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  *
  31  * Implementation of cl_object for OSC layer.
  32  *
  33  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  34  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSC
  38 #include <lustre_osc.h>
  39
  40 #include "osc_internal.h"
  41
  42 /** \addtogroup osc
  43  *  @{
  44  */
  45
  46 /*****************************************************************************
  47  *
  48  * Object operations.
  49  *
  50  */
  51 static void osc_obj_build_res_name(struct osc_object *osc,
  52                                    struct ldlm_res_id *resname)
  53 {
  54         ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
  55 }
  56
  57 static const struct osc_object_operations osc_object_ops = {
  58         .oto_build_res_name = osc_obj_build_res_name,
  59         .oto_dlmlock_at_pgoff = osc_obj_dlmlock_at_pgoff,
  60 };
  61
  62 int osc_object_init(const struct lu_env *env, struct lu_object *obj,
  63                     const struct lu_object_conf *conf)
  64 {
  65         struct osc_object           *osc   = lu2osc(obj);
  66         const struct cl_object_conf *cconf = lu2cl_conf(conf);
  67
  68         osc->oo_oinfo = cconf->u.coc_oinfo;
  69 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
  70         mutex_init(&osc->oo_debug_mutex);
  71 #endif
  72         INIT_LIST_HEAD(&osc->oo_ready_item);
  73         INIT_LIST_HEAD(&osc->oo_hp_ready_item);
  74         INIT_LIST_HEAD(&osc->oo_write_item);
  75         INIT_LIST_HEAD(&osc->oo_read_item);
  76
  77         osc->oo_root.rb_node = NULL;
  78         INIT_LIST_HEAD(&osc->oo_hp_exts);
  79         INIT_LIST_HEAD(&osc->oo_urgent_exts);
  80         INIT_LIST_HEAD(&osc->oo_full_exts);
  81         INIT_LIST_HEAD(&osc->oo_reading_exts);
  82         atomic_set(&osc->oo_nr_reads, 0);
  83         atomic_set(&osc->oo_nr_writes, 0);
  84         spin_lock_init(&osc->oo_lock);
  85         spin_lock_init(&osc->oo_tree_lock);
  86         spin_lock_init(&osc->oo_ol_spin);
  87         INIT_LIST_HEAD(&osc->oo_ol_list);
  88
  89         atomic_set(&osc->oo_nr_ios, 0);
  90         init_waitqueue_head(&osc->oo_io_waitq);
  91         init_waitqueue_head(&osc->oo_group_waitq);
  92         mutex_init(&osc->oo_group_mutex);
  93         osc->oo_group_users = 0;
  94         osc->oo_group_gid = 0;
  95
  96         LASSERT(osc->oo_obj_ops != NULL);
  97
  98         cl_object_page_init(lu2cl(obj), sizeof(struct osc_page));
  99
 100         return 0;
 101 }
 102 EXPORT_SYMBOL(osc_object_init);
 103
 104 void osc_object_free(const struct lu_env *env, struct lu_object *obj)
 105 {
 106         struct osc_object *osc = lu2osc(obj);
 107
 108         LASSERT(list_empty(&osc->oo_ready_item));
 109         LASSERT(list_empty(&osc->oo_hp_ready_item));
 110         LASSERT(list_empty(&osc->oo_write_item));
 111         LASSERT(list_empty(&osc->oo_read_item));
 112
 113         LASSERT(osc->oo_root.rb_node == NULL);
 114         LASSERT(list_empty(&osc->oo_hp_exts));
 115         LASSERT(list_empty(&osc->oo_urgent_exts));
 116         LASSERT(list_empty(&osc->oo_full_exts));
 117         LASSERT(list_empty(&osc->oo_reading_exts));
 118         LASSERT(atomic_read(&osc->oo_nr_reads) == 0);
 119         LASSERT(atomic_read(&osc->oo_nr_writes) == 0);
 120         LASSERT(list_empty(&osc->oo_ol_list));
 121         LASSERT(atomic_read(&osc->oo_nr_ios) == 0);
