1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_page for OSC layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_OSC
43 #include "osc_cl_internal.h"
50 * Comment out osc_page_protected because it may sleep inside the
51 * the client_obd_list_lock.
52 * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
53 * -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
54 * -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
57 static int osc_page_is_dlocked(const struct lu_env *env,
58 const struct osc_page *opg,
59 enum cl_lock_mode mode, int pending, int unref)
62 struct osc_object *obj;
63 struct osc_thread_info *info;
64 struct ldlm_res_id *resname;
65 struct lustre_handle *lockh;
66 ldlm_policy_data_t *policy;
72 info = osc_env_info(env);
73 resname = &info->oti_resname;
74 policy = &info->oti_policy;
75 lockh = &info->oti_handle;
76 page = opg->ops_cl.cpl_page;
77 obj = cl2osc(opg->ops_cl.cpl_obj);
79 flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
81 flags |= LDLM_FL_CBPENDING;
83 dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
84 osc_lock_build_res(env, obj, resname);
85 osc_index2policy(policy, page->cp_obj, page->cp_index, page->cp_index);
86 return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
87 dlmmode, &flags, NULL, lockh, unref);
91 * Checks an invariant that a page in the cache is covered by a lock, as
94 static int osc_page_protected(const struct lu_env *env,
95 const struct osc_page *opg,
96 enum cl_lock_mode mode, int unref)
98 struct cl_object_header *hdr;
100 struct cl_page *page;
101 struct cl_lock_descr *descr;
104 LINVRNT(!opg->ops_temp);
106 page = opg->ops_cl.cpl_page;
107 if (page->cp_owner != NULL &&
108 cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
110 * If IO is done without locks (liblustre, or lloop), lock is
115 /* otherwise check for a DLM lock */
116 result = osc_page_is_dlocked(env, opg, mode, 1, unref);
118 /* maybe this page is a part of a lockless io? */
119 hdr = cl_object_header(opg->ops_cl.cpl_obj);
120 descr = &osc_env_info(env)->oti_descr;
121 descr->cld_mode = mode;
122 descr->cld_start = page->cp_index;
123 descr->cld_end = page->cp_index;
124 cfs_spin_lock(&hdr->coh_lock_guard);
125 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
127 * Lock-less sub-lock has to be either in HELD state
128 * (when io is actively going on), or in CACHED state,
129 * when top-lock is being unlocked:
130 * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
132 if ((scan->cll_state == CLS_HELD ||
133 scan->cll_state == CLS_CACHED) &&
134 cl_lock_ext_match(&scan->cll_descr, descr)) {
135 struct osc_lock *olck;
137 olck = osc_lock_at(scan);
138 result = osc_lock_is_lockless(olck);
142 cfs_spin_unlock(&hdr->coh_lock_guard);
147 static int osc_page_protected(const struct lu_env *env,
148 const struct osc_page *opg,
149 enum cl_lock_mode mode, int unref)
155 /*****************************************************************************
160 static void osc_page_fini(const struct lu_env *env,
161 struct cl_page_slice *slice)
163 struct osc_page *opg = cl2osc_page(slice);
164 CDEBUG(D_TRACE, "%p\n", opg);
165 LASSERT(opg->ops_lock == NULL);
166 OBD_SLAB_FREE_PTR(opg, osc_page_kmem);
169 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
171 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
173 LASSERT(!opg->ops_transfer_pinned);
175 lu_ref_add_atomic(&page->cp_reference, label, page);
176 opg->ops_transfer_pinned = 1;
179 static void osc_page_transfer_put(const struct lu_env *env,
180 struct osc_page *opg)
182 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
184 if (opg->ops_transfer_pinned) {
185 lu_ref_del(&page->cp_reference, "transfer", page);
186 opg->ops_transfer_pinned = 0;
187 cl_page_put(env, page);
192 * This is called once for every page when it is submitted for a transfer
193 * either opportunistic (osc_page_cache_add()), or immediate
194 * (osc_page_submit()).
