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, 2015, 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
  53 static int osc_object_init(const struct lu_env *env, struct lu_object *obj,
  54                            const struct lu_object_conf *conf)
  55 {
  56         struct osc_object           *osc   = lu2osc(obj);
  57         const struct cl_object_conf *cconf = lu2cl_conf(conf);
  58
  59         osc->oo_oinfo = cconf->u.coc_oinfo;
  60 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
  61         mutex_init(&osc->oo_debug_mutex);
  62 #endif
  63         INIT_LIST_HEAD(&osc->oo_ready_item);
  64         INIT_LIST_HEAD(&osc->oo_hp_ready_item);
  65         INIT_LIST_HEAD(&osc->oo_write_item);
  66         INIT_LIST_HEAD(&osc->oo_read_item);
  67
  68         osc->oo_root.rb_node = NULL;
  69         INIT_LIST_HEAD(&osc->oo_hp_exts);
  70         INIT_LIST_HEAD(&osc->oo_urgent_exts);
  71         INIT_LIST_HEAD(&osc->oo_full_exts);
  72         INIT_LIST_HEAD(&osc->oo_reading_exts);
  73         atomic_set(&osc->oo_nr_reads, 0);
  74         atomic_set(&osc->oo_nr_writes, 0);
  75         spin_lock_init(&osc->oo_lock);
  76         spin_lock_init(&osc->oo_tree_lock);
  77         spin_lock_init(&osc->oo_ol_spin);
  78         INIT_LIST_HEAD(&osc->oo_ol_list);
  79
  80         atomic_set(&osc->oo_nr_ios, 0);
  81         init_waitqueue_head(&osc->oo_io_waitq);
  82
  83         cl_object_page_init(lu2cl(obj), sizeof(struct osc_page));
  84
  85         return 0;
  86 }
  87
  88 static void osc_object_free(const struct lu_env *env, struct lu_object *obj)
  89 {
  90         struct osc_object *osc = lu2osc(obj);
  91
  92         LASSERT(list_empty(&osc->oo_ready_item));
  93         LASSERT(list_empty(&osc->oo_hp_ready_item));
  94         LASSERT(list_empty(&osc->oo_write_item));
  95         LASSERT(list_empty(&osc->oo_read_item));
  96
  97         LASSERT(osc->oo_root.rb_node == NULL);
  98         LASSERT(list_empty(&osc->oo_hp_exts));
  99         LASSERT(list_empty(&osc->oo_urgent_exts));
 100         LASSERT(list_empty(&osc->oo_full_exts));
 101         LASSERT(list_empty(&osc->oo_reading_exts));
 102         LASSERT(atomic_read(&osc->oo_nr_reads) == 0);
 103         LASSERT(atomic_read(&osc->oo_nr_writes) == 0);
 104         LASSERT(list_empty(&osc->oo_ol_list));
 105         LASSERT(atomic_read(&osc->oo_nr_ios) == 0);
 106
 107         lu_object_fini(obj);
 108         OBD_SLAB_FREE_PTR(osc, osc_object_kmem);
 109 }
 110
 111 int osc_lvb_print(const struct lu_env *env, void *cookie,
 112                   lu_printer_t p, const struct ost_lvb *lvb)
 113 {
 114         return (*p)(env, cookie, "size: %llu mtime: %llu atime: %llu "
 115                     "ctime: %llu blocks: %llu",
 116                     lvb->lvb_size, lvb->lvb_mtime, lvb->lvb_atime,
 117                     lvb->lvb_ctime, lvb->lvb_blocks);
 118 }
 119
 120 static int osc_object_print(const struct lu_env *env, void *cookie,
 121                             lu_printer_t p, const struct lu_object *obj)
 122 {
 123         struct osc_object   *osc   = lu2osc(obj);
 124         struct lov_oinfo    *oinfo = osc->oo_oinfo;
 125         struct osc_async_rc *ar    = &oinfo->loi_ar;
 126
 127         (*p)(env, cookie, "id: "DOSTID" "
 128              "idx: %d gen: %d kms_valid: %u kms %llu "
 129              "rc: %d force_sync: %d min_xid: %llu ",
 130              POSTID(&oinfo->loi_oi), oinfo->loi_ost_idx,
 131              oinfo->loi_ost_gen, oinfo->loi_kms_valid, oinfo->loi_kms,
 132              ar->ar_rc, ar->ar_force_sync, ar->ar_min_xid);
 133         osc_lvb_print(env, cookie, p, &oinfo->loi_lvb);
 134         return 0;
 135 }
 136
 137
 138 static int osc_attr_get(const struct lu_env *env, struct cl_object *obj,
 139                         struct cl_attr *attr)
 140 {
 141         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 142
 143         cl_lvb2attr(attr, &oinfo->loi_lvb);
 144         attr->cat_kms = oinfo->loi_kms_valid ? oinfo->loi_kms : 0;
 145         return 0;
 146 }
 147
 148 static int osc_attr_update(const struct lu_env *env, struct cl_object *obj,
 149                            const struct cl_attr *attr, unsigned valid)
 150 {
