ksm: fix endless loop on oom
authorHugh Dickins <hugh.dickins@tiscali.co.uk>
Tue, 22 Sep 2009 00:02:16 +0000 (17:02 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 22 Sep 2009 14:17:32 +0000 (07:17 -0700)
break_ksm has been looping endlessly ignoring VM_FAULT_OOM: that should
only be a problem for ksmd when a memory control group imposes limits
(normally the OOM killer will kill others with an mm until it succeeds);
but in general (especially for MADV_UNMERGEABLE and KSM_RUN_UNMERGE) we
do need to route the error (or kill) back to the caller (or sighandling).

Test signal_pending in unmerge_ksm_pages, which could be a lengthy
procedure if it has to spill into swap: returning -ERESTARTSYS so that
trivial signals will restart but fatals will terminate (is that right?
we do different things in different places in mm, none exactly this).

unmerge_and_remove_all_rmap_items was forgetting to lock when going
down the mm_list: fix that.  Whether it's successful or not, reset
ksm_scan cursor to head; but only if it's successful, reset seqnr
(shown in full_scans) - page counts will have gone down to zero.

This patch leaves a significant OOM deadlock, but it's a good step
on the way, and that deadlock is fixed in a subsequent patch.

Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Acked-by: Izik Eidus <ieidus@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/ksm.c

index c49bb71..d9e3cfc 100644 (file)
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -294,10 +294,10 @@ static inline int in_stable_tree(struct rmap_item *rmap_item)
  * Could a ksm page appear anywhere else?  Actually yes, in a VM_PFNMAP
  * mmap of /dev/mem or /dev/kmem, where we would not want to touch it.
  */
-static void break_ksm(struct vm_area_struct *vma, unsigned long addr)
+static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
 {
        struct page *page;
-       int ret;
+       int ret = 0;
 
        do {
                cond_resched();
@@ -310,9 +310,36 @@ static void break_ksm(struct vm_area_struct *vma, unsigned long addr)
                else
                        ret = VM_FAULT_WRITE;
                put_page(page);
-       } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS)));
-
-       /* Which leaves us looping there if VM_FAULT_OOM: hmmm... */
+       } while (!(ret & (VM_FAULT_WRITE | VM_FAULT_SIGBUS | VM_FAULT_OOM)));
+       /*
+        * We must loop because handle_mm_fault() may back out if there's
+        * any difficulty e.g. if pte accessed bit gets updated concurrently.
+        *
+        * VM_FAULT_WRITE is what we have been hoping for: it indicates that
+        * COW has been broken, even if the vma does not permit VM_WRITE;
+        * but note that a concurrent fault might break PageKsm for us.
+        *
+        * VM_FAULT_SIGBUS could occur if we race with truncation of the
+        * backing file, which also invalidates anonymous pages: that's
+        * okay, that truncation will have unmapped the PageKsm for us.
+        *
+        * VM_FAULT_OOM: at the time of writing (late July 2009), setting
+        * aside mem_cgroup limits, VM_FAULT_OOM would only be set if the
+        * current task has TIF_MEMDIE set, and will be OOM killed on return
+        * to user; and ksmd, having no mm, would never be chosen for that.
+        *
+        * But if the mm is in a limited mem_cgroup, then the fault may fail
+        * with VM_FAULT_OOM even if the current task is not TIF_MEMDIE; and
+        * even ksmd can fail in this way - though it's usually breaking ksm
+        * just to undo a merge it made a moment before, so unlikely to oom.
+        *
+        * That's a pity: we might therefore have more kernel pages allocated
+        * than we're counting as nodes in the stable tree; but ksm_do_scan
+        * will retry to break_cow on each pass, so should recover the page
+        * in due course.  The important thing is to not let VM_MERGEABLE
+        * be cleared while any such pages might remain in the area.
+        */
+       return (ret & VM_FAULT_OOM) ? -ENOMEM : 0;
 }
 
 static void break_cow(struct mm_struct *mm, unsigned long addr)
@@ -462,39 +489,61 @@ static void remove_trailing_rmap_items(struct mm_slot *mm_slot,
  * to the next pass of ksmd - consider, for example, how ksmd might be
  * in cmp_and_merge_page on one of the rmap_items we would be removing.
  */
-static void unmerge_ksm_pages(struct vm_area_struct *vma,
-                             unsigned long start, unsigned long end)
+static int unmerge_ksm_pages(struct vm_area_struct *vma,
+                            unsigned long start, unsigned long end)
 {
        unsigned long addr;
+       int err = 0;
 
-       for (addr = start; addr < end; addr += PAGE_SIZE)
-               break_ksm(vma, addr);
+       for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
+               if (signal_pending(current))
+                       err = -ERESTARTSYS;
+               else
+                       err = break_ksm(vma, addr);
+       }
+       return err;
 }
 
