4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, Whamcloud, Inc.
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 obj = cl2osc(opg->ops_cl.cpl_obj);
202 cfs_spin_lock(&obj->oo_seatbelt);
203 cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
204 opg->ops_submitter = cfs_current();
205 cfs_spin_unlock(&obj->oo_seatbelt);
208 static int osc_page_cache_add(const struct lu_env *env,
209 const struct cl_page_slice *slice,
210 struct cl_io *unused)
212 struct osc_page *opg = cl2osc_page(slice);
216 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
218 osc_page_transfer_get(opg, "transfer\0cache");
219 result = osc_queue_async_io(env, opg);
221 osc_page_transfer_put(env, opg);
223 osc_page_transfer_add(env, opg, CRT_WRITE);
227 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
228 pgoff_t start, pgoff_t end)
230 memset(policy, 0, sizeof *policy);
231 policy->l_extent.start = cl_offset(obj, start);
232 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
235 static int osc_page_addref_lock(const struct lu_env *env,
236 struct osc_page *opg,
237 struct cl_lock *lock)
239 struct osc_lock *olock;
242 LASSERT(opg->ops_lock == NULL);
244 olock = osc_lock_at(lock);
245 if (cfs_atomic_inc_return(&olock->ols_pageref) <= 0) {
246 cfs_atomic_dec(&olock->ols_pageref);
247 cl_lock_put(env, lock);
250 opg->ops_lock = lock;
256 static void osc_page_putref_lock(const struct lu_env *env,
257 struct osc_page *opg)
259 struct cl_lock *lock = opg->ops_lock;
260 struct osc_lock *olock;
262 LASSERT(lock != NULL);
263 olock = osc_lock_at(lock);
265 cfs_atomic_dec(&olock->ols_pageref);
266 opg->ops_lock = NULL;
268 cl_lock_put(env, lock);
271 static int osc_page_is_under_lock(const struct lu_env *env,
272 const struct cl_page_slice *slice,
273 struct cl_io *unused)
275 struct cl_lock *lock;
279 lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
282 osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
289 static void osc_page_disown(const struct lu_env *env,
290 const struct cl_page_slice *slice,
293 struct osc_page *opg = cl2osc_page(slice);
295 if (unlikely(opg->ops_lock))
296 osc_page_putref_lock(env, opg);
299 static void osc_page_completion_read(const struct lu_env *env,
300 const struct cl_page_slice *slice,
303 struct osc_page *opg = cl2osc_page(slice);
305 if (likely(opg->ops_lock))
306 osc_page_putref_lock(env, opg);
309 static int osc_page_fail(const struct lu_env *env,
310 const struct cl_page_slice *slice,
311 struct cl_io *unused)
321 static const char *osc_list(cfs_list_t *head)
323 return cfs_list_empty(head) ? "-" : "+";
326 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
328 if (opg->ops_submit_time == 0)
331 return (cfs_time_current() - opg->ops_submit_time);
334 static int osc_page_print(const struct lu_env *env,
335 const struct cl_page_slice *slice,
336 void *cookie, lu_printer_t printer)
338 struct osc_page *opg = cl2osc_page(slice);
339 struct osc_async_page *oap = &opg->ops_oap;
340 struct osc_object *obj = cl2osc(slice->cpl_obj);
341 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
343 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
344 "1< %#x %d %u %s %s %s > "
345 "2< "LPU64" %u %u %#x %#x | %p %p %p > "
346 "3< %s %p %d %lu %d > "
347 "4< %d %d %d %lu %s | %s %s %s %s > "
348 "5< %s %s %s %s | %d %s %s | %d %s %s>\n",
351 oap->oap_magic, oap->oap_cmd,
352 oap->oap_interrupted,
353 osc_list(&oap->oap_pending_item),
354 osc_list(&oap->oap_urgent_item),
355 osc_list(&oap->oap_rpc_item),
357 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
358 oap->oap_async_flags, oap->oap_brw_flags,
359 oap->oap_request, oap->oap_cli, obj,
361 osc_list(&opg->ops_inflight),
362 opg->ops_submitter, opg->ops_transfer_pinned,
363 osc_submit_duration(opg), opg->ops_srvlock,
365 cli->cl_r_in_flight, cli->cl_w_in_flight,
366 cli->cl_max_rpcs_in_flight,
368 osc_list(&cli->cl_cache_waiters),
369 osc_list(&cli->cl_loi_ready_list),
