xfs: use xfs_get_buf_noaddr for iclogs

Currently xlog_alloc allocates memory for the iclogs first, then allocates a buffer using xfs_buf_get_empty and finally assigns the memory to the buffer. We don't really want to do this, but rather allocate a buffer with memory attached to it using xfs_buf_get_noaddr. There's a subtile change because xfs_buf_get_empty returns the buffer locked, but xfs_buf_get_noaddr returns it unlocked. From my auditing and testing nothing in the log I/O code cares about this distincition, but I'd be happy if someone could try to prove this independently.

Signed-off-by Christoph Hellwig <[email protected]>

Index: linux-2.6/fs/xfs/xfs_log.c
===================================================================

fs/xfs/xfs_log.c +8 -10

@@ -1199,11 +1199,16 @@
 		*iclogp = (xlog_in_core_t *)
 			  kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
 		iclog = *iclogp;
-		iclog->hic_data = (xlog_in_core_2_t *)
-			  kmem_zalloc(iclogsize, KM_SLEEP | KM_LARGE);
-
 		iclog->ic_prev = prev_iclog;
 		prev_iclog = iclog;
+
+		bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
+		XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
+		XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
+		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
+		iclog->ic_bp = bp;
+		iclog->hic_data = bp->b_addr;
+
 		log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
 
 		head = &iclog->ic_header;

@@ -1216,11 +1221,6 @@
 		INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
 
-		bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
-		XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
-		XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
-		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
-		iclog->ic_bp = bp;
 
 		iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
 		iclog->ic_state = XLOG_STATE_ACTIVE;

@@ -1229,7 +1229,6 @@
 		iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
 
 		ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
-		ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
 		sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
 		sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
 

@@ -1528,7 +1527,6 @@
 		}
 #endif
 		next_iclog = iclog->ic_next;
-		kmem_free(iclog->hic_data, log->l_iclog_size);
 		kmem_free(iclog, sizeof(xlog_in_core_t));
 		iclog = next_iclog;
 	}

xfs_buf_get_noaddr. There's a subtile change because xfs_buf_get_empty returns the buffer locked, but xfs_buf_get_noaddr returns it unlocked. From my auditing and testing nothing in the log I/O code cares about this distincition, but I'd be happy if someone could try to prove this independently.

fs/xfs/xfs_log.c +0 -5

@@ -1216,11 +1221,6 @@
 		INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
 
-		bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
-		XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
-		XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
-		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
-		iclog->ic_bp = bp;
 
 		iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
 		iclog->ic_state = XLOG_STATE_ACTIVE;

OTOH, all other buffers are supposed to be locked when under I/O. This change makes a special case for the log buffers, and I'd prefer not to have to remember that this behaviour changed fo log buffers at some point in time.

I suggest that adding:

(quoted diff) +1 -0

+		XFS_BUF_PSEMA(bp, PRIBIO);

To lock the buffer should be added here. That way we don't change any semantics of the code at all.