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 2008 Sun Microsystems, Inc. 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 /** \addtogroup osc osc @{ */
43 #define DEBUG_SUBSYSTEM S_OSC
45 #include "osc_cl_internal.h"
48 * Comment out osc_page_protected because it may sleep inside the
49 * the client_obd_list_lock.
50 * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
51 * -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
52 * -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
55 static int osc_page_is_dlocked(const struct lu_env *env,
56 const struct osc_page *opg,
57 enum cl_lock_mode mode, int pending, int unref)
60 struct osc_object *obj;
61 struct osc_thread_info *info;
62 struct ldlm_res_id *resname;
63 struct lustre_handle *lockh;
64 ldlm_policy_data_t *policy;
70 info = osc_env_info(env);
71 resname = &info->oti_resname;
72 policy = &info->oti_policy;
73 lockh = &info->oti_handle;
74 page = opg->ops_cl.cpl_page;
75 obj = cl2osc(opg->ops_cl.cpl_obj);
77 flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
79 flags |= LDLM_FL_CBPENDING;
81 dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
82 osc_lock_build_res(env, obj, resname);
83 osc_index2policy(policy, page->cp_obj, page->cp_index, page->cp_index);
84 return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
85 dlmmode, &flags, NULL, lockh, unref);
89 * Checks an invariant that a page in the cache is covered by a lock, as
92 static int osc_page_protected(const struct lu_env *env,
93 const struct osc_page *opg,
94 enum cl_lock_mode mode, int unref)
96 struct cl_object_header *hdr;
99 struct cl_lock_descr *descr;
102 LINVRNT(!opg->ops_temp);
104 page = opg->ops_cl.cpl_page;
105 if (page->cp_owner != NULL &&
106 cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
108 * If IO is done without locks (liblustre, or lloop), lock is
113 /* otherwise check for a DLM lock */
114 result = osc_page_is_dlocked(env, opg, mode, 1, unref);
116 /* maybe this page is a part of a lockless io? */
117 hdr = cl_object_header(opg->ops_cl.cpl_obj);
118 descr = &osc_env_info(env)->oti_descr;
119 descr->cld_mode = mode;
120 descr->cld_start = page->cp_index;
121 descr->cld_end = page->cp_index;
122 spin_lock(&hdr->coh_lock_guard);
123 list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
125 * Lock-less sub-lock has to be either in HELD state
126 * (when io is actively going on), or in CACHED state,
127 * when top-lock is being unlocked:
128 * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
130 if ((scan->cll_state == CLS_HELD ||
131 scan->cll_state == CLS_CACHED) &&
132 cl_lock_ext_match(&scan->cll_descr, descr)) {
133 struct osc_lock *olck;
135 olck = osc_lock_at(scan);
136 result = osc_lock_is_lockless(olck);
140 spin_unlock(&hdr->coh_lock_guard);
145 static int osc_page_protected(const struct lu_env *env,
146 const struct osc_page *opg,
147 enum cl_lock_mode mode, int unref)
153 /*****************************************************************************
158 static void osc_page_fini(const struct lu_env *env,
159 struct cl_page_slice *slice)
161 struct osc_page *opg = cl2osc_page(slice);
162 CDEBUG(D_TRACE, "%p\n", opg);
163 OBD_SLAB_FREE_PTR(opg, osc_page_kmem);
166 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
168 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
170 LASSERT(!opg->ops_transfer_pinned);
172 lu_ref_add_atomic(&page->cp_reference, label, page);
173 opg->ops_transfer_pinned = 1;
176 static void osc_page_transfer_put(const struct lu_env *env,
177 struct osc_page *opg)
179 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
181 if (opg->ops_transfer_pinned) {
182 lu_ref_del(&page->cp_reference, "transfer", page);
183 opg->ops_transfer_pinned = 0;
184 cl_page_put(env, page);
189 * This is called once for every page when it is submitted for a transfer
190 * either opportunistic (osc_page_cache_add()), or immediate
191 * (osc_page_submit()).
