lustre/osc/osc_page.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   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.
  11  *
  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).
  17  *
  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
  21  *
  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
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_page for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 /** \addtogroup osc osc @{ */
  42
  43 #define DEBUG_SUBSYSTEM S_OSC
  44
  45 #include "osc_cl_internal.h"
  46
  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)
  50 {
  51         struct cl_page         *page;
  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;
  57         ldlm_mode_t             dlmmode;
  58         int                     flags;
  59
  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);
  66
  67         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  68         if (pending)
  69                 flags |= LDLM_FL_CBPENDING;
  70
  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);
  76 }
  77
  78 /**
  79  * Checks an invariant that a page in the cache is covered by a lock, as
  80  * needed.
  81  */
  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)
  85 {
  86         struct cl_object_header *hdr;
  87         struct cl_lock          *scan;
  88         struct cl_page          *page;
  89         struct cl_lock_descr    *descr;
  90         int result;
  91
  92         LINVRNT(!opg->ops_temp);
  93
  94         page = opg->ops_cl.cpl_page;
  95         if (page->cp_owner != NULL &&
  96             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
  97                 /*
  98                  * If IO is done without locks (liblustre, or lloop), lock is
  99                  * not required.
 100                  */
 101                 result = 1;
 102         else
 103                 /* otherwise check for a DLM lock */
 104         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 105         if (result == 0) {
 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) {
 114                         /*
 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().
 119                          */
 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;
 124
 125                                 olck = osc_lock_at(scan);
 126                                 result = osc_lock_is_lockless(olck);
 127                                 break;
 128                         }
 129                 }
 130                 spin_unlock(&hdr->coh_lock_guard);
 131         }
 132         return result;
 133 }
 134
 135 /*****************************************************************************
 136  *
 137  * Page operations.
 138  *
 139  */
 140 static void osc_page_fini(const struct lu_env *env,
 141                           struct cl_page_slice *slice)
 142 {
 143         struct osc_page *opg = cl2osc_page(slice);
 144         CDEBUG(D_TRACE, "%p\n", opg);
 145         OBD_SLAB_FREE_PTR(opg, osc_page_kmem);
 146 }
 147
 148 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 149 {
 150         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 151
 152         LASSERT(!opg->ops_transfer_pinned);
 153         cl_page_get(page);
 154         lu_ref_add_atomic(&page->cp_reference, label, page);
 155         opg->ops_transfer_pinned = 1;
 156 }
 157
 158 static void osc_page_transfer_put(const struct lu_env *env,
 159                                   struct osc_page *opg)
 160 {
 161         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 162
 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);
 167         }
 168 }
 169
 170 /**
 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()).
 174  */
 175 static void osc_page_transfer_add(const struct lu_env *env,
 176                                   struct osc_page *opg, enum cl_req_type crt)
 177 {
 178         struct osc_object *obj;
 179
 180         LINVRNT(cl_page_is_vmlocked(env, opg->ops_cl.cpl_page));
 181
 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);
 187 }
 188
 189 static int osc_page_cache_add(const struct lu_env *env,
 190                               const struct cl_page_slice *slice,
 191                               struct cl_io *_)
 192 {
 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);
 196         int result;
 197         int brw_flags;
 198         int noquota = 0;
 199
 200         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 201         ENTRY;
 202
 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;
 209         }
 210
 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,
 214                                     0, 0, brw_flags, 0);
 215         if (result != 0)
 216                 osc_page_transfer_put(env, opg);
 217         else
 218                 osc_page_transfer_add(env, opg, CRT_WRITE);
 219         RETURN(result);
 220 }
 221
 222 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 223                       pgoff_t start, pgoff_t end)
 224 {
 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;
 228 }
 229
 230 static int osc_page_is_under_lock(const struct lu_env *env,
 231                                   const struct cl_page_slice *slice,
 232                                   struct cl_io *_)
 233 {
 234         struct cl_lock *lock;
 235         int             result;
 236
 237         ENTRY;
 238         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 239                                NULL, 1, 0);
 240         if (lock != NULL) {
 241                 cl_lock_put(env, lock);
 242                 result = -EBUSY;
 243         } else
 244                 result = -ENODATA;
 245         RETURN(result);
 246 }
 247
 248 static int osc_page_fail(const struct lu_env *env,
 249                          const struct cl_page_slice *slice, struct cl_io *_)
 250 {
 251         /*
 252          * Cached read?
