superseded Minchan Kim 5 participants 25 messages Jun 10, 2019 – Jun 20, 2019

Subsystems (16)

v2 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

v2 cover Introduce MADV_COLD and MADV_PAGEOUT

This patch is part of previous series:
https://lore.kernel.org/lkml/[email protected]/T/#u
Originally, it was created for external madvise hinting feature.

https://lkml.org/lkml/2019/5/31/463
Michal wanted to separte the discussion from external hinting interface so this patchset includes only first part of my entire patchset

- introduce MADV_COLD and MADV_PAGEOUT hint to madvise.

However, I keep entire description for others for easier understanding why this kinds of hint was born.

Thanks.

This patchset is against on next-20190530.

Below is description of previous entire patchset.
================= &< =====================

- Background

The Android terminology used for forking a new process and starting an app from scratch is a cold start, while resuming an existing app is a hot start. While we continually try to improve the performance of cold starts, hot starts will always be significantly less power hungry as well as faster so we are trying to make hot start more likely than cold start.

To increase hot start, Android userspace manages the order that apps should be killed in a process called ActivityManagerService. ActivityManagerService tracks every Android app or service that the user could be interacting with at any time and translates that into a ranked list for lmkd(low memory killer daemon). They are likely to be killed by lmkd if the system has to reclaim memory. In that sense they are similar to entries in any other cache. Those apps are kept alive for opportunistic performance improvements but those performance improvements will vary based on the memory requirements of individual workloads.

- Problem

Naturally, cached apps were dominant consumers of memory on the system. However, they were not significant consumers of swap even though they are good candidate for swap. Under investigation, swapping out only begins once the low zone watermark is hit and kswapd wakes up, but the overall allocation rate in the system might trip lmkd thresholds and cause a cached process to be killed(we measured performance swapping out vs. zapping the memory by killing a process. Unsurprisingly, zapping is 10x times faster even though we use zram which is much faster than real storage) so kill from lmkd will often satisfy the high zone watermark, resulting in very few pages actually being moved to swap.

- Approach

The approach we chose was to use a new interface to allow userspace to proactively reclaim entire processes by leveraging platform information. This allowed us to bypass the inaccuracy of the kernel’s LRUs for pages that are known to be cold from userspace and to avoid races with lmkd by reclaiming apps as soon as they entered the cached state. Additionally, it could provide many chances for platform to use much information to optimize memory efficiency.

To achieve the goal, the patchset introduce two new options for madvise. One is MADV_COLD which will deactivate activated pages and the other is MADV_PAGEOUT which will reclaim private pages instantly. These new options complement MADV_DONTNEED and MADV_FREE by adding non-destructive ways to gain some free memory space. MADV_PAGEOUT is similar to MADV_DONTNEED in a way that it hints the kernel that memory region is not currently needed and should be reclaimed immediately; MADV_COLD is similar to MADV_FREE in a way that it hints the kernel that memory region is not currently needed and should be reclaimed when memory pressure rises.

This approach is similar in spirit to madvise(MADV_WONTNEED), but the information required to make the reclaim decision is not known to the app. Instead, it is known to a centralized userspace daemon, and that daemon must be able to initiate reclaim on its own without any app involvement. To solve the concern, this patch introduces new syscall -

    struct pr_madvise_param {
            int size;               /* the size of this structure */
            int cookie;             /* reserved to support atomicity */
            int nr_elem;            /* count of below arrary fields */
            int __user *hints;      /* hints for each range */
            /* to store result of each operation */
            const struct iovec __user *results;
            /* input address ranges */
            const struct iovec __user *ranges;
    };
    
    int process_madvise(int pidfd, struct pr_madvise_param *u_param,
                            unsigned long flags);

The syscall get pidfd to give hints to external process and provides pair of result/ranges vector arguments so that it could give several hints to each address range all at once. It also has cookie variable to support atomicity of the API for address ranges operations. IOW, if target process changes address space since monitor process has parsed address ranges via map_files or maps, the API can detect the race so could cancel entire address space operation. It's not implemented yet. Daniel Colascione suggested a idea(Please read description in patch[6/6]) and this patchset adds cookie a variable for the future.

- Experiment

We did bunch of testing with several hundreds of real users, not artificial benchmark on android. We saw about 17% cold start decreasement without any significant battery/app startup latency issues. And with artificial benchmark which launches and switching apps, we saw average 7% app launching improvement, 18% less lmkd kill and good stat from vmstat.

A is vanilla and B is process_madvise.

                                       A          B      delta   ratio(%)
               allocstall_dma          0          0          0       0.00
           allocstall_movable       1464        457      -1007     -69.00
            allocstall_normal     263210     190763     -72447     -28.00
             allocstall_total     264674     191220     -73454     -28.00
          compact_daemon_wake      26912      25294      -1618      -7.00
                 compact_fail      17885      14151      -3734     -21.00
         compact_free_scanned 4204766409 3835994922 -368771487      -9.00
             compact_isolated    3446484    2967618    -478866     -14.00
      compact_migrate_scanned 1621336411 1324695710 -296640701     -19.00
                compact_stall      19387      15343      -4044     -21.00
              compact_success       1502       1192       -310     -21.00
kswapd_high_wmark_hit_quickly        234        184        -50     -22.00
            kswapd_inodesteal     221635     233093      11458       5.00
 kswapd_low_wmark_hit_quickly      66065      54009     -12056     -19.00
                   nr_dirtied     259934     296476      36542      14.00
  nr_vmscan_immediate_reclaim       2587       2356       -231      -9.00
              nr_vmscan_write    1274232    2661733    1387501     108.00
                   nr_written    1514060    2937560    1423500      94.00
                   pageoutrun      67561      55133     -12428     -19.00
                   pgactivate    2335060    1984882    -350178     -15.00
                  pgalloc_dma   13743011   14096463     353452       2.00
              pgalloc_movable          0          0          0       0.00
               pgalloc_normal   18742440   16802065   -1940375     -11.00
                pgalloc_total   32485451   30898528   -1586923      -5.00
                 pgdeactivate    4262210    2930670   -1331540     -32.00
                      pgfault   30812334   31085065     272731       0.00
                       pgfree   33553970   31765164   -1788806      -6.00
                 pginodesteal      33411      15084     -18327     -55.00
                  pglazyfreed          0          0          0       0.00
                   pgmajfault     551312    1508299     956987     173.00
               pgmigrate_fail      43927      29330     -14597     -34.00
            pgmigrate_success    1399851    1203922    -195929     -14.00
                       pgpgin   24141776   19032156   -5109620     -22.00
                      pgpgout     959344    1103316     143972      15.00
                 pgpgoutclean    4639732    3765868    -873864     -19.00
                     pgrefill    4884560    3006938   -1877622     -39.00
                    pgrotated      37828      25897     -11931     -32.00
                pgscan_direct    1456037     957567    -498470     -35.00
       pgscan_direct_throttle          0          0          0       0.00
                pgscan_kswapd    6667767    5047360   -1620407     -25.00
                 pgscan_total    8123804    6004927   -2118877     -27.00
                   pgskip_dma          0          0          0       0.00
               pgskip_movable          0          0          0       0.00
                pgskip_normal      14907      25382      10475      70.00
                 pgskip_total      14907      25382      10475      70.00
               pgsteal_direct    1118986     690215    -428771     -39.00
               pgsteal_kswapd    4750223    3657107   -1093116     -24.00
                pgsteal_total    5869209    4347322   -1521887     -26.00
                       pswpin     417613    1392647     975034     233.00
                      pswpout    1274224    2661731    1387507     108.00
                slabs_scanned   13686905   10807200   -2879705     -22.00
          workingset_activate     668966     569444     -99522     -15.00
       workingset_nodereclaim      38957      32621      -6336     -17.00
           workingset_refault    2816795    2179782    -637013     -23.00
           workingset_restore     294320     168601    -125719     -43.00

pgmajfault is increased by 173% because swapin is increased by 200% by process_madvise hint. However, swap read based on zram is much cheaper than file IO in performance point of view and app hot start by swapin is also cheaper than cold start from the beginning of app which needs many IO from storage and initialization steps.

Brian Geffon in ChromeOS team had an experiment with process_madvise(2) Quote form him:
"What I found is that by using process_madvise after a tab has been back grounded for more than 45 seconds reduced the average tab switch times by 25%! This is a huge result and very obvious validation that process_madvise hints works well for the ChromeOS use case."

This patchset is against on next-20190607.

Minchan Kim (5):
mm: introduce MADV_COLD
mm: change PAGEREF_RECLAIM_CLEAN with PAGE_REFRECLAIM
mm: account nr_isolated_xxx in [isolate|putback]_lru_page
mm: introduce MADV_PAGEOUT
mm: factor out pmd young/dirty bit handling and THP split

 include/linux/huge_mm.h                |   3 -
 include/linux/swap.h                   |   2 +
 include/uapi/asm-generic/mman-common.h |   2 +
 mm/compaction.c                        |   2 -
 mm/gup.c                               |   7 +-
 mm/huge_memory.c                       |  74 -----
 mm/internal.h                          |   2 +-
 mm/khugepaged.c                        |   3 -
 mm/madvise.c                           | 358 ++++++++++++++++++++++++-
 mm/memory-failure.c                    |   3 -
 mm/memory_hotplug.c                    |   4 -
 mm/mempolicy.c                         |   6 +-
 mm/migrate.c                           |  37 +--
 mm/oom_kill.c                          |   2 +-
 mm/swap.c                              |  42 +++
 mm/vmscan.c                            |  86 +++++-
 16 files changed, 486 insertions(+), 147 deletions(-)
Signature · 1 line
2.22.0.rc2.383.gf4fbbf30c2-goog
#1 lore raw

PATCH v2 1/5 mm: introduce MADV_COLD

When a process expects no accesses to a certain memory range, it could give a hint to kernel that the pages can be reclaimed when memory pressure happens but data should be preserved for future use. This could reduce workingset eviction so it ends up increasing performance.

