2 * Copyright (c) 2015 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
35 * This module implements the hammer2 helper thread API, including
36 * the frontend/backend XOP API.
41 * Signal that the thread has work.
44 hammer2_thr_signal(hammer2_thread_t *thr, uint32_t flags)
51 if (oflags & HAMMER2_THREAD_WAITING) {
52 if (atomic_cmpset_int(&thr->flags, oflags,
53 (oflags | flags) & ~HAMMER2_THREAD_WAITING)) {
58 if (atomic_cmpset_int(&thr->flags, oflags,
67 * Return status to waiting client(s)
70 hammer2_thr_return(hammer2_thread_t *thr, uint32_t flags)
78 nflags = (oflags | flags) & ~HAMMER2_THREAD_CLIENTWAIT;
80 if (oflags & HAMMER2_THREAD_CLIENTWAIT) {
81 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) {
86 if (atomic_cmpset_int(&thr->flags, oflags, nflags))
93 * Wait until the bits in flags are set.
96 hammer2_thr_wait(hammer2_thread_t *thr, uint32_t flags)
104 if ((oflags & flags) == flags)
106 nflags = oflags | HAMMER2_THREAD_CLIENTWAIT;
107 tsleep_interlock(thr, 0);
108 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) {
109 tsleep(thr, PINTERLOCKED, "h2twait", hz*60);
115 * Wait until the bits in flags are clear.
118 hammer2_thr_wait_neg(hammer2_thread_t *thr, uint32_t flags)
126 if ((oflags & flags) == 0)
128 nflags = oflags | HAMMER2_THREAD_CLIENTWAIT;
129 tsleep_interlock(thr, 0);
130 if (atomic_cmpset_int(&thr->flags, oflags, nflags)) {
131 tsleep(thr, PINTERLOCKED, "h2twait", hz*60);
137 * Initialize the supplied thread structure, starting the specified
141 hammer2_thr_create(hammer2_thread_t *thr, hammer2_pfs_t *pmp,
142 const char *id, int clindex, int repidx,
143 void (*func)(void *arg))
146 thr->clindex = clindex;
147 thr->repidx = repidx;
148 TAILQ_INIT(&thr->xopq);
150 lwkt_create(func, thr, &thr->td, NULL, 0, repidx % ncpus,
151 "%s-%s.%02d", id, pmp->pfs_names[clindex], repidx);
153 lwkt_create(func, thr, &thr->td, NULL, 0, -1,
154 "%s-%s", id, pmp->pfs_names[clindex]);
159 * Terminate a thread. This function will silently return if the thread
160 * was never initialized or has already been deleted.
162 * This is accomplished by setting the STOP flag and waiting for the td
163 * structure to become NULL.
166 hammer2_thr_delete(hammer2_thread_t *thr)
170 hammer2_thr_signal(thr, HAMMER2_THREAD_STOP);
171 hammer2_thr_wait(thr, HAMMER2_THREAD_STOPPED);
173 KKASSERT(TAILQ_EMPTY(&thr->xopq));
177 * Asynchronous remaster request. Ask the synchronization thread to
178 * start over soon (as if it were frozen and unfrozen, but without waiting).
179 * The thread always recalculates mastership relationships when restarting.
182 hammer2_thr_remaster(hammer2_thread_t *thr)
186 hammer2_thr_signal(thr, HAMMER2_THREAD_REMASTER);
190 hammer2_thr_freeze_async(hammer2_thread_t *thr)
192 hammer2_thr_signal(thr, HAMMER2_THREAD_FREEZE);
196 hammer2_thr_freeze(hammer2_thread_t *thr)
200 hammer2_thr_signal(thr, HAMMER2_THREAD_FREEZE);
201 hammer2_thr_wait(thr, HAMMER2_THREAD_FROZEN);
205 hammer2_thr_unfreeze(hammer2_thread_t *thr)
209 hammer2_thr_signal(thr, HAMMER2_THREAD_UNFREEZE);
210 hammer2_thr_wait_neg(thr, HAMMER2_THREAD_FROZEN);
214 hammer2_thr_break(hammer2_thread_t *thr)
216 if (thr->flags & (HAMMER2_THREAD_STOP |
217 HAMMER2_THREAD_REMASTER |
218 HAMMER2_THREAD_FREEZE)) {
224 /****************************************************************************
226 ****************************************************************************/
229 hammer2_xop_group_init(hammer2_pfs_t *pmp, hammer2_xop_group_t *xgrp)
231 /* no extra fields in structure at the moment */
235 * Allocate a XOP request.
