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 static int osc_page_protected(const struct lu_env *env,
  79                               const struct osc_page *opg,
  80                               enum cl_lock_mode mode, int unref)
  81 {
  82         struct cl_object_header *hdr;
  83         struct cl_lock          *scan;
  84         struct cl_page          *page;
  85         struct cl_lock_descr    *descr;
  86         int result;
  87
  88         LINVRNT(!opg->ops_temp);
  89
  90         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
  91         if (result == 0) {
  92                 /* maybe this page is a part of a lockless io? */
  93                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
  94                 page = opg->ops_cl.cpl_page;
  95                 descr = &osc_env_info(env)->oti_descr;
  96                 descr->cld_mode = mode;
  97                 descr->cld_start = page->cp_index;
  98                 descr->cld_end   = page->cp_index;
  99                 spin_lock(&hdr->coh_lock_guard);
 100                 list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 101                         /*
 102                          * Lock-less sub-lock has to be either in HELD state
 103                          * (when io is actively going on), or in CACHED state,
 104                          * when top-lock is being unlocked:
 105                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 106                          */
 107                         if ((scan->cll_state == CLS_HELD ||
 108                              scan->cll_state == CLS_CACHED) &&
 109                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 110                                 struct osc_lock *olck;
 111
 112                                 olck = osc_lock_at(scan);
 113                                 result = osc_lock_is_lockless(olck);
 114                                 break;
 115                         }
 116                 }
 117                 spin_unlock(&hdr->coh_lock_guard);
 118         }
 119         return result;
 120 }
 121
 122 /*****************************************************************************
 123  *
 124  * Page operations.
 125  *
 126  */
 127 static void osc_page_fini(const struct lu_env *env,
 128                           struct cl_page_slice *slice)
 129 {
 130         struct osc_page *opg = cl2osc_page(slice);
 131         CDEBUG(D_TRACE, "%p\n", opg);
 132         OBD_SLAB_FREE_PTR(opg, osc_page_kmem);
 133 }
 134
 135 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 136 {
 137         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 138
 139         LASSERT(!opg->ops_transfer_pinned);
 140         cl_page_get(page);
 141         lu_ref_add_atomic(&page->cp_reference, label, page);
 142         opg->ops_transfer_pinned = 1;
 143 }
 144
 145 static void osc_page_transfer_put(const struct lu_env *env,
 146                                   struct osc_page *opg)
 147 {
 148         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 149
 150         if (opg->ops_transfer_pinned) {
 151                 lu_ref_del(&page->cp_reference, "transfer", page);
 152                 opg->ops_transfer_pinned = 0;
 153                 cl_page_put(env, page);
 154         }
 155 }
 156
 157 /**
 158  * This is called once for every page when it is submitted for a transfer
 159  * either opportunistic (osc_page_cache_add()), or immediate
 160  * (osc_page_submit()).
 161  */
 162 static void osc_page_transfer_add(const struct lu_env *env,
 163                                   struct osc_page *opg, enum cl_req_type crt)
 164 {
 165         struct osc_object *obj;
 166
 167         LINVRNT(cl_page_is_vmlocked(env, opg->ops_cl.cpl_page));
 168
 169         obj = cl2osc(opg->ops_cl.cpl_obj);
 170         spin_lock(&obj->oo_seatbelt);
 171         list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 172         opg->ops_submitter = cfs_current();
 173         spin_unlock(&obj->oo_seatbelt);
 174 }
 175
 176 static int osc_page_cache_add(const struct lu_env *env,
 177                               const struct cl_page_slice *slice,
 178                               struct cl_io *_)
 179 {
 180         struct osc_page   *opg = cl2osc_page(slice);
 181         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 182         struct osc_io     *oio = osc_env_io(env);
 183         int result;
 184         int brw_flags;
 185         int noquota = 0;
 186
 187         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 188         ENTRY;
 189
 190         /* Set the OBD_BRW_SRVLOCK before the page is queued. */
 191         brw_flags = oio->oi_lockless ? OBD_BRW_SRVLOCK : 0;
 192         if (!client_is_remote(osc_export(obj)) &&
 193             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 194                 brw_flags |= OBD_BRW_NOQUOTA;
 195                 noquota = OBD_BRW_NOQUOTA;
 196         }
 197
 198         osc_page_transfer_get(opg, "transfer\0cache");
 199         result = osc_queue_async_io(env, osc_export(obj), NULL, obj->oo_oinfo,
 200                                     &opg->ops_oap, OBD_BRW_WRITE | noquota,
 201                                     0, 0, brw_flags, 0);
 202         if (result != 0)
 203                 osc_page_transfer_put(env, opg);
 204         else
 205                 osc_page_transfer_add(env, opg, CRT_WRITE);
 206         RETURN(result);
 207 }
 208
 209 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 210                       pgoff_t start, pgoff_t end)
 211 {
 212         memset(policy, 0, sizeof *policy);
 213         policy->l_extent.start = cl_offset(obj, start);
 214         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 215 }
 216
 217 static int osc_page_is_under_lock(const struct lu_env *env,
 218                                   const struct cl_page_slice *slice,
 219                                   struct cl_io *_)
 220 {
 221         struct cl_lock *lock;
 222         int             result;
 223
 224         ENTRY;
 225         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 226                                NULL, 1, 0);
 227         if (lock != NULL) {
 228                 cl_lock_put(env, lock);
 229                 result = -EBUSY;
 230         } else
 231                 result = -ENODATA;
 232         RETURN(result);
 233 }
 234
 235 static int osc_page_fail(const struct lu_env *env,
 236                          const struct cl_page_slice *slice, struct cl_io *_)
 237 {
 238         /*
 239          * Cached read?