This patch introduces the new MADV_COLD hint to madvise(2) syscall. MADV_COLD can be used by a process to mark a memory range as not expected to be used in the near future. The hint can help kernel in deciding which pages to evict early during memory pressure.

It works for every LRU pages like MADV_[DONTNEED|FREE]. IOW, It moves

	active file page -> inactive file LRU
	active anon page -> inacdtive anon LRU

Unlike MADV_FREE, it doesn't move active anonymous pages to inactive file LRU's head because MADV_COLD is a little bit different symantic. MADV_FREE means it's okay to discard when the memory pressure because the content of the page is *garbage* so freeing such pages is almost zero overhead since we don't need to swap out and access afterward causes just minor fault. Thus, it would make sense to put those freeable pages in inactive file LRU to compete other used-once pages. It makes sense for implmentaion point of view, too because it's not swapbacked memory any longer until it would be re-dirtied. Even, it could give a bonus to make them be reclaimed on swapless system. However, MADV_COLD doesn't mean garbage so reclaiming them requires swap-out/in in the end so it's bigger cost. Since we have designed VM LRU aging based on cost-model, anonymous cold pages would be better to position inactive anon's LRU list, not file LRU. Furthermore, it would help to avoid unnecessary scanning if system doesn't have a swap device. Let's start simpler way without adding complexity at this moment.

All of error rule is same with MADV_DONTNEED.

* v1
* remove page_mapcount filter - hannes, mhocko
* fix idle page handling - joelaf

* RFCv2
* add more description - mhocko

* RFCv1
* renaming from MADV_COOL to MADV_COLD - hannes

* internal review
* use clear_page_youn in deactivate_page - joelaf
* Revise the description - surenb
* Renaming from MADV_WARM to MADV_COOL - surenb

---
 include/linux/swap.h                   |   1 +
 include/uapi/asm-generic/mman-common.h |   1 +
 mm/internal.h                          |   2 +-
 mm/madvise.c                           | 151 ++++++++++++++++++++++++-
 mm/oom_kill.c                          |   2 +-
 mm/swap.c                              |  42 +++++++
 6 files changed, 195 insertions(+), 4 deletions(-)
include/linux/swap.h +1 -0
index de2c67a33b7e..0ce997edb8bb 100644
@@ -340,6 +340,7 @@ extern void lru_add_drain_cpu(int cpu);
 extern void lru_add_drain_all(void);
 extern void rotate_reclaimable_page(struct page *page);
 extern void deactivate_file_page(struct page *page);
+extern void deactivate_page(struct page *page);
 extern void mark_page_lazyfree(struct page *page);
 extern void swap_setup(void);
 
include/uapi/asm-generic/mman-common.h +1 -0
index ef4623f03156..d7b4231eea63 100644
@@ -47,6 +47,7 @@
 #define MADV_SEQUENTIAL	2		/* expect sequential page references */
 #define MADV_WILLNEED	3		/* will need these pages */
 #define MADV_DONTNEED	4		/* don't need these pages */
+#define MADV_COLD	5		/* deactivatie these pages */
 
 /* common parameters: try to keep these consistent across architectures */
 #define MADV_FREE	8		/* free pages only if memory pressure */
mm/internal.h +1 -1
index e32390802fd3..0d5f720c75ab 100644
@@ -39,7 +39,7 @@ vm_fault_t do_swap_page(struct vm_fault *vmf);
 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *start_vma,
 		unsigned long floor, unsigned long ceiling);
 
-static inline bool can_madv_dontneed_vma(struct vm_area_struct *vma)
+static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
 {
 	return !(vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP));
 }
mm/madvise.c +149 -2

187 lines not on this page · raw message

mm/oom_kill.c +1 -1
index 5a58778c91d4..f73d5f5145f0 100644
@@ -515,7 +515,7 @@ bool __oom_reap_task_mm(struct mm_struct *mm)
 	set_bit(MMF_UNSTABLE, &mm->flags);
 
 	for (vma = mm->mmap ; vma; vma = vma->vm_next) {
-		if (!can_madv_dontneed_vma(vma))
+		if (!can_madv_lru_vma(vma))
 			continue;
 
 		/*
mm/swap.c +42 -0

72 lines not on this page · raw message

Signature · 1 line
2.22.0.rc2.383.gf4fbbf30c2-goog
reviewer
#2 lore raw

Re: [PATCH v2 1/5] mm: introduce MADV_COLD

Quoted 15 lines · Minchan Kim

When a process expects no accesses to a certain memory range, it could give a hint to kernel that the pages can be reclaimed when memory pressure happens but data should be preserved for future use. This could reduce workingset eviction so it ends up increasing performance.

This patch introduces the new MADV_COLD hint to madvise(2) syscall. MADV_COLD can be used by a process to mark a memory range as not expected to be used in the near future. The hint can help kernel in deciding which pages to evict early during memory pressure.

It works for every LRU pages like MADV_[DONTNEED|FREE]. IOW, It moves

	active file page -> inactive file LRU
	active anon page -> inacdtive anon LRU
Quoted 16 lines

Unlike MADV_FREE, it doesn't move active anonymous pages to inactive file LRU's head because MADV_COLD is a little bit different symantic. MADV_FREE means it's okay to discard when the memory pressure because the content of the page is *garbage* so freeing such pages is almost zero overhead since we don't need to swap out and access afterward causes just minor fault. Thus, it would make sense to put those freeable pages in inactive file LRU to compete other used-once pages. It makes sense for implmentaion point of view, too because it's not swapbacked memory any longer until it would be re-dirtied. Even, it could give a bonus to make them be reclaimed on swapless system. However, MADV_COLD doesn't mean garbage so reclaiming them requires swap-out/in in the end so it's bigger cost. Since we have designed VM LRU aging based on cost-model, anonymous cold pages would be better to position inactive anon's LRU list, not file LRU. Furthermore, it would help to avoid unnecessary scanning if system doesn't have a swap device. Let's start simpler way without adding complexity at this moment.

I would only add that it is a caveat that workloads with a lot of page cache are likely to ignore MADV_COLD on anonymous memory because we rarely age anonymous LRU lists.

[...]

mm/madvise.c static long madvise_willneed(struct vm_area_struct *vma, · line 313

Quoted 3 lines
mm/madvise.c +3 -0
+static int madvise_cold_pte_range(pmd_t *pmd, unsigned long addr,
+				unsigned long end, struct mm_walk *walk)
+{

This is duplicating a large part of madvise_free_pte_range with some subtle differences which are not explained anywhere (e.g. why does madvise_free_huge_pmd need try_lock on a page while not here? etc.).

Why cannot we reuse a large part of that code and differ essentially on the reclaim target check and action? Have you considered to consolidate the code to share as much as possible? Maybe that is easier said than done because the devil is always in details...

I would definitely feel much more comfortable to review the code without thinking about all those subtle details that have been already solved before. Especially all the THP ones.

Other than that the patch looks sane to me.

Quoted 105 lines
mm/madvise.c +105 -0
+	struct mmu_gather *tlb = walk->private;
+	struct mm_struct *mm = tlb->mm;
+	struct vm_area_struct *vma = walk->vma;
+	pte_t *orig_pte, *pte, ptent;
+	spinlock_t *ptl;
+	struct page *page;
+	unsigned long next;
+
+	next = pmd_addr_end(addr, end);
+	if (pmd_trans_huge(*pmd)) {
+		pmd_t orig_pmd;
+
+		tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
+		ptl = pmd_trans_huge_lock(pmd, vma);
+		if (!ptl)
+			return 0;
+
+		orig_pmd = *pmd;
+		if (is_huge_zero_pmd(orig_pmd))
+			goto huge_unlock;
+
+		if (unlikely(!pmd_present(orig_pmd))) {
+			VM_BUG_ON(thp_migration_supported() &&
+					!is_pmd_migration_entry(orig_pmd));
+			goto huge_unlock;
+		}
+
+		page = pmd_page(orig_pmd);
+		if (next - addr != HPAGE_PMD_SIZE) {
+			int err;
+
+			if (page_mapcount(page) != 1)
+				goto huge_unlock;
+
+			get_page(page);
+			spin_unlock(ptl);
+			lock_page(page);
+			err = split_huge_page(page);
+			unlock_page(page);
+			put_page(page);
+			if (!err)
+				goto regular_page;
+			return 0;
+		}
+
+		if (pmd_young(orig_pmd)) {
+			pmdp_invalidate(vma, addr, pmd);
+			orig_pmd = pmd_mkold(orig_pmd);
+
+			set_pmd_at(mm, addr, pmd, orig_pmd);
+			tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
+		}
+
+		test_and_clear_page_young(page);
+		deactivate_page(page);
+huge_unlock:
+		spin_unlock(ptl);
+		return 0;
+	}
+
+	if (pmd_trans_unstable(pmd))
+		return 0;
+
+regular_page:
+	tlb_change_page_size(tlb, PAGE_SIZE);
+	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+	flush_tlb_batched_pending(mm);
+	arch_enter_lazy_mmu_mode();
+	for (; addr < end; pte++, addr += PAGE_SIZE) {
+		ptent = *pte;
+
+		if (pte_none(ptent))
+			continue;
+
+		if (!pte_present(ptent))
+			continue;
+
+		page = vm_normal_page(vma, addr, ptent);
+		if (!page)
+			continue;
+
+		if (pte_young(ptent)) {
+			ptent = ptep_get_and_clear_full(mm, addr, pte,
+							tlb->fullmm);
+			ptent = pte_mkold(ptent);
+			set_pte_at(mm, addr, pte, ptent);
+			tlb_remove_tlb_entry(tlb, pte, addr);
+		}
+
+		/*
+		 * We are deactivating a page for accelerating reclaiming.
+		 * VM couldn't reclaim the page unless we clear PG_young.
+		 * As a side effect, it makes confuse idle-page tracking
+		 * because they will miss recent referenced history.
+		 */
+		test_and_clear_page_young(page);
+		deactivate_page(page);
+	}
+
+	arch_enter_lazy_mmu_mode();
+	pte_unmap_unlock(orig_pte, ptl);
+	cond_resched();
+
+	return 0;
+}
Signature · 2 lines
Michal Hocko
SUSE Labs
#3 lore raw

Re: [PATCH v2 1/5] mm: introduce MADV_COLD

Quoted 33 lines · Michal Hocko
Quoted 15 lines · Minchan Kim

When a process expects no accesses to a certain memory range, it could give a hint to kernel that the pages can be reclaimed when memory pressure happens but data should be preserved for future use. This could reduce workingset eviction so it ends up increasing performance.

