1 /* 2 * Copyright (c) 2004 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * $DragonFly: src/sys/kern/vfs_journal.c,v 1.18 2005/07/13 01:58:20 dillon Exp $ 35 */ 36 /* 37 * Each mount point may have zero or more independantly configured journals 38 * attached to it. Each journal is represented by a memory FIFO and worker 39 * thread. Journal events are streamed through the FIFO to the thread, 40 * batched up (typically on one-second intervals), and written out by the 41 * thread. 42 * 43 * Journal vnode ops are executed instead of mnt_vn_norm_ops when one or 44 * more journals have been installed on a mount point. It becomes the 45 * responsibility of the journal op to call the underlying normal op as 46 * appropriate. 47 * 48 * The journaling protocol is intended to evolve into a two-way stream 49 * whereby transaction IDs can be acknowledged by the journaling target 50 * when the data has been committed to hard storage. Both implicit and 51 * explicit acknowledgement schemes will be supported, depending on the 52 * sophistication of the journaling stream, plus resynchronization and 53 * restart when a journaling stream is interrupted. This information will 54 * also be made available to journaling-aware filesystems to allow better 55 * management of their own physical storage synchronization mechanisms as 56 * well as to allow such filesystems to take direct advantage of the kernel's 57 * journaling layer so they don't have to roll their own. 58 * 59 * In addition, the worker thread will have access to much larger 60 * spooling areas then the memory buffer is able to provide by e.g. 61 * reserving swap space, in order to absorb potentially long interruptions 62 * of off-site journaling streams, and to prevent 'slow' off-site linkages 63 * from radically slowing down local filesystem operations. 64 * 65 * Because of the non-trivial algorithms the journaling system will be 66 * required to support, use of a worker thread is mandatory. Efficiencies 67 * are maintained by utilitizing the memory FIFO to batch transactions when 68 * possible, reducing the number of gratuitous thread switches and taking 69 * advantage of cpu caches through the use of shorter batched code paths 70 * rather then trying to do everything in the context of the process 71 * originating the filesystem op. In the future the memory FIFO can be 72 * made per-cpu to remove BGL or other locking requirements. 73 */ 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/buf.h> 77 #include <sys/conf.h> 78 #include <sys/kernel.h> 79 #include <sys/queue.h> 80 #include <sys/lock.h> 81 #include <sys/malloc.h> 82 #include <sys/mount.h> 83 #include <sys/unistd.h> 84 #include <sys/vnode.h> 85 #include <sys/poll.h> 86 #include <sys/mountctl.h> 87 #include <sys/journal.h> 88 #include <sys/file.h> 89 #include <sys/proc.h> 90 #include <sys/msfbuf.h> 91 #include <sys/socket.h> 92 #include <sys/socketvar.h> 93 94 #include <machine/limits.h> 95 96 #include <vm/vm.h> 97 #include <vm/vm_object.h> 98 #include <vm/vm_page.h> 99 #include <vm/vm_pager.h> 100 #include <vm/vnode_pager.h> 101 102 #include <sys/file2.h> 103 #include <sys/thread2.h> 104 105 static int journal_attach(struct mount *mp); 106 static void journal_detach(struct mount *mp); 107 static int journal_install_vfs_journal(struct mount *mp, struct file *fp, 108 const struct mountctl_install_journal *info); 109 static int journal_restart_vfs_journal(struct mount *mp, struct file *fp, 110 const struct mountctl_restart_journal *info); 111 static int journal_remove_vfs_journal(struct mount *mp, 112 const struct mountctl_remove_journal *info); 113 static int journal_restart(struct mount *mp, struct file *fp, 114 struct journal *jo, int flags); 115 static int journal_destroy(struct mount *mp, struct journal *jo, int flags); 116 static int journal_resync_vfs_journal(struct mount *mp, const void *ctl); 117 static int journal_status_vfs_journal(struct mount *mp, 118 const struct mountctl_status_journal *info, 119 struct mountctl_journal_ret_status *rstat, 120 int buflen, int *res); 121 static void journal_create_threads(struct journal *jo); 122 static void journal_destroy_threads(struct journal *jo, int flags); 123 static void journal_wthread(void *info); 124 static void journal_rthread(void *info); 125 126 static void *journal_reserve(struct journal *jo, 127 struct journal_rawrecbeg **rawpp, 128 int16_t streamid, int bytes); 129 static void *journal_extend(struct journal *jo, 130 struct journal_rawrecbeg **rawpp, 131 int truncbytes, int bytes, int *newstreamrecp); 132 static void journal_abort(struct journal *jo, 133 struct journal_rawrecbeg **rawpp); 134 static void journal_commit(struct journal *jo, 135 struct journal_rawrecbeg **rawpp, 136 int bytes, int closeout); 137 138 static void jrecord_init(struct journal *jo, 139 struct jrecord *jrec, int16_t streamid); 140 static struct journal_subrecord *jrecord_push( 141 struct jrecord *jrec, int16_t rectype); 142 static void jrecord_pop(struct jrecord *jrec, struct journal_subrecord *parent); 143 static struct journal_subrecord *jrecord_write(struct jrecord *jrec, 144 int16_t rectype, int bytes); 145 static void jrecord_data(struct jrecord *jrec, const void *buf, int bytes); 146 static void jrecord_done(struct jrecord *jrec, int abortit); 147 148 static int journal_setattr(struct vop_setattr_args *ap); 149 static int journal_write(struct vop_write_args *ap); 150 static int journal_fsync(struct vop_fsync_args *ap); 151 static int journal_putpages(struct vop_putpages_args *ap); 152 static int journal_setacl(struct vop_setacl_args *ap); 153 static int journal_setextattr(struct vop_setextattr_args *ap); 154 static int journal_ncreate(struct vop_ncreate_args *ap); 155 static int journal_nmknod(struct vop_nmknod_args *ap); 156 static int journal_nlink(struct vop_nlink_args *ap); 157 static int journal_nsymlink(struct vop_nsymlink_args *ap); 158 static int journal_nwhiteout(struct vop_nwhiteout_args *ap); 159 static int journal_nremove(struct vop_nremove_args *ap); 160 static int journal_nmkdir(struct vop_nmkdir_args *ap); 161 static int journal_nrmdir(struct vop_nrmdir_args *ap); 162 static int journal_nrename(struct vop_nrename_args *ap); 163 164 static struct vnodeopv_entry_desc journal_vnodeop_entries[] = { 165 { &vop_default_desc, vop_journal_operate_ap }, 166 { &vop_mountctl_desc, (void *)journal_mountctl }, 167 { &vop_setattr_desc, (void *)journal_setattr }, 168 { &vop_write_desc, (void *)journal_write }, 169 { &vop_fsync_desc, (void *)journal_fsync }, 170 { &vop_putpages_desc, (void *)journal_putpages }, 171 { &vop_setacl_desc, (void *)journal_setacl }, 172 { &vop_setextattr_desc, (void *)journal_setextattr }, 173 { &vop_ncreate_desc, (void *)journal_ncreate }, 174 { &vop_nmknod_desc, (void *)journal_nmknod }, 175 { &vop_nlink_desc, (void *)journal_nlink }, 176 { &vop_nsymlink_desc, (void *)journal_nsymlink }, 177 { &vop_nwhiteout_desc, (void *)journal_nwhiteout }, 178 { &vop_nremove_desc, (void *)journal_nremove }, 179 { &vop_nmkdir_desc, (void *)journal_nmkdir }, 180 { &vop_nrmdir_desc, (void *)journal_nrmdir }, 181 { &vop_nrename_desc, (void *)journal_nrename }, 182 { NULL, NULL } 183 }; 184 185 static MALLOC_DEFINE(M_JOURNAL, "journal", "Journaling structures"); 186 static MALLOC_DEFINE(M_JFIFO, "journal-fifo", "Journal FIFO"); 187 188 int 189 journal_mountctl(struct vop_mountctl_args *ap) 190 { 191 struct mount *mp; 192 int error = 0; 193 194 mp = ap->a_head.a_ops->vv_mount; 195 KKASSERT(mp); 196 197 if (mp->mnt_vn_journal_ops == NULL) { 198 switch(ap->a_op) { 199 case MOUNTCTL_INSTALL_VFS_JOURNAL: 200 error = journal_attach(mp); 201 if (error == 0 && ap->a_ctllen != sizeof(struct mountctl_install_journal)) 202 error = EINVAL; 203 if (error == 0 && ap->a_fp == NULL) 204 error = EBADF; 205 if (error == 0) 206 error = journal_install_vfs_journal(mp, ap->a_fp, ap->a_ctl); 207 if (TAILQ_EMPTY(&mp->mnt_jlist)) 208 journal_detach(mp); 209 break; 210 case MOUNTCTL_RESTART_VFS_JOURNAL: 211 case MOUNTCTL_REMOVE_VFS_JOURNAL: 212 case MOUNTCTL_RESYNC_VFS_JOURNAL: 213 case MOUNTCTL_STATUS_VFS_JOURNAL: 214 error = ENOENT; 215 break; 216 default: 217 error = EOPNOTSUPP; 218 break; 219 } 220 } else { 221 switch(ap->a_op) { 222 case MOUNTCTL_INSTALL_VFS_JOURNAL: 223 if (ap->a_ctllen != sizeof(struct mountctl_install_journal)) 224 error = EINVAL; 225 if (error == 0 && ap->a_fp == NULL) 226 error = EBADF; 227 if (error == 0) 228 error = journal_install_vfs_journal(mp, ap->a_fp, ap->a_ctl); 229 break; 230 case MOUNTCTL_RESTART_VFS_JOURNAL: 231 if (ap->a_ctllen != sizeof(struct mountctl_restart_journal)) 232 error = EINVAL; 233 if (error == 0 && ap->a_fp == NULL) 234 error = EBADF; 235 if (error == 0) 236 error = journal_restart_vfs_journal(mp, ap->a_fp, ap->a_ctl); 237 break; 238 case MOUNTCTL_REMOVE_VFS_JOURNAL: 239 if (ap->a_ctllen != sizeof(struct mountctl_remove_journal)) 240 error = EINVAL; 241 if (error == 0) 242 error = journal_remove_vfs_journal(mp, ap->a_ctl); 243 if (TAILQ_EMPTY(&mp->mnt_jlist)) 244 journal_detach(mp); 245 break; 246 case MOUNTCTL_RESYNC_VFS_JOURNAL: 247 if (ap->a_ctllen != 0) 248 error = EINVAL; 249 error = journal_resync_vfs_journal(mp, ap->a_ctl); 250 break; 251 case MOUNTCTL_STATUS_VFS_JOURNAL: 252 if (ap->a_ctllen != sizeof(struct mountctl_status_journal)) 253 error = EINVAL; 254 if (error == 0) { 255 error = journal_status_vfs_journal(mp, ap->a_ctl, 256 ap->a_buf, ap->a_buflen, ap->a_res); 257 } 258 break; 259 default: 260 error = EOPNOTSUPP; 261 break; 262 } 263 } 264 return (error); 265 } 266 267 /* 268 * High level mount point setup. When a 269 */ 270 static int 271 journal_attach(struct mount *mp) 272 { 273 vfs_add_vnodeops(mp, &mp->mnt_vn_journal_ops, journal_vnodeop_entries); 274 return(0); 275 } 276 277 static void 278 journal_detach(struct mount *mp) 279 { 280 if (mp->mnt_vn_journal_ops) 281 vfs_rm_vnodeops(&mp->mnt_vn_journal_ops); 282 } 283 284 /* 285 * Install a journal on a mount point. Each journal has an associated worker 286 * thread which is responsible for buffering and spooling the data to the 287 * target. A mount point may have multiple journals attached to it. An 288 * initial start record is generated when the journal is associated. 