 253          */
 254         LBUG();
 255         return 0;
 256 }
 257
 258
 259 static const char *osc_list(struct list_head *head)
 260 {
 261         return list_empty(head) ? "-" : "+";
 262 }
 263
 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)
 267 {
 268         struct osc_page       *opg = cl2osc_page(slice);
 269         struct osc_async_page *oap = &opg->ops_oap;
 270
 271         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
 272                           "%#x %d %u %s %s %s %llu %u %#x %p %p %p %p %p\n",
 273                           opg, oap->oap_magic, oap->oap_cmd,
 274                           oap->oap_interrupted,
 275                           osc_list(&oap->oap_pending_item),
 276                           osc_list(&oap->oap_urgent_item),
 277                           osc_list(&oap->oap_rpc_item),
 278                           oap->oap_obj_off, oap->oap_page_off,
 279                           oap->oap_async_flags, oap->oap_request,
 280                           oap->oap_cli, oap->oap_loi, oap->oap_caller_ops,
 281                           oap->oap_caller_data);
 282 }
 283
 284 static void osc_page_delete(const struct lu_env *env,
 285                             const struct cl_page_slice *slice)
 286 {
 287         struct osc_page       *opg = cl2osc_page(slice);
 288         struct osc_object     *obj = cl2osc(opg->ops_cl.cpl_obj);
 289         struct osc_async_page *oap = &opg->ops_oap;
 290         int rc;
 291
 292         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 293
 294         ENTRY;
 295         CDEBUG(D_TRACE, "%p\n", opg);
 296         osc_page_transfer_put(env, opg);
 297         rc = osc_teardown_async_page(osc_export(obj), NULL, obj->oo_oinfo, oap);
 298         LASSERTF(rc == 0, "%i\n", rc);
 299         spin_lock(&obj->oo_seatbelt);
 300         list_del_init(&opg->ops_inflight);
 301         spin_unlock(&obj->oo_seatbelt);
 302         EXIT;
 303 }
 304
 305 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 306                    int from, int to)
 307 {
 308         struct osc_page       *opg = cl2osc_page(slice);
 309         struct osc_async_page *oap = &opg->ops_oap;
 310
 311         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 312
 313         opg->ops_from = from;
 314         opg->ops_to   = to;
 315         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 316 }
 317
 318 static int osc_page_cancel(const struct lu_env *env,
 319                            const struct cl_page_slice *slice)
 320 {
 321         struct osc_page *opg       = cl2osc_page(slice);
 322         struct osc_async_page *oap = &opg->ops_oap;
 323         int rc = 0;
 324
 325         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 326
 327         client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
 328         /* Check if the transferring against this page
 329          * is completed, or not even queued. */
 330         if (opg->ops_transfer_pinned)
 331                 /* FIXME: may not be interrupted.. */
 332                 rc = osc_oap_interrupted(env, oap);
 333         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 334         client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
 335         return rc;
 336 }
 337
 338 static const struct cl_page_operations osc_page_ops = {
 339         .cpo_fini          = osc_page_fini,
 340         .cpo_print         = osc_page_print,
 341         .cpo_delete        = osc_page_delete,
 342         .cpo_is_under_lock = osc_page_is_under_lock,
 343         .io = {
 344                 [CRT_READ] = {
 345                         .cpo_cache_add = osc_page_fail
 346                 },
 347                 [CRT_WRITE] = {
 348                         .cpo_cache_add = osc_page_cache_add
 349                 }
 350         },
 351         .cpo_clip           = osc_page_clip,
 352         .cpo_cancel         = osc_page_cancel
 353 };
 354
 355 static int osc_make_ready(const struct lu_env *env, void *data, int cmd)
 356 {
 357         struct osc_page *opg  = data;
