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_setip4(hammer2_xop_head_t *xop, hammer2_inode_t *ip4) 405 { 406 xop->ip4 = ip4; 407 hammer2_inode_ref(ip4); 408 } 409 410 void 411 hammer2_xop_reinit(hammer2_xop_head_t *xop) 412 { 413 xop->state = 0; 414 xop->error = 0; 415 xop->collect_key = 0; 416 xop->run_mask = HAMMER2_XOPMASK_VOP; 417 } 418 419 /* 420 * A mounted PFS needs Xops threads to support frontend operations. 421 */ 422 void 423 hammer2_xop_helper_create(hammer2_pfs_t *pmp) 424 { 425 int i; 426 int j; 427 428 lockmgr(&pmp->lock, LK_EXCLUSIVE); 429 pmp->has_xop_threads = 1; 430 431 pmp->xop_groups = kmalloc(hammer2_xopgroups * 432 sizeof(hammer2_xop_group_t), 433 M_HAMMER2, M_WAITOK | M_ZERO); 434 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) { 435 for (j = 0; j < hammer2_xopgroups; ++j) { 436 if (pmp->xop_groups[j].thrs[i].td) 437 continue; 438 hammer2_thr_create(&pmp->xop_groups[j].thrs[i], 439 pmp, NULL, 440 "h2xop", i, j, 441 hammer2_primary_xops_thread); 442 } 443 } 444 lockmgr(&pmp->lock, LK_RELEASE); 445 } 446 447 void 448 hammer2_xop_helper_cleanup(hammer2_pfs_t *pmp) 449 { 450 int i; 451 int j; 452 453 if (pmp->xop_groups == NULL) { 454 KKASSERT(pmp->has_xop_threads == 0); 455 return; 456 } 457 458 for (i = 0; i < pmp->pfs_nmasters; ++i) { 459 for (j = 0; j < hammer2_xopgroups; ++j) { 460 if (pmp->xop_groups[j].thrs[i].td) 461 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]); 462 } 463 } 464 pmp->has_xop_threads = 0; 465 kfree(pmp->xop_groups, M_HAMMER2); 466 pmp->xop_groups = NULL; 467 } 468 469 /* 470 * Start a XOP request, queueing it to all nodes in the cluster to 471 * execute the cluster op. 472 * 473 * XXX optimize single-target case. 474 */ 475 void 476 hammer2_xop_start_except(hammer2_xop_head_t *xop, hammer2_xop_desc_t *desc, 477 int notidx) 478 { 479 hammer2_inode_t *ip1; 480 hammer2_pfs_t *pmp; 481 hammer2_thread_t *thr; 482 int i; 483 int ng; 484 int nchains; 485 486 ip1 = xop->ip1; 487 pmp = ip1->pmp; 488 if (pmp->has_xop_threads == 0) 489 hammer2_xop_helper_create(pmp); 490 491 /* 492 * The intent of the XOP sequencer is to ensure that ops on the same 493 * inode execute in the same order. This is necessary when issuing 494 * modifying operations to multiple targets because some targets might 495 * get behind and the frontend is allowed to complete the moment a 496 * quorum of targets succeed. 497 * 498 * Strategy operations: 499 * 500 * (1) Must be segregated from non-strategy operations to 501 * avoid a deadlock. A vfsync and a bread/bwrite can 502 * deadlock the vfsync's buffer list scan. 503 * 504 * (2) Reads are separated from writes to avoid write stalls 505 * from excessively intefering with reads. Reads are allowed 506 * to wander across multiple worker threads for potential 507 * single-file concurrency improvements. 508 * 509 * (3) Writes are serialized to a single worker thread (for any 510 * given inode) in order to try to improve block allocation 511 * sequentiality and to reduce lock contention. 512 * 513 * TODO - RENAME fails here because it is potentially modifying 514 * three different inodes, but we triple-lock the inodes 515 * involved so it shouldn't create a sequencing schism. 