289 */ 290 static int 291 journal_install_vfs_journal(struct mount *mp, struct file *fp, 292 const struct mountctl_install_journal *info) 293 { 294 struct journal *jo; 295 struct jrecord jrec; 296 int error = 0; 297 int size; 298 299 jo = malloc(sizeof(struct journal), M_JOURNAL, M_WAITOK|M_ZERO); 300 bcopy(info->id, jo->id, sizeof(jo->id)); 301 jo->flags = info->flags & ~(MC_JOURNAL_WACTIVE | MC_JOURNAL_RACTIVE | 302 MC_JOURNAL_STOP_REQ); 303 304 /* 305 * Memory FIFO size, round to nearest power of 2 306 */ 307 if (info->membufsize) { 308 if (info->membufsize < 65536) 309 size = 65536; 310 else if (info->membufsize > 128 * 1024 * 1024) 311 size = 128 * 1024 * 1024; 312 else 313 size = (int)info->membufsize; 314 } else { 315 size = 1024 * 1024; 316 } 317 jo->fifo.size = 1; 318 while (jo->fifo.size < size) 319 jo->fifo.size <<= 1; 320 321 /* 322 * Other parameters. If not specified the starting transaction id 323 * will be the current date. 324 */ 325 if (info->transid) { 326 jo->transid = info->transid; 327 } else { 328 struct timespec ts; 329 getnanotime(&ts); 330 jo->transid = ((int64_t)ts.tv_sec << 30) | ts.tv_nsec; 331 } 332 333 jo->fp = fp; 334 335 /* 336 * Allocate the memory FIFO 337 */ 338 jo->fifo.mask = jo->fifo.size - 1; 339 jo->fifo.membase = malloc(jo->fifo.size, M_JFIFO, M_WAITOK|M_ZERO|M_NULLOK); 340 if (jo->fifo.membase == NULL) 341 error = ENOMEM; 342 343 /* 344 * Create the worker threads and generate the association record. 345 */ 346 if (error) { 347 free(jo, M_JOURNAL); 348 } else { 349 fhold(fp); 350 journal_create_threads(jo); 351 jrecord_init(jo, &jrec, JREC_STREAMID_DISCONT); 352 jrecord_write(&jrec, JTYPE_ASSOCIATE, 0); 353 jrecord_done(&jrec, 0); 354 TAILQ_INSERT_TAIL(&mp->mnt_jlist, jo, jentry); 355 } 356 return(error); 357 } 358 359 /* 360 * Restart a journal with a new descriptor. The existing reader and writer 361 * threads are terminated and a new descriptor is associated with the 362 * journal. The FIFO rindex is reset to xindex and the threads are then 363 * restarted. 364 */ 365 static int 366 journal_restart_vfs_journal(struct mount *mp, struct file *fp, 367 const struct mountctl_restart_journal *info) 368 { 369 struct journal *jo; 370 int error; 371 372 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 373 if (bcmp(jo->id, info->id, sizeof(jo->id)) == 0) 374 break; 375 } 376 if (jo) 377 error = journal_restart(mp, fp, jo, info->flags); 378 else 379 error = EINVAL; 380 return (error); 381 } 382 383 static int 384 journal_restart(struct mount *mp, struct file *fp, 385 struct journal *jo, int flags) 386 { 387 /* 388 * XXX lock the jo 389 */ 390 391 #if 0 392 /* 393 * Record the fact that we are doing a restart in the journal. 394 * XXX it isn't safe to do this if the journal is being restarted 395 * because it was locked up and the writer thread has already exited. 396 */ 397 jrecord_init(jo, &jrec, JREC_STREAMID_RESTART); 398 jrecord_write(&jrec, JTYPE_DISASSOCIATE, 0); 399 jrecord_done(&jrec, 0); 400 #endif 401 402 /* 403 * Stop the reader and writer threads and clean up the current 404 * descriptor. 405 */ 406 printf("RESTART WITH FP %p KILLING %p\n", fp, jo->fp); 407 journal_destroy_threads(jo, flags); 408 409 if (jo->fp) 410 fdrop(jo->fp, curthread); 411 412 /* 413 * Associate the new descriptor, reset the FIFO index, and recreate 414 * the threads. 415 */ 416 fhold(fp); 417 jo->fp = fp; 418 jo->fifo.rindex = jo->fifo.xindex; 419 journal_create_threads(jo); 420 421 return(0); 422 } 423 424 /* 425 * Disassociate a journal from a mount point and terminate its worker thread. 426 * A final termination record is written out before the file pointer is 427 * dropped. 428 */ 429 static int 430 journal_remove_vfs_journal(struct mount *mp, 431 const struct mountctl_remove_journal *info) 432 { 433 struct journal *jo; 434 int error; 435 436 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 437 if (bcmp(jo->id, info->id, sizeof(jo->id)) == 0) 438 break; 439 } 440 if (jo) 441 error = journal_destroy(mp, jo, info->flags); 442 else 443 error = EINVAL; 444 return (error); 445 } 446 447 /* 448 * Remove all journals associated with a mount point. Usually called 449 * by the umount code. 450 */ 451 void 452 journal_remove_all_journals(struct mount *mp, int flags) 453 { 454 struct journal *jo; 455 456 while ((jo = TAILQ_FIRST(&mp->mnt_jlist)) != NULL) { 457 journal_destroy(mp, jo, flags); 458 } 459 } 460 461 static int 462 journal_destroy(struct mount *mp, struct journal *jo, int flags) 463 { 464 struct jrecord jrec; 465 466 TAILQ_REMOVE(&mp->mnt_jlist, jo, jentry); 467 468 jrecord_init(jo, &jrec, JREC_STREAMID_DISCONT); 469 jrecord_write(&jrec, JTYPE_DISASSOCIATE, 0); 470 jrecord_done(&jrec, 0); 471 472 journal_destroy_threads(jo, flags); 473 474 if (jo->fp) 475 fdrop(jo->fp, curthread); 476 if (jo->fifo.membase) 477 free(jo->fifo.membase, M_JFIFO); 478 free(jo, M_JOURNAL); 479 return(0); 480 } 481 482 static int 483 journal_resync_vfs_journal(struct mount *mp, const void *ctl) 484 { 485 return(EINVAL); 486 } 487 488 static int 489 journal_status_vfs_journal(struct mount *mp, 490 const struct mountctl_status_journal *info, 491 struct mountctl_journal_ret_status *rstat, 492 int buflen, int *res) 493 { 494 struct journal *jo; 495 int error = 0; 496 int index; 497 498 index = 0; 499 *res = 0; 500 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 501 if (info->index == MC_JOURNAL_INDEX_ID) { 502 if (bcmp(jo->id, info->id, sizeof(jo->id)) != 0) 503 continue; 504 } else if (info->index >= 0) { 505 if (info->index < index) 506 continue; 507 } else if (info->index != MC_JOURNAL_INDEX_ALL) { 508 continue; 509 } 510 if (buflen < sizeof(*rstat)) { 511 if (*res) 512 rstat[-1].flags |= MC_JOURNAL_STATUS_MORETOCOME; 513 else 514 error = EINVAL; 515 break; 516 } 517 bzero(rstat, sizeof(*rstat)); 518 rstat->recsize = sizeof(*rstat); 519 bcopy(jo->id, rstat->id, sizeof(jo->id)); 520 rstat->index = index; 521 rstat->membufsize = jo->fifo.size; 522 rstat->membufused = jo->fifo.windex - jo->fifo.xindex; 523 rstat->membufunacked = jo->fifo.rindex - jo->fifo.xindex; 524 rstat->bytessent = jo->total_acked; 525 rstat->fifostalls = jo->fifostalls; 526 ++rstat; 527 ++index; 528 *res += sizeof(*rstat); 529 buflen -= sizeof(*rstat); 530 } 531 return(error); 532 } 533 534 static void 535 journal_create_threads(struct journal *jo) 536 { 537 jo->flags &= ~(MC_JOURNAL_STOP_REQ | MC_JOURNAL_STOP_IMM); 538 jo->flags |= MC_JOURNAL_WACTIVE; 539 lwkt_create(journal_wthread, jo, NULL, &jo->wthread, 540 TDF_STOPREQ, -1, "journal w:%.*s", JIDMAX, jo->id); 541 lwkt_setpri(&jo->wthread, TDPRI_KERN_DAEMON); 542 lwkt_schedule(&jo->wthread); 543 544 if (jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) { 545 jo->flags |= MC_JOURNAL_RACTIVE; 546 lwkt_create(journal_rthread, jo, NULL, &jo->rthread, 547 TDF_STOPREQ, -1, "journal r:%.*s", JIDMAX, jo->id); 548 lwkt_setpri(&jo->rthread, TDPRI_KERN_DAEMON); 549 lwkt_schedule(&jo->rthread); 550 } 551 } 552 553 static void 554 journal_destroy_threads(struct journal *jo, int flags) 555 { 556 int wcount; 557 558 jo->flags |= MC_JOURNAL_STOP_REQ | (flags & MC_JOURNAL_STOP_IMM); 559 wakeup(&jo->fifo); 560 wcount = 0; 561 while (jo->flags & (MC_JOURNAL_WACTIVE | MC_JOURNAL_RACTIVE)) { 562 tsleep(jo, 0, "jwait", hz); 563 if (++wcount % 10 == 0) { 564 printf("Warning: journal %s waiting for descriptors to close\n", 565 jo->id); 566 } 567 } 568 569 /* 570 * XXX SMP - threads should move to cpu requesting the restart or 571 * termination before finishing up to properly interlock. 572 */ 573 tsleep(jo, 0, "jwait", hz); 574 lwkt_free_thread(&jo->wthread); 575 if (jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) 576 lwkt_free_thread(&jo->rthread); 577 } 578 579 /* 580 * The per-journal worker thread is responsible for writing out the 581 * journal's FIFO to the target stream. 582 */ 583 static void 584 journal_wthread(void *info) 585 { 586 struct journal *jo = info; 587 struct journal_rawrecbeg *rawp; 588 int bytes; 589 int error; 590 int avail; 591 int res; 592 593 for (;;) { 594 /* 595 * Calculate the number of bytes available to write. This buffer 596 * area may contain reserved records so we can't just write it out 597 * without further checks. 598 */ 599 bytes = jo->fifo.windex - jo->fifo.rindex; 600 601 /* 602 * sleep if no bytes are available or if an incomplete record is 603 * encountered (it needs to be filled in before we can write it 604 * out), and skip any pad records that we encounter. 