 240          */
 241         LBUG();
 242         return 0;
 243 }
 244
 245
 246 static const char *osc_list(struct list_head *head)
 247 {
 248         return list_empty(head) ? "-" : "+";
 249 }
 250
 251 static int osc_page_print(const struct lu_env *env,
 252                           const struct cl_page_slice *slice,
 253                           void *cookie, lu_printer_t printer)
 254 {
 255         struct osc_page       *opg = cl2osc_page(slice);
 256         struct osc_async_page *oap = &opg->ops_oap;
 257
 258         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
 259                           "%#x %d %u %s %s %s %llu %u %#x %p %p %p %p %p\n",
 260                           opg, oap->oap_magic, oap->oap_cmd,
 261                           oap->oap_interrupted,
 262                           osc_list(&oap->oap_pending_item),
 263                           osc_list(&oap->oap_urgent_item),
 264                           osc_list(&oap->oap_rpc_item),
 265                           oap->oap_obj_off, oap->oap_page_off,
 266                           oap->oap_async_flags, oap->oap_request,
 267                           oap->oap_cli, oap->oap_loi, oap->oap_caller_ops,
 268                           oap->oap_caller_data);
 269 }
 270
 271 static void osc_page_delete(const struct lu_env *env,
 272                             const struct cl_page_slice *slice)
 273 {
 274         struct osc_page       *opg = cl2osc_page(slice);
 275         struct osc_object     *obj = cl2osc(opg->ops_cl.cpl_obj);
 276         struct osc_async_page *oap = &opg->ops_oap;
 277         int rc;
 278
 279         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 280
 281         ENTRY;
 282         CDEBUG(D_TRACE, "%p\n", opg);
 283         osc_page_transfer_put(env, opg);
 284         rc = osc_teardown_async_page(osc_export(obj), NULL, obj->oo_oinfo, oap);
 285         LASSERTF(rc == 0, "%i\n", rc);
 286         spin_lock(&obj->oo_seatbelt);
 287         list_del_init(&opg->ops_inflight);
 288         spin_unlock(&obj->oo_seatbelt);
 289         EXIT;
 290 }
 291
 292 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 293                    int from, int to)
 294 {
 295         struct osc_page       *opg = cl2osc_page(slice);
 296         struct osc_async_page *oap = &opg->ops_oap;
 297
 298         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 299
 300         opg->ops_from = from;
 301         opg->ops_to   = to;
 302         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 303 }
 304
 305 static int osc_page_cancel(const struct lu_env *env,
 306                            const struct cl_page_slice *slice)
 307 {
 308         struct osc_page *opg       = cl2osc_page(slice);
 309         struct osc_async_page *oap = &opg->ops_oap;
 310         int rc = 0;
 311
 312         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 313
 314         client_obd_list_lock(&oap->oap_cli->cl_loi_list_lock);
 315         /* Check if the transferring against this page
 316          * is completed, or not even queued. */
 317         if (opg->ops_transfer_pinned)
 318                 /* FIXME: may not be interrupted.. */
 319                 rc = osc_oap_interrupted(env, oap);
 320         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 321         client_obd_list_unlock(&oap->oap_cli->cl_loi_list_lock);
 322         return rc;