237 * Once allocated a XOP request can be started, collected, and retired,
238 * and can be retired early if desired.
240 * NOTE: Fifo indices might not be zero but ri == wi on objcache_get().
243 hammer2_xop_alloc(hammer2_inode_t *ip, int flags)
247 xop = objcache_get(cache_xops, M_WAITOK);
248 KKASSERT(xop->head.cluster.array[0].chain == NULL);
251 xop->head.func = NULL;
252 xop->head.flags = flags;
255 xop->head.collect_key = 0;
256 xop->head.check_counter = 0;
257 if (flags & HAMMER2_XOP_MODIFYING)
258 xop->head.mtid = hammer2_trans_sub(ip->pmp);
262 xop->head.cluster.nchains = ip->cluster.nchains;
263 xop->head.cluster.pmp = ip->pmp;
264 xop->head.cluster.flags = HAMMER2_CLUSTER_LOCKED;
267 * run_mask - Active thread (or frontend) associated with XOP
269 xop->head.run_mask = HAMMER2_XOPMASK_VOP;
271 hammer2_inode_ref(ip);
277 hammer2_xop_setname(hammer2_xop_head_t *xop, const char *name, size_t name_len)
279 xop->name1 = kmalloc(name_len + 1, M_HAMMER2, M_WAITOK | M_ZERO);
280 xop->name1_len = name_len;
281 bcopy(name, xop->name1, name_len);
285 hammer2_xop_setname2(hammer2_xop_head_t *xop, const char *name, size_t name_len)
287 xop->name2 = kmalloc(name_len + 1, M_HAMMER2, M_WAITOK | M_ZERO);
288 xop->name2_len = name_len;
289 bcopy(name, xop->name2, name_len);
293 hammer2_xop_setname_inum(hammer2_xop_head_t *xop, hammer2_key_t inum)
295 const size_t name_len = 18;
297 xop->name1 = kmalloc(name_len + 1, M_HAMMER2, M_WAITOK | M_ZERO);
298 xop->name1_len = name_len;
299 ksnprintf(xop->name1, name_len + 1, "0x%016jx", (intmax_t)inum);
306 hammer2_xop_setip2(hammer2_xop_head_t *xop, hammer2_inode_t *ip2)
309 hammer2_inode_ref(ip2);
313 hammer2_xop_setip3(hammer2_xop_head_t *xop, hammer2_inode_t *ip3)
316 hammer2_inode_ref(ip3);
320 hammer2_xop_reinit(hammer2_xop_head_t *xop)
324 xop->collect_key = 0;
325 xop->run_mask = HAMMER2_XOPMASK_VOP;
329 * A mounted PFS needs Xops threads to support frontend operations.
332 hammer2_xop_helper_create(hammer2_pfs_t *pmp)
337 lockmgr(&pmp->lock, LK_EXCLUSIVE);
338 pmp->has_xop_threads = 1;
340 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
341 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
342 if (pmp->xop_groups[j].thrs[i].td)
344 hammer2_thr_create(&pmp->xop_groups[j].thrs[i], pmp,
346 hammer2_primary_xops_thread);
349 lockmgr(&pmp->lock, LK_RELEASE);
353 hammer2_xop_helper_cleanup(hammer2_pfs_t *pmp)
358 for (i = 0; i < pmp->pfs_nmasters; ++i) {
359 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
360 if (pmp->xop_groups[j].thrs[i].td)
361 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
367 * Start a XOP request, queueing it to all nodes in the cluster to
368 * execute the cluster op.
370 * XXX optimize single-target case.
373 hammer2_xop_start_except(hammer2_xop_head_t *xop, hammer2_xop_func_t func,
376 hammer2_inode_t *ip1;
378 hammer2_thread_t *thr;
385 if (pmp->has_xop_threads == 0)
386 hammer2_xop_helper_create(pmp);
389 * The intent of the XOP sequencer is to ensure that ops on the same inode
390 * execute in the same order. This is necessary when issuing modifying operations
391 * to multiple targets because some targets might get behind and the frontend is
392 * allowed to complete the moment a quorum of targets succeed.
