]> git.ipfire.org Git - thirdparty/linux.git/commitdiff
btrfs: use irq safe locking when running and adding delayed iputs
authorFilipe Manana <fdmanana@suse.com>
Wed, 5 Jul 2023 23:41:16 +0000 (00:41 +0100)
committerDavid Sterba <dsterba@suse.com>
Tue, 18 Jul 2023 01:13:10 +0000 (03:13 +0200)
Running delayed iputs, which never happens in an irq context, needs to
lock the spinlock fs_info->delayed_iput_lock. When finishing bios for
data writes (irq context, bio.c) we call btrfs_put_ordered_extent() which
needs to add a delayed iput and for that it needs to acquire the spinlock
fs_info->delayed_iput_lock. Without disabling irqs when running delayed
iputs we can therefore deadlock on that spinlock. The same deadlock can
also happen when adding an inode to the delayed iputs list, since this
can be done outside an irq context as well.

Syzbot recently reported this, which results in the following trace:

  ================================
  WARNING: inconsistent lock state
  6.4.0-syzkaller-09904-ga507db1d8fdc #0 Not tainted
  --------------------------------
  inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-W} usage.
  btrfs-cleaner/16079 [HC0[0]:SC0[0]:HE1:SE1] takes:
  ffff888107804d20 (&fs_info->delayed_iput_lock){+.?.}-{2:2}, at: spin_lock include/linux/spinlock.h:350 [inline]
  ffff888107804d20 (&fs_info->delayed_iput_lock){+.?.}-{2:2}, at: btrfs_run_delayed_iputs+0x28/0xe0 fs/btrfs/inode.c:3523
  {IN-SOFTIRQ-W} state was registered at:
    lock_acquire kernel/locking/lockdep.c:5761 [inline]
    lock_acquire+0x1b1/0x520 kernel/locking/lockdep.c:5726
    __raw_spin_lock include/linux/spinlock_api_smp.h:133 [inline]
    _raw_spin_lock+0x2e/0x40 kernel/locking/spinlock.c:154
    spin_lock include/linux/spinlock.h:350 [inline]
    btrfs_add_delayed_iput+0x128/0x390 fs/btrfs/inode.c:3490
    btrfs_put_ordered_extent fs/btrfs/ordered-data.c:559 [inline]
    btrfs_put_ordered_extent+0x2f6/0x610 fs/btrfs/ordered-data.c:547
    __btrfs_bio_end_io fs/btrfs/bio.c:118 [inline]
    __btrfs_bio_end_io+0x136/0x180 fs/btrfs/bio.c:112
    btrfs_orig_bbio_end_io+0x86/0x2b0 fs/btrfs/bio.c:163
    btrfs_simple_end_io+0x105/0x380 fs/btrfs/bio.c:378
    bio_endio+0x589/0x690 block/bio.c:1617
    req_bio_endio block/blk-mq.c:766 [inline]
    blk_update_request+0x5c5/0x1620 block/blk-mq.c:911
    blk_mq_end_request+0x59/0x680 block/blk-mq.c:1032
    lo_complete_rq+0x1c6/0x280 drivers/block/loop.c:370
    blk_complete_reqs+0xb3/0xf0 block/blk-mq.c:1110
    __do_softirq+0x1d4/0x905 kernel/softirq.c:553
    run_ksoftirqd kernel/softirq.c:921 [inline]
    run_ksoftirqd+0x31/0x60 kernel/softirq.c:913
    smpboot_thread_fn+0x659/0x9e0 kernel/smpboot.c:164
    kthread+0x344/0x440 kernel/kthread.c:389
    ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:308
  irq event stamp: 39
  hardirqs last  enabled at (39): [<ffffffff81d5ebc4>] __do_kmem_cache_free mm/slab.c:3558 [inline]
  hardirqs last  enabled at (39): [<ffffffff81d5ebc4>] kmem_cache_free mm/slab.c:3582 [inline]
  hardirqs last  enabled at (39): [<ffffffff81d5ebc4>] kmem_cache_free+0x244/0x370 mm/slab.c:3575
  hardirqs last disabled at (38): [<ffffffff81d5eb5e>] __do_kmem_cache_free mm/slab.c:3553 [inline]
  hardirqs last disabled at (38): [<ffffffff81d5eb5e>] kmem_cache_free mm/slab.c:3582 [inline]
  hardirqs last disabled at (38): [<ffffffff81d5eb5e>] kmem_cache_free+0x1de/0x370 mm/slab.c:3575
  softirqs last  enabled at (0): [<ffffffff814ac99f>] copy_process+0x227f/0x75c0 kernel/fork.c:2448
  softirqs last disabled at (0): [<0000000000000000>] 0x0

  other info that might help us debug this:
   Possible unsafe locking scenario:

         CPU0
         ----
    lock(&fs_info->delayed_iput_lock);
    <Interrupt>
      lock(&fs_info->delayed_iput_lock);

