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,
212 struct osc_io *oio = osc_env_io(env);
213 struct osc_page *opg = cl2osc_page(slice);
217 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
219 osc_page_transfer_get(opg, "transfer\0cache");
220 result = osc_queue_async_io(env, io, opg);
222 osc_page_transfer_put(env, opg);
224 osc_page_transfer_add(env, opg, CRT_WRITE);
226 /* for sync write, kernel will wait for this page to be flushed before
227 * osc_io_end() is called, so release it earlier.
228 * for mkwrite(), it's known there is no further pages. */
229 if (cl_io_is_sync_write(io) || cl_io_is_mkwrite(io)) {
230 if (oio->oi_active != NULL) {
231 osc_extent_release(env, oio->oi_active);
232 oio->oi_active = NULL;
239 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
240 pgoff_t start, pgoff_t end)
242 memset(policy, 0, sizeof *policy);
243 policy->l_extent.start = cl_offset(obj, start);
244 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
247 static int osc_page_addref_lock(const struct lu_env *env,
248 struct osc_page *opg,
249 struct cl_lock *lock)
251 struct osc_lock *olock;
254 LASSERT(opg->ops_lock == NULL);
256 olock = osc_lock_at(lock);
257 if (cfs_atomic_inc_return(&olock->ols_pageref) <= 0) {
258 cfs_atomic_dec(&olock->ols_pageref);
259 cl_lock_put(env, lock);
262 opg->ops_lock = lock;
268 static void osc_page_putref_lock(const struct lu_env *env,
269 struct osc_page *opg)
271 struct cl_lock *lock = opg->ops_lock;
272 struct osc_lock *olock;
274 LASSERT(lock != NULL);
275 olock = osc_lock_at(lock);
277 cfs_atomic_dec(&olock->ols_pageref);
278 opg->ops_lock = NULL;
280 cl_lock_put(env, lock);
283 static int osc_page_is_under_lock(const struct lu_env *env,
284 const struct cl_page_slice *slice,
285 struct cl_io *unused)
287 struct cl_lock *lock;
291 lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
294 osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
301 static void osc_page_disown(const struct lu_env *env,
302 const struct cl_page_slice *slice,
305 struct osc_page *opg = cl2osc_page(slice);
307 if (unlikely(opg->ops_lock))
308 osc_page_putref_lock(env, opg);
311 static void osc_page_completion_read(const struct lu_env *env,
312 const struct cl_page_slice *slice,
315 struct osc_page *opg = cl2osc_page(slice);
317 if (likely(opg->ops_lock))
318 osc_page_putref_lock(env, opg);
321 static int osc_page_fail(const struct lu_env *env,
322 const struct cl_page_slice *slice,
323 struct cl_io *unused)
333 static const char *osc_list(cfs_list_t *head)
335 return cfs_list_empty(head) ? "-" : "+";
338 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
340 if (opg->ops_submit_time == 0)
343 return (cfs_time_current() - opg->ops_submit_time);
346 static int osc_page_print(const struct lu_env *env,
347 const struct cl_page_slice *slice,
348 void *cookie, lu_printer_t printer)
350 struct osc_page *opg = cl2osc_page(slice);
351 struct osc_async_page *oap = &opg->ops_oap;
352 struct osc_object *obj = cl2osc(slice->cpl_obj);
353 struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
355 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
356 "1< %#x %d %u %s %s > "
357 "2< "LPU64" %u %u %#x %#x | %p %p %p > "
358 "3< %s %p %d %lu %d > "
359 "4< %d %d %d %lu %s | %s %s %s %s > "
360 "5< %s %s %s %s | %d %s | %d %s %s>\n",
363 oap->oap_magic, oap->oap_cmd,
364 oap->oap_interrupted,
365 osc_list(&oap->oap_pending_item),
366 osc_list(&oap->oap_rpc_item),
368 oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
369 oap->oap_async_flags, oap->oap_brw_flags,
370 oap->oap_request, oap->oap_cli, obj,
372 osc_list(&opg->ops_inflight),
373 opg->ops_submitter, opg->ops_transfer_pinned,
374 osc_submit_duration(opg), opg->ops_srvlock,
376 cli->cl_r_in_flight, cli->cl_w_in_flight,
377 cli->cl_max_rpcs_in_flight,
379 osc_list(&cli->cl_cache_waiters),
380 osc_list(&cli->cl_loi_ready_list),
381 osc_list(&cli->cl_loi_hp_ready_list),
382 osc_list(&cli->cl_loi_write_list),
383 osc_list(&cli->cl_loi_read_list),
385 osc_list(&obj->oo_ready_item),
386 osc_list(&obj->oo_hp_ready_item),
387 osc_list(&obj->oo_write_item),
388 osc_list(&obj->oo_read_item),
389 cfs_atomic_read(&obj->oo_nr_reads),
