lustre/osc/osc_object.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   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.
   9  *
  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).
  15  *
  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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_object for OSC layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39 #include <lustre_osc.h>
  40
  41 #include "osc_internal.h"
  42
  43 /** \addtogroup osc
  44  *  @{
  45  */
  46
  47 /*****************************************************************************
  48  *
  49  * Object operations.
  50  *
  51  */
  52 static void osc_obj_build_res_name(struct osc_object *osc,
  53                                    struct ldlm_res_id *resname)
  54 {
  55         ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
  56 }
  57
  58 static const struct osc_object_operations osc_object_ops = {
  59         .oto_build_res_name = osc_obj_build_res_name,
  60         .oto_dlmlock_at_pgoff = osc_obj_dlmlock_at_pgoff,
  61 };
  62
  63 int osc_object_init(const struct lu_env *env, struct lu_object *obj,
  64                     const struct lu_object_conf *conf)
  65 {
  66         struct osc_object           *osc   = lu2osc(obj);
  67         const struct cl_object_conf *cconf = lu2cl_conf(conf);
  68
  69         osc->oo_oinfo = cconf->u.coc_oinfo;
  70 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
  71         mutex_init(&osc->oo_debug_mutex);
  72 #endif
  73         INIT_LIST_HEAD(&osc->oo_ready_item);
  74         INIT_LIST_HEAD(&osc->oo_hp_ready_item);
  75         INIT_LIST_HEAD(&osc->oo_write_item);
  76         INIT_LIST_HEAD(&osc->oo_read_item);
  77
  78         osc->oo_root.rb_node = NULL;
  79         INIT_LIST_HEAD(&osc->oo_hp_exts);
  80         INIT_LIST_HEAD(&osc->oo_urgent_exts);
  81         INIT_LIST_HEAD(&osc->oo_full_exts);
  82         INIT_LIST_HEAD(&osc->oo_reading_exts);
  83         atomic_set(&osc->oo_nr_reads, 0);
  84         atomic_set(&osc->oo_nr_writes, 0);
  85         spin_lock_init(&osc->oo_lock);
  86         spin_lock_init(&osc->oo_tree_lock);
  87         spin_lock_init(&osc->oo_ol_spin);
  88         INIT_LIST_HEAD(&osc->oo_ol_list);
  89
  90         atomic_set(&osc->oo_nr_ios, 0);
  91         init_waitqueue_head(&osc->oo_io_waitq);
  92
  93         LASSERT(osc->oo_obj_ops != NULL);
  94
  95         cl_object_page_init(lu2cl(obj), sizeof(struct osc_page));
  96
  97         return 0;
  98 }
  99 EXPORT_SYMBOL(osc_object_init);
 100
 101 void osc_object_free(const struct lu_env *env, struct lu_object *obj)
 102 {
 103         struct osc_object *osc = lu2osc(obj);
 104
 105         LASSERT(list_empty(&osc->oo_ready_item));
 106         LASSERT(list_empty(&osc->oo_hp_ready_item));
 107         LASSERT(list_empty(&osc->oo_write_item));
 108         LASSERT(list_empty(&osc->oo_read_item));
 109
 110         LASSERT(osc->oo_root.rb_node == NULL);
 111         LASSERT(list_empty(&osc->oo_hp_exts));
 112         LASSERT(list_empty(&osc->oo_urgent_exts));
 113         LASSERT(list_empty(&osc->oo_full_exts));
 114         LASSERT(list_empty(&osc->oo_reading_exts));
 115         LASSERT(atomic_read(&osc->oo_nr_reads) == 0);
 116         LASSERT(atomic_read(&osc->oo_nr_writes) == 0);
 117         LASSERT(list_empty(&osc->oo_ol_list));
 118         LASSERT(atomic_read(&osc->oo_nr_ios) == 0);
 119
 120         lu_object_fini(obj);
 121         OBD_SLAB_FREE_PTR(osc, osc_object_kmem);
 122 }
 123 EXPORT_SYMBOL(osc_object_free);
 124
 125 int osc_lvb_print(const struct lu_env *env, void *cookie,
 126                   lu_printer_t p, const struct ost_lvb *lvb)
 127 {
 128         return (*p)(env, cookie, "size: %llu mtime: %llu atime: %llu "
 129                     "ctime: %llu blocks: %llu",
 130                     lvb->lvb_size, lvb->lvb_mtime, lvb->lvb_atime,
