Hi David,
On 8/24/26 6:33 PM, David Hildenbrand (Arm) wrote:
> On 8/23/26 20:43, Zenghui Yu wrote:
> > From: "Zenghui Yu (Huawei)" <[email protected]>
> >
> > The "KSM NUMA merging" test allocates two identical pages on two NUMA nodes
> > and verifies KSM will merge these two pages after 2 scans (see
> > ksm_merge_pages()). But when smart scan is enabled, pages that have
> > previously not been de-duplicated may get skipped for some scans. Verifying
> > KSM behavior after only 2 scans may not be enough.
> >
> > $ ./ksm_tests -N -d
> > TAP version 13
> > 1..1
> > pages_shared : 0
> > pages_sharing : 0
> > max_page_sharing : 256
> > full_scans : 211
> > pages_unshared : 1
> > pages_volatile : 2
> > stable_node_chains: 0
> > stable_node_dups : 0
> > general_profit : -128
> > ksm_rmap_items 2
> > ksm_zero_pages 0
> > ksm_merging_pages 0
> > ksm_process_profit -128
> > ksm_merge_any: no
> > ksm_mergeable: yes
> > not ok 1 KSM NUMA merging
> > # Totals: pass:0 fail:1 xfail:0 xpass:0 skip:0 error:0
>
> How did you reproduce this? Is this an actual test result?
I reproduced this in a guest, using arm64's virtconfig. I don't think there
is any particular configuration. This is an actual test result but this is
not 100% reproducible.
> ~/linux/tools/testing/selftests/mm$ cat /sys/kernel/mm/ksm/smart_scan
> 1
> ~/linux/tools/testing/selftests/mm$ sudo ./ksm_tests -N -d
> TAP version 13
> 1..1
> pages_shared : 1
> pages_sharing : 1
> max_page_sharing : 256
> full_scans : 1283
> pages_unshared : 0
> pages_volatile : 0
> stable_node_chains: 0
> stable_node_dups : 0
> general_profit : 3968
> ksm_rmap_items 2
> ksm_zero_pages 0
> ksm_merging_pages 2
> ksm_process_profit 8064
> ksm_merge_any: no
> ksm_mergeable: yes
> ok 1 KSM NUMA merging
> # Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
>
>
> >
> > This specific test fails because KSM started scanning the second page
> > whilst the first page had already been scanned for 16 times
>
> How is that supposed to happen? The sequence we have is:
>
> numa1_map_ptr = numa_alloc_onnode(page_size, first_node);
> numa2_map_ptr = numa_alloc_onnode(page_size, second_node);
> ...
> memset(numa1_map_ptr, '*', page_size);
> memset(numa2_map_ptr, '*', page_size);
> ...
> if (ksm_merge_pages(merge_type, numa1_map_ptr, page_size, start_time,
> timeout) ...
>
> KSM smart scan operates on rmap entries. rmap entries are per MM.
This is what I had for debugging:
diff --git a/mm/ksm.c b/mm/ksm.c
index 49d48d1e0998..aec2a292cba5 100644
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -2495,6 +2495,8 @@ static bool should_skip_rmap_item(struct folio *folio,
if (age != U8_MAX)
rmap_item->age++;
+ pr_info("addr=%lx age=%u r_skips=%u\n", rmap_item->address, age,
rmap_item->remaining_skips);
+
/*
* Smaller ages are not skipped, they need to get a chance to go
* through the different phases of the KSM merging.
@@ -2820,8 +2822,10 @@ static void ksm_do_scan(unsigned int scan_npages)
while (scan_npages-- && likely(!freezing(current))) {
cond_resched();
rmap_item = scan_get_next_rmap_item(&page);
- if (!rmap_item)
+ if (!rmap_item) {
+ pr_info("\n");
return;
+ }
cmp_and_merge_page(page, rmap_item);
put_page(page);
ksm_pages_scanned++;
$ dmesg
[ 309.111144] addr=7fff959d4000 age=0 r_skips=0
[ 309.111215] addr=7fff959d4000 age=1 r_skips=0
[ 309.111223] addr=7fff959d4101 age=2 r_skips=0
[ 309.111230] addr=7fff959d4102 age=3 r_skips=0
[ 309.111237] addr=7fff959d4103 age=4 r_skips=1
[ 309.111241] addr=7fff959d4000 age=5 r_skips=0
[ 309.111248] addr=7fff959d4105 age=6 r_skips=2
[ 309.111251] addr=7fff959d4000 age=7 r_skips=1
[ 309.111255] addr=7fff959d4000 age=8 r_skips=0
[ 309.111262] addr=7fff959d4108 age=9 r_skips=4
[ 309.111265] addr=7fff959d4000 age=10 r_skips=3
[ 309.111269] addr=7fff959d4000 age=11 r_skips=2
[ 309.111272] addr=7fff959d4000 age=12 r_skips=1
[ 309.111297] addr=7fff959ac000 age=0 r_skips=0
[ 309.111303] addr=7fff959d4000 age=13 r_skips=0 -> page 0's age is 13
greater that page 1's
[ 309.111311] addr=7fff959ac000 age=1 r_skips=0
[ 309.111316] addr=7fff959d410d age=14 r_skips=8
[ 309.111320] addr=7fff959ac10e age=2 r_skips=0
[ 309.111325] addr=7fff959d4000 age=15 r_skips=7
[ 309.111329] addr=7fff959ac10f age=3 r_skips=0
[ 309.111334] addr=7fff959d4000 age=16 r_skips=6
[ 309.111338] addr=7fff959ac110 age=4 r_skips=1
[ 309.111339] addr=7fff959d4000 age=17 r_skips=5
[ 309.111343] addr=7fff959ac000 age=5 r_skips=0
[ 309.111348] addr=7fff959d4000 age=18 r_skips=4
[ 309.111352] addr=7fff959ac112 age=6 r_skips=2
[ 309.111353] addr=7fff959d4000 age=19 r_skips=3
[ 309.111357] addr=7fff959ac000 age=7 r_skips=1
[ 309.111358] addr=7fff959d4000 age=20 r_skips=2
[ 309.111362] addr=7fff959ac000 age=8 r_skips=0
[ 309.111367] addr=7fff959d4000 age=21 r_skips=1
[ 309.111372] addr=7fff959ac115 age=9 r_skips=4
[ 309.111373] addr=7fff959d4000 age=22 r_skips=0 -> remaining_skips can
not be 0 at the same time for both pages
Not sure if I had misunderstood something.
Thanks,
Zenghui