-static void unmerge_and_remove_all_rmap_items(void)
+static int unmerge_and_remove_all_rmap_items(void)
 {
        struct mm_slot *mm_slot;
        struct mm_struct *mm;
        struct vm_area_struct *vma;
+       int err = 0;
+
+       spin_lock(&ksm_mmlist_lock);
+       mm_slot = list_entry(ksm_mm_head.mm_list.next,
+                                               struct mm_slot, mm_list);
+       spin_unlock(&ksm_mmlist_lock);
 
-       list_for_each_entry(mm_slot, &ksm_mm_head.mm_list, mm_list) {
+       while (mm_slot != &ksm_mm_head) {
                mm = mm_slot->mm;
                down_read(&mm->mmap_sem);
                for (vma = mm->mmap; vma; vma = vma->vm_next) {
                        if (!(vma->vm_flags & VM_MERGEABLE) || !vma->anon_vma)
                                continue;
-                       unmerge_ksm_pages(vma, vma->vm_start, vma->vm_end);
+                       err = unmerge_ksm_pages(vma,
+                                               vma->vm_start, vma->vm_end);
+                       if (err) {
+                               up_read(&mm->mmap_sem);
+                               goto out;
+                       }
                }
                remove_trailing_rmap_items(mm_slot, mm_slot->rmap_list.next);
                up_read(&mm->mmap_sem);
+
+               spin_lock(&ksm_mmlist_lock);
+               mm_slot = list_entry(mm_slot->mm_list.next,
+                                               struct mm_slot, mm_list);
+               spin_unlock(&ksm_mmlist_lock);
        }
 
+       ksm_scan.seqnr = 0;
+out:
        spin_lock(&ksm_mmlist_lock);
-       if (ksm_scan.mm_slot != &ksm_mm_head) {
-               ksm_scan.mm_slot = &ksm_mm_head;
-               ksm_scan.seqnr++;
-       }
+       ksm_scan.mm_slot = &ksm_mm_head;
        spin_unlock(&ksm_mmlist_lock);
+       return err;
 }
 
 static void remove_mm_from_lists(struct mm_struct *mm)
@@ -1051,6 +1100,8 @@ static void cmp_and_merge_page(struct page *page, struct rmap_item *rmap_item)
        /*
         * A ksm page might have got here by fork, but its other
         * references have already been removed from the stable tree.
+        * Or it might be left over from a break_ksm which failed
+        * when the mem_cgroup had reached its limit: try again now.
         */
        if (PageKsm(page))
                break_cow(rmap_item->mm, rmap_item->address);
@@ -1286,6 +1337,7 @@ int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
                unsigned long end, int advice, unsigned long *vm_flags)
 {
        struct mm_struct *mm = vma->vm_mm;
+       int err;
 
        switch (advice) {
        case MADV_MERGEABLE:
@@ -1298,9 +1350,11 @@ int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
                                 VM_MIXEDMAP  | VM_SAO))
                        return 0;               /* just ignore the advice */
 
-               if (!test_bit(MMF_VM_MERGEABLE, &mm->flags))
-                       if (__ksm_enter(mm) < 0)
-                               return -EAGAIN;
+               if (!test_bit(MMF_VM_MERGEABLE, &mm->flags)) {
+                       err = __ksm_enter(mm);
+                       if (err)
+                               return err;
+               }
 
                *vm_flags |= VM_MERGEABLE;
                break;
@@ -1309,8 +1363,11 @@ int ksm_madvise(struct vm_area_struct *vma, unsigned long start,
                if (!(*vm_flags & VM_MERGEABLE))
                        return 0;               /* just ignore the advice */
 
-               if (vma->anon_vma)
-                       unmerge_ksm_pages(vma, start, end);
+               if (vma->anon_vma) {
+                       err = unmerge_ksm_pages(vma, start, end);
+                       if (err)
+                               return err;
+               }
 
                *vm_flags &= ~VM_MERGEABLE;
                break;
@@ -1441,8 +1498,13 @@ static ssize_t run_store(struct kobject *kobj, struct kobj_attribute *attr,
        mutex_lock(&ksm_thread_mutex);
        if (ksm_run != flags) {
                ksm_run = flags;
-               if (flags & KSM_RUN_UNMERGE)
-                       unmerge_and_remove_all_rmap_items();
+               if (flags & KSM_RUN_UNMERGE) {
+                       err = unmerge_and_remove_all_rmap_items();
+                       if (err) {
+                               ksm_run = KSM_RUN_STOP;
+                               count = err;
+                       }
+               }
        }
        mutex_unlock(&ksm_thread_mutex);