 122         LASSERT(osc->oo_group_users == 0);
 123
 124         lu_object_fini(obj);
 125         /* osc doen't contain an lu_object_header, so we don't need call_rcu */
 126         OBD_SLAB_FREE_PTR(osc, osc_object_kmem);
 127 }
 128 EXPORT_SYMBOL(osc_object_free);
 129
 130 int osc_lvb_print(const struct lu_env *env, void *cookie,
 131                   lu_printer_t p, const struct ost_lvb *lvb)
 132 {
 133         return (*p)(env, cookie, "size: %llu mtime: %llu atime: %llu "
 134                     "ctime: %llu blocks: %llu",
 135                     lvb->lvb_size, lvb->lvb_mtime, lvb->lvb_atime,
 136                     lvb->lvb_ctime, lvb->lvb_blocks);
 137 }
 138 EXPORT_SYMBOL(osc_lvb_print);
 139
 140 int osc_object_print(const struct lu_env *env, void *cookie,
 141                      lu_printer_t p, const struct lu_object *obj)
 142 {
 143         struct osc_object *osc = lu2osc(obj);
 144         struct lov_oinfo *oinfo = osc->oo_oinfo;
 145         struct osc_async_rc *ar = &oinfo->loi_ar;
 146
 147         (*p)(env, cookie, "id: "DOSTID" "
 148              "idx: %d gen: %d kms_valid: %u kms %llu "
 149              "rc: %d force_sync: %d min_xid: %llu ",
 150              POSTID(&oinfo->loi_oi), oinfo->loi_ost_idx,
 151              oinfo->loi_ost_gen, oinfo->loi_kms_valid, oinfo->loi_kms,
 152              ar->ar_rc, ar->ar_force_sync, ar->ar_min_xid);
 153         osc_lvb_print(env, cookie, p, &oinfo->loi_lvb);
 154         return 0;
 155 }
 156 EXPORT_SYMBOL(osc_object_print);
 157
 158
 159 int osc_attr_get(const struct lu_env *env, struct cl_object *obj,
 160                  struct cl_attr *attr)
 161 {
 162         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 163
 164         cl_lvb2attr(attr, &oinfo->loi_lvb);
 165         attr->cat_kms = oinfo->loi_kms_valid ? oinfo->loi_kms : 0;
 166         return 0;
 167 }
 168 EXPORT_SYMBOL(osc_attr_get);
 169
 170 int osc_attr_update(const struct lu_env *env, struct cl_object *obj,
 171                     const struct cl_attr *attr, unsigned valid)
 172 {
 173         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 174         struct ost_lvb   *lvb   = &oinfo->loi_lvb;
 175
 176         if (valid & CAT_SIZE)
 177                 lvb->lvb_size = attr->cat_size;
 178         if (valid & CAT_MTIME)
 179                 lvb->lvb_mtime = attr->cat_mtime;
 180         if (valid & CAT_ATIME)
 181                 lvb->lvb_atime = attr->cat_atime;
 182         if (valid & CAT_CTIME)
 183                 lvb->lvb_ctime = attr->cat_ctime;
 184         if (valid & CAT_BLOCKS)
 185                 lvb->lvb_blocks = attr->cat_blocks;
 186         if (valid & CAT_KMS) {
 187                 CDEBUG(D_CACHE, "set kms from %lluto %llu\n",
 188                        oinfo->loi_kms, (__u64)attr->cat_kms);
 189                 loi_kms_set(oinfo, attr->cat_kms);
 190         }
 191         return 0;
 192 }
 193 EXPORT_SYMBOL(osc_attr_update);
 194
 195 int osc_object_glimpse(const struct lu_env *env, const struct cl_object *obj,
 196                        struct ost_lvb *lvb)
 197 {
 198         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 199
 200         lvb->lvb_size = oinfo->loi_kms;
 201         lvb->lvb_blocks = oinfo->loi_lvb.lvb_blocks;
 202         return 0;
 203 }
 204 EXPORT_SYMBOL(osc_object_glimpse);
 205
 206 static int osc_object_ast_clear(struct ldlm_lock *lock, void *data)