This patch introduces the new MADV_COLD hint to madvise(2) syscall. MADV_COLD can be used by a process to mark a memory range as not expected to be used in the near future. The hint can help kernel in deciding which pages to evict early during memory pressure.

It works for every LRU pages like MADV_[DONTNEED|FREE]. IOW, It moves

	active file page -> inactive file LRU
	active anon page -> inacdtive anon LRU
Quoted 16 lines

Unlike MADV_FREE, it doesn't move active anonymous pages to inactive file LRU's head because MADV_COLD is a little bit different symantic. MADV_FREE means it's okay to discard when the memory pressure because the content of the page is *garbage* so freeing such pages is almost zero overhead since we don't need to swap out and access afterward causes just minor fault. Thus, it would make sense to put those freeable pages in inactive file LRU to compete other used-once pages. It makes sense for implmentaion point of view, too because it's not swapbacked memory any longer until it would be re-dirtied. Even, it could give a bonus to make them be reclaimed on swapless system. However, MADV_COLD doesn't mean garbage so reclaiming them requires swap-out/in in the end so it's bigger cost. Since we have designed VM LRU aging based on cost-model, anonymous cold pages would be better to position inactive anon's LRU list, not file LRU. Furthermore, it would help to avoid unnecessary scanning if system doesn't have a swap device. Let's start simpler way without adding complexity at this moment.

Quoted 3 lines

I would only add that it is a caveat that workloads with a lot of page cache are likely to ignore MADV_COLD on anonymous memory because we rarely age anonymous LRU lists.

Okay, I will add some more.

Quoted 9 lines

[...]

mm/madvise.c static long madvise_willneed(struct vm_area_struct *vma, · line 313

Quoted 3 lines
mm/madvise.c +3 -0
+static int madvise_cold_pte_range(pmd_t *pmd, unsigned long addr,
+				unsigned long end, struct mm_walk *walk)
+{

This is duplicating a large part of madvise_free_pte_range with some subtle differences which are not explained anywhere (e.g. why does madvise_free_huge_pmd need try_lock on a page while not here? etc.).

madvise_free_huge_pmd handle dirty bit but this is not.

Quoted 5 lines

Why cannot we reuse a large part of that code and differ essentially on the reclaim target check and action? Have you considered to consolidate the code to share as much as possible? Maybe that is easier said than done because the devil is always in details...

Yub, it was not pretty when I tried. Please see last patch in this patchset.

reviewer
#4 lore raw

Re: [PATCH v2 1/5] mm: introduce MADV_COLD

Quoted 1 line · Minchan Kim

On Wed, Jun 19, 2019 at 02:56:12PM +0200, Michal Hocko wrote:

[...]

Quoted 7 lines
Quoted 4 lines

Why cannot we reuse a large part of that code and differ essentially on the reclaim target check and action? Have you considered to consolidate the code to share as much as possible? Maybe that is easier said than done because the devil is always in details...

Yub, it was not pretty when I tried. Please see last patch in this patchset.

That is bad because this code is quite subtle - especially the THP part of it. I will be staring at the code some more. Maybe some simplification pops out.

Signature · 2 lines
Michal Hocko
SUSE Labs
#5 lore raw

Re: [PATCH v2 1/5] mm: introduce MADV_COLD

Quoted 8 lines · Michal Hocko
Quoted 1 line · Minchan Kim

On Wed, Jun 19, 2019 at 02:56:12PM +0200, Michal Hocko wrote:

[...]

Quoted 5 lines
Quoted 4 lines

Why cannot we reuse a large part of that code and differ essentially on the reclaim target check and action? Have you considered to consolidate the code to share as much as possible? Maybe that is easier said than done because the devil is always in details...

Quoted 6 lines
Quoted 2 lines

Yub, it was not pretty when I tried. Please see last patch in this patchset.

That is bad because this code is quite subtle - especially the THP part of it. I will be staring at the code some more. Maybe some simplification pops out.

Yeah, I couldn't come up with better idea. Actually, I wanted to be left. More suggestion to make simple/readable would be great.

#6 lore raw

PATCH v2 2/5 mm: change PAGEREF_RECLAIM_CLEAN with PAGE_REFRECLAIM

In reply to Minchan Kim

The local variable references in shrink_page_list is PAGEREF_RECLAIM_CLEAN as default. It is for preventing to reclaim dirty pages when CMA try to migrate pages. Strictly speaking, we don't need it because CMA didn't allow to write out by .may_writepage = 0 in reclaim_clean_pages_from_list.

Moreover, it has a problem to prevent anonymous pages's swap out even though force_reclaim = true in shrink_page_list on upcoming patch. So this patch makes references's default value to PAGEREF_RECLAIM and rename force_reclaim with ignore_references to make it more clear.

This is a preparatory work for next patch.

* RFCv1
* use ignore_referecnes as parameter name - hannes

---
 mm/vmscan.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
mm/vmscan.c +3 -3
index 84dcb651d05c..0973a46a0472 100644
@@ -1102,7 +1102,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 				      struct scan_control *sc,
 				      enum ttu_flags ttu_flags,
 				      struct reclaim_stat *stat,
-				      bool force_reclaim)
+				      bool ignore_references)
 {
 	LIST_HEAD(ret_pages);
 	LIST_HEAD(free_pages);
@@ -1116,7 +1116,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 		struct address_space *mapping;
 		struct page *page;
 		int may_enter_fs;
-		enum page_references references = PAGEREF_RECLAIM_CLEAN;
+		enum page_references references = PAGEREF_RECLAIM;
 		bool dirty, writeback;
 		unsigned int nr_pages;
 
@@ -1247,7 +1247,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 			}
 		}
 
-		if (!force_reclaim)
+		if (!ignore_references)
 			references = page_check_references(page, sc);
 
 		switch (references) {
Signature · 1 line
2.22.0.rc2.383.gf4fbbf30c2-goog
reviewer Acked-by on v2 2/5
#7 lore raw

Re: [PATCH v2 2/5] mm: change PAGEREF_RECLAIM_CLEAN with PAGE_REFRECLAIM

Quoted 10 lines · Minchan Kim

The local variable references in shrink_page_list is PAGEREF_RECLAIM_CLEAN as default. It is for preventing to reclaim dirty pages when CMA try to migrate pages. Strictly speaking, we don't need it because CMA didn't allow to write out by .may_writepage = 0 in reclaim_clean_pages_from_list.

Moreover, it has a problem to prevent anonymous pages's swap out even though force_reclaim = true in shrink_page_list on upcoming patch. So this patch makes references's default value to PAGEREF_RECLAIM and rename force_reclaim with ignore_references to make it more clear.

Quoted 7 lines

This is a preparatory work for next patch.

* RFCv1
* use ignore_referecnes as parameter name - hannes

The code path is quite tricky to follow but the patch looks OK to me.

Quoted 37 lines
---
 mm/vmscan.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
mm/vmscan.c +3 -3
index 84dcb651d05c..0973a46a0472 100644
@@ -1102,7 +1102,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 				      struct scan_control *sc,
 				      enum ttu_flags ttu_flags,
 				      struct reclaim_stat *stat,
-				      bool force_reclaim)
+				      bool ignore_references)
 {
 	LIST_HEAD(ret_pages);
 	LIST_HEAD(free_pages);
@@ -1116,7 +1116,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 		struct address_space *mapping;
 		struct page *page;
 		int may_enter_fs;
-		enum page_references references = PAGEREF_RECLAIM_CLEAN;
+		enum page_references references = PAGEREF_RECLAIM;
 		bool dirty, writeback;
 		unsigned int nr_pages;
 
@@ -1247,7 +1247,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 			}
 		}
 
-		if (!force_reclaim)
+		if (!ignore_references)
 			references = page_check_references(page, sc);
 
 		switch (references) {
Signature · 1 line
2.22.0.rc2.383.gf4fbbf30c2-goog
Signature · 2 lines
Michal Hocko
SUSE Labs
#8 lore raw

PATCH v2 3/5 mm: account nr_isolated_xxx in [isolate|putback]_lru_page

In reply to Minchan Kim

The isolate counting is pecpu counter so it would be not huge gain to work them by batch. Rather than complicating to make them batch, let's make it more stright-foward via adding the counting logic into [isolate|putback]_lru_page API.

