Hello! This changeset implements a fix for iget4() race possible on reiserfs filesystems. This is minimalistic fix that is intended to be replaced with iget5_locked() implementation later on. The iget5_locked patch is too much of a change for 2.4.21-rc2, I afraid. The race itself may lead to pretty much disasters, we've already seen it may lead to direntries pointing to nowhere, incorrect nlink counts, NFS problems. Please pull from bk://namesys.com/bk/reiser3-linux-2.4-race-fix Diffstat: inode.c | 26 ++++++++++++++++++++++++-- 1 files changed, 24 insertions(+), 2 deletions(-) Plain text patch: # This is a BitKeeper generated patch for the following project: # Project Name: Linux kernel tree # This patch format is intended for GNU patch command version 2.5 or higher. # This patch includes the following deltas: # ChangeSet 1.988 -> 1.989 # fs/reiserfs/inode.c 1.42 -> 1.43 # # The following is the BitKeeper ChangeSet Log # -------------------------------------------- # 03/05/03 green@angband.namesys.com 1.989 # reiserfs: iget4() race fix # -------------------------------------------- # diff -Nru a/fs/reiserfs/inode.c b/fs/reiserfs/inode.c --- a/fs/reiserfs/inode.c Sat May 3 16:38:17 2003 +++ b/fs/reiserfs/inode.c Sat May 3 16:38:17 2003 @@ -20,6 +20,10 @@ static int reiserfs_get_block (struct inode * inode, long block, struct buffer_head * bh_result, int create); +/* This spinlock guards inode pkey in private part of inode + against race between find_actor() vs reiserfs_read_inode2 */ +static spinlock_t keycopy_lock = SPIN_LOCK_UNLOCKED; + void reiserfs_delete_inode (struct inode * inode) { int jbegin_count = JOURNAL_PER_BALANCE_CNT * 2; @@ -898,8 +902,9 @@ bh = PATH_PLAST_BUFFER (path); ih = PATH_PITEM_HEAD (path); - + spin_lock(&keycopy_lock); copy_key (INODE_PKEY (inode), &(ih->ih_key)); + spin_unlock(&keycopy_lock); inode->i_blksize = PAGE_SIZE; INIT_LIST_HEAD(&inode->u.reiserfs_i.i_prealloc_list) ; @@ -1220,10 +1225,27 @@ unsigned long inode_no, void *opaque ) { struct reiserfs_iget4_args *args; + int retval; args = opaque; + /* We protect against possible parallel init_inode() on another CPU here. */ + spin_lock(&keycopy_lock); /* args is already in CPU order */ - return le32_to_cpu(INODE_PKEY(inode)->k_dir_id) == args -> objectid; + if (le32_to_cpu(INODE_PKEY(inode)->k_dir_id) == args -> objectid) + retval = 1; + else + /* If The key does not match, lets see if we are racing + with another iget4, that already progressed so far + to reiserfs_read_inode2() and was preempted in + call to search_by_key(). The signs of that are: + Inode is locked + dirid and object id are zero (not yet initialized)*/ + retval = (inode->i_state & I_LOCK) && + !INODE_PKEY(inode)->k_dir_id && + !INODE_PKEY(inode)->k_objectid; + + spin_unlock(&keycopy_lock); + return retval; } struct inode * reiserfs_iget (struct super_block * s, const struct cpu_key * key)