f2fs: remove i_xattr_sem to avoid deadlock and fix the original issue

This reverts commit 27161f1 "f2fs: avoid race in between read xattr & write xattr".

That introduced a deadlock case:

Thread #1:

[122554.641906][   T92]  f2fs_getxattr+0xd4/0x5fc
    -> waiting for f2fs_down_read(&F2FS_I(inode)->i_xattr_sem);

[122554.641927][   T92]  __f2fs_get_acl+0x50/0x284
[122554.641948][   T92]  f2fs_init_acl+0x84/0x54c
[122554.641969][   T92]  f2fs_init_inode_metadata+0x460/0x5f0
[122554.641990][   T92]  f2fs_add_inline_entry+0x11c/0x350
    -> Locked dir->inode_page by f2fs_get_node_page()

[122554.642009][   T92]  f2fs_do_add_link+0x100/0x1e4
[122554.642025][   T92]  f2fs_create+0xf4/0x22c
[122554.642047][   T92]  vfs_create+0x130/0x1f4

Thread #2:

[123996.386358][   T92]  __get_node_page+0x8c/0x504
    -> waiting for dir->inode_page lock

[123996.386383][   T92]  read_all_xattrs+0x11c/0x1f4
[123996.386405][   T92]  __f2fs_setxattr+0xcc/0x528
[123996.386424][   T92]  f2fs_setxattr+0x158/0x1f4
    -> f2fs_down_write(&F2FS_I(inode)->i_xattr_sem);

[123996.386443][   T92]  __f2fs_set_acl+0x328/0x430
[123996.386618][   T92]  f2fs_set_acl+0x38/0x50
[123996.386642][   T92]  posix_acl_chmod+0xc8/0x1c8
[123996.386669][   T92]  f2fs_setattr+0x5e0/0x6bc
[123996.386689][   T92]  notify_change+0x4d8/0x580
[123996.386717][   T92]  chmod_common+0xd8/0x184
[123996.386748][   T92]  do_fchmodat+0x60/0x124
[123996.386766][   T92]  __arm64_sys_fchmodat+0x28/0x3c

Let's take a look at the original issue back.

Thread A:                                       Thread B:
-f2fs_getxattr
   -lookup_all_xattrs
      -xnid = F2FS_I(inode)->i_xattr_nid;
                                                -f2fs_setxattr
                                                    -__f2fs_setxattr
                                                        -write_all_xattrs
                                                            -truncate_xattr_node
                                                                  ...  ...
                                                -write_checkpoint
                                                                  ...  ...
                                                -alloc_nid   <- nid reuse
          -get_node_page
              -f2fs_bug_on  <- nid != node_footer->nid

I think we don't need to truncate xattr pages eagerly which introduces lots of
data races without big benefits.

Cc: <[email protected]>
Signed-off-by: Jaegeuk Kim <[email protected]>

fs/f2fs/f2fs.h

@@ -838,7 +838,6 @@ struct f2fs_inode_info {
 
 	/* avoid racing between foreground op and gc */
 	struct f2fs_rwsem i_gc_rwsem[2];
-	struct f2fs_rwsem i_xattr_sem; /* avoid racing between reading and changing EAs */
 
 	int i_extra_isize;		/* size of extra space located in i_addr */
 	kprojid_t i_projid;		/* id for project quota */

fs/f2fs/super.c

@@ -1418,7 +1418,6 @@ static struct inode *f2fs_alloc_inode(struct super_block *sb)
 	INIT_LIST_HEAD(&fi->gdirty_list);
 	init_f2fs_rwsem(&fi->i_gc_rwsem[READ]);
 	init_f2fs_rwsem(&fi->i_gc_rwsem[WRITE]);
-	init_f2fs_rwsem(&fi->i_xattr_sem);
 
