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"
47 static int osc_page_is_dlocked(const struct lu_env *env,
48 const struct osc_page *opg,
49 enum cl_lock_mode mode, int pending, int unref)
52 struct osc_object *obj;
53 struct osc_thread_info *info;
54 struct ldlm_res_id *resname;
55 struct lustre_handle *lockh;
56 ldlm_policy_data_t *policy;
60 info = osc_env_info(env);
61 resname = &info->oti_resname;
62 policy = &info->oti_policy;
63 lockh = &info->oti_handle;
64 page = opg->ops_cl.cpl_page;
65 obj = cl2osc(opg->ops_cl.cpl_obj);
67 flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
69 flags |= LDLM_FL_CBPENDING;
71 dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
72 osc_lock_build_res(env, obj, resname);
73 osc_index2policy(policy, page->cp_obj, page->cp_index, page->cp_index);
74 return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
75 dlmmode, &flags, NULL, lockh, unref);
79 * Checks an invariant that a page in the cache is covered by a lock, as
82 static int osc_page_protected(const struct lu_env *env,
83 const struct osc_page *opg,
84 enum cl_lock_mode mode, int unref)
86 struct cl_object_header *hdr;
89 struct cl_lock_descr *descr;
92 LINVRNT(!opg->ops_temp);
94 page = opg->ops_cl.cpl_page;
95 if (page->cp_owner != NULL &&
96 cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
98 * If IO is done without locks (liblustre, or lloop), lock is
103 /* otherwise check for a DLM lock */
104 result = osc_page_is_dlocked(env, opg, mode, 1, unref);
106 /* maybe this page is a part of a lockless io? */
107 hdr = cl_object_header(opg->ops_cl.cpl_obj);
108 descr = &osc_env_info(env)->oti_descr;
109 descr->cld_mode = mode;
110 descr->cld_start = page->cp_index;
111 descr->cld_end = page->cp_index;
112 spin_lock(&hdr->coh_lock_guard);
113 list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
115 * Lock-less sub-lock has to be either in HELD state
116 * (when io is actively going on), or in CACHED state,
117 * when top-lock is being unlocked:
118 * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
120 if ((scan->cll_state == CLS_HELD ||
121 scan->cll_state == CLS_CACHED) &&
122 cl_lock_ext_match(&scan->cll_descr, descr)) {
123 struct osc_lock *olck;
125 olck = osc_lock_at(scan);
126 result = osc_lock_is_lockless(olck);
130 spin_unlock(&hdr->coh_lock_guard);
135 /*****************************************************************************
140 static void osc_page_fini(const struct lu_env *env,
141 struct cl_page_slice *slice)
143 struct osc_page *opg = cl2osc_page(slice);
144 CDEBUG(D_TRACE, "%p\n", opg);
145 OBD_SLAB_FREE_PTR(opg, osc_page_kmem);
148 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
150 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
152 LASSERT(!opg->ops_transfer_pinned);
154 lu_ref_add_atomic(&page->cp_reference, label, page);
155 opg->ops_transfer_pinned = 1;
158 static void osc_page_transfer_put(const struct lu_env *env,
159 struct osc_page *opg)
161 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
163 if (opg->ops_transfer_pinned) {
164 lu_ref_del(&page->cp_reference, "transfer", page);
165 opg->ops_transfer_pinned = 0;
166 cl_page_put(env, page);
171 * This is called once for every page when it is submitted for a transfer
172 * either opportunistic (osc_page_cache_add()), or immediate
173 * (osc_page_submit()).
