| 1 | /* |
| 2 | * (MPSAFE) |
| 3 | * |
| 4 | * Copyright (c) 1989, 1993 |
| 5 | * The Regents of the University of California. All rights reserved. |
| 6 | * (c) UNIX System Laboratories, Inc. |
| 7 | * All or some portions of this file are derived from material licensed |
| 8 | * to the University of California by American Telephone and Telegraph |
| 9 | * Co. or Unix System Laboratories, Inc. and are reproduced herein with |
| 10 | * the permission of UNIX System Laboratories, Inc. |
| 11 | * |
| 12 | * Redistribution and use in source and binary forms, with or without |
| 13 | * modification, are permitted provided that the following conditions |
| 14 | * are met: |
| 15 | * 1. Redistributions of source code must retain the above copyright |
| 16 | * notice, this list of conditions and the following disclaimer. |
| 17 | * 2. Redistributions in binary form must reproduce the above copyright |
| 18 | * notice, this list of conditions and the following disclaimer in the |
| 19 | * documentation and/or other materials provided with the distribution. |
| 20 | * 3. All advertising materials mentioning features or use of this software |
| 21 | * must display the following acknowledgement: |
| 22 | * This product includes software developed by the University of |
| 23 | * California, Berkeley and its contributors. |
| 24 | * 4. Neither the name of the University nor the names of its contributors |
| 25 | * may be used to endorse or promote products derived from this software |
| 26 | * without specific prior written permission. |
| 27 | * |
| 28 | * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND |
| 29 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 30 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 31 | * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE |
| 32 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 33 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 34 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 35 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 36 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 37 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 38 | * SUCH DAMAGE. |
| 39 | * |
| 40 | * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 |
| 41 | * $FreeBSD: src/sys/kern/vfs_subr.c,v 1.249.2.30 2003/04/04 20:35:57 tegge Exp $ |
| 42 | * $DragonFly: src/sys/kern/vfs_sync.c,v 1.18 2008/05/18 05:54:25 dillon Exp $ |
| 43 | */ |
| 44 | |
| 45 | /* |
| 46 | * External virtual filesystem routines |
| 47 | */ |
| 48 | #include "opt_ddb.h" |
| 49 | |
| 50 | #include <sys/param.h> |
| 51 | #include <sys/systm.h> |
| 52 | #include <sys/buf.h> |
| 53 | #include <sys/conf.h> |
| 54 | #include <sys/dirent.h> |
| 55 | #include <sys/domain.h> |
| 56 | #include <sys/eventhandler.h> |
| 57 | #include <sys/fcntl.h> |
| 58 | #include <sys/kernel.h> |
| 59 | #include <sys/kthread.h> |
| 60 | #include <sys/malloc.h> |
| 61 | #include <sys/mbuf.h> |
| 62 | #include <sys/mount.h> |
| 63 | #include <sys/proc.h> |
| 64 | #include <sys/namei.h> |
| 65 | #include <sys/reboot.h> |
| 66 | #include <sys/socket.h> |
| 67 | #include <sys/stat.h> |
| 68 | #include <sys/sysctl.h> |
| 69 | #include <sys/syslog.h> |
| 70 | #include <sys/vmmeter.h> |
| 71 | #include <sys/vnode.h> |
| 72 | |
| 73 | #include <machine/limits.h> |
| 74 | |
| 75 | #include <vm/vm.h> |
| 76 | #include <vm/vm_object.h> |
| 77 | #include <vm/vm_extern.h> |
| 78 | #include <vm/vm_kern.h> |
| 79 | #include <vm/pmap.h> |
| 80 | #include <vm/vm_map.h> |
| 81 | #include <vm/vm_page.h> |
| 82 | #include <vm/vm_pager.h> |
| 83 | #include <vm/vnode_pager.h> |
| 84 | |
| 85 | #include <sys/buf2.h> |