196 static void osc_page_transfer_add(const struct lu_env *env,
197 struct osc_page *opg, enum cl_req_type crt)
199 struct osc_object *obj;
201 LINVRNT(cl_page_is_vmlocked(env, opg->ops_cl.cpl_page));
203 obj = cl2osc(opg->ops_cl.cpl_obj);
204 cfs_spin_lock(&obj->oo_seatbelt);
205 cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
206 opg->ops_submitter = cfs_current();
207 cfs_spin_unlock(&obj->oo_seatbelt);
210 static int osc_page_cache_add(const struct lu_env *env,
211 const struct cl_page_slice *slice,
212 struct cl_io *unused)
214 struct osc_page *opg = cl2osc_page(slice);
215 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
217 /* All cacheable IO is async-capable */
218 int brw_flags = OBD_BRW_ASYNC;
221 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
224 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
225 brw_flags |= opg->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
226 if (!client_is_remote(osc_export(obj)) &&
227 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
228 brw_flags |= OBD_BRW_NOQUOTA;
229 noquota = OBD_BRW_NOQUOTA;
232 osc_page_transfer_get(opg, "transfer\0cache");
233 result = osc_queue_async_io(env, osc_export(obj), NULL, obj->oo_oinfo,
234 &opg->ops_oap, OBD_BRW_WRITE | noquota,
235 opg->ops_from, opg->ops_to - opg->ops_from,
238 osc_page_transfer_put(env, opg);
240 osc_page_transfer_add(env, opg, CRT_WRITE);
244 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
245 pgoff_t start, pgoff_t end)
247 memset(policy, 0, sizeof *policy);
248 policy->l_extent.start = cl_offset(obj, start);
249 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
252 static int osc_page_addref_lock(const struct lu_env *env,
253 struct osc_page *opg,
254 struct cl_lock *lock)
256 struct osc_lock *olock;
259 LASSERT(opg->ops_lock == NULL);
261 olock = osc_lock_at(lock);
262 if (cfs_atomic_inc_return(&olock->ols_pageref) <= 0) {
263 cfs_atomic_dec(&olock->ols_pageref);
264 cl_lock_put(env, lock);
267 opg->ops_lock = lock;
273 static void osc_page_putref_lock(const struct lu_env *env,
274 struct osc_page *opg)
276 struct cl_lock *lock = opg->ops_lock;
277 struct osc_lock *olock;
279 LASSERT(lock != NULL);
280 olock = osc_lock_at(lock);
282 cfs_atomic_dec(&olock->ols_pageref);
283 opg->ops_lock = NULL;
286 * Note: usually this won't be the last reference of the lock, but if
287 * it is, then all the lock_put do is at most just freeing some memory,
288 * so it would be OK that caller is holding spinlocks.
290 LASSERT(cfs_atomic_read(&lock->cll_ref) > 1 || olock->ols_hold == 0);
291 cl_lock_put(env, lock);
294 static int osc_page_is_under_lock(const struct lu_env *env,
295 const struct cl_page_slice *slice,
296 struct cl_io *unused)
298 struct cl_lock *lock;
302 lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
305 osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
312 static void osc_page_disown(const struct lu_env *env,
313 const struct cl_page_slice *slice,
316 struct osc_page *opg = cl2osc_page(slice);
318 if (unlikely(opg->ops_lock))
319 osc_page_putref_lock(env, opg);
322 static void osc_page_completion_read(const struct lu_env *env,
323 const struct cl_page_slice *slice,
326 struct osc_page *opg = cl2osc_page(slice);
328 if (likely(opg->ops_lock))
329 osc_page_putref_lock(env, opg);
332 static int osc_page_fail(const struct lu_env *env,
333 const struct cl_page_slice *slice,
334 struct cl_io *unused)
344 static const char *osc_list(cfs_list_t *head)
346 return cfs_list_empty(head) ? "-" : "+";
349 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
351 if (opg->ops_submit_time == 0)
354 return (cfs_time_current() - opg->ops_submit_time);
357 static int osc_page_print(const struct lu_env *env,
358 const struct cl_page_slice *slice,
359 void *cookie, lu_printer_t printer)
361 struct osc_page *opg = cl2osc_page(slice);
362 struct osc_async_page *oap = &opg->ops_oap;
363 struct osc_object *obj = cl2osc(slice->cpl_obj);
364 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
365 struct lov_oinfo *loi = obj->oo_oinfo;
367 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
368 "1< %#x %d %u %s %s %s > "
369 "2< "LPU64" %u %u %#x %#x | %p %p %p %p %p > "
370 "3< %s %p %d %lu %d > "
371 "4< %d %d %d %lu %s | %s %s %s %s > "
372 "5< %s %s %s %s | %d %s %s | %d %s %s>\n",
375 oap->oap_magic, oap->oap_cmd,
376 oap->oap_interrupted,
377 osc_list(&oap->oap_pending_item),
378 osc_list(&oap->oap_urgent_item),
379 osc_list(&oap->oap_rpc_item),
381 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
382 oap->oap_async_flags, oap->oap_brw_flags,
384 oap->oap_cli, oap->oap_loi, oap->oap_caller_ops,
385 oap->oap_caller_data,
387 osc_list(&opg->ops_inflight),
388 opg->ops_submitter, opg->ops_transfer_pinned,
