1 /* 2 * Copyright (c) 2015-2018 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@dragonflybsd.org> 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 /* 35 * This module implements the hammer2 helper thread API, including 36 * the frontend/backend XOP API. 37 */ 38 #include "hammer2.h" 39 40 #define H2XOPDESCRIPTOR(label) \ 41 hammer2_xop_desc_t hammer2_##label##_desc = { \ 42 .storage_func = hammer2_xop_##label, \ 43 .id = #label \ 44 } 45 46 H2XOPDESCRIPTOR(ipcluster); 47 H2XOPDESCRIPTOR(readdir); 48 H2XOPDESCRIPTOR(nresolve); 49 H2XOPDESCRIPTOR(unlink); 50 H2XOPDESCRIPTOR(nrename); 51 H2XOPDESCRIPTOR(scanlhc); 52 H2XOPDESCRIPTOR(scanall); 53 H2XOPDESCRIPTOR(lookup); 54 H2XOPDESCRIPTOR(delete); 55 H2XOPDESCRIPTOR(inode_mkdirent); 56 H2XOPDESCRIPTOR(inode_create); 57 H2XOPDESCRIPTOR(inode_create_det); 58 H2XOPDESCRIPTOR(inode_create_ins); 59 H2XOPDESCRIPTOR(inode_destroy); 60 H2XOPDESCRIPTOR(inode_chain_sync); 61 H2XOPDESCRIPTOR(inode_unlinkall); 62 H2XOPDESCRIPTOR(inode_connect); 63 H2XOPDESCRIPTOR(inode_flush); 64 H2XOPDESCRIPTOR(strategy_read); 65 H2XOPDESCRIPTOR(strategy_write); 66 67 /* 68 * Set flags and wakeup any waiters. 69 * 70 * WARNING! During teardown (thr) can disappear the instant our cmpset 71 * succeeds. 72 */ 73 void 74 hammer2_thr_signal(hammer2_thread_t *thr, uint32_t flags) 75 { 76 uint32_t oflags; 77 uint32_t nflags; 78 79 for (;;) { 80 oflags = thr->flags; 81 cpu_ccfence(); 82 nflags = (oflags | flags) & ~HAMMER2_THREAD_WAITING; 83 84 if (oflags & HAMMER2_THREAD_WAITING) { 85 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) { 86 wakeup(&thr->flags); 87 break; 88 } 89 } else { 90 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) 91 break; 92 } 93 } 94 } 95 96 /* 97 * Set and clear flags and wakeup any waiters. 98 * 99 * WARNING! During teardown (thr) can disappear the instant our cmpset 100 * succeeds. 101 */ 102 void 103 hammer2_thr_signal2(hammer2_thread_t *thr, uint32_t posflags, uint32_t negflags) 104 { 105 uint32_t oflags; 106 uint32_t nflags; 107 108 for (;;) { 109 oflags = thr->flags; 110 cpu_ccfence(); 111 nflags = (oflags | posflags) & 112 ~(negflags | HAMMER2_THREAD_WAITING); 113 if (oflags & HAMMER2_THREAD_WAITING) { 114 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) { 115 wakeup(&thr->flags); 116 break; 117 } 118 } else { 119 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) 120 break; 121 } 122 } 123 } 124 125 /* 126 * Wait until all the bits in flags are set. 127 * 128 * WARNING! During teardown (thr) can disappear the instant our cmpset 129 * succeeds. 130 */ 131 void 132 hammer2_thr_wait(hammer2_thread_t *thr, uint32_t flags) 133 { 134 uint32_t oflags; 135 uint32_t nflags; 136 137 for (;;) { 138 oflags = thr->flags; 139 cpu_ccfence(); 140 if ((oflags & flags) == flags) 141 break; 142 nflags = oflags | HAMMER2_THREAD_WAITING; 143 tsleep_interlock(&thr->flags, 0); 144 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) { 145 tsleep(&thr->flags, PINTERLOCKED, "h2twait", hz*60); 146 } 147 } 148 } 149 150 /* 151 * Wait until any of the bits in flags are set, with timeout. 152 * 153 * WARNING! During teardown (thr) can disappear the instant our cmpset 154 * succeeds. 155 */ 156 int 157 hammer2_thr_wait_any(hammer2_thread_t *thr, uint32_t flags, int timo) 158 { 159 uint32_t oflags; 160 uint32_t nflags; 161 int error; 162 163 error = 0; 164 for (;;) { 165 oflags = thr->flags; 166 cpu_ccfence(); 167 if (oflags & flags) 168 break; 169 nflags = oflags | HAMMER2_THREAD_WAITING; 170 tsleep_interlock(&thr->flags, 0); 171 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) { 172 error = tsleep(&thr->flags, PINTERLOCKED, 173 "h2twait", timo); 174 } 175 if (error == ETIMEDOUT) { 176 error = HAMMER2_ERROR_ETIMEDOUT; 177 break; 178 } 179 } 180 return error; 181 } 182 183 /* 184 * Wait until the bits in flags are clear. 185 * 186 * WARNING! During teardown (thr) can disappear the instant our cmpset 187 * succeeds. 