370 osc_list(&cli->cl_loi_hp_ready_list),
371 osc_list(&cli->cl_loi_write_list),
372 osc_list(&cli->cl_loi_read_list),
374 osc_list(&obj->oo_ready_item),
375 osc_list(&obj->oo_hp_ready_item),
376 osc_list(&obj->oo_write_item),
377 osc_list(&obj->oo_read_item),
378 obj->oo_read_pages.oop_num_pending,
379 osc_list(&obj->oo_read_pages.oop_pending),
380 osc_list(&obj->oo_read_pages.oop_urgent),
381 obj->oo_write_pages.oop_num_pending,
382 osc_list(&obj->oo_write_pages.oop_pending),
383 osc_list(&obj->oo_write_pages.oop_urgent));
386 static void osc_page_delete(const struct lu_env *env,
387 const struct cl_page_slice *slice)
389 struct osc_page *opg = cl2osc_page(slice);
390 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
393 LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
396 CDEBUG(D_TRACE, "%p\n", opg);
397 osc_page_transfer_put(env, opg);
398 rc = osc_teardown_async_page(obj, opg);
400 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
401 "Trying to teardown failed: %d\n", rc);
404 cfs_spin_lock(&obj->oo_seatbelt);
405 cfs_list_del_init(&opg->ops_inflight);
406 cfs_spin_unlock(&obj->oo_seatbelt);
410 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
413 struct osc_page *opg = cl2osc_page(slice);
414 struct osc_async_page *oap = &opg->ops_oap;
416 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
418 opg->ops_from = from;
420 cfs_spin_lock(&oap->oap_lock);
421 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
422 cfs_spin_unlock(&oap->oap_lock);
425 static int osc_page_cancel(const struct lu_env *env,
426 const struct cl_page_slice *slice)
428 struct osc_page *opg = cl2osc_page(slice);
429 struct osc_async_page *oap = &opg->ops_oap;
432 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
434 client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
435 /* Check if the transferring against this page
436 * is completed, or not even queued. */
437 if (opg->ops_transfer_pinned)
438 /* FIXME: may not be interrupted.. */
439 rc = osc_cancel_async_page(env, opg);
440 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
441 client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
445 static const struct cl_page_operations osc_page_ops = {
446 .cpo_fini = osc_page_fini,
447 .cpo_print = osc_page_print,
448 .cpo_delete = osc_page_delete,
449 .cpo_is_under_lock = osc_page_is_under_lock,
450 .cpo_disown = osc_page_disown,
453 .cpo_cache_add = osc_page_fail,
454 .cpo_completion = osc_page_completion_read
457 .cpo_cache_add = osc_page_cache_add
460 .cpo_clip = osc_page_clip,
461 .cpo_cancel = osc_page_cancel
464 struct cl_page *osc_page_init(const struct lu_env *env,
465 struct cl_object *obj,
466 struct cl_page *page, cfs_page_t *vmpage)
468 struct osc_object *osc = cl2osc(obj);
469 struct osc_page *opg;
472 OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO);
475 opg->ops_to = CFS_PAGE_SIZE;
477 result = osc_prep_async_page(osc, opg, vmpage,
478 cl_offset(obj, page->cp_index));
480 struct osc_io *oio = osc_env_io(env);
481 opg->ops_srvlock = osc_io_srvlock(oio);
482 cl_page_slice_add(page, &opg->ops_cl, obj,
486 * Cannot assert osc_page_protected() here as read-ahead
487 * creates temporary pages outside of a lock.
489 #ifdef INVARIANT_CHECK
490 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
492 CFS_INIT_LIST_HEAD(&opg->ops_inflight);
495 return ERR_PTR(result);
499 * Helper function called by osc_io_submit() for every page in an immediate
500 * transfer (i.e., transferred synchronously).
502 void osc_io_submit_page(const struct lu_env *env,
503 struct osc_io *oio, struct osc_page *opg,
504 enum cl_req_type crt)
506 LINVRNT(osc_page_protected(env, opg,
507 crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
509 osc_queue_sync_page(env, opg,
510 crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
511 osc_io_srvlock(oio) ? OBD_BRW_SRVLOCK : 0);
513 osc_page_transfer_get(opg, "transfer\0imm");
514 osc_page_transfer_add(env, opg, crt);