394 * Strategy operations must be segregated from non-strategy operations to avoid
395 * a deadlock. For example, if a vfsync and a bread/bwrite were queued to
396 * the same worker thread, the locked buffer in the strategy operation can deadlock
397 * the vfsync's buffer list scan.
399 * TODO - RENAME fails here because it is potentially modifying three different
402 if (xop->flags & HAMMER2_XOP_STRATEGY) {
403 hammer2_xop_strategy_t *xopst;
405 xopst = &((hammer2_xop_t *)xop)->xop_strategy;
406 ng = (int)(hammer2_icrc32(&xop->ip1, sizeof(xop->ip1)) ^
407 hammer2_icrc32(&xopst->lbase, sizeof(xopst->lbase)));
408 ng = ng & (HAMMER2_XOPGROUPS_MASK >> 1);
409 ng += HAMMER2_XOPGROUPS / 2;
411 ng = (int)(hammer2_icrc32(&xop->ip1, sizeof(xop->ip1)));
412 ng = ng & (HAMMER2_XOPGROUPS_MASK >> 1);
417 * The instant xop is queued another thread can pick it off. In the
418 * case of asynchronous ops, another thread might even finish and
421 hammer2_spin_ex(&pmp->xop_spin);
422 nchains = ip1->cluster.nchains;
423 for (i = 0; i < nchains; ++i) {
425 * XXX ip1->cluster.array* not stable here. This temporary
426 * hack fixes basic issues in target XOPs which need to
427 * obtain a starting chain from the inode but does not
428 * address possible races against inode updates which
429 * might NULL-out a chain.
431 if (i != notidx && ip1->cluster.array[i].chain) {
432 thr = &pmp->xop_groups[ng].thrs[i];
433 atomic_set_int(&xop->run_mask, 1U << i);
434 atomic_set_int(&xop->chk_mask, 1U << i);
435 TAILQ_INSERT_TAIL(&thr->xopq, xop, collect[i].entry);
438 hammer2_spin_unex(&pmp->xop_spin);
439 /* xop can become invalid at this point */
442 * Each thread has its own xopq
444 for (i = 0; i < nchains; ++i) {
446 thr = &pmp->xop_groups[ng].thrs[i];
447 hammer2_thr_signal(thr, HAMMER2_THREAD_XOPQ);
453 hammer2_xop_start(hammer2_xop_head_t *xop, hammer2_xop_func_t func)
455 hammer2_xop_start_except(xop, func, -1);
459 * Retire a XOP. Used by both the VOP frontend and by the XOP backend.
462 hammer2_xop_retire(hammer2_xop_head_t *xop, uint32_t mask)
464 hammer2_chain_t *chain;
469 * Remove the frontend collector or remove a backend feeder.
470 * When removing the frontend we must wakeup any backend feeders
471 * who are waiting for FIFO space.
473 * XXX optimize wakeup.
475 KKASSERT(xop->run_mask & mask);
476 nmask = atomic_fetchadd_int(&xop->run_mask, -mask);
477 if ((nmask & ~HAMMER2_XOPMASK_FIFOW) != mask) {
478 if (mask == HAMMER2_XOPMASK_VOP) {
479 if (nmask & HAMMER2_XOPMASK_FIFOW)
484 /* else nobody else left, we can ignore FIFOW */
487 * All collectors are gone, we can cleanup and dispose of the XOP.
488 * Note that this can wind up being a frontend OR a backend.
489 * Pending chains are locked shared and not owned by any thread.
493 * Cache the terminating cluster.
496 if ((ip = xop->ip1) != NULL) {
497 hammer2_cluster_t *tmpclu;
499 tmpclu = hammer2_cluster_copy(&xop->cluster);
500 hammer2_spin_ex(&ip->cluster_spin);
501 tmpclu = atomic_swap_ptr((volatile void **)&ip->cluster_cache,
503 hammer2_spin_unex(&ip->cluster_spin);
505 hammer2_cluster_drop(tmpclu);
510 * Cleanup the collection cluster.
