lfs_vnops.c revision 1.239.2.2
1 /* $NetBSD: lfs_vnops.c,v 1.239.2.2 2016/08/27 14:13:18 bouyer 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 * 19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 /* 32 * Copyright (c) 1986, 1989, 1991, 1993, 1995 33 * The Regents of the University of California. All rights reserved. 34 * 35 * Redistribution and use in source and binary forms, with or without 36 * modification, are permitted provided that the following conditions 37 * are met: 38 * 1. Redistributions of source code must retain the above copyright 39 * notice, this list of conditions and the following disclaimer. 40 * 2. Redistributions in binary form must reproduce the above copyright 41 * notice, this list of conditions and the following disclaimer in the 42 * documentation and/or other materials provided with the distribution. 43 * 3. Neither the name of the University nor the names of its contributors 44 * may be used to endorse or promote products derived from this software 45 * without specific prior written permission. 46 * 47 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 48 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 49 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 50 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 51 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 52 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 53 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 54 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 55 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 56 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 57 * SUCH DAMAGE. 58 * 59 * @(#)lfs_vnops.c 8.13 (Berkeley) 6/10/95 60 */ 61 62 #include <sys/cdefs.h> 63 __KERNEL_RCSID(0, "$NetBSD: lfs_vnops.c,v 1.239.2.2 2016/08/27 14:13:18 bouyer Exp $"); 64 65 #ifdef _KERNEL_OPT 66 #include "opt_compat_netbsd.h" 67 #include "opt_uvm_page_trkown.h" 68 #endif 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/namei.h> 73 #include <sys/resourcevar.h> 74 #include <sys/kernel.h> 75 #include <sys/file.h> 76 #include <sys/stat.h> 77 #include <sys/buf.h> 78 #include <sys/proc.h> 79 #include <sys/mount.h> 80 #include <sys/vnode.h> 81 #include <sys/pool.h> 82 #include <sys/signalvar.h> 83 #include <sys/kauth.h> 84 #include <sys/syslog.h> 85 #include <sys/fstrans.h> 86 87 #include <miscfs/fifofs/fifo.h> 88 #include <miscfs/genfs/genfs.h> 89 #include <miscfs/specfs/specdev.h> 90 91 #include <ufs/ufs/inode.h> 92 #include <ufs/ufs/dir.h> 93 #include <ufs/ufs/ufsmount.h> 94 #include <ufs/ufs/ufs_bswap.h> 95 #include <ufs/ufs/ufs_extern.h> 96 97 #include <uvm/uvm.h> 98 #include <uvm/uvm_pmap.h> 99 #include <uvm/uvm_stat.h> 100 #include <uvm/uvm_pager.h> 101 102 #include <ufs/lfs/lfs.h> 103 #include <ufs/lfs/lfs_extern.h> 104 105 extern pid_t lfs_writer_daemon; 106 int lfs_ignore_lazy_sync = 1; 107 108 /* Global vfs data structures for lfs. */ 109 int (**lfs_vnodeop_p)(void *); 110 const struct vnodeopv_entry_desc lfs_vnodeop_entries[] = { 111 { &vop_default_desc, vn_default_error }, 112 { &vop_lookup_desc, ufs_lookup }, /* lookup */ 113 { &vop_create_desc, lfs_create }, /* create */ 114 { &vop_whiteout_desc, ufs_whiteout }, /* whiteout */ 115 { &vop_mknod_desc, lfs_mknod }, /* mknod */ 116 { &vop_open_desc, ufs_open }, /* open */ 117 { &vop_close_desc, lfs_close }, /* close */ 118 { &vop_access_desc, ufs_access }, /* access */ 119 { &vop_getattr_desc, lfs_getattr }, /* getattr */ 120 { &vop_setattr_desc, lfs_setattr }, /* setattr */ 121 { &vop_read_desc, lfs_read }, /* read */ 122 { &vop_write_desc, lfs_write }, /* write */ 123 { &vop_ioctl_desc, ufs_ioctl }, /* ioctl */ 124 { &vop_fcntl_desc, lfs_fcntl }, /* fcntl */ 125 { &vop_poll_desc, ufs_poll }, /* poll */ 126 { &vop_kqfilter_desc, genfs_kqfilter }, /* kqfilter */ 127 { &vop_revoke_desc, ufs_revoke }, /* revoke */ 128 { &vop_mmap_desc, lfs_mmap }, /* mmap */ 129 { &vop_fsync_desc, lfs_fsync }, /* fsync */ 130 { &vop_seek_desc, ufs_seek }, /* seek */ 131 { &vop_remove_desc, lfs_remove }, /* remove */ 132 { &vop_link_desc, lfs_link }, /* link */ 133 { &vop_rename_desc, lfs_rename }, /* rename */ 134 { &vop_mkdir_desc, lfs_mkdir }, /* mkdir */ 135 { &vop_rmdir_desc, lfs_rmdir }, /* rmdir */ 136 { &vop_symlink_desc, lfs_symlink }, /* symlink */ 137 { &vop_readdir_desc, ufs_readdir }, /* readdir */ 138 { &vop_readlink_desc, ufs_readlink }, /* readlink */ 139 { &vop_abortop_desc, ufs_abortop }, /* abortop */ 140 { &vop_inactive_desc, lfs_inactive }, /* inactive */ 141 { &vop_reclaim_desc, lfs_reclaim }, /* reclaim */ 142 { &vop_lock_desc, ufs_lock }, /* lock */ 143 { &vop_unlock_desc, ufs_unlock }, /* unlock */ 144 { &vop_bmap_desc, ufs_bmap }, /* bmap */ 145 { &vop_strategy_desc, lfs_strategy }, /* strategy */ 146 { &vop_print_desc, ufs_print }, /* print */ 147 { &vop_islocked_desc, ufs_islocked }, /* islocked */ 148 { &vop_pathconf_desc, ufs_pathconf }, /* pathconf */ 149 { &vop_advlock_desc, ufs_advlock }, /* advlock */ 150 { &vop_bwrite_desc, lfs_bwrite }, /* bwrite */ 151 { &vop_getpages_desc, lfs_getpages }, /* getpages */ 152 { &vop_putpages_desc, lfs_putpages }, /* putpages */ 153 { NULL, NULL } 154 }; 155 const struct vnodeopv_desc lfs_vnodeop_opv_desc = 156 { &lfs_vnodeop_p, lfs_vnodeop_entries }; 157 158 int (**lfs_specop_p)(void *); 159 const struct vnodeopv_entry_desc lfs_specop_entries[] = { 160 { &vop_default_desc, vn_default_error }, 161 { &vop_lookup_desc, spec_lookup }, /* lookup */ 162 { &vop_create_desc, spec_create }, /* create */ 163 { &vop_mknod_desc, spec_mknod }, /* mknod */ 164 { &vop_open_desc, spec_open }, /* open */ 165 { &vop_close_desc, lfsspec_close }, /* close */ 166 { &vop_access_desc, ufs_access }, /* access */ 167 { &vop_getattr_desc, lfs_getattr }, /* getattr */ 168 { &vop_setattr_desc, lfs_setattr }, /* setattr */ 169 { &vop_read_desc, ufsspec_read }, /* read */ 170 { &vop_write_desc, ufsspec_write }, /* write */ 171 { &vop_ioctl_desc, spec_ioctl }, /* ioctl */ 172 { &vop_fcntl_desc, ufs_fcntl }, /* fcntl */ 173 { &vop_poll_desc, spec_poll }, /* poll */ 174 { &vop_kqfilter_desc, spec_kqfilter }, /* kqfilter */ 175 { &vop_revoke_desc, spec_revoke }, /* revoke */ 176 { &vop_mmap_desc, spec_mmap }, /* mmap */ 177 { &vop_fsync_desc, spec_fsync }, /* fsync */ 178 { &vop_seek_desc, spec_seek }, /* seek */ 179 { &vop_remove_desc, spec_remove }, /* remove */ 180 { &vop_link_desc, spec_link }, /* link */ 181 { &vop_rename_desc, spec_rename }, /* rename */ 182 { &vop_mkdir_desc, spec_mkdir }, /* mkdir */ 183 { &vop_rmdir_desc, spec_rmdir }, /* rmdir */ 184 { &vop_symlink_desc, spec_symlink }, /* symlink */ 185 { &vop_readdir_desc, spec_readdir }, /* readdir */ 186 { &vop_readlink_desc, spec_readlink }, /* readlink */ 187 { &vop_abortop_desc, spec_abortop }, /* abortop */ 188 { &vop_inactive_desc, lfs_inactive }, /* inactive */ 189 { &vop_reclaim_desc, lfs_reclaim }, /* reclaim */ 190 { &vop_lock_desc, ufs_lock }, /* lock */ 191 { &vop_unlock_desc, ufs_unlock }, /* unlock */ 192 { &vop_bmap_desc, spec_bmap }, /* bmap */ 193 { &vop_strategy_desc, spec_strategy }, /* strategy */ 194 { &vop_print_desc, ufs_print }, /* print */ 195 { &vop_islocked_desc, ufs_islocked }, /* islocked */ 196 { &vop_pathconf_desc, spec_pathconf }, /* pathconf */ 197 { &vop_advlock_desc, spec_advlock }, /* advlock */ 198 { &vop_bwrite_desc, vn_bwrite }, /* bwrite */ 199 { &vop_getpages_desc, spec_getpages }, /* getpages */ 200 { &vop_putpages_desc, spec_putpages }, /* putpages */ 201 { NULL, NULL } 202 }; 203 const struct vnodeopv_desc lfs_specop_opv_desc = 204 { &lfs_specop_p, lfs_specop_entries }; 205 206 int (**lfs_fifoop_p)(void *); 207 const struct vnodeopv_entry_desc lfs_fifoop_entries[] = { 208 { &vop_default_desc, vn_default_error }, 209 { &vop_lookup_desc, vn_fifo_bypass }, /* lookup */ 210 { &vop_create_desc, vn_fifo_bypass }, /* create */ 211 { &vop_mknod_desc, vn_fifo_bypass }, /* mknod */ 212 { &vop_open_desc, vn_fifo_bypass }, /* open */ 213 { &vop_close_desc, lfsfifo_close }, /* close */ 214 { &vop_access_desc, ufs_access }, /* access */ 215 { &vop_getattr_desc, lfs_getattr }, /* getattr */ 216 { &vop_setattr_desc, lfs_setattr }, /* setattr */ 217 { &vop_read_desc, ufsfifo_read }, /* read */ 218 { &vop_write_desc, ufsfifo_write }, /* write */ 219 { &vop_ioctl_desc, vn_fifo_bypass }, /* ioctl */ 220 { &vop_fcntl_desc, ufs_fcntl }, /* fcntl */ 221 { &vop_poll_desc, vn_fifo_bypass }, /* poll */ 222 { &vop_kqfilter_desc, vn_fifo_bypass }, /* kqfilter */ 223 { &vop_revoke_desc, vn_fifo_bypass }, /* revoke */ 224 { &vop_mmap_desc, vn_fifo_bypass }, /* mmap */ 225 { &vop_fsync_desc, vn_fifo_bypass }, /* fsync */ 226 { &vop_seek_desc, vn_fifo_bypass }, /* seek */ 227 { &vop_remove_desc, vn_fifo_bypass }, /* remove */ 228 { &vop_link_desc, vn_fifo_bypass }, /* link */ 229 { &vop_rename_desc, vn_fifo_bypass }, /* rename */ 230 { &vop_mkdir_desc, vn_fifo_bypass }, /* mkdir */ 231 { &vop_rmdir_desc, vn_fifo_bypass }, /* rmdir */ 232 { &vop_symlink_desc, vn_fifo_bypass }, /* symlink */ 233 { &vop_readdir_desc, vn_fifo_bypass }, /* readdir */ 234 { &vop_readlink_desc, vn_fifo_bypass }, /* readlink */ 235 { &vop_abortop_desc, vn_fifo_bypass }, /* abortop */ 236 { &vop_inactive_desc, lfs_inactive }, /* inactive */ 237 { &vop_reclaim_desc, lfs_reclaim }, /* reclaim */ 238 { &vop_lock_desc, ufs_lock }, /* lock */ 239 { &vop_unlock_desc, ufs_unlock }, /* unlock */ 240 { &vop_bmap_desc, vn_fifo_bypass }, /* bmap */ 241 { &vop_strategy_desc, vn_fifo_bypass }, /* strategy */ 242 { &vop_print_desc, ufs_print }, /* print */ 243 { &vop_islocked_desc, ufs_islocked }, /* islocked */ 244 { &vop_pathconf_desc, vn_fifo_bypass }, /* pathconf */ 245 { &vop_advlock_desc, vn_fifo_bypass }, /* advlock */ 246 { &vop_bwrite_desc, lfs_bwrite }, /* bwrite */ 247 { &vop_putpages_desc, vn_fifo_bypass }, /* putpages */ 248 { NULL, NULL } 249 }; 250 const struct vnodeopv_desc lfs_fifoop_opv_desc = 251 { &lfs_fifoop_p, lfs_fifoop_entries }; 252 253 static int check_dirty(struct lfs *, struct vnode *, off_t, off_t, off_t, int, int, struct vm_page **); 254 255 #define LFS_READWRITE 256 #include <ufs/ufs/ufs_readwrite.c> 257 #undef LFS_READWRITE 258 259 /* 260 * Synch an open file. 261 */ 262 /* ARGSUSED */ 263 int 264 lfs_fsync(void *v) 265 { 266 struct vop_fsync_args /* { 267 struct vnode *a_vp; 268 kauth_cred_t a_cred; 269 int a_flags; 270 off_t offlo; 271 off_t offhi; 272 } */ *ap = v; 273 struct vnode *vp = ap->a_vp; 274 int error, wait; 275 struct inode *ip = VTOI(vp); 276 struct lfs *fs = ip->i_lfs; 277 278 /* If we're mounted read-only, don't try to sync. */ 279 if (fs->lfs_ronly) 280 return 0; 281 282 /* If a removed vnode is being cleaned, no need to sync here. */ 283 if ((ap->a_flags & FSYNC_RECLAIM) != 0 && ip->i_mode == 0) 284 return 0; 285 286 /* 287 * Trickle sync simply adds this vnode to the pager list, as if 288 * the pagedaemon had requested a pageout. 289 */ 290 if (ap->a_flags & FSYNC_LAZY) { 291 if (lfs_ignore_lazy_sync == 0) { 292 mutex_enter(&lfs_lock); 293 if (!(ip->i_flags & IN_PAGING)) { 294 ip->i_flags |= IN_PAGING; 295 TAILQ_INSERT_TAIL(&fs->lfs_pchainhd, ip, 296 i_lfs_pchain); 297 } 298 wakeup(&lfs_writer_daemon); 299 mutex_exit(&lfs_lock); 300 } 301 return 0; 302 } 303 304 /* 305 * If a vnode is bring cleaned, flush it out before we try to 306 * reuse it. This prevents the cleaner from writing files twice 307 * in the same partial segment, causing an accounting underflow. 308 */ 309 if (ap->a_flags & FSYNC_RECLAIM && ip->i_flags & IN_CLEANING) { 310 lfs_vflush(vp); 311 } 312 313 wait = (ap->a_flags & FSYNC_WAIT); 314 do { 315 mutex_enter(vp->v_interlock); 316 error = VOP_PUTPAGES(vp, trunc_page(ap->a_offlo), 317 round_page(ap->a_offhi), 318 PGO_CLEANIT | (wait ? PGO_SYNCIO : 0)); 319 if (error == EAGAIN) { 320 mutex_enter(&lfs_lock); 321 mtsleep(&fs->lfs_avail, PCATCH | PUSER, "lfs_fsync", 322 hz / 100 + 1, &lfs_lock); 323 mutex_exit(&lfs_lock); 324 } 325 } while (error == EAGAIN); 326 if (error) 327 return error; 328 329 if ((ap->a_flags & FSYNC_DATAONLY) == 0) 330 error = lfs_update(vp, NULL, NULL, wait ? UPDATE_WAIT : 0); 331 332 if (error == 0 && ap->a_flags & FSYNC_CACHE) { 333 int l = 0; 334 error = VOP_IOCTL(ip->i_devvp, DIOCCACHESYNC, &l, FWRITE, 335 curlwp->l_cred); 336 } 337 if (wait && !VPISEMPTY(vp)) 338 LFS_SET_UINO(ip, IN_MODIFIED); 339 340 return error; 341 } 342 343 /* 344 * Take IN_ADIROP off, then call ufs_inactive. 