389 osc_submit_duration(opg), opg->ops_srvlock,
391 cli->cl_r_in_flight, cli->cl_w_in_flight,
392 cli->cl_max_rpcs_in_flight,
394 osc_list(&cli->cl_cache_waiters),
395 osc_list(&cli->cl_loi_ready_list),
396 osc_list(&cli->cl_loi_hp_ready_list),
397 osc_list(&cli->cl_loi_write_list),
398 osc_list(&cli->cl_loi_read_list),
400 osc_list(&loi->loi_ready_item),
401 osc_list(&loi->loi_hp_ready_item),
402 osc_list(&loi->loi_write_item),
403 osc_list(&loi->loi_read_item),
404 loi->loi_read_lop.lop_num_pending,
405 osc_list(&loi->loi_read_lop.lop_pending),
406 osc_list(&loi->loi_read_lop.lop_urgent),
407 loi->loi_write_lop.lop_num_pending,
408 osc_list(&loi->loi_write_lop.lop_pending),
409 osc_list(&loi->loi_write_lop.lop_urgent));
412 static void osc_page_delete(const struct lu_env *env,
413 const struct cl_page_slice *slice)
415 struct osc_page *opg = cl2osc_page(slice);
416 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
417 struct osc_async_page *oap = &opg->ops_oap;
420 LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
423 CDEBUG(D_TRACE, "%p\n", opg);
424 osc_page_transfer_put(env, opg);
425 rc = osc_teardown_async_page(osc_export(obj), NULL, obj->oo_oinfo, oap);
427 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
428 "Trying to teardown failed: %d\n", rc);
431 cfs_spin_lock(&obj->oo_seatbelt);
432 cfs_list_del_init(&opg->ops_inflight);
433 cfs_spin_unlock(&obj->oo_seatbelt);
437 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
440 struct osc_page *opg = cl2osc_page(slice);
441 struct osc_async_page *oap = &opg->ops_oap;
443 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
445 opg->ops_from = from;
447 cfs_spin_lock(&oap->oap_lock);
448 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
449 cfs_spin_unlock(&oap->oap_lock);
452 static int osc_page_cancel(const struct lu_env *env,
453 const struct cl_page_slice *slice)
455 struct osc_page *opg = cl2osc_page(slice);
456 struct osc_async_page *oap = &opg->ops_oap;
459 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
461 client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
462 /* Check if the transferring against this page
463 * is completed, or not even queued. */
464 if (opg->ops_transfer_pinned)
465 /* FIXME: may not be interrupted.. */
466 rc = osc_oap_interrupted(env, oap);
467 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
468 client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
472 static const struct cl_page_operations osc_page_ops = {
473 .cpo_fini = osc_page_fini,
474 .cpo_print = osc_page_print,
475 .cpo_delete = osc_page_delete,
476 .cpo_is_under_lock = osc_page_is_under_lock,
477 .cpo_disown = osc_page_disown,
480 .cpo_cache_add = osc_page_fail,
481 .cpo_completion = osc_page_completion_read
484 .cpo_cache_add = osc_page_cache_add
487 .cpo_clip = osc_page_clip,
488 .cpo_cancel = osc_page_cancel
491 static int osc_make_ready(const struct lu_env *env, void *data, int cmd)
493 struct osc_page *opg = data;
494 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
497 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
498 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 1));
501 result = cl_page_make_ready(env, page, CRT_WRITE);
503 opg->ops_submit_time = cfs_time_current();
507 static int osc_refresh_count(const struct lu_env *env, void *data, int cmd)
509 struct cl_page *page;
510 struct osc_page *osc = data;
511 struct cl_object *obj;
512 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
517 LINVRNT(osc_page_protected(env, osc, CLM_READ, 1));
519 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
520 LASSERT(!(cmd & OBD_BRW_READ));
521 LASSERT(osc != NULL);
522 page = osc->ops_cl.cpl_page;
523 obj = osc->ops_cl.cpl_obj;
525 cl_object_attr_lock(obj);
526 result = cl_object_attr_get(env, obj, attr);
527 cl_object_attr_unlock(obj);
531 if (cl_offset(obj, page->cp_index) >= kms)
532 /* catch race with truncate */
534 else if (cl_offset(obj, page->cp_index + 1) > kms)
535 /* catch sub-page write at end of file */
536 return kms % CFS_PAGE_SIZE;
538 return CFS_PAGE_SIZE;
541 static int osc_completion(const struct lu_env *env,
542 void *data, int cmd, struct obdo *oa, int rc)
544 struct osc_page *opg = data;
545 struct osc_async_page *oap = &opg->ops_oap;
546 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
547 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
548 enum cl_req_type crt;
550 LINVRNT(osc_page_protected(env, opg, CLM_READ, 1));
551 LINVRNT(cl_page_is_vmlocked(env, page));
555 cmd &= ~OBD_BRW_NOQUOTA;
556 LASSERT(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ));
557 LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
558 LASSERT(opg->ops_transfer_pinned);
561 * page->cp_req can be NULL if io submission failed before
562 * cl_req was allocated.
