ext4: serialize unaligned asynchronous DIO
authorEric Sandeen <sandeen@redhat.com>
Sat, 12 Feb 2011 13:17:34 +0000 (08:17 -0500)
committerTheodore Ts'o <tytso@mit.edu>
Sat, 12 Feb 2011 13:17:34 +0000 (08:17 -0500)
ext4 has a data corruption case when doing non-block-aligned
asynchronous direct IO into a sparse file, as demonstrated
by xfstest 240.

The root cause is that while ext4 preallocates space in the
hole, mappings of that space still look "new" and
dio_zero_block() will zero out the unwritten portions.  When
more than one AIO thread is going, they both find this "new"
block and race to zero out their portion; this is uncoordinated
and causes data corruption.

Dave Chinner fixed this for xfs by simply serializing all
unaligned asynchronous direct IO.  I've done the same here.
The difference is that we only wait on conversions, not all IO.
This is a very big hammer, and I'm not very pleased with
stuffing this into ext4_file_write().  But since ext4 is
DIO_LOCKING, we need to serialize it at this high level.

I tried to move this into ext4_ext_direct_IO, but by then
we have the i_mutex already, and we will wait on the
work queue to do conversions - which must also take the
i_mutex.  So that won't work.

This was originally exposed by qemu-kvm installing to
a raw disk image with a normal sector-63 alignment.  I've
tested a backport of this patch with qemu, and it does
avoid the corruption.  It is also quite a lot slower
(14 min for package installs, vs. 8 min for well-aligned)
but I'll take slow correctness over fast corruption any day.

Mingming suggested that we can track outstanding
conversions, and wait on those so that non-sparse
files won't be affected, and I've implemented that here;
unaligned AIO to nonsparse files won't take a perf hit.

[tytso@mit.edu: Keep the mutex as a hashed array instead
 of bloating the ext4 inode]

[tytso@mit.edu: Fix up namespace issues so that global
 variables are protected with an "ext4_" prefix.]

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fs/ext4/ext4.h
fs/ext4/extents.c
fs/ext4/file.c
fs/ext4/page-io.c
fs/ext4/super.c

index 0c8d97b56f34858c09943407916aefdbcc80ef7f..3aa0b72b3b94b42bea4a499e836068ed991814ce 100644 (file)
@@ -848,6 +848,7 @@ struct ext4_inode_info {
        atomic_t i_ioend_count; /* Number of outstanding io_end structs */
        /* current io_end structure for async DIO write*/
        ext4_io_end_t *cur_aio_dio;
+       atomic_t i_aiodio_unwritten; /* Nr. of inflight conversions pending */
 
        spinlock_t i_block_reservation_lock;
 
@@ -2119,6 +2120,15 @@ static inline void set_bitmap_uptodate(struct buffer_head *bh)
 
 #define in_range(b, first, len)        ((b) >= (first) && (b) <= (first) + (len) - 1)
 
+/* For ioend & aio unwritten conversion wait queues */
+#define EXT4_WQ_HASH_SZ                37
+#define ext4_ioend_wq(v)   (&ext4__ioend_wq[((unsigned long)(v)) %\
+                                           EXT4_WQ_HASH_SZ])
+#define ext4_aio_mutex(v)  (&ext4__aio_mutex[((unsigned long)(v)) %\
+                                            EXT4_WQ_HASH_SZ])
+extern wait_queue_head_t ext4__ioend_wq[EXT4_WQ_HASH_SZ];
+extern struct mutex ext4__aio_mutex[EXT4_WQ_HASH_SZ];
+
 #endif /* __KERNEL__ */
 