188 */ 189 void 190 hammer2_thr_wait_neg(hammer2_thread_t *thr, uint32_t flags) 191 { 192 uint32_t oflags; 193 uint32_t nflags; 194 195 for (;;) { 196 oflags = thr->flags; 197 cpu_ccfence(); 198 if ((oflags & flags) == 0) 199 break; 200 nflags = oflags | HAMMER2_THREAD_WAITING; 201 tsleep_interlock(&thr->flags, 0); 202 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) { 203 tsleep(&thr->flags, PINTERLOCKED, "h2twait", hz*60); 204 } 205 } 206 } 207 208 /* 209 * Initialize the supplied thread structure, starting the specified 210 * thread. 211 * 212 * NOTE: thr structure can be retained across mounts and unmounts for this 213 * pmp, so make sure the flags are in a sane state. 214 */ 215 void 216 hammer2_thr_create(hammer2_thread_t *thr, hammer2_pfs_t *pmp, 217 hammer2_dev_t *hmp, 218 const char *id, int clindex, int repidx, 219 void (*func)(void *arg)) 220 { 221 thr->pmp = pmp; /* xop helpers */ 222 thr->hmp = hmp; /* bulkfree */ 223 thr->clindex = clindex; 224 thr->repidx = repidx; 225 TAILQ_INIT(&thr->xopq); 226 atomic_clear_int(&thr->flags, HAMMER2_THREAD_STOP | 227 HAMMER2_THREAD_STOPPED | 228 HAMMER2_THREAD_FREEZE | 229 HAMMER2_THREAD_FROZEN); 230 if (thr->scratch == NULL) 231 thr->scratch = kmalloc(MAXPHYS, M_HAMMER2, M_WAITOK | M_ZERO); 232 if (repidx >= 0) { 233 lwkt_create(func, thr, &thr->td, NULL, 0, repidx % ncpus, 234 "%s-%s.%02d", id, pmp->pfs_names[clindex], repidx); 235 } else if (pmp) { 236 lwkt_create(func, thr, &thr->td, NULL, 0, -1, 237 "%s-%s", id, pmp->pfs_names[clindex]); 238 } else { 239 lwkt_create(func, thr, &thr->td, NULL, 0, -1, "%s", id); 240 } 241 } 242 243 /* 244 * Terminate a thread. This function will silently return if the thread 245 * was never initialized or has already been deleted. 246 * 247 * This is accomplished by setting the STOP flag and waiting for the td 248 * structure to become NULL. 249 */ 250 void 251 hammer2_thr_delete(hammer2_thread_t *thr) 252 { 253 if (thr->td == NULL) 254 return; 255 hammer2_thr_signal(thr, HAMMER2_THREAD_STOP); 256 hammer2_thr_wait(thr, HAMMER2_THREAD_STOPPED); 257 thr->pmp = NULL; 258 if (thr->scratch) { 259 kfree(thr->scratch, M_HAMMER2); 260 thr->scratch = NULL; 261 } 262 KKASSERT(TAILQ_EMPTY(&thr->xopq)); 263 } 264 265 /* 266 * Asynchronous remaster request. Ask the synchronization thread to 267 * start over soon (as if it were frozen and unfrozen, but without waiting). 268 * The thread always recalculates mastership relationships when restarting. 269 */ 270 void 271 hammer2_thr_remaster(hammer2_thread_t *thr) 272 { 273 if (thr->td == NULL) 274 return; 275 hammer2_thr_signal(thr, HAMMER2_THREAD_REMASTER); 276 } 277 278 void 279 hammer2_thr_freeze_async(hammer2_thread_t *thr) 280 { 281 hammer2_thr_signal(thr, HAMMER2_THREAD_FREEZE); 282 } 283 284 void 285 hammer2_thr_freeze(hammer2_thread_t *thr) 286 { 287 if (thr->td == NULL) 288 return; 289 hammer2_thr_signal(thr, HAMMER2_THREAD_FREEZE); 290 hammer2_thr_wait(thr, HAMMER2_THREAD_FROZEN); 291 } 292 293 void 294 hammer2_thr_unfreeze(hammer2_thread_t *thr) 295 { 296 if (thr->td == NULL) 297 return; 298 hammer2_thr_signal(thr, HAMMER2_THREAD_UNFREEZE); 299 hammer2_thr_wait_neg(thr, HAMMER2_THREAD_FROZEN); 300 } 301 302 int 303 hammer2_thr_break(hammer2_thread_t *thr) 304 { 305 if (thr->flags & (HAMMER2_THREAD_STOP | 306 HAMMER2_THREAD_REMASTER | 307 HAMMER2_THREAD_FREEZE)) { 308 return 1; 309 } 310 return 0; 311 } 312 313 /**************************************************************************** 314 * HAMMER2 XOPS API * 315 ****************************************************************************/ 316 317 /* 318 * Allocate a XOP request. 319 * 320 * Once allocated a XOP request can be started, collected, and retired, 321 * and can be retired early if desired. 322 * 323 * NOTE: Fifo indices might not be zero but ri == wi on objcache_get(). 