diff options
-rw-r--r-- | fs/xfs/xfs_vfsops.c | 198 |
1 files changed, 23 insertions, 175 deletions
diff --git a/fs/xfs/xfs_vfsops.c b/fs/xfs/xfs_vfsops.c index a1544597bcd3..0f237f9f0c3a 100644 --- a/fs/xfs/xfs_vfsops.c +++ b/fs/xfs/xfs_vfsops.c | |||
@@ -981,8 +981,6 @@ xfs_sync_inodes( | |||
981 | int *bypassed) | 981 | int *bypassed) |
982 | { | 982 | { |
983 | xfs_inode_t *ip = NULL; | 983 | xfs_inode_t *ip = NULL; |
984 | xfs_inode_t *ip_next; | ||
985 | xfs_buf_t *bp; | ||
986 | bhv_vnode_t *vp = NULL; | 984 | bhv_vnode_t *vp = NULL; |
987 | int error; | 985 | int error; |
988 | int last_error; | 986 | int last_error; |
@@ -992,7 +990,6 @@ xfs_sync_inodes( | |||
992 | boolean_t mount_locked; | 990 | boolean_t mount_locked; |
993 | boolean_t vnode_refed; | 991 | boolean_t vnode_refed; |
994 | int preempt; | 992 | int preempt; |
995 | xfs_dinode_t *dip; | ||
996 | xfs_iptr_t *ipointer; | 993 | xfs_iptr_t *ipointer; |
997 | #ifdef DEBUG | 994 | #ifdef DEBUG |
998 | boolean_t ipointer_in = B_FALSE; | 995 | boolean_t ipointer_in = B_FALSE; |
@@ -1045,6 +1042,8 @@ xfs_sync_inodes( | |||
1045 | 1042 | ||
1046 | #define XFS_PREEMPT_MASK 0x7f | 1043 | #define XFS_PREEMPT_MASK 0x7f |
1047 | 1044 | ||
1045 | ASSERT(!(flags & SYNC_BDFLUSH)); | ||
1046 | |||
1048 | if (bypassed) | 1047 | if (bypassed) |
1049 | *bypassed = 0; | 1048 | *bypassed = 0; |
1050 | if (mp->m_flags & XFS_MOUNT_RDONLY) | 1049 | if (mp->m_flags & XFS_MOUNT_RDONLY) |
@@ -1057,7 +1056,7 @@ xfs_sync_inodes( | |||
1057 | ipointer = (xfs_iptr_t *)kmem_zalloc(sizeof(xfs_iptr_t), KM_SLEEP); | 1056 | ipointer = (xfs_iptr_t *)kmem_zalloc(sizeof(xfs_iptr_t), KM_SLEEP); |
1058 | 1057 | ||
1059 | fflag = XFS_B_ASYNC; /* default is don't wait */ | 1058 | fflag = XFS_B_ASYNC; /* default is don't wait */ |
1060 | if (flags & (SYNC_BDFLUSH | SYNC_DELWRI)) | 1059 | if (flags & SYNC_DELWRI) |
1061 | fflag = XFS_B_DELWRI; | 1060 | fflag = XFS_B_DELWRI; |
1062 | if (flags & SYNC_WAIT) | 1061 | if (flags & SYNC_WAIT) |
1063 | fflag = 0; /* synchronous overrides all */ | 1062 | fflag = 0; /* synchronous overrides all */ |
@@ -1147,24 +1146,6 @@ xfs_sync_inodes( | |||
1147 | } | 1146 | } |
1148 | 1147 | ||
1149 | /* | 1148 | /* |
1150 | * If this is just vfs_sync() or pflushd() calling | ||
1151 | * then we can skip inodes for which it looks like | ||
1152 | * there is nothing to do. Since we don't have the | ||
1153 | * inode locked this is racy, but these are periodic | ||
1154 | * calls so it doesn't matter. For the others we want | ||
1155 | * to know for sure, so we at least try to lock them. | ||
1156 | */ | ||
1157 | if (flags & SYNC_BDFLUSH) { | ||
1158 | if (((ip->i_itemp == NULL) || | ||
1159 | !(ip->i_itemp->ili_format.ilf_fields & | ||
1160 | XFS_ILOG_ALL)) && | ||
1161 | (ip->i_update_core == 0)) { | ||
1162 | ip = ip->i_mnext; | ||
1163 | continue; | ||
1164 | } | ||
1165 | } | ||
1166 | |||
1167 | /* | ||
1168 | * Try to lock without sleeping. We're out of order with | 1149 | * Try to lock without sleeping. We're out of order with |
1169 | * the inode list lock here, so if we fail we need to drop | 1150 | * the inode list lock here, so if we fail we need to drop |
1170 | * the mount lock and try again. If we're called from | 1151 | * the mount lock and try again. If we're called from |