| 86 | #include <sys/thread2.h> |
| 87 | #include <sys/mplock2.h> |
| 88 | |
| 89 | /* |
| 90 | * The workitem queue. |
| 91 | */ |
| 92 | #define SYNCER_MAXDELAY 32 |
| 93 | static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ |
| 94 | time_t syncdelay = 30; /* max time to delay syncing data */ |
| 95 | SYSCTL_INT(_kern, OID_AUTO, syncdelay, CTLFLAG_RW, |
| 96 | &syncdelay, 0, "VFS data synchronization delay"); |
| 97 | time_t filedelay = 30; /* time to delay syncing files */ |
| 98 | SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, |
| 99 | &filedelay, 0, "File synchronization delay"); |
| 100 | time_t dirdelay = 29; /* time to delay syncing directories */ |
| 101 | SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, |
| 102 | &dirdelay, 0, "Directory synchronization delay"); |
| 103 | time_t metadelay = 28; /* time to delay syncing metadata */ |
| 104 | SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, |
| 105 | &metadelay, 0, "VFS metadata synchronization delay"); |
| 106 | static int rushjob; /* number of slots to run ASAP */ |
| 107 | static int stat_rush_requests; /* number of times I/O speeded up */ |
| 108 | SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, |
| 109 | &stat_rush_requests, 0, ""); |
| 110 | |
| 111 | static int syncer_delayno = 0; |
| 112 | static long syncer_mask; |
| 113 | static struct lwkt_token syncer_token; |
| 114 | LIST_HEAD(synclist, vnode); |
| 115 | static struct synclist *syncer_workitem_pending; |
| 116 | |
| 117 | /* |
| 118 | * Called from vfsinit() |
| 119 | */ |
| 120 | void |
| 121 | vfs_sync_init(void) |
| 122 | { |
| 123 | syncer_workitem_pending = hashinit(syncer_maxdelay, M_DEVBUF, |
| 124 | &syncer_mask); |
| 125 | syncer_maxdelay = syncer_mask + 1; |
| 126 | lwkt_token_init(&syncer_token, 1, "syncer"); |
| 127 | } |
| 128 | |
| 129 | /* |
| 130 | * The workitem queue. |
| 131 | * |
| 132 | * It is useful to delay writes of file data and filesystem metadata |
| 133 | * for tens of seconds so that quickly created and deleted files need |
| 134 | * not waste disk bandwidth being created and removed. To realize this, |
| 135 | * we append vnodes to a "workitem" queue. When running with a soft |
| 136 | * updates implementation, most pending metadata dependencies should |
| 137 | * not wait for more than a few seconds. Thus, mounted on block devices |
| 138 | * are delayed only about a half the time that file data is delayed. |
| 139 | * Similarly, directory updates are more critical, so are only delayed |
| 140 | * about a third the time that file data is delayed. Thus, there are |
| 141 | * SYNCER_MAXDELAY queues that are processed round-robin at a rate of |
| 142 | * one each second (driven off the filesystem syncer process). The |
| 143 | * syncer_delayno variable indicates the next queue that is to be processed. |
| 144 | * Items that need to be processed soon are placed in this queue: |
| 145 | * |
| 146 | * syncer_workitem_pending[syncer_delayno] |
| 147 | * |
| 148 | * A delay of fifteen seconds is done by placing the request fifteen |
| 149 | * entries later in the queue: |
| 150 | * |
| 151 | * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] |
| 152 | * |
| 153 | */ |
| 154 | |
| 155 | /* |
| 156 | * Add an item to the syncer work queue. |
| 157 | * |
| 158 | * WARNING: Cannot get vp->v_token here if not already held, we must |
| 159 | * depend on the syncer_token (which might already be held by |
| 160 | * the caller) to protect v_synclist and VONWORKLST. |
| 161 | * |
| 162 | * MPSAFE |
| 163 | */ |
| 164 | void |
| 165 | vn_syncer_add(struct vnode *vp, int delay) |
| 166 | { |
| 167 | int slot; |
| 168 | |