324 */ 325 void * 326 hammer2_xop_alloc(hammer2_inode_t *ip, int flags) 327 { 328 hammer2_xop_t *xop; 329 330 xop = objcache_get(cache_xops, M_WAITOK); 331 KKASSERT(xop->head.cluster.array[0].chain == NULL); 332 333 xop->head.ip1 = ip; 334 xop->head.desc = NULL; 335 xop->head.flags = flags; 336 xop->head.state = 0; 337 xop->head.error = 0; 338 xop->head.collect_key = 0; 339 xop->head.focus_dio = NULL; 340 341 if (flags & HAMMER2_XOP_MODIFYING) 342 xop->head.mtid = hammer2_trans_sub(ip->pmp); 343 else 344 xop->head.mtid = 0; 345 346 xop->head.cluster.nchains = ip->cluster.nchains; 347 xop->head.cluster.pmp = ip->pmp; 348 xop->head.cluster.flags = HAMMER2_CLUSTER_LOCKED; 349 350 /* 351 * run_mask - Active thread (or frontend) associated with XOP 352 */ 353 xop->head.run_mask = HAMMER2_XOPMASK_VOP; 354 355 hammer2_inode_ref(ip); 356 357 return xop; 358 } 359 360 void 361 hammer2_xop_setname(hammer2_xop_head_t *xop, const char *name, size_t name_len) 362 { 363 xop->name1 = kmalloc(name_len + 1, M_HAMMER2, M_WAITOK | M_ZERO); 364 xop->name1_len = name_len; 365 bcopy(name, xop->name1, name_len); 366 } 367 368 void 369 hammer2_xop_setname2(hammer2_xop_head_t *xop, const char *name, size_t name_len) 370 { 371 xop->name2 = kmalloc(name_len + 1, M_HAMMER2, M_WAITOK | M_ZERO); 372 xop->name2_len = name_len; 373 bcopy(name, xop->name2, name_len); 374 } 375 376 size_t 377 hammer2_xop_setname_inum(hammer2_xop_head_t *xop, hammer2_key_t inum) 378 { 379 const size_t name_len = 18; 380 381 xop->name1 = kmalloc(name_len + 1, M_HAMMER2, M_WAITOK | M_ZERO); 382 xop->name1_len = name_len; 383 ksnprintf(xop->name1, name_len + 1, "0x%016jx", (intmax_t)inum); 384 385 return name_len; 386 } 387 388 389 void 390 hammer2_xop_setip2(hammer2_xop_head_t *xop, hammer2_inode_t *ip2) 391 { 392 xop->ip2 = ip2; 393 hammer2_inode_ref(ip2); 394 } 395 396 void 397 hammer2_xop_setip3(hammer2_xop_head_t *xop, hammer2_inode_t *ip3) 398 { 399 xop->ip3 = ip3; 400 hammer2_inode_ref(ip3); 401 } 402 403 void 404 hammer2_xop_reinit(hammer2_xop_head_t *xop) 405 { 406 xop->state = 0; 407 xop->error = 0; 408 xop->collect_key = 0; 409 xop->run_mask = HAMMER2_XOPMASK_VOP; 410 } 411 412 /* 413 * A mounted PFS needs Xops threads to support frontend operations. 414 */ 415 void 416 hammer2_xop_helper_create(hammer2_pfs_t *pmp) 417 { 418 int i; 419 int j; 420 421 lockmgr(&pmp->lock, LK_EXCLUSIVE); 422 pmp->has_xop_threads = 1; 423 424 pmp->xop_groups = kmalloc(hammer2_xopgroups * 425 sizeof(hammer2_xop_group_t), 426 M_HAMMER2, M_WAITOK | M_ZERO); 427 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 428 for (j = 0; j < hammer2_xopgroups; ++j) { 429 if (pmp->xop_groups[j].thrs[i].td) 430 continue; 431 hammer2_thr_create(&pmp->xop_groups[j].thrs[i], 432 pmp, NULL, 433 "h2xop", i, j, 434 hammer2_primary_xops_thread); 435 } 436 } 437 lockmgr(&pmp->lock, LK_RELEASE); 438 } 439 440 void 441 hammer2_xop_helper_cleanup(hammer2_pfs_t *pmp) 442 { 443 int i; 444 int j; 445 446 if (pmp->xop_groups == NULL) { 447 KKASSERT(pmp->has_xop_threads == 0); 448 return; 449 } 450 451 for (i = 0; i < pmp->pfs_nmasters; ++i) { 452 for (j = 0; j < hammer2_xopgroups; ++j) { 453 if (pmp->xop_groups[j].thrs[i].td) 454 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]); 455 } 456 } 457 pmp->has_xop_threads = 0; 458 kfree(pmp->xop_groups, M_HAMMER2); 459 pmp->xop_groups = NULL; 460 } 461 462 /* 463 * Start a XOP request, queueing it to all nodes in the cluster to 464 * execute the cluster op. 465 * 466 * XXX optimize single-target case. 467 */ 468 void 469 hammer2_xop_start_except(hammer2_xop_head_t *xop, hammer2_xop_desc_t *desc, 470 int notidx) 471 { 472 hammer2_inode_t *ip1; 473 hammer2_pfs_t *pmp; 474 hammer2_thread_t *thr; 475 int i; 476 int ng; 477 int nchains; 478 479 ip1 = xop->ip1; 480 pmp = ip1->pmp; 481 if (pmp->has_xop_threads == 0) 482 hammer2_xop_helper_create(pmp); 483 484 /* 485 * The intent of the XOP sequencer is to ensure that ops on the same 486 * inode execute in the same order. This is necessary when issuing 487 * modifying operations to multiple targets because some targets might 488 * get behind and the frontend is allowed to complete the moment a 489 * quorum of targets succeed. 490 * 491 * Strategy operations: 492 * 493 * (1) Must be segregated from non-strategy operations to 494 * avoid a deadlock. A vfsync and a bread/bwrite can 495 * deadlock the vfsync's buffer list scan. 496 * 497 * (2) Reads are separated from writes to avoid write stalls 498 * from excessively intefering with reads. Reads are allowed 499 * to wander across multiple worker threads for potential 500 * single-file concurrency improvements. 