xfs: introduce xfs_rw_lock() helpers for locking the inode
authorDave Chinner <dchinner@redhat.com>
Wed, 12 Jan 2011 00:37:10 +0000 (11:37 +1100)
committerDave Chinner <david@fromorbit.com>
Wed, 12 Jan 2011 00:37:10 +0000 (11:37 +1100)
We need to obtain the i_mutex, i_iolock and i_ilock during the read
and write paths. Add a set of wrapper functions to neatly
encapsulate the lock ordering and shared/exclusive semantics to make
the locking easier to follow and get right.

Note that this changes some of the exclusive locking serialisation in
that serialisation will occur against the i_mutex instead of the
XFS_IOLOCK_EXCL. This does not change any behaviour, and it is
arguably more efficient to use the mutex for such serialisation than
the rw_sem.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
fs/xfs/linux-2.6/xfs_file.c

index c47d7dc0a3071a4af6b7aa730de5c0c565a4794b..b5e13fbb7386b8b179b646c8ae7747ae1f988928 100644 (file)
 
 static const struct vm_operations_struct xfs_file_vm_ops;
 
+/*
+ * Locking primitives for read and write IO paths to ensure we consistently use
+ * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
+ */
+static inline void
+xfs_rw_ilock(
+       struct xfs_inode        *ip,
+       int                     type)
+{
+       if (type & XFS_IOLOCK_EXCL)
+               mutex_lock(&VFS_I(ip)->i_mutex);
+       xfs_ilock(ip, type);
+}
+
+static inline void
+xfs_rw_iunlock(
+       struct xfs_inode        *ip,
+       int                     type)
+{
+       xfs_iunlock(ip, type);
+       if (type & XFS_IOLOCK_EXCL)
+               mutex_unlock(&VFS_I(ip)->i_mutex);
+}
+
+static inline void
+xfs_rw_ilock_demote(
+       struct xfs_inode        *ip,
+       int                     type)
+{
+       xfs_ilock_demote(ip, type);
+       if (type & XFS_IOLOCK_EXCL)
+               mutex_unlock(&VFS_I(ip)->i_mutex);
+}
+
 /*
  *     xfs_iozero
  *
@@ -262,22 +296,21 @@ xfs_file_aio_read(
        if (XFS_FORCED_SHUTDOWN(mp))
                return -EIO;
 
-       if (unlikely(ioflags & IO_ISDIRECT))
-               mutex_lock(&inode->i_mutex);
-       xfs_ilock(ip, XFS_IOLOCK_SHARED);
-
        if (unlikely(ioflags & IO_ISDIRECT)) {
+               xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
+
                if (inode->i_mapping->nrpages) {
                        ret = -xfs_flushinval_pages(ip,
                                        (iocb->ki_pos & PAGE_CACHE_MASK),
                                        -1, FI_REMAPF_LOCKED);
+                       if (ret) {
+                               xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
+                               return ret;
+                       }
                }
-               mutex_unlock(&inode->i_mutex);
-               if (ret) {
-                       xfs_iunlock(ip, XFS_IOLOCK_SHARED);
-                       return ret;
-               }
-       }
+               xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
+       } else
+               xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
 
        trace_xfs_file_read(ip, size, iocb->ki_pos, ioflags);
 
@@ -285,7 +318,7 @@ xfs_file_aio_read(
        if (ret > 0)
                XFS_STATS_ADD(xs_read_bytes, ret);
 
-       xfs_iunlock(ip, XFS_IOLOCK_SHARED);
+       xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
        return ret;
 }
 
@@ -309,7 +342,7 @@ xfs_file_splice_read(
        if (XFS_FORCED_SHUTDOWN(ip->i_mount))
                return -EIO;
 
-       xfs_ilock(ip, XFS_IOLOCK_SHARED);
+       xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
 
        trace_xfs_file_splice_read(ip, count, *ppos, ioflags);
 
@@ -317,7 +350,7 @@ xfs_file_splice_read(
        if (ret > 0)
                XFS_STATS_ADD(xs_read_bytes, ret);
 
-       xfs_iunlock(ip, XFS_IOLOCK_SHARED);
+       xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
        return ret;
 }
 
@@ -338,10 +371,10 @@ xfs_aio_write_isize_update(
                *ppos = isize;
 
        if (*ppos > ip->i_size) {
-               xfs_ilock(ip, XFS_ILOCK_EXCL);
+               xfs_rw_ilock(ip, XFS_ILOCK_EXCL);
                if (*ppos > ip->i_size)
                        ip->i_size = *ppos;
-               xfs_iunlock(ip, XFS_ILOCK_EXCL);
+               xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
        }
 }
 
@@ -356,14 +389,22 @@ xfs_aio_write_newsize_update(
        struct xfs_inode        *ip)
 {
        if (ip->i_new_size) {
-               xfs_ilock(ip, XFS_ILOCK_EXCL);
+               xfs_rw_ilock(ip, XFS_ILOCK_EXCL);
                ip->i_new_size = 0;
                if (ip->i_d.di_size > ip->i_size)
                        ip->i_d.di_size = ip->i_size;
-               xfs_iunlock(ip, XFS_ILOCK_EXCL);
+               xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
        }
 }
 