345 */ 346 int 347 lfs_inactive(void *v) 348 { 349 struct vop_inactive_args /* { 350 struct vnode *a_vp; 351 } */ *ap = v; 352 353 lfs_unmark_vnode(ap->a_vp); 354 355 /* 356 * The Ifile is only ever inactivated on unmount. 357 * Streamline this process by not giving it more dirty blocks. 358 */ 359 if (VTOI(ap->a_vp)->i_number == LFS_IFILE_INUM) { 360 mutex_enter(&lfs_lock); 361 LFS_CLR_UINO(VTOI(ap->a_vp), IN_ALLMOD); 362 mutex_exit(&lfs_lock); 363 VOP_UNLOCK(ap->a_vp); 364 return 0; 365 } 366 367 #ifdef DEBUG 368 /* 369 * This might happen on unmount. 370 * XXX If it happens at any other time, it should be a panic. 371 */ 372 if (ap->a_vp->v_uflag & VU_DIROP) { 373 struct inode *ip = VTOI(ap->a_vp); 374 printf("lfs_inactive: inactivating VU_DIROP? ino = %d\n", (int)ip->i_number); 375 } 376 #endif /* DIAGNOSTIC */ 377 378 return ufs_inactive(v); 379 } 380 381 /* 382 * These macros are used to bracket UFS directory ops, so that we can 383 * identify all the pages touched during directory ops which need to 384 * be ordered and flushed atomically, so that they may be recovered. 385 * 386 * Because we have to mark nodes VU_DIROP in order to prevent 387 * the cache from reclaiming them while a dirop is in progress, we must 388 * also manage the number of nodes so marked (otherwise we can run out). 389 * We do this by setting lfs_dirvcount to the number of marked vnodes; it 390 * is decremented during segment write, when VU_DIROP is taken off. 391 */ 392 #define MARK_VNODE(vp) lfs_mark_vnode(vp) 393 #define UNMARK_VNODE(vp) lfs_unmark_vnode(vp) 394 #define SET_DIROP_CREATE(dvp, vpp) lfs_set_dirop_create((dvp), (vpp)) 395 #define SET_DIROP_REMOVE(dvp, vp) lfs_set_dirop((dvp), (vp)) 396 static int lfs_set_dirop_create(struct vnode *, struct vnode **); 397 static int lfs_set_dirop(struct vnode *, struct vnode *); 398 399 static int 400 lfs_set_dirop(struct vnode *dvp, struct vnode *vp) 401 { 402 struct lfs *fs; 403 int error; 404 405 KASSERT(VOP_ISLOCKED(dvp)); 406 KASSERT(vp == NULL || VOP_ISLOCKED(vp)); 407 408 fs = VTOI(dvp)->i_lfs; 409 410 ASSERT_NO_SEGLOCK(fs); 411 /* 412 * LFS_NRESERVE calculates direct and indirect blocks as well 413 * as an inode block; an overestimate in most cases. 414 */ 415 if ((error = lfs_reserve(fs, dvp, vp, LFS_NRESERVE(fs))) != 0) 416 return (error); 417 418 restart: 419 mutex_enter(&lfs_lock); 420 if (fs->lfs_dirops == 0) { 421 mutex_exit(&lfs_lock); 422 lfs_check(dvp, LFS_UNUSED_LBN, 0); 423 mutex_enter(&lfs_lock); 424 } 425 while (fs->lfs_writer) { 426 error = mtsleep(&fs->lfs_dirops, (PRIBIO + 1) | PCATCH, 427 "lfs_sdirop", 0, &lfs_lock); 428 if (error == EINTR) { 429 mutex_exit(&lfs_lock); 430 goto unreserve; 431 } 432 } 433 if (lfs_dirvcount > LFS_MAX_DIROP && fs->lfs_dirops == 0) { 434 wakeup(&lfs_writer_daemon); 435 mutex_exit(&lfs_lock); 436 preempt(); 437 goto restart; 438 } 439 440 if (lfs_dirvcount > LFS_MAX_DIROP) { 441 DLOG((DLOG_DIROP, "lfs_set_dirop: sleeping with dirops=%d, " 442 "dirvcount=%d\n", fs->lfs_dirops, lfs_dirvcount)); 443 if ((error = mtsleep(&lfs_dirvcount, 444 PCATCH | PUSER | PNORELOCK, "lfs_maxdirop", 0, 445 &lfs_lock)) != 0) { 446 mutex_exit(&lfs_lock); 447 goto unreserve; 448 } 449 mutex_exit(&lfs_lock); 450 goto restart; 451 } 452 453 ++fs->lfs_dirops; 454 /* fs->lfs_doifile = 1; */ /* XXX why? --ks */ 455 mutex_exit(&lfs_lock); 456 457 /* Hold a reference so SET_ENDOP will be happy */ 458 vref(dvp); 459 if (vp) { 460 vref(vp); 461 MARK_VNODE(vp); 462 } 463 464 MARK_VNODE(dvp); 465 return 0; 466 467 unreserve: 468 lfs_reserve(fs, dvp, vp, -LFS_NRESERVE(fs)); 469 return error; 470 } 471 472 /* 473 * Get a new vnode *before* adjusting the dirop count, to avoid a deadlock 474 * in getnewvnode(), if we have a stacked filesystem mounted on top 475 * of us. 476 * 477 * NB: this means we have to clear the new vnodes on error. Fortunately 478 * SET_ENDOP is there to do that for us. 479 */ 480 static int 481 lfs_set_dirop_create(struct vnode *dvp, struct vnode **vpp) 482 { 483 int error; 484 struct lfs *fs; 485 486 fs = VFSTOUFS(dvp->v_mount)->um_lfs; 487 ASSERT_NO_SEGLOCK(fs); 488 if (fs->lfs_ronly) 489 return EROFS; 490 if (vpp == NULL) { 491 return lfs_set_dirop(dvp, NULL); 492 } 493 error = getnewvnode(VT_LFS, dvp->v_mount, lfs_vnodeop_p, NULL, vpp); 494 if (error) { 495 DLOG((DLOG_ALLOC, "lfs_set_dirop_create: dvp %p error %d\n", 496 dvp, error)); 497 return error; 498 } 499 if ((error = lfs_set_dirop(dvp, NULL)) != 0) { 500 ungetnewvnode(*vpp); 501 *vpp = NULL; 502 return error; 503 } 504 return 0; 505 } 506 507 #define SET_ENDOP_BASE(fs, dvp, str) \ 508 do { \ 509 mutex_enter(&lfs_lock); \ 510 --(fs)->lfs_dirops; \ 511 if (!(fs)->lfs_dirops) { \ 512 if ((fs)->lfs_nadirop) { \ 513 panic("SET_ENDOP: %s: no dirops but " \ 514 " nadirop=%d", (str), \ 515 (fs)->lfs_nadirop); \ 516 } \ 517 wakeup(&(fs)->lfs_writer); \ 518 mutex_exit(&lfs_lock); \ 519 lfs_check((dvp), LFS_UNUSED_LBN, 0); \ 520 } else \ 521 mutex_exit(&lfs_lock); \ 522 } while(0) 523 #define SET_ENDOP_CREATE(fs, dvp, nvpp, str) \ 524 do { \ 525 UNMARK_VNODE(dvp); \ 526 if (nvpp && *nvpp) \ 527 UNMARK_VNODE(*nvpp); \ 528 /* Check for error return to stem vnode leakage */ \ 529 if (nvpp && *nvpp && !((*nvpp)->v_uflag & VU_DIROP)) \ 530 ungetnewvnode(*(nvpp)); \ 531 SET_ENDOP_BASE((fs), (dvp), (str)); \ 532 lfs_reserve((fs), (dvp), NULL, -LFS_NRESERVE(fs)); \ 533 vrele(dvp); \ 534 } while(0) 535 #define SET_ENDOP_CREATE_AP(ap, str) \ 536 SET_ENDOP_CREATE(VTOI((ap)->a_dvp)->i_lfs, (ap)->a_dvp, \ 537 (ap)->a_vpp, (str)) 538 #define SET_ENDOP_REMOVE(fs, dvp, ovp, str) \ 539 do { \ 540 UNMARK_VNODE(dvp); \ 541 if (ovp) \ 542 UNMARK_VNODE(ovp); \ 543 SET_ENDOP_BASE((fs), (dvp), (str)); \ 544 lfs_reserve((fs), (dvp), (ovp), -LFS_NRESERVE(fs)); \ 545 vrele(dvp); \ 546 if (ovp) \ 547 vrele(ovp); \ 548 } while(0) 549 550 void 551 lfs_mark_vnode(struct vnode *vp) 552 { 553 struct inode *ip = VTOI(vp); 554 struct lfs *fs = ip->i_lfs; 555 556 mutex_enter(&lfs_lock); 557 if (!(ip->i_flag & IN_ADIROP)) { 558 if (!(vp->v_uflag & VU_DIROP)) { 559 mutex_exit(&lfs_lock); 560 mutex_enter(vp->v_interlock); 561 if (lfs_vref(vp) != 0) 562 panic("lfs_mark_vnode: could not vref"); 563 mutex_enter(&lfs_lock); 564 ++lfs_dirvcount; 565 ++fs->lfs_dirvcount; 566 TAILQ_INSERT_TAIL(&fs->lfs_dchainhd, ip, i_lfs_dchain); 567 vp->v_uflag |= VU_DIROP; 568 } 569 ++fs->lfs_nadirop; 570 ip->i_flag &= ~IN_CDIROP; 571 ip->i_flag |= IN_ADIROP; 572 } else 573 KASSERT(vp->v_uflag & VU_DIROP); 574 mutex_exit(&lfs_lock); 575 } 576 577 void 578 lfs_unmark_vnode(struct vnode *vp) 579 { 580 struct inode *ip = VTOI(vp); 581 582 mutex_enter(&lfs_lock); 583 if (ip && (ip->i_flag & IN_ADIROP)) { 584 KASSERT(vp->v_uflag & VU_DIROP); 585 --ip->i_lfs->lfs_nadirop; 586 ip->i_flag &= ~IN_ADIROP; 587 } 588 mutex_exit(&lfs_lock); 589 } 590 591 int 592 lfs_symlink(void *v) 593 { 594 struct vop_symlink_args /* { 595 struct vnode *a_dvp; 596 struct vnode **a_vpp; 597 struct componentname *a_cnp; 598 struct vattr *a_vap; 599 char *a_target; 600 } */ *ap = v; 601 int error; 602 603 if ((error = SET_DIROP_CREATE(ap->a_dvp, ap->a_vpp)) != 0) { 604 vput(ap->a_dvp); 605 return error; 606 } 607 error = ufs_symlink(ap); 608 SET_ENDOP_CREATE_AP(ap, "symlink"); 609 return (error); 610 } 611 612 int 613 lfs_mknod(void *v) 614 { 615 struct vop_mknod_args /* { 616 struct vnode *a_dvp; 617 struct vnode **a_vpp; 618 struct componentname *a_cnp; 619 struct vattr *a_vap; 620 } */ *ap = v; 621 struct vattr *vap = ap->a_vap; 622 struct vnode **vpp = ap->a_vpp; 623 struct inode *ip; 624 int error; 625 struct mount *mp; 626 ino_t ino; 627 struct ufs_lookup_results *ulr; 628 629 /* XXX should handle this material another way */ 630 ulr = &VTOI(ap->a_dvp)->i_crap; 631 UFS_CHECK_CRAPCOUNTER(VTOI(ap->a_dvp)); 632 633 if ((error = SET_DIROP_CREATE(ap->a_dvp, ap->a_vpp)) != 0) { 634 vput(ap->a_dvp); 635 return error; 636 } 637 error = ufs_makeinode(MAKEIMODE(vap->va_type, vap->va_mode), 638 ap->a_dvp, ulr, vpp, ap->a_cnp); 639 640 /* Either way we're done with the dirop at this point */ 641 SET_ENDOP_CREATE_AP(ap, "mknod"); 642 643 if (error) 644 return (error); 645 646 ip = VTOI(*vpp); 647 mp = (*vpp)->v_mount; 648 ino = ip->i_number; 649 ip->i_flag |= IN_ACCESS | IN_CHANGE | IN_UPDATE; 650 if (vap->va_rdev != VNOVAL) { 651 /* 652 * Want to be able to use this to make badblock 653 * inodes, so don't truncate the dev number. 654 */ 655 #if 0 656 ip->i_ffs1_rdev = ufs_rw32(vap->va_rdev, 657 UFS_MPNEEDSWAP((*vpp)->v_mount)); 658 #else 659 ip->i_ffs1_rdev = vap->va_rdev; 660 #endif 661 } 662 663 /* 664 * Call fsync to write the vnode so that we don't have to deal with 665 * flushing it when it's marked VU_DIROP|VI_XLOCK. 666 * 667 * XXX KS - If we can't flush we also can't call vgone(), so must 668 * return. But, that leaves this vnode in limbo, also not good. 669 * Can this ever happen (barring hardware failure)? 670 */ 671 if ((error = VOP_FSYNC(*vpp, NOCRED, FSYNC_WAIT, 0, 0)) != 0) { 672 panic("lfs_mknod: couldn't fsync (ino %llu)", 673 (unsigned long long)ino); 674 /* return (error); */ 675 } 676 /* 677 * Remove vnode so that it will be reloaded by VFS_VGET and 678 * checked to see if it is an alias of an existing entry in 679 * the inode cache. 680 */ 681 /* Used to be vput, but that causes us to call VOP_INACTIVE twice. */ 682 683 VOP_UNLOCK(*vpp); 684 (*vpp)->v_type = VNON; 685 vgone(*vpp); 686 error = VFS_VGET(mp, ino, vpp); 687 688 if (error != 0) { 689 *vpp = NULL; 690 return (error); 691 } 692 return (0); 693 } 694 695 int 696 lfs_create(void *v) 697 { 698 struct vop_create_args /* { 699 struct vnode *a_dvp; 700 struct vnode **a_vpp; 701 struct componentname *a_cnp; 702 struct vattr *a_vap; 703 } */ *ap = v; 704 int error; 705 706 if ((error = SET_DIROP_CREATE(ap->a_dvp, ap->a_vpp)) != 0) { 707 vput(ap->a_dvp); 708 return error; 709 } 710 error = ufs_create(ap); 711 SET_ENDOP_CREATE_AP(ap, "create"); 712 return (error); 713 } 714 715 int 716 lfs_mkdir(void *v) 717 { 718 struct vop_mkdir_args /* { 719 struct vnode *a_dvp; 720 struct vnode **a_vpp; 721 struct componentname *a_cnp; 722 struct vattr *a_vap; 723 } */ *ap = v; 724 int error; 725 726 if ((error = SET_DIROP_CREATE(ap->a_dvp, ap->a_vpp)) != 0) { 727 vput(ap->a_dvp); 728 return error; 729 } 730 error = ufs_mkdir(ap); 731 SET_ENDOP_CREATE_AP(ap, "mkdir"); 732 return (error); 733 } 734 735 int 736 lfs_remove(void *v) 737 { 738 struct vop_remove_args /* { 739 struct vnode *a_dvp; 740 struct vnode *a_vp; 741 struct componentname *a_cnp; 742 } */ *ap = v; 743 struct vnode *dvp, *vp; 744 struct inode *ip; 745 int error; 746 747 dvp = ap->a_dvp; 748 vp = ap->a_vp; 749 ip = VTOI(vp); 750 if ((error = SET_DIROP_REMOVE(dvp, vp)) != 0) { 751 if (dvp == vp) 752 vrele(vp); 753 else 754 vput(vp); 755 vput(dvp); 756 return error; 757 } 758 error = ufs_remove(ap); 759 if (ip->i_nlink == 0) 760 lfs_orphan(ip->i_lfs, ip->i_number); 761 SET_ENDOP_REMOVE(ip->i_lfs, dvp, ap->a_vp, "remove"); 762 return (error); 763 } 764 765 int 766 lfs_rmdir(void *v) 767 { 768 struct vop_rmdir_args /* { 769 struct vnodeop_desc *a_desc; 770 struct vnode *a_dvp; 771 struct vnode *a_vp; 772 struct componentname *a_cnp; 773 } */ *ap = v; 774 struct vnode *vp; 775 struct inode *ip; 776 int error; 777 778 vp = ap->a_vp; 779 ip = VTOI(vp); 780 if ((error = SET_DIROP_REMOVE(ap->a_dvp, ap->a_vp)) != 0) { 781 if (ap->a_dvp == vp) 782 vrele(ap->a_dvp); 783 else 784 vput(ap->a_dvp); 785 vput(vp); 786 return error; 787 } 788 error = ufs_rmdir(ap); 789 if (ip->i_nlink == 0) 790 lfs_orphan(ip->i_lfs, ip->i_number); 791 SET_ENDOP_REMOVE(ip->i_lfs, ap->a_dvp, ap->a_vp, "rmdir"); 792 return (error); 793 } 794 795 int 796 lfs_link(void *v) 797 { 798 struct vop_link_args /* { 799 struct vnode *a_dvp; 800 struct vnode *a_vp; 801 struct componentname *a_cnp; 802 } */ *ap = v; 803 int error; 804 struct vnode **vpp = NULL; 805 806 if ((error = SET_DIROP_CREATE(ap->a_dvp, vpp)) != 0) { 807 vput(ap->a_dvp); 808 return error; 809 } 810 error = ufs_link(ap); 811 SET_ENDOP_CREATE(VTOI(ap->a_dvp)->i_lfs, ap->a_dvp, vpp, "link"); 812 return (error); 813 } 814 815 /* XXX following lifted from ufs_lookup.c */ 816 #define FSFMT(vp) (((vp)->v_mount->mnt_iflag & IMNT_DTYPE) == 0) 817 818 /* 819 * Check if either entry referred to by FROM_ULR is within the range 820 * of entries named by TO_ULR. 821 */ 822 static int 823 ulr_overlap(const struct ufs_lookup_results *from_ulr, 824 const struct ufs_lookup_results *to_ulr) 825 { 826 doff_t from_start, from_prevstart; 827 doff_t to_start, to_end; 828 829 /* 830 * FROM is a DELETE result; offset points to the entry to 831 * remove and subtracting count gives the previous entry. 