ocfs2: serialize unaligned aio
Mark Fasheh [Wed, 22 Jun 2011 21:23:38 +0000 (14:23 -0700)]
Fix a corruption that can happen when we have (two or more) outstanding
aio's to an overlapping unaligned region.  Ext4
(e9e3bcecf44c04b9e6b505fd8e2eb9cea58fb94d) and xfs recently had to fix
similar issues.

In our case what happens is that we can have an outstanding aio on a region
and if a write comes in with some bytes overlapping the original aio we may
decide to read that region into a page before continuing (typically because
of buffered-io fallback).  Since we have no ordering guarantees with the
aio, we can read stale or bad data into the page and then write it back out.

If the i/o is page and block aligned, then we avoid this issue as there
won't be any need to read data from disk.

I took the same approach as Eric in the ext4 patch and introduced some
serialization of unaligned async direct i/o.  I don't expect this to have an
effect on the most common cases of AIO.  Unaligned aio will be slower
though, but that's far more acceptable than data corruption.

Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <jlbec@evilplan.org>

fs/ocfs2/aops.c
fs/ocfs2/aops.h
fs/ocfs2/file.c
fs/ocfs2/inode.h
fs/ocfs2/super.c

index ac97bca..4c1ec8f 100644 (file)
@@ -564,6 +564,7 @@ static void ocfs2_dio_end_io(struct kiocb *iocb,
 {
        struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
        int level;
+       wait_queue_head_t *wq = ocfs2_ioend_wq(inode);
 
        /* this io's submitter should not have unlocked this before we could */
        BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
@@ -573,6 +574,15 @@ static void ocfs2_dio_end_io(struct kiocb *iocb,
                ocfs2_iocb_clear_sem_locked(iocb);
        }
 
+       if (ocfs2_iocb_is_unaligned_aio(iocb)) {
+               ocfs2_iocb_clear_unaligned_aio(iocb);
+
+               if (atomic_dec_and_test(&OCFS2_I(inode)->ip_unaligned_aio) &&
+                   waitqueue_active(wq)) {
+                       wake_up_all(wq);
+               }
+       }
+
        ocfs2_iocb_clear_rw_locked(iocb);
 
        level = ocfs2_iocb_rw_locked_level(iocb);
index 75cf3ad..ffb2da3 100644 (file)
@@ -78,6 +78,7 @@ enum ocfs2_iocb_lock_bits {
        OCFS2_IOCB_RW_LOCK = 0,
        OCFS2_IOCB_RW_LOCK_LEVEL,
        OCFS2_IOCB_SEM,
+       OCFS2_IOCB_UNALIGNED_IO,
        OCFS2_IOCB_NUM_LOCKS
 };
 
@@ -91,4 +92,17 @@ enum ocfs2_iocb_lock_bits {
        clear_bit(OCFS2_IOCB_SEM, (unsigned long *)&iocb->private)
 #define ocfs2_iocb_is_sem_locked(iocb) \
        test_bit(OCFS2_IOCB_SEM, (unsigned long *)&iocb->private)
+
+#define ocfs2_iocb_set_unaligned_aio(iocb) \
+       set_bit(OCFS2_IOCB_UNALIGNED_IO, (unsigned long *)&iocb->private)
+#define ocfs2_iocb_clear_unaligned_aio(iocb) \
+       clear_bit(OCFS2_IOCB_UNALIGNED_IO, (unsigned long *)&iocb->private)
+#define ocfs2_iocb_is_unaligned_aio(iocb) \
+       test_bit(OCFS2_IOCB_UNALIGNED_IO, (unsigned long *)&iocb->private)
+
+#define OCFS2_IOEND_WQ_HASH_SZ 37
+#define ocfs2_ioend_wq(v)   (&ocfs2__ioend_wq[((unsigned long)(v)) %\
+                                           OCFS2_IOEND_WQ_HASH_SZ])
+extern wait_queue_head_t ocfs2__ioend_wq[OCFS2_IOEND_WQ_HASH_SZ];
+
 #endif /* OCFS2_FILE_H */
index b1e35a3..145f453 100644 (file)
@@ -2038,6 +2038,23 @@ out:
        return ret;
 }
 