 131                     lvb->lvb_ctime, lvb->lvb_blocks);
 132 }
 133 EXPORT_SYMBOL(osc_lvb_print);
 134
 135 int osc_object_print(const struct lu_env *env, void *cookie,
 136                      lu_printer_t p, const struct lu_object *obj)
 137 {
 138         struct osc_object *osc = lu2osc(obj);
 139         struct lov_oinfo *oinfo = osc->oo_oinfo;
 140         struct osc_async_rc *ar = &oinfo->loi_ar;
 141
 142         (*p)(env, cookie, "id: "DOSTID" "
 143              "idx: %d gen: %d kms_valid: %u kms %llu "
 144              "rc: %d force_sync: %d min_xid: %llu ",
 145              POSTID(&oinfo->loi_oi), oinfo->loi_ost_idx,
 146              oinfo->loi_ost_gen, oinfo->loi_kms_valid, oinfo->loi_kms,
 147              ar->ar_rc, ar->ar_force_sync, ar->ar_min_xid);
 148         osc_lvb_print(env, cookie, p, &oinfo->loi_lvb);
 149         return 0;
 150 }
 151 EXPORT_SYMBOL(osc_object_print);
 152
 153
 154 int osc_attr_get(const struct lu_env *env, struct cl_object *obj,
 155                  struct cl_attr *attr)
 156 {
 157         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 158
 159         cl_lvb2attr(attr, &oinfo->loi_lvb);
 160         attr->cat_kms = oinfo->loi_kms_valid ? oinfo->loi_kms : 0;
 161         return 0;
 162 }
 163 EXPORT_SYMBOL(osc_attr_get);
 164
 165 int osc_attr_update(const struct lu_env *env, struct cl_object *obj,
 166                     const struct cl_attr *attr, unsigned valid)
 167 {
 168         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 169         struct ost_lvb   *lvb   = &oinfo->loi_lvb;
 170
 171         if (valid & CAT_SIZE)
 172                 lvb->lvb_size = attr->cat_size;
 173         if (valid & CAT_MTIME)
 174                 lvb->lvb_mtime = attr->cat_mtime;
 175         if (valid & CAT_ATIME)
 176                 lvb->lvb_atime = attr->cat_atime;
 177         if (valid & CAT_CTIME)
 178                 lvb->lvb_ctime = attr->cat_ctime;
 179         if (valid & CAT_BLOCKS)
 180                 lvb->lvb_blocks = attr->cat_blocks;
 181         if (valid & CAT_KMS) {
 182                 CDEBUG(D_CACHE, "set kms from %lluto %llu\n",
 183                        oinfo->loi_kms, (__u64)attr->cat_kms);
 184                 loi_kms_set(oinfo, attr->cat_kms);
 185         }
 186         return 0;
 187 }
 188 EXPORT_SYMBOL(osc_attr_update);
 189
 190 int osc_object_glimpse(const struct lu_env *env, const struct cl_object *obj,
 191                        struct ost_lvb *lvb)
 192 {
 193         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 194
 195         lvb->lvb_size = oinfo->loi_kms;
 196         lvb->lvb_blocks = oinfo->loi_lvb.lvb_blocks;
 197         return 0;
 198 }
 199 EXPORT_SYMBOL(osc_object_glimpse);
 200
 201 static int osc_object_ast_clear(struct ldlm_lock *lock, void *data)
 202 {
 203         struct osc_object *osc = (struct osc_object *)data;
 204         struct ost_lvb *lvb = lock->l_lvb_data;
 205         struct lov_oinfo *oinfo;
 206         ENTRY;
 207
 208         if (lock->l_ast_data == data) {
 209                 lock->l_ast_data = NULL;
 210
 211                 LASSERT(osc != NULL);
 212                 LASSERT(osc->oo_oinfo != NULL);
 213                 LASSERT(lvb != NULL);
 214
 215                 /* Updates lvb in lock by the cached oinfo */
 216                 oinfo = osc->oo_oinfo;
 217
 218                 LDLM_DEBUG(lock, "update lock size %llu blocks %llu [cma]time: "
 219                            "%llu %llu %llu by oinfo size %llu blocks %llu "
 220                            "[cma]time %llu %llu %llu", lvb->lvb_size,
 221                            lvb->lvb_blocks, lvb->lvb_ctime, lvb->lvb_mtime,
 222                            lvb->lvb_atime, oinfo->loi_lvb.lvb_size,
 223                            oinfo->loi_lvb.lvb_blocks, oinfo->loi_lvb.lvb_ctime,
