shmem: fix faulting into a hole, not taking i_mutex
authorHugh Dickins <hughd@google.com>
Wed, 23 Jul 2014 21:00:10 +0000 (14:00 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 23 Jul 2014 22:10:54 +0000 (15:10 -0700)
Commit f00cdc6df7d7 ("shmem: fix faulting into a hole while it's
punched") was buggy: Sasha sent a lockdep report to remind us that
grabbing i_mutex in the fault path is a no-no (write syscall may already
hold i_mutex while faulting user buffer).

We tried a completely different approach (see following patch) but that
proved inadequate: good enough for a rational workload, but not good
enough against trinity - which forks off so many mappings of the object
that contention on i_mmap_mutex while hole-puncher holds i_mutex builds
into serious starvation when concurrent faults force the puncher to fall
back to single-page unmap_mapping_range() searches of the i_mmap tree.

So return to the original umbrella approach, but keep away from i_mutex
this time.  We really don't want to bloat every shmem inode with a new
mutex or completion, just to protect this unlikely case from trinity.
So extend the original with wait_queue_head on stack at the hole-punch
end, and wait_queue item on the stack at the fault end.

This involves further use of i_lock to guard against the races: lockdep
has been happy so far, and I see fs/inode.c:unlock_new_inode() holds
i_lock around wake_up_bit(), which is comparable to what we do here.
i_lock is more convenient, but we could switch to shmem's info->lock.

This issue has been tagged with CVE-2014-4171, which will require commit
f00cdc6df7d7 and this and the following patch to be backported: we
suggest to 3.1+, though in fact the trinity forkbomb effect might go
back as far as 2.6.16, when madvise(,,MADV_REMOVE) came in - or might
not, since much has changed, with i_mmap_mutex a spinlock before 3.0.
Anyone running trinity on 3.0 and earlier? I don't think we need care.

Signed-off-by: Hugh Dickins <hughd@google.com>
Reported-by: Sasha Levin <sasha.levin@oracle.com>
Tested-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Lukas Czerner <lczerner@redhat.com>
Cc: Dave Jones <davej@redhat.com>
Cc: <stable@vger.kernel.org> [3.1+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/shmem.c

index 1140f49b6ded6f7a72d89d2e89f9fce0df1a940d..c0719f0822460a84a7791b7d61ea3934558f0787 100644 (file)
@@ -85,7 +85,7 @@ static struct vfsmount *shm_mnt;
  * a time): we would prefer not to enlarge the shmem inode just for that.
  */
 struct shmem_falloc {
-       int     mode;           /* FALLOC_FL mode currently operating */
+       wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
        pgoff_t start;          /* start of range currently being fallocated */
        pgoff_t next;           /* the next page offset to be fallocated */
        pgoff_t nr_falloced;    /* how many new pages have been fallocated */
@@ -760,7 +760,7 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
                        spin_lock(&inode->i_lock);
                        shmem_falloc = inode->i_private;
                        if (shmem_falloc &&
-                           !shmem_falloc->mode &&
+                           !shmem_falloc->waitq &&
                            index >= shmem_falloc->start &&
                            index < shmem_falloc->next)
                                shmem_falloc->nr_unswapped++;
@@ -1248,38 +1248,58 @@ static int shmem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
         * Trinity finds that probing a hole which tmpfs is punching can
         * prevent the hole-punch from ever completing: which in turn
         * locks writers out with its hold on i_mutex.  So refrain from
-        * faulting pages into the hole while it's being punched, and
-        * wait on i_mutex to be released if vmf->flags permits.
+        * faulting pages into the hole while it's being punched.  Although
+        * shmem_undo_range() does remove the additions, it may be unable to
+        * keep up, as each new page needs its own unmap_mapping_range() call,
+        * and the i_mmap tree grows ever slower to scan if new vmas are added.
+        *
+        * It does not matter if we sometimes reach this check just before the
+        * hole-punch begins, so that one fault then races with the punch:
+        * we just need to make racing faults a rare case.
+        *
+        * The implementation below would be much simpler if we just used a
+        * standard mutex or completion: but we cannot take i_mutex in fault,
+        * and bloating every shmem inode for this unlikely case would be sad.
         */
        if (unlikely(inode->i_private)) {
                struct shmem_falloc *shmem_falloc;
 
