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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  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_object for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45
  46 /** \addtogroup osc
  47  *  @{
  48  */
  49
  50 /*****************************************************************************
  51  *
  52  * Type conversions.
  53  *
  54  */
  55
  56 static struct lu_object *osc2lu(struct osc_object *osc)
  57 {
  58         return &osc->oo_cl.co_lu;
  59 }
  60
  61 static struct osc_object *lu2osc(const struct lu_object *obj)
  62 {
  63         LINVRNT(osc_is_object(obj));
  64         return container_of0(obj, struct osc_object, oo_cl.co_lu);
  65 }
  66
  67 /*****************************************************************************
  68  *
  69  * Object operations.
  70  *
  71  */
  72
  73 static int osc_object_init(const struct lu_env *env, struct lu_object *obj,
  74                            const struct lu_object_conf *conf)
  75 {
  76         struct osc_object           *osc   = lu2osc(obj);
  77         const struct cl_object_conf *cconf = lu2cl_conf(conf);
  78
  79         osc->oo_oinfo = cconf->u.coc_oinfo;
  80 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
  81         mutex_init(&osc->oo_debug_mutex);
  82 #endif
  83         INIT_LIST_HEAD(&osc->oo_ready_item);
  84         INIT_LIST_HEAD(&osc->oo_hp_ready_item);
  85         INIT_LIST_HEAD(&osc->oo_write_item);
  86         INIT_LIST_HEAD(&osc->oo_read_item);
  87
  88         osc->oo_root.rb_node = NULL;
  89         INIT_LIST_HEAD(&osc->oo_hp_exts);
  90         INIT_LIST_HEAD(&osc->oo_urgent_exts);
  91         INIT_LIST_HEAD(&osc->oo_rpc_exts);
  92         INIT_LIST_HEAD(&osc->oo_reading_exts);
  93         atomic_set(&osc->oo_nr_reads, 0);
  94         atomic_set(&osc->oo_nr_writes, 0);
  95         spin_lock_init(&osc->oo_lock);
  96         spin_lock_init(&osc->oo_tree_lock);
  97         spin_lock_init(&osc->oo_ol_spin);
  98         INIT_LIST_HEAD(&osc->oo_ol_list);
  99
 100         atomic_set(&osc->oo_nr_ios, 0);
 101         init_waitqueue_head(&osc->oo_io_waitq);
 102
 103         cl_object_page_init(lu2cl(obj), sizeof(struct osc_page));
 104
 105         return 0;
 106 }
 107
 108 static void osc_object_free(const struct lu_env *env, struct lu_object *obj)
 109 {
 110         struct osc_object *osc = lu2osc(obj);
 111
 112         LASSERT(list_empty(&osc->oo_ready_item));
 113         LASSERT(list_empty(&osc->oo_hp_ready_item));
 114         LASSERT(list_empty(&osc->oo_write_item));
 115         LASSERT(list_empty(&osc->oo_read_item));
 116
 117         LASSERT(osc->oo_root.rb_node == NULL);
 118         LASSERT(list_empty(&osc->oo_hp_exts));
 119         LASSERT(list_empty(&osc->oo_urgent_exts));
 120         LASSERT(list_empty(&osc->oo_rpc_exts));
 121         LASSERT(list_empty(&osc->oo_reading_exts));
 122         LASSERT(atomic_read(&osc->oo_nr_reads) == 0);
 123         LASSERT(atomic_read(&osc->oo_nr_writes) == 0);
 124         LASSERT(list_empty(&osc->oo_ol_list));
 125         LASSERT(atomic_read(&osc->oo_nr_ios) == 0);
 126
 127         lu_object_fini(obj);
 128         OBD_SLAB_FREE_PTR(osc, osc_object_kmem);
 129 }
 130
 131 int osc_lvb_print(const struct lu_env *env, void *cookie,
 132                   lu_printer_t p, const struct ost_lvb *lvb)