516 */ 517 if (xop->flags & HAMMER2_XOP_STRATEGY) { 518 hammer2_xop_strategy_t *xopst; 519 520 xopst = &((hammer2_xop_t *)xop)->xop_strategy; 521 ng = mycpu->gd_cpuid % (hammer2_xopgroups >> 1); 522 #if 0 523 hammer2_off_t off; 524 int cdr; 525 526 ng = (int)(hammer2_icrc32(&xop->ip1, sizeof(xop->ip1))); 527 if (desc == &hammer2_strategy_read_desc) { 528 off = xopst->lbase / HAMMER2_PBUFSIZE; 529 cdr = hammer2_cluster_data_read; 530 /* sysctl race, load into var */ 531 cpu_ccfence(); 532 if (cdr) 533 off /= cdr; 534 ng ^= hammer2_icrc32(&off, sizeof(off)) & 535 (hammer2_worker_rmask << 1); 536 ng |= 1; 537 } else { 538 #if 0 539 off = xopst->lbase >> 21; 540 ng ^= hammer2_icrc32(&off, sizeof(off)) & 3; 541 #endif 542 ng &= ~1; 543 } 544 ng = ng % (hammer2_xopgroups >> 1); 545 ng += (hammer2_xopgroups >> 1); 546 #endif 547 } else { 548 ng = (int)(hammer2_icrc32(&xop->ip1, sizeof(xop->ip1))); 549 ng = (unsigned int)ng % (hammer2_xopgroups >> 1); 550 } 551 xop->desc = desc; 552 553 /* 554 * The instant xop is queued another thread can pick it off. In the 555 * case of asynchronous ops, another thread might even finish and 556 * deallocate it. 557 */ 558 hammer2_spin_ex(&pmp->xop_spin); 559 nchains = ip1->cluster.nchains; 560 for (i = 0; i < nchains; ++i) { 561 /* 562 * XXX ip1->cluster.array* not stable here. This temporary 563 * hack fixes basic issues in target XOPs which need to 564 * obtain a starting chain from the inode but does not 565 * address possible races against inode updates which 566 * might NULL-out a chain. 567 */ 568 if (i != notidx && ip1->cluster.array[i].chain) { 569 thr = &pmp->xop_groups[ng].thrs[i]; 570 atomic_set_64(&xop->run_mask, 1LLU << i); 571 atomic_set_64(&xop->chk_mask, 1LLU << i); 572 xop->collect[i].thr = thr; 573 TAILQ_INSERT_TAIL(&thr->xopq, xop, collect[i].entry); 574 } 575 } 576 hammer2_spin_unex(&pmp->xop_spin); 577 /* xop can become invalid at this point */ 578 579 /* 580 * Each thread has its own xopq 581 */ 582 for (i = 0; i < nchains; ++i) { 583 if (i != notidx) { 584 thr = &pmp->xop_groups[ng].thrs[i]; 585 hammer2_thr_signal(thr, HAMMER2_THREAD_XOPQ); 586 } 587 } 588 } 589 590 void 591 hammer2_xop_start(hammer2_xop_head_t *xop, hammer2_xop_desc_t *desc) 592 { 593 hammer2_xop_start_except(xop, desc, -1); 594 } 595 596 /* 597 * Retire a XOP. Used by both the VOP frontend and by the XOP backend. 598 */ 599 void 600 hammer2_xop_retire(hammer2_xop_head_t *xop, uint64_t mask) 601 { 602 hammer2_chain_t *chain; 603 uint64_t nmask; 604 int i; 605 606 /* 607 * Remove the frontend collector or remove a backend feeder. 608 * 609 * When removing the frontend we must wakeup any backend feeders 610 * who are waiting for FIFO space. 611 * 612 * When removing the last backend feeder we must wakeup any waiting 613 * frontend. 614 */ 615 KKASSERT(xop->run_mask & mask); 616 nmask = atomic_fetchadd_64(&xop->run_mask, 617 -mask + HAMMER2_XOPMASK_FEED); 618 619 /* 620 * More than one entity left 621 */ 622 if ((nmask & HAMMER2_XOPMASK_ALLDONE) != mask) { 623 /* 624 * Frontend terminating, wakeup any backends waiting on 625 * fifo full. 626 * 627 * NOTE!!! The xop can get ripped out from under us at 628 * this point, so do not reference it again. 629 * The wakeup(xop) doesn't touch the xop and 630 * is ok. 631 */ 632 if (mask == HAMMER2_XOPMASK_VOP) { 633 if (nmask & HAMMER2_XOPMASK_FIFOW) 634 wakeup(xop); 635 } 636 637 /* 638 * Wakeup frontend if the last backend is terminating. 