 207 {
 208         struct osc_object *osc = (struct osc_object *)data;
 209         struct ost_lvb *lvb = lock->l_lvb_data;
 210         struct lov_oinfo *oinfo;
 211         ENTRY;
 212
 213         if (lock->l_ast_data == data) {
 214                 lock->l_ast_data = NULL;
 215
 216                 LASSERT(osc != NULL);
 217                 LASSERT(osc->oo_oinfo != NULL);
 218                 LASSERT(lvb != NULL);
 219
 220                 /* Updates lvb in lock by the cached oinfo */
 221                 oinfo = osc->oo_oinfo;
 222
 223                 LDLM_DEBUG(lock, "update lock size %llu blocks %llu [cma]time: "
 224                            "%llu %llu %llu by oinfo size %llu blocks %llu "
 225                            "[cma]time %llu %llu %llu", lvb->lvb_size,
 226                            lvb->lvb_blocks, lvb->lvb_ctime, lvb->lvb_mtime,
 227                            lvb->lvb_atime, oinfo->loi_lvb.lvb_size,
 228                            oinfo->loi_lvb.lvb_blocks, oinfo->loi_lvb.lvb_ctime,
 229                            oinfo->loi_lvb.lvb_mtime, oinfo->loi_lvb.lvb_atime);
 230                 LASSERTF(oinfo->loi_lvb.lvb_size >= oinfo->loi_kms,
 231                          "lvb_size %#llx, loi_kms %#llx\n",
 232                          oinfo->loi_lvb.lvb_size, oinfo->loi_kms);
 233
 234                 cl_object_attr_lock(&osc->oo_cl);
 235                 memcpy(lvb, &oinfo->loi_lvb, sizeof(oinfo->loi_lvb));
 236                 cl_object_attr_unlock(&osc->oo_cl);
 237                 ldlm_clear_lvb_cached(lock);
 238
 239                 /**
 240                  * Object is being destroyed and gets unlinked from the lock,
 241                  * IO is finished and no cached data is left under the lock. As
 242                  * grouplock is immediately marked CBPENDING it is not reused.
 243                  * It will also be not possible to flush data later due to a
 244                  * NULL l_ast_data - enough conditions to let new grouplocks to
 245                  * be enqueued even if the lock still exists on client.
 246                  */
 247                 if (lock->l_req_mode == LCK_GROUP)
 248                         osc_grouplock_dec(osc, lock);
 249         }
 250         RETURN(LDLM_ITER_CONTINUE);
 251 }
 252
 253 int osc_object_prune(const struct lu_env *env, struct cl_object *obj)
 254 {
 255         struct osc_object  *osc = cl2osc(obj);
 256         struct ldlm_res_id *resname = &osc_env_info(env)->oti_resname;
 257
 258         /* DLM locks don't hold a reference of osc_object so we have to
 259          * clear it before the object is being destroyed. */
 260         osc_build_res_name(osc, resname);
 261         ldlm_resource_iterate(osc_export(osc)->exp_obd->obd_namespace, resname,
 262                               osc_object_ast_clear, osc);
 263         return 0;
 264 }
 265 EXPORT_SYMBOL(osc_object_prune);
 266
 267 static int osc_object_fiemap(const struct lu_env *env, struct cl_object *obj,
 268                              struct ll_fiemap_info_key *fmkey,
 269                              struct fiemap *fiemap, size_t *buflen)
 270 {
 271         struct obd_export *exp = osc_export(cl2osc(obj));
 272         struct ldlm_res_id resid;
 273         union ldlm_policy_data policy;
 274         struct lustre_handle lockh;
 275         enum ldlm_mode mode = LCK_MINMODE;