501 * 502 * (3) Writes are serialized to a single worker thread (for any 503 * given inode) in order to try to improve block allocation 504 * sequentiality and to reduce lock contention. 505 * 506 * TODO - RENAME fails here because it is potentially modifying 507 * three different inodes, but we triple-lock the inodes 508 * involved so it shouldn't create a sequencing schism. 509 */ 510 if (xop->flags & HAMMER2_XOP_STRATEGY) { 511 hammer2_xop_strategy_t *xopst; 512 513 xopst = &((hammer2_xop_t *)xop)->xop_strategy; 514 ng = mycpu->gd_cpuid % (hammer2_xopgroups >> 1); 515 #if 0 516 hammer2_off_t off; 517 int cdr; 518 519 ng = (int)(hammer2_icrc32(&xop->ip1, sizeof(xop->ip1))); 520 if (desc == &hammer2_strategy_read_desc) { 521 off = xopst->lbase / HAMMER2_PBUFSIZE; 522 cdr = hammer2_cluster_data_read; 523 /* sysctl race, load into var */ 524 cpu_ccfence(); 525 if (cdr) 526 off /= cdr; 527 ng ^= hammer2_icrc32(&off, sizeof(off)) & 528 (hammer2_worker_rmask << 1); 529 ng |= 1; 530 } else { 531 #if 0 532 off = xopst->lbase >> 21; 533 ng ^= hammer2_icrc32(&off, sizeof(off)) & 3; 534 #endif 535 ng &= ~1; 536 } 537 ng = ng % (hammer2_xopgroups >> 1); 538 ng += (hammer2_xopgroups >> 1); 539 #endif 540 } else { 541 ng = (int)(hammer2_icrc32(&xop->ip1, sizeof(xop->ip1))); 542 ng = (unsigned int)ng % (hammer2_xopgroups >> 1); 543 } 544 xop->desc = desc; 545 546 /* 547 * The instant xop is queued another thread can pick it off. In the 548 * case of asynchronous ops, another thread might even finish and 549 * deallocate it. 550 */ 551 hammer2_spin_ex(&pmp->xop_spin); 552 nchains = ip1->cluster.nchains; 553 for (i = 0; i < nchains; ++i) { 554 /* 555 * XXX ip1->cluster.array* not stable here. This temporary 556 * hack fixes basic issues in target XOPs which need to 557 * obtain a starting chain from the inode but does not 558 * address possible races against inode updates which 559 * might NULL-out a chain. 560 */ 561 if (i != notidx && ip1->cluster.array[i].chain) { 562 thr = &pmp->xop_groups[ng].thrs[i]; 563 atomic_set_64(&xop->run_mask, 1LLU << i); 564 atomic_set_64(&xop->chk_mask, 1LLU << i); 565 xop->collect[i].thr = thr; 566 TAILQ_INSERT_TAIL(&thr->xopq, xop, collect[i].entry); 567 } 568 } 569 hammer2_spin_unex(&pmp->xop_spin); 570 /* xop can become invalid at this point */ 571 572 /* 573 * Each thread has its own xopq 574 */ 575 for (i = 0; i < nchains; ++i) { 576 if (i != notidx) { 577 thr = &pmp->xop_groups[ng].thrs[i]; 578 hammer2_thr_signal(thr, HAMMER2_THREAD_XOPQ); 579 } 580 } 581 } 582 583 void 584 hammer2_xop_start(hammer2_xop_head_t *xop, hammer2_xop_desc_t *desc) 585 { 586 hammer2_xop_start_except(xop, desc, -1); 587 } 588 589 /* 590 * Retire a XOP. Used by both the VOP frontend and by the XOP backend. 591 */ 592 void 593 hammer2_xop_retire(hammer2_xop_head_t *xop, uint64_t mask) 594 { 595 hammer2_chain_t *chain; 596 uint64_t nmask; 597 int i; 598 599 /* 600 * Remove the frontend collector or remove a backend feeder. 601 * 602 * When removing the frontend we must wakeup any backend feeders 603 * who are waiting for FIFO space. 604 * 605 * When removing the last backend feeder we must wakeup any waiting 606 * frontend. 607 */ 608 KKASSERT(xop->run_mask & mask); 609 nmask = atomic_fetchadd_64(&xop->run_mask, 610 -mask + HAMMER2_XOPMASK_FEED); 611 612 /* 613 * More than one entity left 614 */ 615 if ((nmask & HAMMER2_XOPMASK_ALLDONE) != mask) { 616 /* 617 * Frontend terminating, wakeup any backends waiting on 618 * fifo full. 619 * 620 * NOTE!!! The xop can get ripped out from under us at 621 * this point, so do not reference it again. 622 * The wakeup(xop) doesn't touch the xop and 623 * is ok. 624 */ 625 if (mask == HAMMER2_XOPMASK_VOP) { 626 if (nmask & HAMMER2_XOPMASK_FIFOW) 627 wakeup(xop); 628 } 629 630 /* 631 * Wakeup frontend if the last backend is terminating. 