605 */ 606 if (bytes == 0) { 607 if (jo->flags & MC_JOURNAL_STOP_REQ) 608 break; 609 tsleep(&jo->fifo, 0, "jfifo", hz); 610 continue; 611 } 612 613 /* 614 * Sleep if we can not go any further due to hitting an incomplete 615 * record. This case should occur rarely but may have to be better 616 * optimized XXX. 617 */ 618 rawp = (void *)(jo->fifo.membase + (jo->fifo.rindex & jo->fifo.mask)); 619 if (rawp->begmagic == JREC_INCOMPLETEMAGIC) { 620 tsleep(&jo->fifo, 0, "jpad", hz); 621 continue; 622 } 623 624 /* 625 * Skip any pad records. We do not write out pad records if we can 626 * help it. 627 */ 628 if (rawp->streamid == JREC_STREAMID_PAD) { 629 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) { 630 if (jo->fifo.rindex == jo->fifo.xindex) { 631 jo->fifo.xindex += (rawp->recsize + 15) & ~15; 632 jo->total_acked += (rawp->recsize + 15) & ~15; 633 } 634 } 635 jo->fifo.rindex += (rawp->recsize + 15) & ~15; 636 jo->total_acked += bytes; 637 KKASSERT(jo->fifo.windex - jo->fifo.rindex >= 0); 638 continue; 639 } 640 641 /* 642 * 'bytes' is the amount of data that can potentially be written out. 643 * Calculate 'res', the amount of data that can actually be written 644 * out. res is bounded either by hitting the end of the physical 645 * memory buffer or by hitting an incomplete record. Incomplete 646 * records often occur due to the way the space reservation model 647 * works. 648 */ 649 res = 0; 650 avail = jo->fifo.size - (jo->fifo.rindex & jo->fifo.mask); 651 while (res < bytes && rawp->begmagic == JREC_BEGMAGIC) { 652 res += (rawp->recsize + 15) & ~15; 653 if (res >= avail) { 654 KKASSERT(res == avail); 655 break; 656 } 657 rawp = (void *)((char *)rawp + ((rawp->recsize + 15) & ~15)); 658 } 659 660 /* 661 * Issue the write and deal with any errors or other conditions. 662 * For now assume blocking I/O. Since we are record-aware the 663 * code cannot yet handle partial writes. 664 * 665 * We bump rindex prior to issuing the write to avoid racing 666 * the acknowledgement coming back (which could prevent the ack 667 * from bumping xindex). Restarts are always based on xindex so 668 * we do not try to undo the rindex if an error occurs. 669 * 670 * XXX EWOULDBLOCK/NBIO 671 * XXX notification on failure 672 * XXX permanent verses temporary failures 673 * XXX two-way acknowledgement stream in the return direction / xindex 674 */ 675 bytes = res; 676 jo->fifo.rindex += bytes; 677 error = fp_write(jo->fp, 678 jo->fifo.membase + ((jo->fifo.rindex - bytes) & jo->fifo.mask), 679 bytes, &res); 680 if (error) { 681 printf("journal_thread(%s) write, error %d\n", jo->id, error); 682 /* XXX */ 683 } else { 684 KKASSERT(res == bytes); 685 } 686 687 /* 688 * Advance rindex. If the journal stream is not full duplex we also 689 * advance xindex, otherwise the rjournal thread is responsible for 690 * advancing xindex. 691 */ 692 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) { 693 jo->fifo.xindex += bytes; 694 jo->total_acked += bytes; 695 } 696 KKASSERT(jo->fifo.windex - jo->fifo.rindex >= 0); 697 if ((jo->flags & MC_JOURNAL_WANT_FULLDUPLEX) == 0) { 698 if (jo->flags & MC_JOURNAL_WWAIT) { 699 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */ 700 wakeup(&jo->fifo.windex); 701 } 702 } 703 } 704 fp_shutdown(jo->fp, SHUT_WR); 705 jo->flags &= ~MC_JOURNAL_WACTIVE; 706 wakeup(jo); 707 wakeup(&jo->fifo.windex); 708 } 709 710 /* 711 * A second per-journal worker thread is created for two-way journaling 712 * streams to deal with the return acknowledgement stream. 713 */ 714 static void 715 journal_rthread(void *info) 716 { 717 struct journal_rawrecbeg *rawp; 718 struct journal_ackrecord ack; 719 struct journal *jo = info; 720 int64_t transid; 721 int error; 722 int count; 723 int bytes; 724 725 transid = 0; 726 error = 0; 727 728 for (;;) { 729 /* 730 * We have been asked to stop 731 */ 732 if (jo->flags & MC_JOURNAL_STOP_REQ) 733 break; 734 735 /* 736 * If we have no active transaction id, get one from the return 737 * stream. 738 */ 739 if (transid == 0) { 740 error = fp_read(jo->fp, &ack, sizeof(ack), &count, 1); 741 #if 0 742 printf("fp_read ack error %d count %d\n", error, count); 743 #endif 744 if (error || count != sizeof(ack)) 745 break; 746 if (error) { 747 printf("read error %d on receive stream\n", error); 748 break; 749 } 750 if (ack.rbeg.begmagic != JREC_BEGMAGIC || 751 ack.rend.endmagic != JREC_ENDMAGIC 752 ) { 753 printf("bad begmagic or endmagic on receive stream\n"); 754 break; 755 } 756 transid = ack.rbeg.transid; 757 } 758 759 /* 760 * Calculate the number of unacknowledged bytes. If there are no 761 * unacknowledged bytes then unsent data was acknowledged, report, 762 * sleep a bit, and loop in that case. This should not happen 763 * normally. The ack record is thrown away. 764 */ 765 bytes = jo->fifo.rindex - jo->fifo.xindex; 766 767 if (bytes == 0) { 768 printf("warning: unsent data acknowledged transid %08llx\n", transid); 769 tsleep(&jo->fifo.xindex, 0, "jrseq", hz); 770 transid = 0; 771 continue; 772 } 773 774 /* 775 * Since rindex has advanced, the record pointed to by xindex 776 * must be a valid record. 777 */ 778 rawp = (void *)(jo->fifo.membase + (jo->fifo.xindex & jo->fifo.mask)); 779 KKASSERT(rawp->begmagic == JREC_BEGMAGIC); 780 KKASSERT(rawp->recsize <= bytes); 781 782 /* 783 * The target can acknowledge several records at once. 784 */ 785 if (rawp->transid < transid) { 786 #if 1 787 printf("ackskip %08llx/%08llx\n", rawp->transid, transid); 788 #endif 789 jo->fifo.xindex += (rawp->recsize + 15) & ~15; 790 jo->total_acked += (rawp->recsize + 15) & ~15; 791 if (jo->flags & MC_JOURNAL_WWAIT) { 792 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */ 793 wakeup(&jo->fifo.windex); 794 } 795 continue; 796 } 797 if (rawp->transid == transid) { 798 #if 1 799 printf("ackskip %08llx/%08llx\n", rawp->transid, transid); 800 #endif 801 jo->fifo.xindex += (rawp->recsize + 15) & ~15; 802 jo->total_acked += (rawp->recsize + 15) & ~15; 803 if (jo->flags & MC_JOURNAL_WWAIT) { 804 jo->flags &= ~MC_JOURNAL_WWAIT; /* XXX hysteresis */ 805 wakeup(&jo->fifo.windex); 806 } 807 transid = 0; 808 continue; 809 } 810 printf("warning: unsent data(2) acknowledged transid %08llx\n", transid); 811 transid = 0; 812 } 813 jo->flags &= ~MC_JOURNAL_RACTIVE; 814 wakeup(jo); 815 wakeup(&jo->fifo.windex); 816 } 817 818 /* 819 * This builds a pad record which the journaling thread will skip over. Pad 820 * records are required when we are unable to reserve sufficient stream space 821 * due to insufficient space at the end of the physical memory fifo. 822 * 823 * Even though the record is not transmitted, a normal transid must be 824 * assigned to it so link recovery operations after a failure work properly. 825 */ 826 static 827 void 828 journal_build_pad(struct journal_rawrecbeg *rawp, int recsize, int64_t transid) 829 { 830 struct journal_rawrecend *rendp; 831 832 KKASSERT((recsize & 15) == 0 && recsize >= 16); 833 834 rawp->streamid = JREC_STREAMID_PAD; 835 rawp->recsize = recsize; /* must be 16-byte aligned */ 836 rawp->transid = transid; 837 /* 838 * WARNING, rendp may overlap rawp->seqno. This is necessary to 839 * allow PAD records to fit in 16 bytes. Use cpu_ccfence() to 840 * hopefully cause the compiler to not make any assumptions. 841 */ 842 rendp = (void *)((char *)rawp + rawp->recsize - sizeof(*rendp)); 843 rendp->endmagic = JREC_ENDMAGIC; 844 rendp->check = 0; 845 rendp->recsize = rawp->recsize; 846 847 /* 848 * Set the begin magic last. This is what will allow the journal 849 * thread to write the record out. Use a store fence to prevent 850 * compiler and cpu reordering of the writes. 851 */ 852 cpu_sfence(); 853 rawp->begmagic = JREC_BEGMAGIC; 854 } 855 856 /* 857 * Wake up the worker thread if the FIFO is more then half full or if 858 * someone is waiting for space to be freed up. Otherwise let the 859 * heartbeat deal with it. Being able to avoid waking up the worker 860 * is the key to the journal's cpu performance. 861 */ 862 static __inline 863 void 864 journal_commit_wakeup(struct journal *jo) 865 { 866 int avail; 867 868 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex); 869 KKASSERT(avail >= 0); 870 if ((avail < (jo->fifo.size >> 1)) || (jo->flags & MC_JOURNAL_WWAIT)) 871 wakeup(&jo->fifo); 872 } 873 874 /* 875 * Create a new BEGIN stream record with the specified streamid and the 876 * specified amount of payload space. *rawpp will be set to point to the 877 * base of the new stream record and a pointer to the base of the payload 878 * space will be returned. *rawpp does not need to be pre-NULLd prior to 879 * making this call. The raw record header will be partially initialized. 880 * 881 * A stream can be extended, aborted, or committed by other API calls 882 * below. This may result in a sequence of potentially disconnected 883 * stream records to be output to the journaling target. The first record 884 * (the one created by this function) will be marked JREC_STREAMCTL_BEGIN, 885 * while the last record on commit or abort will be marked JREC_STREAMCTL_END 886 * (and possibly also JREC_STREAMCTL_ABORTED). The last record could wind 887 * up being the same as the first, in which case the bits are all set in 888 * the first record. 889 * 890 * The stream record is created in an incomplete state by setting the begin 891 * magic to JREC_INCOMPLETEMAGIC. This prevents the worker thread from 892 * flushing the fifo past our record until we have finished populating it. 893 * Other threads can reserve and operate on their own space without stalling 894 * but the stream output will stall until we have completed operations. The 895 * memory FIFO is intended to be large enough to absorb such situations 896 * without stalling out other threads. 897 */ 898 static 899 void * 900 journal_reserve(struct journal *jo, struct journal_rawrecbeg **rawpp, 901 int16_t streamid, int bytes) 902 { 903 struct journal_rawrecbeg *rawp; 904 int avail; 905 int availtoend; 906 int req; 907 908 /* 909 * Add header and trailer overheads to the passed payload. Note that 910 * the passed payload size need not be aligned in any way. 911 */ 912 bytes += sizeof(struct journal_rawrecbeg); 913 bytes += sizeof(struct journal_rawrecend); 914 915 for (;;) { 916 /* 917 * First, check boundary conditions. If the request would wrap around 918 * we have to skip past the ending block and return to the beginning 919 * of the FIFO's buffer. Calculate 'req' which is the actual number 920 * of bytes being reserved, including wrap-around dead space. 