564 if (page->cp_req != NULL)
565 cl_req_page_done(env, page);
566 LASSERT(page->cp_req == NULL);
568 /* As the transfer for this page is being done, clear the flags */
569 cfs_spin_lock(&oap->oap_lock);
570 oap->oap_async_flags = 0;
571 cfs_spin_unlock(&oap->oap_lock);
573 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
574 /* Clear opg->ops_transfer_pinned before VM lock is released. */
575 opg->ops_transfer_pinned = 0;
577 cfs_spin_lock(&obj->oo_seatbelt);
578 LASSERT(opg->ops_submitter != NULL);
579 LASSERT(!cfs_list_empty(&opg->ops_inflight));
580 cfs_list_del_init(&opg->ops_inflight);
581 cfs_spin_unlock(&obj->oo_seatbelt);
583 opg->ops_submit_time = 0;
585 cl_page_completion(env, page, crt, rc);
588 if (rc == 0 && oap->oap_brw_flags & OBD_BRW_SRVLOCK) {
589 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
590 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
591 int bytes = oap->oap_count;
594 stats->os_lockless_reads += bytes;
596 stats->os_lockless_writes += bytes;
600 * This has to be the last operation with the page, as locks are
601 * released in cl_page_completion() and nothing except for the
602 * reference counter protects page from concurrent reclaim.
604 lu_ref_del(&page->cp_reference, "transfer", page);
606 * As page->cp_obj is pinned by a reference from page->cp_req, it is
607 * safe to call cl_page_put() without risking object destruction in a
608 * non-blocking context.
610 cl_page_put(env, page);
614 const static struct obd_async_page_ops osc_async_page_ops = {
615 .ap_make_ready = osc_make_ready,
616 .ap_refresh_count = osc_refresh_count,
617 .ap_completion = osc_completion
620 struct cl_page *osc_page_init(const struct lu_env *env,
621 struct cl_object *obj,
622 struct cl_page *page, cfs_page_t *vmpage)
624 struct osc_object *osc = cl2osc(obj);
625 struct osc_page *opg;
628 OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO);
630 void *oap = &opg->ops_oap;
633 opg->ops_to = CFS_PAGE_SIZE;
635 result = osc_prep_async_page(osc_export(osc),
636 NULL, osc->oo_oinfo, vmpage,
637 cl_offset(obj, page->cp_index),
639 opg, (void **)&oap, 1, NULL);
641 struct osc_io *oio = osc_env_io(env);
642 opg->ops_srvlock = osc_io_srvlock(oio);
643 cl_page_slice_add(page, &opg->ops_cl, obj,
647 * Cannot assert osc_page_protected() here as read-ahead
648 * creates temporary pages outside of a lock.
650 #ifdef INVARIANT_CHECK
651 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
653 CFS_INIT_LIST_HEAD(&opg->ops_inflight);
656 return ERR_PTR(result);
660 * Helper function called by osc_io_submit() for every page in an immediate
661 * transfer (i.e., transferred synchronously).
663 void osc_io_submit_page(const struct lu_env *env,
664 struct osc_io *oio, struct osc_page *opg,
665 enum cl_req_type crt)
667 struct osc_async_page *oap = &opg->ops_oap;
668 struct client_obd *cli = oap->oap_cli;
671 LINVRNT(osc_page_protected(env, opg,
672 crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
674 oap->oap_page_off = opg->ops_from;
675 oap->oap_count = opg->ops_to - opg->ops_from;
676 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
677 if (cfs_memory_pressure_get())
678 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
679 oap->oap_brw_flags |= OBD_BRW_SYNC;
680 if (osc_io_srvlock(oio))
681 oap->oap_brw_flags |= OBD_BRW_SRVLOCK;
683 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
684 if (!client_is_remote(osc_export(cl2osc(opg->ops_cl.cpl_obj))) &&
685 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
686 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
687 oap->oap_cmd |= OBD_BRW_NOQUOTA;
690 if (oap->oap_cmd & OBD_BRW_READ)
691 flags = ASYNC_COUNT_STABLE;
692 else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT))
693 osc_enter_cache_try(env, cli, oap->oap_loi, oap, 1);
695 cfs_spin_lock(&oap->oap_lock);
696 oap->oap_async_flags |= OSC_FLAGS | flags;
697 cfs_spin_unlock(&oap->oap_lock);
699 osc_oap_to_pending(oap);
700 osc_page_transfer_get(opg, "transfer\0imm");
701 osc_page_transfer_add(env, opg, crt);