On Mon, 2009-04-13 at 09:40 -0700, Sage Weil wrote: > Starting in b7ec40d7845bffca8bb3af2ea3f192d6257bbe21, drop_dirty_roots() > tries to avoid generating delayed refs in a transaction that is currently > closing (and trying to flush dirty refs out) by waiting for it to close. > However, if a transaction is held open by a user space TRANS_START ioctl, > that will deadlock against throttle_on_drops(). > > The underlying problem is that by the time we are ready to wait in > wait_transaction_pre_flush(), drop_dirty_roots() has already signaled its > intent to drop in fs_info->throttles, and the process calling > throttle_on_drops() will block. If the throttling process is part of a > user transaction, the transaction will never close and we deadlock. > Thanks for this patch, I think that even if non-userland ioctl code doesn't deadlock it can wait too long for throttle_on_drops in this case. Would something like the patch below also work?: diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c index 2869b33..01b1436 100644 --- a/fs/btrfs/transaction.c +++ b/fs/btrfs/transaction.c @@ -687,7 +687,13 @@ static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info) prepare_to_wait(&info->transaction_wait, &wait, TASK_UNINTERRUPTIBLE); mutex_unlock(&info->trans_mutex); + + atomic_dec(&info->throttles); + wake_up(&info->transaction_throttle); + schedule(); + + atomic_inc(&info->throttles); mutex_lock(&info->trans_mutex); finish_wait(&info->transaction_wait, &wait); } -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
