lfs_vnops.c revision 1.137.2.25
1 /* $NetBSD: lfs_vnops.c,v 1.137.2.25 2006/08/10 12:16:47 tron Exp $ */ 2 3 /*- 4 * Copyright (c) 1999, 2000, 2001, 2002, 2003 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Konrad E. Schroder <perseant (at) hhhh.org>. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the NetBSD 21 * Foundation, Inc. and its contributors. 22 * 4. Neither the name of The NetBSD Foundation nor the names of its 23 * contributors may be used to endorse or promote products derived 24 * from this software without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 36 * POSSIBILITY OF SUCH DAMAGE. 37 */ 38 /* 39 * Copyright (c) 1986, 1989, 1991, 1993, 1995 40 * The Regents of the University of California. All rights reserved. 41 * 42 * Redistribution and use in source and binary forms, with or without 43 * modification, are permitted provided that the following conditions 44 * are met: 45 * 1. Redistributions of source code must retain the above copyright 46 * notice, this list of conditions and the following disclaimer. 47 * 2. Redistributions in binary form must reproduce the above copyright 48 * notice, this list of conditions and the following disclaimer in the 49 * documentation and/or other materials provided with the distribution. 50 * 3. Neither the name of the University nor the names of its contributors 51 * may be used to endorse or promote products derived from this software 52 * without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 64 * SUCH DAMAGE. 65 * 66 * @(#)lfs_vnops.c 8.13 (Berkeley) 6/10/95 67 */ 68 69 #include <sys/cdefs.h> 70 __KERNEL_RCSID(0, "$NetBSD: lfs_vnops.c,v 1.137.2.25 2006/08/10 12:16:47 tron Exp $"); 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/namei.h> 75 #include <sys/resourcevar.h> 76 #include <sys/kernel.h> 77 #include <sys/file.h> 78 #include <sys/stat.h> 79 #include <sys/buf.h> 80 #include <sys/proc.h> 81 #include <sys/mount.h> 82 #include <sys/vnode.h> 83 #include <sys/pool.h> 84 #include <sys/signalvar.h> 85 86 #include <miscfs/fifofs/fifo.h> 87 #include <miscfs/genfs/genfs.h> 88 #include <miscfs/specfs/specdev.h> 89 90 #include <ufs/ufs/inode.h> 91 #include <ufs/ufs/dir.h> 92 #include <ufs/ufs/ufsmount.h> 93 #include <ufs/ufs/ufs_extern.h> 94 95 #include <uvm/uvm.h> 96 #include <uvm/uvm_pmap.h> 97 #include <uvm/uvm_stat.h> 98 #include <uvm/uvm_pager.h> 99 100 #include <ufs/lfs/lfs.h> 101 #include <ufs/lfs/lfs_extern.h> 102 103 extern pid_t lfs_writer_daemon; 104 105 /* Global vfs data structures for lfs. */ 106 int (**lfs_vnodeop_p)(void *); 107 const struct vnodeopv_entry_desc lfs_vnodeop_entries[] = { 108 { &vop_default_desc, vn_default_error }, 109 { &vop_lookup_desc, ufs_lookup }, /* lookup */ 110 { &vop_create_desc, lfs_create }, /* create */ 111 { &vop_whiteout_desc, ufs_whiteout }, /* whiteout */ 112 { &vop_mknod_desc, lfs_mknod }, /* mknod */ 113 { &vop_open_desc, ufs_open }, /* open */ 114 { &vop_close_desc, lfs_close }, /* close */ 115 { &vop_access_desc, ufs_access }, /* access */ 116 { &vop_getattr_desc, lfs_getattr }, /* getattr */ 117 { &vop_setattr_desc, lfs_setattr }, /* setattr */ 118 { &vop_read_desc, lfs_read }, /* read */ 119 { &vop_write_desc, lfs_write }, /* write */ 120 { &vop_lease_desc, ufs_lease_check }, /* lease */ 121 { &vop_ioctl_desc, ufs_ioctl }, /* ioctl */ 122 { &vop_fcntl_desc, lfs_fcntl }, /* fcntl */ 123 { &vop_poll_desc, ufs_poll }, /* poll */ 124 { &vop_kqfilter_desc, genfs_kqfilter }, /* kqfilter */ 125 { &vop_revoke_desc, ufs_revoke }, /* revoke */ 126 { &vop_mmap_desc, lfs_mmap }, /* mmap */ 127 { &vop_fsync_desc, lfs_fsync }, /* fsync */ 128 { &vop_seek_desc, ufs_seek }, /* seek */ 129 { &vop_remove_desc, lfs_remove }, /* remove */ 130 { &vop_link_desc, lfs_link }, /* link */ 131 { &vop_rename_desc, lfs_rename }, /* rename */ 132 { &vop_mkdir_desc, lfs_mkdir }, /* mkdir */ 133 { &vop_rmdir_desc, lfs_rmdir }, /* rmdir */ 134 { &vop_symlink_desc, lfs_symlink }, /* symlink */ 135 { &vop_readdir_desc, ufs_readdir }, /* readdir */ 136 { &vop_readlink_desc, ufs_readlink }, /* readlink */ 137 { &vop_abortop_desc, ufs_abortop }, /* abortop */ 138 { &vop_inactive_desc, lfs_inactive }, /* inactive */ 139 { &vop_reclaim_desc, lfs_reclaim }, /* reclaim */ 140 { &vop_lock_desc, ufs_lock }, /* lock */ 141 { &vop_unlock_desc, ufs_unlock }, /* unlock */ 142 { &vop_bmap_desc, ufs_bmap }, /* bmap */ 143 { &vop_strategy_desc, lfs_strategy }, /* strategy */ 144 { &vop_print_desc, ufs_print }, /* print */ 145 { &vop_islocked_desc, ufs_islocked }, /* islocked */ 146 { &vop_pathconf_desc, ufs_pathconf }, /* pathconf */ 147 { &vop_advlock_desc, ufs_advlock }, /* advlock */ 148 { &vop_blkatoff_desc, lfs_blkatoff }, /* blkatoff */ 149 { &vop_valloc_desc, lfs_valloc }, /* valloc */ 150 { &vop_balloc_desc, lfs_balloc }, /* balloc */ 151 { &vop_vfree_desc, lfs_vfree }, /* vfree */ 152 { &vop_truncate_desc, lfs_truncate }, /* truncate */ 153 { &vop_update_desc, lfs_update }, /* update */ 154 { &vop_bwrite_desc, lfs_bwrite }, /* bwrite */ 155 { &vop_getpages_desc, lfs_getpages }, /* getpages */ 156 { &vop_putpages_desc, lfs_putpages }, /* putpages */ 157 { NULL, NULL } 158 }; 159 const struct vnodeopv_desc lfs_vnodeop_opv_desc = 160 { &lfs_vnodeop_p, lfs_vnodeop_entries }; 161 162 int (**lfs_specop_p)(void *); 163 const struct vnodeopv_entry_desc lfs_specop_entries[] = { 164 { &vop_default_desc, vn_default_error }, 165 { &vop_lookup_desc, spec_lookup }, /* lookup */ 166 { &vop_create_desc, spec_create }, /* create */ 167 { &vop_mknod_desc, spec_mknod }, /* mknod */ 168 { &vop_open_desc, spec_open }, /* open */ 169 { &vop_close_desc, lfsspec_close }, /* close */ 170 { &vop_access_desc, ufs_access }, /* access */ 171 { &vop_getattr_desc, lfs_getattr }, /* getattr */ 172 { &vop_setattr_desc, lfs_setattr }, /* setattr */ 173 { &vop_read_desc, ufsspec_read }, /* read */ 174 { &vop_write_desc, ufsspec_write }, /* write */ 175 { &vop_lease_desc, spec_lease_check }, /* lease */ 176 { &vop_ioctl_desc, spec_ioctl }, /* ioctl */ 177 { &vop_fcntl_desc, ufs_fcntl }, /* fcntl */ 178 { &vop_poll_desc, spec_poll }, /* poll */ 179 { &vop_kqfilter_desc, spec_kqfilter }, /* kqfilter */ 180 { &vop_revoke_desc, spec_revoke }, /* revoke */ 181 { &vop_mmap_desc, spec_mmap }, /* mmap */ 182 { &vop_fsync_desc, spec_fsync }, /* fsync */ 183 { &vop_seek_desc, spec_seek }, /* seek */ 184 { &vop_remove_desc, spec_remove }, /* remove */ 185 { &vop_link_desc, spec_link }, /* link */ 186 { &vop_rename_desc, spec_rename }, /* rename */ 187 { &vop_mkdir_desc, spec_mkdir }, /* mkdir */ 188 { &vop_rmdir_desc, spec_rmdir }, /* rmdir */ 189 { &vop_symlink_desc, spec_symlink }, /* symlink */ 190 { &vop_readdir_desc, spec_readdir }, /* readdir */ 191 { &vop_readlink_desc, spec_readlink }, /* readlink */ 192 { &vop_abortop_desc, spec_abortop }, /* abortop */ 193 { &vop_inactive_desc, lfs_inactive }, /* inactive */ 194 { &vop_reclaim_desc, lfs_reclaim }, /* reclaim */ 195 { &vop_lock_desc, ufs_lock }, /* lock */ 196 { &vop_unlock_desc, ufs_unlock }, /* unlock */ 197 { &vop_bmap_desc, spec_bmap }, /* bmap */ 198 { &vop_strategy_desc, spec_strategy }, /* strategy */ 199 { &vop_print_desc, ufs_print }, /* print */ 200 { &vop_islocked_desc, ufs_islocked }, /* islocked */ 201 { &vop_pathconf_desc, spec_pathconf }, /* pathconf */ 202 { &vop_advlock_desc, spec_advlock }, /* advlock */ 203 { &vop_blkatoff_desc, spec_blkatoff }, /* blkatoff */ 204 { &vop_valloc_desc, spec_valloc }, /* valloc */ 205 { &vop_vfree_desc, lfs_vfree }, /* vfree */ 206 { &vop_truncate_desc, spec_truncate }, /* truncate */ 207 { &vop_update_desc, lfs_update }, /* update */ 208 { &vop_bwrite_desc, vn_bwrite }, /* bwrite */ 209 { &vop_getpages_desc, spec_getpages }, /* getpages */ 210 { &vop_putpages_desc, spec_putpages }, /* putpages */ 211 { NULL, NULL } 212 }; 213 const struct vnodeopv_desc lfs_specop_opv_desc = 214 { &lfs_specop_p, lfs_specop_entries }; 215 216 int (**lfs_fifoop_p)(void *); 217 const struct vnodeopv_entry_desc lfs_fifoop_entries[] = { 218 { &vop_default_desc, vn_default_error }, 219 { &vop_lookup_desc, fifo_lookup }, /* lookup */ 220 { &vop_create_desc, fifo_create }, /* create */ 221 { &vop_mknod_desc, fifo_mknod }, /* mknod */ 222 { &vop_open_desc, fifo_open }, /* open */ 223 { &vop_close_desc, lfsfifo_close }, /* close */ 224 { &vop_access_desc, ufs_access }, /* access */ 225 { &vop_getattr_desc, lfs_getattr }, /* getattr */ 226 { &vop_setattr_desc, lfs_setattr }, /* setattr */ 227 { &vop_read_desc, ufsfifo_read }, /* read */ 228 { &vop_write_desc, ufsfifo_write }, /* write */ 229 { &vop_lease_desc, fifo_lease_check }, /* lease */ 230 { &vop_ioctl_desc, fifo_ioctl }, /* ioctl */ 231 { &vop_fcntl_desc, ufs_fcntl }, /* fcntl */ 232 { &vop_poll_desc, fifo_poll }, /* poll */ 233 { &vop_kqfilter_desc, fifo_kqfilter }, /* kqfilter */ 234 { &vop_revoke_desc, fifo_revoke }, /* revoke */ 235 { &vop_mmap_desc, fifo_mmap }, /* mmap */ 236 { &vop_fsync_desc, fifo_fsync }, /* fsync */ 237 { &vop_seek_desc, fifo_seek }, /* seek */ 238 { &vop_remove_desc, fifo_remove }, /* remove */ 239 { &vop_link_desc, fifo_link }, /* link */ 240 { &vop_rename_desc, fifo_rename }, /* rename */ 241 { &vop_mkdir_desc, fifo_mkdir }, /* mkdir */ 242 { &vop_rmdir_desc, fifo_rmdir }, /* rmdir */ 243 { &vop_symlink_desc, fifo_symlink }, /* symlink */ 244 { &vop_readdir_desc, fifo_readdir }, /* readdir */ 245 { &vop_readlink_desc, fifo_readlink }, /* readlink */ 246 { &vop_abortop_desc, fifo_abortop }, /* abortop */ 247 { &vop_inactive_desc, lfs_inactive }, /* inactive */ 248 { &vop_reclaim_desc, lfs_reclaim }, /* reclaim */ 249 { &vop_lock_desc, ufs_lock }, /* lock */ 250 { &vop_unlock_desc, ufs_unlock }, /* unlock */ 251 { &vop_bmap_desc, fifo_bmap }, /* bmap */ 252 { &vop_strategy_desc, fifo_strategy }, /* strategy */ 253 { &vop_print_desc, ufs_print }, /* print */ 254 { &vop_islocked_desc, ufs_islocked }, /* islocked */ 255 { &vop_pathconf_desc, fifo_pathconf }, /* pathconf */ 256 { &vop_advlock_desc, fifo_advlock }, /* advlock */ 257 { &vop_blkatoff_desc, fifo_blkatoff }, /* blkatoff */ 258 { &vop_valloc_desc, fifo_valloc }, /* valloc */ 259 { &vop_vfree_desc, lfs_vfree }, /* vfree */ 260 { &vop_truncate_desc, fifo_truncate }, /* truncate */ 261 { &vop_update_desc, lfs_update }, /* update */ 262 { &vop_bwrite_desc, lfs_bwrite }, /* bwrite */ 263 { &vop_putpages_desc, fifo_putpages }, /* putpages */ 264 { NULL, NULL } 265 }; 266 const struct vnodeopv_desc lfs_fifoop_opv_desc = 267 { &lfs_fifoop_p, lfs_fifoop_entries }; 268 269 static int check_dirty(struct lfs *, struct vnode *, off_t, off_t, off_t, int, int); 270 271 /* 272 * A function version of LFS_ITIMES, for the UFS functions which call ITIMES 273 */ 274 void 275 lfs_itimes(struct inode *ip, struct timespec *acc, struct timespec *mod, struct timespec *cre) 276 { 277 LFS_ITIMES(ip, acc, mod, cre); 278 } 279 280 #define LFS_READWRITE 281 #include <ufs/ufs/ufs_readwrite.c> 282 #undef LFS_READWRITE 283 284 /* 285 * Synch an open file. 286 */ 287 /* ARGSUSED */ 288 int 289 lfs_fsync(void *v) 290 { 291 struct vop_fsync_args /* { 292 struct vnode *a_vp; 293 struct ucred *a_cred; 294 int a_flags; 295 off_t offlo; 296 off_t offhi; 297 struct proc *a_p; 298 } */ *ap = v; 299 struct vnode *vp = ap->a_vp; 300 int error, wait; 301 302 /* If we're mounted read-only, don't try to sync. */ 303 if (VTOI(vp)->i_lfs->lfs_ronly) 304 return 0; 305 306 /* 307 * Trickle sync checks for need to do a checkpoint after possible 308 * activity from the pagedaemon. 309 */ 310 if (ap->a_flags & FSYNC_LAZY) { 311 simple_lock(&lfs_subsys_lock); 312 wakeup(&lfs_writer_daemon); 313 simple_unlock(&lfs_subsys_lock); 314 return 0; 315 } 316 317 /* 318 * Don't reclaim any vnodes that are being cleaned. 319 * This prevents the cleaner from writing files twice 320 * in the same partial segment, causing an accounting 321 * underflow. 322 */ 323 if (ap->a_flags & FSYNC_RECLAIM) { 324 if (VTOI(vp)->i_flags & IN_CLEANING) 325 return EAGAIN; 326 } 327 328 wait = (ap->a_flags & FSYNC_WAIT); 329 simple_lock(&vp->v_interlock); 330 error = VOP_PUTPAGES(vp, trunc_page(ap->a_offlo), 331 round_page(ap->a_offhi), 332 PGO_CLEANIT | (wait ? PGO_SYNCIO : 0)); 333 if (error) 334 return error; 335 error = VOP_UPDATE(vp, NULL, NULL, wait ? UPDATE_WAIT : 0); 336 if (error == 0 && ap->a_flags & FSYNC_CACHE) { 337 int l = 0; 338 error = VOP_IOCTL(VTOI(vp)->i_devvp, DIOCCACHESYNC, &l, FWRITE, 339 ap->a_p->p_ucred, ap->a_p); 340 } 341 if (wait && !VPISEMPTY(vp)) 342 LFS_SET_UINO(VTOI(vp), IN_MODIFIED); 343 344 return error; 345 } 346 347 /* 348 * Take IN_ADIROP off, then call ufs_inactive. 349 */ 350 int 351 lfs_inactive(void *v) 352 { 353 struct vop_inactive_args /* { 354 struct vnode *a_vp; 355 struct proc *a_p; 356 } */ *ap = v; 357 358 KASSERT(VTOI(ap->a_vp)->i_nlink == VTOI(ap->a_vp)->i_ffs_effnlink); 359 360 lfs_unmark_vnode(ap->a_vp); 361 362 /* 363 * The Ifile is only ever inactivated on unmount. 364 * Streamline this process by not giving it more dirty blocks. 365 */ 366 if (VTOI(ap->a_vp)->i_number == LFS_IFILE_INUM) { 367 LFS_CLR_UINO(VTOI(ap->a_vp), IN_ALLMOD); 368 VOP_UNLOCK(ap->a_vp, 0); 369 return 0; 370 } 371 372 return ufs_inactive(v); 373 } 374 375 /* 376 * These macros are used to bracket UFS directory ops, so that we can 377 * identify all the pages touched during directory ops which need to 378 * be ordered and flushed atomically, so that they may be recovered. 379 * 380 * Because we have to mark nodes VDIROP in order to prevent 381 * the cache from reclaiming them while a dirop is in progress, we must 382 * also manage the number of nodes so marked (otherwise we can run out). 383 * We do this by setting lfs_dirvcount to the number of marked vnodes; it 384 * is decremented during segment write, when VDIROP is taken off. 385 */ 386 #define MARK_VNODE(vp) lfs_mark_vnode(vp) 387 #define UNMARK_VNODE(vp) lfs_unmark_vnode(vp) 388 #define SET_DIROP_CREATE(dvp, vpp) lfs_set_dirop_create((dvp), (vpp)) 389 #define SET_DIROP_REMOVE(dvp, vp) lfs_set_dirop((dvp), (vp)) 390 static int lfs_set_dirop_create(struct vnode *, struct vnode **); 391 static int lfs_set_dirop(struct vnode *, struct vnode *); 392 393 static int 394 lfs_set_dirop(struct vnode *dvp, struct vnode *vp) 395 { 396 struct lfs *fs; 397 int error; 398 399 KASSERT(VOP_ISLOCKED(dvp)); 400 KASSERT(vp == NULL || VOP_ISLOCKED(vp)); 401 402 fs = VTOI(dvp)->i_lfs; 403 404 ASSERT_NO_SEGLOCK(fs); 405 /* 406 * LFS_NRESERVE calculates direct and indirect blocks as well 407 * as an inode block; an overestimate in most cases. 408 */ 409 if ((error = lfs_reserve(fs, dvp, vp, LFS_NRESERVE(fs))) != 0) 410 return (error); 411 412 restart: 413 simple_lock(&fs->lfs_interlock); 414 if (fs->lfs_dirops == 0) { 415 simple_unlock(&fs->lfs_interlock); 416 lfs_check(dvp, LFS_UNUSED_LBN, 0); 417 simple_lock(&fs->lfs_interlock); 418 } 419 while (fs->lfs_writer) 420 ltsleep(&fs->lfs_dirops, (PRIBIO + 1), "lfs_sdirop", 0, 421 &fs->lfs_interlock); 422 simple_lock(&lfs_subsys_lock); 423 if (lfs_dirvcount > LFS_MAX_DIROP && fs->lfs_dirops == 0) { 424 wakeup(&lfs_writer_daemon); 425 simple_unlock(&lfs_subsys_lock); 426 simple_unlock(&fs->lfs_interlock); 427 preempt(1); 428 goto restart; 429 } 430 431 if (lfs_dirvcount > LFS_MAX_DIROP) { 432 simple_unlock(&fs->lfs_interlock); 433 DLOG((DLOG_DIROP, "lfs_set_dirop: sleeping with dirops=%d, " 434 "dirvcount=%d\n", fs->lfs_dirops, lfs_dirvcount)); 435 if ((error = ltsleep(&lfs_dirvcount, 436 PCATCH | PUSER | PNORELOCK, "lfs_maxdirop", 0, 437 &lfs_subsys_lock)) != 0) { 438 goto unreserve; 439 } 440 goto restart; 441 } 442 simple_unlock(&lfs_subsys_lock); 443 444 ++fs->lfs_dirops; 445 fs->lfs_doifile = 1; 446 simple_unlock(&fs->lfs_interlock); 447 448 /* Hold a reference so SET_ENDOP will be happy */ 449 vref(dvp); 450 if (vp) { 451 vref(vp); 452 MARK_VNODE(vp); 453 } 454 455 MARK_VNODE(dvp); 456 return 0; 457 458 unreserve: 459 lfs_reserve(fs, dvp, vp, -LFS_NRESERVE(fs)); 460 return error; 461 } 462 463 /* 464 * Get a new vnode *before* adjusting the dirop count, to avoid a deadlock 465 * in getnewvnode(), if we have a stacked filesystem mounted on top 466 * of us. 467 * 468 * NB: this means we have to clear the new vnodes on error. Fortunately 469 * SET_ENDOP is there to do that for us. 470 */ 471 static int 472 lfs_set_dirop_create(struct vnode *dvp, struct vnode **vpp) 473 { 474 int error; 475 struct lfs *fs; 476 477 fs = VFSTOUFS(dvp->v_mount)->um_lfs; 478 ASSERT_NO_SEGLOCK(fs); 479 if (fs->lfs_ronly) 480 return EROFS; 481 if (vpp && (error = getnewvnode(VT_LFS, dvp->v_mount, lfs_vnodeop_p, vpp))) { 482 DLOG((DLOG_ALLOC, "lfs_set_dirop_create: dvp %p error %d\n", 483 dvp, error)); 484 return error; 485 } 486 if ((error = lfs_set_dirop(dvp, NULL)) != 0) { 487 if (vpp) { 488 ungetnewvnode(*vpp); 489 *vpp = NULL; 490 } 491 return error; 492 } 493 return 0; 494 } 495 496 #define SET_ENDOP_BASE(fs, dvp, str) \ 497 do { \ 498 simple_lock(&(fs)->lfs_interlock); \ 499 --(fs)->lfs_dirops; \ 500 if (!