193 static void osc_page_transfer_add(const struct lu_env *env,
194 struct osc_page *opg, enum cl_req_type crt)
196 struct osc_object *obj;
198 LINVRNT(cl_page_is_vmlocked(env, opg->ops_cl.cpl_page));
200 obj = cl2osc(opg->ops_cl.cpl_obj);
201 spin_lock(&obj->oo_seatbelt);
202 list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
203 opg->ops_submitter = cfs_current();
204 spin_unlock(&obj->oo_seatbelt);
207 static int osc_page_cache_add(const struct lu_env *env,
208 const struct cl_page_slice *slice,
209 struct cl_io *unused)
211 struct osc_page *opg = cl2osc_page(slice);
212 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
213 struct osc_io *oio = osc_env_io(env);
218 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
221 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
222 brw_flags = osc_io_srvlock(oio) ? OBD_BRW_SRVLOCK : 0;
223 if (!client_is_remote(osc_export(obj)) &&
224 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
225 brw_flags |= OBD_BRW_NOQUOTA;
226 noquota = OBD_BRW_NOQUOTA;
229 osc_page_transfer_get(opg, "transfer\0cache");
230 result = osc_queue_async_io(env, osc_export(obj), NULL, obj->oo_oinfo,
231 &opg->ops_oap, OBD_BRW_WRITE | noquota,
234 osc_page_transfer_put(env, opg);
236 osc_page_transfer_add(env, opg, CRT_WRITE);
240 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
241 pgoff_t start, pgoff_t end)
243 memset(policy, 0, sizeof *policy);
244 policy->l_extent.start = cl_offset(obj, start);
245 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
248 static int osc_page_is_under_lock(const struct lu_env *env,
249 const struct cl_page_slice *slice,
250 struct cl_io *unused)
252 struct cl_lock *lock;
256 lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
259 cl_lock_put(env, lock);
266 static int osc_page_fail(const struct lu_env *env,
267 const struct cl_page_slice *slice,
268 struct cl_io *unused)
278 static const char *osc_list(struct list_head *head)
280 return list_empty(head) ? "-" : "+";
283 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
285 if (opg->ops_submit_time == 0)
288 return (cfs_time_current() - opg->ops_submit_time);
291 static int osc_page_print(const struct lu_env *env,
292 const struct cl_page_slice *slice,
293 void *cookie, lu_printer_t printer)
295 struct osc_page *opg = cl2osc_page(slice);
296 struct osc_async_page *oap = &opg->ops_oap;
297 struct osc_object *obj = cl2osc(slice->cpl_obj);
298 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
299 struct lov_oinfo *loi = obj->oo_oinfo;
301 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
302 "1< %#x %d %u %s %s %s > "
303 "2< %llu %u %#x %#x | %p %p %p %p %p > "
305 "4< %d %d %d %lu %s | %s %s %s %s > "
306 "5< %s %s %s %s | %d %s %s | %d %s %s>\n",
309 oap->oap_magic, oap->oap_cmd,
310 oap->oap_interrupted,
311 osc_list(&oap->oap_pending_item),
312 osc_list(&oap->oap_urgent_item),
313 osc_list(&oap->oap_rpc_item),
315 oap->oap_obj_off, oap->oap_page_off,
316 oap->oap_async_flags, oap->oap_brw_flags,
318 oap->oap_cli, oap->oap_loi, oap->oap_caller_ops,
319 oap->oap_caller_data,
321 osc_list(&opg->ops_inflight),
322 opg->ops_submitter, opg->ops_transfer_pinned,
323 osc_submit_duration(opg),
325 cli->cl_r_in_flight, cli->cl_w_in_flight,
326 cli->cl_max_rpcs_in_flight,
328 osc_list(&cli->cl_cache_waiters),
329 osc_list(&cli->cl_loi_ready_list),
330 osc_list(&cli->cl_loi_hp_ready_list),
331 osc_list(&cli->cl_loi_write_list),
332 osc_list(&cli->cl_loi_read_list),
334 osc_list(&loi->loi_ready_item),
335 osc_list(&loi->loi_hp_ready_item),
336 osc_list(&loi->loi_write_item),
337 osc_list(&loi->loi_read_item),
338 loi->loi_read_lop.lop_num_pending,
339 osc_list(&loi->loi_read_lop.lop_pending),
340 osc_list(&loi->loi_read_lop.lop_urgent),
341 loi->loi_write_lop.lop_num_pending,
342 osc_list(&loi->loi_write_lop.lop_pending),
343 osc_list(&loi->loi_write_lop.lop_urgent));
346 static void osc_page_delete(const struct lu_env *env,