@@ -1181,7 +1162,7 @@ xfs_sync_inodes( | |||
1181 | * it. | 1162 | * it. |
1182 | */ | 1163 | */ |
1183 | if (xfs_ilock_nowait(ip, lock_flags) == 0) { | 1164 | if (xfs_ilock_nowait(ip, lock_flags) == 0) { |
1184 | if ((flags & SYNC_BDFLUSH) || (vp == NULL)) { | 1165 | if (vp == NULL) { |
1185 | ip = ip->i_mnext; | 1166 | ip = ip->i_mnext; |
1186 | continue; | 1167 | continue; |
1187 | } | 1168 | } |
@@ -1242,160 +1223,27 @@ xfs_sync_inodes( | |||
1242 | xfs_ilock(ip, XFS_ILOCK_SHARED); | 1223 | xfs_ilock(ip, XFS_ILOCK_SHARED); |
1243 | } | 1224 | } |
1244 | 1225 | ||
1245 | if (flags & SYNC_BDFLUSH) { | 1226 | if ((flags & SYNC_ATTR) && |
1246 | if ((flags & SYNC_ATTR) && | 1227 | (ip->i_update_core || |
1247 | ((ip->i_update_core) || | 1228 | (ip->i_itemp && ip->i_itemp->ili_format.ilf_fields))) { |
1248 | ((ip->i_itemp != NULL) && | 1229 | if (mount_locked) |
1249 | (ip->i_itemp->ili_format.ilf_fields != 0)))) { | 1230 | IPOINTER_INSERT(ip, mp); |
1250 | |||
1251 | /* Insert marker and drop lock if not already | ||
1252 | * done. | ||
1253 | */ | ||
1254 | if (mount_locked) { | ||
1255 | IPOINTER_INSERT(ip, mp); | ||
1256 | } | ||
1257 | |||
1258 | /* | ||
1259 | * We don't want the periodic flushing of the | ||
1260 | * inodes by vfs_sync() to interfere with | ||
1261 | * I/O to the file, especially read I/O | ||
1262 | * where it is only the access time stamp | ||
1263 | * that is being flushed out. To prevent | ||
1264 | * long periods where we have both inode | ||
1265 | * locks held shared here while reading the | ||
1266 | * inode's buffer in from disk, we drop the | ||
1267 | * inode lock while reading in the inode | ||
1268 | * buffer. We have to release the buffer | ||
1269 | * and reacquire the inode lock so that they | ||
1270 | * are acquired in the proper order (inode | ||
1271 | * locks first). The buffer will go at the | ||
1272 | * end of the lru chain, though, so we can | ||
1273 | * expect it to still be there when we go | ||
1274 | * for it again in xfs_iflush(). | ||
1275 | */ | ||
1276 | if ((xfs_ipincount(ip) == 0) && | ||
1277 | xfs_iflock_nowait(ip)) { | ||
1278 | |||
1279 | xfs_ifunlock(ip); | ||
1280 | xfs_iunlock(ip, XFS_ILOCK_SHARED); | ||
1281 | |||
1282 | error = xfs_itobp(mp, NULL, ip, | ||
1283 | &dip, &bp, 0, 0); | ||
1284 | if (!error) { | ||
1285 | xfs_buf_relse(bp); | ||
1286 | } else { | ||
1287 | /* Bailing out, remove the | ||
1288 | * marker and free it. | ||
1289 | */ | ||
1290 | XFS_MOUNT_ILOCK(mp); | ||
1291 | IPOINTER_REMOVE(ip, mp); | ||
1292 | XFS_MOUNT_IUNLOCK(mp); | ||
1293 | |||
1294 | ASSERT(!(lock_flags & | ||
1295 | XFS_IOLOCK_SHARED)); | ||
1296 | |||
1297 | kmem_free(ipointer, | ||
1298 | sizeof(xfs_iptr_t)); | ||
1299 | return (0); | ||
1300 | } | ||
1301 | |||
1302 | /* | ||
1303 | * Since we dropped the inode lock, | ||
1304 | * the inode may have been reclaimed. | ||
1305 | * Therefore, we reacquire the mount | ||
1306 | * lock and check to see if we were the | ||
1307 | * inode reclaimed. If this happened | ||
1308 | * then the ipointer marker will no | ||
1309 | * longer point back at us. In this | ||
1310 | * case, move ip along to the inode | ||
1311 | * after the marker, remove the marker | ||
1312 | * and continue. | ||