390 osc_list(&obj->oo_reading_exts),
391 cfs_atomic_read(&obj->oo_nr_writes),
392 osc_list(&obj->oo_hp_exts),
393 osc_list(&obj->oo_urgent_exts));
396 static void osc_page_delete(const struct lu_env *env,
397 const struct cl_page_slice *slice)
399 struct osc_page *opg = cl2osc_page(slice);
400 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
403 LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
406 CDEBUG(D_TRACE, "%p\n", opg);
407 osc_page_transfer_put(env, opg);
408 rc = osc_teardown_async_page(env, obj, opg);
410 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
411 "Trying to teardown failed: %d\n", rc);
414 cfs_spin_lock(&obj->oo_seatbelt);
415 cfs_list_del_init(&opg->ops_inflight);
416 cfs_spin_unlock(&obj->oo_seatbelt);
420 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
423 struct osc_page *opg = cl2osc_page(slice);
424 struct osc_async_page *oap = &opg->ops_oap;
426 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
428 opg->ops_from = from;
430 cfs_spin_lock(&oap->oap_lock);
431 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
432 cfs_spin_unlock(&oap->oap_lock);
435 static int osc_page_cancel(const struct lu_env *env,
436 const struct cl_page_slice *slice)
438 struct osc_page *opg = cl2osc_page(slice);
441 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
443 /* Check if the transferring against this page
444 * is completed, or not even queued. */
445 if (opg->ops_transfer_pinned)
446 /* FIXME: may not be interrupted.. */
447 rc = osc_cancel_async_page(env, opg);
448 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
452 static int osc_page_flush(const struct lu_env *env,
453 const struct cl_page_slice *slice,
456 struct osc_page *opg = cl2osc_page(slice);
459 rc = osc_flush_async_page(env, io, opg);
463 static const struct cl_page_operations osc_page_ops = {
464 .cpo_fini = osc_page_fini,
465 .cpo_print = osc_page_print,
466 .cpo_delete = osc_page_delete,
467 .cpo_is_under_lock = osc_page_is_under_lock,
468 .cpo_disown = osc_page_disown,
471 .cpo_cache_add = osc_page_fail,
472 .cpo_completion = osc_page_completion_read
475 .cpo_cache_add = osc_page_cache_add
478 .cpo_clip = osc_page_clip,
479 .cpo_cancel = osc_page_cancel,
480 .cpo_flush = osc_page_flush
483 struct cl_page *osc_page_init(const struct lu_env *env,
484 struct cl_object *obj,
485 struct cl_page *page, cfs_page_t *vmpage)
487 struct osc_object *osc = cl2osc(obj);
488 struct osc_page *opg;
491 OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO);
494 opg->ops_to = CFS_PAGE_SIZE;
496 result = osc_prep_async_page(osc, opg, vmpage,
497 cl_offset(obj, page->cp_index));
499 struct osc_io *oio = osc_env_io(env);
500 opg->ops_srvlock = osc_io_srvlock(oio);
501 cl_page_slice_add(page, &opg->ops_cl, obj,
505 * Cannot assert osc_page_protected() here as read-ahead
506 * creates temporary pages outside of a lock.
508 #ifdef INVARIANT_CHECK
509 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
511 CFS_INIT_LIST_HEAD(&opg->ops_inflight);
514 return ERR_PTR(result);
518 * Helper function called by osc_io_submit() for every page in an immediate
519 * transfer (i.e., transferred synchronously).
521 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
522 enum cl_req_type crt, int brw_flags)
524 struct osc_async_page *oap = &opg->ops_oap;
525 struct osc_object *obj = oap->oap_obj;
527 LINVRNT(osc_page_protected(env, opg,
528 crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
530 LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
531 "magic 0x%x\n", oap, oap->oap_magic);
532 LASSERT(oap->oap_async_flags & ASYNC_READY);
533 LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
535 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
536 oap->oap_page_off = opg->ops_from;
537 oap->oap_count = opg->ops_to - opg->ops_from;
538 oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
540 if (!client_is_remote(osc_export(obj)) &&
541 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
542 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
543 oap->oap_cmd |= OBD_BRW_NOQUOTA;
546 opg->ops_submit_time = cfs_time_current();
547 osc_page_transfer_get(opg, "transfer\0imm");
548 osc_page_transfer_add(env, opg, crt);