 #endif /* _EXT4_H */
index 63a75810b7c3ab1b58b802d1d456275160180fa9..ccce8a7e94edc9ca2d23b76e94bd45339869970d 100644 (file)
@@ -3174,9 +3174,10 @@ ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
                 * that this IO needs to convertion to written when IO is
                 * completed
                 */
-               if (io)
+               if (io && !(io->flag & EXT4_IO_END_UNWRITTEN)) {
                        io->flag = EXT4_IO_END_UNWRITTEN;
-               else
+                       atomic_inc(&EXT4_I(inode)->i_aiodio_unwritten);
+               } else
                        ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
                if (ext4_should_dioread_nolock(inode))
                        map->m_flags |= EXT4_MAP_UNINIT;
@@ -3463,9 +3464,10 @@ int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
                 * that we need to perform convertion when IO is done.
                 */
                if ((flags & EXT4_GET_BLOCKS_PRE_IO)) {
-                       if (io)
+                       if (io && !(io->flag & EXT4_IO_END_UNWRITTEN)) {
                                io->flag = EXT4_IO_END_UNWRITTEN;
-                       else
+                               atomic_inc(&EXT4_I(inode)->i_aiodio_unwritten);
+                       } else
                                ext4_set_inode_state(inode,
                                                     EXT4_STATE_DIO_UNWRITTEN);
                }
index 2e8322c8aa882f977410322b62d88346104b86c6..7b80d543b89e71c41764b54a5d6eee59992c712c 100644 (file)
@@ -55,11 +55,47 @@ static int ext4_release_file(struct inode *inode, struct file *filp)
        return 0;
 }
 
+static void ext4_aiodio_wait(struct inode *inode)
+{
+       wait_queue_head_t *wq = ext4_ioend_wq(inode);
+
+       wait_event(*wq, (atomic_read(&EXT4_I(inode)->i_aiodio_unwritten) == 0));
+}
+
+/*
+ * This tests whether the IO in question is block-aligned or not.
+ * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
+ * are converted to written only after the IO is complete.  Until they are
+ * mapped, these blocks appear as holes, so dio_zero_block() will assume that
+ * it needs to zero out portions of the start and/or end block.  If 2 AIO
+ * threads are at work on the same unwritten block, they must be synchronized
+ * or one thread will zero the other's data, causing corruption.
+ */
+static int
+ext4_unaligned_aio(struct inode *inode, const struct iovec *iov,
+                  unsigned long nr_segs, loff_t pos)
+{
+       struct super_block *sb = inode->i_sb;
+       int blockmask = sb->s_blocksize - 1;
+       size_t count = iov_length(iov, nr_segs);
+       loff_t final_size = pos + count;
+
+       if (pos >= inode->i_size)
+               return 0;
+
+       if ((pos & blockmask) || (final_size & blockmask))
+               return 1;
+
+       return 0;
+}
+
 static ssize_t
 ext4_file_write(struct kiocb *iocb, const struct iovec *iov,
                unsigned long nr_segs, loff_t pos)
 {
        struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
+       int unaligned_aio = 0;
+       int ret;
 
        /*
         * If we have encountered a bitmap-format file, the size limit
@@ -78,9 +114,31 @@ ext4_file_write(struct kiocb *iocb, const struct iovec *iov,
                        nr_segs = iov_shorten((struct iovec *)iov, nr_segs,
                                              sbi->s_bitmap_maxbytes - pos);
                }
+       } else if (unlikely((iocb->ki_filp->f_flags & O_DIRECT) &&
+                  !is_sync_kiocb(iocb))) {
+               unaligned_aio = ext4_unaligned_aio(inode, iov, nr_segs, pos);
        }
 
-       return generic_file_aio_write(iocb, iov, nr_segs, pos);
+       /* Unaligned direct AIO must be serialized; see comment above */
+       if (unaligned_aio) {
+               static unsigned long unaligned_warn_time;
+
+               /* Warn about this once per day */
+               if (printk_timed_ratelimit(&unaligned_warn_time, 60*60*24*HZ))
+                       ext4_msg(inode->i_sb, KERN_WARNING,
+                                "Unaligned AIO/DIO on inode %ld by %s; "
+                                "performance will be poor.",
+                                inode->i_ino, current->comm);
+               mutex_lock(ext4_aio_mutex(inode));
+               ext4_aiodio_wait(inode);
+       }
+
+       ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
+
+       if (unaligned_aio)
+               mutex_unlock(ext4_aio_mutex(inode));
+
+       return ret;
 }
 
 static const struct vm_operations_struct ext4_file_vm_ops = {
index 4e9b0a242f4c8a53ce7f5d1c6817e74eebf35afc..955cc309142fe983dc41d95467b1f5a66d7a6b69 100644 (file)
 
 static struct kmem_cache *io_page_cachep, *io_end_cachep;
 