(fs)->lfs_dirops) { \ 501 if ((fs)->lfs_nadirop) { \ 502 panic("SET_ENDOP: %s: no dirops but " \ 503 " nadirop=%d", (str), \ 504 (fs)->lfs_nadirop); \ 505 } \ 506 wakeup(&(fs)->lfs_writer); \ 507 simple_unlock(&(fs)->lfs_interlock); \ 508 lfs_check((dvp), LFS_UNUSED_LBN, 0); \ 509 } else \ 510 simple_unlock(&(fs)->lfs_interlock); \ 511 } while(0) 512 #define SET_ENDOP_CREATE(fs, dvp, nvpp, str) \ 513 do { \ 514 UNMARK_VNODE(dvp); \ 515 if (nvpp && *nvpp) \ 516 UNMARK_VNODE(*nvpp); \ 517 /* Check for error return to stem vnode leakage */ \ 518 if (nvpp && *nvpp && !((*nvpp)->v_flag & VDIROP)) \ 519 ungetnewvnode(*(nvpp)); \ 520 SET_ENDOP_BASE((fs), (dvp), (str)); \ 521 lfs_reserve((fs), (dvp), NULL, -LFS_NRESERVE(fs)); \ 522 vrele(dvp); \ 523 } while(0) 524 #define SET_ENDOP_CREATE_AP(ap, str) \ 525 SET_ENDOP_CREATE(VTOI((ap)->a_dvp)->i_lfs, (ap)->a_dvp, \ 526 (ap)->a_vpp, (str)) 527 #define SET_ENDOP_REMOVE(fs, dvp, ovp, str) \ 528 do { \ 529 UNMARK_VNODE(dvp); \ 530 if (ovp) \ 531 UNMARK_VNODE(ovp); \ 532 SET_ENDOP_BASE((fs), (dvp), (str)); \ 533 lfs_reserve((fs), (dvp), (ovp), -LFS_NRESERVE(fs)); \ 534 vrele(dvp); \ 535 if (ovp) \ 536 vrele(ovp); \ 537 } while(0) 538 539 void 540 lfs_mark_vnode(struct vnode *vp) 541 { 542 struct inode *ip = VTOI(vp); 543 struct lfs *fs = ip->i_lfs; 544 545 simple_lock(&fs->lfs_interlock); 546 if (!(ip->i_flag & IN_ADIROP)) { 547 if (!(vp->v_flag & VDIROP)) { 548 (void)lfs_vref(vp); 549 simple_lock(&lfs_subsys_lock); 550 ++lfs_dirvcount; 551 ++fs->lfs_dirvcount; 552 simple_unlock(&lfs_subsys_lock); 553 TAILQ_INSERT_TAIL(&fs->lfs_dchainhd, ip, i_lfs_dchain); 554 vp->v_flag |= VDIROP; 555 } 556 ++fs->lfs_nadirop; 557 ip->i_flag |= IN_ADIROP; 558 } else 559 KASSERT(vp->v_flag & VDIROP); 560 simple_unlock(&fs->lfs_interlock); 561 } 562 563 void 564 lfs_unmark_vnode(struct vnode *vp) 565 { 566 struct inode *ip = VTOI(vp); 567 568 if (ip && (ip->i_flag & IN_ADIROP)) { 569 KASSERT(vp->v_flag & VDIROP); 570 simple_lock(&ip->i_lfs->lfs_interlock); 571 --ip->i_lfs->lfs_nadirop; 572 simple_unlock(&ip->i_lfs->lfs_interlock); 573 ip->i_flag &= ~IN_ADIROP; 574 } 575 } 576 577 int 578 lfs_symlink(void *v) 579 { 580 struct vop_symlink_args /* { 581 struct vnode *a_dvp; 582 struct vnode **a_vpp; 583 struct componentname *a_cnp; 584 struct vattr *a_vap; 585 char *a_target; 586 } */ *ap = v; 587 int error; 588 589 if ((error = SET_DIROP_CREATE(ap->a_dvp, ap->a_vpp)) != 0) { 590 vput(ap->a_dvp); 591 return error; 592 } 593 error = ufs_symlink(ap); 594 SET_ENDOP_CREATE_AP(ap, "symlink"); 595 return (error); 596 } 597 598 int 599 lfs_mknod(void *v) 600 { 601 struct vop_mknod_args /* { 602 struct vnode *a_dvp; 603 struct vnode **a_vpp; 604 struct componentname *a_cnp; 605 struct vattr *a_vap; 606 } */ *ap = v; 607 struct vattr *vap = ap->a_vap; 608 struct vnode **vpp = ap->a_vpp; 609 struct inode *ip; 610 int error; 611 struct mount *mp; 612 ino_t ino; 613 614 if ((error = SET_DIROP_CREATE(ap->a_dvp, ap->a_vpp)) != 0) { 615 vput(ap->a_dvp); 616 return error; 617 } 618 error = ufs_makeinode(MAKEIMODE(vap->va_type, vap->va_mode), 619 ap->a_dvp, vpp, ap->a_cnp); 620 621 /* Either way we're done with the dirop at this point */ 622 SET_ENDOP_CREATE_AP(ap, "mknod"); 623 624 if (error) 625 return (error); 626 627 ip = VTOI(*vpp); 628 mp = (*vpp)->v_mount; 629 ino = ip->i_number; 630 ip->i_flag |= IN_ACCESS | IN_CHANGE | IN_UPDATE; 631 if (vap->va_rdev != VNOVAL) { 632 /* 633 * Want to be able to use this to make badblock 634 * inodes, so don't truncate the dev number. 635 */ 636 #if 0 637 ip->i_ffs1_rdev = ufs_rw32(vap->va_rdev, 638 UFS_MPNEEDSWAP((*vpp)->v_mount)); 639 #else 640 ip->i_ffs1_rdev = vap->va_rdev; 641 #endif 642 } 643 644 /* 645 * Call fsync to write the vnode so that we don't have to deal with 646 * flushing it when it's marked VDIROP|VXLOCK. 647 * 648 * XXX KS - If we can't flush we also can't call vgone(), so must 649 * return. But, that leaves this vnode in limbo, also not good. 650 * Can this ever happen (barring hardware failure)? 651 */ 652 if ((error = VOP_FSYNC(*vpp, NOCRED, FSYNC_WAIT, 0, 0, 653 curproc)) != 0) { 654 panic("lfs_mknod: couldn't fsync (ino %llu)", 655 (unsigned long long)ino); 656 /* return (error); */ 657 } 658 /* 659 * Remove vnode so that it will be reloaded by VFS_VGET and 660 * checked to see if it is an alias of an existing entry in 661 * the inode cache. 662 */ 663 /* Used to be vput, but that causes us to call VOP_INACTIVE twice. */ 664 665 VOP_UNLOCK(*vpp, 0); 666 lfs_vunref(*vpp); 667 (*vpp)->v_type = VNON; 668 vgone(*vpp); 669 error = VFS_VGET(mp, ino, vpp); 670 671 if (error != 0) { 672 *vpp = NULL; 673 return (error); 674 } 675 return (0); 676 } 677 678 int 679 lfs_create(void *v) 680 { 681 struct vop_create_args /* { 682 struct vnode *a_dvp; 683 struct vnode **a_vpp; 684 struct componentname *a_cnp; 685 struct vattr *a_vap; 686 } */ *ap = v; 687 int error; 688 689 if ((error = SET_DIROP_CREATE(ap->a_dvp, ap->a_vpp)) != 0) { 690 vput(ap->a_dvp); 691 return error; 692 } 693 error = ufs_create(ap); 694 SET_ENDOP_CREATE_AP(ap, "create"); 695 return (error); 696 } 697 698 int 699 lfs_mkdir(void *v) 700 { 701 struct vop_mkdir_args /* { 702 struct vnode *a_dvp; 703 struct vnode **a_vpp; 704 struct componentname *a_cnp; 705 struct vattr *a_vap; 706 } */ *ap = v; 707 int error; 708 709 if ((error = SET_DIROP_CREATE(ap->a_dvp, ap->a_vpp)) != 0) { 710 vput(ap->a_dvp); 711 return error; 712 } 713 error = ufs_mkdir(ap); 714 SET_ENDOP_CREATE_AP(ap, "mkdir"); 715 return (error); 716 } 717 718 int 719 lfs_remove(void *v) 720 { 721 struct vop_remove_args /* { 722 struct vnode *a_dvp; 723 struct vnode *a_vp; 724 struct componentname *a_cnp; 725 } */ *ap = v; 726 struct vnode *dvp, *vp; 727 int error; 728 729 dvp = ap->a_dvp; 730 vp = ap->a_vp; 731 if ((error = SET_DIROP_REMOVE(dvp, vp)) != 0) { 732 if (dvp == vp) 733 vrele(vp); 734 else 735 vput(vp); 736 vput(dvp); 737 return error; 738 } 739 error = ufs_remove(ap); 740 SET_ENDOP_REMOVE(VTOI(dvp)->i_lfs, dvp, ap->a_vp, "remove"); 741 return (error); 742 } 743 744 int 745 lfs_rmdir(void *v) 746 { 747 struct vop_rmdir_args /* { 748 struct vnodeop_desc *a_desc; 749 struct vnode *a_dvp; 750 struct vnode *a_vp; 751 struct componentname *a_cnp; 752 } */ *ap = v; 753 struct vnode *vp; 754 int error; 755 756 vp = ap->a_vp; 757 if ((error = SET_DIROP_REMOVE(ap->a_dvp, ap->a_vp)) != 0) { 758 vrele(ap->a_dvp); 759 if (ap->a_vp != ap->a_dvp) 760 VOP_UNLOCK(ap->a_dvp, 0); 761 vput(vp); 762 return error; 763 } 764 error = ufs_rmdir(ap); 765 SET_ENDOP_REMOVE(VTOI(ap->a_dvp)->i_lfs, ap->a_dvp, ap->a_vp, "rmdir"); 766 return (error); 767 } 768 769 int 770 lfs_link(void *v) 771 { 772 struct vop_link_args /* { 773 struct vnode *a_dvp; 774 struct vnode *a_vp; 775 struct componentname *a_cnp; 776 } */ *ap = v; 777 int error; 778 struct vnode **vpp = NULL; 779 780 if ((error = SET_DIROP_CREATE(ap->a_dvp, vpp)) != 0) { 781 vput(ap->a_dvp); 782 return error; 783 } 784 error = ufs_link(ap); 785 SET_ENDOP_CREATE(VTOI(ap->a_dvp)->i_lfs, ap->a_dvp, vpp, "link"); 786 return (error); 787 } 788 789 int 790 lfs_rename(void *v) 791 { 792 struct vop_rename_args /* { 793 struct vnode *a_fdvp; 794 struct vnode *a_fvp; 795 struct componentname *a_fcnp; 796 struct vnode *a_tdvp; 797 struct vnode *a_tvp; 798 struct componentname *a_tcnp; 799 } */ *ap = v; 800 struct vnode *tvp, *fvp, *tdvp, *fdvp; 801 struct componentname *tcnp, *fcnp; 802 int error; 803 struct lfs *fs; 804 805 fs = VTOI(ap->a_fdvp)->i_lfs; 806 tvp = ap->a_tvp; 807 tdvp = ap->a_tdvp; 808 tcnp = ap->a_tcnp; 809 fvp = ap->a_fvp; 810 fdvp = ap->a_fdvp; 811 fcnp = ap->a_fcnp; 812 813 /* 814 * Check for cross-device rename. 