 133 {
 134         return (*p)(env, cookie, "size: "LPU64" mtime: "LPU64" atime: "LPU64" "
 135                     "ctime: "LPU64" blocks: "LPU64,
 136                     lvb->lvb_size, lvb->lvb_mtime, lvb->lvb_atime,
 137                     lvb->lvb_ctime, lvb->lvb_blocks);
 138 }
 139
 140 static 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 "LPU64" "
 149              "rc: %d force_sync: %d min_xid: "LPU64" ",
 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
 157
 158 static int osc_attr_get(const struct lu_env *env, struct cl_object *obj,
 159                         struct cl_attr *attr)
 160 {
 161         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 162
 163         cl_lvb2attr(attr, &oinfo->loi_lvb);
 164         attr->cat_kms = oinfo->loi_kms_valid ? oinfo->loi_kms : 0;
 165         return 0;
 166 }
 167
 168 static int osc_attr_update(const struct lu_env *env, struct cl_object *obj,
 169                            const struct cl_attr *attr, unsigned valid)
 170 {
 171         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 172         struct ost_lvb   *lvb   = &oinfo->loi_lvb;
 173
 174         if (valid & CAT_SIZE)
 175                 lvb->lvb_size = attr->cat_size;
 176         if (valid & CAT_MTIME)
 177                 lvb->lvb_mtime = attr->cat_mtime;
 178         if (valid & CAT_ATIME)
 179                 lvb->lvb_atime = attr->cat_atime;
 180         if (valid & CAT_CTIME)
 181                 lvb->lvb_ctime = attr->cat_ctime;
 182         if (valid & CAT_BLOCKS)
 183                 lvb->lvb_blocks = attr->cat_blocks;
 184         if (valid & CAT_KMS) {
 185                 CDEBUG(D_CACHE, "set kms from "LPU64"to "LPU64"\n",
 186                        oinfo->loi_kms, (__u64)attr->cat_kms);
 187                 loi_kms_set(oinfo, attr->cat_kms);
 188         }
 189         return 0;
 190 }
 191
 192 static int osc_object_glimpse(const struct lu_env *env,
 193                               const struct cl_object *obj, struct ost_lvb *lvb)
 194 {
 195         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
 196
 197         ENTRY;
 198         lvb->lvb_size   = oinfo->loi_kms;
 199         lvb->lvb_blocks = oinfo->loi_lvb.lvb_blocks;
 200         RETURN(0);
 201 }
 202
 203 static int osc_object_ast_clear(struct ldlm_lock *lock, void *data)
 204 {
 205         ENTRY;
 206
 207         if (lock->l_ast_data == data)
 208                 lock->l_ast_data = NULL;
 209         RETURN(LDLM_ITER_CONTINUE);
 210 }
 211
 212 static int osc_object_prune(const struct lu_env *env, struct cl_object *obj)
 213 {
 214         struct osc_object       *osc = cl2osc(obj);
 215         struct ldlm_res_id      *resname = &osc_env_info(env)->oti_resname;
 216
 217         LASSERTF(osc->oo_npages == 0,
 218                  DFID "still have %lu pages, obj: %p, osc: %p\n",
 219                  PFID(lu_object_fid(&obj->co_lu)), osc->oo_npages, obj, osc);
 220
 221         /* DLM locks don't hold a reference of osc_object so we have to
 222          * clear it before the object is being destroyed. */
 223         ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
 224         ldlm_resource_iterate(osc_export(osc)->exp_obd->obd_namespace, resname,
 225                               osc_object_ast_clear, osc);
 226         return 0;
 227 }
 228 /**
 229  * Find any ldlm lock covers the osc object.