+static void ocfs2_aiodio_wait(struct inode *inode)
+{
+       wait_queue_head_t *wq = ocfs2_ioend_wq(inode);
+
+       wait_event(*wq, (atomic_read(&OCFS2_I(inode)->ip_unaligned_aio) == 0));
+}
+
+static int ocfs2_is_io_unaligned(struct inode *inode, size_t count, loff_t pos)
+{
+       int blockmask = inode->i_sb->s_blocksize - 1;
+       loff_t final_size = pos + count;
+
+       if ((pos & blockmask) || (final_size & blockmask))
+               return 1;
+       return 0;
+}
+
 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
                                            struct file *file,
                                            loff_t pos, size_t count,
@@ -2216,6 +2233,7 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
        struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
        int full_coherency = !(osb->s_mount_opt &
                               OCFS2_MOUNT_COHERENCY_BUFFERED);
+       int unaligned_dio = 0;
 
        trace_ocfs2_file_aio_write(inode, file, file->f_path.dentry,
                (unsigned long long)OCFS2_I(inode)->ip_blkno,
@@ -2284,6 +2302,10 @@ relock:
                goto out;
        }
 
+       if (direct_io && !is_sync_kiocb(iocb))
+               unaligned_dio = ocfs2_is_io_unaligned(inode, iocb->ki_left,
+                                                     *ppos);
+
        /*
         * We can't complete the direct I/O as requested, fall back to
         * buffered I/O.
@@ -2299,6 +2321,18 @@ relock:
                goto relock;
        }
 
+       if (unaligned_dio) {
+               /*
+                * Wait on previous unaligned aio to complete before
+                * proceeding.
+                */
+               ocfs2_aiodio_wait(inode);
+
+               /* Mark the iocb as needing a decrement in ocfs2_dio_end_io */
+               atomic_inc(&OCFS2_I(inode)->ip_unaligned_aio);
+               ocfs2_iocb_set_unaligned_aio(iocb);
+       }
+
        /*
         * To later detect whether a journal commit for sync writes is
         * necessary, we sample i_size, and cluster count here.
@@ -2371,8 +2405,12 @@ out_dio:
        if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
                rw_level = -1;
                have_alloc_sem = 0;
+               unaligned_dio = 0;
        }
 
+       if (unaligned_dio)
+               atomic_dec(&OCFS2_I(inode)->ip_unaligned_aio);
+
 out:
        if (rw_level != -1)
                ocfs2_rw_unlock(inode, rw_level);
index 1c508b1..88924a3 100644 (file)
@@ -43,6 +43,9 @@ struct ocfs2_inode_info
        /* protects extended attribute changes on this inode */
        struct rw_semaphore             ip_xattr_sem;
 
+       /* Number of outstanding AIO's which are not page aligned */
+       atomic_t                        ip_unaligned_aio;
+
        /* These fields are protected by ip_lock */
        spinlock_t                      ip_lock;
        u32                             ip_open_count;
index 029c4cd..603f5fe 100644 (file)
@@ -54,6 +54,7 @@
 #include "ocfs1_fs_compat.h"
 
 #include "alloc.h"
+#include "aops.h"
 #include "blockcheck.h"
 #include "dlmglue.h"
 #include "export.h"
@@ -1616,12 +1617,17 @@ static int ocfs2_show_options(struct seq_file *s, struct vfsmount *mnt)
        return 0;
 }
 
+wait_queue_head_t ocfs2__ioend_wq[OCFS2_IOEND_WQ_HASH_SZ];
+
 static int __init ocfs2_init(void)
 {
-       int status;
+       int status, i;
 
        ocfs2_print_version();
 
+       for (i = 0; i < OCFS2_IOEND_WQ_HASH_SZ; i++)
+               init_waitqueue_head(&ocfs2__ioend_wq[i]);
+
        status = init_ocfs2_uptodate_cache();
        if (status < 0) {
                mlog_errno(status);
@@ -1760,7 +1766,7 @@ static void ocfs2_inode_init_once(void *data)
        ocfs2_extent_map_init(&oi->vfs_inode);
        INIT_LIST_HEAD(&oi->ip_io_markers);
        oi->ip_dir_start_lookup = 0;
-
+       atomic_set(&oi->ip_unaligned_aio, 0);
        init_rwsem(&oi->ip_alloc_sem);
        init_rwsem(&oi->ip_xattr_sem);
        mutex_init(&oi->ip_io_mutex);