832 */ 833 from_start = from_ulr->ulr_offset - from_ulr->ulr_count; 834 from_prevstart = from_ulr->ulr_offset; 835 836 /* 837 * TO is a RENAME (thus non-DELETE) result; offset points 838 * to the beginning of a region to write in, and adding 839 * count gives the end of the region. 840 */ 841 to_start = to_ulr->ulr_offset; 842 to_end = to_ulr->ulr_offset + to_ulr->ulr_count; 843 844 if (from_prevstart >= to_start && from_prevstart < to_end) { 845 return 1; 846 } 847 if (from_start >= to_start && from_start < to_end) { 848 return 1; 849 } 850 return 0; 851 } 852 853 /* 854 * A virgin directory (no blushing please). 855 */ 856 static const struct dirtemplate mastertemplate = { 857 0, 12, DT_DIR, 1, ".", 858 0, DIRBLKSIZ - 12, DT_DIR, 2, ".." 859 }; 860 861 /* 862 * Wrapper for relookup that also updates the supplemental results. 863 */ 864 static int 865 do_relookup(struct vnode *dvp, struct ufs_lookup_results *ulr, 866 struct vnode **vp, struct componentname *cnp) 867 { 868 int error; 869 870 error = relookup(dvp, vp, cnp, 0); 871 if (error) { 872 return error; 873 } 874 /* update the supplemental reasults */ 875 *ulr = VTOI(dvp)->i_crap; 876 UFS_CHECK_CRAPCOUNTER(VTOI(dvp)); 877 return 0; 878 } 879 880 /* 881 * Lock and relookup a sequence of two directories and two children. 882 * 883 */ 884 static int 885 lock_vnode_sequence(struct vnode *d1, struct ufs_lookup_results *ulr1, 886 struct vnode **v1_ret, struct componentname *cn1, 887 int v1_missing_ok, 888 int overlap_error, 889 struct vnode *d2, struct ufs_lookup_results *ulr2, 890 struct vnode **v2_ret, struct componentname *cn2, 891 int v2_missing_ok) 892 { 893 struct vnode *v1, *v2; 894 int error; 895 896 KASSERT(d1 != d2); 897 898 vn_lock(d1, LK_EXCLUSIVE | LK_RETRY); 899 if (VTOI(d1)->i_size == 0) { 900 /* d1 has been rmdir'd */ 901 VOP_UNLOCK(d1); 902 return ENOENT; 903 } 904 error = do_relookup(d1, ulr1, &v1, cn1); 905 if (v1_missing_ok) { 906 if (error == ENOENT) { 907 /* 908 * Note: currently if the name doesn't exist, 909 * relookup succeeds (it intercepts the 910 * EJUSTRETURN from VOP_LOOKUP) and sets tvp 911 * to NULL. Therefore, we will never get 912 * ENOENT and this branch is not needed. 913 * However, in a saner future the EJUSTRETURN 914 * garbage will go away, so let's DTRT. 915 */ 916 v1 = NULL; 917 error = 0; 918 } 919 } else { 920 if (error == 0 && v1 == NULL) { 921 /* This is what relookup sets if v1 disappeared. */ 922 error = ENOENT; 923 } 924 } 925 if (error) { 926 VOP_UNLOCK(d1); 927 return error; 928 } 929 if (v1 && v1 == d2) { 930 VOP_UNLOCK(d1); 931 VOP_UNLOCK(v1); 932 vrele(v1); 933 return overlap_error; 934 } 935 936 /* 937 * The right way to do this is to do lookups without locking 938 * the results, and lock the results afterwards; then at the 939 * end we can avoid trying to lock v2 if v2 == v1. 940 * 941 * However, for the reasons described in the fdvp == tdvp case 942 * in rename below, we can't do that safely. So, in the case 943 * where v1 is not a directory, unlock it and lock it again 944 * afterwards. This is safe in locking order because a 945 * non-directory can't be above anything else in the tree. If 946 * v1 *is* a directory, that's not true, but then because d1 947 * != d2, v1 != v2. 948 */ 949 if (v1 && v1->v_type != VDIR) { 950 VOP_UNLOCK(v1); 951 } 952 vn_lock(d2, LK_EXCLUSIVE | LK_RETRY); 953 if (VTOI(d2)->i_size == 0) { 954 /* d2 has been rmdir'd */ 955 VOP_UNLOCK(d2); 956 if (v1 && v1->v_type == VDIR) { 957 VOP_UNLOCK(v1); 958 } 959 VOP_UNLOCK(d1); 960 if (v1) { 961 vrele(v1); 962 } 963 return ENOENT; 964 } 965 error = do_relookup(d2, ulr2, &v2, cn2); 966 if (v2_missing_ok) { 967 if (error == ENOENT) { 968 /* as above */ 969 v2 = NULL; 970 error = 0; 971 } 972 } else { 973 if (error == 0 && v2 == NULL) { 974 /* This is what relookup sets if v2 disappeared. */ 975 error = ENOENT; 976 } 977 } 978 if (error) { 979 VOP_UNLOCK(d2); 980 if (v1 && v1->v_type == VDIR) { 981 VOP_UNLOCK(v1); 982 } 983 VOP_UNLOCK(d1); 984 if (v1) { 985 vrele(v1); 986 } 987 return error; 988 } 989 if (v1 && v1->v_type != VDIR && v1 != v2) { 990 vn_lock(v1, LK_EXCLUSIVE | LK_RETRY); 991 } 992 *v1_ret = v1; 993 *v2_ret = v2; 994 return 0; 995 } 996 997 int 998 lfs_rename(void *v) 999 { 1000 struct vop_rename_args /* { 1001 struct vnode *a_fdvp; 1002 struct vnode *a_fvp; 1003 struct componentname *a_fcnp; 1004 struct vnode *a_tdvp; 1005 struct vnode *a_tvp; 1006 struct componentname *a_tcnp; 1007 } */ *ap = v; 1008 struct vnode *tvp, *tdvp, *fvp, *fdvp; 1009 struct componentname *tcnp, *fcnp; 1010 struct inode *ip, *txp, *fxp, *tdp, *fdp; 1011 struct mount *mp; 1012 struct direct *newdir; 1013 int doingdirectory, error, marked; 1014 ino_t oldparent, newparent; 1015 1016 struct ufs_lookup_results from_ulr, to_ulr; 1017 struct lfs *fs = VTOI(ap->a_fvp)->i_lfs; 1018 1019 tvp = ap->a_tvp; 1020 tdvp = ap->a_tdvp; 1021 fvp = ap->a_fvp; 1022 fdvp = ap->a_fdvp; 1023 tcnp = ap->a_tcnp; 1024 fcnp = ap->a_fcnp; 1025 doingdirectory = error = 0; 1026 oldparent = newparent = 0; 1027 marked = 0; 1028 1029 /* save the supplemental lookup results as they currently exist */ 1030 from_ulr = VTOI(fdvp)->i_crap; 1031 to_ulr = VTOI(tdvp)->i_crap; 1032 UFS_CHECK_CRAPCOUNTER(VTOI(fdvp)); 1033 UFS_CHECK_CRAPCOUNTER(VTOI(tdvp)); 1034 1035 /* 1036 * Owing to VFS oddities we are currently called with tdvp/tvp 1037 * locked and not fdvp/fvp. In a sane world we'd be passed 1038 * tdvp and fdvp only, unlocked, and two name strings. Pretend 1039 * we have a sane world and unlock tdvp and tvp. 1040 */ 1041 VOP_UNLOCK(tdvp); 1042 if (tvp && tvp != tdvp) { 1043 VOP_UNLOCK(tvp); 1044 } 1045 1046 /* Also pretend we have a sane world and vrele fvp/tvp. */ 1047 vrele(fvp); 1048 fvp = NULL; 1049 if (tvp) { 1050 vrele(tvp); 1051 tvp = NULL; 1052 } 1053 1054 /* 1055 * Check for cross-device rename. 1056 */ 1057 if (fdvp->v_mount != tdvp->v_mount) { 1058 error = EXDEV; 1059 goto abort; 1060 } 1061 1062 /* 1063 * Reject "." and ".." 1064 */ 1065 if ((fcnp->cn_flags & ISDOTDOT) || (tcnp->cn_flags & ISDOTDOT) || 1066 (fcnp->cn_namelen == 1 && fcnp->cn_nameptr[0] == '.') || 1067 (tcnp->cn_namelen == 1 && tcnp->cn_nameptr[0] == '.')) { 1068 error = EINVAL; 1069 goto abort; 1070 } 1071 1072 /* 1073 * Get locks. 1074 */ 1075 1076 /* paranoia */ 1077 fcnp->cn_flags |= LOCKPARENT|LOCKLEAF; 1078 tcnp->cn_flags |= LOCKPARENT|LOCKLEAF; 1079 1080 if (fdvp == tdvp) { 1081 /* One directory. Lock it and relookup both children. */ 1082 vn_lock(fdvp, LK_EXCLUSIVE | LK_RETRY); 1083 1084 if (VTOI(fdvp)->i_size == 0) { 1085 /* directory has been rmdir'd */ 1086 VOP_UNLOCK(fdvp); 1087 error = ENOENT; 1088 goto abort; 1089 } 1090 1091 error = do_relookup(fdvp, &from_ulr, &fvp, fcnp); 1092 if (error == 0 && fvp == NULL) { 1093 /* relookup may produce this if fvp disappears */ 1094 error = ENOENT; 1095 } 1096 if (error) { 1097 VOP_UNLOCK(fdvp); 1098 goto abort; 1099 } 1100 1101 /* 1102 * The right way to do this is to look up both children 1103 * without locking either, and then lock both unless they 1104 * turn out to be the same. However, due to deep-seated 1105 * VFS-level issues all lookups lock the child regardless 1106 * of whether LOCKLEAF is set (if LOCKLEAF is not set, 1107 * the child is locked during lookup and then unlocked) 1108 * so it is not safe to look up tvp while fvp is locked. 1109 * 1110 * Unlocking fvp here temporarily is more or less safe, 1111 * because with the directory locked there's not much 1112 * that can happen to it. However, ideally it wouldn't 1113 * be necessary. XXX. 1114 */ 1115 VOP_UNLOCK(fvp); 1116 /* remember fdvp == tdvp so tdvp is locked */ 1117 error = do_relookup(tdvp, &to_ulr, &tvp, tcnp); 1118 if (error && error != ENOENT) { 1119 VOP_UNLOCK(fdvp); 1120 goto abort; 1121 } 1122 if (error == ENOENT) { 1123 /* 1124 * Note: currently if the name doesn't exist, 1125 * relookup succeeds (it intercepts the 1126 * EJUSTRETURN from VOP_LOOKUP) and sets tvp 1127 * to NULL. Therefore, we will never get 1128 * ENOENT and this branch is not needed. 1129 * However, in a saner future the EJUSTRETURN 1130 * garbage will go away, so let's DTRT. 1131 */ 1132 tvp = NULL; 1133 } 1134 1135 /* tvp is locked; lock fvp if necessary */ 1136 if (!tvp || tvp != fvp) { 1137 vn_lock(fvp, LK_EXCLUSIVE | LK_RETRY); 1138 } 1139 } else { 1140 int found_fdvp; 1141 struct vnode *illegal_fvp; 1142 1143 /* 1144 * The source must not be above the destination. (If 1145 * it were, the rename would detach a section of the 1146 * tree.) 1147 * 1148 * Look up the tree from tdvp to see if we find fdvp, 1149 * and if so, return the immediate child of fdvp we're 1150 * under; that must not turn out to be the same as 1151 * fvp. 1152 * 1153 * The per-volume rename lock guarantees that the 1154 * result of this check remains true until we finish 1155 * looking up and locking. 1156 */ 1157 error = ufs_parentcheck(fdvp, tdvp, fcnp->cn_cred, 1158 &found_fdvp, &illegal_fvp); 1159 if (error) { 1160 goto abort; 1161 } 1162 1163 /* Must lock in tree order. */ 1164 1165 if (found_fdvp) { 1166 /* fdvp -> fvp -> tdvp -> tvp */ 1167 error = lock_vnode_sequence(fdvp, &from_ulr, 1168 &fvp, fcnp, 0, 1169 EINVAL, 1170 tdvp, &to_ulr, 1171 &tvp, tcnp, 1); 1172 } else { 1173 /* tdvp -> tvp -> fdvp -> fvp */ 1174 error = lock_vnode_sequence(tdvp, &to_ulr, 1175 &tvp, tcnp, 1, 1176 ENOTEMPTY, 1177 fdvp, &from_ulr, 1178 &fvp, fcnp, 0); 1179 } 1180 if (error) { 1181 if (illegal_fvp) { 1182 vrele(illegal_fvp); 1183 } 1184 goto abort; 1185 } 1186 KASSERT(fvp != NULL); 1187 1188 if (illegal_fvp && fvp == illegal_fvp) { 1189 vrele(illegal_fvp); 1190 error = EINVAL; 1191 goto abort_withlocks; 1192 } 1193 1194 if (illegal_fvp) { 1195 vrele(illegal_fvp); 1196 } 1197 } 1198 1199 KASSERT(fdvp && VOP_ISLOCKED(fdvp)); 1200 KASSERT(fvp && VOP_ISLOCKED(fvp)); 1201 KASSERT(tdvp && VOP_ISLOCKED(tdvp)); 1202 KASSERT(tvp == NULL || VOP_ISLOCKED(tvp)); 1203 1204 /* --- everything is now locked --- */ 1205 1206 if (tvp && ((VTOI(tvp)->i_flags & (IMMUTABLE | APPEND)) || 1207 (VTOI(tdvp)->i_flags & APPEND))) { 1208 error = EPERM; 1209 goto abort_withlocks; 1210 } 1211 1212 /* 1213 * Check if just deleting a link name. 1214 */ 1215 if (fvp == tvp) { 1216 if (fvp->v_type == VDIR) { 1217 error = EINVAL; 1218 goto abort_withlocks; 1219 } 1220 1221 /* Release destination completely. Leave fdvp locked. */ 1222 VOP_ABORTOP(tdvp, tcnp); 1223 if (fdvp != tdvp) { 1224 VOP_UNLOCK(tdvp); 1225 } 1226 VOP_UNLOCK(tvp); 1227 vrele(tdvp); 1228 vrele(tvp); 1229 1230 /* Delete source. */ 1231 /* XXX: do we really need to relookup again? */ 1232 1233 /* 1234 * fdvp is still locked, but we just unlocked fvp 1235 * (because fvp == tvp) so just decref fvp 1236 */ 1237 vrele(fvp); 1238 fcnp->cn_flags &= ~(MODMASK); 1239 fcnp->cn_flags |= LOCKPARENT | LOCKLEAF; 1240 fcnp->cn_nameiop = DELETE; 1241 if ((error = relookup(fdvp, &fvp, fcnp, 0))) { 1242 vput(fdvp); 1243 return (error); 1244 } 1245 return (VOP_REMOVE(fdvp, fvp, fcnp)); 1246 } 1247 1248 /* The tiny bit of actual LFS code in this function */ 1249 if ((error = SET_DIROP_REMOVE(tdvp, tvp)) != 0) 1250 goto abort_withlocks; 1251 MARK_VNODE(fdvp); 1252 MARK_VNODE(fvp); 1253 marked = 1; 1254 1255 fdp = VTOI(fdvp); 1256 ip = VTOI(fvp); 1257 if ((nlink_t) ip->i_nlink >= LINK_MAX) { 1258 error = EMLINK; 1259 goto abort_withlocks; 1260 } 1261 if ((ip->i_flags & (IMMUTABLE | APPEND)) || 1262 (fdp->i_flags & APPEND)) { 1263 error = EPERM; 1264 goto abort_withlocks; 1265 } 1266 if ((ip->i_mode & IFMT) == IFDIR) { 1267 /* 1268 * Avoid ".", "..", and aliases of "." for obvious reasons. 