639 */ 640 nmask -= mask; 641 if ((nmask & HAMMER2_XOPMASK_ALLDONE) == HAMMER2_XOPMASK_VOP) { 642 if (nmask & HAMMER2_XOPMASK_WAIT) 643 wakeup(xop); 644 } 645 646 return; 647 } 648 /* else nobody else left, we can ignore FIFOW */ 649 650 /* 651 * All collectors are gone, we can cleanup and dispose of the XOP. 652 * Note that this can wind up being a frontend OR a backend. 653 * Pending chains are locked shared and not owned by any thread. 654 * 655 * Cleanup the collection cluster. 656 */ 657 for (i = 0; i < xop->cluster.nchains; ++i) { 658 xop->cluster.array[i].flags = 0; 659 chain = xop->cluster.array[i].chain; 660 if (chain) { 661 xop->cluster.array[i].chain = NULL; 662 hammer2_chain_drop_unhold(chain); 663 } 664 } 665 666 /* 667 * Cleanup the fifos. Since we are the only entity left on this 668 * xop we don't have to worry about fifo flow control, and one 669 * lfence() will do the job. 670 */ 671 cpu_lfence(); 672 mask = xop->chk_mask; 673 for (i = 0; mask && i < HAMMER2_MAXCLUSTER; ++i) { 674 hammer2_xop_fifo_t *fifo = &xop->collect[i]; 675 while (fifo->ri != fifo->wi) { 676 chain = fifo->array[fifo->ri & HAMMER2_XOPFIFO_MASK]; 677 if (chain) 678 hammer2_chain_drop_unhold(chain); 679 ++fifo->ri; 680 } 681 mask &= ~(1U << i); 682 } 683 684 /* 685 * The inode is only held at this point, simply drop it. 686 */ 687 if (xop->ip1) { 688 hammer2_inode_drop(xop->ip1); 689 xop->ip1 = NULL; 690 } 691 if (xop->ip2) { 692 hammer2_inode_drop(xop->ip2); 693 xop->ip2 = NULL; 694 } 695 if (xop->ip3) { 696 hammer2_inode_drop(xop->ip3); 697 xop->ip3 = NULL; 698 } 699 if (xop->ip4) { 700 hammer2_inode_drop(xop->ip4); 701 xop->ip4 = NULL; 702 } 703 if (xop->name1) { 704 kfree(xop->name1, M_HAMMER2); 705 xop->name1 = NULL; 706 xop->name1_len = 0; 707 } 708 if (xop->name2) { 709 kfree(xop->name2, M_HAMMER2); 710 xop->name2 = NULL; 711 xop->name2_len = 0; 712 } 713 714 objcache_put(cache_xops, xop); 715 } 716 717 /* 718 * (Backend) Returns non-zero if the frontend is still attached. 719 */ 720 int 721 hammer2_xop_active(hammer2_xop_head_t *xop) 722 { 723 if (xop->run_mask & HAMMER2_XOPMASK_VOP) 724 return 1; 725 else 726 return 0; 727 } 728 729 /* 730 * (Backend) Feed chain data through the cluster validator and back to 731 * the frontend. Chains are fed from multiple nodes concurrently 732 * and pipelined via per-node FIFOs in the XOP. 733 * 734 * The chain must be locked (either shared or exclusive). The caller may 735 * unlock and drop the chain on return. This function will add an extra 736 * ref and hold the chain's data for the pass-back. 737 * 738 * No xop lock is needed because we are only manipulating fields under 739 * our direct control. 740 * 741 * Returns 0 on success and a hammer2 error code if sync is permanently 742 * lost. The caller retains a ref on the chain but by convention 743 * the lock is typically inherited by the xop (caller loses lock). 744 * 745 * Returns non-zero on error. In this situation the caller retains a 746 * ref on the chain but loses the lock (we unlock here). 747 */ 748 int 749 hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain, 750 int clindex, int error) 751 { 752 hammer2_xop_fifo_t *fifo; 753 uint64_t mask; 754 755 /* 756 * Early termination (typicaly of xop_readir) 757 */ 758 if (hammer2_xop_active(xop) == 0) { 759 error = HAMMER2_ERROR_ABORTED; 760 goto done; 761 } 762 763 /* 764 * Multi-threaded entry into the XOP collector. We own the 765 * fifo->wi for our clindex. 