 224                            oinfo->loi_lvb.lvb_mtime, oinfo->loi_lvb.lvb_atime);
 225                 LASSERT(oinfo->loi_lvb.lvb_size >= oinfo->loi_kms);
 226
 227                 cl_object_attr_lock(&osc->oo_cl);
 228                 memcpy(lvb, &oinfo->loi_lvb, sizeof(oinfo->loi_lvb));
 229                 cl_object_attr_unlock(&osc->oo_cl);
 230                 ldlm_clear_lvb_cached(lock);
 231         }
 232         RETURN(LDLM_ITER_CONTINUE);
 233 }
 234
 235 int osc_object_prune(const struct lu_env *env, struct cl_object *obj)
 236 {
 237         struct osc_object  *osc = cl2osc(obj);
 238         struct ldlm_res_id *resname = &osc_env_info(env)->oti_resname;
 239
 240         /* DLM locks don't hold a reference of osc_object so we have to
 241          * clear it before the object is being destroyed. */
 242         osc_build_res_name(osc, resname);
 243         ldlm_resource_iterate(osc_export(osc)->exp_obd->obd_namespace, resname,
 244                               osc_object_ast_clear, osc);
 245         return 0;
 246 }
 247 EXPORT_SYMBOL(osc_object_prune);
 248
 249 static int osc_object_fiemap(const struct lu_env *env, struct cl_object *obj,
 250                              struct ll_fiemap_info_key *fmkey,
 251                              struct fiemap *fiemap, size_t *buflen)
 252 {
 253         struct obd_export *exp = osc_export(cl2osc(obj));
 254         struct ldlm_res_id resid;
 255         union ldlm_policy_data policy;
 256         struct lustre_handle lockh;
 257         enum ldlm_mode mode = LCK_MINMODE;
 258         struct ptlrpc_request *req;
 259         struct fiemap *reply;
 260         char *tmp;
 261         int rc;
 262         ENTRY;
 263
 264         fmkey->lfik_oa.o_oi = cl2osc(obj)->oo_oinfo->loi_oi;
 265         if (!(fmkey->lfik_fiemap.fm_flags & FIEMAP_FLAG_SYNC))
 266                 goto skip_locking;
 267
 268         policy.l_extent.start = fmkey->lfik_fiemap.fm_start & PAGE_MASK;
 269
 270         if (OBD_OBJECT_EOF - fmkey->lfik_fiemap.fm_length <=
 271             fmkey->lfik_fiemap.fm_start + PAGE_SIZE - 1)
 272                 policy.l_extent.end = OBD_OBJECT_EOF;
 273         else
 274                 policy.l_extent.end = (fmkey->lfik_fiemap.fm_start +
 275                                        fmkey->lfik_fiemap.fm_length +
 276                                        PAGE_SIZE - 1) & PAGE_MASK;
 277
 278         ostid_build_res_name(&fmkey->lfik_oa.o_oi, &resid);
 279         mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
 280                                LDLM_FL_BLOCK_GRANTED | LDLM_FL_LVB_READY,
 281                                &resid, LDLM_EXTENT, &policy,
 282                                LCK_PR | LCK_PW, &lockh, 0);
 283         if (mode) { /* lock is cached on client */
 284                 if (mode != LCK_PR) {
 285                         ldlm_lock_addref(&lockh, LCK_PR);
 286                         ldlm_lock_decref(&lockh, LCK_PW);
 287                 }
 288         } else { /* no cached lock, needs acquire lock on server side */
 289                 fmkey->lfik_oa.o_valid |= OBD_MD_FLFLAGS;
 290                 fmkey->lfik_oa.o_flags |= OBD_FL_SRVLOCK;
 291         }
 292
 293 skip_locking:
 294         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 295                                    &RQF_OST_GET_INFO_FIEMAP);
 296         if (req == NULL)
 297                 GOTO(drop_lock, rc = -ENOMEM);
 298
 299         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY, RCL_CLIENT,
 300                              sizeof(*fmkey));
 301         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_CLIENT,
 302                              *buflen);
 303         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_SERVER,
 304                              *buflen);
 305