 	/* Will be used by directory only */
 	fi->i_dir_level = F2FS_SB(sb)->dir_level;

fs/f2fs/xattr.c

@@ -432,7 +432,7 @@ static inline int write_all_xattrs(struct inode *inode, __u32 hsize,
 {
 	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
 	size_t inline_size = inline_xattr_size(inode);
-	struct page *in_page = NULL;
+	struct page *in_page = ipage;
 	void *xattr_addr;
 	void *inline_addr = NULL;
 	struct page *xpage;
@@ -445,29 +445,19 @@ static inline int write_all_xattrs(struct inode *inode, __u32 hsize,
 
 	/* write to inline xattr */
 	if (inline_size) {
-		if (ipage) {
-			inline_addr = inline_xattr_addr(inode, ipage);
-		} else {
+		if (!in_page) {
 			in_page = f2fs_get_node_page(sbi, inode->i_ino);
 			if (IS_ERR(in_page)) {
 				f2fs_alloc_nid_failed(sbi, new_nid);
 				return PTR_ERR(in_page);
 			}
-			inline_addr = inline_xattr_addr(inode, in_page);
 		}
+		inline_addr = inline_xattr_addr(inode, in_page);
 
-		f2fs_wait_on_page_writeback(ipage ? ipage : in_page,
-							NODE, true, true);
-		/* no need to use xattr node block */
+		f2fs_wait_on_page_writeback(in_page, NODE, true, true);
 		if (hsize <= inline_size) {
-			err = f2fs_truncate_xattr_node(inode);
-			f2fs_alloc_nid_failed(sbi, new_nid);
-			if (err) {
-				f2fs_put_page(in_page, 1);
-				return err;
-			}
 			memcpy(inline_addr, txattr_addr, inline_size);
-			set_page_dirty(ipage ? ipage : in_page);
+			set_page_dirty(in_page);
 			goto in_page_out;
 		}
 	}
@@ -501,12 +491,13 @@ static inline int write_all_xattrs(struct inode *inode, __u32 hsize,
 	memcpy(xattr_addr, txattr_addr + inline_size, VALID_XATTR_BLOCK_SIZE);
 
 	if (inline_size)
-		set_page_dirty(ipage ? ipage : in_page);
+		set_page_dirty(in_page);
 	set_page_dirty(xpage);
 
 	f2fs_put_page(xpage, 1);
 in_page_out:
-	f2fs_put_page(in_page, 1);
+	if (in_page != ipage)
+		f2fs_put_page(in_page, 1);
 	return err;
 }
 
@@ -527,10 +518,8 @@ int f2fs_getxattr(struct inode *inode, int index, const char *name,
 	if (len > F2FS_NAME_LEN)
 		return -ERANGE;
 
-	f2fs_down_read(&F2FS_I(inode)->i_xattr_sem);
 	error = lookup_all_xattrs(inode, ipage, index, len, name,
 				&entry, &base_addr, &base_size, &is_inline);
-	f2fs_up_read(&F2FS_I(inode)->i_xattr_sem);
 	if (error)
 		return error;
 
@@ -564,9 +553,7 @@ ssize_t f2fs_listxattr(struct dentry *dentry, char *buffer, size_t buffer_size)
 	int error;
 	size_t rest = buffer_size;
 
-	f2fs_down_read(&F2FS_I(inode)->i_xattr_sem);
 	error = read_all_xattrs(inode, NULL, &base_addr);
-	f2fs_up_read(&F2FS_I(inode)->i_xattr_sem);
 	if (error)
 		return error;
 
@@ -794,9 +781,7 @@ int f2fs_setxattr(struct inode *inode, int index, const char *name,
 	f2fs_balance_fs(sbi, true);
 
 	f2fs_lock_op(sbi);
-	f2fs_down_write(&F2FS_I(inode)->i_xattr_sem);
 	err = __f2fs_setxattr(inode, index, name, value, size, ipage, flags);
-	f2fs_up_write(&F2FS_I(inode)->i_xattr_sem);
 	f2fs_unlock_op(sbi);
 
 	f2fs_update_time(sbi, REQ_TIME);