 323 }
 324
 325 static const struct cl_page_operations osc_page_ops = {
 326         .cpo_fini          = osc_page_fini,
 327         .cpo_print         = osc_page_print,
 328         .cpo_delete        = osc_page_delete,
 329         .cpo_is_under_lock = osc_page_is_under_lock,
 330         .io = {
 331                 [CRT_READ] = {
 332                         .cpo_cache_add = osc_page_fail
 333                 },
 334                 [CRT_WRITE] = {
 335                         .cpo_cache_add = osc_page_cache_add
 336                 }
 337         },
 338         .cpo_clip           = osc_page_clip,
 339         .cpo_cancel         = osc_page_cancel
 340 };
 341
 342 static int osc_make_ready(const struct lu_env *env, void *data, int cmd)
 343 {
 344         struct osc_page *opg  = data;
 345         struct cl_page  *page = cl_page_top(opg->ops_cl.cpl_page);
 346         int result;
 347
 348         LASSERT(cmd == OBD_BRW_WRITE); /* no cached reads */
 349         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 1));
 350
 351         ENTRY;
 352         result = cl_page_make_ready(env, page, CRT_WRITE);
 353         RETURN(result);
 354 }
 355
 356 static int osc_refresh_count(const struct lu_env *env, void *data, int cmd)
 357 {
 358         struct cl_page   *page;
 359         struct osc_page  *osc = data;
 360         struct cl_object *obj;
 361         struct cl_attr   *attr = &osc_env_info(env)->oti_attr;
 362
 363         int result;
 364         loff_t kms;
 365
 366         LINVRNT(osc_page_protected(env, osc, CLM_READ, 1));
 367
 368         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 369         LASSERT(!(cmd & OBD_BRW_READ));
 370         LASSERT(osc != NULL);
 371         page = osc->ops_cl.cpl_page;
 372         obj = osc->ops_cl.cpl_obj;
 373
 374         cl_object_attr_lock(obj);
 375         result = cl_object_attr_get(env, obj, attr);
 376         cl_object_attr_unlock(obj);
 377         if (result < 0)
 378                 return result;
 379         kms = attr->cat_kms;
 380         if (cl_offset(obj, page->cp_index) >= kms)
 381                 /* catch race with truncate */
 382                 return 0;
 383         else if (cl_offset(obj, page->cp_index + 1) > kms)
 384                 /* catch sub-page write at end of file */
 385                 return kms % CFS_PAGE_SIZE;
 386         else
 387                 return CFS_PAGE_SIZE;
 388 }
 389
 390 static int osc_completion(const struct lu_env *env,
 391                           void *data, int cmd, struct obdo *oa, int rc)
 392 {
 393         struct osc_page       *opg  = data;
 394         struct osc_async_page *oap  = &opg->ops_oap;
 395         struct cl_page        *page = cl_page_top(opg->ops_cl.cpl_page);
 396         struct osc_object     *obj  = cl2osc(opg->ops_cl.cpl_obj);
 397         enum cl_req_type crt;
 398
 399         LINVRNT(osc_page_protected(env, opg, CLM_READ, 1));
 400         LINVRNT(cl_page_is_vmlocked(env, page));
 401
 402         ENTRY;
 403
 404         cmd &= ~OBD_BRW_NOQUOTA;
 405         LASSERT(equi(page->cp_state == CPS_PAGEIN,  cmd == OBD_BRW_READ));
 406         LASSERT(equi(page->cp_state == CPS_PAGEOUT, cmd == OBD_BRW_WRITE));
 407         LASSERT(opg->ops_transfer_pinned);
 408
 409         /*
 410          * page->cp_req can be NULL if io submission failed before
 411          * cl_req was allocated.