* v1
* fix accounting bug - Hillf

---
 mm/compaction.c     |  2 --
 mm/gup.c            |  7 +------
 mm/khugepaged.c     |  3 ---
 mm/memory-failure.c |  3 ---
 mm/memory_hotplug.c |  4 ----
 mm/mempolicy.c      |  6 +-----
 mm/migrate.c        | 37 ++++++++-----------------------------
 mm/vmscan.c         | 22 ++++++++++++++++------
 8 files changed, 26 insertions(+), 58 deletions(-)
mm/compaction.c +0 -2
index 9e1b9acb116b..c6591682deda 100644
@@ -982,8 +982,6 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 
 		/* Successfully isolated */
 		del_page_from_lru_list(page, lruvec, page_lru(page));
-		inc_node_page_state(page,
-				NR_ISOLATED_ANON + page_is_file_cache(page));
 
 isolate_success:
 		list_add(&page->lru, &cc->migratepages);
mm/gup.c +1 -6
index 63ac50e48072..2d9a9bc358c7 100644
@@ -1360,13 +1360,8 @@ static long check_and_migrate_cma_pages(struct task_struct *tsk,
 					drain_allow = false;
 				}
 
-				if (!isolate_lru_page(head)) {
+				if (!isolate_lru_page(head))
 					list_add_tail(&head->lru, &cma_page_list);
-					mod_node_page_state(page_pgdat(head),
-							    NR_ISOLATED_ANON +
-							    page_is_file_cache(head),
-							    hpage_nr_pages(head));
-				}
 			}
 		}
 	}
mm/khugepaged.c +0 -3
index a335f7c1fac4..3359df994fb4 100644
@@ -503,7 +503,6 @@ void __khugepaged_exit(struct mm_struct *mm)
 
 static void release_pte_page(struct page *page)
 {
-	dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
 	unlock_page(page);
 	putback_lru_page(page);
 }
@@ -602,8 +601,6 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 			result = SCAN_DEL_PAGE_LRU;
 			goto out;
 		}
-		inc_node_page_state(page,
-				NR_ISOLATED_ANON + page_is_file_cache(page));
 		VM_BUG_ON_PAGE(!PageLocked(page), page);
 		VM_BUG_ON_PAGE(PageLRU(page), page);
 
mm/memory-failure.c +0 -3
index b9cc36a284f9..430946cf9c8a 100644
@@ -1796,9 +1796,6 @@ static int __soft_offline_page(struct page *page, int flags)
 		 * so use !__PageMovable instead for LRU page's mapping
 		 * cannot have PAGE_MAPPING_MOVABLE.
 		 */
-		if (!__PageMovable(page))
-			inc_node_page_state(page, NR_ISOLATED_ANON +
-						page_is_file_cache(page));
 		list_add(&page->lru, &pagelist);
 		ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
 					MIGRATE_SYNC, MR_MEMORY_FAILURE);
mm/memory_hotplug.c +0 -4
index a88c5f334e5a..a41bea24d0c9 100644
@@ -1390,10 +1390,6 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
 			ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
 		if (!ret) { /* Success */
 			list_add_tail(&page->lru, &source);
-			if (!__PageMovable(page))
-				inc_node_page_state(page, NR_ISOLATED_ANON +
-						    page_is_file_cache(page));
-
 		} else {
 			pr_warn("failed to isolate pfn %lx\n", pfn);
 			dump_page(page, "isolation failed");
mm/mempolicy.c +1 -5
index fdcb73536319..89bb25fe7553 100644
@@ -948,12 +948,8 @@ static void migrate_page_add(struct page *page, struct list_head *pagelist,
 	 * Avoid migrating a page that is shared with others.
 	 */
 	if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
-		if (!isolate_lru_page(head)) {
+		if (!isolate_lru_page(head))
 			list_add_tail(&head->lru, pagelist);
-			mod_node_page_state(page_pgdat(head),
-				NR_ISOLATED_ANON + page_is_file_cache(head),
-				hpage_nr_pages(head));
-		}
 	}
 }
 
mm/migrate.c +8 -29
index 572b4bc85d76..5583324c01e7 100644
@@ -190,8 +190,6 @@ void putback_movable_pages(struct list_head *l)
 			unlock_page(page);
 			put_page(page);
 		} else {
-			mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
-					page_is_file_cache(page), -hpage_nr_pages(page));
 			putback_lru_page(page);
 		}
 	}
@@ -1181,10 +1179,17 @@ static ICE_noinline int unmap_and_move(new_page_t get_new_page,
 		return -ENOMEM;
 
 	if (page_count(page) == 1) {
+		bool is_lru = !__PageMovable(page);
+
 		/* page was freed from under us. So we are done. */
 		ClearPageActive(page);
 		ClearPageUnevictable(page);
-		if (unlikely(__PageMovable(page))) {
+		if (likely(is_lru))
+			mod_node_page_state(page_pgdat(page),
+						NR_ISOLATED_ANON +
+						page_is_file_cache(page),
+						-hpage_nr_pages(page));
+		else {
 			lock_page(page);
 			if (!PageMovable(page))
 				__ClearPageIsolated(page);
@@ -1210,15 +1215,6 @@ static ICE_noinline int unmap_and_move(new_page_t get_new_page,
 		 * restored.
 		 */
 		list_del(&page->lru);
-
-		/*
-		 * Compaction can migrate also non-LRU pages which are
-		 * not accounted to NR_ISOLATED_*. They can be recognized
-		 * as __PageMovable
-		 */
-		if (likely(!__PageMovable(page)))
-			mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
-					page_is_file_cache(page), -hpage_nr_pages(page));
 	}
 
 	/*
@@ -1572,9 +1568,6 @@ static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
 
 		err = 0;
 		list_add_tail(&head->lru, pagelist);
-		mod_node_page_state(page_pgdat(head),
-			NR_ISOLATED_ANON + page_is_file_cache(head),
-			hpage_nr_pages(head));
 	}
 out_putpage:
 	/*
@@ -1890,8 +1883,6 @@ static struct page *alloc_misplaced_dst_page(struct page *page,
 
 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
 {
-	int page_lru;
-
 	VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
 
 	/* Avoid migrating to a node that is nearly full */
@@ -1913,10 +1904,6 @@ static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
 		return 0;
 	}
 
-	page_lru = page_is_file_cache(page);
-	mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
-				hpage_nr_pages(page));
-
 	/*
 	 * Isolating the page has taken another reference, so the
 	 * caller's reference can be safely dropped without the page
@@ -1971,8 +1958,6 @@ int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
 	if (nr_remaining) {
 		if (!list_empty(&migratepages)) {
 			list_del(&page->lru);
-			dec_node_page_state(page, NR_ISOLATED_ANON +
-					page_is_file_cache(page));
 			putback_lru_page(page);
 		}
 		isolated = 0;
@@ -2002,7 +1987,6 @@ int migrate_misplaced_transhuge_page(struct mm_struct *mm,
 	pg_data_t *pgdat = NODE_DATA(node);
 	int isolated = 0;
 	struct page *new_page = NULL;
-	int page_lru = page_is_file_cache(page);
 	unsigned long start = address & HPAGE_PMD_MASK;
 
 	new_page = alloc_pages_node(node,
@@ -2048,8 +2032,6 @@ int migrate_misplaced_transhuge_page(struct mm_struct *mm,
 		/* Retake the callers reference and putback on LRU */
 		get_page(page);
 		putback_lru_page(page);
-		mod_node_page_state(page_pgdat(page),
-			 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
 
 		goto out_unlock;
 	}
@@ -2099,9 +2081,6 @@ int migrate_misplaced_transhuge_page(struct mm_struct *mm,
 	count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
 	count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
 
-	mod_node_page_state(page_pgdat(page),
-			NR_ISOLATED_ANON + page_lru,
-			-HPAGE_PMD_NR);
 	return isolated;
 
 out_fail:
mm/vmscan.c +16 -6
index 0973a46a0472..56df55e8afcd 100644
@@ -999,6 +999,9 @@ int remove_mapping(struct address_space *mapping, struct page *page)
 void putback_lru_page(struct page *page)
 {
 	lru_cache_add(page);
+	mod_node_page_state(page_pgdat(page),
+				NR_ISOLATED_ANON + page_is_file_cache(page),
+				-hpage_nr_pages(page));
 	put_page(page);		/* drop ref from isolate */
 }
 
@@ -1464,6 +1467,9 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 		 */
 		nr_reclaimed += nr_pages;
 
+		mod_node_page_state(pgdat, NR_ISOLATED_ANON +
+						page_is_file_cache(page),
+						-nr_pages);
 		/*
 		 * Is there need to periodically free_page_list? It would
 		 * appear not as the counts should be low
@@ -1539,7 +1545,6 @@ unsigned long reclaim_clean_pages_from_list(struct zone *zone,
 	ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
 			TTU_IGNORE_ACCESS, &dummy_stat, true);
 	list_splice(&clean_pages, page_list);
-	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
 	return ret;
 }
 
@@ -1615,6 +1620,9 @@ int __isolate_lru_page(struct page *page, isolate_mode_t mode)
 		 */
 		ClearPageLRU(page);
 		ret = 0;
+		__mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
+						page_is_file_cache(page),
+						hpage_nr_pages(page));
 	}
 
 	return ret;
@@ -1746,6 +1754,7 @@ static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
 	trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
 				    total_scan, skipped, nr_taken, mode, lru);
 	update_lru_sizes(lruvec, lru, nr_zone_taken);
+
 	return nr_taken;
 }
 
@@ -1794,6 +1803,9 @@ int isolate_lru_page(struct page *page)
 			ClearPageLRU(page);
 			del_page_from_lru_list(page, lruvec, lru);
 			ret = 0;
+			mod_node_page_state(pgdat, NR_ISOLATED_ANON +
+						page_is_file_cache(page),
+						hpage_nr_pages(page));
 		}
 		spin_unlock_irq(&pgdat->lru_lock);
 	}
@@ -1885,6 +1897,9 @@ static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
 		update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
 		list_move(&page->lru, &lruvec->lists[lru]);
 
+		__mod_node_page_state(pgdat, NR_ISOLATED_ANON +
+						page_is_file_cache(page),
+						-hpage_nr_pages(page));
 		if (put_page_testzero(page)) {
 			__ClearPageLRU(page);
 			__ClearPageActive(page);
@@ -1962,7 +1977,6 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
 	nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
 				     &nr_scanned, sc, lru);
 
-	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
 	reclaim_stat->recent_scanned[file] += nr_taken;
 
 	item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
@@ -1988,8 +2002,6 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
 
 	move_pages_to_lru(lruvec, &page_list);
 
-	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
-
 	spin_unlock_irq(&pgdat->lru_lock);
 
 	mem_cgroup_uncharge_list(&page_list);
@@ -2048,7 +2060,6 @@ static void shrink_active_list(unsigned long nr_to_scan,
 	nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
 				     &nr_scanned, sc, lru);
 
-	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
 	reclaim_stat->recent_scanned[file] += nr_taken;
 
 	__count_vm_events(PGREFILL, nr_scanned);
@@ -2117,7 +2128,6 @@ static void shrink_active_list(unsigned long nr_to_scan,
 	__count_vm_events(PGDEACTIVATE, nr_deactivate);
 	__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
 
-	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
 	spin_unlock_irq(&pgdat->lru_lock);
 
 	mem_cgroup_uncharge_list(&l_active);
Signature · 1 line
2.22.0.rc2.383.gf4fbbf30c2-goog
#9 lore raw

PATCH v2 4/5 mm: introduce MADV_PAGEOUT

In reply to Minchan Kim

When a process expects no accesses to a certain memory range for a long time, it could hint kernel that the pages can be reclaimed instantly but data should be preserved for future use. This could reduce workingset eviction so it ends up increasing performance.