 151         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 152         struct ost_lvb   *lvb   = &oinfo->loi_lvb;
 153
 154         if (valid & CAT_SIZE)
 155                 lvb->lvb_size = attr->cat_size;
 156         if (valid & CAT_MTIME)
 157                 lvb->lvb_mtime = attr->cat_mtime;
 158         if (valid & CAT_ATIME)
 159                 lvb->lvb_atime = attr->cat_atime;
 160         if (valid & CAT_CTIME)
 161                 lvb->lvb_ctime = attr->cat_ctime;
 162         if (valid & CAT_BLOCKS)
 163                 lvb->lvb_blocks = attr->cat_blocks;
 164         if (valid & CAT_KMS) {
 165                 CDEBUG(D_CACHE, "set kms from %lluto %llu\n",
 166                        oinfo->loi_kms, (__u64)attr->cat_kms);
 167                 loi_kms_set(oinfo, attr->cat_kms);
 168         }
 169         return 0;
 170 }
 171
 172 static int osc_object_glimpse(const struct lu_env *env,
 173                               const struct cl_object *obj, struct ost_lvb *lvb)
 174 {
 175         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 176
 177         ENTRY;
 178         lvb->lvb_size   = oinfo->loi_kms;
 179         lvb->lvb_blocks = oinfo->loi_lvb.lvb_blocks;
 180         RETURN(0);
 181 }
 182
 183 static int osc_object_ast_clear(struct ldlm_lock *lock, void *data)
 184 {
 185         ENTRY;
 186
 187         if (lock->l_ast_data == data)
 188                 lock->l_ast_data = NULL;
 189         RETURN(LDLM_ITER_CONTINUE);
 190 }
 191
 192 static int osc_object_prune(const struct lu_env *env, struct cl_object *obj)
 193 {
 194         struct osc_object       *osc = cl2osc(obj);
 195         struct ldlm_res_id      *resname = &osc_env_info(env)->oti_resname;
 196
 197         /* DLM locks don't hold a reference of osc_object so we have to
 198          * clear it before the object is being destroyed. */
 199         ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
 200         ldlm_resource_iterate(osc_export(osc)->exp_obd->obd_namespace, resname,
 201                               osc_object_ast_clear, osc);
 202         return 0;
 203 }
 204
 205 static int osc_object_fiemap(const struct lu_env *env, struct cl_object *obj,
 206                              struct ll_fiemap_info_key *fmkey,
 207                              struct fiemap *fiemap, size_t *buflen)
 208 {
 209         struct obd_export *exp = osc_export(cl2osc(obj));
 210         struct ldlm_res_id resid;
 211         union ldlm_policy_data policy;
 212         struct lustre_handle lockh;
 213         enum ldlm_mode mode = LCK_MINMODE;
 214         struct ptlrpc_request *req;
 215         struct fiemap *reply;
 216         char *tmp;
 217         int rc;
 218         ENTRY;
 219
 220         fmkey->lfik_oa.o_oi = cl2osc(obj)->oo_oinfo->loi_oi;
 221         if (!(fmkey->lfik_fiemap.fm_flags & FIEMAP_FLAG_SYNC))
 222                 goto skip_locking;
 223
 224         policy.l_extent.start = fmkey->lfik_fiemap.fm_start & PAGE_MASK;
 225
 226         if (OBD_OBJECT_EOF - fmkey->lfik_fiemap.fm_length <=
 227             fmkey->lfik_fiemap.fm_start + PAGE_SIZE - 1)
 228                 policy.l_extent.end = OBD_OBJECT_EOF;
 229         else
 230                 policy.l_extent.end = (fmkey->lfik_fiemap.fm_start +
 231                                        fmkey->lfik_fiemap.fm_length +
 232                                        PAGE_SIZE - 1) & PAGE_MASK;
 233
 234         ostid_build_res_name(&fmkey->lfik_oa.o_oi, &resid);
 235         mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
 236                                LDLM_FL_BLOCK_GRANTED | LDLM_FL_LVB_READY,
 237                                &resid, LDLM_EXTENT, &policy,
 238                                LCK_PR | LCK_PW, &lockh, 0);
 239         if (mode) { /* lock is cached on client */
 240                 if (mode != LCK_PR) {
 241                         ldlm_lock_addref(&lockh, LCK_PR);
 242                         ldlm_lock_decref(&lockh, LCK_PW);
 243                 }
 244         } else { /* no cached lock, needs acquire lock on server side */
 245                 fmkey->lfik_oa.o_valid |= OBD_MD_FLFLAGS;
 246                 fmkey->lfik_oa.o_flags |= OBD_FL_SRVLOCK;
 247         }
 248
 249 skip_locking:
 250         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 251                                    &RQF_OST_GET_INFO_FIEMAP);
 252         if (req == NULL)
 253                 GOTO(drop_lock, rc = -ENOMEM);
 254
 255         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY, RCL_CLIENT,
 256                              sizeof(*fmkey));
 257         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_CLIENT,
 258                              *buflen);
 259         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_SERVER,
 260                              *buflen);
 261
 262         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 263         if (rc != 0) {
 264                 ptlrpc_request_free(req);
 265                 GOTO(drop_lock, rc);
 266         }
 267         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
 268         memcpy(tmp, fmkey, sizeof(*fmkey));
 269         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
 270         memcpy(tmp, fiemap, *buflen);
 271         ptlrpc_request_set_replen(req);
 272
 273         rc = ptlrpc_queue_wait(req);
 274         if (rc != 0)
 275                 GOTO(fini_req, rc);
 276
 277         reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
 278         if (reply == NULL)
 279                 GOTO(fini_req, rc = -EPROTO);
 280
 281         memcpy(fiemap, reply, *buflen);
 282 fini_req:
 283         ptlrpc_req_finished(req);
 284 drop_lock:
 285         if (mode)
 286                 ldlm_lock_decref(&lockh, LCK_PR);
 287         RETURN(rc);
 288 }
 289
 290 void osc_object_set_contended(struct osc_object *obj)
 291 {
 292         obj->oo_contention_time = cfs_time_current();
 293         /* mb(); */
 294         obj->oo_contended = 1;
 295 }
 296
 297 void osc_object_clear_contended(struct osc_object *obj)
 298 {
 299         obj->oo_contended = 0;
 300 }
 301
 302 int osc_object_is_contended(struct osc_object *obj)
 303 {
 304         struct osc_device *dev  = lu2osc_dev(obj->oo_cl.co_lu.lo_dev);
 305         int osc_contention_time = dev->od_contention_time;
 306         cfs_time_t cur_time     = cfs_time_current();
 307         cfs_time_t retry_time;
 308
 309         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_OBJECT_CONTENTION))
 310                 return 1;
 311
 312         if (!obj->oo_contended)
 313                 return 0;
 314
 315         /*
 316          * I like copy-paste. the code is copied from
 317          * ll_file_is_contended.
 318          */
 319         retry_time = cfs_time_add(obj->oo_contention_time,
 320                                   cfs_time_seconds(osc_contention_time));
 321         if (cfs_time_after(cur_time, retry_time)) {
 322                 osc_object_clear_contended(obj);
 323                 return 0;
 324         }
 325         return 1;
 326 }
 327
 328 /**
 329  * Implementation of struct cl_object_operations::coo_req_attr_set() for osc
 330  * layer. osc is responsible for struct obdo::o_id and struct obdo::o_seq
 331  * fields.
 332  */
 333 static void osc_req_attr_set(const struct lu_env *env, struct cl_object *obj,
 334                              struct cl_req_attr *attr)
 335 {
 336         struct lov_oinfo *oinfo;
 337         struct obdo      *oa;
 338         struct ost_lvb   *lvb;
 339         u64               flags = attr->cra_flags;
 340
 341         oinfo   = cl2osc(obj)->oo_oinfo;
 342         lvb     = &oinfo->loi_lvb;
 343         oa      = attr->cra_oa;
 344
 345         if ((flags & OBD_MD_FLMTIME) != 0) {
 346                 oa->o_mtime = lvb->lvb_mtime;
 347                 oa->o_valid |= OBD_MD_FLMTIME;
 348         }
 349         if ((flags & OBD_MD_FLATIME) != 0) {
 350                 oa->o_atime = lvb->lvb_atime;
 351                 oa->o_valid |= OBD_MD_FLATIME;
 352         }
 353         if ((flags & OBD_MD_FLCTIME) != 0) {
 354                 oa->o_ctime = lvb->lvb_ctime;
 355                 oa->o_valid |= OBD_MD_FLCTIME;
 356         }
 357         if (flags & OBD_MD_FLGROUP) {
 358                 ostid_set_seq(&oa->o_oi, ostid_seq(&oinfo->loi_oi));
 359                 oa->o_valid |= OBD_MD_FLGROUP;
 360         }
 361         if (flags & OBD_MD_FLID) {
 362                 int rc;
 363
 364                 rc = ostid_set_id(&oa->o_oi, ostid_id(&oinfo->loi_oi));