 276         struct ptlrpc_request *req;
 277         struct fiemap *reply;
 278         char *tmp;
 279         int rc;
 280         ENTRY;
 281
 282         fmkey->lfik_oa.o_oi = cl2osc(obj)->oo_oinfo->loi_oi;
 283         if (!(fmkey->lfik_fiemap.fm_flags & FIEMAP_FLAG_SYNC))
 284                 goto skip_locking;
 285
 286         policy.l_extent.start = fmkey->lfik_fiemap.fm_start & PAGE_MASK;
 287
 288         if (OBD_OBJECT_EOF - fmkey->lfik_fiemap.fm_length <=
 289             fmkey->lfik_fiemap.fm_start + PAGE_SIZE - 1)
 290                 policy.l_extent.end = OBD_OBJECT_EOF;
 291         else
 292                 policy.l_extent.end = (fmkey->lfik_fiemap.fm_start +
 293                                        fmkey->lfik_fiemap.fm_length +
 294                                        PAGE_SIZE - 1) & PAGE_MASK;
 295
 296         ostid_build_res_name(&fmkey->lfik_oa.o_oi, &resid);
 297         mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
 298                                LDLM_FL_BLOCK_GRANTED | LDLM_FL_LVB_READY,
 299                                &resid, LDLM_EXTENT, &policy,
 300                                LCK_PR | LCK_PW, &lockh);
 301         if (mode) { /* lock is cached on client */
 302                 if (mode != LCK_PR) {
 303                         ldlm_lock_addref(&lockh, LCK_PR);
 304                         ldlm_lock_decref(&lockh, LCK_PW);
 305                 }
 306         } else { /* no cached lock, needs acquire lock on server side */
 307                 fmkey->lfik_oa.o_valid |= OBD_MD_FLFLAGS;
 308                 fmkey->lfik_oa.o_flags |= OBD_FL_SRVLOCK;
 309         }
 310
 311 skip_locking:
 312         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 313                                    &RQF_OST_GET_INFO_FIEMAP);
 314         if (req == NULL)
 315                 GOTO(drop_lock, rc = -ENOMEM);
 316
 317         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY, RCL_CLIENT,
 318                              sizeof(*fmkey));
 319         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_CLIENT,
 320                              *buflen);
 321         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_SERVER,
 322                              *buflen);
 323
 324         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 325         if (rc != 0) {
 326                 ptlrpc_request_free(req);
 327                 GOTO(drop_lock, rc);
 328         }
 329         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
 330         memcpy(tmp, fmkey, sizeof(*fmkey));
 331         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
 332         memcpy(tmp, fiemap, *buflen);
 333         ptlrpc_request_set_replen(req);
 334
 335         rc = ptlrpc_queue_wait(req);
 336         if (rc != 0)
 337                 GOTO(fini_req, rc);
 338
 339         reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
 340         if (reply == NULL)
 341                 GOTO(fini_req, rc = -EPROTO);
 342
 343         memcpy(fiemap, reply, *buflen);
 344 fini_req:
 345         ptlrpc_req_finished(req);
 346 drop_lock:
 347         if (mode)
 348                 ldlm_lock_decref(&lockh, LCK_PR);
 349         RETURN(rc);
 350 }
 351
 352 /**
 353  * Implementation of struct cl_object_operations::coo_req_attr_set() for osc
 354  * layer. osc is responsible for struct obdo::o_id and struct obdo::o_seq
 355  * fields.