1269 */ 1270 if ((fcnp->cn_namelen == 1 && fcnp->cn_nameptr[0] == '.') || 1271 fdp == ip || 1272 (fcnp->cn_flags & ISDOTDOT) || 1273 (tcnp->cn_flags & ISDOTDOT) || 1274 (ip->i_flag & IN_RENAME)) { 1275 error = EINVAL; 1276 goto abort_withlocks; 1277 } 1278 ip->i_flag |= IN_RENAME; 1279 doingdirectory = 1; 1280 } 1281 oldparent = fdp->i_number; 1282 VN_KNOTE(fdvp, NOTE_WRITE); /* XXXLUKEM/XXX: right place? */ 1283 1284 /* 1285 * Both the directory 1286 * and target vnodes are locked. 1287 */ 1288 tdp = VTOI(tdvp); 1289 txp = NULL; 1290 if (tvp) 1291 txp = VTOI(tvp); 1292 1293 mp = fdvp->v_mount; 1294 fstrans_start(mp, FSTRANS_SHARED); 1295 1296 if (oldparent != tdp->i_number) 1297 newparent = tdp->i_number; 1298 1299 /* 1300 * If ".." must be changed (ie the directory gets a new 1301 * parent) the user must have write permission in the source 1302 * so as to be able to change "..". 1303 */ 1304 if (doingdirectory && newparent) { 1305 error = VOP_ACCESS(fvp, VWRITE, tcnp->cn_cred); 1306 if (error) 1307 goto out; 1308 } 1309 1310 KASSERT(fdvp != tvp); 1311 1312 if (newparent) { 1313 /* Check for the rename("foo/foo", "foo") case. */ 1314 if (fdvp == tvp) { 1315 error = doingdirectory ? ENOTEMPTY : EISDIR; 1316 goto out; 1317 } 1318 } 1319 1320 fxp = VTOI(fvp); 1321 fdp = VTOI(fdvp); 1322 1323 error = UFS_WAPBL_BEGIN(fdvp->v_mount); 1324 if (error) 1325 goto out2; 1326 1327 /* 1328 * 1) Bump link count while we're moving stuff 1329 * around. If we crash somewhere before 1330 * completing our work, the link count 1331 * may be wrong, but correctable. 1332 */ 1333 ip->i_nlink++; 1334 DIP_ASSIGN(ip, nlink, ip->i_nlink); 1335 ip->i_flag |= IN_CHANGE; 1336 if ((error = UFS_UPDATE(fvp, NULL, NULL, UPDATE_DIROP)) != 0) { 1337 goto bad; 1338 } 1339 1340 /* 1341 * 2) If target doesn't exist, link the target 1342 * to the source and unlink the source. 1343 * Otherwise, rewrite the target directory 1344 * entry to reference the source inode and 1345 * expunge the original entry's existence. 1346 */ 1347 if (txp == NULL) { 1348 if (tdp->i_dev != ip->i_dev) 1349 panic("rename: EXDEV"); 1350 /* 1351 * Account for ".." in new directory. 1352 * When source and destination have the same 1353 * parent we don't fool with the link count. 1354 */ 1355 if (doingdirectory && newparent) { 1356 if ((nlink_t)tdp->i_nlink >= LINK_MAX) { 1357 error = EMLINK; 1358 goto bad; 1359 } 1360 tdp->i_nlink++; 1361 DIP_ASSIGN(tdp, nlink, tdp->i_nlink); 1362 tdp->i_flag |= IN_CHANGE; 1363 if ((error = UFS_UPDATE(tdvp, NULL, NULL, 1364 UPDATE_DIROP)) != 0) { 1365 tdp->i_nlink--; 1366 DIP_ASSIGN(tdp, nlink, tdp->i_nlink); 1367 tdp->i_flag |= IN_CHANGE; 1368 goto bad; 1369 } 1370 } 1371 newdir = pool_cache_get(ufs_direct_cache, PR_WAITOK); 1372 ufs_makedirentry(ip, tcnp, newdir); 1373 error = ufs_direnter(tdvp, &to_ulr, 1374 NULL, newdir, tcnp, NULL); 1375 pool_cache_put(ufs_direct_cache, newdir); 1376 if (error != 0) { 1377 if (doingdirectory && newparent) { 1378 tdp->i_nlink--; 1379 DIP_ASSIGN(tdp, nlink, tdp->i_nlink); 1380 tdp->i_flag |= IN_CHANGE; 1381 (void)UFS_UPDATE(tdvp, NULL, NULL, 1382 UPDATE_WAIT | UPDATE_DIROP); 1383 } 1384 goto bad; 1385 } 1386 VN_KNOTE(tdvp, NOTE_WRITE); 1387 } else { 1388 if (txp->i_dev != tdp->i_dev || txp->i_dev != ip->i_dev) 1389 panic("rename: EXDEV"); 1390 /* 1391 * Short circuit rename(foo, foo). 1392 */ 1393 if (txp->i_number == ip->i_number) 1394 panic("rename: same file"); 1395 /* 1396 * If the parent directory is "sticky", then the user must 1397 * own the parent directory, or the destination of the rename, 1398 * otherwise the destination may not be changed (except by 1399 * root). This implements append-only directories. 1400 */ 1401 if ((tdp->i_mode & S_ISTXT) && 1402 kauth_authorize_generic(tcnp->cn_cred, 1403 KAUTH_GENERIC_ISSUSER, NULL) != 0 && 1404 kauth_cred_geteuid(tcnp->cn_cred) != tdp->i_uid && 1405 txp->i_uid != kauth_cred_geteuid(tcnp->cn_cred)) { 1406 error = EPERM; 1407 goto bad; 1408 } 1409 /* 1410 * Target must be empty if a directory and have no links 1411 * to it. Also, ensure source and target are compatible 1412 * (both directories, or both not directories). 1413 */ 1414 if ((txp->i_mode & IFMT) == IFDIR) { 1415 if (txp->i_nlink > 2 || 1416 !ufs_dirempty(txp, tdp->i_number, tcnp->cn_cred)) { 1417 error = ENOTEMPTY; 1418 goto bad; 1419 } 1420 if (!doingdirectory) { 1421 error = ENOTDIR; 1422 goto bad; 1423 } 1424 cache_purge(tdvp); 1425 } else if (doingdirectory) { 1426 error = EISDIR; 1427 goto bad; 1428 } 1429 if ((error = ufs_dirrewrite(tdp, to_ulr.ulr_offset, 1430 txp, ip->i_number, 1431 IFTODT(ip->i_mode), doingdirectory && newparent ? 1432 newparent : doingdirectory, IN_CHANGE | IN_UPDATE)) != 0) 1433 goto bad; 1434 if (doingdirectory) { 1435 /* 1436 * Truncate inode. The only stuff left in the directory 1437 * is "." and "..". The "." reference is inconsequential 1438 * since we are quashing it. We have removed the "." 1439 * reference and the reference in the parent directory, 1440 * but there may be other hard links. 1441 */ 1442 if (!newparent) { 1443 tdp->i_nlink--; 1444 DIP_ASSIGN(tdp, nlink, tdp->i_nlink); 1445 tdp->i_flag |= IN_CHANGE; 1446 UFS_WAPBL_UPDATE(tdvp, NULL, NULL, 0); 1447 } 1448 txp->i_nlink--; 1449 DIP_ASSIGN(txp, nlink, txp->i_nlink); 1450 txp->i_flag |= IN_CHANGE; 1451 if ((error = UFS_TRUNCATE(tvp, (off_t)0, IO_SYNC, 1452 tcnp->cn_cred))) 1453 goto bad; 1454 } 1455 VN_KNOTE(tdvp, NOTE_WRITE); 1456 VN_KNOTE(tvp, NOTE_DELETE); 1457 } 1458 1459 /* 1460 * Handle case where the directory entry we need to remove, 1461 * which is/was at from_ulr.ulr_offset, or the one before it, 1462 * which is/was at from_ulr.ulr_offset - from_ulr.ulr_count, 1463 * may have been moved when the directory insertion above 1464 * performed compaction. 1465 */ 1466 if (tdp->i_number == fdp->i_number && 1467 ulr_overlap(&from_ulr, &to_ulr)) { 1468 1469 struct buf *bp; 1470 struct direct *ep; 1471 struct ufsmount *ump = fdp->i_ump; 1472 doff_t curpos; 1473 doff_t endsearch; /* offset to end directory search */ 1474 uint32_t prev_reclen; 1475 int dirblksiz = ump->um_dirblksiz; 1476 const int needswap = UFS_MPNEEDSWAP(ump); 1477 u_long bmask; 1478 int namlen, entryoffsetinblock; 1479 char *dirbuf; 1480 1481 bmask = fdvp->v_mount->mnt_stat.f_iosize - 1; 1482 1483 /* 1484 * The fcnp entry will be somewhere between the start of 1485 * compaction (to_ulr.ulr_offset) and the original location 1486 * (from_ulr.ulr_offset). 1487 */ 1488 curpos = to_ulr.ulr_offset; 1489 endsearch = from_ulr.ulr_offset + from_ulr.ulr_reclen; 1490 entryoffsetinblock = 0; 1491 1492 /* 1493 * Get the directory block containing the start of 1494 * compaction. 1495 */ 1496 error = ufs_blkatoff(fdvp, (off_t)to_ulr.ulr_offset, &dirbuf, 1497 &bp, false); 1498 if (error) 1499 goto bad; 1500 1501 /* 1502 * Keep existing ulr_count (length of previous record) 1503 * for the case where compaction did not include the 1504 * previous entry but started at the from-entry. 1505 */ 1506 prev_reclen = from_ulr.ulr_count; 1507 1508 while (curpos < endsearch) { 1509 uint32_t reclen; 1510 1511 /* 1512 * If necessary, get the next directory block. 1513 * 1514 * dholland 7/13/11 to the best of my understanding 1515 * this should never happen; compaction occurs only 1516 * within single blocks. I think. 1517 */ 1518 if ((curpos & bmask) == 0) { 1519 if (bp != NULL) 1520 brelse(bp, 0); 1521 error = ufs_blkatoff(fdvp, (off_t)curpos, 1522 &dirbuf, &bp, false); 1523 if (error) 1524 goto bad; 1525 entryoffsetinblock = 0; 1526 } 1527 1528 KASSERT(bp != NULL); 1529 ep = (struct direct *)(dirbuf + entryoffsetinblock); 1530 reclen = ufs_rw16(ep->d_reclen, needswap); 1531 1532 #if (BYTE_ORDER == LITTLE_ENDIAN) 1533 if (FSFMT(fdvp) && needswap == 0) 1534 namlen = ep->d_type; 1535 else 1536 namlen = ep->d_namlen; 1537 #else 1538 if (FSFMT(fdvp) && needswap != 0) 1539 namlen = ep->d_type; 1540 else 1541 namlen = ep->d_namlen; 1542 #endif 1543 if ((ep->d_ino != 0) && 1544 (ufs_rw32(ep->d_ino, needswap) != WINO) && 1545 (namlen == fcnp->cn_namelen) && 1546 memcmp(ep->d_name, fcnp->cn_nameptr, namlen) == 0) { 1547 from_ulr.ulr_reclen = reclen; 1548 break; 1549 } 1550 curpos += reclen; 1551 entryoffsetinblock += reclen; 1552 prev_reclen = reclen; 1553 } 1554 1555 from_ulr.ulr_offset = curpos; 1556 from_ulr.ulr_count = prev_reclen; 1557 1558 KASSERT(curpos <= endsearch); 1559 1560 /* 1561 * If ulr_offset points to start of a directory block, 1562 * clear ulr_count so ufs_dirremove() doesn't try to 1563 * merge free space over a directory block boundary. 1564 */ 1565 if ((from_ulr.ulr_offset & (dirblksiz - 1)) == 0) 1566 from_ulr.ulr_count = 0; 1567 1568 brelse(bp, 0); 1569 } 1570 1571 /* 1572 * 3) Unlink the source. 1573 */ 1574 1575 #if 0 1576 /* 1577 * Ensure that the directory entry still exists and has not 1578 * changed while the new name has been entered. If the source is 1579 * a file then the entry may have been unlinked or renamed. In 1580 * either case there is no further work to be done. If the source 1581 * is a directory then it cannot have been rmdir'ed; The IRENAME 1582 * flag ensures that it cannot be moved by another rename or removed 1583 * by a rmdir. 1584 */ 1585 #endif 1586 KASSERT(fxp == ip); 1587 1588 /* 1589 * If the source is a directory with a new parent, the link 1590 * count of the old parent directory must be decremented and 1591 * ".." set to point to the new parent. 1592 */ 1593 if (doingdirectory && newparent) { 1594 KASSERT(fdp != NULL); 1595 ufs_dirrewrite(fxp, mastertemplate.dot_reclen, 1596 fdp, newparent, DT_DIR, 0, IN_CHANGE); 1597 cache_purge(fdvp); 1598 } 1599 error = ufs_dirremove(fdvp, &from_ulr, 1600 fxp, fcnp->cn_flags, 0); 1601 fxp->i_flag &= ~IN_RENAME; 1602 1603 VN_KNOTE(fvp, NOTE_RENAME); 1604 goto done; 1605 1606 out: 1607 goto out2; 1608 1609 /* exit routines from steps 1 & 2 */ 1610 bad: 1611 if (doingdirectory) 1612 ip->i_flag &= ~IN_RENAME; 1613 ip->i_nlink--; 1614 DIP_ASSIGN(ip, nlink, ip->i_nlink); 1615 ip->i_flag |= IN_CHANGE; 1616 ip->i_flag &= ~IN_RENAME; 1617 UFS_WAPBL_UPDATE(fvp, NULL, NULL, 0); 1618 done: 1619 UFS_WAPBL_END(fdvp->v_mount); 1620 out2: 1621 /* 1622 * clear IN_RENAME - some exit paths happen too early to go 1623 * through the cleanup done in the "bad" case above, so we 1624 * always do this mini-cleanup here. 1625 */ 1626 ip->i_flag &= ~IN_RENAME; 1627 1628 VOP_UNLOCK(fdvp); 1629 if (tdvp != fdvp) { 1630 VOP_UNLOCK(tdvp); 1631 } 1632 VOP_UNLOCK(fvp); 1633 if (tvp && tvp != fvp) { 1634 VOP_UNLOCK(tvp); 1635 } 1636 1637 vrele(fdvp); 1638 vrele(tdvp); 1639 vrele(fvp); 1640 if (tvp) { 1641 vrele(tvp); 1642 } 1643 1644 fstrans_done(mp); 1645 if (marked) { 1646 UNMARK_VNODE(fdvp); 1647 UNMARK_VNODE(fvp); 1648 SET_ENDOP_REMOVE(fs, tdvp, tvp, "rename"); 1649 } 1650 return (error); 1651 1652 abort_withlocks: 1653 VOP_UNLOCK(fdvp); 1654 if (tdvp != fdvp) { 1655 VOP_UNLOCK(tdvp); 1656 } 1657 VOP_UNLOCK(fvp); 1658 if (tvp && tvp != fvp) { 1659 VOP_UNLOCK(tvp); 1660 } 1661 1662 abort: 1663 VOP_ABORTOP(fdvp, fcnp); /* XXX, why not in NFS? */ 1664 VOP_ABORTOP(tdvp, tcnp); /* XXX, why not in NFS? */ 1665 vrele(tdvp); 1666 if (tvp) { 1667 vrele(tvp); 1668 } 1669 vrele(fdvp); 1670 if (fvp) { 1671 vrele(fvp); 1672 } 1673 if (marked) { 1674 UNMARK_VNODE(fdvp); 1675 UNMARK_VNODE(fvp); 1676 SET_ENDOP_REMOVE(fs, tdvp, tvp, "rename"); 1677 } 1678 return (error); 1679 } 1680 1681 /* XXX hack to avoid calling ITIMES in getattr */ 1682 int 1683 lfs_getattr(void *v) 1684 { 1685 struct vop_getattr_args /* { 1686 struct vnode *a_vp; 1687 struct vattr *a_vap; 1688 kauth_cred_t a_cred; 1689 } */ *ap = v; 1690 struct vnode *vp = ap->a_vp; 1691 struct inode *ip = VTOI(vp); 1692 struct vattr *vap = ap->a_vap; 1693 struct lfs *fs = ip->i_lfs; 1694 /* 1695 * Copy from inode table 1696 */ 1697 vap->va_fsid = ip->i_dev; 1698 vap->va_fileid = ip->i_number; 1699 vap->va_mode = ip->i_mode & ~IFMT; 1700 vap->va_nlink = ip->i_nlink; 1701 vap->va_uid = ip->i_uid; 1702 vap->va_gid = ip->i_gid; 1703 vap->va_rdev = (dev_t)ip->i_ffs1_rdev; 1704 vap->va_size = vp->v_size; 1705 vap->va_atime.tv_sec = ip->i_ffs1_atime; 1706 vap->va_atime.tv_nsec = ip->i_ffs1_atimensec; 1707 vap->va_mtime.tv_sec = ip->i_ffs1_mtime; 1708 vap->va_mtime.tv_nsec = ip->i_ffs1_mtimensec; 1709 vap->va_ctime.tv_sec = ip->i_ffs1_ctime; 1710 vap->va_ctime.tv_nsec = ip->i_ffs1_ctimensec; 1711 vap->va_flags = ip->i_flags; 1712 vap->va_gen = ip->i_gen; 1713 /* this doesn't belong here */ 1714 if (vp->v_type == VBLK) 1715 vap->va_blocksize = BLKDEV_IOSIZE; 1716 else if (vp->v_type == VCHR) 1717 vap->va_blocksize = MAXBSIZE; 1718 else 1719 vap->va_blocksize = vp->v_mount->mnt_stat.f_iosize; 1720 vap->va_bytes = fsbtob(fs, (u_quad_t)ip->i_lfs_effnblks); 1721 vap->va_type = vp->v_type; 1722 vap->va_filerev = ip->i_modrev; 1723 return (0); 1724 } 1725 1726 /* 1727 * Check to make sure the inode blocks won't choke the buffer 1728 * cache, then call ufs_setattr as usual. 