512 for (i = 0; i < xop->cluster.nchains; ++i) {
513 xop->cluster.array[i].flags = 0;
514 chain = xop->cluster.array[i].chain;
516 xop->cluster.array[i].chain = NULL;
517 hammer2_chain_drop_unhold(chain);
522 * Cleanup the fifos, use check_counter to optimize the loop.
523 * Since we are the only entity left on this xop we don't have
524 * to worry about fifo flow control, and one lfence() will do the
528 mask = xop->chk_mask;
529 for (i = 0; mask && i < HAMMER2_MAXCLUSTER; ++i) {
530 hammer2_xop_fifo_t *fifo = &xop->collect[i];
531 while (fifo->ri != fifo->wi) {
532 chain = fifo->array[fifo->ri & HAMMER2_XOPFIFO_MASK];
534 hammer2_chain_drop_unhold(chain);
541 * The inode is only held at this point, simply drop it.
544 hammer2_inode_drop(xop->ip1);
548 hammer2_inode_drop(xop->ip2);
552 hammer2_inode_drop(xop->ip3);
556 kfree(xop->name1, M_HAMMER2);
561 kfree(xop->name2, M_HAMMER2);
566 objcache_put(cache_xops, xop);
570 * (Backend) Returns non-zero if the frontend is still attached.
573 hammer2_xop_active(hammer2_xop_head_t *xop)
575 if (xop->run_mask & HAMMER2_XOPMASK_VOP)
582 * (Backend) Feed chain data through the cluster validator and back to
583 * the frontend. Chains are fed from multiple nodes concurrently
584 * and pipelined via per-node FIFOs in the XOP.
586 * The chain must be locked (either shared or exclusive). The caller may
587 * unlock and drop the chain on return. This function will add an extra
588 * ref and hold the chain's data for the pass-back.
590 * No xop lock is needed because we are only manipulating fields under
591 * our direct control.
593 * Returns 0 on success and a hammer error code if sync is permanently
594 * lost. The caller retains a ref on the chain but by convention
595 * the lock is typically inherited by the xop (caller loses lock).
597 * Returns non-zero on error. In this situation the caller retains a
598 * ref on the chain but loses the lock (we unlock here).
601 hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain,
602 int clindex, int error)
604 hammer2_xop_fifo_t *fifo;
608 * Early termination (typicaly of xop_readir)
610 if (hammer2_xop_active(xop) == 0) {
616 * Multi-threaded entry into the XOP collector. We own the
617 * fifo->wi for our clindex.
619 fifo = &xop->collect[clindex];
621 if (fifo->ri == fifo->wi - HAMMER2_XOPFIFO)
623 while (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) {
624 atomic_set_int(&fifo->flags, HAMMER2_XOP_FIFO_STALL);
625 mask = xop->run_mask;
626 if ((mask & HAMMER2_XOPMASK_VOP) == 0) {
630 tsleep_interlock(xop, 0);
631 if (atomic_cmpset_int(&xop->run_mask, mask,
632 mask | HAMMER2_XOPMASK_FIFOW)) {
633 if (fifo->ri == fifo->wi - HAMMER2_XOPFIFO) {
634 tsleep(xop, PINTERLOCKED, "h2feed", hz*60);
639 atomic_clear_int(&fifo->flags, HAMMER2_XOP_FIFO_STALL);
641 hammer2_chain_ref_hold(chain);
642 if (error == 0 && chain)
643 error = chain->error;
644 fifo->errors[fifo->wi & HAMMER2_XOPFIFO_MASK] = error;
645 fifo->array[fifo->wi & HAMMER2_XOPFIFO_MASK] = chain;
648 if (atomic_fetchadd_int(&xop->check_counter, HAMMER2_XOP_CHKINC) &
649 HAMMER2_XOP_CHKWAIT) {
650 atomic_clear_int(&xop->check_counter, HAMMER2_XOP_CHKWAIT);
651 wakeup(&xop->check_counter);
656 * Cleanup. If an error occurred we eat the lock. If no error
657 * occurred the fifo inherits the lock and gains an additional ref.
659 * The caller's ref remains in both cases.