 230  * \retval 0    not found
 231  * \retval 1    find one
 232  * \retval < 0  error
 233  */
 234 static int osc_object_find_cbdata(const struct lu_env *env,
 235                                   struct cl_object *obj, ldlm_iterator_t iter,
 236                                   void *data)
 237 {
 238         struct ldlm_res_id              res_id;
 239         struct obd_device               *obd;
 240         int                             rc = 0;
 241
 242         ostid_build_res_name(&cl2osc(obj)->oo_oinfo->loi_oi, &res_id);
 243         obd = obj->co_lu.lo_dev->ld_obd;
 244         rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, iter, data);
 245         if (rc == LDLM_ITER_STOP)
 246                 return 1;
 247
 248         if (rc == LDLM_ITER_CONTINUE)
 249                 return 0;
 250
 251         return rc;
 252 }
 253
 254 static int osc_object_fiemap(const struct lu_env *env, struct cl_object *obj,
 255                              struct ll_fiemap_info_key *fmkey,
 256                              struct fiemap *fiemap, size_t *buflen)
 257 {
 258         struct obd_export *exp = osc_export(cl2osc(obj));
 259         struct ldlm_res_id resid;
 260         union ldlm_policy_data policy;
 261         struct lustre_handle lockh;
 262         enum ldlm_mode mode = LCK_MINMODE;
 263         struct ptlrpc_request *req;
 264         struct fiemap *reply;
 265         char *tmp;
 266         int rc;
 267         ENTRY;
 268
 269         fmkey->lfik_oa.o_oi = cl2osc(obj)->oo_oinfo->loi_oi;
 270         if (!(fmkey->lfik_fiemap.fm_flags & FIEMAP_FLAG_SYNC))
 271                 goto skip_locking;
 272
 273         policy.l_extent.start = fmkey->lfik_fiemap.fm_start & PAGE_CACHE_MASK;
 274
 275         if (OBD_OBJECT_EOF - fmkey->lfik_fiemap.fm_length <=
 276             fmkey->lfik_fiemap.fm_start + PAGE_CACHE_SIZE - 1)
 277                 policy.l_extent.end = OBD_OBJECT_EOF;
 278         else
 279                 policy.l_extent.end = (fmkey->lfik_fiemap.fm_start +
 280                                        fmkey->lfik_fiemap.fm_length +
 281                                        PAGE_CACHE_SIZE - 1) & PAGE_CACHE_MASK;
 282
 283         ostid_build_res_name(&fmkey->lfik_oa.o_oi, &resid);
 284         mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
 285                                LDLM_FL_BLOCK_GRANTED | LDLM_FL_LVB_READY,
 286                                &resid, LDLM_EXTENT, &policy,
 287                                LCK_PR | LCK_PW, &lockh, 0);
 288         if (mode) { /* lock is cached on client */
 289                 if (mode != LCK_PR) {
 290                         ldlm_lock_addref(&lockh, LCK_PR);
 291                         ldlm_lock_decref(&lockh, LCK_PW);
 292                 }
 293         } else { /* no cached lock, needs acquire lock on server side */
 294                 fmkey->lfik_oa.o_valid |= OBD_MD_FLFLAGS;
 295                 fmkey->lfik_oa.o_flags |= OBD_FL_SRVLOCK;
 296         }
 297
 298 skip_locking:
 299         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 300                                    &RQF_OST_GET_INFO_FIEMAP);
 301         if (req == NULL)
 302                 GOTO(drop_lock, rc = -ENOMEM);
 303
 304         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY, RCL_CLIENT,
 305                              sizeof(*fmkey));
 306         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_CLIENT,
 307                              *buflen);
 308         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_SERVER,
 309                              *buflen);
 310
 311         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 312         if (rc != 0) {
 313                 ptlrpc_request_free(req);
 314                 GOTO(drop_lock, rc);