921 * 922 * Neither 'bytes' or 'req' are aligned. 923 * 924 * Note that availtoend is not truncated to avail and so cannot be 925 * used to determine whether the reservation is possible by itself. 926 * Also, since all fifo ops are 16-byte aligned, we can check 927 * the size before calculating the aligned size. 928 */ 929 availtoend = jo->fifo.size - (jo->fifo.windex & jo->fifo.mask); 930 KKASSERT((availtoend & 15) == 0); 931 if (bytes > availtoend) 932 req = bytes + availtoend; /* add pad to end */ 933 else 934 req = bytes; 935 936 /* 937 * Next calculate the total available space and see if it is 938 * sufficient. We cannot overwrite previously buffered data 939 * past xindex because otherwise we would not be able to restart 940 * a broken link at the target's last point of commit. 941 */ 942 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex); 943 KKASSERT(avail >= 0 && (avail & 15) == 0); 944 945 if (avail < req) { 946 /* XXX MC_JOURNAL_STOP_IMM */ 947 jo->flags |= MC_JOURNAL_WWAIT; 948 ++jo->fifostalls; 949 tsleep(&jo->fifo.windex, 0, "jwrite", 0); 950 continue; 951 } 952 953 /* 954 * Create a pad record for any dead space and create an incomplete 955 * record for the live space, then return a pointer to the 956 * contiguous buffer space that was requested. 957 * 958 * NOTE: The worker thread will not flush past an incomplete 959 * record, so the reserved space can be filled in at-will. The 960 * journaling code must also be aware the reserved sections occuring 961 * after this one will also not be written out even if completed 962 * until this one is completed. 963 * 964 * The transaction id must accomodate real and potential pad creation. 965 */ 966 rawp = (void *)(jo->fifo.membase + (jo->fifo.windex & jo->fifo.mask)); 967 if (req != bytes) { 968 journal_build_pad(rawp, availtoend, jo->transid); 969 ++jo->transid; 970 rawp = (void *)jo->fifo.membase; 971 } 972 rawp->begmagic = JREC_INCOMPLETEMAGIC; /* updated by abort/commit */ 973 rawp->recsize = bytes; /* (unaligned size) */ 974 rawp->streamid = streamid | JREC_STREAMCTL_BEGIN; 975 rawp->transid = jo->transid; 976 jo->transid += 2; 977 978 /* 979 * Issue a memory barrier to guarentee that the record data has been 980 * properly initialized before we advance the write index and return 981 * a pointer to the reserved record. Otherwise the worker thread 982 * could accidently run past us. 983 * 984 * Note that stream records are always 16-byte aligned. 985 */ 986 cpu_sfence(); 987 jo->fifo.windex += (req + 15) & ~15; 988 *rawpp = rawp; 989 return(rawp + 1); 990 } 991 /* not reached */ 992 *rawpp = NULL; 993 return(NULL); 994 } 995 996 /* 997 * Attempt to extend the stream record by <bytes> worth of payload space. 998 * 999 * If it is possible to extend the existing stream record no truncation 1000 * occurs and the record is extended as specified. A pointer to the 1001 * truncation offset within the payload space is returned. 1002 * 1003 * If it is not possible to do this the existing stream record is truncated 1004 * and committed, and a new stream record of size <bytes> is created. A 1005 * pointer to the base of the new stream record's payload space is returned. 1006 * 1007 * *rawpp is set to the new reservation in the case of a new record but 1008 * the caller cannot depend on a comparison with the old rawp to determine if 1009 * this case occurs because we could end up using the same memory FIFO 1010 * offset for the new stream record. Use *newstreamrecp instead. 1011 */ 1012 static void * 1013 journal_extend(struct journal *jo, struct journal_rawrecbeg **rawpp, 1014 int truncbytes, int bytes, int *newstreamrecp) 1015 { 1016 struct journal_rawrecbeg *rawp; 1017 int16_t streamid; 1018 int availtoend; 1019 int avail; 1020 int osize; 1021 int nsize; 1022 int wbase; 1023 void *rptr; 1024 1025 *newstreamrecp = 0; 1026 rawp = *rawpp; 1027 osize = (rawp->recsize + 15) & ~15; 1028 nsize = (rawp->recsize + bytes + 15) & ~15; 1029 wbase = (char *)rawp - jo->fifo.membase; 1030 1031 /* 1032 * If the aligned record size does not change we can trivially adjust 1033 * the record size. 1034 */ 1035 if (nsize == osize) { 1036 rawp->recsize += bytes; 1037 return((char *)(rawp + 1) + truncbytes); 1038 } 1039 1040 /* 1041 * If the fifo's write index hasn't been modified since we made the 1042 * reservation and we do not hit any boundary conditions, we can 1043 * trivially make the record smaller or larger. 1044 */ 1045 if ((jo->fifo.windex & jo->fifo.mask) == wbase + osize) { 1046 availtoend = jo->fifo.size - wbase; 1047 avail = jo->fifo.size - (jo->fifo.windex - jo->fifo.xindex) + osize; 1048 KKASSERT((availtoend & 15) == 0); 1049 KKASSERT((avail & 15) == 0); 1050 if (nsize <= avail && nsize <= availtoend) { 1051 jo->fifo.windex += nsize - osize; 1052 rawp->recsize += bytes; 1053 return((char *)(rawp + 1) + truncbytes); 1054 } 1055 } 1056 1057 /* 1058 * It was not possible to extend the buffer. Commit the current 1059 * buffer and create a new one. We manually clear the BEGIN mark that 1060 * journal_reserve() creates (because this is a continuing record, not 1061 * the start of a new stream). 1062 */ 1063 streamid = rawp->streamid & JREC_STREAMID_MASK; 1064 journal_commit(jo, rawpp, truncbytes, 0); 1065 rptr = journal_reserve(jo, rawpp, streamid, bytes); 1066 rawp = *rawpp; 1067 rawp->streamid &= ~JREC_STREAMCTL_BEGIN; 1068 *newstreamrecp = 1; 1069 return(rptr); 1070 } 1071 1072 /* 1073 * Abort a journal record. If the transaction record represents a stream 1074 * BEGIN and we can reverse the fifo's write index we can simply reverse 1075 * index the entire record, as if it were never reserved in the first place. 1076 * 1077 * Otherwise we set the JREC_STREAMCTL_ABORTED bit and commit the record 1078 * with the payload truncated to 0 bytes. 1079 */ 1080 static void 1081 journal_abort(struct journal *jo, struct journal_rawrecbeg **rawpp) 1082 { 1083 struct journal_rawrecbeg *rawp; 1084 int osize; 1085 1086 rawp = *rawpp; 1087 osize = (rawp->recsize + 15) & ~15; 1088 1089 if ((rawp->streamid & JREC_STREAMCTL_BEGIN) && 1090 (jo->fifo.windex & jo->fifo.mask) == 1091 (char *)rawp - jo->fifo.membase + osize) 1092 { 1093 jo->fifo.windex -= osize; 1094 *rawpp = NULL; 1095 } else { 1096 rawp->streamid |= JREC_STREAMCTL_ABORTED; 1097 journal_commit(jo, rawpp, 0, 1); 1098 } 1099 } 1100 1101 /* 1102 * Commit a journal record and potentially truncate it to the specified 1103 * number of payload bytes. If you do not want to truncate the record, 1104 * simply pass -1 for the bytes parameter. Do not pass rawp->recsize, that 1105 * field includes header and trailer and will not be correct. Note that 1106 * passing 0 will truncate the entire data payload of the record. 1107 * 1108 * The logical stream is terminated by this function. 1109 * 1110 * If truncation occurs, and it is not possible to physically optimize the 1111 * memory FIFO due to other threads having reserved space after ours, 1112 * the remaining reserved space will be covered by a pad record. 1113 */ 1114 static void 1115 journal_commit(struct journal *jo, struct journal_rawrecbeg **rawpp, 1116 int bytes, int closeout) 1117 { 1118 struct journal_rawrecbeg *rawp; 1119 struct journal_rawrecend *rendp; 1120 int osize; 1121 int nsize; 1122 1123 rawp = *rawpp; 1124 *rawpp = NULL; 1125 1126 KKASSERT((char *)rawp >= jo->fifo.membase && 1127 (char *)rawp + rawp->recsize <= jo->fifo.membase + jo->fifo.size); 1128 KKASSERT(((intptr_t)rawp & 15) == 0); 1129 1130 /* 1131 * Truncate the record if necessary. If the FIFO write index as still 1132 * at the end of our record we can optimally backindex it. Otherwise 1133 * we have to insert a pad record to cover the dead space. 1134 * 1135 * We calculate osize which is the 16-byte-aligned original recsize. 1136 * We calculate nsize which is the 16-byte-aligned new recsize. 1137 * 1138 * Due to alignment issues or in case the passed truncation bytes is 1139 * the same as the original payload, nsize may be equal to osize even 1140 * if the committed bytes is less then the originally reserved bytes. 1141 */ 1142 if (bytes >= 0) { 1143 KKASSERT(bytes >= 0 && bytes <= rawp->recsize - sizeof(struct journal_rawrecbeg) - sizeof(struct journal_rawrecend)); 1144 osize = (rawp->recsize + 15) & ~15; 1145 rawp->recsize = bytes + sizeof(struct journal_rawrecbeg) + 1146 sizeof(struct journal_rawrecend); 1147 nsize = (rawp->recsize + 15) & ~15; 1148 KKASSERT(nsize <= osize); 1149 if (osize == nsize) { 1150 /* do nothing */ 1151 } else if ((jo->fifo.windex & jo->fifo.mask) == (char *)rawp - jo->fifo.membase + osize) { 1152 /* we are able to backindex the fifo */ 1153 jo->fifo.windex -= osize - nsize; 1154 } else { 1155 /* we cannot backindex the fifo, emplace a pad in the dead space */ 1156 journal_build_pad((void *)((char *)rawp + nsize), osize - nsize, 1157 rawp->transid + 1); 1158 } 1159 } 1160 1161 /* 1162 * Fill in the trailer. Note that unlike pad records, the trailer will 1163 * never overlap the header. 1164 */ 1165 rendp = (void *)((char *)rawp + 1166 ((rawp->recsize + 15) & ~15) - sizeof(*rendp)); 1167 rendp->endmagic = JREC_ENDMAGIC; 1168 rendp->recsize = rawp->recsize; 1169 rendp->check = 0; /* XXX check word, disabled for now */ 1170 1171 /* 1172 * Fill in begmagic last. This will allow the worker thread to proceed. 