ksm: use range-walk function to jump over holes in scan_get_next_rmap_item
PATCH v5 1/1 ksm: use range-walk function to jump over holes in scan_get_next_rmap_item
Currently, scan_get_next_rmap_item() walks every page address in a VMA to locate mergeable pages. This becomes highly inefficient when scanning large virtual memory areas that contain mostly unmapped regions, causing ksmd to use large amount of cpu without deduplicating much pages.
This patch replaces the per-address lookup with a range walk using walk_page_range(). The range walker allows KSM to skip over entire unmapped holes in a VMA, avoiding unnecessary lookups. This problem was previously discussed in [1].
Consider the following test program which creates a 32 TiB mapping in the virtual address space but only populates a single page:
#include <unistd.h>
#include <stdio.h>
#include <sys/mman.h>
/* 32 TiB */
const size_t size = 32ul * 1024 * 1024 * 1024 * 1024;
int main() {
char *area = mmap(NULL, size, PROT_READ | PROT_WRITE,
MAP_NORESERVE | MAP_PRIVATE | MAP_ANON, -1, 0);
if (area == MAP_FAILED) {
perror("mmap() failed\n");
return -1;
}
/* Populate a single page such that we get an anon_vma. */
*area = 0;
/* Enable KSM. */
madvise(area, size, MADV_MERGEABLE);
pause();
return 0;
}$ ./ksm-sparse & $ echo 1 > /sys/kernel/mm/ksm/run
Without this patch ksmd uses 100% of the cpu for a long time (more then 1 hour in my test machine) scanning all the 32 TiB virtual address space that contain only one mapped page. This makes ksmd essentially deadlocked not able to deduplicate anything of value.
With this patch ksmd walks only the one mapped page and skips the rest of the 32 TiB virtual address space, making the scan fast using little cpu.
[1] https://lore.kernel.org/linux-mm/[email protected]/
---
v5:
- Improve patch description
v4: https://lore.kernel.org/linux-mm/[email protected]/
- Make minimal changes to replace folio_walk by walk_page_range_vma
v3: https://lore.kernel.org/all/[email protected]/
- Treat THPs in ksm_pmd_entry
- Update ksm_scan.address outside walk_page_range
- Change goto to while loop
v2: https://lore.kernel.org/all/[email protected]/
- Use pmd_entry to walk page range
- Use cond_resched inside pmd_entry()
- walk_page_range returns page+folio
v1: https://lore.kernel.org/all/[email protected]/
--- mm/ksm.c | 113 ++++++++++++++++++++++++++++++++++++++++++++++++++----- 1 file changed, 104 insertions(+), 9 deletions(-)
mm/ksm.c +104 -9
@@ -2455,6 +2455,95 @@ static bool should_skip_rmap_item(struct folio *folio,return true; } +struct ksm_next_page_arg { +struct folio *folio; +struct page *page; +unsigned long addr; +}; + +static int ksm_next_page_pmd_entry (pmd_t *pmdp,unsigned long addr,unsigned long end, +struct mm_walk *walk) +{ +struct ksm_next_page_arg *private = walk->private; +struct vm_area_struct *vma = walk->vma; +pte_t *start_ptep =NULL , *ptep, pte; +struct mm_struct *mm = walk->mm; +struct folio *folio; +struct page *page; +spinlock_t *ptl; +pmd_t pmd; + +if (ksm_test_exit (mm)) +return 0 ; + +cond_resched (); + + pmd =pmdp_get_lockless (pmdp); +if (!pmd_present (pmd)) +return 0 ; + +if (IS_ENABLED (CONFIG_TRANSPARENT_HUGEPAGE ) &&pmd_leaf (pmd)) { + ptl =pmd_lock (mm, pmdp); + pmd =pmdp_get (pmdp); + +if (!pmd_present (pmd)) { + goto not_found_unlock; + }else if (pmd_leaf (pmd)) { + page =vm_normal_page_pmd (vma, addr, pmd); +if (!page) + goto not_found_unlock; + folio =page_folio (page); + +if (folio_is_zone_device (folio) || !folio_test_anon (folio)) + goto not_found_unlock; + + page += ((addr & (PMD_SIZE -1 )) >>PAGE_SHIFT ); + goto found_unlock; + } +spin_unlock (ptl); + } + + start_ptep =pte_offset_map_lock (mm, pmdp, addr, &ptl); +if (!start_ptep) +return 0 ; + +for (ptep = start_ptep; addr < end; ptep++, addr +=PAGE_SIZE ) { + pte =ptep_get (ptep); + +if (!pte_present (pte)) +continue ; + + page =vm_normal_page (vma, addr, pte); +if (!page) +continue ; + folio =page_folio (page); + +if (folio_is_zone_device (folio) || !folio_test_anon (folio)) +continue ; + goto found_unlock; + } + +not_found_unlock: +spin_unlock (ptl); +if (start_ptep) +pte_unmap (start_ptep); +return 0 ; +found_unlock: +folio_get (folio); +spin_unlock (ptl); +if (start_ptep) +pte_unmap (start_ptep); + private->page = page; + private->folio = folio; + private->addr = addr; +return 1 ; +} + +static struct mm_walk_ops ksm_next_page_ops = { + .pmd_entry = ksm_next_page_pmd_entry, + .walk_lock =PGWALK_RDLOCK , +}; +static struct ksm_rmap_item *scan_get_next_rmap_item (struct page **page) {struct mm_struct *mm;
@@ -2542,21 +2631,27 @@ static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) ksm_scan.address = vma->vm_end;while (ksm_scan.address < vma->vm_end) { +struct ksm_next_page_arg ksm_next_page_arg;struct page *tmp_page =NULL ; -struct folio_walk fw;struct folio *folio;if (ksm_test_exit (mm))break ; - folio =folio_walk_start (&fw, vma, ksm_scan.address,0 ); -if (folio) { -if (!folio_is_zone_device (folio) && -folio_test_anon (folio)) { -folio_get (folio); - tmp_page = fw.page; - } -folio_walk_end (&fw, vma); +int found; + + found =walk_page_range_vma (vma, ksm_scan.address, + vma->vm_end, + &ksm_next_page_ops, + &ksm_next_page_arg); + +if (found >0 ) { + folio = ksm_next_page_arg.folio; + tmp_page = ksm_next_page_arg.page; + ksm_scan.address = ksm_next_page_arg.addr; + }else { +VM_WARN_ON_ONCE (found <0 ); + ksm_scan.address = vma->vm_end -PAGE_SIZE ; }if (tmp_page) {
Signature · 1 line
2.43.0
Quoted 207 lines · Pedro Demarchi Gomes
Currently,
scan_get_next_rmap_item()walks every page address in a VMA to locate mergeable pages. This becomes highly inefficient when scanning large virtual memory areas that contain mostly unmapped regions, causing ksmd to use large amount of cpu without deduplicating much pages.This patch replaces the per-address lookup with a range walk using
walk_page_range(). The range walker allows KSM to skip over entire unmapped holes in a VMA, avoiding unnecessary lookups. This problem was previously discussed in [1].Consider the following test program which creates a 32 TiB mapping in the virtual address space but only populates a single page:
#include <
unistd.h>
#include <stdio.h>
#include <sys/mman.h>/* 32 TiB */
constsize_tsize = 32ul * 1024 * 1024 * 1024 * 1024;int main() { char *area = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_NORESERVE | MAP_PRIVATE | MAP_ANON, -1, 0); if (area == MAP_FAILED) { perror("mmap() failed\n"); return -1; } /* Populate a single page such that we get an anon_vma. */ *area = 0; /* Enable KSM. */ madvise(area, size, MADV_MERGEABLE); pause(); return 0; }$ ./ksm-sparse & $ echo 1 > /sys/kernel/mm/ksm/runWithout this patch ksmd uses 100% of the cpu for a long time (more then 1 hour in my test machine) scanning all the 32 TiB virtual address space that contain only one mapped page. This makes ksmd essentially deadlocked not able to deduplicate anything of value.