 315         }
 316         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
 317         memcpy(tmp, fmkey, sizeof(*fmkey));
 318         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
 319         memcpy(tmp, fiemap, *buflen);
 320         ptlrpc_request_set_replen(req);
 321
 322         rc = ptlrpc_queue_wait(req);
 323         if (rc != 0)
 324                 GOTO(fini_req, rc);
 325
 326         reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
 327         if (reply == NULL)
 328                 GOTO(fini_req, rc = -EPROTO);
 329
 330         memcpy(fiemap, reply, *buflen);
 331 fini_req:
 332         ptlrpc_req_finished(req);
 333 drop_lock:
 334         if (mode)
 335                 ldlm_lock_decref(&lockh, LCK_PR);
 336         RETURN(rc);
 337 }
 338
 339 void osc_object_set_contended(struct osc_object *obj)
 340 {
 341         obj->oo_contention_time = cfs_time_current();
 342         /* mb(); */
 343         obj->oo_contended = 1;
 344 }
 345
 346 void osc_object_clear_contended(struct osc_object *obj)
 347 {
 348         obj->oo_contended = 0;
 349 }
 350
 351 int osc_object_is_contended(struct osc_object *obj)
 352 {
 353         struct osc_device *dev  = lu2osc_dev(obj->oo_cl.co_lu.lo_dev);
 354         int osc_contention_time = dev->od_contention_time;
 355         cfs_time_t cur_time     = cfs_time_current();
 356         cfs_time_t retry_time;
 357
 358         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_OBJECT_CONTENTION))
 359                 return 1;
 360
 361         if (!obj->oo_contended)
 362                 return 0;
 363
 364         /*
 365          * I like copy-paste. the code is copied from
 366          * ll_file_is_contended.
 367          */
 368         retry_time = cfs_time_add(obj->oo_contention_time,
 369                                   cfs_time_seconds(osc_contention_time));
 370         if (cfs_time_after(cur_time, retry_time)) {
 371                 osc_object_clear_contended(obj);
 372                 return 0;
 373         }
 374         return 1;
 375 }
 376
 377 /**
 378  * Implementation of struct cl_object_operations::coo_req_attr_set() for osc
 379  * layer. osc is responsible for struct obdo::o_id and struct obdo::o_seq
 380  * fields.
 381  */
 382 static void osc_req_attr_set(const struct lu_env *env, struct cl_object *obj,
 383                              struct cl_req_attr *attr)
 384 {
 385         struct lov_oinfo *oinfo;
 386         struct obdo      *oa;
 387         struct ost_lvb   *lvb;
 388         u64               flags = attr->cra_flags;
 389
 390         oinfo   = cl2osc(obj)->oo_oinfo;
 391         lvb     = &oinfo->loi_lvb;
 392         oa      = attr->cra_oa;
 393
 394         if ((flags & OBD_MD_FLMTIME) != 0) {
 395                 oa->o_mtime = lvb->lvb_mtime;
 396                 oa->o_valid |= OBD_MD_FLMTIME;
 397         }
 398         if ((flags & OBD_MD_FLATIME) != 0) {
 399                 oa->o_atime = lvb->lvb_atime;
 400                 oa->o_valid |= OBD_MD_FLATIME;
 401         }
 402         if ((flags & OBD_MD_FLCTIME) != 0) {
 403                 oa->o_ctime = lvb->lvb_ctime;
 404                 oa->o_valid |= OBD_MD_FLCTIME;
 405         }
 406         if (flags & OBD_MD_FLGROUP) {
 407                 ostid_set_seq(&oa->o_oi, ostid_seq(&oinfo->loi_oi));
 408                 oa->o_valid |= OBD_MD_FLGROUP;
 409         }
 410         if (flags & OBD_MD_FLID) {
 411                 ostid_set_id(&oa->o_oi, ostid_id(&oinfo->loi_oi));
 412                 oa->o_valid |= OBD_MD_FLID;
 413         }
 414         if (flags & OBD_MD_FLHANDLE) {
 415                 struct ldlm_lock *lock;
 416                 struct osc_page *opg;
 417
 418                 opg = osc_cl_page_osc(attr->cra_page, cl2osc(obj));