815 * If it is, we don't want to set dirops, just error out. 816 * (In particular note that MARK_VNODE(tdvp) will DTWT on 817 * a cross-device rename.) 818 * 819 * Copied from ufs_rename. 820 */ 821 if ((fvp->v_mount != tdvp->v_mount) || 822 (tvp && (fvp->v_mount != tvp->v_mount))) { 823 error = EXDEV; 824 goto errout; 825 } 826 827 /* 828 * Check to make sure we're not renaming a vnode onto itself 829 * (deleting a hard link by renaming one name onto another); 830 * if we are we can't recursively call VOP_REMOVE since that 831 * would leave us with an unaccounted-for number of live dirops. 832 * 833 * Inline the relevant section of ufs_rename here, *before* 834 * calling SET_DIROP_REMOVE. 835 */ 836 if (tvp && ((VTOI(tvp)->i_flags & (IMMUTABLE | APPEND)) || 837 (VTOI(tdvp)->i_flags & APPEND))) { 838 error = EPERM; 839 goto errout; 840 } 841 if (fvp == tvp) { 842 if (fvp->v_type == VDIR) { 843 error = EINVAL; 844 goto errout; 845 } 846 847 /* Release destination completely. */ 848 VOP_ABORTOP(tdvp, tcnp); 849 vput(tdvp); 850 vput(tvp); 851 852 /* Delete source. */ 853 vrele(fvp); 854 fcnp->cn_flags &= ~(MODMASK | SAVESTART); 855 fcnp->cn_flags |= LOCKPARENT | LOCKLEAF; 856 fcnp->cn_nameiop = DELETE; 857 if ((error = relookup(fdvp, &fvp, fcnp))){ 858 /* relookup blew away fdvp */ 859 return (error); 860 } 861 return (VOP_REMOVE(fdvp, fvp, fcnp)); 862 } 863 864 if ((error = SET_DIROP_REMOVE(tdvp, tvp)) != 0) 865 goto errout; 866 MARK_VNODE(fdvp); 867 MARK_VNODE(fvp); 868 869 error = ufs_rename(ap); 870 UNMARK_VNODE(fdvp); 871 UNMARK_VNODE(fvp); 872 SET_ENDOP_REMOVE(fs, tdvp, tvp, "rename"); 873 return (error); 874 875 errout: 876 VOP_ABORTOP(tdvp, ap->a_tcnp); /* XXX, why not in NFS? */ 877 if (tdvp == tvp) 878 vrele(tdvp); 879 else 880 vput(tdvp); 881 if (tvp) 882 vput(tvp); 883 VOP_ABORTOP(fdvp, ap->a_fcnp); /* XXX, why not in NFS? */ 884 vrele(fdvp); 885 vrele(fvp); 886 return (error); 887 } 888 889 /* XXX hack to avoid calling ITIMES in getattr */ 890 int 891 lfs_getattr(void *v) 892 { 893 struct vop_getattr_args /* { 894 struct vnode *a_vp; 895 struct vattr *a_vap; 896 struct ucred *a_cred; 897 struct proc *a_p; 898 } */ *ap = v; 899 struct vnode *vp = ap->a_vp; 900 struct inode *ip = VTOI(vp); 901 struct vattr *vap = ap->a_vap; 902 struct lfs *fs = ip->i_lfs; 903 /* 904 * Copy from inode table 905 */ 906 vap->va_fsid = ip->i_dev; 907 vap->va_fileid = ip->i_number; 908 vap->va_mode = ip->i_mode & ~IFMT; 909 vap->va_nlink = ip->i_nlink; 910 vap->va_uid = ip->i_uid; 911 vap->va_gid = ip->i_gid; 912 vap->va_rdev = (dev_t)ip->i_ffs1_rdev; 913 vap->va_size = vp->v_size; 914 vap->va_atime.tv_sec = ip->i_ffs1_atime; 915 vap->va_atime.tv_nsec = ip->i_ffs1_atimensec; 916 vap->va_mtime.tv_sec = ip->i_ffs1_mtime; 917 vap->va_mtime.tv_nsec = ip->i_ffs1_mtimensec; 918 vap->va_ctime.tv_sec = ip->i_ffs1_ctime; 919 vap->va_ctime.tv_nsec = ip->i_ffs1_ctimensec; 920 vap->va_flags = ip->i_flags; 921 vap->va_gen = ip->i_gen; 922 /* this doesn't belong here */ 923 if (vp->v_type == VBLK) 924 vap->va_blocksize = BLKDEV_IOSIZE; 925 else if (vp->v_type == VCHR) 926 vap->va_blocksize = MAXBSIZE; 927 else 928 vap->va_blocksize = vp->v_mount->mnt_stat.f_iosize; 929 vap->va_bytes = fsbtob(fs, (u_quad_t)ip->i_lfs_effnblks); 930 vap->va_type = vp->v_type; 931 vap->va_filerev = ip->i_modrev; 932 return (0); 933 } 934 935 /* 936 * Check to make sure the inode blocks won't choke the buffer 937 * cache, then call ufs_setattr as usual. 938 */ 939 int 940 lfs_setattr(void *v) 941 { 942 struct vop_setattr_args /* { 943 struct vnode *a_vp; 944 struct vattr *a_vap; 945 struct ucred *a_cred; 946 struct proc *a_p; 947 } */ *ap = v; 948 struct vnode *vp = ap->a_vp; 949 950 lfs_check(vp, LFS_UNUSED_LBN, 0); 951 return ufs_setattr(v); 952 } 953 954 /* 955 * Close called 956 * 957 * XXX -- we were using ufs_close, but since it updates the 958 * times on the inode, we might need to bump the uinodes 959 * count. 960 */ 961 /* ARGSUSED */ 962 int 963 lfs_close(void *v) 964 { 965 struct vop_close_args /* { 966 struct vnode *a_vp; 967 int a_fflag; 968 struct ucred *a_cred; 969 struct proc *a_p; 970 } */ *ap = v; 971 struct vnode *vp = ap->a_vp; 972 struct inode *ip = VTOI(vp); 973 struct timespec ts; 974 975 if (vp == ip->i_lfs->lfs_ivnode && 976 vp->v_mount->mnt_iflag & IMNT_UNMOUNT) 977 return 0; 978 979 if (vp->v_usecount > 1 && vp != ip->i_lfs->lfs_ivnode) { 980 TIMEVAL_TO_TIMESPEC(&time, &ts); 981 LFS_ITIMES(ip, &ts, &ts, &ts); 982 } 983 return (0); 984 } 985 986 /* 987 * Close wrapper for special devices. 988 * 989 * Update the times on the inode then do device close. 990 */ 991 int 992 lfsspec_close(void *v) 993 { 994 struct vop_close_args /* { 995 struct vnode *a_vp; 996 int a_fflag; 997 struct ucred *a_cred; 998 struct proc *a_p; 999 } */ *ap = v; 1000 struct vnode *vp; 1001 struct inode *ip; 1002 struct timespec ts; 1003 1004 vp = ap->a_vp; 1005 ip = VTOI(vp); 1006 if (vp->v_usecount > 1) { 1007 TIMEVAL_TO_TIMESPEC(&time, &ts); 1008 LFS_ITIMES(ip, &ts, &ts, &ts); 1009 } 1010 return (VOCALL (spec_vnodeop_p, VOFFSET(vop_close), ap)); 1011 } 1012 1013 /* 1014 * Close wrapper for fifo's. 1015 * 1016 * Update the times on the inode then do device close. 1017 */ 1018 int 1019 lfsfifo_close(void *v) 1020 { 1021 struct vop_close_args /* { 1022 struct vnode *a_vp; 1023 int a_fflag; 1024 struct ucred *a_cred; 1025 struct proc *a_p; 1026 } */ *ap = v; 1027 struct vnode *vp; 1028 struct inode *ip; 1029 struct timespec ts; 1030 1031 vp = ap->a_vp; 1032 ip = VTOI(vp); 1033 if (ap->a_vp->v_usecount > 1) { 1034 TIMEVAL_TO_TIMESPEC(&time, &ts); 1035 LFS_ITIMES(ip, &ts, &ts, &ts); 1036 } 1037 return (VOCALL (fifo_vnodeop_p, VOFFSET(vop_close), ap)); 1038 } 1039 1040 /* 1041 * Reclaim an inode so that it can be used for other purposes. 1042 */ 1043 1044 int 1045 lfs_reclaim(void *v) 1046 { 1047 struct vop_reclaim_args /* { 1048 struct vnode *a_vp; 1049 struct proc *a_p; 1050 } */ *ap = v; 1051 struct vnode *vp = ap->a_vp; 1052 struct inode *ip = VTOI(vp); 1053 int error; 1054 1055 KASSERT(ip->i_nlink == ip->i_ffs_effnlink); 1056 1057 LFS_CLR_UINO(ip, IN_ALLMOD); 1058 if ((error = ufs_reclaim(vp, ap->a_p))) 1059 return (error); 1060 pool_put(&lfs_dinode_pool, ip->i_din.ffs1_din); 1061 lfs_deregister_all(vp); 1062 pool_put(&lfs_inoext_pool, ip->inode_ext.lfs); 1063 ip->inode_ext.lfs = NULL; 1064 pool_put(&lfs_inode_pool, vp->v_data); 1065 vp->v_data = NULL; 1066 return (0); 1067 } 1068 1069 /* 1070 * Read a block from a storage device. 1071 * In order to avoid reading blocks that are in the process of being 1072 * written by the cleaner---and hence are not mutexed by the normal 1073 * buffer cache / page cache mechanisms---check for collisions before 1074 * reading. 1075 * 1076 * We inline ufs_strategy to make sure that the VOP_BMAP occurs *before* 1077 * the active cleaner test. 1078 * 1079 * XXX This code assumes that lfs_markv makes synchronous checkpoints. 1080 */ 1081 int 1082 lfs_strategy(void *v) 1083 { 1084 struct vop_strategy_args /* { 1085 struct vnode *a_vp; 1086 struct buf *a_bp; 1087 } */ *ap = v; 1088 struct buf *bp; 1089 struct lfs *fs; 1090 struct vnode *vp; 1091 struct inode *ip; 1092 daddr_t tbn; 1093 int i, sn, error, slept; 1094 1095 bp = ap->a_bp; 1096 vp = ap->a_vp; 1097 ip = VTOI(vp); 1098 fs = ip->i_lfs; 1099 1100 /* lfs uses its strategy routine only for read */ 1101 KASSERT(bp->b_flags & B_READ); 1102 1103 if (vp->v_type == VBLK || vp->v_type == VCHR) 1104 panic("lfs_strategy: spec"); 1105 KASSERT(bp->b_bcount != 0); 1106 if (bp->b_blkno == bp->b_lblkno) { 1107 error = VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, 1108 NULL); 1109 if (error) { 1110 bp->b_error = error; 1111 bp->b_flags |= B_ERROR; 1112 biodone(bp); 1113 return (error); 1114 } 1115 if ((long)bp->b_blkno == -1) /* no valid data */ 1116 clrbuf(bp); 1117 } 1118 if ((long)bp->b_blkno < 0) { /* block is not on disk */ 1119 biodone(bp); 1120 return (0); 1121 } 1122 1123 slept = 1; 1124 simple_lock(&fs->lfs_interlock); 1125 while (slept && fs->lfs_seglock) { 1126 simple_unlock(&fs->lfs_interlock); 1127 /* 1128 * Look through list of intervals. 