| 169 | lwkt_gettoken(&syncer_token); |
| 170 | |
| 171 | if (vp->v_flag & VONWORKLST) |
| 172 | LIST_REMOVE(vp, v_synclist); |
| 173 | if (delay > syncer_maxdelay - 2) |
| 174 | delay = syncer_maxdelay - 2; |
| 175 | slot = (syncer_delayno + delay) & syncer_mask; |
| 176 | |
| 177 | LIST_INSERT_HEAD(&syncer_workitem_pending[slot], vp, v_synclist); |
| 178 | vsetflags(vp, VONWORKLST); |
| 179 | |
| 180 | lwkt_reltoken(&syncer_token); |
| 181 | } |
| 182 | |
| 183 | /* |
| 184 | * Removes the vnode from the syncer list. Since we might block while |
| 185 | * acquiring the syncer_token we have to recheck conditions. |
| 186 | * |
| 187 | * vp->v_token held on call |
| 188 | */ |
| 189 | void |
| 190 | vn_syncer_remove(struct vnode *vp) |
| 191 | { |
| 192 | lwkt_gettoken(&syncer_token); |
| 193 | |
| 194 | if ((vp->v_flag & VONWORKLST) && RB_EMPTY(&vp->v_rbdirty_tree)) { |
| 195 | vclrflags(vp, VONWORKLST); |
| 196 | LIST_REMOVE(vp, v_synclist); |
| 197 | } |
| 198 | |
| 199 | lwkt_reltoken(&syncer_token); |
| 200 | } |
| 201 | |
| 202 | struct thread *updatethread; |
| 203 | |
| 204 | /* |
| 205 | * System filesystem synchronizer daemon. |
| 206 | */ |
| 207 | static void |
| 208 | syncer_thread(void) |
| 209 | { |
| 210 | struct thread *td = curthread; |
| 211 | struct synclist *slp; |
| 212 | struct vnode *vp; |
| 213 | long starttime; |
| 214 | |
| 215 | EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td, |
| 216 | SHUTDOWN_PRI_LAST); |
| 217 | for (;;) { |
| 218 | kproc_suspend_loop(); |
| 219 | |
| 220 | starttime = time_second; |
| 221 | lwkt_gettoken(&syncer_token); |
| 222 | |
| 223 | /* |
| 224 | * Push files whose dirty time has expired. Be careful |
| 225 | * of interrupt race on slp queue. |
| 226 | */ |
| 227 | slp = &syncer_workitem_pending[syncer_delayno]; |
| 228 | syncer_delayno += 1; |
| 229 | if (syncer_delayno == syncer_maxdelay) |
| 230 | syncer_delayno = 0; |
| 231 | |
| 232 | while ((vp = LIST_FIRST(slp)) != NULL) { |
| 233 | if (vget(vp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { |
| 234 | VOP_FSYNC(vp, MNT_LAZY, 0); |
| 235 | vput(vp); |
| 236 | } |
| 237 | |
| 238 | /* |
| 239 | * vp is stale but can still be used if we can |
| 240 | * verify that it remains at the head of the list. |
| 241 | * Be careful not to try to get vp->v_token as |
| 242 | * vp can become stale if this blocks. |
| 243 | * |
| 244 | * If the vp is still at the head of the list were |
| 245 | * unable to completely flush it and move it to |
| 246 | * a later slot to give other vnodes a fair shot. |
| 247 | * |
| 248 | * Note that v_tag VT_VFS vnodes can remain on the |
| 249 | * worklist with no dirty blocks, but sync_fsync() |
| 250 | * moves it to a later slot so we will never see it |
| 251 | * here. |
| 252 | * |
| 253 | * It is possible to race a vnode with no dirty |
| 254 | * buffers being removed from the list. If this |
| 255 | * occurs we will move the vnode in the synclist |
| 256 | * and then the other thread will remove it. Do |
| 257 | * not try to remove it here. |
| 258 | */ |
| 259 | if (LIST_FIRST(slp) == vp) |
| 260 | vn_syncer_add(vp, syncdelay); |
| 261 | } |
| 262 | lwkt_reltoken(&syncer_token); |
| 263 | |
| 264 | /* |
| 265 | * Do sync processing for each mount. |
| 266 | */ |
| 267 | bio_ops_sync(NULL); |
| 268 | |
| 269 | /* |
| 270 | * The variable rushjob allows the kernel to speed up the |
| 271 | * processing of the filesystem syncer process. A rushjob |
| 272 | * value of N tells the filesystem syncer to process the next |
| 273 | * N seconds worth of work on its queue ASAP. Currently rushjob |
| 274 | * is used by the soft update code to speed up the filesystem |