                spin_lock(&inode->i_lock);
                shmem_falloc = inode->i_private;
-               if (!shmem_falloc ||
-                   shmem_falloc->mode != FALLOC_FL_PUNCH_HOLE ||
-                   vmf->pgoff < shmem_falloc->start ||
-                   vmf->pgoff >= shmem_falloc->next)
-                       shmem_falloc = NULL;
-               spin_unlock(&inode->i_lock);
-               /*
-                * i_lock has protected us from taking shmem_falloc seriously
-                * once return from shmem_fallocate() went back up that stack.
-                * i_lock does not serialize with i_mutex at all, but it does
-                * not matter if sometimes we wait unnecessarily, or sometimes
-                * miss out on waiting: we just need to make those cases rare.
-                */
-               if (shmem_falloc) {
+               if (shmem_falloc &&
+                   shmem_falloc->waitq &&
+                   vmf->pgoff >= shmem_falloc->start &&
+                   vmf->pgoff < shmem_falloc->next) {
+                       wait_queue_head_t *shmem_falloc_waitq;
+                       DEFINE_WAIT(shmem_fault_wait);
+
+                       ret = VM_FAULT_NOPAGE;
                        if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
                           !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
+                               /* It's polite to up mmap_sem if we can */
                                up_read(&vma->vm_mm->mmap_sem);
-                               mutex_lock(&inode->i_mutex);
-                               mutex_unlock(&inode->i_mutex);
-                               return VM_FAULT_RETRY;
+                               ret = VM_FAULT_RETRY;
                        }
-                       /* cond_resched? Leave that to GUP or return to user */
-                       return VM_FAULT_NOPAGE;
+
+                       shmem_falloc_waitq = shmem_falloc->waitq;
+                       prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
+                                       TASK_UNINTERRUPTIBLE);
+                       spin_unlock(&inode->i_lock);
+                       schedule();
+
+                       /*
+                        * shmem_falloc_waitq points into the shmem_fallocate()
+                        * stack of the hole-punching task: shmem_falloc_waitq
+                        * is usually invalid by the time we reach here, but
+                        * finish_wait() does not dereference it in that case;
+                        * though i_lock needed lest racing with wake_up_all().
+                        */
+                       spin_lock(&inode->i_lock);
+                       finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
+                       spin_unlock(&inode->i_lock);
+                       return ret;
                }
+               spin_unlock(&inode->i_lock);
        }
 
        error = shmem_getpage(inode, vmf->pgoff, &vmf->page, SGP_CACHE, &ret);
@@ -1774,13 +1794,13 @@ static long shmem_fallocate(struct file *file, int mode, loff_t offset,
 
        mutex_lock(&inode->i_mutex);
 
-       shmem_falloc.mode = mode & ~FALLOC_FL_KEEP_SIZE;
-
        if (mode & FALLOC_FL_PUNCH_HOLE) {
                struct address_space *mapping = file->f_mapping;
                loff_t unmap_start = round_up(offset, PAGE_SIZE);
                loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
+               DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
 
+               shmem_falloc.waitq = &shmem_falloc_waitq;
                shmem_falloc.start = unmap_start >> PAGE_SHIFT;
                shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
                spin_lock(&inode->i_lock);
@@ -1792,8 +1812,13 @@ static long shmem_fallocate(struct file *file, int mode, loff_t offset,
                                            1 + unmap_end - unmap_start, 0);
                shmem_truncate_range(inode, offset, offset + len - 1);
                /* No need to unmap again: hole-punching leaves COWed pages */
+
+               spin_lock(&inode->i_lock);
+               inode->i_private = NULL;
+               wake_up_all(&shmem_falloc_waitq);
+               spin_unlock(&inode->i_lock);
                error = 0;
-               goto undone;
+               goto out;
        }
 
        /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
@@ -1809,6 +1834,7 @@ static long shmem_fallocate(struct file *file, int mode, loff_t offset,
                goto out;
        }
 
+       shmem_falloc.waitq = NULL;
        shmem_falloc.start = start;
        shmem_falloc.next  = start;
        shmem_falloc.nr_falloced = 0;