1129 * There will only be intervals to look through 1130 * if the cleaner holds the seglock. 1131 * Since the cleaner is synchronous, we can trust 1132 * the list of intervals to be current. 1133 */ 1134 tbn = dbtofsb(fs, bp->b_blkno); 1135 sn = dtosn(fs, tbn); 1136 slept = 0; 1137 for (i = 0; i < fs->lfs_cleanind; i++) { 1138 if (sn == dtosn(fs, fs->lfs_cleanint[i]) && 1139 tbn >= fs->lfs_cleanint[i]) { 1140 DLOG((DLOG_CLEAN, 1141 "lfs_strategy: ino %d lbn %" PRId64 1142 " ind %d sn %d fsb %" PRIx32 1143 " given sn %d fsb %" PRIx64 "\n", 1144 ip->i_number, bp->b_lblkno, i, 1145 dtosn(fs, fs->lfs_cleanint[i]), 1146 fs->lfs_cleanint[i], sn, tbn)); 1147 DLOG((DLOG_CLEAN, 1148 "lfs_strategy: sleeping on ino %d lbn %" 1149 PRId64 "\n", ip->i_number, bp->b_lblkno)); 1150 simple_lock(&fs->lfs_interlock); 1151 if (LFS_SEGLOCK_HELD(fs) && fs->lfs_iocount) { 1152 /* Cleaner can't wait for itself */ 1153 ltsleep(&fs->lfs_iocount, 1154 (PRIBIO + 1) | PNORELOCK, 1155 "clean2", 0, 1156 &fs->lfs_interlock); 1157 slept = 1; 1158 break; 1159 } else if (fs->lfs_seglock) { 1160 ltsleep(&fs->lfs_seglock, 1161 (PRIBIO + 1) | PNORELOCK, 1162 "clean1", 0, 1163 &fs->lfs_interlock); 1164 slept = 1; 1165 break; 1166 } 1167 simple_unlock(&fs->lfs_interlock); 1168 } 1169 } 1170 simple_lock(&fs->lfs_interlock); 1171 } 1172 simple_unlock(&fs->lfs_interlock); 1173 1174 vp = ip->i_devvp; 1175 VOP_STRATEGY(vp, bp); 1176 return (0); 1177 } 1178 1179 void 1180 lfs_flush_dirops(struct lfs *fs) 1181 { 1182 struct inode *ip, *nip; 1183 struct vnode *vp; 1184 extern int lfs_dostats; 1185 struct segment *sp; 1186 int waslocked; 1187 1188 ASSERT_MAYBE_SEGLOCK(fs); 1189 KASSERT(fs->lfs_nadirop == 0); 1190 1191 if (fs->lfs_ronly) 1192 return; 1193 1194 simple_lock(&fs->lfs_interlock); 1195 if (TAILQ_FIRST(&fs->lfs_dchainhd) == NULL) { 1196 simple_unlock(&fs->lfs_interlock); 1197 return; 1198 } else 1199 simple_unlock(&fs->lfs_interlock); 1200 1201 if (lfs_dostats) 1202 ++lfs_stats.flush_invoked; 1203 1204 /* 1205 * Inline lfs_segwrite/lfs_writevnodes, but just for dirops. 1206 * Technically this is a checkpoint (the on-disk state is valid) 1207 * even though we are leaving out all the file data. 1208 */ 1209 lfs_imtime(fs); 1210 lfs_seglock(fs, SEGM_CKP); 1211 sp = fs->lfs_sp; 1212 1213 /* 1214 * lfs_writevnodes, optimized to get dirops out of the way. 1215 * Only write dirops, and don't flush files' pages, only 1216 * blocks from the directories. 1217 * 1218 * We don't need to vref these files because they are 1219 * dirops and so hold an extra reference until the 1220 * segunlock clears them of that status. 1221 * 1222 * We don't need to check for IN_ADIROP because we know that 1223 * no dirops are active. 1224 * 1225 */ 1226 simple_lock(&fs->lfs_interlock); 1227 for (ip = TAILQ_FIRST(&fs->lfs_dchainhd); ip != NULL; ip = nip) { 1228 nip = TAILQ_NEXT(ip, i_lfs_dchain); 1229 simple_unlock(&fs->lfs_interlock); 1230 vp = ITOV(ip); 1231 1232 KASSERT((ip->i_flag & IN_ADIROP) == 0); 1233 1234 /* 1235 * All writes to directories come from dirops; all 1236 * writes to files' direct blocks go through the page 1237 * cache, which we're not touching. Reads to files 1238 * and/or directories will not be affected by writing 1239 * directory blocks inodes and file inodes. So we don't 1240 * really need to lock. If we don't lock, though, 1241 * make sure that we don't clear IN_MODIFIED 1242 * unnecessarily. 1243 */ 1244 if (vp->v_flag & VXLOCK) { 1245 simple_lock(&fs->lfs_interlock); 1246 continue; 1247 } 1248 waslocked = VOP_ISLOCKED(vp); 1249 if (vp->v_type != VREG && 1250 ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp))) { 1251 lfs_writefile(fs, sp, vp); 1252 if (!VPISEMPTY(vp) && !WRITEINPROG(vp) && 1253 !(ip->i_flag & IN_ALLMOD)) { 1254 LFS_SET_UINO(ip, IN_MODIFIED); 1255 } 1256 } 1257 (void) lfs_writeinode(fs, sp, ip); 1258 if (waslocked) 1259 LFS_SET_UINO(ip, IN_MODIFIED); 1260 simple_lock(&fs->lfs_interlock); 1261 } 1262 simple_unlock(&fs->lfs_interlock); 1263 /* We've written all the dirops there are */ 1264 ((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT); 1265 lfs_finalize_fs_seguse(fs); 1266 (void) lfs_writeseg(fs, sp); 1267 lfs_segunlock(fs); 1268 } 1269 1270 /* 1271 * Flush all vnodes for which the pagedaemon has requested pageouts. 1272 * Skip over any files that are marked VDIROP (since lfs_flush_dirop() 1273 * has just run, this would be an error). If we have to skip a vnode 1274 * for any reason, just skip it; if we have to wait for the cleaner, 1275 * abort. The writer daemon will call us again later. 1276 */ 1277 void 1278 lfs_flush_pchain(struct lfs *fs) 1279 { 1280 struct inode *ip, *nip; 1281 struct vnode *vp; 1282 extern int lfs_dostats; 1283 struct segment *sp; 1284 int error; 1285 1286 ASSERT_NO_SEGLOCK(fs); 1287 1288 if (fs->lfs_ronly) 1289 return; 1290 1291 simple_lock(&fs->lfs_interlock); 1292 if (TAILQ_FIRST(&fs->lfs_pchainhd) == NULL) { 1293 simple_unlock(&fs->lfs_interlock); 1294 return; 1295 } else 1296 simple_unlock(&fs->lfs_interlock); 1297 1298 /* Get dirops out of the way */ 1299 lfs_flush_dirops(fs); 1300 1301 if (lfs_dostats) 1302 ++lfs_stats.flush_invoked; 1303 1304 /* 1305 * Inline lfs_segwrite/lfs_writevnodes, but just for pageouts. 1306 */ 1307 lfs_imtime(fs); 1308 lfs_seglock(fs, 0); 1309 sp = fs->lfs_sp; 1310 1311 /* 1312 * lfs_writevnodes, optimized to clear pageout requests. 1313 * Only write non-dirop files that are in the pageout queue. 1314 * We're very conservative about what we write; we want to be 1315 * fast and async. 1316 */ 1317 simple_lock(&fs->lfs_interlock); 1318 top: 1319 for (ip = TAILQ_FIRST(&fs->lfs_pchainhd); ip != NULL; ip = nip) { 1320 nip = TAILQ_NEXT(ip, i_lfs_pchain); 1321 vp = ITOV(ip); 1322 1323 if (!(ip->i_flags & IN_PAGING)) 1324 goto top; 1325 1326 if (vp->v_flag & (VXLOCK|VDIROP)) 1327 continue; 1328 if (vp->v_type != VREG) 1329 continue; 1330 if (lfs_vref(vp)) 1331 continue; 1332 simple_unlock(&fs->lfs_interlock); 1333 1334 if (VOP_ISLOCKED(vp)) { 1335 lfs_vunref(vp); 1336 simple_lock(&fs->lfs_interlock); 1337 continue; 1338 } 1339 1340 error = lfs_writefile(fs, sp, vp); 1341 if (!VPISEMPTY(vp) && !WRITEINPROG(vp) && 1342 !(ip->i_flag & IN_ALLMOD)) { 1343 LFS_SET_UINO(ip, IN_MODIFIED); 1344 } 1345 (void) lfs_writeinode(fs, sp, ip); 1346 1347 lfs_vunref(vp); 1348 1349 if (error == EAGAIN) { 1350 lfs_writeseg(fs, sp); 1351 simple_lock(&fs->lfs_interlock); 1352 break; 1353 } 1354 simple_lock(&fs->lfs_interlock); 1355 } 1356 simple_unlock(&fs->lfs_interlock); 1357 (void) lfs_writeseg(fs, sp); 1358 lfs_segunlock(fs); 1359 } 1360 1361 /* 1362 * Provide a fcntl interface to sys_lfs_{segwait,bmapv,markv}. 1363 */ 1364 int 1365 lfs_fcntl(void *v) 1366 { 1367 struct vop_fcntl_args /* { 1368 struct vnode *a_vp; 1369 u_long a_command; 1370 caddr_t a_data; 1371 int a_fflag; 1372 struct ucred *a_cred; 1373 struct proc *a_p; 1374 } */ *ap = v; 1375 struct timeval *tvp; 1376 BLOCK_INFO *blkiov; 1377 CLEANERINFO *cip; 1378 SEGUSE *sup; 1379 int blkcnt, error, oclean; 1380 struct lfs_fcntl_markv blkvp; 1381 fsid_t *fsidp; 1382 struct lfs *fs; 1383 struct buf *bp; 1384 fhandle_t *fhp; 1385 daddr_t off; 1386 1387 /* Only respect LFS fcntls on fs root or Ifile */ 1388 if (VTOI(ap->a_vp)->i_number != ROOTINO && 1389 VTOI(ap->a_vp)->i_number != LFS_IFILE_INUM) { 1390 return ufs_fcntl(v); 1391 } 1392 1393 /* Avoid locking a draining lock */ 1394 if (ap->a_vp->v_mount->mnt_iflag & IMNT_UNMOUNT) { 1395 return ESHUTDOWN; 1396 } 1397 1398 fs = VTOI(ap->a_vp)->i_lfs; 1399 fsidp = &ap->a_vp->v_mount->mnt_stat.f_fsidx; 1400 1401 switch (ap->a_command) { 1402 case LFCNSEGWAITALL: 1403 case LFCNSEGWAITALL_COMPAT: 1404 fsidp = NULL; 1405 /* FALLSTHROUGH */ 1406 case LFCNSEGWAIT: 1407 case LFCNSEGWAIT_COMPAT: 1408 tvp = (struct timeval *)ap->a_data; 1409 simple_lock(&fs->lfs_interlock); 1410 ++fs->lfs_sleepers; 1411 simple_unlock(&fs->lfs_interlock); 1412 VOP_UNLOCK(ap->a_vp, 0); 1413 1414 error = lfs_segwait(fsidp, tvp); 