632 */ 633 nmask -= mask; 634 if ((nmask & HAMMER2_XOPMASK_ALLDONE) == HAMMER2_XOPMASK_VOP) { 635 if (nmask & HAMMER2_XOPMASK_WAIT) 636 wakeup(xop); 637 } 638 639 return; 640 } 641 /* else nobody else left, we can ignore FIFOW */ 642 643 /* 644 * All collectors are gone, we can cleanup and dispose of the XOP. 645 * Note that this can wind up being a frontend OR a backend. 646 * Pending chains are locked shared and not owned by any thread. 647 * 648 * Cleanup the collection cluster. 649 */ 650 for (i = 0; i < xop->cluster.nchains; ++i) { 651 xop->cluster.array[i].flags = 0; 652 chain = xop->cluster.array[i].chain; 653 if (chain) { 654 xop->cluster.array[i].chain = NULL; 655 hammer2_chain_drop_unhold(chain); 656 } 657 } 658 659 /* 660 * Cleanup the fifos. Since we are the only entity left on this 661 * xop we don't have to worry about fifo flow control, and one 662 * lfence() will do the job. 663 */ 664 cpu_lfence(); 665 mask = xop->chk_mask; 666 for (i = 0; mask && i < HAMMER2_MAXCLUSTER; ++i) { 667 hammer2_xop_fifo_t *fifo = &xop->collect[i]; 668 while (fifo->ri != fifo->wi) { 669 chain = fifo->array[fifo->ri & HAMMER2_XOPFIFO_MASK]; 670 if (chain) 671 hammer2_chain_drop_unhold(chain); 672 ++fifo->ri; 673 } 674 mask &= ~(1U << i); 675 } 676 677 /* 678 * The inode is only held at this point, simply drop it. 679 */ 680 if (xop->ip1) { 681 hammer2_inode_drop(xop->ip1); 682 xop->ip1 = NULL; 683 } 684 if (xop->ip2) { 685 hammer2_inode_drop(xop->ip2); 686 xop->ip2 = NULL; 687 } 688 if (xop->ip3) { 689 hammer2_inode_drop(xop->ip3); 690 xop->ip3 = NULL; 691 } 692 if (xop->name1) { 693 kfree(xop->name1, M_HAMMER2); 694 xop->name1 = NULL; 695 xop->name1_len = 0; 696 } 697 if (xop->name2) { 698 kfree(xop->name2, M_HAMMER2); 699 xop->name2 = NULL; 700 xop->name2_len = 0; 701 } 702 703 objcache_put(cache_xops, xop); 704 } 705 706 /* 707 * (Backend) Returns non-zero if the frontend is still attached. 708 */ 709 int 710 hammer2_xop_active(hammer2_xop_head_t *xop) 711 { 712 if (xop->run_mask & HAMMER2_XOPMASK_VOP) 713 return 1; 714 else 715 return 0; 716 } 717 718 /* 719 * (Backend) Feed chain data through the cluster validator and back to 720 * the frontend. Chains are fed from multiple nodes concurrently 721 * and pipelined via per-node FIFOs in the XOP. 722 * 723 * The chain must be locked (either shared or exclusive). The caller may 724 * unlock and drop the chain on return. This function will add an extra 725 * ref and hold the chain's data for the pass-back. 726 * 727 * No xop lock is needed because we are only manipulating fields under 728 * our direct control. 729 * 730 * Returns 0 on success and a hammer2 error code if sync is permanently 731 * lost. The caller retains a ref on the chain but by convention 732 * the lock is typically inherited by the xop (caller loses lock). 733 * 734 * Returns non-zero on error. In this situation the caller retains a 735 * ref on the chain but loses the lock (we unlock here). 736 */ 737 int 738 hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain, 739 int clindex, int error) 740 { 741 hammer2_xop_fifo_t *fifo; 742 uint64_t mask; 743 744 /* 745 * Early termination (typicaly of xop_readir) 746 */ 747 if (hammer2_xop_active(xop) == 0) { 748 error = HAMMER2_ERROR_ABORTED; 749 goto done; 750 } 751 752 /* 753 * Multi-threaded entry into the XOP collector. We own the 754 * fifo->wi for our clindex. 755 */ 756 fifo = &xop->collect[clindex]; 757 758 if (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) 759 lwkt_yield(); 760 while (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) { 761 atomic_set_int(&fifo->flags, HAMMER2_XOP_FIFO_STALL); 762 mask = xop->run_mask; 763 if ((mask & HAMMER2_XOPMASK_VOP) == 0) { 764 error = HAMMER2_ERROR_ABORTED; 765 goto done; 766 } 767 tsleep_interlock(xop, 0); 768 if (atomic_cmpset_64(&xop->run_mask, mask, 769 mask | HAMMER2_XOPMASK_FIFOW)) { 770 if (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) { 771 tsleep(xop, PINTERLOCKED, "h2feed", hz*60); 772 } 773 } 774 /* retry */ 775 } 776 atomic_clear_int(&fifo->flags, HAMMER2_XOP_FIFO_STALL); 777 if (chain) 778 hammer2_chain_ref_hold(chain); 779 if (error == 0 && chain) 780 error = chain->error; 781 fifo->errors[fifo->wi & HAMMER2_XOPFIFO_MASK] = error; 782 fifo->array[fifo->wi & HAMMER2_XOPFIFO_MASK] = chain; 783 cpu_sfence(); 784 ++fifo->wi; 785 786 mask = atomic_fetchadd_64(&xop->run_mask, HAMMER2_XOPMASK_FEED); 787 if (mask & HAMMER2_XOPMASK_WAIT) { 788 atomic_clear_64(&xop->run_mask, HAMMER2_XOPMASK_WAIT); 789 wakeup(xop); 790 } 791 error = 0; 792 793 /* 794 * Cleanup. If an error occurred we eat the lock. If no error 795 * occurred the fifo inherits the lock and gains an additional ref. 796 * 797 * The caller's ref remains in both cases. 