-#define WQ_HASH_SZ             37
-#define to_ioend_wq(v) (&ioend_wq[((unsigned long)v) % WQ_HASH_SZ])
-static wait_queue_head_t ioend_wq[WQ_HASH_SZ];
-
 int __init ext4_init_pageio(void)
 {
-       int i;
-
        io_page_cachep = KMEM_CACHE(ext4_io_page, SLAB_RECLAIM_ACCOUNT);
        if (io_page_cachep == NULL)
                return -ENOMEM;
@@ -48,9 +42,6 @@ int __init ext4_init_pageio(void)
                kmem_cache_destroy(io_page_cachep);
                return -ENOMEM;
        }
-       for (i = 0; i < WQ_HASH_SZ; i++)
-               init_waitqueue_head(&ioend_wq[i]);
-
        return 0;
 }
 
@@ -62,7 +53,7 @@ void ext4_exit_pageio(void)
 
 void ext4_ioend_wait(struct inode *inode)
 {
-       wait_queue_head_t *wq = to_ioend_wq(inode);
+       wait_queue_head_t *wq = ext4_ioend_wq(inode);
 
        wait_event(*wq, (atomic_read(&EXT4_I(inode)->i_ioend_count) == 0));
 }
@@ -87,7 +78,7 @@ void ext4_free_io_end(ext4_io_end_t *io)
        for (i = 0; i < io->num_io_pages; i++)
                put_io_page(io->pages[i]);
        io->num_io_pages = 0;
-       wq = to_ioend_wq(io->inode);
+       wq = ext4_ioend_wq(io->inode);
        if (atomic_dec_and_test(&EXT4_I(io->inode)->i_ioend_count) &&
            waitqueue_active(wq))
                wake_up_all(wq);
@@ -102,6 +93,7 @@ int ext4_end_io_nolock(ext4_io_end_t *io)
        struct inode *inode = io->inode;
        loff_t offset = io->offset;
        ssize_t size = io->size;
+       wait_queue_head_t *wq;
        int ret = 0;
 
        ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
@@ -126,7 +118,16 @@ int ext4_end_io_nolock(ext4_io_end_t *io)
        if (io->iocb)
                aio_complete(io->iocb, io->result, 0);
        /* clear the DIO AIO unwritten flag */
-       io->flag &= ~EXT4_IO_END_UNWRITTEN;
+       if (io->flag & EXT4_IO_END_UNWRITTEN) {
+               io->flag &= ~EXT4_IO_END_UNWRITTEN;
+               /* Wake up anyone waiting on unwritten extent conversion */
+               wq = ext4_ioend_wq(io->inode);
+               if (atomic_dec_and_test(&EXT4_I(inode)->i_aiodio_unwritten) &&
+                   waitqueue_active(wq)) {
+                       wake_up_all(wq);
+               }
+       }
+
        return ret;
 }
 
index 86b05486dc631af4292ff2d70a0051bd692a131b..f6a318f836b2cd7d5c2e70af3ae1bde6b6d301ce 100644 (file)
@@ -833,6 +833,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
        ei->i_sync_tid = 0;
        ei->i_datasync_tid = 0;
        atomic_set(&ei->i_ioend_count, 0);
+       atomic_set(&ei->i_aiodio_unwritten, 0);
 
        return &ei->vfs_inode;
 }
@@ -4800,11 +4801,21 @@ static void ext4_exit_feat_adverts(void)
        kfree(ext4_feat);
 }
 
+/* Shared across all ext4 file systems */
+wait_queue_head_t ext4__ioend_wq[EXT4_WQ_HASH_SZ];
+struct mutex ext4__aio_mutex[EXT4_WQ_HASH_SZ];
+
 static int __init ext4_init_fs(void)
 {
-       int err;
+       int i, err;
 
        ext4_check_flag_values();
+
+       for (i = 0; i < EXT4_WQ_HASH_SZ; i++) {
+               mutex_init(&ext4__aio_mutex[i]);
+               init_waitqueue_head(&ext4__ioend_wq[i]);
+       }
+
        err = ext4_init_pageio();
        if (err)
                return err;