 306         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 307         if (rc != 0) {
 308                 ptlrpc_request_free(req);
 309                 GOTO(drop_lock, rc);
 310         }
 311         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
 312         memcpy(tmp, fmkey, sizeof(*fmkey));
 313         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
 314         memcpy(tmp, fiemap, *buflen);
 315         ptlrpc_request_set_replen(req);
 316
 317         rc = ptlrpc_queue_wait(req);
 318         if (rc != 0)
 319                 GOTO(fini_req, rc);
 320
 321         reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
 322         if (reply == NULL)
 323                 GOTO(fini_req, rc = -EPROTO);
 324
 325         memcpy(fiemap, reply, *buflen);
 326 fini_req:
 327         ptlrpc_req_finished(req);
 328 drop_lock:
 329         if (mode)
 330                 ldlm_lock_decref(&lockh, LCK_PR);
 331         RETURN(rc);
 332 }
 333
 334 int osc_object_is_contended(struct osc_object *obj)
 335 {
 336         struct osc_device *dev = lu2osc_dev(obj->oo_cl.co_lu.lo_dev);
 337         time64_t osc_contention_time = dev->od_contention_time;
 338         ktime_t retry_time;
 339
 340         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_OBJECT_CONTENTION))
 341                 return 1;
 342
 343         if (!obj->oo_contended)
 344                 return 0;
 345
 346         /*
 347          * I like copy-paste. the code is copied from
 348          * ll_file_is_contended.
 349          */
 350         retry_time = ktime_add_ns(obj->oo_contention_time,
 351                                   osc_contention_time * NSEC_PER_SEC);
 352         if (ktime_after(ktime_get(), retry_time)) {
 353                 osc_object_clear_contended(obj);
 354                 return 0;
 355         }
 356         return 1;
 357 }
 358 EXPORT_SYMBOL(osc_object_is_contended);
 359
 360 /**
 361  * Implementation of struct cl_object_operations::coo_req_attr_set() for osc
 362  * layer. osc is responsible for struct obdo::o_id and struct obdo::o_seq
 363  * fields.
 364  */
 365 static void osc_req_attr_set(const struct lu_env *env, struct cl_object *obj,
 366                              struct cl_req_attr *attr)
 367 {
 368         struct lov_oinfo *oinfo;
 369         struct obdo      *oa;
 370         struct ost_lvb   *lvb;
 371         u64               flags = attr->cra_flags;
 372
 373         oinfo   = cl2osc(obj)->oo_oinfo;
 374         lvb     = &oinfo->loi_lvb;
 375         oa      = attr->cra_oa;
 376
 377         if ((flags & OBD_MD_FLMTIME) != 0) {
 378                 oa->o_mtime = lvb->lvb_mtime;
 379                 oa->o_valid |= OBD_MD_FLMTIME;
 380         }
 381         if ((flags & OBD_MD_FLATIME) != 0) {
 382                 oa->o_atime = lvb->lvb_atime;
 383                 oa->o_valid |= OBD_MD_FLATIME;
 384         }
 385         if ((flags & OBD_MD_FLCTIME) != 0) {
 386                 oa->o_ctime = lvb->lvb_ctime;
 387                 oa->o_valid |= OBD_MD_FLCTIME;
 388         }
 389         if (flags & OBD_MD_FLGROUP) {
 390                 ostid_set_seq(&oa->o_oi, ostid_seq(&oinfo->loi_oi));
 391                 oa->o_valid |= OBD_MD_FLGROUP;
 392         }
 393         if (flags & OBD_MD_FLID) {
 394                 int rc;
 395
 396                 rc = ostid_set_id(&oa->o_oi, ostid_id(&oinfo->loi_oi));
 397                 if (rc) {
 398                         CERROR("Bad %llu to set " DOSTID " : rc %d\n",
 399                                (unsigned long long)ostid_id(&oinfo->loi_oi),
 400                                POSTID(&oa->o_oi), rc);
 401                 }
 402                 oa->o_valid |= OBD_MD_FLID;
 403         }
 404         if (flags & OBD_MD_FLHANDLE) {
 405                 struct ldlm_lock *lock;
 406                 struct osc_page *opg;
 407
