On 8/5/25 10:33 PM, Wei Yang wrote:
On Tue, Aug 05, 2025 at 11:39:15AM +0530, Donet Tom wrote:
On 8/4/25 2:41 PM, Wei Yang wrote:
On Tue, Jul 29, 2025 at 11:03:59AM +0530, Aboorva Devarajan wrote:
From: Donet Tom donettom@linux.ibm.com
[...]
diff --git a/tools/testing/selftests/mm/ksm_functional_tests.c b/tools/testing/selftests/mm/ksm_functional_tests.c index d8bd1911dfc0..996dc6645570 100644 --- a/tools/testing/selftests/mm/ksm_functional_tests.c +++ b/tools/testing/selftests/mm/ksm_functional_tests.c @@ -46,6 +46,8 @@ static int ksm_use_zero_pages_fd; static int pagemap_fd; static size_t pagesize;
+static void init_global_file_handles(void);
static bool range_maps_duplicates(char *addr, unsigned long size) { unsigned long offs_a, offs_b, pfn_a, pfn_b; @@ -274,6 +276,7 @@ static void test_unmerge(void) ksft_test_result(!range_maps_duplicates(map, size), "Pages were unmerged\n"); unmap:
- ksm_unmerge();
In __mmap_and_merge_range(), we call ksm_unmerge(). Why this one not help?
Not very familiar with ksm stuff. Would you mind giving more on how this fix the failure you see?
The issue I was facing here was test_prctl_fork was failing.
# [RUN] test_prctl_fork # Still pages merged #
This issue occurred because the previous test performed a merge, causing the value of /proc/self/ksm_merging_pages to reflect the number of deduplicated pages. After that, a fork() was called. Post-fork, the child process inherited the parent's ksm_merging_pages value.
Yes, this one is fixed by calling init_global_file_handles() in child.
Then, the child process invoked __mmap_and_merge_range(), which resulted in unmerging the pages and resetting the value. However, since the parent process had performed the merge, its ksm_merging_pages value also got reset to 0. Meanwhile, the child process had not performed any merge itself, so the inherited
I assume the behavior described here is after the change to call init_global_file_handles() in child.
Yes
Child process inherit the ksm_merging_pages from parent, which is reasonable to me. But I am confused why ksm_unmerge() would just reset ksm_merging_pages for parent and leave ksm_merging_pages in child process unchanged.
ksm_unmerge() writes to /sys/kernel/mm/ksm/run, which is a system wide sysfs interface. I expect it applies to both parent and child.
I am not very familiar with the KSM code, but from what I understand:
The ksm_merging_pages counter is maintained per mm_struct. When we write to /sys/kernel/mm/ksm/run, unmerging is triggered, and the counters are updated for all mm_structs present in the ksm_mm_slot list.
A mm_struct gets added to this list when MADV_MERGEABLE is called. In the case of the child process, since MADV_MERGEABLE has not been invoked yet, its mm_struct is not part of the list. As a result, its ksm_merging_pages counter is not reset.
value remained unchanged. That’s why get_my_merging_page() in the child was returning a non-zero value.
I guess you mean the get_my_merging_page() in __mmap_and_merge_range() return a non-zero value. But there is ksm_unmerge() before it. Why this ksm_unmerge() couldn't reset the value, but a ksm_unmerge() in parent could.
Initially, I fixed the issue by calling ksm_unmerge() before the fork(), and that resolved the problem. Later, I decided it would be cleaner to move the ksm_unmerge() call to the test cleanup phase.
Also all the tests before test_prctl_fork(), except test_prctl(), calls
ksft_test_result(!range_maps_duplicates());
If the previous tests succeed, it means there is no duplicate pages. This means ksm_merging_pages should be 0 before test_prctl_fork() if other tests pass. And the child process would inherit a 0 ksm_merging_pages. (A quick test proves it.)
If I understand correctly, all the tests are calling MADV_UNMERGEABLE, which internally calls break_ksm() in the kernel. This function replaces the KSM page with an exclusive anonymous page. However, the ksm_merging_pages counters are not updated at this point.
The function range_maps_duplicates(map, size) checks whether the pages have been unmerged. Since break_ksm() does perform the unmerge, this function returns false, and the test passes.
The ksm_merging_pages update happens later via the ksm_scan_thread(). That’s why we observe that ksm_merging_pages values are not reset immediately after the test finishes.
If we add a sleep(1) after the MADV_UNMERGEABLE call, we can see that the ksm_merging_pages values are reset after the sleep.
Once the test completes successfully, we can call ksm_unmerge(), which will immediately reset the ksm_merging_pages value. This way, in the fork test, the child process will also see the correct value.
So which part of the story I missed?
So, during the cleanup phase after a successful test, we can call ksm_unmerge() to reset the counter. Do you see any issue with this approach?