   *** DEADLOCK ***

  1 lock held by btrfs-cleaner/16079:
   #0: ffff888107804860 (&fs_info->cleaner_mutex){+.+.}-{3:3}, at: cleaner_kthread+0x103/0x4b0 fs/btrfs/disk-io.c:1463

  stack backtrace:
  CPU: 3 PID: 16079 Comm: btrfs-cleaner Not tainted 6.4.0-syzkaller-09904-ga507db1d8fdc #0
  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-2 04/01/2014
  Call Trace:
   <TASK>
   __dump_stack lib/dump_stack.c:88 [inline]
   dump_stack_lvl+0xd9/0x150 lib/dump_stack.c:106
   print_usage_bug kernel/locking/lockdep.c:3978 [inline]
   valid_state kernel/locking/lockdep.c:4020 [inline]
   mark_lock_irq kernel/locking/lockdep.c:4223 [inline]
   mark_lock.part.0+0x1102/0x1960 kernel/locking/lockdep.c:4685
   mark_lock kernel/locking/lockdep.c:4649 [inline]
   mark_usage kernel/locking/lockdep.c:4598 [inline]
   __lock_acquire+0x8e4/0x5e20 kernel/locking/lockdep.c:5098
   lock_acquire kernel/locking/lockdep.c:5761 [inline]
   lock_acquire+0x1b1/0x520 kernel/locking/lockdep.c:5726
   __raw_spin_lock include/linux/spinlock_api_smp.h:133 [inline]
   _raw_spin_lock+0x2e/0x40 kernel/locking/spinlock.c:154
   spin_lock include/linux/spinlock.h:350 [inline]
   btrfs_run_delayed_iputs+0x28/0xe0 fs/btrfs/inode.c:3523
   cleaner_kthread+0x2e5/0x4b0 fs/btrfs/disk-io.c:1478
   kthread+0x344/0x440 kernel/kthread.c:389
   ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:308
   </TASK>

So fix this by using spin_lock_irq() and spin_unlock_irq() when running
delayed iputs, and using spin_lock_irqsave() and spin_unlock_irqrestore()
when adding a delayed iput().

Reported-by: syzbot+da501a04be5ff533b102@syzkaller.appspotmail.com
Fixes: ec63b84d4611 ("btrfs: add an ordered_extent pointer to struct btrfs_bio")
Link: https://lore.kernel.org/linux-btrfs/000000000000d5c89a05ffbd39dd@google.com/
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/inode.c

index c8921589e2f3009b92c7d55a9f7e4efceb31a7f7..b12850b31cb3cd26dbb2020580871da36e5c8bcb 100644 (file)
@@ -3482,15 +3482,21 @@ zeroit:
 void btrfs_add_delayed_iput(struct btrfs_inode *inode)
 {
        struct btrfs_fs_info *fs_info = inode->root->fs_info;
+       unsigned long flags;
 
        if (atomic_add_unless(&inode->vfs_inode.i_count, -1, 1))
                return;
 
        atomic_inc(&fs_info->nr_delayed_iputs);
-       spin_lock(&fs_info->delayed_iput_lock);
+       /*
+        * Need to be irq safe here because we can be called from either an irq
+        * context (see bio.c and btrfs_put_ordered_extent()) or a non-irq
+        * context.
+        */
+       spin_lock_irqsave(&fs_info->delayed_iput_lock, flags);
        ASSERT(list_empty(&inode->delayed_iput));
        list_add_tail(&inode->delayed_iput, &fs_info->delayed_iputs);
-       spin_unlock(&fs_info->delayed_iput_lock);
+       spin_unlock_irqrestore(&fs_info->delayed_iput_lock, flags);
        if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
                wake_up_process(fs_info->cleaner_kthread);
 }
@@ -3499,37 +3505,46 @@ static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
                                    struct btrfs_inode *inode)
 {
        list_del_init(&inode->delayed_iput);
-       spin_unlock(&fs_info->delayed_iput_lock);
+       spin_unlock_irq(&fs_info->delayed_iput_lock);
        iput(&inode->vfs_inode);
        if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
                wake_up(&fs_info->delayed_iputs_wait);
-       spin_lock(&fs_info->delayed_iput_lock);
+       spin_lock_irq(&fs_info->delayed_iput_lock);
 }
 
 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
                                   struct btrfs_inode *inode)
 {
        if (!list_empty(&inode->delayed_iput)) {
-               spin_lock(&fs_info->delayed_iput_lock);
+               spin_lock_irq(&fs_info->delayed_iput_lock);
                if (!list_empty(&inode->delayed_iput))
                        run_delayed_iput_locked(fs_info, inode);
-               spin_unlock(&fs_info->delayed_iput_lock);
+               spin_unlock_irq(&fs_info->delayed_iput_lock);
        }
 }
 
 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
 {
-
-       spin_lock(&fs_info->delayed_iput_lock);
+       /*
+        * btrfs_put_ordered_extent() can run in irq context (see bio.c), which
+        * calls btrfs_add_delayed_iput() and that needs to lock
+        * fs_info->delayed_iput_lock. So we need to disable irqs here to
+        * prevent a deadlock.
+        */
+       spin_lock_irq(&fs_info->delayed_iput_lock);
        while (!list_empty(&fs_info->delayed_iputs)) {
                struct btrfs_inode *inode;
 
                inode = list_first_entry(&fs_info->delayed_iputs,
                                struct btrfs_inode, delayed_iput);
                run_delayed_iput_locked(fs_info, inode);
-               cond_resched_lock(&fs_info->delayed_iput_lock);
+               if (need_resched()) {
+                       spin_unlock_irq(&fs_info->delayed_iput_lock);
+                       cond_resched();
+                       spin_lock_irq(&fs_info->delayed_iput_lock);
+               }
        }
-       spin_unlock(&fs_info->delayed_iput_lock);
+       spin_unlock_irq(&fs_info->delayed_iput_lock);
 }
 
 /*