This patch introduces the new MADV_PAGEOUT hint to madvise(2) syscall. MADV_PAGEOUT can be used by a process to mark a memory range as not expected to be used for a long time so that kernel reclaims *any LRU* pages instantly. The hint can help kernel in deciding which pages to evict proactively.

All of error rule is same with MADV_DONTNEED.

* v1
* change pte to old and rely on the other's reference - hannes
* remove page_mapcount to check shared page - mhocko

* RFC v2
* make reclaim_pages simple via factoring out isolate logic - hannes

* RFCv1
* rename from MADV_COLD to MADV_PAGEOUT - hannes
* bail out if process is being killed - Hillf
* fix reclaim_pages bugs - Hillf

---
 include/linux/swap.h                   |   1 +
 include/uapi/asm-generic/mman-common.h |   1 +
 mm/madvise.c                           | 161 +++++++++++++++++++++++++
 mm/vmscan.c                            |  58 +++++++++
 4 files changed, 221 insertions(+)
include/linux/swap.h +1 -0
index 0ce997edb8bb..063c0c1e112b 100644
@@ -365,6 +365,7 @@ extern int vm_swappiness;
 extern int remove_mapping(struct address_space *mapping, struct page *page);
 extern unsigned long vm_total_pages;
 
+extern unsigned long reclaim_pages(struct list_head *page_list);
 #ifdef CONFIG_NUMA
 extern int node_reclaim_mode;
 extern int sysctl_min_unmapped_ratio;
include/uapi/asm-generic/mman-common.h +1 -0
index d7b4231eea63..f545e159b472 100644
@@ -48,6 +48,7 @@
 #define MADV_WILLNEED	3		/* will need these pages */
 #define MADV_DONTNEED	4		/* don't need these pages */
 #define MADV_COLD	5		/* deactivatie these pages */
+#define MADV_PAGEOUT	6		/* reclaim these pages */
 
 /* common parameters: try to keep these consistent across architectures */
 #define MADV_FREE	8		/* free pages only if memory pressure */
mm/madvise.c +161 -0

191 lines not on this page · raw message

mm/vmscan.c +58 -0

64 lines not on this page · raw message

Signature · 1 line
2.22.0.rc2.383.gf4fbbf30c2-goog
reviewer
#10 lore raw

Re: [PATCH v2 4/5] mm: introduce MADV_PAGEOUT

On Mon 10-06-19 20:12:51, Minchan Kim wrote:
[...]

mm/madvise.c static long madvise_cold(struct vm_area_struct *vma, · line 457

Quoted 2 lines
mm/madvise.c +2 -0
+static int madvise_pageout_pte_range(pmd_t *pmd, unsigned long addr,
+				unsigned long end, struct mm_walk *walk)

Again the same question about a potential code reuse... [...]

Quoted 19 lines
mm/madvise.c +19 -0
+regular_page:
+	tlb_change_page_size(tlb, PAGE_SIZE);
+	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+	flush_tlb_batched_pending(mm);
+	arch_enter_lazy_mmu_mode();
+	for (; addr < end; pte++, addr += PAGE_SIZE) {
+		ptent = *pte;
+		if (!pte_present(ptent))
+			continue;
+
+		page = vm_normal_page(vma, addr, ptent);
+		if (!page)
+			continue;
+
+		if (isolate_lru_page(page))
+			continue;
+
+		isolated++;
+		if (pte_young(ptent)) {

mm/madvise.c static long madvise_cold(struct vm_area_struct *vma, · line 558

Quoted 10 lines
mm/madvise.c +10 -0
+			ptent = ptep_get_and_clear_full(mm, addr, pte,
+							tlb->fullmm);
+			ptent = pte_mkold(ptent);
+			set_pte_at(mm, addr, pte, ptent);
+			tlb_remove_tlb_entry(tlb, pte, addr);
+		}
+		ClearPageReferenced(page);
+		test_and_clear_page_young(page);
+		list_add(&page->lru, &page_list);
+		if (isolated >= SWAP_CLUSTER_MAX) {

Why do we need SWAP_CLUSTER_MAX batching? Especially when we need ... [...]

Quoted 22 lines
mm/vmscan.c +22 -0
+unsigned long reclaim_pages(struct list_head *page_list)
+{
+	int nid = -1;
+	unsigned long nr_reclaimed = 0;
+	LIST_HEAD(node_page_list);
+	struct reclaim_stat dummy_stat;
+	struct scan_control sc = {
+		.gfp_mask = GFP_KERNEL,
+		.priority = DEF_PRIORITY,
+		.may_writepage = 1,
+		.may_unmap = 1,
+		.may_swap = 1,
+	};
+
+	while (!list_empty(page_list)) {
+		struct page *page;
+
+		page = lru_to_page(page_list);
+		if (nid == -1) {
+			nid = page_to_nid(page);
+			INIT_LIST_HEAD(&node_page_list);
+		}

mm/vmscan.c static void shrink_active_list(unsigned long nr_to_scan, · line 2170

Quoted 10 lines
mm/vmscan.c +10 -0
+
+		if (nid == page_to_nid(page)) {
+			list_move(&page->lru, &node_page_list);
+			continue;
+		}
+
+		nr_reclaimed += shrink_page_list(&node_page_list,
+						NODE_DATA(nid),
+						&sc, 0,
+						&dummy_stat, false);

per-node batching in fact. Other than that nothing really jumped at me. Except for the shared page cache side channel timing aspect not being considered AFAICS. To be more specific. Pushing out a shared page cache is possible even now but this interface gives a much easier tool to evict shared state and perform all sorts of timing attacks. Unless I am missing something we should be doing something similar to mincore and ignore shared pages without a writeable access or at least document why we do not care.

Signature · 2 lines
Michal Hocko
SUSE Labs
#11 lore raw

Re: [PATCH v2 4/5] mm: introduce MADV_PAGEOUT

Quoted 31 lines · Michal Hocko

On Mon 10-06-19 20:12:51, Minchan Kim wrote:
[...]

mm/madvise.c static long madvise_cold(struct vm_area_struct *vma, · line 457

Quoted 2 lines
mm/madvise.c +2 -0
+static int madvise_pageout_pte_range(pmd_t *pmd, unsigned long addr,
+				unsigned long end, struct mm_walk *walk)

Again the same question about a potential code reuse... [...]

Quoted 24 lines
mm/madvise.c +24 -0
+regular_page:
+	tlb_change_page_size(tlb, PAGE_SIZE);
+	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+	flush_tlb_batched_pending(mm);
+	arch_enter_lazy_mmu_mode();
+	for (; addr < end; pte++, addr += PAGE_SIZE) {
+		ptent = *pte;
+		if (!pte_present(ptent))
+			continue;
+
+		page = vm_normal_page(vma, addr, ptent);
+		if (!page)
+			continue;
+
+		if (isolate_lru_page(page))
+			continue;
+
+		isolated++;
+		if (pte_young(ptent)) {
+			ptent = ptep_get_and_clear_full(mm, addr, pte,
+							tlb->fullmm);
+			ptent = pte_mkold(ptent);
+			set_pte_at(mm, addr, pte, ptent);
+			tlb_remove_tlb_entry(tlb, pte, addr);
Quoted 8 lines

mm/madvise.c static long madvise_cold(struct vm_area_struct *vma, · line 558

Quoted 5 lines
mm/madvise.c +5 -0
+		}
+		ClearPageReferenced(page);
+		test_and_clear_page_young(page);
+		list_add(&page->lru, &page_list);
+		if (isolated >= SWAP_CLUSTER_MAX) {

Why do we need SWAP_CLUSTER_MAX batching? Especially when we need ... [...]

It aims for preventing early OOM kill since we isolate too many LRU pages concurrently.

Quoted 34 lines
Quoted 23 lines
mm/vmscan.c +22 -0
+unsigned long reclaim_pages(struct list_head *page_list)
+{
+	int nid = -1;
+	unsigned long nr_reclaimed = 0;
+	LIST_HEAD(node_page_list);
+	struct reclaim_stat dummy_stat;
+	struct scan_control sc = {
+		.gfp_mask = GFP_KERNEL,
+		.priority = DEF_PRIORITY,
+		.may_writepage = 1,
+		.may_unmap = 1,
+		.may_swap = 1,
+	};
+
+	while (!list_empty(page_list)) {
+		struct page *page;
+
+		page = lru_to_page(page_list);
+		if (nid == -1) {
+			nid = page_to_nid(page);
+			INIT_LIST_HEAD(&node_page_list);
+		}

mm/vmscan.c static void shrink_active_list(unsigned long nr_to_scan, · line 2170

Quoted 10 lines
mm/vmscan.c +10 -0
+
+		if (nid == page_to_nid(page)) {
+			list_move(&page->lru, &node_page_list);
+			continue;
+		}
+
+		nr_reclaimed += shrink_page_list(&node_page_list,
+						NODE_DATA(nid),
+						&sc, 0,
+						&dummy_stat, false);
Quoted 8 lines

per-node batching in fact. Other than that nothing really jumped at me. Except for the shared page cache side channel timing aspect not being considered AFAICS. To be more specific. Pushing out a shared page cache is possible even now but this interface gives a much easier tool to evict shared state and perform all sorts of timing attacks. Unless I am missing something we should be doing something similar to mincore and ignore shared pages without a writeable access or at least document why we do not care.

