Received-SPF: pass (google.com: domain of linux-ext4-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Jan Kara <jack@suse.cz>
To:     Ted Tso <tytso@mit.edu>
Cc:     <linux-ext4@vger.kernel.org>,
        Ritesh Harjani <ritesh.list@gmail.com>, Jan Kara <jack@suse.cz>
Subject: [PATCH 08/10] ext2: Avoid deleting xattr block that is being reused
Date:   Tue, 12 Jul 2022 12:54:27 +0200
Message-Id: <20220712105436.32204-8-jack@suse.cz>
In-Reply-To: <20220712104519.29887-1-jack@suse.cz>
References: <20220712104519.29887-1-jack@suse.cz>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Currently when we decide to reuse xattr block we detect the case when
the last reference to xattr block is being dropped at the same time and
cancel the reuse attempt. Convert ext2 to a new scheme when as soon as
matching mbcache entry is found, we wait with dropping the last xattr
block reference until mbcache entry reference is dropped (meaning either
the xattr block reference is increased or we decided not to reuse the
block).

Signed-off-by: Jan Kara <jack@suse.cz>
---
 fs/ext2/xattr.c | 58 ++++++++++++++++++++++++-------------------------
 1 file changed, 29 insertions(+), 29 deletions(-)

diff --git a/fs/ext2/xattr.c b/fs/ext2/xattr.c
index 37ce495eb279..641abfa4b718 100644
--- a/fs/ext2/xattr.c
+++ b/fs/ext2/xattr.c
@@ -522,17 +522,18 @@ ext2_xattr_set(struct inode *inode, int name_index, const char *name,
 		lock_buffer(bh);
 		if (header->h_refcount == cpu_to_le32(1)) {
 			__u32 hash = le32_to_cpu(header->h_hash);
+			struct mb_cache_entry *oe;
 
-			ea_bdebug(bh, "modifying in-place");
+			oe = mb_cache_entry_delete_or_get(EA_BLOCK_CACHE(inode),
+					hash, bh->b_blocknr);
+			if (!oe) {
+				ea_bdebug(bh, "modifying in-place");
+				goto update_block;
+			}
 			/*
-			 * This must happen under buffer lock for
-			 * ext2_xattr_set2() to reliably detect modified block
+			 * Someone is trying to reuse the block, leave it alone
 			 */
-			mb_cache_entry_delete(EA_BLOCK_CACHE(inode), hash,
-					      bh->b_blocknr);
-
-			/* keep the buffer locked while modifying it. */
-			goto update_block;
+			mb_cache_entry_put(EA_BLOCK_CACHE(inode), oe);
 		}
 		unlock_buffer(bh);
 		ea_bdebug(bh, "cloning");
@@ -656,16 +657,29 @@ static void ext2_xattr_release_block(struct inode *inode,
 {
 	struct mb_cache *ea_block_cache = EA_BLOCK_CACHE(inode);
 
+retry_ref:
 	lock_buffer(bh);
 	if (HDR(bh)->h_refcount == cpu_to_le32(1)) {
 		__u32 hash = le32_to_cpu(HDR(bh)->h_hash);
+		struct mb_cache_entry *oe;
 
 		/*
-		 * This must happen under buffer lock for
-		 * ext2_xattr_set2() to reliably detect freed block
+		 * This must happen under buffer lock to properly
+		 * serialize with ext2_xattr_set() reusing the block.
 		 */
-		mb_cache_entry_delete(ea_block_cache, hash,
-				      bh->b_blocknr);
+		oe = mb_cache_entry_delete_or_get(ea_block_cache, hash,
+						  bh->b_blocknr);
+		if (oe) {
+			/*
+			 * Someone is trying to reuse the block. Wait
+			 * and retry.
+			 */
+			unlock_buffer(bh);
+			mb_cache_entry_wait_unused(oe);
+			mb_cache_entry_put(ea_block_cache, oe);
+			goto retry_ref;
+		}
+
 		/* Free the old block. */
 		ea_bdebug(bh, "freeing");
 		ext2_free_blocks(inode, bh->b_blocknr, 1);
@@ -929,7 +943,7 @@ ext2_xattr_cache_find(struct inode *inode, struct ext2_xattr_header *header)
 	if (!header->h_hash)
 		return NULL;  /* never share */
 	ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
-again:
+
 	ce = mb_cache_entry_find_first(ea_block_cache, hash);
 	while (ce) {
 		struct buffer_head *bh;
@@ -941,22 +955,8 @@ ext2_xattr_cache_find(struct inode *inode, struct ext2_xattr_header *header)
 				inode->i_ino, (unsigned long) ce->e_value);
 		} else {
 			lock_buffer(bh);
-			/*
-			 * We have to be careful about races with freeing or
-			 * rehashing of xattr block. Once we hold buffer lock
-			 * xattr block's state is stable so we can check
-			 * whether the block got freed / rehashed or not.
-			 * Since we unhash mbcache entry under buffer lock when
-			 * freeing / rehashing xattr block, checking whether
-			 * entry is still hashed is reliable.
-			 */
-			if (hlist_bl_unhashed(&ce->e_hash_list)) {
-				mb_cache_entry_put(ea_block_cache, ce);
-				unlock_buffer(bh);
-				brelse(bh);
-				goto again;
-			} else if (le32_to_cpu(HDR(bh)->h_refcount) >
-				   EXT2_XATTR_REFCOUNT_MAX) {
+			if (le32_to_cpu(HDR(bh)->h_refcount) >
+			    EXT2_XATTR_REFCOUNT_MAX) {
 				ea_idebug(inode, "block %ld refcount %d>%d",
 					  (unsigned long) ce->e_value,
 					  le32_to_cpu(HDR(bh)->h_refcount),
-- 
2.35.3