1729 */ 1730 int 1731 lfs_setattr(void *v) 1732 { 1733 struct vop_setattr_args /* { 1734 struct vnode *a_vp; 1735 struct vattr *a_vap; 1736 kauth_cred_t a_cred; 1737 } */ *ap = v; 1738 struct vnode *vp = ap->a_vp; 1739 1740 lfs_check(vp, LFS_UNUSED_LBN, 0); 1741 return ufs_setattr(v); 1742 } 1743 1744 /* 1745 * Release the block we hold on lfs_newseg wrapping. Called on file close, 1746 * or explicitly from LFCNWRAPGO. Called with the interlock held. 1747 */ 1748 static int 1749 lfs_wrapgo(struct lfs *fs, struct inode *ip, int waitfor) 1750 { 1751 if (fs->lfs_stoplwp != curlwp) 1752 return EBUSY; 1753 1754 fs->lfs_stoplwp = NULL; 1755 cv_signal(&fs->lfs_stopcv); 1756 1757 KASSERT(fs->lfs_nowrap > 0); 1758 if (fs->lfs_nowrap <= 0) { 1759 return 0; 1760 } 1761 1762 if (--fs->lfs_nowrap == 0) { 1763 log(LOG_NOTICE, "%s: re-enabled log wrap\n", fs->lfs_fsmnt); 1764 wakeup(&fs->lfs_wrappass); 1765 lfs_wakeup_cleaner(fs); 1766 } 1767 if (waitfor) { 1768 mtsleep(&fs->lfs_nextseg, PCATCH | PUSER, "segment", 1769 0, &lfs_lock); 1770 } 1771 1772 return 0; 1773 } 1774 1775 /* 1776 * Close called 1777 */ 1778 /* ARGSUSED */ 1779 int 1780 lfs_close(void *v) 1781 { 1782 struct vop_close_args /* { 1783 struct vnode *a_vp; 1784 int a_fflag; 1785 kauth_cred_t a_cred; 1786 } */ *ap = v; 1787 struct vnode *vp = ap->a_vp; 1788 struct inode *ip = VTOI(vp); 1789 struct lfs *fs = ip->i_lfs; 1790 1791 if ((ip->i_number == ROOTINO || ip->i_number == LFS_IFILE_INUM) && 1792 fs->lfs_stoplwp == curlwp) { 1793 mutex_enter(&lfs_lock); 1794 log(LOG_NOTICE, "lfs_close: releasing log wrap control\n"); 1795 lfs_wrapgo(fs, ip, 0); 1796 mutex_exit(&lfs_lock); 1797 } 1798 1799 if (vp == ip->i_lfs->lfs_ivnode && 1800 vp->v_mount->mnt_iflag & IMNT_UNMOUNT) 1801 return 0; 1802 1803 if (vp->v_usecount > 1 && vp != ip->i_lfs->lfs_ivnode) { 1804 LFS_ITIMES(ip, NULL, NULL, NULL); 1805 } 1806 return (0); 1807 } 1808 1809 /* 1810 * Close wrapper for special devices. 1811 * 1812 * Update the times on the inode then do device close. 1813 */ 1814 int 1815 lfsspec_close(void *v) 1816 { 1817 struct vop_close_args /* { 1818 struct vnode *a_vp; 1819 int a_fflag; 1820 kauth_cred_t a_cred; 1821 } */ *ap = v; 1822 struct vnode *vp; 1823 struct inode *ip; 1824 1825 vp = ap->a_vp; 1826 ip = VTOI(vp); 1827 if (vp->v_usecount > 1) { 1828 LFS_ITIMES(ip, NULL, NULL, NULL); 1829 } 1830 return (VOCALL (spec_vnodeop_p, VOFFSET(vop_close), ap)); 1831 } 1832 1833 /* 1834 * Close wrapper for fifo's. 1835 * 1836 * Update the times on the inode then do device close. 1837 */ 1838 int 1839 lfsfifo_close(void *v) 1840 { 1841 struct vop_close_args /* { 1842 struct vnode *a_vp; 1843 int a_fflag; 1844 kauth_cred_ a_cred; 1845 } */ *ap = v; 1846 struct vnode *vp; 1847 struct inode *ip; 1848 1849 vp = ap->a_vp; 1850 ip = VTOI(vp); 1851 if (ap->a_vp->v_usecount > 1) { 1852 LFS_ITIMES(ip, NULL, NULL, NULL); 1853 } 1854 return (VOCALL (fifo_vnodeop_p, VOFFSET(vop_close), ap)); 1855 } 1856 1857 /* 1858 * Reclaim an inode so that it can be used for other purposes. 1859 */ 1860 1861 int 1862 lfs_reclaim(void *v) 1863 { 1864 struct vop_reclaim_args /* { 1865 struct vnode *a_vp; 1866 } */ *ap = v; 1867 struct vnode *vp = ap->a_vp; 1868 struct inode *ip = VTOI(vp); 1869 struct lfs *fs = ip->i_lfs; 1870 int error; 1871 1872 /* 1873 * The inode must be freed and updated before being removed 1874 * from its hash chain. Other threads trying to gain a hold 1875 * on the inode will be stalled because it is locked (VI_XLOCK). 1876 */ 1877 if (ip->i_nlink <= 0 && (vp->v_mount->mnt_flag & MNT_RDONLY) == 0) 1878 lfs_vfree(vp, ip->i_number, ip->i_omode); 1879 1880 mutex_enter(&lfs_lock); 1881 LFS_CLR_UINO(ip, IN_ALLMOD); 1882 mutex_exit(&lfs_lock); 1883 if ((error = ufs_reclaim(vp))) 1884 return (error); 1885 1886 /* 1887 * Take us off the paging and/or dirop queues if we were on them. 1888 * We shouldn't be on them. 1889 */ 1890 mutex_enter(&lfs_lock); 1891 if (ip->i_flags & IN_PAGING) { 1892 log(LOG_WARNING, "%s: reclaimed vnode is IN_PAGING\n", 1893 fs->lfs_fsmnt); 1894 ip->i_flags &= ~IN_PAGING; 1895 TAILQ_REMOVE(&fs->lfs_pchainhd, ip, i_lfs_pchain); 1896 } 1897 if (vp->v_uflag & VU_DIROP) { 1898 panic("reclaimed vnode is VU_DIROP"); 1899 vp->v_uflag &= ~VU_DIROP; 1900 TAILQ_REMOVE(&fs->lfs_dchainhd, ip, i_lfs_dchain); 1901 } 1902 mutex_exit(&lfs_lock); 1903 1904 pool_put(&lfs_dinode_pool, ip->i_din.ffs1_din); 1905 lfs_deregister_all(vp); 1906 pool_put(&lfs_inoext_pool, ip->inode_ext.lfs); 1907 ip->inode_ext.lfs = NULL; 1908 genfs_node_destroy(vp); 1909 pool_put(&lfs_inode_pool, vp->v_data); 1910 vp->v_data = NULL; 1911 return (0); 1912 } 1913 1914 /* 1915 * Read a block from a storage device. 1916 * In order to avoid reading blocks that are in the process of being 1917 * written by the cleaner---and hence are not mutexed by the normal 1918 * buffer cache / page cache mechanisms---check for collisions before 1919 * reading. 1920 * 1921 * We inline ufs_strategy to make sure that the VOP_BMAP occurs *before* 1922 * the active cleaner test. 1923 * 1924 * XXX This code assumes that lfs_markv makes synchronous checkpoints. 1925 */ 1926 int 1927 lfs_strategy(void *v) 1928 { 1929 struct vop_strategy_args /* { 1930 struct vnode *a_vp; 1931 struct buf *a_bp; 1932 } */ *ap = v; 1933 struct buf *bp; 1934 struct lfs *fs; 1935 struct vnode *vp; 1936 struct inode *ip; 1937 daddr_t tbn; 1938 #define MAXLOOP 25 1939 int i, sn, error, slept, loopcount; 1940 1941 bp = ap->a_bp; 1942 vp = ap->a_vp; 1943 ip = VTOI(vp); 1944 fs = ip->i_lfs; 1945 1946 /* lfs uses its strategy routine only for read */ 1947 KASSERT(bp->b_flags & B_READ); 1948 1949 if (vp->v_type == VBLK || vp->v_type == VCHR) 1950 panic("lfs_strategy: spec"); 1951 KASSERT(bp->b_bcount != 0); 1952 if (bp->b_blkno == bp->b_lblkno) { 1953 error = VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, 1954 NULL); 1955 if (error) { 1956 bp->b_error = error; 1957 bp->b_resid = bp->b_bcount; 1958 biodone(bp); 1959 return (error); 1960 } 1961 if ((long)bp->b_blkno == -1) /* no valid data */ 1962 clrbuf(bp); 1963 } 1964 if ((long)bp->b_blkno < 0) { /* block is not on disk */ 1965 bp->b_resid = bp->b_bcount; 1966 biodone(bp); 1967 return (0); 1968 } 1969 1970 slept = 1; 1971 loopcount = 0; 1972 mutex_enter(&lfs_lock); 1973 while (slept && fs->lfs_seglock) { 1974 mutex_exit(&lfs_lock); 1975 /* 1976 * Look through list of intervals. 1977 * There will only be intervals to look through 1978 * if the cleaner holds the seglock. 1979 * Since the cleaner is synchronous, we can trust 1980 * the list of intervals to be current. 1981 */ 1982 tbn = dbtofsb(fs, bp->b_blkno); 1983 sn = dtosn(fs, tbn); 1984 slept = 0; 1985 for (i = 0; i < fs->lfs_cleanind; i++) { 1986 if (sn == dtosn(fs, fs->lfs_cleanint[i]) && 1987 tbn >= fs->lfs_cleanint[i]) { 1988 DLOG((DLOG_CLEAN, 1989 "lfs_strategy: ino %d lbn %" PRId64 1990 " ind %d sn %d fsb %" PRIx32 1991 " given sn %d fsb %" PRIx64 "\n", 1992 ip->i_number, bp->b_lblkno, i, 1993 dtosn(fs, fs->lfs_cleanint[i]), 1994 fs->lfs_cleanint[i], sn, tbn)); 1995 DLOG((DLOG_CLEAN, 1996 "lfs_strategy: sleeping on ino %d lbn %" 1997 PRId64 "\n", ip->i_number, bp->b_lblkno)); 1998 mutex_enter(&lfs_lock); 1999 if (LFS_SEGLOCK_HELD(fs) && fs->lfs_iocount) { 2000 /* 2001 * Cleaner can't wait for itself. 2002 * Instead, wait for the blocks 2003 * to be written to disk. 2004 * XXX we need pribio in the test 2005 * XXX here. 2006 */ 2007 mtsleep(&fs->lfs_iocount, 2008 (PRIBIO + 1) | PNORELOCK, 2009 "clean2", hz/10 + 1, 2010 &lfs_lock); 2011 slept = 1; 2012 ++loopcount; 2013 break; 2014 } else if (fs->lfs_seglock) { 2015 mtsleep(&fs->lfs_seglock, 2016 (PRIBIO + 1) | PNORELOCK, 2017 "clean1", 0, 2018 &lfs_lock); 2019 slept = 1; 2020 break; 2021 } 2022 mutex_exit(&lfs_lock); 2023 } 2024 } 2025 mutex_enter(&lfs_lock); 2026 if (loopcount > MAXLOOP) { 2027 printf("lfs_strategy: breaking out of clean2 loop\n"); 2028 break; 2029 } 2030 } 2031 mutex_exit(&lfs_lock); 2032 2033 vp = ip->i_devvp; 2034 VOP_STRATEGY(vp, bp); 2035 return (0); 2036 } 2037 2038 /* 2039 * Inline lfs_segwrite/lfs_writevnodes, but just for dirops. 2040 * Technically this is a checkpoint (the on-disk state is valid) 2041 * even though we are leaving out all the file data. 2042 */ 2043 int 2044 lfs_flush_dirops(struct lfs *fs) 2045 { 2046 struct inode *ip, *nip; 2047 struct vnode *vp; 2048 extern int lfs_dostats; 2049 struct segment *sp; 2050 int flags = 0; 2051 int error = 0; 2052 2053 ASSERT_MAYBE_SEGLOCK(fs); 2054 KASSERT(fs->lfs_nadirop == 0); 2055 2056 if (fs->lfs_ronly) 2057 return EROFS; 2058 2059 mutex_enter(&lfs_lock); 2060 if (TAILQ_FIRST(&fs->lfs_dchainhd) == NULL) { 2061 mutex_exit(&lfs_lock); 2062 return 0; 2063 } else 2064 mutex_exit(&lfs_lock); 2065 2066 if (lfs_dostats) 2067 ++lfs_stats.flush_invoked; 2068 2069 lfs_imtime(fs); 2070 lfs_seglock(fs, flags); 2071 sp = fs->lfs_sp; 2072 2073 /* 2074 * lfs_writevnodes, optimized to get dirops out of the way. 2075 * Only write dirops, and don't flush files' pages, only 2076 * blocks from the directories. 2077 * 2078 * We don't need to vref these files because they are 2079 * dirops and so hold an extra reference until the 2080 * segunlock clears them of that status. 2081 * 2082 * We don't need to check for IN_ADIROP because we know that 2083 * no dirops are active. 2084 * 2085 */ 2086 mutex_enter(&lfs_lock); 2087 for (ip = TAILQ_FIRST(&fs->lfs_dchainhd); ip != NULL; ip = nip) { 2088 nip = TAILQ_NEXT(ip, i_lfs_dchain); 2089 mutex_exit(&lfs_lock); 2090 vp = ITOV(ip); 2091 2092 KASSERT((ip->i_flag & IN_ADIROP) == 0); 2093 KASSERT(vp->v_uflag & VU_DIROP); 2094 KASSERT(!(vp->v_iflag & VI_XLOCK)); 2095 2096 /* 2097 * All writes to directories come from dirops; all 2098 * writes to files' direct blocks go through the page 2099 * cache, which we're not touching. Reads to files 2100 * and/or directories will not be affected by writing 2101 * directory blocks inodes and file inodes. So we don't 2102 * really need to lock. 2103 */ 2104 if (vp->v_iflag & VI_XLOCK) { 2105 mutex_enter(&lfs_lock); 2106 continue; 2107 } 2108 /* XXX see below 2109 * waslocked = VOP_ISLOCKED(vp); 2110 */ 2111 if (vp->v_type != VREG && 2112 ((ip->i_flag & IN_ALLMOD) || !VPISEMPTY(vp))) { 2113 error = lfs_writefile(fs, sp, vp); 2114 if (!VPISEMPTY(vp) && !WRITEINPROG(vp) && 2115 !(ip->i_flag & IN_ALLMOD)) { 2116 mutex_enter(&lfs_lock); 2117 LFS_SET_UINO(ip, IN_MODIFIED); 2118 mutex_exit(&lfs_lock); 2119 } 2120 if (error && (sp->seg_flags & SEGM_SINGLE)) { 2121 mutex_enter(&lfs_lock); 2122 error = EAGAIN; 2123 break; 2124 } 2125 } 2126 KDASSERT(ip->i_number != LFS_IFILE_INUM); 2127 error = lfs_writeinode(fs, sp, ip); 2128 mutex_enter(&lfs_lock); 2129 if (error && (sp->seg_flags & SEGM_SINGLE)) { 2130 error = EAGAIN; 2131 break; 2132 } 2133 2134 /* 2135 * We might need to update these inodes again, 2136 * for example, if they have data blocks to write. 