 419                 lock = osc_dlmlock_at_pgoff(env, cl2osc(obj), osc_index(opg),
 420                                 OSC_DAP_FL_TEST_LOCK | OSC_DAP_FL_CANCELING);
 421                 if (lock == NULL && !opg->ops_srvlock) {
 422                         struct ldlm_resource *res;
 423                         struct ldlm_res_id *resname;
 424
 425                         CL_PAGE_DEBUG(D_ERROR, env, attr->cra_page,
 426                                       "uncovered page!\n");
 427
 428                         resname = &osc_env_info(env)->oti_resname;
 429                         ostid_build_res_name(&oinfo->loi_oi, resname);
 430                         res = ldlm_resource_get(
 431                                 osc_export(cl2osc(obj))->exp_obd->obd_namespace,
 432                                 NULL, resname, LDLM_EXTENT, 0);
 433                         ldlm_resource_dump(D_ERROR, res);
 434
 435                         libcfs_debug_dumpstack(NULL);
 436                         LBUG();
 437                 }
 438
 439                 /* check for lockless io. */
 440                 if (lock != NULL) {
 441                         oa->o_handle = lock->l_remote_handle;
 442                         oa->o_valid |= OBD_MD_FLHANDLE;
 443                         LDLM_LOCK_PUT(lock);
 444                 }
 445         }
 446 }
 447
 448 static const struct cl_object_operations osc_ops = {
 449         .coo_page_init    = osc_page_init,
 450         .coo_lock_init    = osc_lock_init,
 451         .coo_io_init      = osc_io_init,
 452         .coo_attr_get     = osc_attr_get,
 453         .coo_attr_update  = osc_attr_update,
 454         .coo_glimpse      = osc_object_glimpse,
 455         .coo_prune        = osc_object_prune,
 456         .coo_find_cbdata  = osc_object_find_cbdata,
 457         .coo_fiemap       = osc_object_fiemap,
 458         .coo_req_attr_set = osc_req_attr_set
 459 };
 460
 461 static const struct lu_object_operations osc_lu_obj_ops = {
 462         .loo_object_init      = osc_object_init,
 463         .loo_object_release   = NULL,
 464         .loo_object_free      = osc_object_free,
 465         .loo_object_print     = osc_object_print,
 466         .loo_object_invariant = NULL
 467 };
 468
 469 struct lu_object *osc_object_alloc(const struct lu_env *env,
 470                                    const struct lu_object_header *unused,
 471                                    struct lu_device *dev)
 472 {
 473         struct osc_object *osc;
 474         struct lu_object  *obj;
 475
 476         OBD_SLAB_ALLOC_PTR_GFP(osc, osc_object_kmem, GFP_NOFS);
 477         if (osc != NULL) {
 478                 obj = osc2lu(osc);
 479                 lu_object_init(obj, NULL, dev);
 480                 osc->oo_cl.co_ops = &osc_ops;
 481                 obj->lo_ops = &osc_lu_obj_ops;
 482         } else
 483                 obj = NULL;
 484         return obj;
 485 }
 486
 487 int osc_object_invalidate(const struct lu_env *env, struct osc_object *osc)
 488 {
 489         struct l_wait_info lwi = { 0 };
 490         ENTRY;
 491
 492         CDEBUG(D_INODE, "Invalidate osc object: %p, # of active IOs: %d\n",
 493                osc, atomic_read(&osc->oo_nr_ios));
 494
 495         l_wait_event(osc->oo_io_waitq, atomic_read(&osc->oo_nr_ios) == 0, &lwi);
 496
 497         /* Discard all dirty pages of this object. */
 498         osc_cache_truncate_start(env, osc, 0, NULL);
 499
 500         /* Discard all caching pages */
 501         osc_lock_discard_pages(env, osc, 0, CL_PAGE_EOF, CLM_WRITE);
 502
 503         /* Clear ast data of dlm lock. Do this after discarding all pages */
 504         osc_object_prune(env, osc2cl(osc));
 505
 506         RETURN(0);
 507 }
 508
 509 /** @} osc */