1415 1416 VOP_LOCK(ap->a_vp, LK_EXCLUSIVE); 1417 simple_lock(&fs->lfs_interlock); 1418 if (--fs->lfs_sleepers == 0) 1419 wakeup(&fs->lfs_sleepers); 1420 simple_unlock(&fs->lfs_interlock); 1421 return error; 1422 1423 case LFCNBMAPV: 1424 case LFCNMARKV: 1425 if ((error = suser(ap->a_p->p_ucred, &ap->a_p->p_acflag)) != 0) 1426 return (error); 1427 blkvp = *(struct lfs_fcntl_markv *)ap->a_data; 1428 1429 blkcnt = blkvp.blkcnt; 1430 if ((u_int) blkcnt > LFS_MARKV_MAXBLKCNT) 1431 return (EINVAL); 1432 blkiov = lfs_malloc(fs, blkcnt * sizeof(BLOCK_INFO), LFS_NB_BLKIOV); 1433 if ((error = copyin(blkvp.blkiov, blkiov, 1434 blkcnt * sizeof(BLOCK_INFO))) != 0) { 1435 lfs_free(fs, blkiov, LFS_NB_BLKIOV); 1436 return error; 1437 } 1438 1439 simple_lock(&fs->lfs_interlock); 1440 ++fs->lfs_sleepers; 1441 simple_unlock(&fs->lfs_interlock); 1442 VOP_UNLOCK(ap->a_vp, 0); 1443 if (ap->a_command == LFCNBMAPV) 1444 error = lfs_bmapv(ap->a_p, fsidp, blkiov, blkcnt); 1445 else /* LFCNMARKV */ 1446 error = lfs_markv(ap->a_p, fsidp, blkiov, blkcnt); 1447 if (error == 0) 1448 error = copyout(blkiov, blkvp.blkiov, 1449 blkcnt * sizeof(BLOCK_INFO)); 1450 VOP_LOCK(ap->a_vp, LK_EXCLUSIVE); 1451 simple_lock(&fs->lfs_interlock); 1452 if (--fs->lfs_sleepers == 0) 1453 wakeup(&fs->lfs_sleepers); 1454 simple_unlock(&fs->lfs_interlock); 1455 lfs_free(fs, blkiov, LFS_NB_BLKIOV); 1456 return error; 1457 1458 case LFCNRECLAIM: 1459 /* 1460 * Flush dirops and write Ifile, allowing empty segments 1461 * to be immediately reclaimed. 1462 */ 1463 VOP_UNLOCK(ap->a_vp, 0); 1464 lfs_writer_enter(fs, "pndirop"); 1465 off = fs->lfs_offset; 1466 lfs_seglock(fs, SEGM_FORCE_CKP | SEGM_CKP); 1467 lfs_flush_dirops(fs); 1468 LFS_CLEANERINFO(cip, fs, bp); 1469 oclean = cip->clean; 1470 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 1471 lfs_segwrite(ap->a_vp->v_mount, SEGM_FORCE_CKP); 1472 fs->lfs_sp->seg_flags |= SEGM_PROT; 1473 lfs_segunlock(fs); 1474 lfs_writer_leave(fs); 1475 1476 #ifdef DEBUG 1477 LFS_CLEANERINFO(cip, fs, bp); 1478 DLOG((DLOG_CLEAN, "lfs_fcntl: reclaim wrote %" PRId64 1479 " blocks, cleaned %" PRId32 " segments (activesb %d)\n", 1480 fs->lfs_offset - off, cip->clean - oclean, 1481 fs->lfs_activesb)); 1482 LFS_SYNC_CLEANERINFO(cip, fs, bp, 0); 1483 #endif 1484 1485 VOP_LOCK(ap->a_vp, LK_EXCLUSIVE); 1486 return 0; 1487 1488 case LFCNIFILEFH: 1489 /* Return the filehandle of the Ifile */ 1490 if ((error = suser(ap->a_p->p_ucred, &ap->a_p->p_acflag)) != 0) 1491 return (error); 1492 fhp = (struct fhandle *)ap->a_data; 1493 fhp->fh_fsid = *fsidp; 1494 return lfs_vptofh(fs->lfs_ivnode, &(fhp->fh_fid)); 1495 1496 case LFCNREWIND: 1497 /* Move lfs_offset to the lowest-numbered segment */ 1498 return lfs_rewind(fs, *(int *)ap->a_data); 1499 1500 case LFCNINVAL: 1501 /* Mark a segment SEGUSE_INVAL */ 1502 LFS_SEGENTRY(sup, fs, *(int *)ap->a_data, bp); 1503 if (sup->su_nbytes > 0) { 1504 brelse(bp); 1505 lfs_unset_inval_all(fs); 1506 return EBUSY; 1507 } 1508 sup->su_flags |= SEGUSE_INVAL; 1509 VOP_BWRITE(bp); 1510 return 0; 1511 1512 case LFCNRESIZE: 1513 /* Resize the filesystem */ 1514 return lfs_resize_fs(fs, *(int *)ap->a_data); 1515 1516 case LFCNWRAPSTOP: 1517 /* 1518 * Hold lfs_newseg at segment 0; sleep until the filesystem 1519 * wraps around. For debugging purposes, so an external 1520 * agent can log every segment in the filesystem as it 1521 * was written, and we can regression-test checkpoint 1522 * validity in the general case. 1523 */ 1524 VOP_UNLOCK(ap->a_vp, 0); 1525 simple_lock(&fs->lfs_interlock); 1526 fs->lfs_nowrap = 1; 1527 error = ltsleep(&fs->lfs_nowrap, PCATCH | PUSER | PNORELOCK, 1528 "segwrap", 0, &fs->lfs_interlock); 1529 if (error) { 1530 fs->lfs_nowrap = 0; 1531 wakeup(&fs->lfs_nowrap); 1532 } 1533 VOP_LOCK(ap->a_vp, LK_EXCLUSIVE); 1534 return 0; 1535 1536 case LFCNWRAPGO: 1537 /* 1538 * Having done its work, the agent wakes up the writer. 1539 * It sleeps until a new segment is selected. 1540 */ 1541 VOP_UNLOCK(ap->a_vp, 0); 1542 simple_lock(&fs->lfs_interlock); 1543 fs->lfs_nowrap = 0; 1544 wakeup(&fs->lfs_nowrap); 1545 ltsleep(&fs->lfs_nextseg, PCATCH | PUSER | PNORELOCK, 1546 "segment", 0, &fs->lfs_interlock); 1547 VOP_LOCK(ap->a_vp, LK_EXCLUSIVE); 1548 return 0; 1549 1550 default: 1551 return ufs_fcntl(v); 1552 } 1553 return 0; 1554 } 1555 1556 int 1557 lfs_getpages(void *v) 1558 { 1559 struct vop_getpages_args /* { 1560 struct vnode *a_vp; 1561 voff_t a_offset; 1562 struct vm_page **a_m; 1563 int *a_count; 1564 int a_centeridx; 1565 vm_prot_t a_access_type; 1566 int a_advice; 1567 int a_flags; 1568 } */ *ap = v; 1569 1570 if (VTOI(ap->a_vp)->i_number == LFS_IFILE_INUM && 1571 (ap->a_access_type & VM_PROT_WRITE) != 0) { 1572 return EPERM; 1573 } 1574 if ((ap->a_access_type & VM_PROT_WRITE) != 0) { 1575 LFS_SET_UINO(VTOI(ap->a_vp), IN_MODIFIED); 1576 } 1577 1578 /* 1579 * we're relying on the fact that genfs_getpages() always read in 1580 * entire filesystem blocks. 1581 */ 1582 return genfs_getpages(v); 1583 } 1584 1585 /* 1586 * Make sure that for all pages in every block in the given range, 1587 * either all are dirty or all are clean. If any of the pages 1588 * we've seen so far are dirty, put the vnode on the paging chain, 1589 * and mark it IN_PAGING. 1590 * 1591 * If checkfirst != 0, don't check all the pages but return at the 1592 * first dirty page. 1593 */ 1594 static int 1595 check_dirty(struct lfs *fs, struct vnode *vp, 1596 off_t startoffset, off_t endoffset, off_t blkeof, 1597 int flags, int checkfirst) 1598 { 1599 int by_list; 1600 struct vm_page *curpg = NULL; /* XXX: gcc */ 1601 struct vm_page *pgs[MAXBSIZE / PAGE_SIZE], *pg; 1602 off_t soff = 0; /* XXX: gcc */ 1603 voff_t off; 1604 int i; 1605 int nonexistent; 1606 int any_dirty; /* number of dirty pages */ 1607 int dirty; /* number of dirty pages in a block */ 1608 int tdirty; 1609 int pages_per_block = fs->lfs_bsize >> PAGE_SHIFT; 1610 int pagedaemon = (curproc == uvm.pagedaemon_proc); 1611 1612 ASSERT_MAYBE_SEGLOCK(fs); 1613 top: 1614 by_list = (vp->v_uobj.uo_npages <= 1615 ((endoffset - startoffset) >> PAGE_SHIFT) * 1616 UVM_PAGE_HASH_PENALTY); 1617 any_dirty = 0; 1618 1619 if (by_list) { 1620 curpg = TAILQ_FIRST(&vp->v_uobj.memq); 1621 } else { 1622 soff = startoffset; 1623 } 1624 while (by_list || soff < MIN(blkeof, endoffset)) { 1625 if (by_list) { 1626 /* 1627 * Find the first page in a block. Skip 1628 * blocks outside our area of interest or beyond 1629 * the end of file. 1630 */ 1631 if (pages_per_block > 1) { 1632 while (curpg && 1633 ((curpg->offset & fs->lfs_bmask) || 1634 curpg->offset >= vp->v_size || 1635 curpg->offset >= endoffset)) 1636 curpg = TAILQ_NEXT(curpg, listq); 1637 } 1638 if (curpg == NULL) 1639 break; 1640 soff = curpg->offset; 1641 } 1642 1643 /* 1644 * Mark all pages in extended range busy; find out if any 1645 * of them are dirty. 1646 */ 1647 nonexistent = dirty = 0; 1648 for (i = 0; i == 0 || i < pages_per_block; i++) { 1649 if (by_list && pages_per_block <= 1) { 1650 pgs[i] = pg = curpg; 1651 } else { 1652 off = soff + (i << PAGE_SHIFT); 1653 pgs[i] = pg = uvm_pagelookup(&vp->v_uobj, off); 1654 if (pg == NULL) { 1655 ++nonexistent; 1656 continue; 1657 } 1658 } 1659 KASSERT(pg != NULL); 1660 1661 /* 1662 * If we're holding the segment lock, we can deadlock 1663 * against a process that has our page and is waiting 1664 * for the cleaner, while the cleaner waits for the 1665 * segment lock. Just bail in that case. 1666 */ 1667 if ((pg->flags & PG_BUSY) && 1668 (pagedaemon || LFS_SEGLOCK_HELD(fs))) { 1669 if (by_list && i > 0) 1670 uvm_page_unbusy(pgs, i); 1671 DLOG((DLOG_PAGE, "lfs_putpages: avoiding 3-way or pagedaemon deadlock\n")); 1672 return -1; 1673 } 1674 1675 while (pg->flags & PG_BUSY) { 1676 pg->flags |= PG_WANTED; 1677 UVM_UNLOCK_AND_WAIT(pg, &vp->v_interlock, 0, 1678 "lfsput", 0); 1679 simple_lock(&vp->v_interlock); 1680 if (by_list) { 1681 if (i > 0) 1682 uvm_page_unbusy(pgs, i); 1683 goto top; 1684 } 1685 } 1686 pg->flags |= PG_BUSY; 1687 UVM_PAGE_OWN(pg, "lfs_putpages"); 1688 1689 pmap_page_protect(pg, VM_PROT_NONE); 1690 tdirty = (pmap_clear_modify(pg) || 1691 (pg->flags & PG_CLEAN) == 0); 1692 dirty += tdirty; 1693 } 1694 if (pages_per_block > 0 && nonexistent >= pages_per_block) { 1695 if (by_list) { 1696 curpg = TAILQ_NEXT(curpg, listq); 1697 } else { 1698 soff += fs->lfs_bsize; 1699 } 1700 continue; 1701 } 1702 1703 any_dirty += dirty; 1704 KASSERT(nonexistent == 0); 1705 1706 /* 1707 * If any are dirty make all dirty; unbusy them, 1708 * but if we were asked to clean, wire them so that 1709 * the pagedaemon doesn't bother us about them while 1710 * they're on their way to disk. 1711 */ 1712 for (i = 0; i == 0 || i < pages_per_block; i++) { 1713 pg = pgs[i]; 1714 KASSERT(!