FOLL_MIGRATION exists only for the purpose of break_ksm(), and actually, there is not even the need to wait for the migration to finish, we only want to know if we're dealing with a KSM page.
Using follow_page() just to identify a KSM page overcomplicates GUP code. Let's use walk_page_range_vma() instead, because we don't actually care about the page itself, we only need to know a single property -- no need to even grab a reference on the page.
In my setup (AMD Ryzen 9 3900X), running the KSM selftest to test unmerge performance on 2 GiB (taskset 0x8 ./ksm_tests -D -s 2048), this results in a performance degradation of ~4% (old: ~5010 MiB/s, new: ~4800 MiB/s). I don't think we particularly care for now.
Signed-off-by: David Hildenbrand david@redhat.com --- mm/ksm.c | 70 +++++++++++++++++++++++++++++++++++++++++++++++++------- 1 file changed, 62 insertions(+), 8 deletions(-)
diff --git a/mm/ksm.c b/mm/ksm.c index 4d7bcf7da7c3..814c1a37c323 100644 --- a/mm/ksm.c +++ b/mm/ksm.c @@ -39,6 +39,7 @@ #include <linux/freezer.h> #include <linux/oom.h> #include <linux/numa.h> +#include <linux/pagewalk.h>
#include <asm/tlbflush.h> #include "internal.h" @@ -452,6 +453,60 @@ static inline bool ksm_test_exit(struct mm_struct *mm) return atomic_read(&mm->mm_users) == 0; }
+int break_ksm_pud_entry(pud_t *pud, unsigned long addr, unsigned long next, + struct mm_walk *walk) +{ + /* We only care about page tables to walk to a single base page. */ + if (pud_leaf(*pud) || !pud_present(*pud)) + return 1; + return 0; +} + +int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next, + struct mm_walk *walk) +{ + bool *ksm_page = walk->private; + struct page *page = NULL; + pte_t *pte, ptent; + spinlock_t *ptl; + + /* We only care about page tables to walk to a single base page. */ + if (pmd_leaf(*pmd) || !pmd_present(*pmd)) + return 1; + + /* + * We only lookup a single page (a) no need to iterate; and (b) + * always return 1 to exit immediately and not iterate in the caller. + */ + pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); + ptent = *pte; + + if (pte_none(ptent)) + return 1; + if (!pte_present(ptent)) { + swp_entry_t entry = pte_to_swp_entry(ptent); + + /* + * We only care about migration of KSM pages. As KSM pages + * remain KSM pages until freed, no need to wait here for + * migration to end to identify such. + */ + if (is_migration_entry(entry)) + page = pfn_swap_entry_to_page(entry); + } else { + page = vm_normal_page(walk->vma, addr, ptent); + } + if (page && PageKsm(page)) + *ksm_page = true; + pte_unmap_unlock(pte, ptl); + return 1; +} + +static const struct mm_walk_ops break_ksm_ops = { + .pud_entry = break_ksm_pud_entry, + .pmd_entry = break_ksm_pmd_entry, +}; + /* * We use break_ksm to break COW on a ksm page by triggering unsharing, * such that the ksm page will get replaced by an exclusive anonymous page. @@ -467,20 +522,19 @@ static inline bool ksm_test_exit(struct mm_struct *mm) */ static int break_ksm(struct vm_area_struct *vma, unsigned long addr) { - struct page *page; vm_fault_t ret = 0;
+ if (WARN_ON_ONCE(!IS_ALIGNED(addr, PAGE_SIZE))) + return -EINVAL; + do { bool ksm_page = false;
cond_resched(); - page = follow_page(vma, addr, - FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE); - if (IS_ERR_OR_NULL(page)) - break; - if (PageKsm(page)) - ksm_page = true; - put_page(page); + ret = walk_page_range_vma(vma, addr, addr + PAGE_SIZE, + &break_ksm_ops, &ksm_page); + if (WARN_ON_ONCE(ret < 0)) + return ret;
if (!ksm_page) return 0;