347 const struct cl_page_slice *slice)
349 struct osc_page *opg = cl2osc_page(slice);
350 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
351 struct osc_async_page *oap = &opg->ops_oap;
354 LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
357 CDEBUG(D_TRACE, "%p\n", opg);
358 osc_page_transfer_put(env, opg);
359 rc = osc_teardown_async_page(osc_export(obj), NULL, obj->oo_oinfo, oap);
361 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
362 "Trying to teardown failed: %d\n", rc);
365 spin_lock(&obj->oo_seatbelt);
366 list_del_init(&opg->ops_inflight);
367 spin_unlock(&obj->oo_seatbelt);
371 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
374 struct osc_page *opg = cl2osc_page(slice);
375 struct osc_async_page *oap = &opg->ops_oap;
377 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
379 opg->ops_from = from;
381 spin_lock(&oap->oap_lock);
382 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
383 spin_unlock(&oap->oap_lock);
386 static int osc_page_cancel(const struct lu_env *env,
387 const struct cl_page_slice *slice)
389 struct osc_page *opg = cl2osc_page(slice);
390 struct osc_async_page *oap = &opg->ops_oap;
393 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
395 client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
396 /* Check if the transferring against this page
397 * is completed, or not even queued. */
398 if (opg->ops_transfer_pinned)
399 /* FIXME: may not be interrupted.. */
400 rc = osc_oap_interrupted(env, oap);
401 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
402 client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
406 static const struct cl_page_operations osc_page_ops = {
407 .cpo_fini = osc_page_fini,
408 .cpo_print = osc_page_print,
409 .cpo_delete = osc_page_delete,
410 .cpo_is_under_lock = osc_page_is_under_lock,
413 .cpo_cache_add = osc_page_fail
416 .cpo_cache_add = osc_page_cache_add
419 .cpo_clip = osc_page_clip,
420 .cpo_cancel = osc_page_cancel
423 static int osc_make_ready(const struct lu_env *env, void *data, int cmd)
425 struct osc_page *opg = data;
426 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
429 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
430 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 1));
433 result = cl_page_make_ready(env, page, CRT_WRITE);
435 opg->ops_submit_time = cfs_time_current();
439 static int osc_refresh_count(const struct lu_env *env, void *data, int cmd)
441 struct cl_page *page;
442 struct osc_page *osc = data;
443 struct cl_object *obj;
444 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
449 LINVRNT(osc_page_protected(env, osc, CLM_READ, 1));
451 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
452 LASSERT(!(cmd & OBD_BRW_READ));
453 LASSERT(osc != NULL);
454 page = osc->ops_cl.cpl_page;
455 obj = osc->ops_cl.cpl_obj;
457 cl_object_attr_lock(obj);
458 result = cl_object_attr_get(env, obj, attr);
459 cl_object_attr_unlock(obj);
463 if (cl_offset(obj, page->cp_index) >= kms)
464 /* catch race with truncate */
466 else if (cl_offset(obj, page->cp_index + 1) > kms)
467 /* catch sub-page write at end of file */
468 return kms % CFS_PAGE_SIZE;
470 return CFS_PAGE_SIZE;
473 static int osc_completion(const struct lu_env *env,
474 void *data, int cmd, struct obdo *oa, int rc)
476 struct osc_page *opg = data;
477 struct osc_async_page *oap = &opg->ops_oap;
478 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
479 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
480 enum cl_req_type crt;
482 LINVRNT(osc_page_protected(env, opg, CLM_READ, 1));
483 LINVRNT(cl_page_is_vmlocked(env, page));
487 cmd &= ~OBD_BRW_NOQUOTA;
488 LASSERT(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ));
489 LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
490 LASSERT(opg->ops_transfer_pinned);
493 * page->cp_req can be NULL if io submission failed before
494 * cl_req was allocated.