175 static void osc_page_transfer_add(const struct lu_env *env,
176 struct osc_page *opg, enum cl_req_type crt)
178 struct osc_object *obj;
180 LINVRNT(cl_page_is_vmlocked(env, opg->ops_cl.cpl_page));
182 obj = cl2osc(opg->ops_cl.cpl_obj);
183 spin_lock(&obj->oo_seatbelt);
184 list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
185 opg->ops_submitter = cfs_current();
186 spin_unlock(&obj->oo_seatbelt);
189 static int osc_page_cache_add(const struct lu_env *env,
190 const struct cl_page_slice *slice,
193 struct osc_page *opg = cl2osc_page(slice);
194 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
195 struct osc_io *oio = osc_env_io(env);
200 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
203 /* Set the OBD_BRW_SRVLOCK before the page is queued. */
204 brw_flags = osc_io_srvlock(oio) ? OBD_BRW_SRVLOCK : 0;
205 if (!client_is_remote(osc_export(obj)) &&
206 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
207 brw_flags |= OBD_BRW_NOQUOTA;
208 noquota = OBD_BRW_NOQUOTA;
211 osc_page_transfer_get(opg, "transfer\0cache");
212 result = osc_queue_async_io(env, osc_export(obj), NULL, obj->oo_oinfo,
213 &opg->ops_oap, OBD_BRW_WRITE | noquota,
216 osc_page_transfer_put(env, opg);
218 osc_page_transfer_add(env, opg, CRT_WRITE);
222 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
223 pgoff_t start, pgoff_t end)
225 memset(policy, 0, sizeof *policy);
226 policy->l_extent.start = cl_offset(obj, start);
227 policy->l_extent.end = cl_offset(obj, end + 1) - 1;
230 static int osc_page_is_under_lock(const struct lu_env *env,
231 const struct cl_page_slice *slice,
234 struct cl_lock *lock;
238 lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
241 cl_lock_put(env, lock);
248 static int osc_page_fail(const struct lu_env *env,
249 const struct cl_page_slice *slice, struct cl_io *_)
259 static const char *osc_list(struct list_head *head)
261 return list_empty(head) ? "-" : "+";
264 static int osc_page_print(const struct lu_env *env,
265 const struct cl_page_slice *slice,
266 void *cookie, lu_printer_t printer)
268 struct osc_page *opg = cl2osc_page(slice);
269 struct osc_async_page *oap = &opg->ops_oap;
271 return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
272 "< %#x %d %u %s %s %s >"
273 "< %llu %u %#x %#x %p %p %p %p %p >"
277 oap->oap_magic, oap->oap_cmd,
278 oap->oap_interrupted,
279 osc_list(&oap->oap_pending_item),
280 osc_list(&oap->oap_urgent_item),
281 osc_list(&oap->oap_rpc_item),
283 oap->oap_obj_off, oap->oap_page_off,
284 oap->oap_async_flags, oap->oap_brw_flags,
286 oap->oap_cli, oap->oap_loi, oap->oap_caller_ops,
287 oap->oap_caller_data,
289 osc_list(&opg->ops_inflight),
290 opg->ops_submitter, opg->ops_transfer_pinned);
293 static void osc_page_delete(const struct lu_env *env,
294 const struct cl_page_slice *slice)
296 struct osc_page *opg = cl2osc_page(slice);
297 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
298 struct osc_async_page *oap = &opg->ops_oap;
301 LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
304 CDEBUG(D_TRACE, "%p\n", opg);
305 osc_page_transfer_put(env, opg);
306 rc = osc_teardown_async_page(osc_export(obj), NULL, obj->oo_oinfo, oap);
308 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
309 "Trying to teardown failed: %d\n", rc);
312 spin_lock(&obj->oo_seatbelt);
313 list_del_init(&opg->ops_inflight);
314 spin_unlock(&obj->oo_seatbelt);
318 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
321 struct osc_page *opg = cl2osc_page(slice);
322 struct osc_async_page *oap = &opg->ops_oap;
324 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
326 opg->ops_from = from;
328 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
331 static int osc_page_cancel(const struct lu_env *env,
332 const struct cl_page_slice *slice)
334 struct osc_page *opg = cl2osc_page(slice);
335 struct osc_async_page *oap = &opg->ops_oap;
338 LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
340 client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
341 /* Check if the transferring against this page
342 * is completed, or not even queued. */
343 if (opg->ops_transfer_pinned)
344 /* FIXME: may not be interrupted.. */
345 rc = osc_oap_interrupted(env, oap);
346 LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
347 client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
351 static const struct cl_page_operations osc_page_ops = {
352 .cpo_fini = osc_page_fini,
353 .cpo_print = osc_page_print,
354 .cpo_delete = osc_page_delete,
355 .cpo_is_under_lock = osc_page_is_under_lock,
358 .cpo_cache_add = osc_page_fail
361 .cpo_cache_add = osc_page_cache_add
364 .cpo_clip = osc_page_clip,
365 .cpo_cancel = osc_page_cancel
368 static int osc_make_ready(const struct lu_env *env, void *data, int cmd)
370 struct osc_page *opg = data;
371 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
374 LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
375 LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 1));
378 result = cl_page_make_ready(env, page, CRT_WRITE);
382 static int osc_refresh_count(const struct lu_env *env, void *data, int cmd)
384 struct cl_page *page;
385 struct osc_page *osc = data;
386 struct cl_object *obj;
387 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
392 LINVRNT(osc_page_protected(env, osc, CLM_READ, 1));
394 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
395 LASSERT(!(cmd & OBD_BRW_READ));
396 LASSERT(osc != NULL);
397 page = osc->ops_cl.cpl_page;
398 obj = osc->ops_cl.cpl_obj;
400 cl_object_attr_lock(obj);
401 result = cl_object_attr_get(env, obj, attr);
402 cl_object_attr_unlock(obj);
406 if (cl_offset(obj, page->cp_index) >= kms)
407 /* catch race with truncate */
409 else if (cl_offset(obj, page->cp_index + 1) > kms)
410 /* catch sub-page write at end of file */
411 return kms % CFS_PAGE_SIZE;
413 return CFS_PAGE_SIZE;
416 static int osc_completion(const struct lu_env *env,
417 void *data, int cmd, struct obdo *oa, int rc)
419 struct osc_page *opg = data;
420 struct osc_async_page *oap = &opg->ops_oap;
421 struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
422 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
423 enum cl_req_type crt;
425 LINVRNT(osc_page_protected(env, opg, CLM_READ, 1));
426 LINVRNT(cl_page_is_vmlocked(env, page));
430 cmd &= ~OBD_BRW_NOQUOTA;
431 LASSERT(equi(page->cp_state == CPS_PAGEIN, cmd == OBD_BRW_READ));
432 LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
433 LASSERT(opg->ops_transfer_pinned);
436 * page->cp_req can be NULL if io submission failed before
437 * cl_req was allocated.
