reiserfs: don't lock root inode searching
authorFrederic Weisbecker <fweisbec@gmail.com>
Tue, 10 Jan 2012 23:11:11 +0000 (15:11 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 11 Jan 2012 00:30:54 +0000 (16:30 -0800)
Nothing requires that we lock the filesystem until the root inode is
provided.

Also iget5_locked() triggers a warning because we are holding the
filesystem lock while allocating the inode, which result in a lockdep
suspicion that we have a lock inversion against the reclaim path:

[ 1986.896979] =================================
[ 1986.896990] [ INFO: inconsistent lock state ]
[ 1986.896997] 3.1.1-main #8
[ 1986.897001] ---------------------------------
[ 1986.897007] inconsistent {RECLAIM_FS-ON-W} -> {IN-RECLAIM_FS-W} usage.
[ 1986.897016] kswapd0/16 [HC0[0]:SC0[0]:HE1:SE1] takes:
[ 1986.897023]  (&REISERFS_SB(s)->lock){+.+.?.}, at: [<c01f8bd4>] reiserfs_write_lock+0x20/0x2a
[ 1986.897044] {RECLAIM_FS-ON-W} state was registered at:
[ 1986.897050]   [<c014a5b9>] mark_held_locks+0xae/0xd0
[ 1986.897060]   [<c014aab3>] lockdep_trace_alloc+0x7d/0x91
[ 1986.897068]   [<c0190ee0>] kmem_cache_alloc+0x1a/0x93
[ 1986.897078]   [<c01e7728>] reiserfs_alloc_inode+0x13/0x3d
[ 1986.897088]   [<c01a5b06>] alloc_inode+0x14/0x5f
[ 1986.897097]   [<c01a5cb9>] iget5_locked+0x62/0x13a
[ 1986.897106]   [<c01e99e0>] reiserfs_fill_super+0x410/0x8b9
[ 1986.897114]   [<c01953da>] mount_bdev+0x10b/0x159
[ 1986.897123]   [<c01e764d>] get_super_block+0x10/0x12
[ 1986.897131]   [<c0195b38>] mount_fs+0x59/0x12d
[ 1986.897138]   [<c01a80d1>] vfs_kern_mount+0x45/0x7a
[ 1986.897147]   [<c01a83e3>] do_kern_mount+0x2f/0xb0
[ 1986.897155]   [<c01a987a>] do_mount+0x5c2/0x612
[ 1986.897163]   [<c01a9a72>] sys_mount+0x61/0x8f
[ 1986.897170]   [<c044060c>] sysenter_do_call+0x12/0x32
[ 1986.897181] irq event stamp: 7509691
[ 1986.897186] hardirqs last  enabled at (7509691): [<c0190f34>] kmem_cache_alloc+0x6e/0x93
[ 1986.897197] hardirqs last disabled at (7509690): [<c0190eea>] kmem_cache_alloc+0x24/0x93
[ 1986.897209] softirqs last  enabled at (7508896): [<c01294bd>] __do_softirq+0xee/0xfd
[ 1986.897222] softirqs last disabled at (7508859): [<c01030ed>] do_softirq+0x50/0x9d
[ 1986.897234]
[ 1986.897235] other info that might help us debug this:
[ 1986.897242]  Possible unsafe locking scenario:
[ 1986.897244]
[ 1986.897250]        CPU0
[ 1986.897254]        ----
[ 1986.897257]   lock(&REISERFS_SB(s)->lock);
[ 1986.897265] <Interrupt>
[ 1986.897269]     lock(&REISERFS_SB(s)->lock);
[ 1986.897276]
[ 1986.897277]  *** DEADLOCK ***
[ 1986.897278]
[ 1986.897286] no locks held by kswapd0/16.
[ 1986.897291]
[ 1986.897292] stack backtrace:
[ 1986.897299] Pid: 16, comm: kswapd0 Not tainted 3.1.1-main #8
[ 1986.897306] Call Trace:
[ 1986.897314]  [<c0439e76>] ? printk+0xf/0x11
[ 1986.897324]  [<c01482d1>] print_usage_bug+0x20e/0x21a
[ 1986.897332]  [<c01479b8>] ? print_irq_inversion_bug+0x172/0x172
[ 1986.897341]  [<c014855c>] mark_lock+0x27f/0x483
[ 1986.897349]  [<c0148d88>] __lock_acquire+0x628/0x1472
[ 1986.897358]  [<c0149fae>] lock_acquire+0x47/0x5e
[ 1986.897366]  [<c01f8bd4>] ? reiserfs_write_lock+0x20/0x2a
[ 1986.897384]  [<c01f8bd4>] ? reiserfs_write_lock+0x20/0x2a
[ 1986.897397]  [<c043b5ef>] mutex_lock_nested+0x35/0x26f
[ 1986.897409]  [<c01f8bd4>] ? reiserfs_write_lock+0x20/0x2a
[ 1986.897421]  [<c01f8bd4>] reiserfs_write_lock+0x20/0x2a
[ 1986.897433]  [<c01e2edd>] map_block_for_writepage+0xc9/0x590
[ 1986.897448]  [<c01b1706>] ? create_empty_buffers+0x33/0x8f
[ 1986.897461]  [<c0121124>] ? get_parent_ip+0xb/0x31
[ 1986.897472]  [<c043ef7f>] ? sub_preempt_count+0x81/0x8e
[ 1986.897485]  [<c043cae0>] ? _raw_spin_unlock+0x27/0x3d
[ 1986.897496]  [<c0121124>] ? get_parent_ip+0xb/0x31
[ 1986.897508]  [<c01e355d>] reiserfs_writepage+0x1b9/0x3e7
[ 1986.897521]  [<c0173b40>] ? clear_page_dirty_for_io+0xcb/0xde
[ 1986.897533]  [<c014a6e3>] ? trace_hardirqs_on_caller+0x108/0x138
[ 1986.897546]  [<c014a71e>] ? trace_hardirqs_on+0xb/0xd
[ 1986.897559]  [<c0177b38>] shrink_page_list+0x34f/0x5e2
[ 1986.897572]  [<c01780a7>] shrink_inactive_list+0x172/0x22c
[ 1986.897585]  [<c0178464>] shrink_zone+0x303/0x3b1
[ 1986.897597]  [<c043cae0>] ? _raw_spin_unlock+0x27/0x3d
[ 1986.897611]  [<c01788c9>] kswapd+0x3b7/0x5f2