 408                 opg = osc_cl_page_osc(attr->cra_page, cl2osc(obj));
 409                 lock = osc_dlmlock_at_pgoff(env, cl2osc(obj), osc_index(opg),
 410                                 OSC_DAP_FL_TEST_LOCK | OSC_DAP_FL_CANCELING);
 411                 if (lock == NULL && !opg->ops_srvlock) {
 412                         struct ldlm_resource *res;
 413                         struct ldlm_res_id *resname;
 414
 415                         CL_PAGE_DEBUG(D_ERROR, env, attr->cra_page,
 416                                       "uncovered page!\n");
 417
 418                         resname = &osc_env_info(env)->oti_resname;
 419                         ostid_build_res_name(&oinfo->loi_oi, resname);
 420                         res = ldlm_resource_get(
 421                                 osc_export(cl2osc(obj))->exp_obd->obd_namespace,
 422                                 NULL, resname, LDLM_EXTENT, 0);
 423                         ldlm_resource_dump(D_ERROR, res);
 424
 425                         libcfs_debug_dumpstack(NULL);
 426                         LBUG();
 427                 }
 428
 429                 /* check for lockless io. */
 430                 if (lock != NULL) {
 431                         oa->o_handle = lock->l_remote_handle;
 432                         oa->o_valid |= OBD_MD_FLHANDLE;
 433                         LDLM_LOCK_PUT(lock);
 434                 }
 435         }
 436 }
 437
 438 static const struct cl_object_operations osc_ops = {
 439         .coo_page_init    = osc_page_init,
 440         .coo_lock_init    = osc_lock_init,
 441         .coo_io_init      = osc_io_init,
 442         .coo_attr_get     = osc_attr_get,
 443         .coo_attr_update  = osc_attr_update,
 444         .coo_glimpse      = osc_object_glimpse,
 445         .coo_prune        = osc_object_prune,
 446         .coo_fiemap       = osc_object_fiemap,
 447         .coo_req_attr_set = osc_req_attr_set
 448 };
 449
 450 static const struct lu_object_operations osc_lu_obj_ops = {
 451         .loo_object_init      = osc_object_init,
 452         .loo_object_release   = NULL,
 453         .loo_object_free      = osc_object_free,
 454         .loo_object_print     = osc_object_print,
 455         .loo_object_invariant = NULL
 456 };
 457
 458 struct lu_object *osc_object_alloc(const struct lu_env *env,
 459                                    const struct lu_object_header *unused,
 460                                    struct lu_device *dev)
 461 {
 462         struct osc_object *osc;
 463         struct lu_object  *obj;
 464
 465         OBD_SLAB_ALLOC_PTR_GFP(osc, osc_object_kmem, GFP_NOFS);
 466         if (osc != NULL) {
 467                 obj = osc2lu(osc);
 468                 lu_object_init(obj, NULL, dev);
 469                 osc->oo_cl.co_ops = &osc_ops;
 470                 obj->lo_ops = &osc_lu_obj_ops;
 471                 osc->oo_obj_ops = &osc_object_ops;
 472         } else
 473                 obj = NULL;
 474         return obj;
 475 }
 476
 477 int osc_object_invalidate(const struct lu_env *env, struct osc_object *osc)
 478 {
 479         struct l_wait_info lwi = { 0 };
 480         ENTRY;
 481
 482         CDEBUG(D_INODE, "Invalidate osc object: %p, # of active IOs: %d\n",
 483                osc, atomic_read(&osc->oo_nr_ios));
 484
 485         l_wait_event(osc->oo_io_waitq, atomic_read(&osc->oo_nr_ios) == 0, &lwi);
 486
 487         /* Discard all dirty pages of this object. */
 488         osc_cache_truncate_start(env, osc, 0, NULL);
 489
 490         /* Discard all caching pages */
 491         osc_lock_discard_pages(env, osc, 0, CL_PAGE_EOF, true);
 492
 493         /* Clear ast data of dlm lock. Do this after discarding all pages */
 494         osc_object_prune(env, osc2cl(osc));
 495
 496         RETURN(0);
 497 }
 498 EXPORT_SYMBOL(osc_object_invalidate);
 499 /** @} osc */