666 * (Frontend) collect a response from a running cluster op.
668 * Responses are fed from all appropriate nodes concurrently
669 * and collected into a cohesive response >= collect_key.
671 * The collector will return the instant quorum or other requirements
672 * are met, even if some nodes get behind or become non-responsive.
674 * HAMMER2_XOP_COLLECT_NOWAIT - Used to 'poll' a completed collection,
675 * usually called synchronously from the
676 * node XOPs for the strategy code to
677 * fake the frontend collection and complete
678 * the BIO as soon as possible.
680 * HAMMER2_XOP_SYNCHRONIZER - Reqeuest synchronization with a particular
681 * cluster index, prevents looping when that
682 * index is out of sync so caller can act on
683 * the out of sync element. ESRCH and EDEADLK
684 * can be returned if this flag is specified.
686 * Returns 0 on success plus a filled out xop->cluster structure.
687 * Return ENOENT on normal termination.
688 * Otherwise return an error.
691 hammer2_xop_collect(hammer2_xop_head_t *xop, int flags)
693 hammer2_xop_fifo_t *fifo;
694 hammer2_chain_t *chain;
698 int adv; /* advance the element */
700 uint32_t check_counter;
704 * First loop tries to advance pieces of the cluster which
707 lokey = HAMMER2_KEY_MAX;
708 keynull = HAMMER2_CHECK_NULL;
709 check_counter = xop->check_counter;
712 for (i = 0; i < xop->cluster.nchains; ++i) {
713 chain = xop->cluster.array[i].chain;
716 } else if (chain->bref.key < xop->collect_key) {
719 keynull &= ~HAMMER2_CHECK_NULL;
720 if (lokey > chain->bref.key)
721 lokey = chain->bref.key;
728 * Advance element if possible, advanced element may be NULL.
731 hammer2_chain_drop_unhold(chain);
733 fifo = &xop->collect[i];
734 if (fifo->ri != fifo->wi) {
736 chain = fifo->array[fifo->ri & HAMMER2_XOPFIFO_MASK];
737 error = fifo->errors[fifo->ri & HAMMER2_XOPFIFO_MASK];
739 xop->cluster.array[i].chain = chain;
740 xop->cluster.array[i].error = error;
743 xop->cluster.array[i].flags |=
746 if (fifo->wi - fifo->ri <= HAMMER2_XOPFIFO / 2) {
747 if (fifo->flags & HAMMER2_XOP_FIFO_STALL) {
748 atomic_clear_int(&fifo->flags,
749 HAMMER2_XOP_FIFO_STALL);
754 --i; /* loop on same index */
757 * Retain CITEM_NULL flag. If set just repeat EOF.
758 * If not, the NULL,0 combination indicates an
759 * operation in-progress.
761 xop->cluster.array[i].chain = NULL;
762 /* retain any CITEM_NULL setting */
767 * Determine whether the lowest collected key meets clustering
768 * requirements. Returns:
770 * 0 - key valid, cluster can be returned.
772 * ENOENT - normal end of scan, return ENOENT.
774 * ESRCH - sufficient elements collected, quorum agreement
775 * that lokey is not a valid element and should be
778 * EDEADLK - sufficient elements collected, no quorum agreement
779 * (and no agreement possible). In this situation a
780 * repair is needed, for now we loop.
782 * EINPROGRESS - insufficient elements collected to resolve, wait
783 * for event and loop.
785 if ((flags & HAMMER2_XOP_COLLECT_WAITALL) &&
786 xop->run_mask != HAMMER2_XOPMASK_VOP) {
789 error = hammer2_cluster_check(&xop->cluster, lokey, keynull);
791 if (error == EINPROGRESS) {
792 if ((flags & HAMMER2_XOP_COLLECT_NOWAIT) == 0)
793 tsleep_interlock(&xop->check_counter, 0);
794 if (atomic_cmpset_int(&xop->check_counter,
796 check_counter | HAMMER2_XOP_CHKWAIT)) {
797 if (flags & HAMMER2_XOP_COLLECT_NOWAIT)
799 tsleep(&xop->check_counter, PINTERLOCKED, "h2coll", hz*60);
803 if (error == ESRCH) {
804 if (lokey != HAMMER2_KEY_MAX) {
805 xop->collect_key = lokey + 1;
810 if (error == EDEADLK) {
811 kprintf("hammer2: no quorum possible lokey %016jx\n",
813 if (lokey != HAMMER2_KEY_MAX) {
814 xop->collect_key = lokey + 1;
819 if (lokey == HAMMER2_KEY_MAX)
820 xop->collect_key = lokey;
822 xop->collect_key = lokey + 1;
828 * N x M processing threads are available to handle XOPs, N per cluster
829 * index x M cluster nodes. All the threads for any given cluster index
830 * share and pull from the same xopq.