 412          */
 413         if (page->cp_req != NULL)
 414                 cl_req_page_done(env, page);
 415         LASSERT(page->cp_req == NULL);
 416
 417         /* As the transfer for this page is being done, clear the flags */
 418         oap->oap_async_flags = 0;
 419
 420         crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
 421         /* Clear opg->ops_transfer_pinned before VM lock is released. */
 422         opg->ops_transfer_pinned = 0;
 423
 424         spin_lock(&obj->oo_seatbelt);
 425         LASSERT(opg->ops_submitter != NULL);
 426         LASSERT(!list_empty(&opg->ops_inflight));
 427         list_del_init(&opg->ops_inflight);
 428         spin_unlock(&obj->oo_seatbelt);
 429
 430         cl_page_completion(env, page, crt, rc);
 431
 432         /* statistic */
 433         if (rc == 0 && oap->oap_brw_flags & OBD_BRW_SRVLOCK) {
 434                 struct lu_device *ld    = opg->ops_cl.cpl_obj->co_lu.lo_dev;
 435                 struct osc_stats *stats = &lu2osc_dev(ld)->od_stats;
 436                 int bytes = opg->ops_to - opg->ops_from;
 437
 438                 if (crt == CRT_READ)
 439                         stats->os_lockless_reads += bytes;
 440                 else
 441                         stats->os_lockless_writes += bytes;
 442         }
 443
 444         /*
 445          * This has to be the last operation with the page, as locks are
 446          * released in cl_page_completion() and nothing except for the
 447          * reference counter protects page from concurrent reclaim.
 448          */
 449         lu_ref_del(&page->cp_reference, "transfer", page);
 450         /*
 451          * As page->cp_obj is pinned by a reference from page->cp_req, it is
 452          * safe to call cl_page_put() without risking object destruction in a
 453          * non-blocking context.
 454          */
 455         cl_page_put(env, page);
 456         RETURN(0);
 457 }
 458
 459 const static struct obd_async_page_ops osc_async_page_ops = {
 460         .ap_make_ready    = osc_make_ready,
 461         .ap_refresh_count = osc_refresh_count,
 462         .ap_completion    = osc_completion
 463 };
 464
 465 struct cl_page *osc_page_init(const struct lu_env *env,
 466                               struct cl_object *obj,
 467                               struct cl_page *page, cfs_page_t *vmpage)
 468 {
 469         struct osc_object *osc = cl2osc(obj);
 470         struct osc_page   *opg;
 471         int result;
 472
 473         OBD_SLAB_ALLOC_PTR(opg, osc_page_kmem);
 474         if (opg != NULL) {
 475                 void *oap = &opg->ops_oap;
 476
 477                 opg->ops_from = 0;
 478                 opg->ops_to   = CFS_PAGE_SIZE;
 479
 480                 result = osc_prep_async_page(osc_export(osc),
 481                                              NULL, osc->oo_oinfo, vmpage,
 482                                              cl_offset(obj, page->cp_index),
 483                                              &osc_async_page_ops,
 484                                              opg, (void **)&oap, 1, NULL);
 485                 if (result == 0)
 486                         cl_page_slice_add(page, &opg->ops_cl, obj,
 487                                           &osc_page_ops);
 488                 /*
 489                  * Cannot assert osc_page_protected() here as read-ahead
 490                  * creates temporary pages outside of a lock.
 491                  */
 492 #ifdef INVARIANT_CHECK
 493                 opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 494 #endif
 495                 CFS_INIT_LIST_HEAD(&opg->ops_inflight);
 496         } else
 497                 result = -ENOMEM;
 498         return ERR_PTR(result);
 499 }
 500
 501 /**
 502  * Helper function called by osc_io_submit() for every page in an immediate
 503  * transfer (i.e., transferred synchronously).
 504  */
 505 void osc_io_submit_page(const struct lu_env *env,
 506                         struct osc_io *oio, struct osc_page *opg,
 507                         enum cl_req_type crt)
 508 {
 509         struct osc_async_page *oap = &opg->ops_oap;
 510         struct client_obd     *cli = oap->oap_cli;
 511
 512         LINVRNT(osc_page_protected(env, opg,
 513                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 514
 515         oap->oap_page_off   = opg->ops_from;
 516         oap->oap_count      = opg->ops_to - opg->ops_from;
 517         oap->oap_brw_flags |= OBD_BRW_SYNC;
 518         if (oio->oi_lockless)
 519                 oap->oap_brw_flags |= OBD_BRW_SRVLOCK;
 520
 521         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 522         if (!client_is_remote(osc_export(cl2osc(opg->ops_cl.cpl_obj))) &&
 523             cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 524                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 525                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 526         }
 527
 528         oap->oap_async_flags |= OSC_FLAGS;
 529         if (oap->oap_cmd & OBD_BRW_READ)
 530                 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 531         else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT))
 532                 osc_enter_cache_try(env, cli, oap->oap_loi, oap, 1);
 533
 534         osc_oap_to_pending(oap);
 535         osc_page_transfer_get(opg, "transfer\0imm");
 536         osc_page_transfer_add(env, opg, crt);
 537 }
 538
 539 /** @} osc */