((pg->flags & PG_CLEAN) && (pg->flags & PG_DELWRI))); 1715 if (dirty) { 1716 pg->flags &= ~PG_CLEAN; 1717 if (flags & PGO_FREE) { 1718 /* 1719 * Wire the page so that 1720 * pdaemon doesn't see it again. 1721 */ 1722 uvm_lock_pageq(); 1723 uvm_pagewire(pg); 1724 uvm_unlock_pageq(); 1725 1726 /* Suspended write flag */ 1727 pg->flags |= PG_DELWRI; 1728 } 1729 } 1730 if (pg->flags & PG_WANTED) 1731 wakeup(pg); 1732 pg->flags &= ~(PG_WANTED|PG_BUSY); 1733 UVM_PAGE_OWN(pg, NULL); 1734 } 1735 1736 if (checkfirst && any_dirty) 1737 break; 1738 1739 if (by_list) { 1740 curpg = TAILQ_NEXT(curpg, listq); 1741 } else { 1742 soff += MAX(PAGE_SIZE, fs->lfs_bsize); 1743 } 1744 } 1745 1746 return any_dirty; 1747 } 1748 1749 /* 1750 * lfs_putpages functions like genfs_putpages except that 1751 * 1752 * (1) It needs to bounds-check the incoming requests to ensure that 1753 * they are block-aligned; if they are not, expand the range and 1754 * do the right thing in case, e.g., the requested range is clean 1755 * but the expanded range is dirty. 1756 * 1757 * (2) It needs to explicitly send blocks to be written when it is done. 1758 * VOP_PUTPAGES is not ever called with the seglock held, so 1759 * we simply take the seglock and let lfs_segunlock wait for us. 1760 * XXX Actually we can be called with the seglock held, if we have 1761 * XXX to flush a vnode while lfs_markv is in operation. As of this 1762 * XXX writing we panic in this case. 1763 * 1764 * Assumptions: 1765 * 1766 * (1) The caller does not hold any pages in this vnode busy. If it does, 1767 * there is a danger that when we expand the page range and busy the 1768 * pages we will deadlock. 1769 * 1770 * (2) We are called with vp->v_interlock held; we must return with it 1771 * released. 1772 * 1773 * (3) We don't absolutely have to free pages right away, provided that 1774 * the request does not have PGO_SYNCIO. When the pagedaemon gives 1775 * us a request with PGO_FREE, we take the pages out of the paging 1776 * queue and wake up the writer, which will handle freeing them for us. 1777 * 1778 * We ensure that for any filesystem block, all pages for that 1779 * block are either resident or not, even if those pages are higher 1780 * than EOF; that means that we will be getting requests to free 1781 * "unused" pages above EOF all the time, and should ignore them. 1782 * 1783 * (4) If we are called with PGO_LOCKED, the finfo array we are to write 1784 * into has been set up for us by lfs_writefile. If not, we will 1785 * have to handle allocating and/or freeing an finfo entry. 1786 * 1787 * XXX note that we're (ab)using PGO_LOCKED as "seglock held". 1788 */ 1789 1790 int 1791 lfs_putpages(void *v) 1792 { 1793 int error; 1794 struct vop_putpages_args /* { 1795 struct vnode *a_vp; 1796 voff_t a_offlo; 1797 voff_t a_offhi; 1798 int a_flags; 1799 } */ *ap = v; 1800 struct vnode *vp; 1801 struct inode *ip; 1802 struct lfs *fs; 1803 struct segment *sp; 1804 off_t origoffset, startoffset, endoffset, origendoffset, blkeof; 1805 off_t off, max_endoffset; 1806 int s; 1807 boolean_t seglocked, sync, pagedaemon; 1808 struct vm_page *pg; 1809 UVMHIST_FUNC("lfs_putpages"); UVMHIST_CALLED(ubchist); 1810 1811 vp = ap->a_vp; 1812 ip = VTOI(vp); 1813 fs = ip->i_lfs; 1814 sync = (ap->a_flags & PGO_SYNCIO) != 0; 1815 pagedaemon = (curproc == uvm.pagedaemon_proc); 1816 1817 /* Putpages does nothing for metadata. */ 1818 if (vp == fs->lfs_ivnode || vp->v_type != VREG) { 1819 simple_unlock(&vp->v_interlock); 1820 return 0; 1821 } 1822 1823 /* 1824 * If there are no pages, don't do anything. 1825 */ 1826 if (vp->v_uobj.uo_npages == 0) { 1827 s = splbio(); 1828 if (LIST_FIRST(&vp->v_dirtyblkhd) == NULL && 1829 (vp->v_flag & VONWORKLST)) { 1830 vp->v_flag &= ~VONWORKLST; 1831 LIST_REMOVE(vp, v_synclist); 1832 } 1833 splx(s); 1834 simple_unlock(&vp->v_interlock); 1835 1836 /* Remove us from paging queue, if we were on it */ 1837 simple_lock(&fs->lfs_interlock); 1838 if (ip->i_flags & IN_PAGING) { 1839 ip->i_flags &= ~IN_PAGING; 1840 TAILQ_REMOVE(&fs->lfs_pchainhd, ip, i_lfs_pchain); 1841 } 1842 simple_unlock(&fs->lfs_interlock); 1843 return 0; 1844 } 1845 1846 blkeof = blkroundup(fs, ip->i_size); 1847 1848 /* 1849 * Ignore requests to free pages past EOF but in the same block 1850 * as EOF, unless the request is synchronous. (If the request is 1851 * sync, it comes from lfs_truncate.) 1852 * XXXUBC Make these pages look "active" so the pagedaemon won't 1853 * XXXUBC bother us with them again. 1854 */ 1855 if (!sync && ap->a_offlo >= ip->i_size && ap->a_offlo < blkeof) { 1856 origoffset = ap->a_offlo; 1857 for (off = origoffset; off < blkeof; off += fs->lfs_bsize) { 1858 pg = uvm_pagelookup(&vp->v_uobj, off); 1859 KASSERT(pg != NULL); 1860 while (pg->flags & PG_BUSY) { 1861 pg->flags |= PG_WANTED; 1862 UVM_UNLOCK_AND_WAIT(pg, &vp->v_interlock, 0, 1863 "lfsput2", 0); 1864 simple_lock(&vp->v_interlock); 1865 } 1866 uvm_lock_pageq(); 1867 uvm_pageactivate(pg); 1868 uvm_unlock_pageq(); 1869 } 1870 ap->a_offlo = blkeof; 1871 if (ap->a_offhi > 0 && ap->a_offhi <= ap->a_offlo) { 1872 simple_unlock(&vp->v_interlock); 1873 return 0; 1874 } 1875 } 1876 1877 /* 1878 * Extend page range to start and end at block boundaries. 1879 * (For the purposes of VOP_PUTPAGES, fragments don't exist.) 1880 */ 1881 origoffset = ap->a_offlo; 1882 origendoffset = ap->a_offhi; 1883 startoffset = origoffset & ~(fs->lfs_bmask); 1884 max_endoffset = (trunc_page(LLONG_MAX) >> fs->lfs_bshift) 1885 << fs->lfs_bshift; 1886 1887 if (origendoffset == 0 || ap->a_flags & PGO_ALLPAGES) { 1888 endoffset = max_endoffset; 1889 origendoffset = endoffset; 1890 } else { 1891 origendoffset = round_page(ap->a_offhi); 1892 endoffset = round_page(blkroundup(fs, origendoffset)); 1893 } 1894 1895 KASSERT(startoffset > 0 || endoffset >= startoffset); 1896 if (startoffset == endoffset) { 1897 /* Nothing to do, why were we called? */ 1898 simple_unlock(&vp->v_interlock); 1899 DLOG((DLOG_PAGE, "lfs_putpages: startoffset = endoffset = %" 1900 PRId64 "\n", startoffset)); 1901 return 0; 1902 } 1903 1904 ap->a_offlo = startoffset; 1905 ap->a_offhi = endoffset; 1906 1907 if (!(ap->a_flags & PGO_CLEANIT)) 1908 return genfs_putpages(v); 1909 1910 /* 1911 * If there are more than one page per block, we don't want 1912 * to get caught locking them backwards; so set PGO_BUSYFAIL 1913 * to avoid deadlocks. 1914 */ 1915 ap->a_flags |= PGO_BUSYFAIL; 1916 1917 do { 1918 int r; 1919 1920 /* If no pages are dirty, we can just use genfs_putpages. */ 1921 r = check_dirty(fs, vp, startoffset, endoffset, blkeof, 1922 ap->a_flags, 1); 1923 if (r < 0) { 1924 simple_unlock(&vp->v_interlock); 1925 return EDEADLK; 1926 } 1927 if (r > 0) 1928 break; 1929 1930 /* 1931 * Sometimes pages are dirtied between the time that 1932 * we check and the time we try to clean them. 1933 * Instruct lfs_gop_write to return EDEADLK in this case 1934 * so we can write them properly. 1935 */ 1936 ip->i_lfs_iflags |= LFSI_NO_GOP_WRITE; 1937 r = genfs_putpages(v); 1938 ip->i_lfs_iflags &= ~LFSI_NO_GOP_WRITE; 1939 if (r != EDEADLK) 1940 return r; 1941 1942 /* Start over. */ 1943 preempt(1); 1944 simple_lock(&vp->v_interlock); 1945 } while(1); 1946 1947 /* 1948 * Dirty and asked to clean. 1949 * 1950 * Pagedaemon can't actually write LFS pages; wake up 1951 * the writer to take care of that. The writer will 1952 * notice the pager inode queue and act on that. 1953 */ 1954 if (pagedaemon) { 1955 simple_lock(&fs->lfs_interlock); 1956 if (!