1173 * Use a memory barrier to guarentee write ordering. Mark the stream 1174 * as terminated if closeout is set. This is the typical case. 1175 */ 1176 if (closeout) 1177 rawp->streamid |= JREC_STREAMCTL_END; 1178 cpu_sfence(); /* memory and compiler barrier */ 1179 rawp->begmagic = JREC_BEGMAGIC; 1180 1181 journal_commit_wakeup(jo); 1182 } 1183 1184 /************************************************************************ 1185 * TRANSACTION SUPPORT ROUTINES * 1186 ************************************************************************ 1187 * 1188 * JRECORD_*() - routines to create subrecord transactions and embed them 1189 * in the logical streams managed by the journal_*() routines. 1190 */ 1191 1192 static int16_t sid = JREC_STREAMID_JMIN; 1193 1194 /* 1195 * Initialize the passed jrecord structure and start a new stream transaction 1196 * by reserving an initial build space in the journal's memory FIFO. 1197 */ 1198 static void 1199 jrecord_init(struct journal *jo, struct jrecord *jrec, int16_t streamid) 1200 { 1201 bzero(jrec, sizeof(*jrec)); 1202 jrec->jo = jo; 1203 if (streamid < 0) { 1204 streamid = sid++; /* XXX need to track stream ids! */ 1205 if (sid == JREC_STREAMID_JMAX) 1206 sid = JREC_STREAMID_JMIN; 1207 } 1208 jrec->streamid = streamid; 1209 jrec->stream_residual = JREC_DEFAULTSIZE; 1210 jrec->stream_reserved = jrec->stream_residual; 1211 jrec->stream_ptr = 1212 journal_reserve(jo, &jrec->rawp, streamid, jrec->stream_reserved); 1213 } 1214 1215 /* 1216 * Push a recursive record type. All pushes should have matching pops. 1217 * The old parent is returned and the newly pushed record becomes the 1218 * new parent. Note that the old parent's pointer may already be invalid 1219 * or may become invalid if jrecord_write() had to build a new stream 1220 * record, so the caller should not mess with the returned pointer in 1221 * any way other then to save it. 1222 */ 1223 static 1224 struct journal_subrecord * 1225 jrecord_push(struct jrecord *jrec, int16_t rectype) 1226 { 1227 struct journal_subrecord *save; 1228 1229 save = jrec->parent; 1230 jrec->parent = jrecord_write(jrec, rectype|JMASK_NESTED, 0); 1231 jrec->last = NULL; 1232 KKASSERT(jrec->parent != NULL); 1233 ++jrec->pushcount; 1234 ++jrec->pushptrgood; /* cleared on flush */ 1235 return(save); 1236 } 1237 1238 /* 1239 * Pop a previously pushed sub-transaction. We must set JMASK_LAST 1240 * on the last record written within the subtransaction. If the last 1241 * record written is not accessible or if the subtransaction is empty, 1242 * we must write out a pad record with JMASK_LAST set before popping. 1243 * 1244 * When popping a subtransaction the parent record's recsize field 1245 * will be properly set. If the parent pointer is no longer valid 1246 * (which can occur if the data has already been flushed out to the 1247 * stream), the protocol spec allows us to leave it 0. 1248 * 1249 * The saved parent pointer which we restore may or may not be valid, 1250 * and if not valid may or may not be NULL, depending on the value 1251 * of pushptrgood. 1252 */ 1253 static void 1254 jrecord_pop(struct jrecord *jrec, struct journal_subrecord *save) 1255 { 1256 struct journal_subrecord *last; 1257 1258 KKASSERT(jrec->pushcount > 0); 1259 KKASSERT(jrec->residual == 0); 1260 1261 /* 1262 * Set JMASK_LAST on the last record we wrote at the current 1263 * level. If last is NULL we either no longer have access to the 1264 * record or the subtransaction was empty and we must write out a pad 1265 * record. 1266 */ 1267 if ((last = jrec->last) == NULL) { 1268 jrecord_write(jrec, JLEAF_PAD|JMASK_LAST, 0); 1269 last = jrec->last; /* reload after possible flush */ 1270 } else { 1271 last->rectype |= JMASK_LAST; 1272 } 1273 1274 /* 1275 * pushptrgood tells us how many levels of parent record pointers 1276 * are valid. The jrec only stores the current parent record pointer 1277 * (and it is only valid if pushptrgood != 0). The higher level parent 1278 * record pointers are saved by the routines calling jrecord_push() and 1279 * jrecord_pop(). These pointers may become stale and we determine 1280 * that fact by tracking the count of valid parent pointers with 1281 * pushptrgood. Pointers become invalid when their related stream 1282 * record gets pushed out. 1283 * 1284 * If no pointer is available (the data has already been pushed out), 1285 * then no fixup of e.g. the length field is possible for non-leaf 1286 * nodes. The protocol allows for this situation by placing a larger 1287 * burden on the program scanning the stream on the other end. 1288 * 1289 * [parentA] 1290 * [node X] 1291 * [parentB] 1292 * [node Y] 1293 * [node Z] 1294 * (pop B) see NOTE B 1295 * (pop A) see NOTE A 1296 * 1297 * NOTE B: This pop sets LAST in node Z if the node is still accessible, 1298 * else a PAD record is appended and LAST is set in that. 1299 * 1300 * This pop sets the record size in parentB if parentB is still 1301 * accessible, else the record size is left 0 (the scanner must 1302 * deal with that). 1303 * 1304 * This pop sets the new 'last' record to parentB, the pointer 1305 * to which may or may not still be accessible. 1306 * 1307 * NOTE A: This pop sets LAST in parentB if the node is still accessible, 1308 * else a PAD record is appended and LAST is set in that. 1309 * 1310 * This pop sets the record size in parentA if parentA is still 1311 * accessible, else the record size is left 0 (the scanner must 1312 * deal with that). 1313 * 1314 * This pop sets the new 'last' record to parentA, the pointer 1315 * to which may or may not still be accessible. 1316 * 1317 * Also note that the last record in the stream transaction, which in 1318 * the above example is parentA, does not currently have the LAST bit 1319 * set. 1320 * 1321 * The current parent becomes the last record relative to the 1322 * saved parent passed into us. It's validity is based on 1323 * whether pushptrgood is non-zero prior to decrementing. The saved 1324 * parent becomes the new parent, and its validity is based on whether 1325 * pushptrgood is non-zero after decrementing. 1326 * 1327 * The old jrec->parent may be NULL if it is no longer accessible. 1328 * If pushptrgood is non-zero, however, it is guarenteed to not 1329 * be NULL (since no flush occured). 1330 */ 1331 jrec->last = jrec->parent; 1332 --jrec->pushcount; 1333 if (jrec->pushptrgood) { 1334 KKASSERT(jrec->last != NULL && last != NULL); 1335 if (--jrec->pushptrgood == 0) { 1336 jrec->parent = NULL; /* 'save' contains garbage or NULL */ 1337 } else { 1338 KKASSERT(save != NULL); 1339 jrec->parent = save; /* 'save' must not be NULL */ 1340 } 1341 1342 /* 1343 * Set the record size in the old parent. 'last' still points to 1344 * the original last record in the subtransaction being popped, 1345 * jrec->last points to the old parent (which became the last 1346 * record relative to the new parent being popped into). 1347 */ 1348 jrec->last->recsize = (char *)last + last->recsize - (char *)jrec->last; 1349 } else { 1350 jrec->parent = NULL; 1351 KKASSERT(jrec->last == NULL); 1352 } 1353 } 1354 1355 /* 1356 * Write out a leaf record, including associated data. 1357 */ 1358 static 1359 void 1360 jrecord_leaf(struct jrecord *jrec, int16_t rectype, void *ptr, int bytes) 1361 { 1362 jrecord_write(jrec, rectype, bytes); 1363 jrecord_data(jrec, ptr, bytes); 1364 } 1365 1366 /* 1367 * Write a leaf record out and return a pointer to its base. The leaf 1368 * record may contain potentially megabytes of data which is supplied 1369 * in jrecord_data() calls. The exact amount must be specified in this 1370 * call. 1371 * 1372 * THE RETURNED SUBRECORD POINTER IS ONLY VALID IMMEDIATELY AFTER THE 1373 * CALL AND MAY BECOME INVALID AT ANY TIME. ONLY THE PUSH/POP CODE SHOULD 1374 * USE THE RETURN VALUE. 1375 */ 1376 static 1377 struct journal_subrecord * 1378 jrecord_write(struct jrecord *jrec, int16_t rectype, int bytes) 1379 { 1380 struct journal_subrecord *last; 1381 int pusheditout; 1382 1383 /* 1384 * Try to catch some obvious errors. Nesting records must specify a 1385 * size of 0, and there should be no left-overs from previous operations 1386 * (such as incomplete data writeouts). 1387 */ 1388 KKASSERT(bytes == 0 || (rectype & JMASK_NESTED) == 0); 1389 KKASSERT(jrec->residual == 0); 1390 1391 /* 1392 * Check to see if the current stream record has enough room for 1393 * the new subrecord header. If it doesn't we extend the current 1394 * stream record. 1395 * 1396 * This may have the side effect of pushing out the current stream record 1397 * and creating a new one. We must adjust our stream tracking fields 1398 * accordingly. 1399 */ 1400 if (jrec->stream_residual < sizeof(struct journal_subrecord)) { 1401 jrec->stream_ptr = journal_extend(jrec->jo, &jrec->rawp, 1402 jrec->stream_reserved - jrec->stream_residual, 1403 JREC_DEFAULTSIZE, &pusheditout); 1404 if (pusheditout) { 1405 /* 1406 * If a pushout occured, the pushed out stream record was 1407 * truncated as specified and the new record is exactly the 1408 * extension size specified. 1409 */ 1410 jrec->stream_reserved = JREC_DEFAULTSIZE; 1411 jrec->stream_residual = JREC_DEFAULTSIZE; 1412 jrec->parent = NULL; /* no longer accessible */ 1413 jrec->pushptrgood = 0; /* restored parents in pops no good */ 1414 } else { 1415 /* 1416 * If no pushout occured the stream record is NOT truncated and 1417 * IS extended. 