mm: Close races between THP migration and PMD numa clearing
Mel Gorman [Mon, 7 Oct 2013 10:28:46 +0000 (11:28 +0100)]
commit 3f926ab945b60a5824369d21add7710622a2eac0 upstream.

THP migration uses the page lock to guard against parallel allocations
but there are cases like this still open

  Task A Task B
  --------------------- ---------------------
  do_huge_pmd_numa_page do_huge_pmd_numa_page
  lock_page
  mpol_misplaced == -1
  unlock_page
  goto clear_pmdnuma
lock_page
mpol_misplaced == 2
migrate_misplaced_transhuge
  pmd = pmd_mknonnuma
  set_pmd_at

During hours of testing, one crashed with weird errors and while I have
no direct evidence, I suspect something like the race above happened.
This patch extends the page lock to being held until the pmd_numa is
cleared to prevent migration starting in parallel while the pmd_numa is
being cleared. It also flushes the old pmd entry and orders pagetable
insertion before rmap insertion.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1381141781-10992-9-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

mm/huge_memory.c
mm/migrate.c

index f276efe..c403a74 100644 (file)
@@ -1314,24 +1314,25 @@ int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
        target_nid = mpol_misplaced(page, vma, haddr);
        if (target_nid == -1) {
                /* If the page was locked, there are no parallel migrations */
-               if (page_locked) {
-                       unlock_page(page);
+               if (page_locked)
                        goto clear_pmdnuma;
-               }
 
-               /* Otherwise wait for potential migrations and retry fault */
+               /*
+                * Otherwise wait for potential migrations and retry. We do
+                * relock and check_same as the page may no longer be mapped.
+                * As the fault is being retried, do not account for it.
+                */
                spin_unlock(&mm->page_table_lock);
                wait_on_page_locked(page);
+               page_nid = -1;
                goto out;
        }
 
        /* Page is misplaced, serialise migrations and parallel THP splits */
        get_page(page);
        spin_unlock(&mm->page_table_lock);
-       if (!page_locked) {
+       if (!page_locked)
                lock_page(page);
-               page_locked = true;
-       }
        anon_vma = page_lock_anon_vma_read(page);
 
        /* Confirm the PTE did not while locked */
@@ -1339,32 +1340,28 @@ int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
        if (unlikely(!pmd_same(pmd, *pmdp))) {
                unlock_page(page);
                put_page(page);
+               page_nid = -1;
                goto out_unlock;
        }
 
-       /* Migrate the THP to the requested node */
+       /*
+        * Migrate the THP to the requested node, returns with page unlocked
+        * and pmd_numa cleared.
+        */
        spin_unlock(&mm->page_table_lock);
        migrated = migrate_misplaced_transhuge_page(mm, vma,
                                pmdp, pmd, addr, page, target_nid);
        if (migrated)
                page_nid = target_nid;
-       else
-               goto check_same;
 
        goto out;
-
-check_same:
-       spin_lock(&mm->page_table_lock);
-       if (unlikely(!pmd_same(pmd, *pmdp))) {
-               /* Someone else took our fault */
-               page_nid = -1;
-               goto out_unlock;
-       }
 clear_pmdnuma:
+       BUG_ON(!PageLocked(page));
        pmd = pmd_mknonnuma(pmd);
        set_pmd_at(mm, haddr, pmdp, pmd);
        VM_BUG_ON(pmd_numa(*pmdp));
        update_mmu_cache_pmd(vma, addr, pmdp);
+       unlock_page(page);
 out_unlock:
        spin_unlock(&mm->page_table_lock);
 
index 25ca7ca..bf436c1 100644 (file)
@@ -1710,12 +1710,12 @@ int migrate_misplaced_transhuge_page(struct mm_struct *mm,
                unlock_page(new_page);
                put_page(new_page);             /* Free it */
 
-               unlock_page(page);
+               /* Retake the callers reference and putback on LRU */
+               get_page(page);
                putback_lru_page(page);
-
-               count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
-               isolated = 0;
-               goto out;
+               mod_zone_page_state(page_zone(page),
+                        NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
+               goto out_fail;
        }
 
        /*
@@ -1732,9 +1732,9 @@ int migrate_misplaced_transhuge_page(struct mm_struct *mm,
        entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
        entry = pmd_mkhuge(entry);
 
-       page_add_new_anon_rmap(new_page, vma, haddr);
-
+       pmdp_clear_flush(vma, haddr, pmd);
        set_pmd_at(mm, haddr, pmd, entry);
+       page_add_new_anon_rmap(new_page, vma, haddr);
        update_mmu_cache_pmd(vma, address, &entry);
        page_remove_rmap(page);
        /*
@@ -1753,7 +1753,6 @@ int migrate_misplaced_transhuge_page(struct mm_struct *mm,
        count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
        count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
 
-out:
        mod_zone_page_state(page_zone(page),
                        NR_ISOLATED_ANON + page_lru,
                        -HPAGE_PMD_NR);
@@ -1762,6 +1761,10 @@ out:
 out_fail:
        count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
 out_dropref:
+       entry = pmd_mknonnuma(entry);
+       set_pmd_at(mm, haddr, pmd, entry);
+       update_mmu_cache_pmd(vma, address, &entry);
+
        unlock_page(page);
        put_page(page);
        return 0;