766 */ 767 fifo = &xop->collect[clindex]; 768 769 if (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) 770 lwkt_yield(); 771 while (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) { 772 atomic_set_int(&fifo->flags, HAMMER2_XOP_FIFO_STALL); 773 mask = xop->run_mask; 774 if ((mask & HAMMER2_XOPMASK_VOP) == 0) { 775 error = HAMMER2_ERROR_ABORTED; 776 goto done; 777 } 778 tsleep_interlock(xop, 0); 779 if (atomic_cmpset_64(&xop->run_mask, mask, 780 mask | HAMMER2_XOPMASK_FIFOW)) { 781 if (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) { 782 tsleep(xop, PINTERLOCKED, "h2feed", hz*60); 783 } 784 } 785 /* retry */ 786 } 787 atomic_clear_int(&fifo->flags, HAMMER2_XOP_FIFO_STALL); 788 if (chain) 789 hammer2_chain_ref_hold(chain); 790 if (error == 0 && chain) 791 error = chain->error; 792 fifo->errors[fifo->wi & HAMMER2_XOPFIFO_MASK] = error; 793 fifo->array[fifo->wi & HAMMER2_XOPFIFO_MASK] = chain; 794 cpu_sfence(); 795 ++fifo->wi; 796 797 mask = atomic_fetchadd_64(&xop->run_mask, HAMMER2_XOPMASK_FEED); 798 if (mask & HAMMER2_XOPMASK_WAIT) { 799 atomic_clear_64(&xop->run_mask, HAMMER2_XOPMASK_WAIT); 800 wakeup(xop); 801 } 802 error = 0; 803 804 /* 805 * Cleanup. If an error occurred we eat the lock. If no error 806 * occurred the fifo inherits the lock and gains an additional ref. 807 * 808 * The caller's ref remains in both cases. 809 */ 810 done: 811 return error; 812 } 813 814 /* 815 * (Frontend) collect a response from a running cluster op. 816 * 817 * Responses are fed from all appropriate nodes concurrently 818 * and collected into a cohesive response >= collect_key. 819 * 820 * The collector will return the instant quorum or other requirements 821 * are met, even if some nodes get behind or become non-responsive. 822 * 823 * HAMMER2_XOP_COLLECT_NOWAIT - Used to 'poll' a completed collection, 824 * usually called synchronously from the 825 * node XOPs for the strategy code to 826 * fake the frontend collection and complete 827 * the BIO as soon as possible. 828 * 829 * HAMMER2_XOP_SYNCHRONIZER - Reqeuest synchronization with a particular 830 * cluster index, prevents looping when that 831 * index is out of sync so caller can act on 832 * the out of sync element. ESRCH and EDEADLK 833 * can be returned if this flag is specified. 834 * 835 * Returns 0 on success plus a filled out xop->cluster structure. 836 * Return ENOENT on normal termination. 837 * Otherwise return an error. 838 * 839 * WARNING! If the xop returns a cluster with a non-NULL focus, note that 840 * none of the chains in the cluster (or the focus) are either 841 * locked or I/O synchronized with the cpu. hammer2_xop_gdata() 842 * and hammer2_xop_pdata() must be used to safely access the focus 843 * chain's content. 