On Thu 12-01-12 16:57:42, Eric Sandeen wrote:
On 1/11/12 7:20 PM, Jan Kara wrote:
quoted
@@ -183,6 +186,13 @@ static struct super_block *alloc_super(struct file_system_type *type)
s->s_shrink.seeks = DEFAULT_SEEKS;
s->s_shrink.shrink = prune_super;
s->s_shrink.batch = 1024;
+
+ init_waitqueue_head(&s->s_writers_wait);
+#ifdef CONFIG_SMP
+ s->s_page_faults = alloc_percpu(int);
isn't this s->s_writers? s->s_page_faults isn't defined anywhere.
Right. Leftover from original implementation and since I was doing
initial testing only using UML, I didn't spot this. Thanks.
quoted
+#endif
+ lockdep_init_map(&s->s_writers_lock_map, "sb_writers",
+ &sb_writers_key, 0);
}
out:
return s;
@@ -1126,6 +1136,84 @@ out:
}
/**
+ * sb_start_write - drop write access to a superblock
^^^^^^^^^^^^^^
s/b sb_end_write
Fixed.
quoted
@@ -1136,6 +1224,7 @@ out:
int freeze_super(struct super_block *sb)
{
int ret;
+ int writers;
atomic_inc(&sb->s_active);
down_write(&sb->s_umount);@@ -1151,8 +1240,36 @@ int freeze_super(struct super_block *sb)
return 0;
}
+ rwsem_acquire(&sb->s_writers_lock_map, 0, 0, _THIS_IP_);
sb->s_frozen = SB_FREEZE_WRITE;
- smp_wmb();
+ /*
+ * Now wait for all page faults to finish. ->page_mkwrite()
+ * implementations must call vfs_check_frozen() before starting
+ * a fault so that we cannot livelock here. Because of that we
+ * are guaranteed that from this moment on new ->page_mkwrite()
+ * calls will block and we just have to wait for s_page_faults
wait for s_writers, right?
Yes. Fixed.
Thanks for review.
Honza
--
Jan Kara [off-list ref]
SUSE Labs, CR