1418 */ 1419 jrec->stream_reserved += JREC_DEFAULTSIZE; 1420 jrec->stream_residual += JREC_DEFAULTSIZE; 1421 } 1422 } 1423 last = (void *)jrec->stream_ptr; 1424 last->rectype = rectype; 1425 last->reserved = 0; 1426 1427 /* 1428 * We may not know the record size for recursive records and the 1429 * header may become unavailable due to limited FIFO space. Write 1430 * -1 to indicate this special case. 1431 */ 1432 if ((rectype & JMASK_NESTED) && bytes == 0) 1433 last->recsize = -1; 1434 else 1435 last->recsize = sizeof(struct journal_subrecord) + bytes; 1436 jrec->last = last; 1437 jrec->residual = bytes; /* remaining data to be posted */ 1438 jrec->residual_align = -bytes & 7; /* post-data alignment required */ 1439 jrec->stream_ptr += sizeof(*last); /* current write pointer */ 1440 jrec->stream_residual -= sizeof(*last); /* space remaining in stream */ 1441 return(last); 1442 } 1443 1444 /* 1445 * Write out the data associated with a leaf record. Any number of calls 1446 * to this routine may be made as long as the byte count adds up to the 1447 * amount originally specified in jrecord_write(). 1448 * 1449 * The act of writing out the leaf data may result in numerous stream records 1450 * being pushed out. Callers should be aware that even the associated 1451 * subrecord header may become inaccessible due to stream record pushouts. 1452 */ 1453 static void 1454 jrecord_data(struct jrecord *jrec, const void *buf, int bytes) 1455 { 1456 int pusheditout; 1457 int extsize; 1458 1459 KKASSERT(bytes >= 0 && bytes <= jrec->residual); 1460 1461 /* 1462 * Push out stream records as long as there is insufficient room to hold 1463 * the remaining data. 1464 */ 1465 while (jrec->stream_residual < bytes) { 1466 /* 1467 * Fill in any remaining space in the current stream record. 1468 */ 1469 bcopy(buf, jrec->stream_ptr, jrec->stream_residual); 1470 buf = (const char *)buf + jrec->stream_residual; 1471 bytes -= jrec->stream_residual; 1472 /*jrec->stream_ptr += jrec->stream_residual;*/ 1473 jrec->residual -= jrec->stream_residual; 1474 jrec->stream_residual = 0; 1475 1476 /* 1477 * Try to extend the current stream record, but no more then 1/4 1478 * the size of the FIFO. 1479 */ 1480 extsize = jrec->jo->fifo.size >> 2; 1481 if (extsize > bytes) 1482 extsize = (bytes + 15) & ~15; 1483 1484 jrec->stream_ptr = journal_extend(jrec->jo, &jrec->rawp, 1485 jrec->stream_reserved - jrec->stream_residual, 1486 extsize, &pusheditout); 1487 if (pusheditout) { 1488 jrec->stream_reserved = extsize; 1489 jrec->stream_residual = extsize; 1490 jrec->parent = NULL; /* no longer accessible */ 1491 jrec->last = NULL; /* no longer accessible */ 1492 jrec->pushptrgood = 0; /* restored parents in pops no good */ 1493 } else { 1494 jrec->stream_reserved += extsize; 1495 jrec->stream_residual += extsize; 1496 } 1497 } 1498 1499 /* 1500 * Push out any remaining bytes into the current stream record. 1501 */ 1502 if (bytes) { 1503 bcopy(buf, jrec->stream_ptr, bytes); 1504 jrec->stream_ptr += bytes; 1505 jrec->stream_residual -= bytes; 1506 jrec->residual -= bytes; 1507 } 1508 1509 /* 1510 * Handle data alignment requirements for the subrecord. Because the 1511 * stream record's data space is more strictly aligned, it must already 1512 * have sufficient space to hold any subrecord alignment slop. 1513 */ 1514 if (jrec->residual == 0 && jrec->residual_align) { 1515 KKASSERT(jrec->residual_align <= jrec->stream_residual); 1516 bzero(jrec->stream_ptr, jrec->residual_align); 1517 jrec->stream_ptr += jrec->residual_align; 1518 jrec->stream_residual -= jrec->residual_align; 1519 jrec->residual_align = 0; 1520 } 1521 } 1522 1523 /* 1524 * We are finished with the transaction. This closes the transaction created 1525 * by jrecord_init(). 1526 * 1527 * NOTE: If abortit is not set then we must be at the top level with no 1528 * residual subrecord data left to output. 1529 * 1530 * If abortit is set then we can be in any state, all pushes will be 1531 * popped and it is ok for there to be residual data. This works 1532 * because the virtual stream itself is truncated. Scanners must deal 1533 * with this situation. 1534 * 1535 * The stream record will be committed or aborted as specified and jrecord 1536 * resources will be cleaned up. 1537 */ 1538 static void 1539 jrecord_done(struct jrecord *jrec, int abortit) 1540 { 1541 KKASSERT(jrec->rawp != NULL); 1542 1543 if (abortit) { 1544 journal_abort(jrec->jo, &jrec->rawp); 1545 } else { 1546 KKASSERT(jrec->pushcount == 0 && jrec->residual == 0); 1547 journal_commit(jrec->jo, &jrec->rawp, 1548 jrec->stream_reserved - jrec->stream_residual, 1); 1549 } 1550 1551 /* 1552 * jrec should not be used beyond this point without another init, 1553 * but clean up some fields to ensure that we panic if it is. 1554 * 1555 * Note that jrec->rawp is NULLd out by journal_abort/journal_commit. 1556 */ 1557 jrec->jo = NULL; 1558 jrec->stream_ptr = NULL; 1559 } 1560 1561 /************************************************************************ 1562 * LOW LEVEL RECORD SUPPORT ROUTINES * 1563 ************************************************************************ 1564 * 1565 * These routine create low level recursive and leaf subrecords representing 1566 * common filesystem structures. 1567 */ 1568 1569 /* 1570 * Write out a filename path relative to the base of the mount point. 1571 * rectype is typically JLEAF_PATH{1,2,3,4}. 1572 */ 1573 static void 1574 jrecord_write_path(struct jrecord *jrec, int16_t rectype, struct namecache *ncp) 1575 { 1576 char buf[64]; /* local buffer if it fits, else malloced */ 1577 char *base; 1578 int pathlen; 1579 int index; 1580 struct namecache *scan; 1581 1582 /* 1583 * Pass 1 - figure out the number of bytes required. Include terminating 1584 * \0 on last element and '/' separator on other elements. 1585 */ 1586 again: 1587 pathlen = 0; 1588 for (scan = ncp; 1589 scan && (scan->nc_flag & NCF_MOUNTPT) == 0; 1590 scan = scan->nc_parent 1591 ) { 1592 pathlen += scan->nc_nlen + 1; 1593 } 1594 1595 if (pathlen <= sizeof(buf)) 1596 base = buf; 1597 else 1598 base = malloc(pathlen, M_TEMP, M_INTWAIT); 1599 1600 /* 1601 * Pass 2 - generate the path buffer 1602 */ 1603 index = pathlen; 1604 for (scan = ncp; 1605 scan && (scan->nc_flag & NCF_MOUNTPT) == 0; 1606 scan = scan->nc_parent 1607 ) { 1608 if (scan->nc_nlen >= index) { 1609 if (base != buf) 1610 free(base, M_TEMP); 1611 goto again; 1612 } 1613 if (index == pathlen) 1614 base[--index] = 0; 1615 else 1616 base[--index] = '/'; 1617 index -= scan->nc_nlen; 1618 bcopy(scan->nc_name, base + index, scan->nc_nlen); 1619 } 1620 jrecord_leaf(jrec, rectype, base + index, pathlen - index); 1621 if (base != buf) 1622 free(base, M_TEMP); 1623 } 1624 1625 /* 1626 * Write out a file attribute structure. While somewhat inefficient, using 1627 * a recursive data structure is the most portable and extensible way. 1628 */ 1629 static void 1630 jrecord_write_vattr(struct jrecord *jrec, struct vattr *vat) 1631 { 1632 void *save; 1633 1634 save = jrecord_push(jrec, JTYPE_VATTR); 1635 if (vat->va_type != VNON) 1636 jrecord_leaf(jrec, JLEAF_VTYPE, &vat->va_type, sizeof(vat->va_type)); 1637 if (vat->va_mode != (mode_t)VNOVAL) 1638 jrecord_leaf(jrec, JLEAF_MODES, &vat->va_mode, sizeof(vat->va_mode)); 1639 if (vat->va_nlink != VNOVAL) 1640 jrecord_leaf(jrec, JLEAF_NLINK, &vat->va_nlink, sizeof(vat->va_nlink)); 1641 if (vat->va_uid != VNOVAL) 1642 jrecord_leaf(jrec, JLEAF_UID, &vat->va_uid, sizeof(vat->va_uid)); 1643 if (vat->va_gid != VNOVAL) 1644 jrecord_leaf(jrec, JLEAF_GID, &vat->va_gid, sizeof(vat->va_gid)); 1645 if (vat->va_fsid != VNOVAL) 1646 jrecord_leaf(jrec, JLEAF_FSID, &vat->va_fsid, sizeof(vat->va_fsid)); 1647 if (vat->va_fileid != VNOVAL) 1648 jrecord_leaf(jrec, JLEAF_INUM, &vat->va_fileid, sizeof(vat->va_fileid)); 1649 if (vat->va_size != VNOVAL) 1650 jrecord_leaf(jrec, JLEAF_SIZE, &vat->va_size, sizeof(vat->va_size)); 1651 if (vat->va_atime.tv_sec != VNOVAL) 1652 jrecord_leaf(jrec, JLEAF_ATIME, &vat->va_atime, sizeof(vat->va_atime)); 1653 if (vat->va_mtime.tv_sec != VNOVAL) 1654 jrecord_leaf(jrec, JLEAF_MTIME, &vat->va_mtime, sizeof(vat->va_mtime)); 1655 if (vat->va_ctime.tv_sec != VNOVAL) 1656 jrecord_leaf(jrec, JLEAF_CTIME, &vat->va_ctime, sizeof(vat->va_ctime)); 1657 if (vat->va_gen != VNOVAL) 1658 jrecord_leaf(jrec, JLEAF_GEN, &vat->va_gen, sizeof(vat->va_gen)); 1659 if (vat->va_flags != VNOVAL) 1660 jrecord_leaf(jrec, JLEAF_FLAGS, &vat->va_flags, sizeof(vat->va_flags)); 1661 if (vat->va_rdev != VNOVAL) 1662 jrecord_leaf(jrec, JLEAF_UDEV, &vat->va_rdev, sizeof(vat->va_rdev)); 1663 #if 0 1664 if (vat->va_filerev != VNOVAL) 1665 jrecord_leaf(jrec, JLEAF_FILEREV, &vat->va_filerev, sizeof(vat->va_filerev)); 1666 #endif 1667 jrecord_pop(jrec, save); 1668 } 1669 1670 /* 1671 * Write out the creds used to issue a file operation. If a process is 1672 * available write out additional tracking information related to the 1673 * process. 