+/*
+ * xfs_file_splice_write() does not use xfs_rw_ilock() because
+ * generic_file_splice_write() takes the i_mutex itself. This, in theory,
+ * couuld cause lock inversions between the aio_write path and the splice path
+ * if someone is doing concurrent splice(2) based writes and write(2) based
+ * writes to the same inode. The only real way to fix this is to re-implement
+ * the generic code here with correct locking orders.
+ */
 STATIC ssize_t
 xfs_file_splice_write(
        struct pipe_inode_info  *pipe,
@@ -604,7 +645,6 @@ xfs_file_aio_write(
        xfs_fsize_t             new_size;
        int                     iolock;
        size_t                  ocount = 0, count;
-       int                     need_i_mutex;
 
        XFS_STATS_INC(xs_write_calls);
 
@@ -631,21 +671,17 @@ xfs_file_aio_write(
 relock:
        if (ioflags & IO_ISDIRECT) {
                iolock = XFS_IOLOCK_SHARED;
-               need_i_mutex = 0;
        } else {
                iolock = XFS_IOLOCK_EXCL;
-               need_i_mutex = 1;
-               mutex_lock(&inode->i_mutex);
        }
 
-       xfs_ilock(ip, XFS_ILOCK_EXCL|iolock);
-
 start:
+       xfs_rw_ilock(ip, XFS_ILOCK_EXCL|iolock);
        ret = generic_write_checks(file, &pos, &count,
                                        S_ISBLK(inode->i_mode));
        if (ret) {
-               xfs_iunlock(ip, XFS_ILOCK_EXCL|iolock);
-               goto out_unlock_mutex;
+               xfs_rw_iunlock(ip, XFS_ILOCK_EXCL|iolock);
+               return ret;
        }
 
        if (ioflags & IO_ISDIRECT) {
@@ -654,16 +690,20 @@ start:
                                mp->m_rtdev_targp : mp->m_ddev_targp;
 
                if ((pos & target->bt_smask) || (count & target->bt_smask)) {
-                       xfs_iunlock(ip, XFS_ILOCK_EXCL|iolock);
+                       xfs_rw_iunlock(ip, XFS_ILOCK_EXCL|iolock);
                        return XFS_ERROR(-EINVAL);
                }
 
-               if (!need_i_mutex && (mapping->nrpages || pos > ip->i_size)) {
-                       xfs_iunlock(ip, XFS_ILOCK_EXCL|iolock);
+               /*
+                * For direct I/O, if there are cached pages or we're extending
+                * the file, we need IOLOCK_EXCL until we're sure the bytes at
+                * the new EOF have been zeroed and/or the cached pages are
+                * flushed out.  Upgrade the I/O lock and start again.
+                */
+               if (iolock != XFS_IOLOCK_EXCL &&
+                   (mapping->nrpages || pos > ip->i_size)) {
+                       xfs_rw_iunlock(ip, XFS_ILOCK_EXCL|iolock);
                        iolock = XFS_IOLOCK_EXCL;
-                       need_i_mutex = 1;
-                       mutex_lock(&inode->i_mutex);
-                       xfs_ilock(ip, XFS_ILOCK_EXCL|iolock);
                        goto start;
                }
        }
@@ -687,11 +727,11 @@ start:
        if (pos > ip->i_size) {
                ret = -xfs_zero_eof(ip, pos, ip->i_size);
                if (ret) {
-                       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+                       xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
                        goto out_unlock_internal;
                }
        }
-       xfs_iunlock(ip, XFS_ILOCK_EXCL);
+       xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
 
        /*
         * If we're writing the file then make sure to clear the
@@ -708,7 +748,7 @@ start:
 
        if ((ioflags & IO_ISDIRECT)) {
                if (mapping->nrpages) {
-                       WARN_ON(need_i_mutex == 0);
+                       WARN_ON(iolock != XFS_IOLOCK_EXCL);
                        ret = -xfs_flushinval_pages(ip,
                                        (pos & PAGE_CACHE_MASK),
                                        -1, FI_REMAPF_LOCKED);
@@ -716,13 +756,10 @@ start:
                                goto out_unlock_internal;
                }
 
-               if (need_i_mutex) {
+               if (iolock == XFS_IOLOCK_EXCL) {
                        /* demote the lock now the cached pages are gone */
-                       xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
-                       mutex_unlock(&inode->i_mutex);
-
+                       xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
                        iolock = XFS_IOLOCK_SHARED;
-                       need_i_mutex = 0;
                }
 
                trace_xfs_file_direct_write(ip, count, iocb->ki_pos, ioflags);
@@ -740,7 +777,7 @@ start:
                        count -= ret;
 
                        ioflags &= ~IO_ISDIRECT;
-                       xfs_iunlock(ip, iolock);
+                       xfs_rw_iunlock(ip, iolock);
                        goto relock;
                }
        } else {
@@ -775,14 +812,9 @@ write_retry:
                loff_t end = pos + ret - 1;
                int error, error2;
 
-               xfs_iunlock(ip, iolock);
-               if (need_i_mutex)
-                       mutex_unlock(&inode->i_mutex);
-
+               xfs_rw_iunlock(ip, iolock);
                error = filemap_write_and_wait_range(mapping, pos, end);
-               if (need_i_mutex)
-                       mutex_lock(&inode->i_mutex);
-               xfs_ilock(ip, iolock);
+               xfs_rw_ilock(ip, iolock);
 
                error2 = -xfs_file_fsync(file,
                                         (file->f_flags & __O_SYNC) ? 0 : 1);
@@ -794,10 +826,7 @@ write_retry:
 
  out_unlock_internal:
        xfs_aio_write_newsize_update(ip);
-       xfs_iunlock(ip, iolock);
- out_unlock_mutex:
-       if (need_i_mutex)
-               mutex_unlock(&inode->i_mutex);
+       xfs_rw_iunlock(ip, iolock);
        return ret;
 }