2137 * Make sure that after this flush, they are still 2138 * marked IN_MODIFIED so that we don't forget to 2139 * write them. 2140 */ 2141 /* XXX only for non-directories? --KS */ 2142 LFS_SET_UINO(ip, IN_MODIFIED); 2143 } 2144 mutex_exit(&lfs_lock); 2145 /* We've written all the dirops there are */ 2146 ((SEGSUM *)(sp->segsum))->ss_flags &= ~(SS_CONT); 2147 lfs_finalize_fs_seguse(fs); 2148 (void) lfs_writeseg(fs, sp); 2149 lfs_segunlock(fs); 2150 2151 return error; 2152 } 2153 2154 /* 2155 * Flush all vnodes for which the pagedaemon has requested pageouts. 2156 * Skip over any files that are marked VU_DIROP (since lfs_flush_dirop() 2157 * has just run, this would be an error). If we have to skip a vnode 2158 * for any reason, just skip it; if we have to wait for the cleaner, 2159 * abort. The writer daemon will call us again later. 2160 */ 2161 int 2162 lfs_flush_pchain(struct lfs *fs) 2163 { 2164 struct inode *ip, *nip; 2165 struct vnode *vp; 2166 extern int lfs_dostats; 2167 struct segment *sp; 2168 int error, error2; 2169 2170 ASSERT_NO_SEGLOCK(fs); 2171 2172 if (fs->lfs_ronly) 2173 return EROFS; 2174 2175 mutex_enter(&lfs_lock); 2176 if (TAILQ_FIRST(&fs->lfs_pchainhd) == NULL) { 2177 mutex_exit(&lfs_lock); 2178 return 0; 2179 } else 2180 mutex_exit(&lfs_lock); 2181 2182 /* Get dirops out of the way */ 2183 if ((error = lfs_flush_dirops(fs)) != 0) 2184 return error; 2185 2186 if (lfs_dostats) 2187 ++lfs_stats.flush_invoked; 2188 2189 /* 2190 * Inline lfs_segwrite/lfs_writevnodes, but just for pageouts. 2191 */ 2192 lfs_imtime(fs); 2193 lfs_seglock(fs, 0); 2194 sp = fs->lfs_sp; 2195 2196 /* 2197 * lfs_writevnodes, optimized to clear pageout requests. 2198 * Only write non-dirop files that are in the pageout queue. 2199 * We're very conservative about what we write; we want to be 2200 * fast and async. 2201 */ 2202 mutex_enter(&lfs_lock); 2203 top: 2204 for (ip = TAILQ_FIRST(&fs->lfs_pchainhd); ip != NULL; ip = nip) { 2205 nip = TAILQ_NEXT(ip, i_lfs_pchain); 2206 vp = ITOV(ip); 2207 2208 if (!(ip->i_flags & IN_PAGING)) 2209 goto top; 2210 2211 mutex_enter(vp->v_interlock); 2212 if ((vp->v_iflag & VI_XLOCK) || (vp->v_uflag & VU_DIROP) != 0) { 2213 mutex_exit(vp->v_interlock); 2214 continue; 2215 } 2216 if (vp->v_type != VREG) { 2217 mutex_exit(vp->v_interlock); 2218 continue; 2219 } 2220 if (lfs_vref(vp)) 2221 continue; 2222 mutex_exit(&lfs_lock); 2223 2224 if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT | LK_RETRY) != 0) { 2225 lfs_vunref(vp); 2226 mutex_enter(&lfs_lock); 2227 continue; 2228 } 2229 2230 error = lfs_writefile(fs, sp, vp); 2231 if (!VPISEMPTY(vp) && !WRITEINPROG(vp) && 2232 !(ip->i_flag & IN_ALLMOD)) { 2233 mutex_enter(&lfs_lock); 2234 LFS_SET_UINO(ip, IN_MODIFIED); 2235 mutex_exit(&lfs_lock); 2236 } 2237 KDASSERT(ip->i_number != LFS_IFILE_INUM); 2238 error2 = lfs_writeinode(fs, sp, ip); 2239 2240 VOP_UNLOCK(vp); 2241 lfs_vunref(vp); 2242 2243 if (error == EAGAIN || error2 == EAGAIN) { 2244 lfs_writeseg(fs, sp); 2245 mutex_enter(&lfs_lock); 2246 break; 2247 } 2248 mutex_enter(&lfs_lock); 2249 } 2250 mutex_exit(&lfs_lock); 2251 (void) lfs_writeseg(fs, sp); 2252 lfs_segunlock(fs); 2253 2254 return 0; 2255 } 2256 2257 /* 2258 * Provide a fcntl interface to sys_lfs_{segwait,bmapv,markv}. 2259 */ 2260 int 2261 lfs_fcntl(void *v) 2262 { 2263 struct vop_fcntl_args /* { 2264 struct vnode *a_vp; 2265 u_int a_command; 2266 void * a_data; 2267 int a_fflag; 2268 kauth_cred_t a_cred; 2269 } */ *ap = v; 2270 struct timeval tv; 2271 struct timeval *tvp; 2272 BLOCK_INFO *blkiov; 2273 CLEANERINFO *cip; 2274 SEGUSE *sup; 2275 int blkcnt, error, oclean; 2276 size_t fh_size; 2277 struct lfs_fcntl_markv blkvp; 2278 struct lwp *l; 2279 fsid_t *fsidp; 2280 struct lfs *fs; 2281 struct buf *bp; 2282 fhandle_t *fhp; 2283 daddr_t off; 2284 2285 /* Only respect LFS fcntls on fs root or Ifile */ 2286 if (VTOI(ap->a_vp)->i_number != ROOTINO && 2287 VTOI(ap->a_vp)->i_number != LFS_IFILE_INUM) { 2288 return ufs_fcntl(v); 2289 } 2290 2291 /* Avoid locking a draining lock */ 2292 if (ap->a_vp->v_mount->mnt_iflag & IMNT_UNMOUNT) { 2293 return ESHUTDOWN; 2294 } 2295 2296 /* LFS control and monitoring fcntls are available only to root */ 2297 l = curlwp; 2298 if (((ap->a_command & 0xff00) >> 8) == 'L' && 2299 (error = kauth_authorize_generic(l->l_cred, KAUTH_GENERIC_ISSUSER, 2300 NULL)) != 0) 2301 return (error); 2302 2303 fs = VTOI(ap->a_vp)->i_lfs; 2304 fsidp = &ap->a_vp->v_mount->mnt_stat.f_fsidx; 2305 2306 error = 0; 2307 switch ((int)ap->a_command) { 2308 case LFCNSEGWAITALL_COMPAT_50: 2309 case LFCNSEGWAITALL_COMPAT: 2310 fsidp = NULL; 2311 /* FALLSTHROUGH */ 2312 case LFCNSEGWAIT_COMPAT_50: 2313 case LFCNSEGWAIT_COMPAT: 2314 { 2315 struct timeval50 *tvp50 2316 = (struct timeval50 *)ap->a_data; 2317 timeval50_to_timeval(tvp50, &tv); 2318 tvp = &tv; 2319 } 2320 goto segwait_common; 2321 case LFCNSEGWAITALL: 2322 fsidp = NULL; 2323 /* FALLSTHROUGH */ 2324 case LFCNSEGWAIT: 2325 tvp = (struct timeval *)ap->a_data; 2326 segwait_common: 2327 mutex_enter(&lfs_lock); 2328 ++fs->lfs_sleepers; 2329 mutex_exit(&lfs_lock); 2330 2331 error = lfs_segwait(fsidp, tvp); 2332 2333 mutex_enter(&lfs_lock); 2334 if (--fs->lfs_sleepers == 0) 2335 wakeup(&fs->lfs_sleepers); 2336 mutex_exit(&lfs_lock); 2337 return error; 2338 2339 case LFCNBMAPV: 2340 case LFCNMARKV: 2341 blkvp = *(struct lfs_fcntl_markv *)ap->a_data; 2342 2343 blkcnt = blkvp.blkcnt; 2344 if ((u_int) blkcnt > LFS_MARKV_MAXBLKCNT) 2345 return (EINVAL); 2346 blkiov = lfs_malloc(fs, blkcnt * sizeof(BLOCK_INFO), LFS_NB_BLKIOV); 2347 if ((error = copyin(blkvp.blkiov, blkiov, 2348 blkcnt * sizeof(BLOCK_INFO))) != 0) { 2349 lfs_free(fs, blkiov, LFS_NB_BLKIOV); 2350 return error; 2351 } 2352 2353 mutex_enter(&lfs_lock); 2354 ++fs->lfs_sleepers; 2355 mutex_exit(&lfs_lock); 2356 if (ap->a_command == LFCNBMAPV) 2357 error = lfs_bmapv(l->l_proc, fsidp, blkiov, blkcnt); 2358 else /* LFCNMARKV */ 2359 error = lfs_markv(l->l_proc, fsidp, blkiov, blkcnt); 2360 if (error == 0) 2361 error = copyout(blkiov, blkvp.blkiov, 2362 blkcnt * sizeof(BLOCK_INFO)); 2363 mutex_enter(&lfs_lock); 2364 if (--fs->lfs_sleepers == 0) 2365 wakeup(&fs->lfs_sleepers); 2366 mutex_exit(&lfs_lock); 2367 lfs_free(fs, blkiov, LFS_NB_BLKIOV); 2368 return error; 2369 2370 case LFCNRECLAIM: 2371 /* 2372 * Flush dirops and write Ifile, allowing empty segments 2373 * to be immediately reclaimed. 2374 */ 2375 lfs_writer_enter(fs, "pndirop"); 2376 off = fs->lfs_offset; 2377 lfs_seglock(fs, SEGM_FORCE_CKP | SEGM_CKP); 2378 lfs_flush_dirops(fs); 2379 LFS_CLEANERINFO(cip, fs, bp); 2380 oclean = cip->clean; 2381 LFS_SYNC_CLEANERINFO(cip, fs, bp, 1); 2382 lfs_segwrite(ap->a_vp->v_mount, SEGM_FORCE_CKP); 2383 fs->lfs_sp->seg_flags |= SEGM_PROT; 2384 lfs_segunlock(fs); 2385 lfs_writer_leave(fs); 2386 2387 #ifdef DEBUG 2388 LFS_CLEANERINFO(cip, fs, bp); 2389 DLOG((DLOG_CLEAN, "lfs_fcntl: reclaim wrote %" PRId64 2390 " blocks, cleaned %" PRId32 " segments (activesb %d)\n", 2391 fs->lfs_offset - off, cip->clean - oclean, 2392 fs->lfs_activesb)); 2393 LFS_SYNC_CLEANERINFO(cip, fs, bp, 0); 2394 #endif 2395 2396 return 0; 2397 2398 case LFCNIFILEFH_COMPAT: 2399 /* Return the filehandle of the Ifile */ 2400 if ((error = kauth_authorize_system(l->l_cred, 2401 KAUTH_SYSTEM_FILEHANDLE, 0, NULL, NULL, NULL)) != 0) 2402 return (error); 2403 fhp = (struct fhandle *)ap->a_data; 2404 fhp->fh_fsid = *fsidp; 2405 fh_size = 16; /* former VFS_MAXFIDSIZ */ 2406 return lfs_vptofh(fs->lfs_ivnode, &(fhp->fh_fid), &fh_size); 2407 2408 case LFCNIFILEFH_COMPAT2: 2409 case LFCNIFILEFH: 2410 /* Return the filehandle of the Ifile */ 2411 fhp = (struct fhandle *)ap->a_data; 2412 fhp->fh_fsid = *fsidp; 2413 fh_size = sizeof(struct lfs_fhandle) - 2414 offsetof(fhandle_t, fh_fid); 2415 return lfs_vptofh(fs->lfs_ivnode, &(fhp->fh_fid), &fh_size); 2416 2417 case LFCNREWIND: 2418 /* Move lfs_offset to the lowest-numbered segment */ 2419 return lfs_rewind(fs, *(int *)ap->a_data); 2420 2421 case LFCNINVAL: 2422 /* Mark a segment SEGUSE_INVAL */ 2423 LFS_SEGENTRY(sup, fs, *(int *)ap->a_data, bp); 2424 if (sup->su_nbytes > 0) { 2425 brelse(bp, 0); 2426 lfs_unset_inval_all(fs); 2427 return EBUSY; 2428 } 2429 sup->su_flags |= SEGUSE_INVAL; 2430 VOP_BWRITE(bp->b_vp, bp); 2431 return 0; 2432 2433 case LFCNRESIZE: 2434 /* Resize the filesystem */ 2435 return lfs_resize_fs(fs, *(int *)ap->a_data); 2436 2437 case LFCNWRAPSTOP: 2438 case LFCNWRAPSTOP_COMPAT: 2439 /* 2440 * Hold lfs_newseg at segment 0; if requested, sleep until 2441 * the filesystem wraps around. To support external agents 2442 * (dump, fsck-based regression test) that need to look at 2443 * a snapshot of the filesystem, without necessarily 2444 * requiring that all fs activity stops. 2445 */ 2446 if (fs->lfs_stoplwp == curlwp) 2447 return EALREADY; 2448 2449 mutex_enter(&lfs_lock); 2450 while (fs->lfs_stoplwp != NULL) 2451 cv_wait(&fs->lfs_stopcv, &lfs_lock); 2452 fs->lfs_stoplwp = curlwp; 2453 if (fs->lfs_nowrap == 0) 2454 log(LOG_NOTICE, "%s: disabled log wrap\n", fs->lfs_fsmnt); 2455 ++fs->lfs_nowrap; 2456 if (*(int *)ap->a_data == 1 2457 || ap->a_command == LFCNWRAPSTOP_COMPAT) { 2458 log(LOG_NOTICE, "LFCNSTOPWRAP waiting for log wrap\n"); 2459 error = mtsleep(&fs->lfs_nowrap, PCATCH | PUSER, 2460 "segwrap", 0, &lfs_lock); 2461 log(LOG_NOTICE, "LFCNSTOPWRAP done waiting\n"); 2462 if (error) { 2463 lfs_wrapgo(fs, VTOI(ap->a_vp), 0); 2464 } 2465 } 2466 mutex_exit(&lfs_lock); 2467 return 0; 2468 2469 case LFCNWRAPGO: 2470 case LFCNWRAPGO_COMPAT: 2471 /* 2472 * Having done its work, the agent wakes up the writer. 2473 * If the argument is 1, it sleeps until a new segment 2474 * is selected. 2475 */ 2476 mutex_enter(&lfs_lock); 2477 error = lfs_wrapgo(fs, VTOI(ap->a_vp), 2478 ap->a_command == LFCNWRAPGO_COMPAT ? 1 : 2479 *((int *)ap->a_data)); 2480 mutex_exit(&lfs_lock); 2481 return error; 2482 2483 case LFCNWRAPPASS: 2484 if ((VTOI(ap->a_vp)->i_lfs_iflags & LFSI_WRAPWAIT)) 2485 return EALREADY; 2486 mutex_enter(&lfs_lock); 2487 if (fs->lfs_stoplwp != curlwp) { 2488 mutex_exit(&lfs_lock); 2489 return EALREADY; 2490 } 2491 if (fs->lfs_nowrap == 0) { 2492 mutex_exit(&lfs_lock); 2493 return EBUSY; 2494 } 2495 fs->lfs_wrappass = 1; 2496 wakeup(&fs->lfs_wrappass); 2497 /* Wait for the log to wrap, if asked */ 2498 if (*(int *)ap->a_data) { 2499 mutex_enter(ap->a_vp->v_interlock); 2500 if (lfs_vref(ap->a_vp) != 0) 2501 panic("LFCNWRAPPASS: lfs_vref failed"); 2502 VTOI(ap->a_vp)->i_lfs_iflags |= LFSI_WRAPWAIT; 2503 log(LOG_NOTICE, "LFCNPASS waiting for log wrap\n"); 2504 error = mtsleep(&fs->lfs_nowrap, PCATCH | PUSER, 2505 "segwrap", 0, &lfs_lock); 2506 log(LOG_NOTICE, "LFCNPASS done waiting\n"); 2507 VTOI(ap->a_vp)->i_lfs_iflags &= ~LFSI_WRAPWAIT; 2508 lfs_vunref(ap->a_vp); 2509 } 2510 mutex_exit(&lfs_lock); 2511 return error; 2512 2513 case LFCNWRAPSTATUS: 2514 mutex_enter(&lfs_lock); 2515 *(int *)ap->a_data = fs->lfs_wrapstatus; 2516 mutex_exit(&lfs_lock); 2517 return 0; 2518 2519 default: 2520 return ufs_fcntl(v); 2521 } 2522 return 0; 2523 } 2524 2525 int 2526 lfs_getpages(void *v) 2527 { 2528 struct vop_getpages_args /* { 2529 struct vnode *a_vp; 2530 voff_t a_offset; 2531 struct vm_page **a_m; 2532 int *a_count; 2533 int a_centeridx; 2534 vm_prot_t a_access_type; 2535 int a_advice; 2536 int a_flags; 2537 } */ *ap = v; 2538 2539 if (VTOI(ap->a_vp)->i_number == LFS_IFILE_INUM && 2540 (ap->a_access_type & VM_PROT_WRITE) != 0) { 2541 return EPERM; 2542 } 2543 if ((ap->a_access_type & VM_PROT_WRITE) != 0) { 2544 mutex_enter(&lfs_lock); 2545 LFS_SET_UINO(VTOI(ap->a_vp), IN_MODIFIED); 2546 mutex_exit(&lfs_lock); 2547 } 2548 2549 /* 2550 * we're relying on the fact that genfs_getpages() always read in 2551 * entire filesystem blocks. 2552 */ 2553 return genfs_getpages(v); 2554 } 2555 2556 /* 2557 * Wait for a page to become unbusy, possibly printing diagnostic messages 2558 * as well. 2559 * 2560 * Called with vp->v_interlock held; return with it held. 2561 */ 2562 static void 2563 wait_for_page(struct vnode *vp, struct vm_page *pg, const char *label) 2564 { 2565 KASSERT(mutex_owned(vp->v_interlock)); 2566 if ((pg->flags & PG_BUSY) == 0) 2567 return; /* Nothing to wait for! */ 2568 2569 #if defined(DEBUG) && defined(UVM_PAGE_TRKOWN) 2570 static struct vm_page *lastpg; 2571 2572 if (label != NULL && pg != lastpg) { 2573 if (pg->owner_tag) { 2574 printf("lfs_putpages[%d.%d]: %s: page %p owner %d.