I'm not sure IIUC side channel attach. As you mentioned, without this syscall,
1. they already can do that simply by memory hogging
2. If we need fix MADV_PAGEOUT, that means we need to fix MADV_DONTNEED, too?

reviewer
#12 lore raw

Re: [PATCH v2 4/5] mm: introduce MADV_PAGEOUT

Quoted 41 lines · Minchan Kim
Quoted 31 lines · Michal Hocko

On Mon 10-06-19 20:12:51, Minchan Kim wrote:
[...]

mm/madvise.c static long madvise_cold(struct vm_area_struct *vma, · line 457

Quoted 2 lines
mm/madvise.c +2 -0
+static int madvise_pageout_pte_range(pmd_t *pmd, unsigned long addr,
+				unsigned long end, struct mm_walk *walk)

Again the same question about a potential code reuse... [...]

Quoted 24 lines
mm/madvise.c +24 -0
+regular_page:
+	tlb_change_page_size(tlb, PAGE_SIZE);
+	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+	flush_tlb_batched_pending(mm);
+	arch_enter_lazy_mmu_mode();
+	for (; addr < end; pte++, addr += PAGE_SIZE) {
+		ptent = *pte;
+		if (!pte_present(ptent))
+			continue;
+
+		page = vm_normal_page(vma, addr, ptent);
+		if (!page)
+			continue;
+
+		if (isolate_lru_page(page))
+			continue;
+
+		isolated++;
+		if (pte_young(ptent)) {
+			ptent = ptep_get_and_clear_full(mm, addr, pte,
+							tlb->fullmm);
+			ptent = pte_mkold(ptent);
+			set_pte_at(mm, addr, pte, ptent);
+			tlb_remove_tlb_entry(tlb, pte, addr);
Quoted 8 lines

mm/madvise.c static long madvise_cold(struct vm_area_struct *vma, · line 558

Quoted 5 lines
mm/madvise.c +5 -0
+		}
+		ClearPageReferenced(page);
+		test_and_clear_page_young(page);
+		list_add(&page->lru, &page_list);
+		if (isolated >= SWAP_CLUSTER_MAX) {

Why do we need SWAP_CLUSTER_MAX batching? Especially when we need ... [...]

Quoted 2 lines

It aims for preventing early OOM kill since we isolate too many LRU pages concurrently.

This is a good point. For some reason I thought that we consider isolated pages in should_reclaim_retry but we do not anymore (since we move from zone to node LRUs I guess). Please stick a comment there.

Quoted 42 lines
Quoted 33 lines
Quoted 22 lines
mm/vmscan.c +22 -0
+unsigned long reclaim_pages(struct list_head *page_list)
+{
+	int nid = -1;
+	unsigned long nr_reclaimed = 0;
+	LIST_HEAD(node_page_list);
+	struct reclaim_stat dummy_stat;
+	struct scan_control sc = {
+		.gfp_mask = GFP_KERNEL,
+		.priority = DEF_PRIORITY,
+		.may_writepage = 1,
+		.may_unmap = 1,
+		.may_swap = 1,
+	};
+
+	while (!list_empty(page_list)) {
+		struct page *page;
+
+		page = lru_to_page(page_list);
+		if (nid == -1) {
+			nid = page_to_nid(page);
+			INIT_LIST_HEAD(&node_page_list);
+		}

mm/vmscan.c static void shrink_active_list(unsigned long nr_to_scan, · line 2170

Quoted 10 lines
mm/vmscan.c +10 -0
+
+		if (nid == page_to_nid(page)) {
+			list_move(&page->lru, &node_page_list);
+			continue;
+		}
+
+		nr_reclaimed += shrink_page_list(&node_page_list,
+						NODE_DATA(nid),
+						&sc, 0,
+						&dummy_stat, false);
Quoted 8 lines

per-node batching in fact. Other than that nothing really jumped at me. Except for the shared page cache side channel timing aspect not being considered AFAICS. To be more specific. Pushing out a shared page cache is possible even now but this interface gives a much easier tool to evict shared state and perform all sorts of timing attacks. Unless I am missing something we should be doing something similar to mincore and ignore shared pages without a writeable access or at least document why we do not care.

Quoted 2 lines

I'm not sure IIUC side channel attach. As you mentioned, without this syscall,
1. they already can do that simply by memory hogging

This is way much more harder for practical attacks because the reclaim logic is not fully under the attackers control. Having a direct tool to reclaim memory directly then just opens doors to measure the other consumers of that memory and all sorts of side channel.

Quoted 1 line

2. If we need fix MADV_PAGEOUT, that means we need to fix MADV_DONTNEED, too?

nope because MADV_DONTNEED doesn't unmap from other processes.

Signature · 2 lines
Michal Hocko
SUSE Labs
#13 lore raw

Re: [PATCH v2 4/5] mm: introduce MADV_PAGEOUT

Quoted 42 lines · Michal Hocko
Quoted 41 lines · Minchan Kim
Quoted 39 lines · Michal Hocko

On Mon 10-06-19 20:12:51, Minchan Kim wrote:
[...]

mm/madvise.c static long madvise_cold(struct vm_area_struct *vma, · line 457

Quoted 2 lines
mm/madvise.c +2 -0
+static int madvise_pageout_pte_range(pmd_t *pmd, unsigned long addr,
+				unsigned long end, struct mm_walk *walk)

Again the same question about a potential code reuse... [...]

Quoted 19 lines
mm/madvise.c +19 -0
+regular_page:
+	tlb_change_page_size(tlb, PAGE_SIZE);
+	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+	flush_tlb_batched_pending(mm);
+	arch_enter_lazy_mmu_mode();
+	for (; addr < end; pte++, addr += PAGE_SIZE) {
+		ptent = *pte;
+		if (!pte_present(ptent))
+			continue;
+
+		page = vm_normal_page(vma, addr, ptent);
+		if (!page)
+			continue;
+
+		if (isolate_lru_page(page))
+			continue;
+
+		isolated++;
+		if (pte_young(ptent)) {

mm/madvise.c static long madvise_cold(struct vm_area_struct *vma, · line 558

Quoted 10 lines
mm/madvise.c +10 -0
+			ptent = ptep_get_and_clear_full(mm, addr, pte,
+							tlb->fullmm);
+			ptent = pte_mkold(ptent);
+			set_pte_at(mm, addr, pte, ptent);
+			tlb_remove_tlb_entry(tlb, pte, addr);
+		}
+		ClearPageReferenced(page);
+		test_and_clear_page_young(page);
+		list_add(&page->lru, &page_list);
+		if (isolated >= SWAP_CLUSTER_MAX) {

Why do we need SWAP_CLUSTER_MAX batching? Especially when we need ... [...]

Quoted 6 lines
Quoted 2 lines

It aims for preventing early OOM kill since we isolate too many LRU pages concurrently.

This is a good point. For some reason I thought that we consider isolated pages in should_reclaim_retry but we do not anymore (since we move from zone to node LRUs I guess). Please stick a comment there.

Sure.

Quoted 43 lines
Quoted 43 lines
Quoted 42 lines
Quoted 23 lines
mm/vmscan.c +22 -0
+unsigned long reclaim_pages(struct list_head *page_list)
+{
+	int nid = -1;
+	unsigned long nr_reclaimed = 0;
+	LIST_HEAD(node_page_list);
+	struct reclaim_stat dummy_stat;
+	struct scan_control sc = {
+		.gfp_mask = GFP_KERNEL,
+		.priority = DEF_PRIORITY,
+		.may_writepage = 1,
+		.may_unmap = 1,
+		.may_swap = 1,
+	};
+
+	while (!list_empty(page_list)) {
+		struct page *page;
+
+		page = lru_to_page(page_list);
+		if (nid == -1) {
+			nid = page_to_nid(page);
+			INIT_LIST_HEAD(&node_page_list);
+		}

mm/vmscan.c static void shrink_active_list(unsigned long nr_to_scan, · line 2170

Quoted 10 lines
mm/vmscan.c +10 -0
+
+		if (nid == page_to_nid(page)) {
+			list_move(&page->lru, &node_page_list);
+			continue;
+		}
+
+		nr_reclaimed += shrink_page_list(&node_page_list,
+						NODE_DATA(nid),
+						&sc, 0,
+						&dummy_stat, false);

per-node batching in fact. Other than that nothing really jumped at me. Except for the shared page cache side channel timing aspect not being considered AFAICS. To be more specific. Pushing out a shared page cache is possible even now but this interface gives a much easier tool to evict shared state and perform all sorts of timing attacks. Unless I am missing something we should be doing something similar to mincore and ignore shared pages without a writeable access or at least document why we do not care.

Quoted 7 lines
Quoted 2 lines

I'm not sure IIUC side channel attach. As you mentioned, without this syscall,
1. they already can do that simply by memory hogging

This is way much more harder for practical attacks because the reclaim logic is not fully under the attackers control. Having a direct tool to reclaim memory directly then just opens doors to measure the other consumers of that memory and all sorts of side channel.

Not sure it's much more harder. It's really easy on my experience. Just creating new memory hogger and consume memory step by step until you newly allocated pages will be reclaimed.
Anyway, we fixed mincore so attacker cannot see when the page fault-in if he don't enough permission for the file. Right?
What's the concern of you even though we reclaim more aggressively?