 356  */
 357 static void osc_req_attr_set(const struct lu_env *env, struct cl_object *obj,
 358                              struct cl_req_attr *attr)
 359 {
 360         struct lov_oinfo *oinfo;
 361         struct obdo      *oa;
 362         struct ost_lvb   *lvb;
 363         u64               flags = attr->cra_flags;
 364
 365         oinfo   = cl2osc(obj)->oo_oinfo;
 366         lvb     = &oinfo->loi_lvb;
 367         oa      = attr->cra_oa;
 368
 369         if ((flags & OBD_MD_FLMTIME) != 0) {
 370                 oa->o_mtime = lvb->lvb_mtime;
 371                 oa->o_valid |= OBD_MD_FLMTIME;
 372         }
 373         /* XXX:
 374          * I don't understand this part, what for OSC resets atime just
 375          * set by VVP layer to 0 so that OST gets 0 instead of actual
 376          * atime, bzzz. please inspect this place with extra care.
 377          */
 378         if ((flags & OBD_MD_FLATIME) && lvb->lvb_atime > oa->o_atime) {
 379                 oa->o_atime = lvb->lvb_atime;
 380                 oa->o_valid |= OBD_MD_FLATIME;
 381         }
 382         if ((flags & OBD_MD_FLCTIME) != 0) {
 383                 oa->o_ctime = lvb->lvb_ctime;
 384                 oa->o_valid |= OBD_MD_FLCTIME;
 385         }
 386         if (flags & OBD_MD_FLGROUP) {
 387                 ostid_set_seq(&oa->o_oi, ostid_seq(&oinfo->loi_oi));
 388                 oa->o_valid |= OBD_MD_FLGROUP;
 389         }
 390         if (flags & OBD_MD_FLID) {
 391                 int rc;
 392
 393                 rc = ostid_set_id(&oa->o_oi, ostid_id(&oinfo->loi_oi));
 394                 if (rc) {
 395                         CERROR("Bad %llu to set " DOSTID " : rc %d\n",
 396                                (unsigned long long)ostid_id(&oinfo->loi_oi),
 397                                POSTID(&oa->o_oi), rc);
 398                 }
 399                 oa->o_valid |= OBD_MD_FLID;
 400         }
 401         if (flags & OBD_MD_FLHANDLE) {
 402                 struct ldlm_lock *lock;
 403                 struct osc_page *opg;
 404
 405                 opg = osc_cl_page_osc(attr->cra_page, cl2osc(obj));
 406                 lock = osc_dlmlock_at_pgoff(env, cl2osc(obj), osc_index(opg),
 407                                 OSC_DAP_FL_TEST_LOCK | OSC_DAP_FL_CANCELING);
 408                 if (lock == NULL && !opg->ops_srvlock) {
 409                         struct ldlm_resource *res;
 410                         struct ldlm_res_id *resname;
 411
 412                         CL_PAGE_DEBUG(D_ERROR, env, attr->cra_page,
 413                                       "uncovered page!\n");
 414
 415                         resname = &osc_env_info(env)->oti_resname;
 416                         ostid_build_res_name(&oinfo->loi_oi, resname);
 417                         res = ldlm_resource_get(osc_export(cl2osc(obj))->
 418                                                         exp_obd->obd_namespace,
 419                                                 resname, LDLM_EXTENT, 0);
 420                         if (IS_ERR(res))
 421                                 CERROR("No lock resource\n");
 422                         else
 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         ENTRY;
 480
 481         CDEBUG(D_INODE, "Invalidate osc object: %p, # of active IOs: %d\n",
 482                osc, atomic_read(&osc->oo_nr_ios));
 483
 484         wait_event_idle(osc->oo_io_waitq, atomic_read(&osc->oo_nr_ios) == 0);
 485
 486         /* Discard all dirty pages of this object. */
 487         osc_cache_truncate_start(env, osc, 0, NULL);
 488
 489         /* Discard all caching pages */
 490         osc_lock_discard_pages(env, osc, 0, CL_PAGE_EOF, true);
 491
 492         /* Clear ast data of dlm lock. Do this after discarding all pages */
 493         cl_object_prune(env, osc2cl(osc));
 494
 495         RETURN(0);
 496 }
 497 EXPORT_SYMBOL(osc_object_invalidate);
 498 /** @} osc */