844 * 845 * The frontend can make certain assumptions based on higher-level 846 * locking done by the frontend, but data integrity absolutely 847 * requires using the gdata/pdata API. 848 */ 849 int 850 hammer2_xop_collect(hammer2_xop_head_t *xop, int flags) 851 { 852 hammer2_xop_fifo_t *fifo; 853 hammer2_chain_t *chain; 854 hammer2_key_t lokey; 855 uint64_t mask; 856 int error; 857 int keynull; 858 int adv; /* advance the element */ 859 int i; 860 861 loop: 862 /* 863 * First loop tries to advance pieces of the cluster which 864 * are out of sync. 865 */ 866 lokey = HAMMER2_KEY_MAX; 867 keynull = HAMMER2_CHECK_NULL; 868 mask = xop->run_mask; 869 cpu_lfence(); 870 871 for (i = 0; i < xop->cluster.nchains; ++i) { 872 chain = xop->cluster.array[i].chain; 873 if (chain == NULL) { 874 adv = 1; 875 } else if (chain->bref.key < xop->collect_key) { 876 adv = 1; 877 } else { 878 keynull &= ~HAMMER2_CHECK_NULL; 879 if (lokey > chain->bref.key) 880 lokey = chain->bref.key; 881 adv = 0; 882 } 883 if (adv == 0) 884 continue; 885 886 /* 887 * Advance element if possible, advanced element may be NULL. 888 */ 889 if (chain) 890 hammer2_chain_drop_unhold(chain); 891 892 fifo = &xop->collect[i]; 893 if (fifo->ri != fifo->wi) { 894 cpu_lfence(); 895 chain = fifo->array[fifo->ri & HAMMER2_XOPFIFO_MASK]; 896 error = fifo->errors[fifo->ri & HAMMER2_XOPFIFO_MASK]; 897 ++fifo->ri; 898 xop->cluster.array[i].chain = chain; 899 xop->cluster.array[i].error = error; 900 if (chain == NULL) { 901 /* XXX */ 902 xop->cluster.array[i].flags |= 903 HAMMER2_CITEM_NULL; 904 } 905 if (fifo->wi - fifo->ri <= HAMMER2_XOPFIFO / 2) { 906 if (fifo->flags & HAMMER2_XOP_FIFO_STALL) { 907 atomic_clear_int(&fifo->flags, 908 HAMMER2_XOP_FIFO_STALL); 909 wakeup(xop); 910 lwkt_yield(); 911 } 912 } 913 --i; /* loop on same index */ 914 } else { 915 /* 916 * Retain CITEM_NULL flag. If set just repeat EOF. 917 * If not, the NULL,0 combination indicates an 918 * operation in-progress. 919 */ 920 xop->cluster.array[i].chain = NULL; 921 /* retain any CITEM_NULL setting */ 922 } 923 } 924 925 /* 926 * Determine whether the lowest collected key meets clustering 927 * requirements. Returns: 928 * 929 * 0 - key valid, cluster can be returned. 930 * 931 * ENOENT - normal end of scan, return ENOENT. 932 * 933 * ESRCH - sufficient elements collected, quorum agreement 934 * that lokey is not a valid element and should be 935 * skipped. 936 * 937 * EDEADLK - sufficient elements collected, no quorum agreement 938 * (and no agreement possible). In this situation a 939 * repair is needed, for now we loop. 940 * 941 * EINPROGRESS - insufficient elements collected to resolve, wait 942 * for event and loop. 943 */ 944 if ((flags & HAMMER2_XOP_COLLECT_WAITALL) && 945 (mask & HAMMER2_XOPMASK_ALLDONE) != HAMMER2_XOPMASK_VOP) { 946 error = HAMMER2_ERROR_EINPROGRESS; 947 } else { 948 error = hammer2_cluster_check(&xop->cluster, lokey, keynull); 949 } 950 if (error == HAMMER2_ERROR_EINPROGRESS) { 951 if (flags & HAMMER2_XOP_COLLECT_NOWAIT) 952 goto done; 953 tsleep_interlock(xop, 0); 954 if (atomic_cmpset_64(&xop->run_mask, 955 mask, mask | HAMMER2_XOPMASK_WAIT)) { 956 tsleep(xop, PINTERLOCKED, "h2coll", hz*60); 957 } 958 goto loop; 959 } 960 if (error == HAMMER2_ERROR_ESRCH) { 961 if (lokey != HAMMER2_KEY_MAX) { 962 xop->collect_key = lokey + 1; 963 goto loop; 964 } 965 error = HAMMER2_ERROR_ENOENT; 966 } 967 if (error == HAMMER2_ERROR_EDEADLK) { 968 kprintf("hammer2: no quorum possible lokey %016jx\n", 969 lokey); 970 if (lokey != HAMMER2_KEY_MAX) { 971 xop->collect_key = lokey + 1; 972 goto loop; 973 } 974 error = HAMMER2_ERROR_ENOENT; 975 } 976 if (lokey == HAMMER2_KEY_MAX) 977 xop->collect_key = lokey; 978 else 979 xop->collect_key = lokey + 1; 980 done: 981 return error; 982 } 983 984 /* 985 * N x M processing threads are available to handle XOPs, N per cluster 986 * index x M cluster nodes. 987 * 988 * Locate and return the next runnable xop, or NULL if no xops are 989 * present or none of the xops are currently runnable (for various reasons). 990 * The xop is left on the queue and serves to block other dependent xops 991 * from being run. 992 * 993 * Dependent xops will not be returned. 994 * 995 * Sets HAMMER2_XOP_FIFO_RUN on the returned xop or returns NULL. 996 * 997 * NOTE! Xops run concurrently for each cluster index. 998 */ 999 #define XOP_HASH_SIZE 16 1000 #define XOP_HASH_MASK (XOP_HASH_SIZE - 1) 1001 1002 static __inline 1003 int 1004 xop_testhash(hammer2_thread_t *thr, hammer2_inode_t *ip, uint32_t *hash) 1005 { 1006 uint32_t mask; 1007 int hv; 1008 1009 hv = (int)((uintptr_t)ip + (uintptr_t)thr) / sizeof(hammer2_inode_t); 1010 mask = 1U << (hv & 31); 1011 hv >>= 5; 1012 1013 return ((int)(hash[hv & XOP_HASH_MASK] & mask)); 1014 } 1015 1016 static __inline 1017 void 1018 xop_sethash(hammer2_thread_t *thr, hammer2_inode_t *ip, uint32_t *hash) 1019 { 1020 uint32_t mask; 1021 int hv; 1022 1023 hv = (int)((uintptr_t)ip + (uintptr_t)thr) / sizeof(hammer2_inode_t); 1024 mask = 1U << (hv & 31); 1025 hv >>= 5; 1026 1027 hash[hv & XOP_HASH_MASK] |= mask; 1028 } 1029 1030 static 1031 hammer2_xop_head_t * 1032 hammer2_xop_next(hammer2_thread_t *thr) 1033 { 1034 hammer2_pfs_t *pmp = thr->pmp; 1035 int clindex = thr->clindex; 1036 uint32_t hash[XOP_HASH_SIZE] = { 0 }; 1037 hammer2_xop_head_t *xop; 1038 1039 hammer2_spin_ex(&pmp->xop_spin); 1040 TAILQ_FOREACH(xop, &thr->xopq, collect[clindex].entry) { 1041 /* 1042 * Check dependency 1043 */ 1044 if (xop_testhash(thr, xop->ip1, hash) || 1045 (xop->ip2 && xop_testhash(thr, xop->ip2, hash)) || 1046 (xop->ip3 && xop_testhash(thr, xop->ip3, hash)) || 1047 (xop->ip4 && xop_testhash(thr, xop->ip4, hash))) 1048 { 1049 continue; 1050 } 1051 xop_sethash(thr, xop->ip1, hash); 1052 if (xop->ip2) 1053 xop_sethash(thr, xop->ip2, hash); 1054 if (xop->ip3) 1055 xop_sethash(thr, xop->ip3, hash); 1056 if (xop->ip4) 1057 xop_sethash(thr, xop->ip4, hash); 1058 1059 /* 1060 * Check already running 1061 */ 1062 if (xop->collect[clindex].flags & HAMMER2_XOP_FIFO_RUN) 1063 continue; 1064 1065 /* 1066 * Found a good one, return it. 1067 */ 1068 atomic_set_int(&xop->collect[clindex].flags, 1069 HAMMER2_XOP_FIFO_RUN); 1070 break; 1071 } 1072 hammer2_spin_unex(&pmp->xop_spin); 1073 1074 return xop; 1075 } 1076 1077 /* 1078 * Remove the completed XOP from the queue, clear HAMMER2_XOP_FIFO_RUN. 1079 * 1080 * NOTE! Xops run concurrently for each cluster index. 1081 */ 1082 static 1083 void 1084 hammer2_xop_dequeue(hammer2_thread_t *thr, hammer2_xop_head_t *xop) 1085 { 1086 hammer2_pfs_t *pmp = thr->pmp; 1087 int clindex = thr->clindex; 1088 1089 hammer2_spin_ex(&pmp->xop_spin); 1090 TAILQ_REMOVE(&thr->xopq, xop, collect[clindex].entry); 1091 atomic_clear_int(&xop->collect[clindex].flags, 1092 HAMMER2_XOP_FIFO_RUN); 1093 hammer2_spin_unex(&pmp->xop_spin); 1094 if (TAILQ_FIRST(&thr->xopq)) 1095 hammer2_thr_signal(thr, HAMMER2_THREAD_XOPQ); 1096 } 1097 1098 /* 1099 * Primary management thread for xops support. Each node has several such 1100 * threads which replicate front-end operations on cluster nodes. 1101 * 1102 * XOPS thread node operations, allowing the function to focus on a single 1103 * node in the cluster after validating the operation with the cluster. 