btrfs: fix race between balance and cancel/pause

Syzbot reported a panic that looks like this:

  assertion failed: fs_info->exclusive_operation == BTRFS_EXCLOP_BALANCE_PAUSED, in fs/btrfs/ioctl.c:465
  ------------[ cut here ]------------
  kernel BUG at fs/btrfs/messages.c:259!
  RIP: 0010:btrfs_assertfail+0x2c/0x30 fs/btrfs/messages.c:259
  Call Trace:
   <TASK>
   btrfs_exclop_balance fs/btrfs/ioctl.c:465 [inline]
   btrfs_ioctl_balance fs/btrfs/ioctl.c:3564 [inline]
   btrfs_ioctl+0x531e/0x5b30 fs/btrfs/ioctl.c:4632
   vfs_ioctl fs/ioctl.c:51 [inline]
   __do_sys_ioctl fs/ioctl.c:870 [inline]
   __se_sys_ioctl fs/ioctl.c:856 [inline]
   __x64_sys_ioctl+0x197/0x210 fs/ioctl.c:856
   do_syscall_x64 arch/x86/entry/common.c:50 [inline]
   do_syscall_64+0x39/0xb0 arch/x86/entry/common.c:80
   entry_SYSCALL_64_after_hwframe+0x63/0xcd

The reproducer is running a balance and a cancel or pause in parallel.
The way balance finishes is a bit wonky, if we were paused we need to
save the balance_ctl in the fs_info, but clear it otherwise and cleanup.
However we rely on the return values being specific errors, or having a
cancel request or no pause request.  If balance completes and returns 0,
but we have a pause or cancel request we won't do the appropriate
cleanup, and then the next time we try to start a balance we'll trip
this ASSERT.

The error handling is just wrong here, we always want to clean up,
unless we got -ECANCELLED and we set the appropriate pause flag in the
exclusive op.  With this patch the reproducer ran for an hour without
tripping, previously it would trip in less than a few minutes.

Reported-by: [email protected]
CC: [email protected] # 6.1+
Signed-off-by: Josef Bacik <[email protected]>
Reviewed-by: David Sterba <[email protected]>
Signed-off-by: David Sterba <[email protected]>

Author: josefbacik

fs/btrfs/volumes.c

@@ -4081,14 +4081,6 @@ static int alloc_profile_is_valid(u64 flags, int extended)
 	return has_single_bit_set(flags);
 }
 
-static inline int balance_need_close(struct btrfs_fs_info *fs_info)
-{
-	/* cancel requested || normal exit path */
-	return atomic_read(&fs_info->balance_cancel_req) ||
-		(atomic_read(&fs_info->balance_pause_req) == 0 &&
-		 atomic_read(&fs_info->balance_cancel_req) == 0);
-}
-
 /*
  * Validate target profile against allowed profiles and return true if it's OK.
  * Otherwise print the error message and return false.
@@ -4278,6 +4270,7 @@ int btrfs_balance(struct btrfs_fs_info *fs_info,
 	u64 num_devices;
 	unsigned seq;
 	bool reducing_redundancy;
+	bool paused = false;
 	int i;
 
 	if (btrfs_fs_closing(fs_info) ||
@@ -4408,6 +4401,7 @@ int btrfs_balance(struct btrfs_fs_info *fs_info,
 	if (ret == -ECANCELED && atomic_read(&fs_info->balance_pause_req)) {
 		btrfs_info(fs_info, "balance: paused");
 		btrfs_exclop_balance(fs_info, BTRFS_EXCLOP_BALANCE_PAUSED);
+		paused = true;
 	}
 	/*
 	 * Balance can be canceled by:
@@ -4436,8 +4430,8 @@ int btrfs_balance(struct btrfs_fs_info *fs_info,
 		btrfs_update_ioctl_balance_args(fs_info, bargs);
 	}
 
-	if ((ret && ret != -ECANCELED && ret != -ENOSPC) ||
-	    balance_need_close(fs_info)) {
+	/* We didn't pause, we can clean everything up. */
+	if (!paused) {
 		reset_balance_state(fs_info);
 		btrfs_exclop_finish(fs_info);
 	}