798 */ 799 done: 800 return error; 801 } 802 803 /* 804 * (Frontend) collect a response from a running cluster op. 805 * 806 * Responses are fed from all appropriate nodes concurrently 807 * and collected into a cohesive response >= collect_key. 808 * 809 * The collector will return the instant quorum or other requirements 810 * are met, even if some nodes get behind or become non-responsive. 811 * 812 * HAMMER2_XOP_COLLECT_NOWAIT - Used to 'poll' a completed collection, 813 * usually called synchronously from the 814 * node XOPs for the strategy code to 815 * fake the frontend collection and complete 816 * the BIO as soon as possible. 817 * 818 * HAMMER2_XOP_SYNCHRONIZER - Reqeuest synchronization with a particular 819 * cluster index, prevents looping when that 820 * index is out of sync so caller can act on 821 * the out of sync element. ESRCH and EDEADLK 822 * can be returned if this flag is specified. 823 * 824 * Returns 0 on success plus a filled out xop->cluster structure. 825 * Return ENOENT on normal termination. 826 * Otherwise return an error. 827 * 828 * WARNING! If the xop returns a cluster with a non-NULL focus, note that 829 * none of the chains in the cluster (or the focus) are either 830 * locked or I/O synchronized with the cpu. hammer2_xop_gdata() 831 * and hammer2_xop_pdata() must be used to safely access the focus 832 * chain's content. 833 * 834 * The frontend can make certain assumptions based on higher-level 835 * locking done by the frontend, but data integrity absolutely 836 * requires using the gdata/pdata API. 837 */ 838 int 839 hammer2_xop_collect(hammer2_xop_head_t *xop, int flags) 840 { 841 hammer2_xop_fifo_t *fifo; 842 hammer2_chain_t *chain; 843 hammer2_key_t lokey; 844 uint64_t mask; 845 int error; 846 int keynull; 847 int adv; /* advance the element */ 848 int i; 849 850 loop: 851 /* 852 * First loop tries to advance pieces of the cluster which 853 * are out of sync. 854 */ 855 lokey = HAMMER2_KEY_MAX; 856 keynull = HAMMER2_CHECK_NULL; 857 mask = xop->run_mask; 858 cpu_lfence(); 859 860 for (i = 0; i < xop->cluster.nchains; ++i) { 861 chain = xop->cluster.array[i].chain; 862 if (chain == NULL) { 863 adv = 1; 864 } else if (chain->bref.key < xop->collect_key) { 865 adv = 1; 866 } else { 867 keynull &= ~HAMMER2_CHECK_NULL; 868 if (lokey > chain->bref.key) 869 lokey = chain->bref.key; 870 adv = 0; 871 } 872 if (adv == 0) 873 continue; 874 875 /* 876 * Advance element if possible, advanced element may be NULL. 877 */ 878 if (chain) 879 hammer2_chain_drop_unhold(chain); 880 881 fifo = &xop->collect[i]; 882 if (fifo->ri != fifo->wi) { 883 cpu_lfence(); 884 chain = fifo->array[fifo->ri & HAMMER2_XOPFIFO_MASK]; 885 error = fifo->errors[fifo->ri & HAMMER2_XOPFIFO_MASK]; 886 ++fifo->ri; 887 xop->cluster.array[i].chain = chain; 888 xop->cluster.array[i].error = error; 889 if (chain == NULL) { 890 /* XXX */ 891 xop->cluster.array[i].flags |= 892 HAMMER2_CITEM_NULL; 893 } 894 if (fifo->wi - fifo->ri <= HAMMER2_XOPFIFO / 2) { 895 if (fifo->flags & HAMMER2_XOP_FIFO_STALL) { 896 atomic_clear_int(&fifo->flags, 897 HAMMER2_XOP_FIFO_STALL); 898 wakeup(xop); 899 lwkt_yield(); 900 } 901 } 902 --i; /* loop on same index */ 903 } else { 904 /* 905 * Retain CITEM_NULL flag. If set just repeat EOF. 906 * If not, the NULL,0 combination indicates an 907 * operation in-progress. 908 */ 909 xop->cluster.array[i].chain = NULL; 910 /* retain any CITEM_NULL setting */ 911 } 912 } 913 914 /* 915 * Determine whether the lowest collected key meets clustering 916 * requirements. Returns: 917 * 918 * 0 - key valid, cluster can be returned. 