1104 * This is primarily what prevents dead or stalled nodes from stalling 1105 * the front-end. 1106 */ 1107 void 1108 hammer2_primary_xops_thread(void *arg) 1109 { 1110 hammer2_thread_t *thr = arg; 1111 hammer2_pfs_t *pmp; 1112 hammer2_xop_head_t *xop; 1113 uint64_t mask; 1114 uint32_t flags; 1115 uint32_t nflags; 1116 hammer2_xop_desc_t *last_desc = NULL; 1117 1118 pmp = thr->pmp; 1119 /*xgrp = &pmp->xop_groups[thr->repidx]; not needed */ 1120 mask = 1LLU << thr->clindex; 1121 1122 for (;;) { 1123 flags = thr->flags; 1124 1125 /* 1126 * Handle stop request 1127 */ 1128 if (flags & HAMMER2_THREAD_STOP) 1129 break; 1130 1131 /* 1132 * Handle freeze request 1133 */ 1134 if (flags & HAMMER2_THREAD_FREEZE) { 1135 hammer2_thr_signal2(thr, HAMMER2_THREAD_FROZEN, 1136 HAMMER2_THREAD_FREEZE); 1137 continue; 1138 } 1139 1140 if (flags & HAMMER2_THREAD_UNFREEZE) { 1141 hammer2_thr_signal2(thr, 0, 1142 HAMMER2_THREAD_FROZEN | 1143 HAMMER2_THREAD_UNFREEZE); 1144 continue; 1145 } 1146 1147 /* 1148 * Force idle if frozen until unfrozen or stopped. 1149 */ 1150 if (flags & HAMMER2_THREAD_FROZEN) { 1151 hammer2_thr_wait_any(thr, 1152 HAMMER2_THREAD_UNFREEZE | 1153 HAMMER2_THREAD_STOP, 1154 0); 1155 continue; 1156 } 1157 1158 /* 1159 * Reset state on REMASTER request 1160 */ 1161 if (flags & HAMMER2_THREAD_REMASTER) { 1162 hammer2_thr_signal2(thr, 0, HAMMER2_THREAD_REMASTER); 1163 /* reset state here */ 1164 continue; 1165 } 1166 1167 /* 1168 * Process requests. Each request can be multi-queued. 1169 * 1170 * If we get behind and the frontend VOP is no longer active, 1171 * we retire the request without processing it. The callback 1172 * may also abort processing if the frontend VOP becomes 1173 * inactive. 1174 */ 1175 if (flags & HAMMER2_THREAD_XOPQ) { 1176 nflags = flags & ~HAMMER2_THREAD_XOPQ; 1177 if (!atomic_cmpset_int(&thr->flags, flags, nflags)) 1178 continue; 1179 flags = nflags; 1180 /* fall through */ 1181 } 1182 while ((xop = hammer2_xop_next(thr)) != NULL) { 1183 if (hammer2_xop_active(xop)) { 1184 last_desc = xop->desc; 1185 xop->desc->storage_func((hammer2_xop_t *)xop, 1186 thr->scratch, 1187 thr->clindex); 1188 hammer2_xop_dequeue(thr, xop); 1189 hammer2_xop_retire(xop, mask); 1190 } else { 1191 last_desc = xop->desc; 1192 hammer2_xop_feed(xop, NULL, thr->clindex, 1193 ECONNABORTED); 1194 hammer2_xop_dequeue(thr, xop); 1195 hammer2_xop_retire(xop, mask); 1196 } 1197 } 1198 1199 /* 1200 * Wait for event, interlock using THREAD_WAITING and 1201 * THREAD_SIGNAL. 1202 * 1203 * For robustness poll on a 30-second interval, but nominally 1204 * expect to be woken up. 1205 */ 1206 nflags = flags | HAMMER2_THREAD_WAITING; 1207 1208 tsleep_interlock(&thr->flags, 0); 1209 if (atomic_cmpset_int(&thr->flags, flags, nflags)) { 1210 tsleep(&thr->flags, PINTERLOCKED, "h2idle", hz*30); 1211 } 1212 } 1213 1214 #if 0 1215 /* 1216 * Cleanup / termination 1217 */ 1218 while ((xop = TAILQ_FIRST(&thr->xopq)) != NULL) { 1219 kprintf("hammer2_thread: aborting xop %s\n", xop->desc->id); 1220 TAILQ_REMOVE(&thr->xopq, xop, 1221 collect[thr->clindex].entry); 1222 hammer2_xop_retire(xop, mask); 1223 } 1224 #endif 1225 thr->td = NULL; 1226 hammer2_thr_signal(thr, HAMMER2_THREAD_STOPPED); 1227 /* thr structure can go invalid after this point */ 1228 } 1229