ext2: fix race between setxattr and write back

There's an issue when allocating xattrs as follows:
Block Allocation Reservation Windows Map (ext2_try_to_allocate_with_rsv):
reservation window 0x000000006f105382 start: 0, end: 0
reservation window 0x000000008fd1a555 start: 1044, end: 1059
Window map complete.
kernel BUG at fs/ext2/balloc.c:1158!
invalid opcode: 0000 [#1] PREEMPT SMP KASAN
RIP: 0010:ext2_try_to_allocate_with_rsv.isra.0+0x15c4/0x1800
Call Trace:
 <TASK>
 ext2_new_blocks+0x935/0x1690
 ext2_new_block+0x73/0xa0
 ext2_xattr_set2+0x74f/0x1730
 ext2_xattr_set+0x12b6/0x2260
 ext2_xattr_user_set+0x9c/0x110
 __vfs_setxattr+0x139/0x1d0
 __vfs_setxattr_noperm+0xfc/0x370
 __vfs_setxattr_locked+0x205/0x2c0
 vfs_setxattr+0x19d/0x3b0
 do_setxattr+0xff/0x220
 setxattr+0x123/0x150
 path_setxattr+0x193/0x1e0
 __x64_sys_setxattr+0xc8/0x170
 do_syscall_64+0x35/0x80
 entry_SYSCALL_64_after_hwframe+0x63/0xcd

Above issue may happens as follows:
        setxattr                             write back
ext2_xattr_set
  ext2_xattr_set2
    ext2_new_block
      ext2_new_blocks
        ext2_try_to_allocate_with_rsv
          alloc_new_reservation
          --> group=0 [0, 1023] rsv [1016, 1023]
                                        do_writepages
                                          mpage_writepages
                                            write_cache_pages
                                              __mpage_writepage
                                                ext2_get_block
                                                  ext2_get_blocks
                                                   ext2_alloc_branch
                                                    ext2_new_blocks
                                                     ext2_try_to_allocate_with_rsv
                                                       alloc_new_reservation
                                     -->group=1 [1024, 2047] rsv [1044, 1059]
          if ((my_rsv->rsv_start > group_last_block) ||
              (my_rsv->rsv_end < group_first_block)
              rsv_window_dump
              BUG();

Now ext2 mkwrite doesn't allocate new blocks so for these cases we may
be allocating blocks during writeback. However, there is no protection
between ext2_xattr_set() and do_writepages() so these two functions can
conflict on handling the reservation window.  To solve about issue don't
use the reservation window when allocating block for xattr.

Signed-off-by: Ye Bin <yebin10@huawei.com>
Message-Id: <20230815112612.221145-4-yebin10@huawei.com>
Signed-off-by: Jan Kara <jack@suse.cz>

2 files changed, 9 insertions, 8 deletions

diff --git a/fs/ext2/balloc.c b/fs/ext2/balloc.c
index a9648c3137db..dd74c77a4ac1 100644
--- a/fs/ext2/balloc.c
+++ b/fs/ext2/balloc.c
@@ -1244,15 +1244,15 @@ ext2_fsblk_t ext2_new_blocks(struct inode *inode, ext2_fsblk_t goal,
 	es = EXT2_SB(sb)->s_es;
 	ext2_debug("goal=%lu.\n", goal);
 	/*
-	 * Allocate a block from reservation only when
-	 * filesystem is mounted with reservation(default,-o reservation), and
-	 * it's a regular file, and
-	 * the desired window size is greater than 0 (One could use ioctl
-	 * command EXT2_IOC_SETRSVSZ to set the window size to 0 to turn off
-	 * reservation on that particular file)
+	 * Allocate a block from reservation only when the filesystem is
+	 * mounted with reservation(default,-o reservation), and it's a regular
+	 * file, and the desired window size is greater than 0 (One could use
+	 * ioctl command EXT2_IOC_SETRSVSZ to set the window size to 0 to turn
+	 * off reservation on that particular file). Also do not use the
+	 * reservation window if the caller asked us not to do it.
 	 */
 	block_i = EXT2_I(inode)->i_block_alloc_info;
-	if (block_i) {
+	if (!(flags & EXT2_ALLOC_NORESERVE) && block_i) {
 		windowsz = block_i->rsv_window_node.rsv_goal_size;
 		if (windowsz > 0)
 			my_rsv = &block_i->rsv_window_node;
diff --git a/fs/ext2/xattr.c b/fs/ext2/xattr.c
index 3b4ff6974ab8..b7c64fc1c008 100644
--- a/fs/ext2/xattr.c
+++ b/fs/ext2/xattr.c
@@ -744,7 +744,8 @@ ext2_xattr_set2(struct inode *inode, struct buffer_head *old_bh,
 						EXT2_I(inode)->i_block_group);
 			unsigned long count = 1;
 			ext2_fsblk_t block = ext2_new_blocks(inode, goal,
-							     &count, &error, 0);
+						&count, &error,
+						EXT2_ALLOC_NORESERVE);
 			if (error)
 				goto cleanup;
 			ea_idebug(inode, "creating block %lu", block);