1674 * 1675 * XXX additional tracking info 1676 * XXX tty line info 1677 */ 1678 static void 1679 jrecord_write_cred(struct jrecord *jrec, struct thread *td, struct ucred *cred) 1680 { 1681 void *save; 1682 struct proc *p; 1683 1684 save = jrecord_push(jrec, JTYPE_CRED); 1685 jrecord_leaf(jrec, JLEAF_UID, &cred->cr_uid, sizeof(cred->cr_uid)); 1686 jrecord_leaf(jrec, JLEAF_GID, &cred->cr_gid, sizeof(cred->cr_gid)); 1687 if (td && (p = td->td_proc) != NULL) { 1688 jrecord_leaf(jrec, JLEAF_PID, &p->p_pid, sizeof(p->p_pid)); 1689 jrecord_leaf(jrec, JLEAF_COMM, p->p_comm, sizeof(p->p_comm)); 1690 } 1691 jrecord_pop(jrec, save); 1692 } 1693 1694 /* 1695 * Write out information required to identify a vnode 1696 * 1697 * XXX this needs work. We should write out the inode number as well, 1698 * and in fact avoid writing out the file path for seqential writes 1699 * occuring within e.g. a certain period of time. 1700 */ 1701 static void 1702 jrecord_write_vnode_ref(struct jrecord *jrec, struct vnode *vp) 1703 { 1704 struct namecache *ncp; 1705 1706 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 1707 if ((ncp->nc_flag & (NCF_UNRESOLVED|NCF_DESTROYED)) == 0) 1708 break; 1709 } 1710 if (ncp) 1711 jrecord_write_path(jrec, JLEAF_PATH_REF, ncp); 1712 } 1713 1714 static void 1715 jrecord_write_vnode_link(struct jrecord *jrec, struct vnode *vp, 1716 struct namecache *notncp) 1717 { 1718 struct namecache *ncp; 1719 1720 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 1721 if (ncp == notncp) 1722 continue; 1723 if ((ncp->nc_flag & (NCF_UNRESOLVED|NCF_DESTROYED)) == 0) 1724 break; 1725 } 1726 if (ncp) 1727 jrecord_write_path(jrec, JLEAF_PATH_REF, ncp); 1728 } 1729 1730 #if 0 1731 /* 1732 * Write out the current contents of the file within the specified 1733 * range. This is typically called from within an UNDO section. A 1734 * locked vnode must be passed. 1735 */ 1736 static int 1737 jrecord_write_filearea(struct jrecord *jrec, struct vnode *vp, 1738 off_t begoff, off_t endoff) 1739 { 1740 } 1741 #endif 1742 1743 /* 1744 * Write out the data represented by a pagelist 1745 */ 1746 static void 1747 jrecord_write_pagelist(struct jrecord *jrec, int16_t rectype, 1748 struct vm_page **pglist, int *rtvals, int pgcount, 1749 off_t offset) 1750 { 1751 struct msf_buf *msf; 1752 int error; 1753 int b; 1754 int i; 1755 1756 i = 0; 1757 while (i < pgcount) { 1758 /* 1759 * Find the next valid section. Skip any invalid elements 1760 */ 1761 if (rtvals[i] != VM_PAGER_OK) { 1762 ++i; 1763 offset += PAGE_SIZE; 1764 continue; 1765 } 1766 1767 /* 1768 * Figure out how big the valid section is, capping I/O at what the 1769 * MSFBUF can represent. 1770 */ 1771 b = i; 1772 while (i < pgcount && i - b != XIO_INTERNAL_PAGES && 1773 rtvals[i] == VM_PAGER_OK 1774 ) { 1775 ++i; 1776 } 1777 1778 /* 1779 * And write it out. 1780 */ 1781 if (i - b) { 1782 error = msf_map_pagelist(&msf, pglist + b, i - b, 0); 1783 if (error == 0) { 1784 printf("RECORD PUTPAGES %d\n", msf_buf_bytes(msf)); 1785 jrecord_leaf(jrec, JLEAF_SEEKPOS, &offset, sizeof(offset)); 1786 jrecord_leaf(jrec, rectype, 1787 msf_buf_kva(msf), msf_buf_bytes(msf)); 1788 msf_buf_free(msf); 1789 } else { 1790 printf("jrecord_write_pagelist: mapping failure\n"); 1791 } 1792 offset += (off_t)(i - b) << PAGE_SHIFT; 1793 } 1794 } 1795 } 1796 1797 /* 1798 * Write out the data represented by a UIO. 1799 */ 1800 struct jwuio_info { 1801 struct jrecord *jrec; 1802 int16_t rectype; 1803 }; 1804 1805 static int jrecord_write_uio_callback(void *info, char *buf, int bytes); 1806 1807 static void 1808 jrecord_write_uio(struct jrecord *jrec, int16_t rectype, struct uio *uio) 1809 { 1810 struct jwuio_info info = { jrec, rectype }; 1811 int error; 1812 1813 if (uio->uio_segflg != UIO_NOCOPY) { 1814 jrecord_leaf(jrec, JLEAF_SEEKPOS, &uio->uio_offset, 1815 sizeof(uio->uio_offset)); 1816 error = msf_uio_iterate(uio, jrecord_write_uio_callback, &info); 1817 if (error) 1818 printf("XXX warning uio iterate failed %d\n", error); 1819 } 1820 } 1821 1822 static int 1823 jrecord_write_uio_callback(void *info_arg, char *buf, int bytes) 1824 { 1825 struct jwuio_info *info = info_arg; 1826 1827 jrecord_leaf(info->jrec, info->rectype, buf, bytes); 1828 return(0); 1829 } 1830 1831 /************************************************************************ 1832 * JOURNAL VNOPS * 1833 ************************************************************************ 1834 * 1835 * These are function shims replacing the normal filesystem ops. We become 1836 * responsible for calling the underlying filesystem ops. We have the choice 1837 * of executing the underlying op first and then generating the journal entry, 1838 * or starting the journal entry, executing the underlying op, and then 1839 * either completing or aborting it. 1840 * 1841 * The journal is supposed to be a high-level entity, which generally means 1842 * identifying files by name rather then by inode. Supplying both allows 1843 * the journal to be used both for inode-number-compatible 'mirrors' and 1844 * for simple filesystem replication. 1845 * 1846 * Writes are particularly difficult to deal with because a single write may 1847 * represent a hundred megabyte buffer or more, and both writes and truncations 1848 * require the 'old' data to be written out as well as the new data if the 1849 * log is reversable. Other issues: 1850 * 1851 * - How to deal with operations on unlinked files (no path available), 1852 * but which may still be filesystem visible due to hard links. 1853 * 1854 * - How to deal with modifications made via a memory map. 1855 * 1856 * - Future cache coherency support will require cache coherency API calls 1857 * both prior to and after the call to the underlying VFS. 1858 * 1859 * ALSO NOTE: We do not have to shim compatibility VOPs like MKDIR which have 1860 * new VFS equivalents (NMKDIR). 1861 */ 1862 1863 /* 1864 * Journal vop_settattr { a_vp, a_vap, a_cred, a_td } 1865 */ 1866 static 1867 int 1868 journal_setattr(struct vop_setattr_args *ap) 1869 { 1870 struct mount *mp; 1871 struct journal *jo; 1872 struct jrecord jrec; 1873 void *save; /* warning, save pointers do not always remain valid */ 1874 int error; 1875 1876 error = vop_journal_operate_ap(&ap->a_head); 1877 mp = ap->a_head.a_ops->vv_mount; 1878 if (error == 0) { 1879 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 1880 jrecord_init(jo, &jrec, -1); 1881 save = jrecord_push(&jrec, JTYPE_SETATTR); 1882 jrecord_write_cred(&jrec, ap->a_td, ap->a_cred); 1883 jrecord_write_vnode_ref(&jrec, ap->a_vp); 1884 jrecord_write_vattr(&jrec, ap->a_vap); 1885 jrecord_pop(&jrec, save); 1886 jrecord_done(&jrec, 0); 1887 } 1888 } 1889 return (error); 1890 } 1891 1892 /* 1893 * Journal vop_write { a_vp, a_uio, a_ioflag, a_cred } 1894 */ 1895 static 1896 int 1897 journal_write(struct vop_write_args *ap) 1898 { 1899 struct mount *mp; 1900 struct journal *jo; 1901 struct jrecord jrec; 1902 struct uio uio_copy; 1903 struct iovec uio_one_iovec; 1904 void *save; /* warning, save pointers do not always remain valid */ 1905 int error; 1906 1907 /* 1908 * This is really nasty. UIO's don't retain sufficient information to 1909 * be reusable once they've gone through the VOP chain. The iovecs get 1910 * cleared, so we have to copy the UIO. 1911 * 1912 * XXX fix the UIO code to not destroy iov's during a scan so we can 1913 * reuse the uio over and over again. 1914 * 1915 * XXX UNDO code needs to journal the old data prior to the write. 1916 */ 1917 uio_copy = *ap->a_uio; 1918 if (uio_copy.uio_iovcnt == 1) { 1919 uio_one_iovec = ap->a_uio->uio_iov[0]; 1920 uio_copy.uio_iov = &uio_one_iovec; 1921 } else { 1922 uio_copy.uio_iov = malloc(uio_copy.uio_iovcnt * sizeof(struct iovec), 1923 M_JOURNAL, M_WAITOK); 1924 bcopy(ap->a_uio->uio_iov, uio_copy.uio_iov, 1925 uio_copy.uio_iovcnt * sizeof(struct iovec)); 1926 } 1927 1928 error = vop_journal_operate_ap(&ap->a_head); 1929 1930 /* 1931 * XXX bad hack to figure out the offset for O_APPEND writes (note: 1932 * uio field state after the VFS operation). 1933 */ 1934 uio_copy.uio_offset = ap->a_uio->uio_offset - 1935 (uio_copy.uio_resid - ap->a_uio->uio_resid); 1936 1937 mp = ap->a_head.a_ops->vv_mount; 1938 if (error == 0) { 1939 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 1940 jrecord_init(jo, &jrec, -1); 1941 save = jrecord_push(&jrec, JTYPE_WRITE); 1942 jrecord_write_cred(&jrec, NULL, ap->a_cred); 1943 jrecord_write_vnode_ref(&jrec, ap->a_vp); 1944 jrecord_write_uio(&jrec, JLEAF_FILEDATA, &uio_copy); 1945 jrecord_pop(&jrec, save); 1946 jrecord_done(&jrec, 0); 1947 } 1948 } 1949 1950 if (uio_copy.uio_iov != &uio_one_iovec) 1951 free(uio_copy.uio_iov, M_JOURNAL); 1952 1953 1954 return (error); 1955 } 1956 1957 /* 1958 * Journal vop_fsync { a_vp, a_waitfor, a_td } 1959 */ 1960 static 1961 int 1962 journal_fsync(struct vop_fsync_args *ap) 1963 { 1964 struct mount *mp; 1965 struct journal *jo; 1966 int error; 1967 1968 error = vop_journal_operate_ap(&ap->a_head); 1969 mp = ap->a_head.a_ops->vv_mount; 1970 if (error == 0) { 1971 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 1972 /* XXX synchronize pending journal records */ 1973 } 1974 } 1975 return (error); 1976 } 1977 1978 /* 1979 * Journal vop_putpages { a_vp, a_m, a_count, a_sync, a_rtvals, a_offset } 1980 * 1981 * note: a_count is in bytes. 