496 if (page->cp_req != NULL)
497 cl_req_page_done(env, page);
498 LASSERT(page->cp_req == NULL);
500 /* As the transfer for this page is being done, clear the flags */
501 spin_lock(&oap->oap_lock);
502 oap->oap_async_flags = 0;
503 spin_unlock(&oap->oap_lock);
505 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
506 /* Clear opg->ops_transfer_pinned before VM lock is released. */
507 opg->ops_transfer_pinned = 0;
509 spin_lock(&obj->oo_seatbelt);
510 LASSERT(opg->ops_submitter != NULL);
511 LASSERT(!list_empty(&opg->ops_inflight));
512 list_del_init(&opg->ops_inflight);
513 spin_unlock(&obj->oo_seatbelt);
515 opg->ops_submit_time = 0;
517 cl_page_completion(env, page, crt, rc);
520 if (rc == 0 && oap->oap_brw_flags & OBD_BRW_SRVLOCK) {
521 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
522 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
523 int bytes = opg->ops_to - opg->ops_from;
526 stats->os_lockless_reads += bytes;
528 stats->os_lockless_writes += bytes;
532 * This has to be the last operation with the page, as locks are
533 * released in cl_page_completion() and nothing except for the
534 * reference counter protects page from concurrent reclaim.
536 lu_ref_del(&page->cp_reference, "transfer", page);
538 * As page->cp_obj is pinned by a reference from page->cp_req, it is
539 * safe to call cl_page_put() without risking object destruction in a
540 * non-blocking context.
542 cl_page_put(env, page);
546 const static struct obd_async_page_ops osc_async_page_ops = {
547 .ap_make_ready = osc_make_ready,
548 .ap_refresh_count = osc_refresh_count,
549 .ap_completion = osc_completion
552 struct cl_page *osc_page_init(const struct lu_env *env,
553 struct cl_object *obj,
554 struct cl_page *page, cfs_page_t *vmpage)
556 struct osc_object *osc = cl2osc(obj);
557 struct osc_page *opg;
560 OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO);
562 void *oap = &opg->ops_oap;
565 opg->ops_to = CFS_PAGE_SIZE;
567 result = osc_prep_async_page(osc_export(osc),
568 NULL, osc->oo_oinfo, vmpage,
569 cl_offset(obj, page->cp_index),
571 opg, (void **)&oap, 1, NULL);
573 cl_page_slice_add(page, &opg->ops_cl, obj,
576 * Cannot assert osc_page_protected() here as read-ahead
577 * creates temporary pages outside of a lock.
579 #ifdef INVARIANT_CHECK
580 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
582 CFS_INIT_LIST_HEAD(&opg->ops_inflight);
585 return ERR_PTR(result);
589 * Helper function called by osc_io_submit() for every page in an immediate
590 * transfer (i.e., transferred synchronously).
592 void osc_io_submit_page(const struct lu_env *env,
593 struct osc_io *oio, struct osc_page *opg,
594 enum cl_req_type crt)
596 struct osc_async_page *oap = &opg->ops_oap;
597 struct client_obd *cli = oap->oap_cli;
600 LINVRNT(osc_page_protected(env, opg,
601 crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
603 oap->oap_page_off = opg->ops_from;
604 oap->oap_count = opg->ops_to - opg->ops_from;
605 /* Give a hint to OST that requests are coming from kswapd - bug19529 */
606 if (libcfs_memory_pressure_get())
607 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
608 oap->oap_brw_flags |= OBD_BRW_SYNC;
609 if (osc_io_srvlock(oio))
610 oap->oap_brw_flags |= OBD_BRW_SRVLOCK;
612 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
613 if (!client_is_remote(osc_export(cl2osc(opg->ops_cl.cpl_obj))) &&
614 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
615 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
616 oap->oap_cmd |= OBD_BRW_NOQUOTA;
619 if (oap->oap_cmd & OBD_BRW_READ)
620 flags = ASYNC_COUNT_STABLE;
621 else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT))
622 osc_enter_cache_try(env, cli, oap->oap_loi, oap, 1);
624 spin_lock(&oap->oap_lock);
625 oap->oap_async_flags |= OSC_FLAGS | flags;
626 spin_unlock(&oap->oap_lock);
628 osc_oap_to_pending(oap);
629 osc_page_transfer_get(opg, "transfer\0imm");
630 osc_page_transfer_add(env, opg, crt);