832 * Locate and return the next runnable xop, or NULL if no xops are
833 * present or none of the xops are currently runnable (for various reasons).
834 * The xop is left on the queue and serves to block other dependent xops
837 * Dependent xops will not be returned.
839 * Sets HAMMER2_XOP_FIFO_RUN on the returned xop or returns NULL.
841 * NOTE! Xops run concurrently for each cluster index.
843 #define XOP_HASH_SIZE 16
844 #define XOP_HASH_MASK (XOP_HASH_SIZE - 1)
848 xop_testhash(hammer2_thread_t *thr, hammer2_inode_t *ip, uint32_t *hash)
853 hv = (int)((uintptr_t)ip + (uintptr_t)thr) / sizeof(hammer2_inode_t);
854 mask = 1U << (hv & 31);
857 return ((int)(hash[hv & XOP_HASH_MASK] & mask));
862 xop_sethash(hammer2_thread_t *thr, hammer2_inode_t *ip, uint32_t *hash)
867 hv = (int)((uintptr_t)ip + (uintptr_t)thr) / sizeof(hammer2_inode_t);
868 mask = 1U << (hv & 31);
871 hash[hv & XOP_HASH_MASK] |= mask;
876 hammer2_xop_next(hammer2_thread_t *thr)
878 hammer2_pfs_t *pmp = thr->pmp;
879 int clindex = thr->clindex;
880 uint32_t hash[XOP_HASH_SIZE] = { 0 };
881 hammer2_xop_head_t *xop;
883 hammer2_spin_ex(&pmp->xop_spin);
884 TAILQ_FOREACH(xop, &thr->xopq, collect[clindex].entry) {
888 if (xop_testhash(thr, xop->ip1, hash) ||
889 (xop->ip2 && xop_testhash(thr, xop->ip2, hash)) ||
890 (xop->ip3 && xop_testhash(thr, xop->ip3, hash))) {
893 xop_sethash(thr, xop->ip1, hash);
895 xop_sethash(thr, xop->ip2, hash);
897 xop_sethash(thr, xop->ip3, hash);
900 * Check already running
902 if (xop->collect[clindex].flags & HAMMER2_XOP_FIFO_RUN)
906 * Found a good one, return it.
908 atomic_set_int(&xop->collect[clindex].flags,
909 HAMMER2_XOP_FIFO_RUN);
912 hammer2_spin_unex(&pmp->xop_spin);
918 * Remove the completed XOP from the queue, clear HAMMER2_XOP_FIFO_RUN.
920 * NOTE! Xops run concurrently for each cluster index.
924 hammer2_xop_dequeue(hammer2_thread_t *thr, hammer2_xop_head_t *xop)
926 hammer2_pfs_t *pmp = thr->pmp;
927 int clindex = thr->clindex;
929 hammer2_spin_ex(&pmp->xop_spin);
930 TAILQ_REMOVE(&thr->xopq, xop, collect[clindex].entry);
931 atomic_clear_int(&xop->collect[clindex].flags,
932 HAMMER2_XOP_FIFO_RUN);
933 hammer2_spin_unex(&pmp->xop_spin);
934 if (TAILQ_FIRST(&thr->xopq))
935 hammer2_thr_signal(thr, HAMMER2_THREAD_XOPQ);
939 * Primary management thread for xops support. Each node has several such
940 * threads which replicate front-end operations on cluster nodes.
942 * XOPS thread node operations, allowing the function to focus on a single
943 * node in the cluster after validating the operation with the cluster.