The deadlock shouldn't happen since we are doing that allocation in the
mount path, the filesystem is not available for any reclaim.  Still the
warning is annoying.

To solve this, acquire the lock later only where we need it, right before
calling reiserfs_read_locked_inode() that wants to lock to walk the tree.

Reported-by: Knut Petersen <Knut_Petersen@t-online.de>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/reiserfs/super.c

index 61b60380d2ce37c1388f59a05d06e3667f2a6cd7..e12d8b97cd4dbf3c4c7dc249d42c66a8aeadc4bb 100644 (file)
@@ -1519,9 +1519,7 @@ static int read_super_block(struct super_block *s, int offset)
 static int reread_meta_blocks(struct super_block *s)
 {
        ll_rw_block(READ, 1, &(SB_BUFFER_WITH_SB(s)));
-       reiserfs_write_unlock(s);
        wait_on_buffer(SB_BUFFER_WITH_SB(s));
-       reiserfs_write_lock(s);
        if (!buffer_uptodate(SB_BUFFER_WITH_SB(s))) {
                reiserfs_warning(s, "reiserfs-2504", "error reading the super");
                return 1;
@@ -1837,24 +1835,14 @@ static int reiserfs_fill_super(struct super_block *s, void *data, int silent)
                                 */
        }
 
-       /*
-        * This path assumed to be called with the BKL in the old times.
-        * Now we have inherited the big reiserfs lock from it and many
-        * reiserfs helpers called in the mount path and elsewhere require
-        * this lock to be held even if it's not always necessary. Let's be
-        * conservative and hold it early. The window can be reduced after
-        * careful review of the code.
-        */
-       reiserfs_write_lock(s);
-
        if (reread_meta_blocks(s)) {
                SWARN(silent, s, "jmacd-9",
                      "unable to reread meta blocks after journal init");
-               goto error;
+               goto error_unlocked;
        }
 
        if (replay_only(s))
-               goto error;
+               goto error_unlocked;
 
        if (bdev_read_only(s->s_bdev) && !(s->s_flags & MS_RDONLY)) {
                SWARN(silent, s, "clm-7000",
@@ -1868,9 +1856,19 @@ static int reiserfs_fill_super(struct super_block *s, void *data, int silent)
                         reiserfs_init_locked_inode, (void *)(&args));
        if (!root_inode) {
                SWARN(silent, s, "jmacd-10", "get root inode failed");
-               goto error;
+               goto error_unlocked;
        }
 
+       /*
+        * This path assumed to be called with the BKL in the old times.
+        * Now we have inherited the big reiserfs lock from it and many
+        * reiserfs helpers called in the mount path and elsewhere require
+        * this lock to be held even if it's not always necessary. Let's be
+        * conservative and hold it early. The window can be reduced after
+        * careful review of the code.
+        */
+       reiserfs_write_lock(s);
+
        if (root_inode->i_state & I_NEW) {
                reiserfs_read_locked_inode(root_inode, &args);
                unlock_new_inode(root_inode);