 365                 if (rc) {
 366                         CERROR("Bad %llu to set " DOSTID " : rc %d\n",
 367                                (unsigned long long)ostid_id(&oinfo->loi_oi),
 368                                POSTID(&oa->o_oi), rc);
 369                 }
 370                 oa->o_valid |= OBD_MD_FLID;
 371         }
 372         if (flags & OBD_MD_FLHANDLE) {
 373                 struct ldlm_lock *lock;
 374                 struct osc_page *opg;
 375
 376                 opg = osc_cl_page_osc(attr->cra_page, cl2osc(obj));
 377                 lock = osc_dlmlock_at_pgoff(env, cl2osc(obj), osc_index(opg),
 378                                 OSC_DAP_FL_TEST_LOCK | OSC_DAP_FL_CANCELING);
 379                 if (lock == NULL && !opg->ops_srvlock) {
 380                         struct ldlm_resource *res;
 381                         struct ldlm_res_id *resname;
 382
 383                         CL_PAGE_DEBUG(D_ERROR, env, attr->cra_page,
 384                                       "uncovered page!\n");
 385
 386                         resname = &osc_env_info(env)->oti_resname;
 387                         ostid_build_res_name(&oinfo->loi_oi, resname);
 388                         res = ldlm_resource_get(
 389                                 osc_export(cl2osc(obj))->exp_obd->obd_namespace,
 390                                 NULL, resname, LDLM_EXTENT, 0);
 391                         ldlm_resource_dump(D_ERROR, res);
 392
 393                         libcfs_debug_dumpstack(NULL);
 394                         LBUG();
 395                 }
 396
 397                 /* check for lockless io. */
 398                 if (lock != NULL) {
 399                         oa->o_handle = lock->l_remote_handle;
 400                         oa->o_valid |= OBD_MD_FLHANDLE;
 401                         LDLM_LOCK_PUT(lock);
 402                 }
 403         }
 404 }
 405
 406 static const struct cl_object_operations osc_ops = {
 407         .coo_page_init    = osc_page_init,
 408         .coo_lock_init    = osc_lock_init,
 409         .coo_io_init      = osc_io_init,
 410         .coo_attr_get     = osc_attr_get,
 411         .coo_attr_update  = osc_attr_update,
 412         .coo_glimpse      = osc_object_glimpse,
 413         .coo_prune        = osc_object_prune,
 414         .coo_fiemap       = osc_object_fiemap,
 415         .coo_req_attr_set = osc_req_attr_set
 416 };
 417
 418 static const struct lu_object_operations osc_lu_obj_ops = {
 419         .loo_object_init      = osc_object_init,
 420         .loo_object_release   = NULL,
 421         .loo_object_free      = osc_object_free,
 422         .loo_object_print     = osc_object_print,
 423         .loo_object_invariant = NULL
 424 };
 425
 426 struct lu_object *osc_object_alloc(const struct lu_env *env,
 427                                    const struct lu_object_header *unused,
 428                                    struct lu_device *dev)
 429 {
 430         struct osc_object *osc;
 431         struct lu_object  *obj;
 432
 433         OBD_SLAB_ALLOC_PTR_GFP(osc, osc_object_kmem, GFP_NOFS);
 434         if (osc != NULL) {
 435                 obj = osc2lu(osc);
 436                 lu_object_init(obj, NULL, dev);
 437                 osc->oo_cl.co_ops = &osc_ops;
 438                 obj->lo_ops = &osc_lu_obj_ops;
 439         } else
 440                 obj = NULL;
 441         return obj;
 442 }
 443
 444 int osc_object_invalidate(const struct lu_env *env, struct osc_object *osc)
 445 {
 446         struct l_wait_info lwi = { 0 };
 447         ENTRY;
 448
 449         CDEBUG(D_INODE, "Invalidate osc object: %p, # of active IOs: %d\n",
 450                osc, atomic_read(&osc->oo_nr_ios));
 451
 452         l_wait_event(osc->oo_io_waitq, atomic_read(&osc->oo_nr_ios) == 0, &lwi);
 453
 454         /* Discard all dirty pages of this object. */
 455         osc_cache_truncate_start(env, osc, 0, NULL);
 456
 457         /* Discard all caching pages */
 458         osc_lock_discard_pages(env, osc, 0, CL_PAGE_EOF, true);
 459
 460         /* Clear ast data of dlm lock. Do this after discarding all pages */
 461         osc_object_prune(env, osc2cl(osc));
 462
 463         RETURN(0);
 464 }
 465
 466 /** @} osc */