With this patch ksmd walks only the one mapped page and skips the rest of the 32 TiB virtual address space, making the scan fast using little cpu.[1] https://lore.kernel.org/linux-mm/[email protected]/
---
v5:
- Improve patch descriptionv4: https://lore.kernel.org/linux-mm/[email protected]/
- Make minimal changes to replacefolio_walkbywalk_page_range_vmav3: https://lore.kernel.org/all/[email protected]/
- Treat THPs inksm_pmd_entry
- Update ksm_scan.address outsidewalk_page_range
- Change goto to while loopv2: https://lore.kernel.org/all/[email protected]/
- Usepmd_entryto walk page range
- Usecond_reschedinsidepmd_entry()
-walk_page_rangereturns page+foliov1: https://lore.kernel.org/all/[email protected]/
Reported-by craftfever <[email protected]>Suggested-by David Hildenbrand <[email protected]>Co-developed-by David Hildenbrand <[email protected]>Signed-off-by David Hildenbrand <[email protected]>Fixes 31dbd01f3143 ("ksm: Kernel SamePage Merging")Signed-off-by Pedro Demarchi Gomes <[email protected]>--- mm/ksm.c | 113 ++++++++++++++++++++++++++++++++++++++++++++++++++----- 1 file changed, 104 insertions(+), 9 deletions(-)mm/ksm.c +97
-9@@ -2455,6 +2455,95 @@ static bool should_skip_rmap_item(struct folio *folio,return true; } +struct ksm_next_page_arg { +struct folio *folio; +struct page *page; +unsigned long addr; +}; + +static int ksm_next_page_pmd_entry (pmd_t *pmdp,unsigned long addr,unsigned long end, +struct mm_walk *walk) +{ +struct ksm_next_page_arg *private = walk->private; +struct vm_area_struct *vma = walk->vma; +pte_t *start_ptep =NULL , *ptep, pte; +struct mm_struct *mm = walk->mm; +struct folio *folio; +struct page *page; +spinlock_t *ptl; +pmd_t pmd; + +if (ksm_test_exit (mm)) +return 0 ; + +cond_resched (); + + pmd =pmdp_get_lockless (pmdp); +if (!pmd_present (pmd)) +return 0 ; + +if (IS_ENABLED (CONFIG_TRANSPARENT_HUGEPAGE ) &&pmd_leaf (pmd)) { + ptl =pmd_lock (mm, pmdp); + pmd =pmdp_get (pmdp); + +if (!pmd_present (pmd)) { + goto not_found_unlock; + }else if (pmd_leaf (pmd)) { + page =vm_normal_page_pmd (vma, addr, pmd); +if (!page) + goto not_found_unlock; + folio =page_folio (page); + +if (folio_is_zone_device (folio) || !folio_test_anon (folio)) + goto not_found_unlock; + + page += ((addr & (PMD_SIZE -1 )) >>PAGE_SHIFT ); + goto found_unlock; + } +spin_unlock (ptl); + } + + start_ptep =pte_offset_map_lock (mm, pmdp, addr, &ptl); +if (!start_ptep) +return 0 ; + +for (ptep = start_ptep; addr < end; ptep++, addr +=PAGE_SIZE ) { + pte =ptep_get (ptep); + +if (!pte_present (pte)) +continue ; + + page =vm_normal_page (vma, addr, pte); +if (!page) +continue ; + folio =page_folio (page); + +if (folio_is_zone_device (folio) || !folio_test_anon (folio)) +continue ; + goto found_unlock; + } + +not_found_unlock: +spin_unlock (ptl); +if (start_ptep) +pte_unmap (start_ptep); +return 0 ; +found_unlock: +folio_get (folio); +spin_unlock (ptl); +if (start_ptep) +pte_unmap (start_ptep); + private->page = page; + private->folio = folio; + private->addr = addr; +return 1 ; +} + +static struct mm_walk_ops ksm_next_page_ops = { + .pmd_entry = ksm_next_page_pmd_entry, + .walk_lock =PGWALK_RDLOCK , +}; +static struct ksm_rmap_item *scan_get_next_rmap_item (struct page **page) {struct mm_struct *mm;@@ -2542,21 +2631,27 @@ static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) ksm_scan.address = vma->vm_end;while (ksm_scan.address < vma->vm_end) { +struct ksm_next_page_arg ksm_next_page_arg;struct page *tmp_page =NULL ; -struct folio_walk fw;struct folio *folio;if (ksm_test_exit (mm))break ; - folio =folio_walk_start (&fw, vma, ksm_scan.address,0 ); -if (folio) { -if (!folio_is_zone_device (folio) && -folio_test_anon (folio)) { -folio_get (folio); - tmp_page = fw.page; - } -folio_walk_end (&fw, vma); +int found; + + found =walk_page_range_vma (vma, ksm_scan.address, + vma->vm_end, + &ksm_next_page_ops, + &ksm_next_page_arg); +
mm/ksm.c static struct ksm_rmap_item *scan_get_next_rmap_item(struct page **page) · line 2657
Quoted 10 lines
mm/ksm.c +7
-0+if (found >0 ) { + folio = ksm_next_page_arg.folio; + tmp_page = ksm_next_page_arg.page; + ksm_scan.address = ksm_next_page_arg.addr; + }else { +VM_WARN_ON_ONCE (found <0 ); + ksm_scan.address = vma->vm_end -PAGE_SIZE ; }if (tmp_page) {
Thank you, it works magnificently, very pleasent. Very low CPU-consuming and effective, no crashes. Is this a final version?
Quoted 26 lines
Quoted 26 lines
break;mm/ksm.c +14
-8- folio =folio_walk_start (&fw, vma, ksm_scan.address,0 ); -if (folio) { -if (!folio_is_zone_device (folio) && -folio_test_anon (folio)) { -folio_get (folio); - tmp_page = fw.page; - } -folio_walk_end (&fw, vma); +int found; + + found =walk_page_range_vma (vma, ksm_scan.address, + vma->vm_end, + &ksm_next_page_ops, + &ksm_next_page_arg); + +if (found >0 ) { + folio = ksm_next_page_arg.folio; + tmp_page = ksm_next_page_arg.page; + ksm_scan.address = ksm_next_page_arg.addr; + }else { +VM_WARN_ON_ONCE (found <0 ); + ksm_scan.address = vma->vm_end -PAGE_SIZE ; }
Quoted 4 lines
Quoted 1 line
if (tmp_page) {Thank you, it works magnificently, very pleasent. Very low CPU-consuming and effective, no crashes. Is this a final version?
If there is no further feedback, yes.
Signature · 3 lines
Cheers David / dhildenb
Quoted 69 lines · Pedro Demarchi Gomes
Currently,
scan_get_next_rmap_item()walks every page address in a VMA to locate mergeable pages. This becomes highly inefficient when scanning large virtual memory areas that contain mostly unmapped regions, causing ksmd to use large amount of cpu without deduplicating much pages.This patch replaces the per-address lookup with a range walk using
walk_page_range(). The range walker allows KSM to skip over entire unmapped holes in a VMA, avoiding unnecessary lookups. This problem was previously discussed in [1].Consider the following test program which creates a 32 TiB mapping in the virtual address space but only populates a single page:
#include <
unistd.h>
#include <stdio.h>
#include <sys/mman.h>/* 32 TiB */
constsize_tsize = 32ul * 1024 * 1024 * 1024 * 1024;int main() { char *area = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_NORESERVE | MAP_PRIVATE | MAP_ANON, -1, 0); if (area == MAP_FAILED) { perror("mmap() failed\n"); return -1; } /* Populate a single page such that we get an anon_vma. */ *area = 0; /* Enable KSM. */ madvise(area, size, MADV_MERGEABLE); pause(); return 0; }$ ./ksm-sparse & $ echo 1 > /sys/kernel/mm/ksm/runWithout this patch ksmd uses 100% of the cpu for a long time (more then 1 hour in my test machine) scanning all the 32 TiB virtual address space that contain only one mapped page. This makes ksmd essentially deadlocked not able to deduplicate anything of value.