 358         struct cl_page  *page = cl_page_top(opg->ops_cl.cpl_page);
 359         int result;
 360
 361         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
 362         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 1));
 363
 364         ENTRY;
 365         result = cl_page_make_ready(env, page, CRT_WRITE);
 366         RETURN(result);
 367 }
 368
 369 static int osc_refresh_count(const struct lu_env *env, void *data, int cmd)
 370 {
 371         struct cl_page   *page;
 372         struct osc_page  *osc = data;
 373         struct cl_object *obj;
 374         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
 375
 376         int result;
 377         loff_t kms;
 378
 379         LINVRNT(osc_page_protected(env, osc, CLM_READ, 1));
 380
 381         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 382         LASSERT(!(cmd & OBD_BRW_READ));
 383         LASSERT(osc != NULL);
 384         page = osc->ops_cl.cpl_page;
 385         obj = osc->ops_cl.cpl_obj;
 386
 387         cl_object_attr_lock(obj);
 388         result = cl_object_attr_get(env, obj, attr);
 389         cl_object_attr_unlock(obj);
 390         if (result < 0)
 391                 return result;
 392         kms = attr->cat_kms;
 393         if (cl_offset(obj, page->cp_index) >= kms)
 394                 /* catch race with truncate */
 395                 return 0;
 396         else if (cl_offset(obj, page->cp_index + 1) > kms)
 397                 /* catch sub-page write at end of file */
 398                 return kms % CFS_PAGE_SIZE;
 399         else
 400                 return CFS_PAGE_SIZE;
 401 }
 402
 403 static int osc_completion(const struct lu_env *env,
 404                           void *data, int cmd, struct obdo *oa, int rc)
 405 {
 406         struct osc_page       *opg  = data;
 407         struct osc_async_page *oap  = &opg->ops_oap;
 408         struct cl_page        *page = cl_page_top(opg->ops_cl.cpl_page);
 409         struct osc_object     *obj  = cl2osc(opg->ops_cl.cpl_obj);
 410         enum cl_req_type crt;
 411
 412         LINVRNT(osc_page_protected(env, opg, CLM_READ, 1));
 413         LINVRNT(cl_page_is_vmlocked(env, page));
 414
 415         ENTRY;
 416
 417         cmd &= ~OBD_BRW_NOQUOTA;
 418         LASSERT(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ));
 419         LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
 420         LASSERT(opg->ops_transfer_pinned);
 421
 422         /*
 423          * page->cp_req can be NULL if io submission failed before
 424          * cl_req was allocated.
 425          */
 426         if (page->cp_req != NULL)
 427                 cl_req_page_done(env, page);
 428         LASSERT(page->cp_req == NULL);
 429
 430         /* As the transfer for this page is being done, clear the flags */
 431         oap->oap_async_flags = 0;
 432
 433         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
 434         /* Clear opg->ops_transfer_pinned before VM lock is released. */
 435         opg->ops_transfer_pinned = 0;
 436
 437         spin_lock(&obj->oo_seatbelt);
 438         LASSERT(opg->ops_submitter != NULL);
 439         LASSERT(!list_empty(&opg->ops_inflight));
 440         list_del_init(&opg->ops_inflight);
 441         spin_unlock(&obj->oo_seatbelt);
 442
 443         cl_page_completion(env, page, crt, rc);
 444
 445         /* statistic */
 446         if (rc == 0 && oap->oap_brw_flags & OBD_BRW_SRVLOCK) {
 447                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
 448                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
 449                 int bytes = opg->ops_to - opg->ops_from;
 450
 451                 if (crt == CRT_READ)
 452                         stats->os_lockless_reads += bytes;
 453                 else
 454                         stats->os_lockless_writes += bytes;
 455         }
 456
 457         /*
 458          * This has to be the last operation with the page, as locks are
 459          * released in cl_page_completion() and nothing except for the
 460          * reference counter protects page from concurrent reclaim.