(ip->i_flags & IN_PAGING)) { 1957 ip->i_flags |= IN_PAGING; 1958 TAILQ_INSERT_TAIL(&fs->lfs_pchainhd, ip, i_lfs_pchain); 1959 } 1960 simple_lock(&lfs_subsys_lock); 1961 wakeup(&lfs_writer_daemon); 1962 simple_unlock(&lfs_subsys_lock); 1963 simple_unlock(&fs->lfs_interlock); 1964 simple_unlock(&vp->v_interlock); 1965 preempt(1); 1966 return EWOULDBLOCK; 1967 } 1968 1969 /* 1970 * If this is a file created in a recent dirop, we can't flush its 1971 * inode until the dirop is complete. Drain dirops, then flush the 1972 * filesystem (taking care of any other pending dirops while we're 1973 * at it). 1974 */ 1975 if ((ap->a_flags & (PGO_CLEANIT|PGO_LOCKED)) == PGO_CLEANIT && 1976 (vp->v_flag & VDIROP)) { 1977 int locked; 1978 1979 DLOG((DLOG_PAGE, "lfs_putpages: flushing VDIROP\n")); 1980 locked = VOP_ISLOCKED(vp) && /* XXX */ 1981 vp->v_lock.lk_lockholder == curproc->p_pid; 1982 simple_unlock(&vp->v_interlock); 1983 lfs_writer_enter(fs, "ppdirop"); 1984 if (locked) 1985 VOP_UNLOCK(vp, 0); 1986 1987 simple_lock(&fs->lfs_interlock); 1988 lfs_flush_fs(fs, sync ? SEGM_SYNC : 0); 1989 simple_unlock(&fs->lfs_interlock); 1990 1991 simple_lock(&vp->v_interlock); 1992 if (locked) { 1993 VOP_LOCK(vp, LK_EXCLUSIVE | LK_INTERLOCK); 1994 simple_lock(&vp->v_interlock); 1995 } 1996 lfs_writer_leave(fs); 1997 1998 /* XXX the flush should have taken care of this one too! */ 1999 } 2000 2001 /* 2002 * This is it. We are going to write some pages. From here on 2003 * down it's all just mechanics. 2004 * 2005 * Don't let genfs_putpages wait; lfs_segunlock will wait for us. 2006 */ 2007 ap->a_flags &= ~PGO_SYNCIO; 2008 2009 /* 2010 * If we've already got the seglock, flush the node and return. 2011 * The FIP has already been set up for us by lfs_writefile, 2012 * and FIP cleanup and lfs_updatemeta will also be done there, 2013 * unless genfs_putpages returns EDEADLK; then we must flush 2014 * what we have, and correct FIP and segment header accounting. 2015 */ 2016 get_seglock: 2017 seglocked = (ap->a_flags & PGO_LOCKED) != 0; 2018 if (!seglocked) { 2019 simple_unlock(&vp->v_interlock); 2020 /* 2021 * Take the seglock, because we are going to be writing pages. 2022 */ 2023 error = lfs_seglock(fs, SEGM_PROT | (sync ? SEGM_SYNC : 0)); 2024 if (error != 0) 2025 return error; 2026 simple_lock(&vp->v_interlock); 2027 } 2028 2029 /* 2030 * VOP_PUTPAGES should not be called while holding the seglock. 2031 * XXXUBC fix lfs_markv, or do this properly. 2032 */ 2033 #ifdef notyet 2034 KASSERT(fs->lfs_seglock == 1); 2035 #endif /* notyet */ 2036 2037 /* 2038 * We assume we're being called with sp->fip pointing at blank space. 2039 * Account for a new FIP in the segment header, and set sp->vp. 2040 * (This should duplicate the setup at the top of lfs_writefile().) 2041 */ 2042 sp = fs->lfs_sp; 2043 if (!seglocked) 2044 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 2045 KASSERT(sp->vp == NULL); 2046 sp->vp = vp; 2047 2048 if (!seglocked) { 2049 if (vp->v_flag & VDIROP) 2050 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 2051 } 2052 2053 /* 2054 * Loop through genfs_putpages until all pages are gathered. 2055 * genfs_putpages() drops the interlock, so reacquire it if necessary. 2056 * Whenever we lose the interlock we have to rerun check_dirty, as 2057 * well. 2058 */ 2059 again: 2060 if (check_dirty(fs, vp, startoffset, endoffset, blkeof, 2061 ap->a_flags, 0) < 0) { 2062 simple_unlock(&vp->v_interlock); 2063 sp->vp = NULL; 2064 if (!seglocked) { 2065 lfs_release_finfo(fs); 2066 lfs_segunlock(fs); 2067 } 2068 if (pagedaemon) 2069 return EDEADLK; 2070 /* else seglocked == 0 */ 2071 preempt(1); 2072 simple_lock(&vp->v_interlock); 2073 goto get_seglock; 2074 } 2075 2076 error = genfs_putpages(v); 2077 if (error == EDEADLK || error == EAGAIN) { 2078 DLOG((DLOG_PAGE, "lfs_putpages: genfs_putpages returned" 2079 " EDEADLK [2] ino %d off %x (seg %d)\n", 2080 ip->i_number, fs->lfs_offset, 2081 dtosn(fs, fs->lfs_offset))); 2082 /* If nothing to write, short-circuit */ 2083 if (sp->cbpp - sp->bpp > 1) { 2084 /* Write gathered pages */ 2085 lfs_updatemeta(sp); 2086 (void) lfs_writeseg(fs, sp); 2087 2088 /* 2089 * Reinitialize brand new FIP and add us to it. 2090 */ 2091 KASSERT(sp->vp == vp); 2092 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 2093 } 2094 2095 /* Give the write a chance to complete */ 2096 preempt(1); 2097 2098 /* We've lost the interlock. Start over. */ 2099 if (error == EDEADLK) { 2100 simple_lock(&vp->v_interlock); 2101 goto again; 2102 } 2103 } 2104 2105 KASSERT(sp->vp == vp); 2106 if (!seglocked) { 2107 sp->vp = NULL; 2108 2109 /* Write indirect blocks as well */ 2110 lfs_gather(fs, fs->lfs_sp, vp, lfs_match_indir); 2111 lfs_gather(fs, fs->lfs_sp, vp, lfs_match_dindir); 2112 lfs_gather(fs, fs->lfs_sp, vp, lfs_match_tindir); 2113 2114 KASSERT(sp->vp == NULL); 2115 sp->vp = vp; 2116 } 2117 2118 /* 2119 * Blocks are now gathered into a segment waiting to be written. 2120 * All that's left to do is update metadata, and write them. 2121 */ 2122 lfs_updatemeta(sp); 2123 KASSERT(sp->vp == vp); 2124 sp->vp = NULL; 2125 2126 if (seglocked) { 2127 /* we're called by lfs_writefile. */ 2128 return error; 2129 } 2130 2131 /* Clean up FIP and send it to disk. */ 2132 lfs_release_finfo(fs); 2133 lfs_writeseg(fs, fs->lfs_sp); 2134 2135 /* 2136 * Remove us from paging queue, since we've now written all our 2137 * pages. 2138 */ 2139 simple_lock(&fs->lfs_interlock); 2140 if (ip->i_flags & IN_PAGING) { 2141 ip->i_flags &= ~IN_PAGING; 2142 TAILQ_REMOVE(&fs->lfs_pchainhd, ip, i_lfs_pchain); 2143 } 2144 simple_unlock(&fs->lfs_interlock); 2145 2146 /* 2147 * XXX - with the malloc/copy writeseg, the pages are freed by now 2148 * even if we don't wait (e.g. if we hold a nested lock). This 2149 * will not be true if we stop using malloc/copy. 2150 */ 2151 KASSERT(fs->lfs_sp->seg_flags & SEGM_PROT); 2152 lfs_segunlock(fs); 2153 2154 /* 2155 * Wait for v_numoutput to drop to zero. The seglock should 2156 * take care of this, but there is a slight possibility that 2157 * aiodoned might not have got around to our buffers yet. 2158 */ 2159 if (sync) { 2160 s = splbio(); 2161 simple_lock(&global_v_numoutput_slock); 2162 while (vp->v_numoutput > 0) { 2163 DLOG((DLOG_PAGE, "lfs_putpages: ino %d sleeping on" 2164 " num %d\n", ip->i_number, vp->v_numoutput)); 2165 vp->v_flag |= VBWAIT; 2166 ltsleep(&vp->v_numoutput, PRIBIO + 1, "lfs_vn", 0, 2167 &global_v_numoutput_slock); 2168 } 2169 simple_unlock(&global_v_numoutput_slock); 2170 splx(s); 2171 } 2172 return error; 2173 } 2174 2175 /* 2176 * Return the last logical file offset that should be written for this file 2177 * if we're doing a write that ends at "size". If writing, we need to know 2178 * about sizes on disk, i.e. fragments if there are any; if reading, we need 2179 * to know about entire blocks. 2180 */ 2181 void 2182 lfs_gop_size(struct vnode *vp, off_t size, off_t *eobp, int flags) 2183 { 2184 struct inode *ip = VTOI(vp); 2185 struct lfs *fs = ip->i_lfs; 2186 daddr_t olbn, nlbn; 2187 2188 KASSERT(flags & (GOP_SIZE_READ | GOP_SIZE_WRITE)); 2189 KASSERT((flags & (GOP_SIZE_READ | GOP_SIZE_WRITE)) 2190 != (GOP_SIZE_READ | GOP_SIZE_WRITE)); 2191 2192 olbn = lblkno(fs, ip->i_size); 2193 nlbn = lblkno(fs, size); 2194 if (!(flags & GOP_SIZE_MEM) && nlbn < NDADDR && olbn <= nlbn) { 2195 *eobp = fragroundup(fs, size); 2196 } else { 2197 *eobp = blkroundup(fs, size); 2198 } 2199 } 2200 2201 #ifdef DEBUG 2202 void lfs_dump_vop(void *); 2203 2204 void 2205 lfs_dump_vop(void *v) 2206 { 2207 struct vop_putpages_args /* { 2208 struct vnode *a_vp; 2209 voff_t a_offlo; 2210 voff_t a_offhi; 2211 int a_flags; 2212 } */ *ap = v; 2213 2214 #ifdef DDB 2215 vfs_vnode_print(ap->a_vp, 0, printf); 2216 #endif 2217 lfs_dump_dinode(VTOI(ap->a_vp)->i_din.ffs1_din); 2218 } 2219 #endif 2220 2221 int 2222 lfs_mmap(void *v) 2223 { 2224 struct vop_mmap_args /* { 2225 const struct vnodeop_desc *a_desc; 2226 struct vnode *a_vp; 2227 int a_fflags; 2228 struct ucred *a_cred; 2229 struct proc *a_p; 2230 } */ *ap = v; 2231 2232 if (VTOI(ap->a_vp)->i_number == LFS_IFILE_INUM) 2233 return EOPNOTSUPP; 2234 return ufs_mmap(v); 2235 } 2236
Indexes created Sat Oct 18 23:09:54 GMT 2025