Quoted 4 lines
Quoted 2 lines

2. If we need fix MADV_PAGEOUT, that means we need to fix MADV_DONTNEED, too?

nope because MADV_DONTNEED doesn't unmap from other processes.

Hmm, I don't understand. MADV_PAGEOUT doesn't unmap from other processes, either. Could you elborate it a bit more what's your concern?

Quoted 3 lines
Signature · 2 lines
Michal Hocko
SUSE Labs
reviewer
#14 lore raw

Re: [PATCH v2 4/5] mm: introduce MADV_PAGEOUT

Quoted 10 lines · Minchan Kim
Quoted 10 lines
Quoted 9 lines
Quoted 6 lines

Pushing out a shared page cache
is possible even now but this interface gives a much easier tool to evict shared state and perform all sorts of timing attacks. Unless I am missing something we should be doing something similar to mincore and ignore shared pages without a writeable access or at least document why we do not care.

I'm not sure IIUC side channel attach. As you mentioned, without this syscall,
1. they already can do that simply by memory hogging

Quoted 8 lines
Quoted 4 lines

This is way much more harder for practical attacks because the reclaim logic is not fully under the attackers control. Having a direct tool to reclaim memory directly then just opens doors to measure the other consumers of that memory and all sorts of side channel.

Not sure it's much more harder. It's really easy on my experience. Just creating new memory hogger and consume memory step by step until you newly allocated pages will be reclaimed.

You can contain an untrusted application into a memcg and it will only reclaim its own working set.

Quoted 6 lines
Quoted 3 lines
Quoted 1 line

2. If we need fix MADV_PAGEOUT, that means we need to fix MADV_DONTNEED, too?

nope because MADV_DONTNEED doesn't unmap from other processes.

Hmm, I don't understand. MADV_PAGEOUT doesn't unmap from other processes, either.

Either I am confused or missing something. shrink_page_list does try_to_unmap and that unmaps from all processes, right?

Quoted 1 line

Could you elborate it a bit more what's your concern?

If you manage to unmap from a remote process then you can measure delays implied from the refault and that information can be used to infer what the remote application is doing.

Signature · 2 lines
Michal Hocko
SUSE Labs
#15 lore raw

Re: [PATCH v2 4/5] mm: introduce MADV_PAGEOUT

Quoted 25 lines · Michal Hocko
Quoted 10 lines · Minchan Kim
Quoted 10 lines
Quoted 9 lines
Quoted 6 lines

Pushing out a shared page cache
is possible even now but this interface gives a much easier tool to evict shared state and perform all sorts of timing attacks. Unless I am missing something we should be doing something similar to mincore and ignore shared pages without a writeable access or at least document why we do not care.

I'm not sure IIUC side channel attach. As you mentioned, without this syscall,
1. they already can do that simply by memory hogging

Quoted 8 lines
Quoted 4 lines

This is way much more harder for practical attacks because the reclaim logic is not fully under the attackers control. Having a direct tool to reclaim memory directly then just opens doors to measure the other consumers of that memory and all sorts of side channel.

Not sure it's much more harder. It's really easy on my experience. Just creating new memory hogger and consume memory step by step until you newly allocated pages will be reclaimed.

You can contain an untrusted application into a memcg and it will only reclaim its own working set.

Quoted 2 lines
Quoted 2 lines
Quoted 1 line

2. If we need fix MADV_PAGEOUT, that means we need to fix MADV_DONTNEED, too?

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

nope because MADV_DONTNEED doesn't unmap from other processes.

Hmm, I don't understand. MADV_PAGEOUT doesn't unmap from other processes, either.

Either I am confused or missing something. shrink_page_list does try_to_unmap and that unmaps from all processes, right?

You don't miss it. It seems now I undetstand what you pointed out. What you meant is attacker can see what page was faulting-in from other processes via measuring access delay from his address space and MADV_PAGEOUT makes it more easiler. Thus, it's an issue regardless of recent mincore fix. Right? Then, okay, I will add can_do_mincore similar check for the MADV_PAGEOUT syscall if others have different ideas.

Thanks.

reviewer
#16 lore raw

Re: [PATCH v2 4/5] mm: introduce MADV_PAGEOUT

On Thu 20-06-19 19:32:15, Minchan Kim wrote:
[...]

Quoted 2 lines

Then, okay, I will add can_do_mincore similar check for the MADV_PAGEOUT syscall if others have different ideas.

Great that we are on the same page. We can simply skip over those pages.

Signature · 2 lines
Michal Hocko
SUSE Labs
#17 lore raw

PATCH v2 5/5 mm: factor out pmd young/dirty bit handling and THP split

In reply to Minchan Kim

Now, there are common part among MADV_COLD|PAGEOUT|FREE to reset access/dirty bit resetting or split the THP page to handle part of subpages in the THP page. This patch factor out the common part.

Cc "Kirill A. Shutemov" <[email protected]>
Cc Christopher Lameter <[email protected]>
---
 include/linux/huge_mm.h |   3 -
 mm/huge_memory.c        |  74 -------------
 mm/madvise.c            | 234 +++++++++++++++++++++++-----------------
 3 files changed, 135 insertions(+), 176 deletions(-)
include/linux/huge_mm.h +0 -3
index 7cd5c150c21d..2667e1aa3ce5 100644
@@ -29,9 +29,6 @@ extern struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
 					  unsigned long addr,
 					  pmd_t *pmd,
 					  unsigned int flags);
-extern bool madvise_free_huge_pmd(struct mmu_gather *tlb,
-			struct vm_area_struct *vma,
-			pmd_t *pmd, unsigned long addr, unsigned long next);
 extern int zap_huge_pmd(struct mmu_gather *tlb,
 			struct vm_area_struct *vma,
 			pmd_t *pmd, unsigned long addr);
mm/huge_memory.c +0 -74

80 lines not on this page · raw message

mm/madvise.c +135 -99

307 lines not on this page · raw message

Signature · 1 line
2.22.0.rc2.383.gf4fbbf30c2-goog
#19 lore raw
Quoted 2 lines · Dave Hansen

I'd really love to see the manpages for these new flags. The devil is in the details of our promises to userspace.

I'm waiting comments from reviewers since I have fixed what they point out from the previous version.

I will add manpage material in respin after the getting more feedback. Thanks.

#20 lore raw
In reply to Minchan Kim

Hi!

Quoted 12 lines

- Problem

Naturally, cached apps were dominant consumers of memory on the system. However, they were not significant consumers of swap even though they are good candidate for swap. Under investigation, swapping out only begins once the low zone watermark is hit and kswapd wakes up, but the overall allocation rate in the system might trip lmkd thresholds and cause a cached process to be killed(we measured performance swapping out vs. zapping the memory by killing a process. Unsurprisingly, zapping is 10x times faster even though we use zram which is much faster than real storage) so kill from lmkd will often satisfy the high zone watermark, resulting in very few pages actually being moved to swap.

Is it still faster to swap-in the application than to restart it?

Quoted 6 lines

This approach is similar in spirit to madvise(MADV_WONTNEED), but the information required to make the reclaim decision is not known to the app. Instead, it is known to a centralized userspace daemon, and that daemon must be able to initiate reclaim on its own without any app involvement. To solve the concern, this patch introduces new syscall -

Quoted 13 lines
    struct pr_madvise_param {
            int size;               /* the size of this structure */
            int cookie;             /* reserved to support atomicity */
            int nr_elem;            /* count of below arrary fields */
            int __user *hints;      /* hints for each range */
            /* to store result of each operation */
            const struct iovec __user *results;
            /* input address ranges */
            const struct iovec __user *ranges;
    };
    
    int process_madvise(int pidfd, struct pr_madvise_param *u_param,
                            unsigned long flags);

That's quite a complex interface.

Could we simply have feel_free_to_swap_out(int pid) syscall? :-).

Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
#21 lore raw
Quoted 13 lines · Pavel Machek
Quoted 12 lines

- Problem

Naturally, cached apps were dominant consumers of memory on the system. However, they were not significant consumers of swap even though they are good candidate for swap. Under investigation, swapping out only begins once the low zone watermark is hit and kswapd wakes up, but the overall allocation rate in the system might trip lmkd thresholds and cause a cached process to be killed(we measured performance swapping out vs. zapping the memory by killing a process. Unsurprisingly, zapping is 10x times faster even though we use zram which is much faster than real storage) so kill from lmkd will often satisfy the high zone watermark, resulting in very few pages actually being moved to swap.

Quoted 1 line

Is it still faster to swap-in the application than to restart it?

It's the same type of question I was addressing earlier in the remote KSM discussion: making applications aware of all the memory management stuff or delegate the decision to some supervising task.

In this case, we cannot rewrite all the application to handle imaginary SIGRESTART (or whatever you invent to handle restarts gracefully). SIGTERM may require more memory to finish stuff to not lose your data (and I guess you don't want to lose your data, right?), and SIGKILL is pretty much destructive.

Offloading proactive memory management to a process that knows how to do it allows to handle not only throwaway containers/microservices, but also usual desktop/mobile workflow.

Quoted 17 lines
Quoted 17 lines

This approach is similar in spirit to madvise(MADV_WONTNEED), but the information required to make the reclaim decision is not known to the app. Instead, it is known to a centralized userspace daemon, and that daemon must be able to initiate reclaim on its own without any app involvement. To solve the concern, this patch introduces new syscall -

    struct pr_madvise_param {
            int size;               /* the size of this structure */
            int cookie;             /* reserved to support atomicity */
            int nr_elem;            /* count of below arrary fields */
            int __user *hints;      /* hints for each range */
            /* to store result of each operation */
            const struct iovec __user *results;
            /* input address ranges */
            const struct iovec __user *ranges;
    };
Quoted 6 lines
Quoted 2 lines

int process_madvise(int pidfd, struct pr_madvise_param *u_param,
unsigned long flags);

That's quite a complex interface.