919 * 920 * ENOENT - normal end of scan, return ENOENT. 921 * 922 * ESRCH - sufficient elements collected, quorum agreement 923 * that lokey is not a valid element and should be 924 * skipped. 925 * 926 * EDEADLK - sufficient elements collected, no quorum agreement 927 * (and no agreement possible). In this situation a 928 * repair is needed, for now we loop. 929 * 930 * EINPROGRESS - insufficient elements collected to resolve, wait 931 * for event and loop. 932 */ 933 if ((flags & HAMMER2_XOP_COLLECT_WAITALL) && 934 (mask & HAMMER2_XOPMASK_ALLDONE) != HAMMER2_XOPMASK_VOP) { 935 error = HAMMER2_ERROR_EINPROGRESS; 936 } else { 937 error = hammer2_cluster_check(&xop->cluster, lokey, keynull); 938 } 939 if (error == HAMMER2_ERROR_EINPROGRESS) { 940 if (flags & HAMMER2_XOP_COLLECT_NOWAIT) 941 goto done; 942 tsleep_interlock(xop, 0); 943 if (atomic_cmpset_64(&xop->run_mask, 944 mask, mask | HAMMER2_XOPMASK_WAIT)) { 945 tsleep(xop, PINTERLOCKED, "h2coll", hz*60); 946 } 947 goto loop; 948 } 949 if (error == HAMMER2_ERROR_ESRCH) { 950 if (lokey != HAMMER2_KEY_MAX) { 951 xop->collect_key = lokey + 1; 952 goto loop; 953 } 954 error = HAMMER2_ERROR_ENOENT; 955 } 956 if (error == HAMMER2_ERROR_EDEADLK) { 957 kprintf("hammer2: no quorum possible lokey %016jx\n", 958 lokey); 959 if (lokey != HAMMER2_KEY_MAX) { 960 xop->collect_key = lokey + 1; 961 goto loop; 962 } 963 error = HAMMER2_ERROR_ENOENT; 964 } 965 if (lokey == HAMMER2_KEY_MAX) 966 xop->collect_key = lokey; 967 else 968 xop->collect_key = lokey + 1; 969 done: 970 return error; 971 } 972 973 /* 974 * N x M processing threads are available to handle XOPs, N per cluster 975 * index x M cluster nodes. 976 * 977 * Locate and return the next runnable xop, or NULL if no xops are 978 * present or none of the xops are currently runnable (for various reasons). 979 * The xop is left on the queue and serves to block other dependent xops 980 * from being run. 981 * 982 * Dependent xops will not be returned. 983 * 984 * Sets HAMMER2_XOP_FIFO_RUN on the returned xop or returns NULL. 985 * 986 * NOTE! Xops run concurrently for each cluster index. 987 */ 988 #define XOP_HASH_SIZE 16 989 #define XOP_HASH_MASK (XOP_HASH_SIZE - 1) 990 991 static __inline 992 int 993 xop_testhash(hammer2_thread_t *thr, hammer2_inode_t *ip, uint32_t *hash) 994 { 995 uint32_t mask; 996 int hv; 997 998 hv = (int)((uintptr_t)ip + (uintptr_t)thr) / sizeof(hammer2_inode_t); 999 mask = 1U << (hv & 31); 1000 hv >>= 5; 1001 1002 return ((int)(hash[hv & XOP_HASH_MASK] & mask)); 1003 } 1004 1005 static __inline 1006 void 1007 xop_sethash(hammer2_thread_t *thr, hammer2_inode_t *ip, uint32_t *hash) 1008 { 1009 uint32_t mask; 1010 int hv; 1011 1012 hv = (int)((uintptr_t)ip + (uintptr_t)thr) / sizeof(hammer2_inode_t); 1013 mask = 1U << (hv & 31); 1014 hv >>= 5; 1015 1016 hash[hv & XOP_HASH_MASK] |= mask; 1017 } 1018 1019 static 1020 hammer2_xop_head_t * 1021 hammer2_xop_next(hammer2_thread_t *thr) 1022 { 1023 hammer2_pfs_t *pmp = thr->pmp; 1024 int clindex = thr->clindex; 1025 uint32_t hash[XOP_HASH_SIZE] = { 0 }; 1026 hammer2_xop_head_t *xop; 1027 1028 hammer2_spin_ex(&pmp->xop_spin); 1029 TAILQ_FOREACH(xop, &thr->xopq, collect[clindex].entry) { 1030 /* 1031 * Check dependency 1032 */ 1033 if (xop_testhash(thr, xop->ip1, hash) || 1034 (xop->ip2 && xop_testhash(thr, xop->ip2, hash)) || 1035 (xop->ip3 && xop_testhash(thr, xop->ip3, hash))) { 1036 continue; 1037 } 1038 xop_sethash(thr, xop->ip1, hash); 1039 if (xop->ip2) 1040 xop_sethash(thr, xop->ip2, hash); 1041 if (xop->ip3) 1042 xop_sethash(thr, xop->ip3, hash); 1043 1044 /* 1045 * Check already running 1046 */ 1047 if (xop->collect[clindex].flags & HAMMER2_XOP_FIFO_RUN) 1048 continue; 1049 1050 /* 1051 * Found a good one, return it. 1052 */ 1053 atomic_set_int(&xop->collect[clindex].flags, 1054 HAMMER2_XOP_FIFO_RUN); 1055 break; 1056 } 1057 hammer2_spin_unex(&pmp->xop_spin); 1058 1059 return xop; 1060 } 1061 1062 /* 1063 * Remove the completed XOP from the queue, clear HAMMER2_XOP_FIFO_RUN. 1064 * 1065 * NOTE! Xops run concurrently for each cluster index. 