mm/migrate: migrate_vma() unmap page from vma while collecting pages
authorJérôme Glisse <jglisse@redhat.com>
Fri, 8 Sep 2017 23:12:13 +0000 (16:12 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 9 Sep 2017 01:26:46 +0000 (18:26 -0700)
Common case for migration of virtual address range is page are map only
once inside the vma in which migration is taking place.  Because we
already walk the CPU page table for that range we can directly do the
unmap there and setup special migration swap entry.

Link: http://lkml.kernel.org/r/20170817000548.32038-16-jglisse@redhat.com
Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Signed-off-by: Evgeny Baskakov <ebaskakov@nvidia.com>
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Signed-off-by: Mark Hairgrove <mhairgrove@nvidia.com>
Signed-off-by: Sherry Cheung <SCheung@nvidia.com>
Signed-off-by: Subhash Gutti <sgutti@nvidia.com>
Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com>
Cc: Balbir Singh <bsingharora@gmail.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: David Nellans <dnellans@nvidia.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Bob Liu <liubo95@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/migrate.c

index 991e8886093f0c64f162ece6e10972e25ec4218b..652b2c642eed18d765b10c4a46f698133d979764 100644 (file)
@@ -2149,7 +2149,7 @@ static int migrate_vma_collect_pmd(pmd_t *pmdp,
        struct migrate_vma *migrate = walk->private;
        struct vm_area_struct *vma = walk->vma;
        struct mm_struct *mm = vma->vm_mm;
-       unsigned long addr = start;
+       unsigned long addr = start, unmapped = 0;
        spinlock_t *ptl;
        pte_t *ptep;
 
@@ -2194,9 +2194,12 @@ again:
                return migrate_vma_collect_hole(start, end, walk);
 
        ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
+       arch_enter_lazy_mmu_mode();
+
        for (; addr < end; addr += PAGE_SIZE, ptep++) {
                unsigned long mpfn, pfn;
                struct page *page;
+               swp_entry_t entry;
                pte_t pte;
 
                pte = *ptep;
@@ -2228,11 +2231,44 @@ again:
                mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
                mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
 
+               /*
+                * Optimize for the common case where page is only mapped once
+                * in one process. If we can lock the page, then we can safely
+                * set up a special migration page table entry now.
+                */
+               if (trylock_page(page)) {
+                       pte_t swp_pte;
+
+                       mpfn |= MIGRATE_PFN_LOCKED;
+                       ptep_get_and_clear(mm, addr, ptep);
+
+                       /* Setup special migration page table entry */
+                       entry = make_migration_entry(page, pte_write(pte));
+                       swp_pte = swp_entry_to_pte(entry);
+                       if (pte_soft_dirty(pte))
+                               swp_pte = pte_swp_mksoft_dirty(swp_pte);
+                       set_pte_at(mm, addr, ptep, swp_pte);
+
+                       /*
+                        * This is like regular unmap: we remove the rmap and
+                        * drop page refcount. Page won't be freed, as we took
+                        * a reference just above.
+                        */
+                       page_remove_rmap(page, false);
+                       put_page(page);
+                       unmapped++;
+               }
+
 next:
                migrate->src[migrate->npages++] = mpfn;
        }
+       arch_leave_lazy_mmu_mode();
        pte_unmap_unlock(ptep - 1, ptl);
 
+       /* Only flush the TLB if we actually modified any entries */
+       if (unmapped)
+               flush_tlb_range(walk->vma, start, end);
+
        return 0;
 }
 
@@ -2257,7 +2293,13 @@ static void migrate_vma_collect(struct migrate_vma *migrate)
        mm_walk.mm = migrate->vma->vm_mm;
        mm_walk.private = migrate;
 