| 275 | * syncer process when the incore state is getting so far |
| 276 | * ahead of the disk that the kernel memory pool is being |
| 277 | * threatened with exhaustion. |
| 278 | */ |
| 279 | if (rushjob > 0) { |
| 280 | atomic_subtract_int(&rushjob, 1); |
| 281 | continue; |
| 282 | } |
| 283 | /* |
| 284 | * If it has taken us less than a second to process the |
| 285 | * current work, then wait. Otherwise start right over |
| 286 | * again. We can still lose time if any single round |
| 287 | * takes more than two seconds, but it does not really |
| 288 | * matter as we are just trying to generally pace the |
| 289 | * filesystem activity. |
| 290 | */ |
| 291 | if (time_second == starttime) |
| 292 | tsleep(&lbolt_syncer, 0, "syncer", 0); |
| 293 | } |
| 294 | } |
| 295 | |
| 296 | static struct kproc_desc up_kp = { |
| 297 | "syncer", |
| 298 | syncer_thread, |
| 299 | &updatethread |
| 300 | }; |
| 301 | SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) |
| 302 | |
| 303 | /* |
| 304 | * Request the syncer daemon to speed up its work. |
| 305 | * We never push it to speed up more than half of its |
| 306 | * normal turn time, otherwise it could take over the cpu. |
| 307 | */ |
| 308 | int |
| 309 | speedup_syncer(void) |
| 310 | { |
| 311 | /* |
| 312 | * Don't bother protecting the test. unsleep_and_wakeup_thread() |
| 313 | * will only do something real if the thread is in the right state. |
| 314 | */ |
| 315 | wakeup(&lbolt_syncer); |
| 316 | if (rushjob < syncdelay / 2) { |
| 317 | atomic_add_int(&rushjob, 1); |
| 318 | stat_rush_requests += 1; |
| 319 | return (1); |
| 320 | } |
| 321 | return(0); |
| 322 | } |
| 323 | |
| 324 | /* |
| 325 | * Routine to create and manage a filesystem syncer vnode. |
| 326 | */ |
| 327 | static int sync_close(struct vop_close_args *); |
| 328 | static int sync_fsync(struct vop_fsync_args *); |
| 329 | static int sync_inactive(struct vop_inactive_args *); |
| 330 | static int sync_reclaim (struct vop_reclaim_args *); |
| 331 | static int sync_print(struct vop_print_args *); |
| 332 | |
| 333 | static struct vop_ops sync_vnode_vops = { |
| 334 | .vop_default = vop_eopnotsupp, |
| 335 | .vop_close = sync_close, |
| 336 | .vop_fsync = sync_fsync, |
| 337 | .vop_inactive = sync_inactive, |
| 338 | .vop_reclaim = sync_reclaim, |
| 339 | .vop_print = sync_print, |
| 340 | }; |
| 341 | |
| 342 | static struct vop_ops *sync_vnode_vops_p = &sync_vnode_vops; |
| 343 | |
| 344 | VNODEOP_SET(sync_vnode_vops); |
| 345 | |
| 346 | /* |
| 347 | * Create a new filesystem syncer vnode for the specified mount point. |
| 348 | * This vnode is placed on the worklist and is responsible for sync'ing |
| 349 | * the filesystem. |
| 350 | * |
| 351 | * NOTE: read-only mounts are also placed on the worklist. The filesystem |
| 352 | * sync code is also responsible for cleaning up vnodes. |
| 353 | */ |
| 354 | int |
| 355 | vfs_allocate_syncvnode(struct mount *mp) |
| 356 | { |
| 357 | struct vnode *vp; |
| 358 | static long start, incr, next; |
| 359 | int error; |
| 360 | |
| 361 | /* Allocate a new vnode */ |
| 362 | error = getspecialvnode(VT_VFS, mp, &sync_vnode_vops_p, &vp, 0, 0); |
| 363 | if (error) { |
| 364 | mp->mnt_syncer = NULL; |
| 365 | return (error); |
| 366 | } |
| 367 | vp->v_type = VNON; |
| 368 | /* |
| 369 | * Place the vnode onto the syncer worklist. We attempt to |
| 370 | * scatter them about on the list so that they will go off |
| 371 | * at evenly distributed times even if all the filesystems |
| 372 | * are mounted at once. |
| 373 | */ |
| 374 | next += incr; |
| 375 | if (next == 0 || next > syncer_maxdelay) { |
| 376 | start /= 2; |
| 377 | incr /= 2; |