 461          */
 462         lu_ref_del(&page->cp_reference, "transfer", page);
 463         /*
 464          * As page->cp_obj is pinned by a reference from page->cp_req, it is
 465          * safe to call cl_page_put() without risking object destruction in a
 466          * non-blocking context.
 467          */
 468         cl_page_put(env, page);
 469         RETURN(0);
 470 }
 471
 472 const static struct obd_async_page_ops osc_async_page_ops = {
 473         .ap_make_ready    = osc_make_ready,
 474         .ap_refresh_count = osc_refresh_count,
 475         .ap_completion    = osc_completion
 476 };
 477
 478 struct cl_page *osc_page_init(const struct lu_env *env,
 479                               struct cl_object *obj,
 480                               struct cl_page *page, cfs_page_t *vmpage)
 481 {
 482         struct osc_object *osc = cl2osc(obj);
 483         struct osc_page   *opg;
 484         int result;
 485
 486         OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO);
 487         if (opg != NULL) {
 488                 void *oap = &opg->ops_oap;
 489
 490                 opg->ops_from = 0;
 491                 opg->ops_to   = CFS_PAGE_SIZE;
 492
 493                 result = osc_prep_async_page(osc_export(osc),
 494                                              NULL, osc->oo_oinfo, vmpage,
 495                                              cl_offset(obj, page->cp_index),
 496                                              &osc_async_page_ops,
 497                                              opg, (void **)&oap, 1, NULL);
 498                 if (result == 0)
 499                         cl_page_slice_add(page, &opg->ops_cl, obj,
 500                                           &osc_page_ops);
 501                 /*
 502                  * Cannot assert osc_page_protected() here as read-ahead
 503                  * creates temporary pages outside of a lock.
 504                  */
 505 #ifdef INVARIANT_CHECK
 506                 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 507 #endif
 508                 CFS_INIT_LIST_HEAD(&opg->ops_inflight);
 509         } else
 510                 result = -ENOMEM;
 511         return ERR_PTR(result);
 512 }
 513
 514 /**
 515  * Helper function called by osc_io_submit() for every page in an immediate
 516  * transfer (i.e., transferred synchronously).
 517  */
 518 void osc_io_submit_page(const struct lu_env *env,
 519                         struct osc_io *oio, struct osc_page *opg,
 520                         enum cl_req_type crt)
 521 {
 522         struct osc_async_page *oap = &opg->ops_oap;
 523         struct client_obd     *cli = oap->oap_cli;
 524
 525         LINVRNT(osc_page_protected(env, opg,
 526                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 527
 528         oap->oap_page_off   = opg->ops_from;
 529         oap->oap_count      = opg->ops_to - opg->ops_from;
 530         oap->oap_brw_flags |= OBD_BRW_SYNC;
 531         if (osc_io_srvlock(oio))
 532                 oap->oap_brw_flags |= OBD_BRW_SRVLOCK;
 533
 534         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 535         if (!client_is_remote(osc_export(cl2osc(opg->ops_cl.cpl_obj))) &&
 536             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 537                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 538                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 539         }
 540
 541         oap->oap_async_flags |= OSC_FLAGS;
 542         if (oap->oap_cmd & OBD_BRW_READ)
 543                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 544         else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT))
 545                 osc_enter_cache_try(env, cli, oap->oap_loi, oap, 1);
 546
 547         osc_oap_to_pending(oap);
 548         osc_page_transfer_get(opg, "transfer\0imm");
 549         osc_page_transfer_add(env, opg, crt);
 550 }
 551
 552 /** @} osc */