Could we simply have feel_free_to_swap_out(int pid) syscall? :-).

I wonder for how long we'll go on with adding new syscalls each time we need some amendment to existing interfaces. Yes, clone6(), I'm looking at you :(.

In case of process_madvise() keep in mind it will be focused not only on MADV_COLD, but also, potentially, on other MADV_ flags as well. I can hardly imagine we'll add one syscall per each flag.

Signature · 3 lines
  Best regards,
    Oleksandr Natalenko (post-factum)
    Senior Software Maintenance Engineer
#22 lore raw

Hi!

Quoted 24 lines
Quoted 17 lines
Quoted 17 lines

This approach is similar in spirit to madvise(MADV_WONTNEED), but the information required to make the reclaim decision is not known to the app. Instead, it is known to a centralized userspace daemon, and that daemon must be able to initiate reclaim on its own without any app involvement. To solve the concern, this patch introduces new syscall -

    struct pr_madvise_param {
            int size;               /* the size of this structure */
            int cookie;             /* reserved to support atomicity */
            int nr_elem;            /* count of below arrary fields */
            int __user *hints;      /* hints for each range */
            /* to store result of each operation */
            const struct iovec __user *results;
            /* input address ranges */
            const struct iovec __user *ranges;
    };
Quoted 6 lines
Quoted 2 lines

int process_madvise(int pidfd, struct pr_madvise_param *u_param,
unsigned long flags);

That's quite a complex interface.

Could we simply have feel_free_to_swap_out(int pid) syscall? :-).

Quoted 7 lines

I wonder for how long we'll go on with adding new syscalls each time we need some amendment to existing interfaces. Yes, clone6(), I'm looking at you :(.

In case of process_madvise() keep in mind it will be focused not only on MADV_COLD, but also, potentially, on other MADV_ flags as well. I can hardly imagine we'll add one syscall per each flag.

Use case described above talked about whole-process-at-a-time usage, so I'm asking if simpler interface/code is enough. If there's motivation for more complex version, it should be described here...

Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
reviewer
#23 lore raw
In reply to Minchan Kim
Quoted 61 lines · Minchan Kim

This patch is part of previous series:
https://lore.kernel.org/lkml/[email protected]/T/#u
Originally, it was created for external madvise hinting feature.

https://lkml.org/lkml/2019/5/31/463
Michal wanted to separte the discussion from external hinting interface so this patchset includes only first part of my entire patchset

- introduce MADV_COLD and MADV_PAGEOUT hint to madvise.

However, I keep entire description for others for easier understanding why this kinds of hint was born.

Thanks.

This patchset is against on next-20190530.

Below is description of previous entire patchset.
================= &< =====================

- Background

The Android terminology used for forking a new process and starting an app from scratch is a cold start, while resuming an existing app is a hot start. While we continually try to improve the performance of cold starts, hot starts will always be significantly less power hungry as well as faster so we are trying to make hot start more likely than cold start.

To increase hot start, Android userspace manages the order that apps should be killed in a process called ActivityManagerService. ActivityManagerService tracks every Android app or service that the user could be interacting with at any time and translates that into a ranked list for lmkd(low memory killer daemon). They are likely to be killed by lmkd if the system has to reclaim memory. In that sense they are similar to entries in any other cache. Those apps are kept alive for opportunistic performance improvements but those performance improvements will vary based on the memory requirements of individual workloads.

- Problem

Naturally, cached apps were dominant consumers of memory on the system. However, they were not significant consumers of swap even though they are good candidate for swap. Under investigation, swapping out only begins once the low zone watermark is hit and kswapd wakes up, but the overall allocation rate in the system might trip lmkd thresholds and cause a cached process to be killed(we measured performance swapping out vs. zapping the memory by killing a process. Unsurprisingly, zapping is 10x times faster even though we use zram which is much faster than real storage) so kill from lmkd will often satisfy the high zone watermark, resulting in very few pages actually being moved to swap.

- Approach

The approach we chose was to use a new interface to allow userspace to proactively reclaim entire processes by leveraging platform information. This allowed us to bypass the inaccuracy of the kernel’s LRUs for pages that are known to be cold from userspace and to avoid races with lmkd by reclaiming apps as soon as they entered the cached state. Additionally, it could provide many chances for platform to use much information to optimize memory efficiency.

Quoted 9 lines

To achieve the goal, the patchset introduce two new options for madvise. One is MADV_COLD which will deactivate activated pages and the other is MADV_PAGEOUT which will reclaim private pages instantly. These new options complement MADV_DONTNEED and MADV_FREE by adding non-destructive ways to gain some free memory space. MADV_PAGEOUT is similar to MADV_DONTNEED in a way that it hints the kernel that memory region is not currently needed and should be reclaimed immediately; MADV_COLD is similar to MADV_FREE in a way that it hints the kernel that memory region is not currently needed and should be reclaimed when memory pressure rises.

This all is a very good background information suitable for the cover letter.

Quoted 6 lines

This approach is similar in spirit to madvise(MADV_WONTNEED), but the information required to make the reclaim decision is not known to the app. Instead, it is known to a centralized userspace daemon, and that daemon must be able to initiate reclaim on its own without any app involvement. To solve the concern, this patch introduces new syscall -

Quoted 13 lines
    struct pr_madvise_param {
            int size;               /* the size of this structure */
            int cookie;             /* reserved to support atomicity */
            int nr_elem;            /* count of below arrary fields */
            int __user *hints;      /* hints for each range */
            /* to store result of each operation */
            const struct iovec __user *results;
            /* input address ranges */
            const struct iovec __user *ranges;
    };
    
    int process_madvise(int pidfd, struct pr_madvise_param *u_param,
                            unsigned long flags);

But this and the following paragraphs are referring to the later step when the madvise gains a remote process capabilities and that is out of the scope of this patch series so I would simply remove it from here. Andrew tends to put the cover letter into the first patch of the series and that would be indeed
confusing here.

Signature · 2 lines
Michal Hocko
SUSE Labs
#24 lore raw
Quoted 95 lines · Michal Hocko
Quoted 61 lines · Minchan Kim

This patch is part of previous series:
https://lore.kernel.org/lkml/[email protected]/T/#u
Originally, it was created for external madvise hinting feature.

https://lkml.org/lkml/2019/5/31/463
Michal wanted to separte the discussion from external hinting interface so this patchset includes only first part of my entire patchset

- introduce MADV_COLD and MADV_PAGEOUT hint to madvise.

However, I keep entire description for others for easier understanding why this kinds of hint was born.

Thanks.

This patchset is against on next-20190530.

Below is description of previous entire patchset.
================= &< =====================

- Background

The Android terminology used for forking a new process and starting an app from scratch is a cold start, while resuming an existing app is a hot start. While we continually try to improve the performance of cold starts, hot starts will always be significantly less power hungry as well as faster so we are trying to make hot start more likely than cold start.

To increase hot start, Android userspace manages the order that apps should be killed in a process called ActivityManagerService. ActivityManagerService tracks every Android app or service that the user could be interacting with at any time and translates that into a ranked list for lmkd(low memory killer daemon). They are likely to be killed by lmkd if the system has to reclaim memory. In that sense they are similar to entries in any other cache. Those apps are kept alive for opportunistic performance improvements but those performance improvements will vary based on the memory requirements of individual workloads.

- Problem

Naturally, cached apps were dominant consumers of memory on the system. However, they were not significant consumers of swap even though they are good candidate for swap. Under investigation, swapping out only begins once the low zone watermark is hit and kswapd wakes up, but the overall allocation rate in the system might trip lmkd thresholds and cause a cached process to be killed(we measured performance swapping out vs. zapping the memory by killing a process. Unsurprisingly, zapping is 10x times faster even though we use zram which is much faster than real storage) so kill from lmkd will often satisfy the high zone watermark, resulting in very few pages actually being moved to swap.

- Approach

The approach we chose was to use a new interface to allow userspace to proactively reclaim entire processes by leveraging platform information. This allowed us to bypass the inaccuracy of the kernel’s LRUs for pages that are known to be cold from userspace and to avoid races with lmkd by reclaiming apps as soon as they entered the cached state. Additionally, it could provide many chances for platform to use much information to optimize memory efficiency.

Quoted 9 lines

To achieve the goal, the patchset introduce two new options for madvise. One is MADV_COLD which will deactivate activated pages and the other is MADV_PAGEOUT which will reclaim private pages instantly. These new options complement MADV_DONTNEED and MADV_FREE by adding non-destructive ways to gain some free memory space. MADV_PAGEOUT is similar to MADV_DONTNEED in a way that it hints the kernel that memory region is not currently needed and should be reclaimed immediately; MADV_COLD is similar to MADV_FREE in a way that it hints the kernel that memory region is not currently needed and should be reclaimed when memory pressure rises.

This all is a very good background information suitable for the cover letter.

Quoted 6 lines

This approach is similar in spirit to madvise(MADV_WONTNEED), but the information required to make the reclaim decision is not known to the app. Instead, it is known to a centralized userspace daemon, and that daemon must be able to initiate reclaim on its own without any app involvement. To solve the concern, this patch introduces new syscall -

Quoted 13 lines
    struct pr_madvise_param {
            int size;               /* the size of this structure */
            int cookie;             /* reserved to support atomicity */
            int nr_elem;            /* count of below arrary fields */
            int __user *hints;      /* hints for each range */
            /* to store result of each operation */
            const struct iovec __user *results;
            /* input address ranges */
            const struct iovec __user *ranges;
    };
    
    int process_madvise(int pidfd, struct pr_madvise_param *u_param,
                            unsigned long flags);
Quoted 6 lines

But this and the following paragraphs are referring to the later step when the madvise gains a remote process capabilities and that is out of the scope of this patch series so I would simply remove it from here. Andrew tends to put the cover letter into the first patch of the series and that would be indeed
confusing here.

Okay, I will remove the part in next revision.