1982 */ 1983 static 1984 int 1985 journal_putpages(struct vop_putpages_args *ap) 1986 { 1987 struct mount *mp; 1988 struct journal *jo; 1989 struct jrecord jrec; 1990 void *save; /* warning, save pointers do not always remain valid */ 1991 int error; 1992 1993 error = vop_journal_operate_ap(&ap->a_head); 1994 mp = ap->a_head.a_ops->vv_mount; 1995 if (error == 0 && ap->a_count > 0) { 1996 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 1997 jrecord_init(jo, &jrec, -1); 1998 save = jrecord_push(&jrec, JTYPE_PUTPAGES); 1999 jrecord_write_vnode_ref(&jrec, ap->a_vp); 2000 jrecord_write_pagelist(&jrec, JLEAF_FILEDATA, 2001 ap->a_m, ap->a_rtvals, btoc(ap->a_count), ap->a_offset); 2002 jrecord_pop(&jrec, save); 2003 jrecord_done(&jrec, 0); 2004 } 2005 } 2006 return (error); 2007 } 2008 2009 /* 2010 * Journal vop_setacl { a_vp, a_type, a_aclp, a_cred, a_td } 2011 */ 2012 static 2013 int 2014 journal_setacl(struct vop_setacl_args *ap) 2015 { 2016 struct mount *mp; 2017 struct journal *jo; 2018 struct jrecord jrec; 2019 void *save; /* warning, save pointers do not always remain valid */ 2020 int error; 2021 2022 error = vop_journal_operate_ap(&ap->a_head); 2023 mp = ap->a_head.a_ops->vv_mount; 2024 if (error == 0) { 2025 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2026 jrecord_init(jo, &jrec, -1); 2027 save = jrecord_push(&jrec, JTYPE_SETACL); 2028 jrecord_write_cred(&jrec, ap->a_td, ap->a_cred); 2029 jrecord_write_vnode_ref(&jrec, ap->a_vp); 2030 /* XXX type, aclp */ 2031 jrecord_pop(&jrec, save); 2032 jrecord_done(&jrec, 0); 2033 } 2034 } 2035 return (error); 2036 } 2037 2038 /* 2039 * Journal vop_setextattr { a_vp, a_name, a_uio, a_cred, a_td } 2040 */ 2041 static 2042 int 2043 journal_setextattr(struct vop_setextattr_args *ap) 2044 { 2045 struct mount *mp; 2046 struct journal *jo; 2047 struct jrecord jrec; 2048 void *save; /* warning, save pointers do not always remain valid */ 2049 int error; 2050 2051 error = vop_journal_operate_ap(&ap->a_head); 2052 mp = ap->a_head.a_ops->vv_mount; 2053 if (error == 0) { 2054 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2055 jrecord_init(jo, &jrec, -1); 2056 save = jrecord_push(&jrec, JTYPE_SETEXTATTR); 2057 jrecord_write_cred(&jrec, ap->a_td, ap->a_cred); 2058 jrecord_write_vnode_ref(&jrec, ap->a_vp); 2059 jrecord_leaf(&jrec, JLEAF_ATTRNAME, ap->a_name, strlen(ap->a_name)); 2060 jrecord_write_uio(&jrec, JLEAF_FILEDATA, ap->a_uio); 2061 jrecord_pop(&jrec, save); 2062 jrecord_done(&jrec, 0); 2063 } 2064 } 2065 return (error); 2066 } 2067 2068 /* 2069 * Journal vop_ncreate { a_ncp, a_vpp, a_cred, a_vap } 2070 */ 2071 static 2072 int 2073 journal_ncreate(struct vop_ncreate_args *ap) 2074 { 2075 struct mount *mp; 2076 struct journal *jo; 2077 struct jrecord jrec; 2078 void *save; /* warning, save pointers do not always remain valid */ 2079 int error; 2080 2081 error = vop_journal_operate_ap(&ap->a_head); 2082 mp = ap->a_head.a_ops->vv_mount; 2083 if (error == 0) { 2084 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2085 jrecord_init(jo, &jrec, -1); 2086 save = jrecord_push(&jrec, JTYPE_CREATE); 2087 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2088 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2089 if (*ap->a_vpp) 2090 jrecord_write_vnode_ref(&jrec, *ap->a_vpp); 2091 jrecord_write_vattr(&jrec, ap->a_vap); 2092 jrecord_pop(&jrec, save); 2093 jrecord_done(&jrec, 0); 2094 } 2095 } 2096 return (error); 2097 } 2098 2099 /* 2100 * Journal vop_nmknod { a_ncp, a_vpp, a_cred, a_vap } 2101 */ 2102 static 2103 int 2104 journal_nmknod(struct vop_nmknod_args *ap) 2105 { 2106 struct mount *mp; 2107 struct journal *jo; 2108 struct jrecord jrec; 2109 void *save; /* warning, save pointers do not always remain valid */ 2110 int error; 2111 2112 error = vop_journal_operate_ap(&ap->a_head); 2113 mp = ap->a_head.a_ops->vv_mount; 2114 if (error == 0) { 2115 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2116 jrecord_init(jo, &jrec, -1); 2117 save = jrecord_push(&jrec, JTYPE_MKNOD); 2118 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2119 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2120 jrecord_write_vattr(&jrec, ap->a_vap); 2121 if (*ap->a_vpp) 2122 jrecord_write_vnode_ref(&jrec, *ap->a_vpp); 2123 jrecord_pop(&jrec, save); 2124 jrecord_done(&jrec, 0); 2125 } 2126 } 2127 return (error); 2128 } 2129 2130 /* 2131 * Journal vop_nlink { a_ncp, a_vp, a_cred } 2132 */ 2133 static 2134 int 2135 journal_nlink(struct vop_nlink_args *ap) 2136 { 2137 struct mount *mp; 2138 struct journal *jo; 2139 struct jrecord jrec; 2140 void *save; /* warning, save pointers do not always remain valid */ 2141 int error; 2142 2143 error = vop_journal_operate_ap(&ap->a_head); 2144 mp = ap->a_head.a_ops->vv_mount; 2145 if (error == 0) { 2146 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2147 jrecord_init(jo, &jrec, -1); 2148 save = jrecord_push(&jrec, JTYPE_LINK); 2149 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2150 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2151 /* XXX PATH to VP and inode number */ 2152 /* XXX this call may not record the correct path when 2153 * multiple paths are available */ 2154 jrecord_write_vnode_link(&jrec, ap->a_vp, ap->a_ncp); 2155 jrecord_pop(&jrec, save); 2156 jrecord_done(&jrec, 0); 2157 } 2158 } 2159 return (error); 2160 } 2161 2162 /* 2163 * Journal vop_symlink { a_ncp, a_vpp, a_cred, a_vap, a_target } 2164 */ 2165 static 2166 int 2167 journal_nsymlink(struct vop_nsymlink_args *ap) 2168 { 2169 struct mount *mp; 2170 struct journal *jo; 2171 struct jrecord jrec; 2172 void *save; /* warning, save pointers do not always remain valid */ 2173 int error; 2174 2175 error = vop_journal_operate_ap(&ap->a_head); 2176 mp = ap->a_head.a_ops->vv_mount; 2177 if (error == 0) { 2178 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2179 jrecord_init(jo, &jrec, -1); 2180 save = jrecord_push(&jrec, JTYPE_SYMLINK); 2181 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2182 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2183 jrecord_leaf(&jrec, JLEAF_SYMLINKDATA, 2184 ap->a_target, strlen(ap->a_target)); 2185 if (*ap->a_vpp) 2186 jrecord_write_vnode_ref(&jrec, *ap->a_vpp); 2187 jrecord_pop(&jrec, save); 2188 jrecord_done(&jrec, 0); 2189 } 2190 } 2191 return (error); 2192 } 2193 2194 /* 2195 * Journal vop_nwhiteout { a_ncp, a_cred, a_flags } 2196 */ 2197 static 2198 int 2199 journal_nwhiteout(struct vop_nwhiteout_args *ap) 2200 { 2201 struct mount *mp; 2202 struct journal *jo; 2203 struct jrecord jrec; 2204 void *save; /* warning, save pointers do not always remain valid */ 2205 int error; 2206 2207 error = vop_journal_operate_ap(&ap->a_head); 2208 mp = ap->a_head.a_ops->vv_mount; 2209 if (error == 0) { 2210 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2211 jrecord_init(jo, &jrec, -1); 2212 save = jrecord_push(&jrec, JTYPE_WHITEOUT); 2213 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2214 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2215 jrecord_pop(&jrec, save); 2216 jrecord_done(&jrec, 0); 2217 } 2218 } 2219 return (error); 2220 } 2221 2222 /* 2223 * Journal vop_nremove { a_ncp, a_cred } 2224 */ 2225 static 2226 int 2227 journal_nremove(struct vop_nremove_args *ap) 2228 { 2229 struct mount *mp; 2230 struct journal *jo; 2231 struct jrecord jrec; 2232 void *save; /* warning, save pointers do not always remain valid */ 2233 int error; 2234 2235 error = vop_journal_operate_ap(&ap->a_head); 2236 mp = ap->a_head.a_ops->vv_mount; 2237 if (error == 0) { 2238 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2239 jrecord_init(jo, &jrec, -1); 2240 save = jrecord_push(&jrec, JTYPE_REMOVE); 2241 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2242 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2243 jrecord_pop(&jrec, save); 2244 jrecord_done(&jrec, 0); 2245 } 2246 } 2247 return (error); 2248 } 2249 2250 /* 2251 * Journal vop_nmkdir { a_ncp, a_vpp, a_cred, a_vap } 2252 */ 2253 static 2254 int 2255 journal_nmkdir(struct vop_nmkdir_args *ap) 2256 { 2257 struct mount *mp; 2258 struct journal *jo; 2259 struct jrecord jrec; 2260 void *save; /* warning, save pointers do not always remain valid */ 2261 int error; 2262 2263 error = vop_journal_operate_ap(&ap->a_head); 2264 mp = ap->a_head.a_ops->vv_mount; 2265 if (error == 0) { 2266 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2267 jrecord_init(jo, &jrec, -1); 2268 if (jo->flags & MC_JOURNAL_WANT_REVERSABLE) { 2269 save = jrecord_push(&jrec, JTYPE_UNDO); 2270 /* XXX undo operations */ 2271 jrecord_pop(&jrec, save); 2272 } 2273 #if 0 2274 if (jo->flags & MC_JOURNAL_WANT_AUDIT) { 2275 jrecord_write_audit(&jrec); 2276 } 2277 #endif 2278 save = jrecord_push(&jrec, JTYPE_MKDIR); 2279 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2280 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2281 jrecord_write_vattr(&jrec, ap->a_vap); 2282 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2283 if (*ap->a_vpp) 2284 jrecord_write_vnode_ref(&jrec, *ap->a_vpp); 2285 jrecord_pop(&jrec, save); 2286 jrecord_done(&jrec, 0); 2287 } 2288 } 2289 return (error); 2290 } 2291 2292 /* 2293 * Journal vop_nrmdir { a_ncp, a_cred } 2294 */ 2295 static 2296 int 2297 journal_nrmdir(struct vop_nrmdir_args *ap) 2298 { 2299 struct mount *mp; 2300 struct journal *jo; 2301 struct jrecord jrec; 2302 void *save; /* warning, save pointers do not always remain valid */ 2303 int error; 2304 2305 error = vop_journal_operate_ap(&ap->a_head); 2306 mp = ap->a_head.a_ops->vv_mount; 2307 if (error == 0) { 2308 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2309 jrecord_init(jo, &jrec, -1); 2310 save = jrecord_push(&jrec, JTYPE_RMDIR); 2311 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2312 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_ncp); 2313 jrecord_pop(&jrec, save); 2314 jrecord_done(&jrec, 0); 2315 } 2316 } 2317 return (error); 2318 } 2319 2320 /* 2321 * Journal vop_nrename { a_fncp, a_tncp, a_cred } 2322 */ 2323 static 2324 int 2325 journal_nrename(struct vop_nrename_args *ap) 2326 { 2327 struct mount *mp; 2328 struct journal *jo; 2329 struct jrecord jrec; 2330 void *save; /* warning, save pointers do not always remain valid */ 2331 int error; 2332 2333 error = vop_journal_operate_ap(&ap->a_head); 2334 mp = ap->a_head.a_ops->vv_mount; 2335 if (error == 0) { 2336 TAILQ_FOREACH(jo, &mp->mnt_jlist, jentry) { 2337 jrecord_init(jo, &jrec, -1); 2338 save = jrecord_push(&jrec, JTYPE_RENAME); 2339 jrecord_write_cred(&jrec, NULL, ap->a_cred); 2340 jrecord_write_path(&jrec, JLEAF_PATH1, ap->a_fncp); 2341 jrecord_write_path(&jrec, JLEAF_PATH2, ap->a_tncp); 2342 jrecord_pop(&jrec, save); 2343 jrecord_done(&jrec, 0); 2344 } 2345 } 2346 return (error); 2347 } 2348 2349