With this patch ksmd walks only the one mapped page and skips the rest of the 32 TiB virtual address space, making the scan fast using little cpu.[1] https://lore.kernel.org/linux-mm/[email protected]/
---
v5:
- Improve patch descriptionv4: https://lore.kernel.org/linux-mm/[email protected]/
- Make minimal changes to replacefolio_walkbywalk_page_range_vmav3: https://lore.kernel.org/all/[email protected]/
- Treat THPs inksm_pmd_entry
- Update ksm_scan.address outsidewalk_page_range
- Change goto to while loopv2: https://lore.kernel.org/all/[email protected]/
- Usepmd_entryto walk page range
- Usecond_reschedinsidepmd_entry()
-walk_page_rangereturns page+folio
Quoted 7 lines
Reported-by craftfever <[email protected]>Suggested-by David Hildenbrand <[email protected]>Co-developed-by David Hildenbrand <[email protected]>Signed-off-by David Hildenbrand <[email protected]>Fixes 31dbd01f3143 ("ksm: Kernel SamePage Merging")Signed-off-by Pedro Demarchi Gomes <[email protected]>
I think we really want to
Andrew can do that when applying.
As a note, we have similar code that should probably be doing a range walk instead: unmerge_ksm_pages()->break_ksm().
It can be triggered on a range through unmerge_ksm_pages(), which gets called from:
* ksm_madvise() through madvise(MADV_UNMERGEABLE). There are not a lot of users of that function.
* __ksm_del_vma() through ksm_del_vmas(). Effectively called when disabling KSM for a process either through the sysctl or from s390x gmap code when enabling storage keys for a VM.
In both cases, it's not ksmd that's blocked, it's just that the operation (trigger by the app) takes longer.
So both are not as critical as this thing here, but likely we should take care of it at some point.
Interestingly, I converted that from a walk_page_range_vma() to folio_walk_start() after converting it from follow_page() to walk_page_range_vma().
But we never did a range walk, we just walked individual addresses, because that's what break_ksm() does.
We could effectively revert e317a8d8b4f600fc7ec9725e26417030ee594f52 and adjust it to perform an actual range walk by passing a range to break_ksm().
Signature · 3 lines
Cheers David / dhildenb
Quoted 10 lines · David Hildenbrand
As a note, we have similar code that should probably be doing a range walk instead: unmerge_ksm_pages()->break_ksm().
It can be triggered on a range through
unmerge_ksm_pages(), which gets called from:*
ksm_madvise()through madvise(MADV_UNMERGEABLE). There are not a lot of users of that function.*
__ksm_del_vma()throughksm_del_vmas(). Effectively called when disabling KSM for a process either through the sysctl or from s390x gmap code when enabling storage keys for a VM.In both cases, it's not ksmd that's blocked, it's just that the operation (trigger by the app) takes longer.
Quoted 7 lines
So both are not as critical as this thing here, but likely we should take care of it at some point.
Interestingly, I converted that from a
walk_page_range_vma()tofolio_walk_start()after converting it fromfollow_page()towalk_page_range_vma().But we never did a range walk, we just walked individual addresses, because that's what
break_ksm()does.We could effectively revert
e317a8d8b4f600fc7ec9725e26417030ee594f52and adjust it to perform an actual range walk by passing a range tobreak_ksm().
Thanks for letting me know. I will send a patch fixing this issue.