%d [%s]\n", 2575 curproc->p_pid, curlwp->l_lid, label, 2576 pg, pg->owner, pg->lowner, pg->owner_tag); 2577 } else { 2578 printf("lfs_putpages[%d.%d]: %s: page %p unowned?!\n", 2579 curproc->p_pid, curlwp->l_lid, label, pg); 2580 } 2581 } 2582 lastpg = pg; 2583 #endif 2584 2585 pg->flags |= PG_WANTED; 2586 UVM_UNLOCK_AND_WAIT(pg, vp->v_interlock, 0, "lfsput", 0); 2587 mutex_enter(vp->v_interlock); 2588 } 2589 2590 /* 2591 * This routine is called by lfs_putpages() when it can't complete the 2592 * write because a page is busy. This means that either (1) someone, 2593 * possibly the pagedaemon, is looking at this page, and will give it up 2594 * presently; or (2) we ourselves are holding the page busy in the 2595 * process of being written (either gathered or actually on its way to 2596 * disk). We don't need to give up the segment lock, but we might need 2597 * to call lfs_writeseg() to expedite the page's journey to disk. 2598 * 2599 * Called with vp->v_interlock held; return with it held. 2600 */ 2601 /* #define BUSYWAIT */ 2602 static void 2603 write_and_wait(struct lfs *fs, struct vnode *vp, struct vm_page *pg, 2604 int seglocked, const char *label) 2605 { 2606 KASSERT(mutex_owned(vp->v_interlock)); 2607 #ifndef BUSYWAIT 2608 struct inode *ip = VTOI(vp); 2609 struct segment *sp = fs->lfs_sp; 2610 int count = 0; 2611 2612 if (pg == NULL) 2613 return; 2614 2615 while (pg->flags & PG_BUSY && 2616 pg->uobject == &vp->v_uobj) { 2617 mutex_exit(vp->v_interlock); 2618 if (sp->cbpp - sp->bpp > 1) { 2619 /* Write gathered pages */ 2620 lfs_updatemeta(sp); 2621 lfs_release_finfo(fs); 2622 (void) lfs_writeseg(fs, sp); 2623 2624 /* 2625 * Reinitialize FIP 2626 */ 2627 KASSERT(sp->vp == vp); 2628 lfs_acquire_finfo(fs, ip->i_number, 2629 ip->i_gen); 2630 } 2631 ++count; 2632 mutex_enter(vp->v_interlock); 2633 wait_for_page(vp, pg, label); 2634 } 2635 if (label != NULL && count > 1) { 2636 DLOG((DLOG_PAGE, "lfs_putpages[%d]: %s: %sn = %d\n", 2637 curproc->p_pid, label, (count > 0 ? "looping, " : ""), 2638 count)); 2639 } 2640 #else 2641 preempt(1); 2642 #endif 2643 KASSERT(mutex_owned(vp->v_interlock)); 2644 } 2645 2646 /* 2647 * Make sure that for all pages in every block in the given range, 2648 * either all are dirty or all are clean. If any of the pages 2649 * we've seen so far are dirty, put the vnode on the paging chain, 2650 * and mark it IN_PAGING. 2651 * 2652 * If checkfirst != 0, don't check all the pages but return at the 2653 * first dirty page. 2654 */ 2655 static int 2656 check_dirty(struct lfs *fs, struct vnode *vp, 2657 off_t startoffset, off_t endoffset, off_t blkeof, 2658 int flags, int checkfirst, struct vm_page **pgp) 2659 { 2660 int by_list; 2661 struct vm_page *curpg = NULL; /* XXX: gcc */ 2662 struct vm_page *pgs[MAXBSIZE / PAGE_SIZE], *pg; 2663 off_t soff = 0; /* XXX: gcc */ 2664 voff_t off; 2665 int i; 2666 int nonexistent; 2667 int any_dirty; /* number of dirty pages */ 2668 int dirty; /* number of dirty pages in a block */ 2669 int tdirty; 2670 int pages_per_block = fs->lfs_bsize >> PAGE_SHIFT; 2671 int pagedaemon = (curlwp == uvm.pagedaemon_lwp); 2672 2673 KASSERT(mutex_owned(vp->v_interlock)); 2674 ASSERT_MAYBE_SEGLOCK(fs); 2675 top: 2676 by_list = (vp->v_uobj.uo_npages <= 2677 ((endoffset - startoffset) >> PAGE_SHIFT) * 2678 UVM_PAGE_TREE_PENALTY); 2679 any_dirty = 0; 2680 2681 if (by_list) { 2682 curpg = TAILQ_FIRST(&vp->v_uobj.memq); 2683 } else { 2684 soff = startoffset; 2685 } 2686 while (by_list || soff < MIN(blkeof, endoffset)) { 2687 if (by_list) { 2688 /* 2689 * Find the first page in a block. Skip 2690 * blocks outside our area of interest or beyond 2691 * the end of file. 2692 */ 2693 KASSERT(curpg == NULL 2694 || (curpg->flags & PG_MARKER) == 0); 2695 if (pages_per_block > 1) { 2696 while (curpg && 2697 ((curpg->offset & fs->lfs_bmask) || 2698 curpg->offset >= vp->v_size || 2699 curpg->offset >= endoffset)) { 2700 curpg = TAILQ_NEXT(curpg, listq.queue); 2701 KASSERT(curpg == NULL || 2702 (curpg->flags & PG_MARKER) == 0); 2703 } 2704 } 2705 if (curpg == NULL) 2706 break; 2707 soff = curpg->offset; 2708 } 2709 2710 /* 2711 * Mark all pages in extended range busy; find out if any 2712 * of them are dirty. 2713 */ 2714 nonexistent = dirty = 0; 2715 for (i = 0; i == 0 || i < pages_per_block; i++) { 2716 KASSERT(mutex_owned(vp->v_interlock)); 2717 if (by_list && pages_per_block <= 1) { 2718 pgs[i] = pg = curpg; 2719 } else { 2720 off = soff + (i << PAGE_SHIFT); 2721 pgs[i] = pg = uvm_pagelookup(&vp->v_uobj, off); 2722 if (pg == NULL) { 2723 ++nonexistent; 2724 continue; 2725 } 2726 } 2727 KASSERT(pg != NULL); 2728 2729 /* 2730 * If we're holding the segment lock, we can deadlock 2731 * against a process that has our page and is waiting 2732 * for the cleaner, while the cleaner waits for the 2733 * segment lock. Just bail in that case. 2734 */ 2735 if ((pg->flags & PG_BUSY) && 2736 (pagedaemon || LFS_SEGLOCK_HELD(fs))) { 2737 if (i > 0) 2738 uvm_page_unbusy(pgs, i); 2739 DLOG((DLOG_PAGE, "lfs_putpages: avoiding 3-way or pagedaemon deadlock\n")); 2740 if (pgp) 2741 *pgp = pg; 2742 KASSERT(mutex_owned(vp->v_interlock)); 2743 return -1; 2744 } 2745 2746 while (pg->flags & PG_BUSY) { 2747 wait_for_page(vp, pg, NULL); 2748 KASSERT(mutex_owned(vp->v_interlock)); 2749 if (i > 0) 2750 uvm_page_unbusy(pgs, i); 2751 KASSERT(mutex_owned(vp->v_interlock)); 2752 goto top; 2753 } 2754 pg->flags |= PG_BUSY; 2755 UVM_PAGE_OWN(pg, "lfs_putpages"); 2756 2757 pmap_page_protect(pg, VM_PROT_NONE); 2758 tdirty = (pmap_clear_modify(pg) || 2759 (pg->flags & PG_CLEAN) == 0); 2760 dirty += tdirty; 2761 } 2762 if (pages_per_block > 0 && nonexistent >= pages_per_block) { 2763 if (by_list) { 2764 curpg = TAILQ_NEXT(curpg, listq.queue); 2765 } else { 2766 soff += fs->lfs_bsize; 2767 } 2768 continue; 2769 } 2770 2771 any_dirty += dirty; 2772 KASSERT(nonexistent == 0); 2773 KASSERT(mutex_owned(vp->v_interlock)); 2774 2775 /* 2776 * If any are dirty make all dirty; unbusy them, 2777 * but if we were asked to clean, wire them so that 2778 * the pagedaemon doesn't bother us about them while 2779 * they're on their way to disk. 2780 */ 2781 for (i = 0; i == 0 || i < pages_per_block; i++) { 2782 KASSERT(mutex_owned(vp->v_interlock)); 2783 pg = pgs[i]; 2784 KASSERT(!((pg->flags & PG_CLEAN) && (pg->flags & PG_DELWRI))); 2785 KASSERT(pg->flags & PG_BUSY); 2786 if (dirty) { 2787 pg->flags &= ~PG_CLEAN; 2788 if (flags & PGO_FREE) { 2789 /* 2790 * Wire the page so that 2791 * pdaemon doesn't see it again. 2792 */ 2793 mutex_enter(&uvm_pageqlock); 2794 uvm_pagewire(pg); 2795 mutex_exit(&uvm_pageqlock); 2796 2797 /* Suspended write flag */ 2798 pg->flags |= PG_DELWRI; 2799 } 2800 } 2801 if (pg->flags & PG_WANTED) 2802 wakeup(pg); 2803 pg->flags &= ~(PG_WANTED|PG_BUSY); 2804 UVM_PAGE_OWN(pg, NULL); 2805 } 2806 2807 if (checkfirst && any_dirty) 2808 break; 2809 2810 if (by_list) { 2811 curpg = TAILQ_NEXT(curpg, listq.queue); 2812 } else { 2813 soff += MAX(PAGE_SIZE, fs->lfs_bsize); 2814 } 2815 } 2816 2817 KASSERT(mutex_owned(vp->v_interlock)); 2818 return any_dirty; 2819 } 2820 2821 /* 2822 * lfs_putpages functions like genfs_putpages except that 2823 * 2824 * (1) It needs to bounds-check the incoming requests to ensure that 2825 * they are block-aligned; if they are not, expand the range and 2826 * do the right thing in case, e.g., the requested range is clean 2827 * but the expanded range is dirty. 2828 * 2829 * (2) It needs to explicitly send blocks to be written when it is done. 2830 * If VOP_PUTPAGES is called without the seglock held, we simply take 2831 * the seglock and let lfs_segunlock wait for us. 2832 * XXX There might be a bad situation if we have to flush a vnode while 2833 * XXX lfs_markv is in operation. As of this writing we panic in this 2834 * XXX case. 2835 * 2836 * Assumptions: 2837 * 2838 * (1) The caller does not hold any pages in this vnode busy. If it does, 2839 * there is a danger that when we expand the page range and busy the 2840 * pages we will deadlock. 2841 * 2842 * (2) We are called with vp->v_interlock held; we must return with it 2843 * released. 2844 * 2845 * (3) We don't absolutely have to free pages right away, provided that 2846 * the request does not have PGO_SYNCIO. When the pagedaemon gives 2847 * us a request with PGO_FREE, we take the pages out of the paging 2848 * queue and wake up the writer, which will handle freeing them for us. 2849 * 2850 * We ensure that for any filesystem block, all pages for that 2851 * block are either resident or not, even if those pages are higher 2852 * than EOF; that means that we will be getting requests to free 2853 * "unused" pages above EOF all the time, and should ignore them. 2854 * 2855 * (4) If we are called with PGO_LOCKED, the finfo array we are to write 2856 * into has been set up for us by lfs_writefile. If not, we will 2857 * have to handle allocating and/or freeing an finfo entry. 2858 * 2859 * XXX note that we're (ab)using PGO_LOCKED as "seglock held". 2860 */ 2861 2862 /* How many times to loop before we should start to worry */ 2863 #define TOOMANY 4 2864 2865 int 2866 lfs_putpages(void *v) 2867 { 2868 int error; 2869 struct vop_putpages_args /* { 2870 struct vnode *a_vp; 2871 voff_t a_offlo; 2872 voff_t a_offhi; 2873 int a_flags; 2874 } */ *ap = v; 2875 struct vnode *vp; 2876 struct inode *ip; 2877 struct lfs *fs; 2878 struct segment *sp; 2879 off_t origoffset, startoffset, endoffset, origendoffset, blkeof; 2880 off_t off, max_endoffset; 2881 bool seglocked, sync, pagedaemon, reclaim; 2882 struct vm_page *pg, *busypg; 2883 UVMHIST_FUNC("lfs_putpages"); UVMHIST_CALLED(ubchist); 2884 int oreclaim = 0; 2885 int donewriting = 0; 2886 #ifdef DEBUG 2887 int debug_n_again, debug_n_dirtyclean; 2888 #endif 2889 2890 vp = ap->a_vp; 2891 ip = VTOI(vp); 2892 fs = ip->i_lfs; 2893 sync = (ap->a_flags & PGO_SYNCIO) != 0; 2894 reclaim = (ap->a_flags & PGO_RECLAIM) != 0; 2895 pagedaemon = (curlwp == uvm.pagedaemon_lwp); 2896 2897 KASSERT(mutex_owned(vp->v_interlock)); 2898 2899 /* Putpages does nothing for metadata. */ 2900 if (vp == fs->lfs_ivnode || vp->v_type != VREG) { 2901 mutex_exit(vp->v_interlock); 2902 return 0; 2903 } 2904 2905 /* 2906 * If there are no pages, don't do anything. 2907 */ 2908 if (vp->v_uobj.uo_npages == 0) { 2909 if (TAILQ_EMPTY(&vp->v_uobj.memq) && 2910 (vp->v_iflag & VI_ONWORKLST) && 2911 LIST_FIRST(&vp->v_dirtyblkhd) == NULL) { 2912 vp->v_iflag &= ~VI_WRMAPDIRTY; 2913 vn_syncer_remove_from_worklist(vp); 2914 } 2915 mutex_exit(vp->v_interlock); 2916 2917 /* Remove us from paging queue, if we were on it */ 2918 mutex_enter(&lfs_lock); 2919 if (ip->i_flags & IN_PAGING) { 2920 ip->i_flags &= ~IN_PAGING; 2921 TAILQ_REMOVE(&fs->lfs_pchainhd, ip, i_lfs_pchain); 2922 } 2923 mutex_exit(&lfs_lock); 2924 2925 KASSERT(!mutex_owned(vp->v_interlock)); 2926 return 0; 2927 } 2928 2929 blkeof = blkroundup(fs, ip->i_size); 2930 2931 /* 2932 * Ignore requests to free pages past EOF but in the same block 2933 * as EOF, unless the vnode is being reclaimed or the request 2934 * is synchronous. (If the request is sync, it comes from 2935 * lfs_truncate.) 2936 * 2937 * To avoid being flooded with this request, make these pages 2938 * look "active". 2939 */ 2940 if (!sync && !reclaim && 2941 ap->a_offlo >= ip->i_size && ap->a_offlo < blkeof) { 2942 origoffset = ap->a_offlo; 2943 for (off = origoffset; off < blkeof; off += fs->lfs_bsize) { 2944 pg = uvm_pagelookup(&vp->v_uobj, off); 2945 KASSERT(pg != NULL); 2946 while (pg->flags & PG_BUSY) { 2947 pg->flags |= PG_WANTED; 2948 UVM_UNLOCK_AND_WAIT(pg, vp->v_interlock, 0, 2949 "lfsput2", 0); 2950 mutex_enter(vp->v_interlock); 2951 } 2952 mutex_enter(&uvm_pageqlock); 2953 uvm_pageactivate(pg); 2954 mutex_exit(&uvm_pageqlock); 2955 } 2956 ap->a_offlo = blkeof; 2957 if (ap->a_offhi > 0 && ap->a_offhi <= ap->a_offlo) { 2958 mutex_exit(vp->v_interlock); 2959 return 0; 2960 } 2961 } 2962 2963 /* 2964 * Extend page range to start and end at block boundaries. 2965 * (For the purposes of VOP_PUTPAGES, fragments don't exist.) 