1066 */ 1067 static 1068 void 1069 hammer2_xop_dequeue(hammer2_thread_t *thr, hammer2_xop_head_t *xop) 1070 { 1071 hammer2_pfs_t *pmp = thr->pmp; 1072 int clindex = thr->clindex; 1073 1074 hammer2_spin_ex(&pmp->xop_spin); 1075 TAILQ_REMOVE(&thr->xopq, xop, collect[clindex].entry); 1076 atomic_clear_int(&xop->collect[clindex].flags, 1077 HAMMER2_XOP_FIFO_RUN); 1078 hammer2_spin_unex(&pmp->xop_spin); 1079 if (TAILQ_FIRST(&thr->xopq)) 1080 hammer2_thr_signal(thr, HAMMER2_THREAD_XOPQ); 1081 } 1082 1083 /* 1084 * Primary management thread for xops support. Each node has several such 1085 * threads which replicate front-end operations on cluster nodes. 1086 * 1087 * XOPS thread node operations, allowing the function to focus on a single 1088 * node in the cluster after validating the operation with the cluster. 1089 * This is primarily what prevents dead or stalled nodes from stalling 1090 * the front-end. 1091 */ 1092 void 1093 hammer2_primary_xops_thread(void *arg) 1094 { 1095 hammer2_thread_t *thr = arg; 1096 hammer2_pfs_t *pmp; 1097 hammer2_xop_head_t *xop; 1098 uint64_t mask; 1099 uint32_t flags; 1100 uint32_t nflags; 1101 hammer2_xop_desc_t *last_desc = NULL; 1102 1103 pmp = thr->pmp; 1104 /*xgrp = &pmp->xop_groups[thr->repidx]; not needed */ 1105 mask = 1LLU << thr->clindex; 1106 1107 for (;;) { 1108 flags = thr->flags; 1109 1110 /* 1111 * Handle stop request 1112 */ 1113 if (flags & HAMMER2_THREAD_STOP) 1114 break; 1115 1116 /* 1117 * Handle freeze request 1118 */ 1119 if (flags & HAMMER2_THREAD_FREEZE) { 1120 hammer2_thr_signal2(thr, HAMMER2_THREAD_FROZEN, 1121 HAMMER2_THREAD_FREEZE); 1122 continue; 1123 } 1124 1125 if (flags & HAMMER2_THREAD_UNFREEZE) { 1126 hammer2_thr_signal2(thr, 0, 1127 HAMMER2_THREAD_FROZEN | 1128 HAMMER2_THREAD_UNFREEZE); 1129 continue; 1130 } 1131 1132 /* 1133 * Force idle if frozen until unfrozen or stopped. 1134 */ 1135 if (flags & HAMMER2_THREAD_FROZEN) { 1136 hammer2_thr_wait_any(thr, 1137 HAMMER2_THREAD_UNFREEZE | 1138 HAMMER2_THREAD_STOP, 1139 0); 1140 continue; 1141 } 1142 1143 /* 1144 * Reset state on REMASTER request 1145 */ 1146 if (flags & HAMMER2_THREAD_REMASTER) { 1147 hammer2_thr_signal2(thr, 0, HAMMER2_THREAD_REMASTER); 1148 /* reset state here */ 1149 continue; 1150 } 1151 1152 /* 1153 * Process requests. Each request can be multi-queued. 1154 * 1155 * If we get behind and the frontend VOP is no longer active, 1156 * we retire the request without processing it. The callback 1157 * may also abort processing if the frontend VOP becomes 1158 * inactive. 1159 */ 1160 if (flags & HAMMER2_THREAD_XOPQ) { 1161 nflags = flags & ~HAMMER2_THREAD_XOPQ; 1162 if (!atomic_cmpset_int(&thr->flags, flags, nflags)) 1163 continue; 1164 flags = nflags; 1165 /* fall through */ 1166 } 1167 while ((xop = hammer2_xop_next(thr)) != NULL) { 1168 if (hammer2_xop_active(xop)) { 1169 last_desc = xop->desc; 1170 xop->desc->storage_func((hammer2_xop_t *)xop, 1171 thr->scratch, 1172 thr->clindex); 1173 hammer2_xop_dequeue(thr, xop); 1174 hammer2_xop_retire(xop, mask); 1175 } else { 1176 last_desc = xop->desc; 1177 hammer2_xop_feed(xop, NULL, thr->clindex, 1178 ECONNABORTED); 1179 hammer2_xop_dequeue(thr, xop); 1180 hammer2_xop_retire(xop, mask); 1181 } 1182 } 1183 1184 /* 1185 * Wait for event, interlock using THREAD_WAITING and 1186 * THREAD_SIGNAL. 1187 * 1188 * For robustness poll on a 30-second interval, but nominally 1189 * expect to be woken up. 1190 */ 1191 nflags = flags | HAMMER2_THREAD_WAITING; 1192 1193 tsleep_interlock(&thr->flags, 0); 1194 if (atomic_cmpset_int(&thr->flags, flags, nflags)) { 1195 tsleep(&thr->flags, PINTERLOCKED, "h2idle", hz*30); 1196 } 1197 } 1198 1199 #if 0 1200 /* 1201 * Cleanup / termination 1202 */ 1203 while ((xop = TAILQ_FIRST(&thr->xopq)) != NULL) { 1204 kprintf("hammer2_thread: aborting xop %s\n", xop->desc->id); 1205 TAILQ_REMOVE(&thr->xopq, xop, 1206 collect[thr->clindex].entry); 1207 hammer2_xop_retire(xop, mask); 1208 } 1209 #endif 1210 thr->td = NULL; 1211 hammer2_thr_signal(thr, HAMMER2_THREAD_STOPPED); 1212 /* thr structure can go invalid after this point */ 1213 } 1214