439 if (page->cp_req != NULL)
440 cl_req_page_done(env, page);
441 LASSERT(page->cp_req == NULL);
443 /* As the transfer for this page is being done, clear the flags */
444 oap->oap_async_flags = 0;
446 crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
447 /* Clear opg->ops_transfer_pinned before VM lock is released. */
448 opg->ops_transfer_pinned = 0;
450 spin_lock(&obj->oo_seatbelt);
451 LASSERT(opg->ops_submitter != NULL);
452 LASSERT(!list_empty(&opg->ops_inflight));
453 list_del_init(&opg->ops_inflight);
454 spin_unlock(&obj->oo_seatbelt);
456 cl_page_completion(env, page, crt, rc);
459 if (rc == 0 && oap->oap_brw_flags & OBD_BRW_SRVLOCK) {
460 struct lu_device *ld = opg->ops_cl.cpl_obj->co_lu.lo_dev;
461 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
462 int bytes = opg->ops_to - opg->ops_from;
465 stats->os_lockless_reads += bytes;
467 stats->os_lockless_writes += bytes;
471 * This has to be the last operation with the page, as locks are
472 * released in cl_page_completion() and nothing except for the
473 * reference counter protects page from concurrent reclaim.
475 lu_ref_del(&page->cp_reference, "transfer", page);
477 * As page->cp_obj is pinned by a reference from page->cp_req, it is
478 * safe to call cl_page_put() without risking object destruction in a
479 * non-blocking context.
481 cl_page_put(env, page);
485 const static struct obd_async_page_ops osc_async_page_ops = {
486 .ap_make_ready = osc_make_ready,
487 .ap_refresh_count = osc_refresh_count,
488 .ap_completion = osc_completion
491 struct cl_page *osc_page_init(const struct lu_env *env,
492 struct cl_object *obj,
493 struct cl_page *page, cfs_page_t *vmpage)
495 struct osc_object *osc = cl2osc(obj);
496 struct osc_page *opg;
499 OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO);
501 void *oap = &opg->ops_oap;
504 opg->ops_to = CFS_PAGE_SIZE;
506 result = osc_prep_async_page(osc_export(osc),
507 NULL, osc->oo_oinfo, vmpage,
508 cl_offset(obj, page->cp_index),
510 opg, (void **)&oap, 1, NULL);
512 cl_page_slice_add(page, &opg->ops_cl, obj,
515 * Cannot assert osc_page_protected() here as read-ahead
516 * creates temporary pages outside of a lock.
518 #ifdef INVARIANT_CHECK
519 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
521 CFS_INIT_LIST_HEAD(&opg->ops_inflight);
524 return ERR_PTR(result);
528 * Helper function called by osc_io_submit() for every page in an immediate
529 * transfer (i.e., transferred synchronously).
531 void osc_io_submit_page(const struct lu_env *env,
532 struct osc_io *oio, struct osc_page *opg,
533 enum cl_req_type crt)
535 struct osc_async_page *oap = &opg->ops_oap;
536 struct client_obd *cli = oap->oap_cli;
538 LINVRNT(osc_page_protected(env, opg,
539 crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
541 oap->oap_page_off = opg->ops_from;
542 oap->oap_count = opg->ops_to - opg->ops_from;
543 oap->oap_brw_flags |= OBD_BRW_SYNC;
544 if (osc_io_srvlock(oio))
545 oap->oap_brw_flags |= OBD_BRW_SRVLOCK;
547 oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
548 if (!client_is_remote(osc_export(cl2osc(opg->ops_cl.cpl_obj))) &&
549 cfs_capable(CFS_CAP_SYS_RESOURCE)) {
550 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
551 oap->oap_cmd |= OBD_BRW_NOQUOTA;
554 oap->oap_async_flags |= OSC_FLAGS;
555 if (oap->oap_cmd & OBD_BRW_READ)
556 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
557 else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT))
558 osc_enter_cache_try(env, cli, oap->oap_loi, oap, 1);
560 osc_oap_to_pending(oap);
561 osc_page_transfer_get(opg, "transfer\0imm");
562 osc_page_transfer_add(env, opg, crt);