2966 */ 2967 origoffset = ap->a_offlo; 2968 origendoffset = ap->a_offhi; 2969 startoffset = origoffset & ~(fs->lfs_bmask); 2970 max_endoffset = (trunc_page(LLONG_MAX) >> fs->lfs_bshift) 2971 << fs->lfs_bshift; 2972 2973 if (origendoffset == 0 || ap->a_flags & PGO_ALLPAGES) { 2974 endoffset = max_endoffset; 2975 origendoffset = endoffset; 2976 } else { 2977 origendoffset = round_page(ap->a_offhi); 2978 endoffset = round_page(blkroundup(fs, origendoffset)); 2979 } 2980 2981 KASSERT(startoffset > 0 || endoffset >= startoffset); 2982 if (startoffset == endoffset) { 2983 /* Nothing to do, why were we called? */ 2984 mutex_exit(vp->v_interlock); 2985 DLOG((DLOG_PAGE, "lfs_putpages: startoffset = endoffset = %" 2986 PRId64 "\n", startoffset)); 2987 return 0; 2988 } 2989 2990 ap->a_offlo = startoffset; 2991 ap->a_offhi = endoffset; 2992 2993 /* 2994 * If not cleaning, just send the pages through genfs_putpages 2995 * to be returned to the pool. 2996 */ 2997 if (!(ap->a_flags & PGO_CLEANIT)) { 2998 DLOG((DLOG_PAGE, "lfs_putpages: no cleanit vn %p ino %d (flags %x)\n", 2999 vp, (int)ip->i_number, ap->a_flags)); 3000 int r = genfs_putpages(v); 3001 KASSERT(!mutex_owned(vp->v_interlock)); 3002 return r; 3003 } 3004 3005 /* Set PGO_BUSYFAIL to avoid deadlocks */ 3006 ap->a_flags |= PGO_BUSYFAIL; 3007 3008 /* 3009 * Likewise, if we are asked to clean but the pages are not 3010 * dirty, we can just free them using genfs_putpages. 3011 */ 3012 #ifdef DEBUG 3013 debug_n_dirtyclean = 0; 3014 #endif 3015 do { 3016 int r; 3017 KASSERT(mutex_owned(vp->v_interlock)); 3018 3019 /* Count the number of dirty pages */ 3020 r = check_dirty(fs, vp, startoffset, endoffset, blkeof, 3021 ap->a_flags, 1, NULL); 3022 if (r < 0) { 3023 /* Pages are busy with another process */ 3024 mutex_exit(vp->v_interlock); 3025 return EDEADLK; 3026 } 3027 if (r > 0) /* Some pages are dirty */ 3028 break; 3029 3030 /* 3031 * Sometimes pages are dirtied between the time that 3032 * we check and the time we try to clean them. 3033 * Instruct lfs_gop_write to return EDEADLK in this case 3034 * so we can write them properly. 3035 */ 3036 ip->i_lfs_iflags |= LFSI_NO_GOP_WRITE; 3037 r = genfs_do_putpages(vp, startoffset, endoffset, 3038 ap->a_flags & ~PGO_SYNCIO, &busypg); 3039 ip->i_lfs_iflags &= ~LFSI_NO_GOP_WRITE; 3040 if (r != EDEADLK) { 3041 KASSERT(!mutex_owned(vp->v_interlock)); 3042 return r; 3043 } 3044 3045 /* One of the pages was busy. Start over. */ 3046 mutex_enter(vp->v_interlock); 3047 wait_for_page(vp, busypg, "dirtyclean"); 3048 #ifdef DEBUG 3049 ++debug_n_dirtyclean; 3050 #endif 3051 } while(1); 3052 3053 #ifdef DEBUG 3054 if (debug_n_dirtyclean > TOOMANY) 3055 DLOG((DLOG_PAGE, "lfs_putpages: dirtyclean: looping, n = %d\n", 3056 debug_n_dirtyclean)); 3057 #endif 3058 3059 /* 3060 * Dirty and asked to clean. 3061 * 3062 * Pagedaemon can't actually write LFS pages; wake up 3063 * the writer to take care of that. The writer will 3064 * notice the pager inode queue and act on that. 3065 * 3066 * XXX We must drop the vp->interlock before taking the lfs_lock or we 3067 * get a nasty deadlock with lfs_flush_pchain(). 3068 */ 3069 if (pagedaemon) { 3070 mutex_exit(vp->v_interlock); 3071 mutex_enter(&lfs_lock); 3072 if (!(ip->i_flags & IN_PAGING)) { 3073 ip->i_flags |= IN_PAGING; 3074 TAILQ_INSERT_TAIL(&fs->lfs_pchainhd, ip, i_lfs_pchain); 3075 } 3076 wakeup(&lfs_writer_daemon); 3077 mutex_exit(&lfs_lock); 3078 preempt(); 3079 KASSERT(!mutex_owned(vp->v_interlock)); 3080 return EWOULDBLOCK; 3081 } 3082 3083 /* 3084 * If this is a file created in a recent dirop, we can't flush its 3085 * inode until the dirop is complete. Drain dirops, then flush the 3086 * filesystem (taking care of any other pending dirops while we're 3087 * at it). 3088 */ 3089 if ((ap->a_flags & (PGO_CLEANIT|PGO_LOCKED)) == PGO_CLEANIT && 3090 (vp->v_uflag & VU_DIROP)) { 3091 DLOG((DLOG_PAGE, "lfs_putpages: flushing VU_DIROP\n")); 3092 3093 lfs_writer_enter(fs, "ppdirop"); 3094 3095 /* Note if we hold the vnode locked */ 3096 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE) 3097 { 3098 DLOG((DLOG_PAGE, "lfs_putpages: dirop inode already locked\n")); 3099 } else { 3100 DLOG((DLOG_PAGE, "lfs_putpages: dirop inode not locked\n")); 3101 } 3102 mutex_exit(vp->v_interlock); 3103 3104 mutex_enter(&lfs_lock); 3105 lfs_flush_fs(fs, sync ? SEGM_SYNC : 0); 3106 mutex_exit(&lfs_lock); 3107 3108 mutex_enter(vp->v_interlock); 3109 lfs_writer_leave(fs); 3110 3111 /* The flush will have cleaned out this vnode as well, 3112 no need to do more to it. */ 3113 } 3114 3115 /* 3116 * This is it. We are going to write some pages. From here on 3117 * down it's all just mechanics. 3118 * 3119 * Don't let genfs_putpages wait; lfs_segunlock will wait for us. 3120 */ 3121 ap->a_flags &= ~PGO_SYNCIO; 3122 3123 /* 3124 * If we've already got the seglock, flush the node and return. 3125 * The FIP has already been set up for us by lfs_writefile, 3126 * and FIP cleanup and lfs_updatemeta will also be done there, 3127 * unless genfs_putpages returns EDEADLK; then we must flush 3128 * what we have, and correct FIP and segment header accounting. 3129 */ 3130 get_seglock: 3131 /* 3132 * If we are not called with the segment locked, lock it. 3133 * Account for a new FIP in the segment header, and set sp->vp. 3134 * (This should duplicate the setup at the top of lfs_writefile().) 3135 */ 3136 seglocked = (ap->a_flags & PGO_LOCKED) != 0; 3137 if (!seglocked) { 3138 mutex_exit(vp->v_interlock); 3139 error = lfs_seglock(fs, SEGM_PROT | (sync ? SEGM_SYNC : 0)); 3140 if (error != 0) { 3141 KASSERT(!mutex_owned(vp->v_interlock)); 3142 return error; 3143 } 3144 mutex_enter(vp->v_interlock); 3145 lfs_acquire_finfo(fs, ip->i_number, ip->i_gen); 3146 } 3147 sp = fs->lfs_sp; 3148 KASSERT(sp->vp == NULL); 3149 sp->vp = vp; 3150 3151 /* Note segments written by reclaim; only for debugging */ 3152 if ((vp->v_iflag & VI_XLOCK) != 0) { 3153 sp->seg_flags |= SEGM_RECLAIM; 3154 fs->lfs_reclino = ip->i_number; 3155 } 3156 3157 /* 3158 * Ensure that the partial segment is marked SS_DIROP if this 3159 * vnode is a DIROP. 3160 */ 3161 if (!seglocked && vp->v_uflag & VU_DIROP) 3162 ((SEGSUM *)(sp->segsum))->ss_flags |= (SS_DIROP|SS_CONT); 3163 3164 /* 3165 * Loop over genfs_putpages until all pages are gathered. 3166 * genfs_putpages() drops the interlock, so reacquire it if necessary. 3167 * Whenever we lose the interlock we have to rerun check_dirty, as 3168 * well, since more pages might have been dirtied in our absence. 3169 */ 3170 #ifdef DEBUG 3171 debug_n_again = 0; 3172 #endif 3173 do { 3174 busypg = NULL; 3175 KASSERT(mutex_owned(vp->v_interlock)); 3176 if (check_dirty(fs, vp, startoffset, endoffset, blkeof, 3177 ap->a_flags, 0, &busypg) < 0) { 3178 mutex_exit(vp->v_interlock); 3179 /* XXX why? --ks */ 3180 mutex_enter(vp->v_interlock); 3181 write_and_wait(fs, vp, busypg, seglocked, NULL); 3182 if (!seglocked) { 3183 mutex_exit(vp->v_interlock); 3184 lfs_release_finfo(fs); 3185 lfs_segunlock(fs); 3186 mutex_enter(vp->v_interlock); 3187 } 3188 sp->vp = NULL; 3189 goto get_seglock; 3190 } 3191 3192 busypg = NULL; 3193 KASSERT(!mutex_owned(&uvm_pageqlock)); 3194 oreclaim = (ap->a_flags & PGO_RECLAIM); 3195 ap->a_flags &= ~PGO_RECLAIM; 3196 error = genfs_do_putpages(vp, startoffset, endoffset, 3197 ap->a_flags, &busypg); 3198 ap->a_flags |= oreclaim; 3199 3200 if (error == EDEADLK || error == EAGAIN) { 3201 DLOG((DLOG_PAGE, "lfs_putpages: genfs_putpages returned" 3202 " %d ino %d off %x (seg %d)\n", error, 3203 ip->i_number, fs->lfs_offset, 3204 dtosn(fs, fs->lfs_offset))); 3205 3206 if (oreclaim) { 3207 mutex_enter(vp->v_interlock); 3208 write_and_wait(fs, vp, busypg, seglocked, "again"); 3209 mutex_exit(vp->v_interlock); 3210 } else { 3211 if ((sp->seg_flags & SEGM_SINGLE) && 3212 fs->lfs_curseg != fs->lfs_startseg) 3213 donewriting = 1; 3214 } 3215 } else if (error) { 3216 DLOG((DLOG_PAGE, "lfs_putpages: genfs_putpages returned" 3217 " %d ino %d off %x (seg %d)\n", error, 3218 (int)ip->i_number, fs->lfs_offset, 3219 dtosn(fs, fs->lfs_offset))); 3220 } 3221 /* genfs_do_putpages loses the interlock */ 3222 #ifdef DEBUG 3223 ++debug_n_again; 3224 #endif 3225 if (oreclaim && error == EAGAIN) { 3226 DLOG((DLOG_PAGE, "vp %p ino %d vi_flags %x a_flags %x avoiding vclean panic\n", 3227 vp, (int)ip->i_number, vp->v_iflag, ap->a_flags)); 3228 mutex_enter(vp->v_interlock); 3229 } 3230 if (error == EDEADLK) 3231 mutex_enter(vp->v_interlock); 3232 } while (error == EDEADLK || (oreclaim && error == EAGAIN)); 3233 #ifdef DEBUG 3234 if (debug_n_again > TOOMANY) 3235 DLOG((DLOG_PAGE, "lfs_putpages: again: looping, n = %d\n", debug_n_again)); 3236 #endif 3237 3238 KASSERT(sp != NULL && sp->vp == vp); 3239 if (!seglocked && !donewriting) { 3240 sp->vp = NULL; 3241 3242 /* Write indirect blocks as well */ 3243 lfs_gather(fs, fs->lfs_sp, vp, lfs_match_indir); 3244 lfs_gather(fs, fs->lfs_sp, vp, lfs_match_dindir); 3245 lfs_gather(fs, fs->lfs_sp, vp, lfs_match_tindir); 3246 3247 KASSERT(sp->vp == NULL); 3248 sp->vp = vp; 3249 } 3250 3251 /* 3252 * Blocks are now gathered into a segment waiting to be written. 3253 * All that's left to do is update metadata, and write them. 3254 */ 3255 lfs_updatemeta(sp); 3256 KASSERT(sp->vp == vp); 3257 sp->vp = NULL; 3258 3259 /* 3260 * If we were called from lfs_writefile, we don't need to clean up 3261 * the FIP or unlock the segment lock. We're done. 3262 */ 3263 if (seglocked) { 3264 KASSERT(!mutex_owned(vp->v_interlock)); 3265 return error; 3266 } 3267 3268 /* Clean up FIP and send it to disk. */ 3269 lfs_release_finfo(fs); 3270 lfs_writeseg(fs, fs->lfs_sp); 3271 3272 /* 3273 * Remove us from paging queue if we wrote all our pages. 3274 */ 3275 if (origendoffset == 0 || ap->a_flags & PGO_ALLPAGES) { 3276 mutex_enter(&lfs_lock); 3277 if (ip->i_flags & IN_PAGING) { 3278 ip->i_flags &= ~IN_PAGING; 3279 TAILQ_REMOVE(&fs->lfs_pchainhd, ip, i_lfs_pchain); 3280 } 3281 mutex_exit(&lfs_lock); 3282 } 3283 3284 /* 3285 * XXX - with the malloc/copy writeseg, the pages are freed by now 3286 * even if we don't wait (e.g. if we hold a nested lock). This 3287 * will not be true if we stop using malloc/copy. 3288 */ 3289 KASSERT(fs->lfs_sp->seg_flags & SEGM_PROT); 3290 lfs_segunlock(fs); 3291 3292 /* 3293 * Wait for v_numoutput to drop to zero. The seglock should 3294 * take care of this, but there is a slight possibility that 3295 * aiodoned might not have got around to our buffers yet. 3296 */ 3297 if (sync) { 3298 mutex_enter(vp->v_interlock); 3299 while (vp->v_numoutput > 0) { 3300 DLOG((DLOG_PAGE, "lfs_putpages: ino %d sleeping on" 3301 " num %d\n", ip->i_number, vp->v_numoutput)); 3302 cv_wait(&vp->v_cv, vp->v_interlock); 3303 } 3304 mutex_exit(vp->v_interlock); 3305 } 3306 KASSERT(!mutex_owned(vp->v_interlock)); 3307 return error; 3308 } 3309 3310 /* 3311 * Return the last logical file offset that should be written for this file 3312 * if we're doing a write that ends at "size". If writing, we need to know 3313 * about sizes on disk, i.e. fragments if there are any; if reading, we need 3314 * to know about entire blocks. 3315 */ 3316 void 3317 lfs_gop_size(struct vnode *vp, off_t size, off_t *eobp, int flags) 3318 { 3319 struct inode *ip = VTOI(vp); 3320 struct lfs *fs = ip->i_lfs; 3321 daddr_t olbn, nlbn; 3322 3323 olbn = lblkno(fs, ip->i_size); 3324 nlbn = lblkno(fs, size); 3325 if (!(flags & GOP_SIZE_MEM) && nlbn < NDADDR && olbn <= nlbn) { 3326 *eobp = fragroundup(fs, size); 3327 } else { 3328 *eobp = blkroundup(fs, size); 3329 } 3330 } 3331 3332 #ifdef DEBUG 3333 void lfs_dump_vop(void *); 3334 3335 void 3336 lfs_dump_vop(void *v) 3337 { 3338 struct vop_putpages_args /* { 3339 struct vnode *a_vp; 3340 voff_t a_offlo; 3341 voff_t a_offhi; 3342 int a_flags; 3343 } */ *ap = v; 3344 3345 #ifdef DDB 3346 vfs_vnode_print(ap->a_vp, 0, printf); 3347 #endif 3348 lfs_dump_dinode(VTOI(ap->a_vp)->i_din.ffs1_din); 3349 } 3350 #endif 3351 3352 int 3353 lfs_mmap(void *v) 3354 { 3355 struct vop_mmap_args /* { 3356 const struct vnodeop_desc *a_desc; 3357 struct vnode *a_vp; 3358 vm_prot_t a_prot; 3359 kauth_cred_t a_cred; 3360 } */ *ap = v; 3361 3362 if (VTOI(ap->a_vp)->i_number == LFS_IFILE_INUM) 3363 return EOPNOTSUPP; 3364 return ufs_mmap(v); 3365 } 3366
Indexes created Sun Oct 19 02:09:48 GMT 2025