944 * This is primarily what prevents dead or stalled nodes from stalling
948 hammer2_primary_xops_thread(void *arg)
950 hammer2_thread_t *thr = arg;
952 hammer2_xop_head_t *xop;
956 hammer2_xop_func_t last_func = NULL;
959 /*xgrp = &pmp->xop_groups[thr->repidx]; not needed */
960 mask = 1U << thr->clindex;
966 * Handle stop request
968 if (flags & HAMMER2_THREAD_STOP)
972 * Handle freeze request
974 if (flags & HAMMER2_THREAD_FREEZE) {
975 nflags = (flags & ~(HAMMER2_THREAD_FREEZE |
976 HAMMER2_THREAD_CLIENTWAIT)) |
977 HAMMER2_THREAD_FROZEN;
978 if (!atomic_cmpset_int(&thr->flags, flags, nflags))
980 if (flags & HAMMER2_THREAD_CLIENTWAIT)
986 if (flags & HAMMER2_THREAD_UNFREEZE) {
987 nflags = flags & ~(HAMMER2_THREAD_UNFREEZE |
988 HAMMER2_THREAD_FROZEN |
989 HAMMER2_THREAD_CLIENTWAIT);
990 if (!atomic_cmpset_int(&thr->flags, flags, nflags))
992 if (flags & HAMMER2_THREAD_CLIENTWAIT)
999 * Force idle if frozen until unfrozen or stopped.
1001 if (flags & HAMMER2_THREAD_FROZEN) {
1002 nflags = flags | HAMMER2_THREAD_WAITING;
1003 tsleep_interlock(&thr->flags, 0);
1004 if (atomic_cmpset_int(&thr->flags, flags, nflags)) {
1005 tsleep(&thr->flags, PINTERLOCKED, "frozen", 0);
1006 atomic_clear_int(&thr->flags,
1007 HAMMER2_THREAD_WAITING);
1013 * Reset state on REMASTER request
1015 if (flags & HAMMER2_THREAD_REMASTER) {
1016 nflags = flags & ~HAMMER2_THREAD_REMASTER;
1017 if (atomic_cmpset_int(&thr->flags, flags, nflags)) {
1018 /* reset state here */
1024 * Process requests. Each request can be multi-queued.
1026 * If we get behind and the frontend VOP is no longer active,
1027 * we retire the request without processing it. The callback
1028 * may also abort processing if the frontend VOP becomes
1031 if (flags & HAMMER2_THREAD_XOPQ) {
1032 nflags = flags & ~HAMMER2_THREAD_XOPQ;
1033 if (!atomic_cmpset_int(&thr->flags, flags, nflags))
1038 while ((xop = hammer2_xop_next(thr)) != NULL) {
1039 if (hammer2_xop_active(xop)) {
1040 last_func = xop->func;
1041 xop->func((hammer2_xop_t *)xop, thr->clindex);
1042 hammer2_xop_dequeue(thr, xop);
1043 hammer2_xop_retire(xop, mask);
1045 last_func = xop->func;
1046 hammer2_xop_feed(xop, NULL, thr->clindex,
1048 hammer2_xop_dequeue(thr, xop);
1049 hammer2_xop_retire(xop, mask);
1054 * Wait for event, interlock using THREAD_WAITING and
1057 * For robustness poll on a 30-second interval, but nominally
1058 * expect to be woken up.
1060 nflags = flags | HAMMER2_THREAD_WAITING;
1062 tsleep_interlock(&thr->flags, 0);
1063 if (atomic_cmpset_int(&thr->flags, flags, nflags)) {
1064 tsleep(&thr->flags, PINTERLOCKED, "h2idle", hz*30);
1065 atomic_clear_int(&thr->flags, HAMMER2_THREAD_WAITING);
1071 * Cleanup / termination
1073 while ((xop = TAILQ_FIRST(&thr->xopq)) != NULL) {
1074 kprintf("hammer2_thread: aborting xop %p\n", xop->func);
1075 TAILQ_REMOVE(&thr->xopq, xop,
1076 collect[thr->clindex].entry);
1077 hammer2_xop_retire(xop, mask);
1081 hammer2_thr_return(thr, HAMMER2_THREAD_STOPPED);
1082 /* thr structure can go invalid after this point */