| 378 | if (start == 0) { |
| 379 | start = syncer_maxdelay / 2; |
| 380 | incr = syncer_maxdelay; |
| 381 | } |
| 382 | next = start; |
| 383 | } |
| 384 | vn_syncer_add(vp, syncdelay > 0 ? next % syncdelay : 0); |
| 385 | |
| 386 | /* |
| 387 | * The mnt_syncer field inherits the vnode reference, which is |
| 388 | * held until later decomissioning. |
| 389 | */ |
| 390 | mp->mnt_syncer = vp; |
| 391 | vx_unlock(vp); |
| 392 | return (0); |
| 393 | } |
| 394 | |
| 395 | static int |
| 396 | sync_close(struct vop_close_args *ap) |
| 397 | { |
| 398 | return (0); |
| 399 | } |
| 400 | |
| 401 | /* |
| 402 | * Do a lazy sync of the filesystem. |
| 403 | * |
| 404 | * sync_fsync { struct vnode *a_vp, int a_waitfor } |
| 405 | */ |
| 406 | static int |
| 407 | sync_fsync(struct vop_fsync_args *ap) |
| 408 | { |
| 409 | struct vnode *syncvp = ap->a_vp; |
| 410 | struct mount *mp = syncvp->v_mount; |
| 411 | int asyncflag; |
| 412 | |
| 413 | /* |
| 414 | * We only need to do something if this is a lazy evaluation. |
| 415 | */ |
| 416 | if ((ap->a_waitfor & MNT_LAZY) == 0) |
| 417 | return (0); |
| 418 | |
| 419 | /* |
| 420 | * Move ourselves to the back of the sync list. |
| 421 | */ |
| 422 | vn_syncer_add(syncvp, syncdelay); |
| 423 | |
| 424 | /* |
| 425 | * Walk the list of vnodes pushing all that are dirty and |
| 426 | * not already on the sync list, and freeing vnodes which have |
| 427 | * no refs and whos VM objects are empty. vfs_msync() handles |
| 428 | * the VM issues and must be called whether the mount is readonly |
| 429 | * or not. |
| 430 | */ |
| 431 | if (vfs_busy(mp, LK_NOWAIT) != 0) |
| 432 | return (0); |
| 433 | if (mp->mnt_flag & MNT_RDONLY) { |
| 434 | vfs_msync(mp, MNT_NOWAIT); |
| 435 | } else { |
| 436 | asyncflag = mp->mnt_flag & MNT_ASYNC; |
| 437 | mp->mnt_flag &= ~MNT_ASYNC; /* ZZZ hack */ |
| 438 | vfs_msync(mp, MNT_NOWAIT); |
| 439 | VFS_SYNC(mp, MNT_NOWAIT | MNT_LAZY); |
| 440 | if (asyncflag) |
| 441 | mp->mnt_flag |= MNT_ASYNC; |
| 442 | } |
| 443 | vfs_unbusy(mp); |
| 444 | return (0); |
| 445 | } |
| 446 | |
| 447 | /* |
| 448 | * The syncer vnode is no longer referenced. |
| 449 | * |
| 450 | * sync_inactive { struct vnode *a_vp, struct proc *a_p } |
| 451 | */ |
| 452 | static int |
| 453 | sync_inactive(struct vop_inactive_args *ap) |
| 454 | { |
| 455 | vgone_vxlocked(ap->a_vp); |
| 456 | return (0); |
| 457 | } |
| 458 | |
| 459 | /* |
| 460 | * The syncer vnode is no longer needed and is being decommissioned. |
| 461 | * This can only occur when the last reference has been released on |
| 462 | * mp->mnt_syncer, so mp->mnt_syncer had better be NULL. |
| 463 | * |
| 464 | * Modifications to the worklist must be protected with a critical |
| 465 | * section. |
| 466 | * |
| 467 | * sync_reclaim { struct vnode *a_vp } |
| 468 | */ |
| 469 | static int |
| 470 | sync_reclaim(struct vop_reclaim_args *ap) |
| 471 | { |
| 472 | struct vnode *vp = ap->a_vp; |
| 473 | |
| 474 | lwkt_gettoken(&syncer_token); |
| 475 | KKASSERT(vp->v_mount->mnt_syncer != vp); |
| 476 | if (vp->v_flag & VONWORKLST) { |
| 477 | LIST_REMOVE(vp, v_synclist); |
| 478 | vclrflags(vp, VONWORKLST); |
| 479 | } |
| 480 | lwkt_reltoken(&syncer_token); |
| 481 | |
| 482 | return (0); |
| 483 | } |
| 484 | |
| 485 | /* |
| 486 | * Print out a syncer vnode. |
| 487 | * |
| 488 | * sync_print { struct vnode *a_vp } |
| 489 | */ |
| 490 | static int |
| 491 | sync_print(struct vop_print_args *ap) |
| 492 | { |
| 493 | struct vnode *vp = ap->a_vp; |
| 494 | |
| 495 | kprintf("syncer vnode"); |
| 496 | lockmgr_printinfo(&vp->v_lock); |
| 497 | kprintf("\n"); |
| 498 | return (0); |
| 499 | } |
| 500 | |