1313 | */ | ||
1314 | XFS_MOUNT_ILOCK(mp); | ||
1315 | mount_locked = B_TRUE; | ||
1316 | |||
1317 | if (ip != ipointer->ip_mprev) { | ||
1318 | IPOINTER_REMOVE(ip, mp); | ||
1319 | |||
1320 | ASSERT(!vnode_refed); | ||
1321 | ASSERT(!(lock_flags & | ||
1322 | XFS_IOLOCK_SHARED)); | ||
1323 | continue; | ||
1324 | } | ||
1325 | |||
1326 | ASSERT(ip->i_mount == mp); | ||
1327 | |||
1328 | if (xfs_ilock_nowait(ip, | ||
1329 | XFS_ILOCK_SHARED) == 0) { | ||
1330 | ASSERT(ip->i_mount == mp); | ||
1331 | /* | ||
1332 | * We failed to reacquire | ||
1333 | * the inode lock without | ||
1334 | * sleeping, so just skip | ||
1335 | * the inode for now. We | ||
1336 | * clear the ILOCK bit from | ||
1337 | * the lock_flags so that we | ||
1338 | * won't try to drop a lock | ||
1339 | * we don't hold below. | ||
1340 | */ | ||
1341 | lock_flags &= ~XFS_ILOCK_SHARED; | ||
1342 | IPOINTER_REMOVE(ip_next, mp); | ||
1343 | } else if ((xfs_ipincount(ip) == 0) && | ||
1344 | xfs_iflock_nowait(ip)) { | ||
1345 | ASSERT(ip->i_mount == mp); | ||
1346 | /* | ||
1347 | * Since this is vfs_sync() | ||
1348 | * calling we only flush the | ||
1349 | * inode out if we can lock | ||
1350 | * it without sleeping and | ||
1351 | * it is not pinned. Drop | ||
1352 | * the mount lock here so | ||
1353 | * that we don't hold it for | ||
1354 | * too long. We already have | ||
1355 | * a marker in the list here. | ||
1356 | */ | ||
1357 | XFS_MOUNT_IUNLOCK(mp); | ||
1358 | mount_locked = B_FALSE; | ||
1359 | error = xfs_iflush(ip, | ||
1360 | XFS_IFLUSH_DELWRI); | ||
1361 | } else { | ||
1362 | ASSERT(ip->i_mount == mp); | ||
1363 | IPOINTER_REMOVE(ip_next, mp); | ||
1364 | } | ||
1365 | } | ||
1366 | |||
1367 | } | ||
1368 | 1231 | ||
1369 | } else { | 1232 | if (flags & SYNC_WAIT) { |
1370 | if ((flags & SYNC_ATTR) && | 1233 | xfs_iflock(ip); |
1371 | ((ip->i_update_core) || | 1234 | error = xfs_iflush(ip, XFS_IFLUSH_SYNC); |
1372 | ((ip->i_itemp != NULL) && | ||
1373 | (ip->i_itemp->ili_format.ilf_fields != 0)))) { | ||
1374 | if (mount_locked) { | ||
1375 | IPOINTER_INSERT(ip, mp); | ||
1376 | } | ||
1377 | 1235 | ||
1378 | if (flags & SYNC_WAIT) { | 1236 | /* |
1379 | xfs_iflock(ip); | 1237 | * If we can't acquire the flush lock, then the inode |
1380 | error = xfs_iflush(ip, | 1238 | * is already being flushed so don't bother waiting. |
1381 | XFS_IFLUSH_SYNC); | 1239 | * |
1382 | } else { | 1240 | * If we can lock it then do a delwri flush so we can |
1383 | /* | 1241 | * combine multiple inode flushes in each disk write. |
1384 | * If we can't acquire the flush | 1242 | */ |
1385 | * lock, then the inode is already | 1243 | } else if (xfs_iflock_nowait(ip)) { |
1386 | * being flushed so don't bother | 1244 | error = xfs_iflush(ip, XFS_IFLUSH_DELWRI); |
1387 | * waiting. If we can lock it then | 1245 | } else if (bypassed) { |
1388 | * do a delwri flush so we can | 1246 | (*bypassed)++; |
1389 | * combine multiple inode flushes | ||
1390 | * in each disk write. | ||
1391 | */ | ||
1392 | if (xfs_iflock_nowait(ip)) { | ||
1393 | error = xfs_iflush(ip, | ||
1394 | XFS_IFLUSH_DELWRI); | ||
1395 | } | ||
1396 | else if (bypassed) | ||
1397 | (*bypassed)++; | ||
1398 | } | ||
1399 | } | 1247 | } |
1400 | } | 1248 | } |
1401 | 1249 | ||