+       mmu_notifier_invalidate_range_start(mm_walk.mm,
+                                           migrate->start,
+                                           migrate->end);
        walk_page_range(migrate->start, migrate->end, &mm_walk);
+       mmu_notifier_invalidate_range_end(mm_walk.mm,
+                                         migrate->start,
+                                         migrate->end);
 
        migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
 }
@@ -2305,32 +2347,37 @@ static bool migrate_vma_check_page(struct page *page)
 static void migrate_vma_prepare(struct migrate_vma *migrate)
 {
        const unsigned long npages = migrate->npages;
+       const unsigned long start = migrate->start;
+       unsigned long addr, i, restore = 0;
        bool allow_drain = true;
-       unsigned long i;
 
        lru_add_drain();
 
        for (i = 0; (i < npages) && migrate->cpages; i++) {
                struct page *page = migrate_pfn_to_page(migrate->src[i]);
+               bool remap = true;
 
                if (!page)
                        continue;
 
-               /*
-                * Because we are migrating several pages there can be
-                * a deadlock between 2 concurrent migration where each
-                * are waiting on each other page lock.
-                *
-                * Make migrate_vma() a best effort thing and backoff
-                * for any page we can not lock right away.
-                */
-               if (!trylock_page(page)) {
-                       migrate->src[i] = 0;
-                       migrate->cpages--;
-                       put_page(page);
-                       continue;
+               if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
+                       /*
+                        * Because we are migrating several pages there can be
+                        * a deadlock between 2 concurrent migration where each
+                        * are waiting on each other page lock.
+                        *
+                        * Make migrate_vma() a best effort thing and backoff
+                        * for any page we can not lock right away.
+                        */
+                       if (!trylock_page(page)) {
+                               migrate->src[i] = 0;
+                               migrate->cpages--;
+                               put_page(page);
+                               continue;
+                       }
+                       remap = false;
+                       migrate->src[i] |= MIGRATE_PFN_LOCKED;
                }
-               migrate->src[i] |= MIGRATE_PFN_LOCKED;
 
                if (!PageLRU(page) && allow_drain) {
                        /* Drain CPU's pagevec */
@@ -2339,21 +2386,50 @@ static void migrate_vma_prepare(struct migrate_vma *migrate)
                }
 
                if (isolate_lru_page(page)) {
-                       migrate->src[i] = 0;
-                       unlock_page(page);
-                       migrate->cpages--;
-                       put_page(page);
+                       if (remap) {
+                               migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
+                               migrate->cpages--;
+                               restore++;
+                       } else {
+                               migrate->src[i] = 0;
+                               unlock_page(page);
+                               migrate->cpages--;
+                               put_page(page);
+                       }
                        continue;
                }
 
                if (!migrate_vma_check_page(page)) {
-                       migrate->src[i] = 0;
-                       unlock_page(page);
-                       migrate->cpages--;
+                       if (remap) {
+                               migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
+                               migrate->cpages--;
+                               restore++;
 
-                       putback_lru_page(page);
+                               get_page(page);
+                               putback_lru_page(page);
+                       } else {
+                               migrate->src[i] = 0;
+                               unlock_page(page);
+                               migrate->cpages--;
+
+                               putback_lru_page(page);
+                       }
                }
        }
+
+       for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
+               struct page *page = migrate_pfn_to_page(migrate->src[i]);
+
+               if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
+                       continue;
+
+               remove_migration_pte(page, migrate->vma, addr, page);
+
+               migrate->src[i] = 0;
+               unlock_page(page);
+               put_page(page);
+               restore--;
+       }
 }
 
 /*
@@ -2380,12 +2456,19 @@ static void migrate_vma_unmap(struct migrate_vma *migrate)
                if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
                        continue;
 
-               try_to_unmap(page, flags);
-               if (page_mapped(page) || !migrate_vma_check_page(page)) {
-                       migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
-                       migrate->cpages--;
-                       restore++;
+               if (page_mapped(page)) {
+                       try_to_unmap(page, flags);
+                       if (page_mapped(page))
+                               goto restore;
                }
+
+               if (migrate_vma_check_page(page))
+                       continue;
+
+restore:
+               migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
+               migrate->cpages--;
+               restore++;
        }
 
        for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {