lustre/llite/pcc.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) 2017, DDN Storage Corporation.
  24  */
  25 /*
  26  * Persistent Client Cache
  27  *
  28  * PCC is a new framework which provides a group of local cache on Lustre
  29  * client side. It works in two modes: RW-PCC enables a read-write cache on the
  30  * local SSDs of a single client; RO-PCC provides a read-only cache on the
  31  * local SSDs of multiple clients. Less overhead is visible to the applications
  32  * and network latencies and lock conflicts can be significantly reduced.
  33  *
  34  * For RW-PCC, no global namespace will be provided. Each client uses its own
  35  * local storage as a cache for itself. Local file system is used to manage
  36  * the data on local caches. Cached I/O is directed to local file system while
  37  * normal I/O is directed to OSTs. RW-PCC uses HSM for data synchronization.
  38  * It uses HSM copytool to restore file from local caches to Lustre OSTs. Each
  39  * PCC has a copytool instance running with unique archive number. Any remote
  40  * access from another Lustre client would trigger the data synchronization. If
  41  * a client with RW-PCC goes offline, the cached data becomes inaccessible for
  42  * other client temporarily. And after the RW-PCC client reboots and the
  43  * copytool restarts, the data will be accessible again.
  44  *
  45  * Following is what will happen in different conditions for RW-PCC:
  46  *
  47  * > When file is being created on RW-PCC
  48  *
  49  * A normal HSM released file is created on MDT;
  50  * An empty mirror file is created on local cache;
  51  * The HSM status of the Lustre file will be set to archived and released;
  52  * The archive number will be set to the proper value.
  53  *
  54  * > When file is being prefetched to RW-PCC
  55  *
  56  * An file is copied to the local cache;
  57  * The HSM status of the Lustre file will be set to archived and released;
  58  * The archive number will be set to the proper value.
  59  *
  60  * > When file is being accessed from PCC
  61  *
  62  * Data will be read directly from local cache;
  63  * Metadata will be read from MDT, except file size;
  64  * File size will be got from local cache.
  65  *
  66  * > When PCC cached file is being accessed on another client
  67  *
  68  * RW-PCC cached files are automatically restored when a process on another
  69  * client tries to read or modify them. The corresponding I/O will block
  70  * waiting for the released file to be restored. This is transparent to the
  71  * process.
  72  *
  73  * For RW-PCC, when a file is being created, a rule-based policy is used to
  74  * determine whether it will be cached. Rule-based caching of newly created
  75  * files can determine which file can use a cache on PCC directly without any
  76  * admission control.
  77  *
  78  * RW-PCC design can accelerate I/O intensive applications with one-to-one
  79  * mappings between files and accessing clients. However, in several use cases,
  80  * files will never be updated, but need to be read simultaneously from many
  81  * clients. RO-PCC implements a read-only caching on Lustre clients using
  82  * SSDs. RO-PCC is based on the same framework as RW-PCC, expect
  83  * that no HSM mechanism is used.
  84  *
  85  * The main advantages to use this SSD cache on the Lustre clients via PCC
  86  * is that:
  87  * - The I/O stack becomes much simpler for the cached data, as there is no
  88  *   interference with I/Os from other clients, which enables easier
  89  *   performance optimizations;
  90  * - The requirements on the HW inside the client nodes are small, any kind of
  91  *   SSDs or even HDDs can be used as cache devices;
  92  * - Caching reduces the pressure on the object storage targets (OSTs), as
  93  *   small or random I/Os can be regularized to big sequential I/Os and
  94  *   temporary files do not even need to be flushed to OSTs.
  95  *
  96  * PCC can accelerate applications with certain I/O patterns:
  97  * - small-sized random writes (< 1MB) from a single client
  98  * - repeated read of data that is larger than RAM
  99  * - clients with high network latency
 100  *
 101  * Author: Li Xi <lixi@ddn.com>
 102  * Author: Qian Yingjin <qian@ddn.com>
 103  */
 104
 105 #define DEBUG_SUBSYSTEM S_LLITE
 106
 107 #include "pcc.h"
 108 #include <linux/namei.h>
 109 #include <linux/file.h>
 110 #include <lustre_compat.h>
 111 #include "llite_internal.h"
 112
 113 struct kmem_cache *pcc_inode_slab;
 114
 115 int pcc_super_init(struct pcc_super *super)
 116 {
 117         struct cred *cred;
 118
 119         super->pccs_cred = cred = prepare_creds();
 120         if (!cred)
 121                 return -ENOMEM;
 122
 123         /* Never override disk quota limits or use reserved space */
 124         cap_lower(cred->cap_effective, CAP_SYS_RESOURCE);
 125         init_rwsem(&super->pccs_rw_sem);
 126         INIT_LIST_HEAD(&super->pccs_datasets);
 127
 128         return 0;
 129 }
 130
 131 /* Rule based auto caching */
 132 static void pcc_id_list_free(struct list_head *id_list)
 133 {
 134         struct pcc_match_id *id, *n;
 135
 136         list_for_each_entry_safe(id, n, id_list, pmi_linkage) {
 137                 list_del_init(&id->pmi_linkage);
 138                 OBD_FREE_PTR(id);
 139         }
 140 }
 141
 142 static void pcc_fname_list_free(struct list_head *fname_list)
 143 {
 144         struct pcc_match_fname *fname, *n;
 145
 146         list_for_each_entry_safe(fname, n, fname_list, pmf_linkage) {
 147                 OBD_FREE(fname->pmf_name, strlen(fname->pmf_name) + 1);
 148                 list_del_init(&fname->pmf_linkage);
 149                 OBD_FREE_PTR(fname);
 150         }
 151 }
 152
 153 static void pcc_expression_free(struct pcc_expression *expr)
 154 {
 155         LASSERT(expr->pe_field >= PCC_FIELD_UID &&
 156                 expr->pe_field < PCC_FIELD_MAX);
 157         switch (expr->pe_field) {
 158         case PCC_FIELD_UID:
 159         case PCC_FIELD_GID:
 160         case PCC_FIELD_PROJID:
 161                 pcc_id_list_free(&expr->pe_cond);
 162                 break;
 163         case PCC_FIELD_FNAME:
 164                 pcc_fname_list_free(&expr->pe_cond);
 165                 break;
 166         default:
 167                 LBUG();
 168         }
 169         OBD_FREE_PTR(expr);
 170 }
 171
 172 static void pcc_conjunction_free(struct pcc_conjunction *conjunction)
 173 {
 174         struct pcc_expression *expression, *n;
 175
 176         LASSERT(list_empty(&conjunction->pc_linkage));
 177         list_for_each_entry_safe(expression, n,
 178                                  &conjunction->pc_expressions,
 179                                  pe_linkage) {
 180                 list_del_init(&expression->pe_linkage);
 181                 pcc_expression_free(expression);
 182         }
 183         OBD_FREE_PTR(conjunction);
 184 }
 185
 186 static void pcc_rule_conds_free(struct list_head *cond_list)
 187 {
 188         struct pcc_conjunction *conjunction, *n;
 189
 190         list_for_each_entry_safe(conjunction, n, cond_list, pc_linkage) {
 191                 list_del_init(&conjunction->pc_linkage);
 192                 pcc_conjunction_free(conjunction);
 193         }
 194 }
 195
 196 static void pcc_cmd_fini(struct pcc_cmd *cmd)
 197 {
 198         if (cmd->pccc_cmd == PCC_ADD_DATASET) {
 199                 if (!list_empty(&cmd->u.pccc_add.pccc_conds))
 200                         pcc_rule_conds_free(&cmd->u.pccc_add.pccc_conds);
 201                 if (cmd->u.pccc_add.pccc_conds_str)
 202                         OBD_FREE(cmd->u.pccc_add.pccc_conds_str,
 203                                  strlen(cmd->u.pccc_add.pccc_conds_str) + 1);
 204         }
 205 }
 206
 207 #define PCC_DISJUNCTION_DELIM   (',')
 208 #define PCC_CONJUNCTION_DELIM   ('&')
 209 #define PCC_EXPRESSION_DELIM    ('=')
 210
 211 static int
 212 pcc_fname_list_add(struct cfs_lstr *id, struct list_head *fname_list)
 213 {
 214         struct pcc_match_fname *fname;
 215
 216         OBD_ALLOC(fname, sizeof(struct pcc_match_fname));
 217         if (fname == NULL)
 218                 return -ENOMEM;
 219
 220         OBD_ALLOC(fname->pmf_name, id->ls_len + 1);
 221         if (fname->pmf_name == NULL) {
 222                 OBD_FREE(fname, sizeof(struct pcc_match_fname));
 223                 return -ENOMEM;
 224         }
 225
 226         memcpy(fname->pmf_name, id->ls_str, id->ls_len);
 227         list_add_tail(&fname->pmf_linkage, fname_list);
 228         return 0;
 229 }
 230
 231 static int
 232 pcc_fname_list_parse(char *str, int len, struct list_head *fname_list)
 233 {
 234         struct cfs_lstr src;
 235         struct cfs_lstr res;
 236         int rc = 0;
 237
 238         ENTRY;
 239
 240         src.ls_str = str;
 241         src.ls_len = len;
 242         INIT_LIST_HEAD(fname_list);
 243         while (src.ls_str) {
 244                 rc = cfs_gettok(&src, ' ', &res);
 245                 if (rc == 0) {
 246                         rc = -EINVAL;
 247                         break;
 248                 }
 249                 rc = pcc_fname_list_add(&res, fname_list);
 250                 if (rc)
 251                         break;
 252         }
 253         if (rc)
 254                 pcc_fname_list_free(fname_list);
 255         RETURN(rc);
 256 }
 257
 258 static int
 259 pcc_id_list_parse(char *str, int len, struct list_head *id_list,
 260                   enum pcc_field type)
 261 {
 262         struct cfs_lstr src;
 263         struct cfs_lstr res;
 264         int rc = 0;
 265
 266         ENTRY;
 267
 268         if (type != PCC_FIELD_UID && type != PCC_FIELD_GID &&
 269             type != PCC_FIELD_PROJID)
 270                 RETURN(-EINVAL);
 271
 272         src.ls_str = str;
 273         src.ls_len = len;
 274         INIT_LIST_HEAD(id_list);
 275         while (src.ls_str) {
 276                 struct pcc_match_id *id;
 277                 __u32 id_val;
 278
 279                 if (cfs_gettok(&src, ' ', &res) == 0)
 280                         GOTO(out, rc = -EINVAL);
 281
 282                 if (!cfs_str2num_check(res.ls_str, res.ls_len,
 283                                        &id_val, 0, (u32)~0U))
 284                         GOTO(out, rc = -EINVAL);
 285
 286                 OBD_ALLOC_PTR(id);
 287                 if (id == NULL)
 288                         GOTO(out, rc = -ENOMEM);
 289
 290                 id->pmi_id = id_val;
 291                 list_add_tail(&id->pmi_linkage, id_list);
 292         }
 293 out:
 294         if (rc)
 295                 pcc_id_list_free(id_list);
 296         RETURN(rc);
 297 }
 298
 299 static inline bool
 300 pcc_check_field(struct cfs_lstr *field, char *str)
 301 {
 302         int len = strlen(str);
 303
 304         return (field->ls_len == len &&
 305                 strncmp(field->ls_str, str, len) == 0);
 306 }
 307
 308 static int
 309 pcc_expression_parse(struct cfs_lstr *src, struct list_head *cond_list)
 310 {
 311         struct pcc_expression *expr;
 312         struct cfs_lstr field;
 313         int rc = 0;
 314
 315         OBD_ALLOC(expr, sizeof(struct pcc_expression));
 316         if (expr == NULL)
 317                 return -ENOMEM;
 318
 319         rc = cfs_gettok(src, PCC_EXPRESSION_DELIM, &field);
 320         if (rc == 0 || src->ls_len <= 2 || src->ls_str[0] != '{' ||
 321             src->ls_str[src->ls_len - 1] != '}')
 322                 GOTO(out, rc = -EINVAL);
 323
 324         /* Skip '{' and '}' */
 325         src->ls_str++;
 326         src->ls_len -= 2;
 327
 328         if (pcc_check_field(&field, "uid")) {
 329                 if (pcc_id_list_parse(src->ls_str,
 330                                       src->ls_len,
 331                                       &expr->pe_cond,
 332                                       PCC_FIELD_UID) < 0)
 333                         GOTO(out, rc = -EINVAL);
 334                 expr->pe_field = PCC_FIELD_UID;
 335         } else if (pcc_check_field(&field, "gid")) {
 336                 if (pcc_id_list_parse(src->ls_str,
 337                                       src->ls_len,
 338                                       &expr->pe_cond,
 339                                       PCC_FIELD_GID) < 0)
 340                         GOTO(out, rc = -EINVAL);
 341                 expr->pe_field = PCC_FIELD_GID;
 342         } else if (pcc_check_field(&field, "projid")) {
 343                 if (pcc_id_list_parse(src->ls_str,
 344                                       src->ls_len,
 345                                       &expr->pe_cond,
 346                                       PCC_FIELD_PROJID) < 0)
 347                         GOTO(out, rc = -EINVAL);
 348                 expr->pe_field = PCC_FIELD_PROJID;
 349         } else if (pcc_check_field(&field, "fname")) {
 350                 if (pcc_fname_list_parse(src->ls_str,
 351                                          src->ls_len,
 352                                          &expr->pe_cond) < 0)
 353                         GOTO(out, rc = -EINVAL);
 354                 expr->pe_field = PCC_FIELD_FNAME;
 355         } else {
 356                 GOTO(out, rc = -EINVAL);
 357         }
 358
 359         list_add_tail(&expr->pe_linkage, cond_list);
 360         return 0;
 361 out:
 362         OBD_FREE_PTR(expr);
 363         return rc;
 364 }
 365
 366 static int
 367 pcc_conjunction_parse(struct cfs_lstr *src, struct list_head *cond_list)
 368 {
 369         struct pcc_conjunction *conjunction;
 370         struct cfs_lstr expr;
 371         int rc = 0;
 372
 373         OBD_ALLOC(conjunction, sizeof(struct pcc_conjunction));
 374         if (conjunction == NULL)
 375                 return -ENOMEM;
 376
 377         INIT_LIST_HEAD(&conjunction->pc_expressions);
 378         list_add_tail(&conjunction->pc_linkage, cond_list);
 379
 380         while (src->ls_str) {
 381                 rc = cfs_gettok(src, PCC_CONJUNCTION_DELIM, &expr);
 382                 if (rc == 0) {
 383                         rc = -EINVAL;
 384                         break;
 385                 }
 386                 rc = pcc_expression_parse(&expr,
 387                                           &conjunction->pc_expressions);
 388                 if (rc)
 389                         break;
 390         }
 391         return rc;
 392 }
 393
 394 static int pcc_conds_parse(char *str, int len, struct list_head *cond_list)
 395 {
 396         struct cfs_lstr src;
 397         struct cfs_lstr res;
 398         int rc = 0;
 399
 400         src.ls_str = str;
 401         src.ls_len = len;
 402         INIT_LIST_HEAD(cond_list);
 403         while (src.ls_str) {
 404                 rc = cfs_gettok(&src, PCC_DISJUNCTION_DELIM, &res);
 405                 if (rc == 0) {
 406                         rc = -EINVAL;
 407                         break;
 408                 }
 409                 rc = pcc_conjunction_parse(&res, cond_list);
 410                 if (rc)
 411                         break;
 412         }
 413         return rc;
 414 }
 415
 416 static int pcc_id_parse(struct pcc_cmd *cmd, const char *id)
 417 {
 418         int rc;
 419
 420         OBD_ALLOC(cmd->u.pccc_add.pccc_conds_str, strlen(id) + 1);
 421         if (cmd->u.pccc_add.pccc_conds_str == NULL)
 422                 return -ENOMEM;
 423
 424         memcpy(cmd->u.pccc_add.pccc_conds_str, id, strlen(id));
 425
 426         rc = pcc_conds_parse(cmd->u.pccc_add.pccc_conds_str,
 427                              strlen(cmd->u.pccc_add.pccc_conds_str),
 428                              &cmd->u.pccc_add.pccc_conds);
 429         if (rc)
 430                 pcc_cmd_fini(cmd);
 431
 432         return rc;
 433 }
 434
 435 static int
 436 pcc_parse_value_pair(struct pcc_cmd *cmd, char *buffer)
 437 {
 438         char *key, *val;
 439         unsigned long id;
 440         int rc;
 441
 442         val = buffer;
 443         key = strsep(&val, "=");
 444         if (val == NULL || strlen(val) == 0)
 445                 return -EINVAL;
 446
 447         /* Key of the value pair */
 448         if (strcmp(key, "rwid") == 0) {
 449                 rc = kstrtoul(val, 10, &id);
 450                 if (rc)
 451                         return rc;
 452                 if (id <= 0)
 453                         return -EINVAL;
 454                 cmd->u.pccc_add.pccc_rwid = id;
 455         } else if (strcmp(key, "roid") == 0) {
 456                 rc = kstrtoul(val, 10, &id);
 457                 if (rc)
 458                         return rc;
 459                 if (id <= 0)
 460                         return -EINVAL;
 461                 cmd->u.pccc_add.pccc_roid = id;
 462         } else if (strcmp(key, "open_attach") == 0) {
 463                 rc = kstrtoul(val, 10, &id);
 464                 if (rc)
 465                         return rc;
 466                 if (id > 0)
 467                         cmd->u.pccc_add.pccc_flags |= PCC_DATASET_OPEN_ATTACH;
 468         } else if (strcmp(key, "rwpcc") == 0) {
 469                 rc = kstrtoul(val, 10, &id);
 470                 if (rc)
 471                         return rc;
 472                 if (id > 0)
 473                         cmd->u.pccc_add.pccc_flags |= PCC_DATASET_RWPCC;
 474         } else if (strcmp(key, "ropcc") == 0) {
 475                 rc = kstrtoul(val, 10, &id);
 476                 if (rc)
 477                         return rc;
 478                 if (id > 0)
 479                         cmd->u.pccc_add.pccc_flags |= PCC_DATASET_ROPCC;
 480         } else {
 481                 return -EINVAL;
 482         }
 483
 484         return 0;
 485 }
 486
 487 static int
 488 pcc_parse_value_pairs(struct pcc_cmd *cmd, char *buffer)
 489 {
 490         char *val;
 491         char *token;
 492         int rc;
 493
 494         val = buffer;
 495         while (val != NULL && strlen(val) != 0) {
 496                 token = strsep(&val, " ");
 497                 rc = pcc_parse_value_pair(cmd, token);
 498                 if (rc)
 499                         return rc;
 500         }
 501
 502         switch (cmd->pccc_cmd) {
 503         case PCC_ADD_DATASET:
 504                 if (cmd->u.pccc_add.pccc_flags & PCC_DATASET_RWPCC &&
 505                     cmd->u.pccc_add.pccc_flags & PCC_DATASET_ROPCC)
 506                         return -EINVAL;
 507                 /*
 508                  * By default, a PCC backend can provide caching service for
 509                  * both RW-PCC and RO-PCC.
 510                  */
 511                 if ((cmd->u.pccc_add.pccc_flags & PCC_DATASET_PCC_ALL) == 0)
 512                         cmd->u.pccc_add.pccc_flags |= PCC_DATASET_PCC_ALL;
 513                 break;
 514         case PCC_DEL_DATASET:
 515         case PCC_CLEAR_ALL:
 516                 break;
 517         default:
 518                 return -EINVAL;
 519         }
 520         return 0;
 521 }
 522
 523 static void
 524 pcc_dataset_rule_fini(struct pcc_match_rule *rule)
 525 {
 526         if (!list_empty(&rule->pmr_conds))
 527                 pcc_rule_conds_free(&rule->pmr_conds);
 528         LASSERT(rule->pmr_conds_str != NULL);
 529         OBD_FREE(rule->pmr_conds_str, strlen(rule->pmr_conds_str) + 1);
 530 }
 531
 532 static int
 533 pcc_dataset_rule_init(struct pcc_match_rule *rule, struct pcc_cmd *cmd)
 534 {
 535         int rc = 0;
 536
 537         LASSERT(cmd->u.pccc_add.pccc_conds_str);
 538         OBD_ALLOC(rule->pmr_conds_str,
 539                   strlen(cmd->u.pccc_add.pccc_conds_str) + 1);
 540         if (rule->pmr_conds_str == NULL)
 541                 return -ENOMEM;
 542
 543         memcpy(rule->pmr_conds_str,
 544                cmd->u.pccc_add.pccc_conds_str,
 545                strlen(cmd->u.pccc_add.pccc_conds_str));
 546
 547         INIT_LIST_HEAD(&rule->pmr_conds);
 548         if (!list_empty(&cmd->u.pccc_add.pccc_conds))
 549                 rc = pcc_conds_parse(rule->pmr_conds_str,
 550                                           strlen(rule->pmr_conds_str),
 551                                           &rule->pmr_conds);
 552
 553         if (rc)
 554                 pcc_dataset_rule_fini(rule);
 555
 556         return rc;
 557 }
 558
 559 /* Rule Matching */
 560 static int
 561 pcc_id_list_match(struct list_head *id_list, __u32 id_val)
 562 {
 563         struct pcc_match_id *id;
 564
 565         list_for_each_entry(id, id_list, pmi_linkage) {
 566                 if (id->pmi_id == id_val)
 567                         return 1;
 568         }
 569         return 0;
 570 }
 571
 572 static bool
 573 cfs_match_wildcard(const char *pattern, const char *content)
 574 {
 575         if (*pattern == '\0' && *content == '\0')
 576                 return true;
 577
 578         if (*pattern == '*' && *(pattern + 1) != '\0' && *content == '\0')
 579                 return false;
 580
 581         while (*pattern == *content) {
 582                 pattern++;
 583                 content++;
 584                 if (*pattern == '\0' && *content == '\0')
 585                         return true;
 586
 587                 if (*pattern == '*' && *(pattern + 1) != '\0' &&
 588                     *content == '\0')
 589                         return false;
 590         }
 591
 592         if (*pattern == '*')
 593                 return (cfs_match_wildcard(pattern + 1, content) ||
 594                         cfs_match_wildcard(pattern, content + 1));
 595
 596         return false;
 597 }
 598
 599 static int
 600 pcc_fname_list_match(struct list_head *fname_list, const char *name)
 601 {
 602         struct pcc_match_fname *fname;
 603
 604         list_for_each_entry(fname, fname_list, pmf_linkage) {
 605                 if (cfs_match_wildcard(fname->pmf_name, name))
 606                         return 1;
 607         }
 608         return 0;
 609 }
 610
 611 static int
 612 pcc_expression_match(struct pcc_expression *expr, struct pcc_matcher *matcher)
 613 {
 614         switch (expr->pe_field) {
 615         case PCC_FIELD_UID:
 616                 return pcc_id_list_match(&expr->pe_cond, matcher->pm_uid);
 617         case PCC_FIELD_GID:
 618                 return pcc_id_list_match(&expr->pe_cond, matcher->pm_gid);
 619         case PCC_FIELD_PROJID:
 620                 return pcc_id_list_match(&expr->pe_cond, matcher->pm_projid);
 621         case PCC_FIELD_FNAME:
 622                 return pcc_fname_list_match(&expr->pe_cond,
 623                                             matcher->pm_name->name);
 624         default:
 625                 return 0;
 626         }
 627 }
 628
 629 static int
 630 pcc_conjunction_match(struct pcc_conjunction *conjunction,
 631                       struct pcc_matcher *matcher)
 632 {
 633         struct pcc_expression *expr;
 634         int matched;
 635
 636         list_for_each_entry(expr, &conjunction->pc_expressions, pe_linkage) {
 637                 matched = pcc_expression_match(expr, matcher);
 638                 if (!matched)
 639                         return 0;
 640         }
 641
 642         return 1;
 643 }
 644
 645 static int
 646 pcc_cond_match(struct pcc_match_rule *rule, struct pcc_matcher *matcher)
 647 {
 648         struct pcc_conjunction *conjunction;
 649         int matched;
 650
 651         list_for_each_entry(conjunction, &rule->pmr_conds, pc_linkage) {
 652                 matched = pcc_conjunction_match(conjunction, matcher);
 653                 if (matched)
 654                         return 1;
 655         }
 656
 657         return 0;
 658 }
 659
 660 struct pcc_dataset*
 661 pcc_dataset_match_get(struct pcc_super *super, struct pcc_matcher *matcher)
 662 {
 663         struct pcc_dataset *dataset;
 664         struct pcc_dataset *selected = NULL;
 665
 666         down_read(&super->pccs_rw_sem);
 667         list_for_each_entry(dataset, &super->pccs_datasets, pccd_linkage) {
 668                 if (!(dataset->pccd_flags & PCC_DATASET_RWPCC))
 669                         continue;
 670
 671                 if (pcc_cond_match(&dataset->pccd_rule, matcher)) {
 672                         atomic_inc(&dataset->pccd_refcount);
 673                         selected = dataset;
 674                         break;
 675                 }
 676         }
 677         up_read(&super->pccs_rw_sem);
 678         if (selected)
 679                 CDEBUG(D_CACHE, "PCC create, matched %s - %d:%d:%d:%s\n",
 680                        dataset->pccd_rule.pmr_conds_str,
 681                        matcher->pm_uid, matcher->pm_gid,
 682                        matcher->pm_projid, matcher->pm_name->name);
 683
 684         return selected;
 685 }
 686
 687 /**
 688  * pcc_dataset_add - Add a Cache policy to control which files need be
 689  * cached and where it will be cached.
 690  *
 691  * @super:      superblock of pcc
 692  * @cmd:        pcc command
 693  */
 694 static int
 695 pcc_dataset_add(struct pcc_super *super, struct pcc_cmd *cmd)
 696 {
 697         char *pathname = cmd->pccc_pathname;
 698         struct pcc_dataset *dataset;
 699         struct pcc_dataset *tmp;
 700         bool found = false;
 701         int rc;
 702
 703         OBD_ALLOC_PTR(dataset);
 704         if (dataset == NULL)
 705                 return -ENOMEM;
 706
 707         rc = kern_path(pathname, LOOKUP_DIRECTORY, &dataset->pccd_path);
 708         if (unlikely(rc)) {
 709                 OBD_FREE_PTR(dataset);
 710                 return rc;
 711         }
 712         strncpy(dataset->pccd_pathname, pathname, PATH_MAX);
 713         dataset->pccd_rwid = cmd->u.pccc_add.pccc_rwid;
 714         dataset->pccd_roid = cmd->u.pccc_add.pccc_roid;
 715         dataset->pccd_flags = cmd->u.pccc_add.pccc_flags;
 716         atomic_set(&dataset->pccd_refcount, 1);
 717
 718         rc = pcc_dataset_rule_init(&dataset->pccd_rule, cmd);
 719         if (rc) {
 720                 pcc_dataset_put(dataset);
 721                 return rc;
 722         }
 723
 724         down_write(&super->pccs_rw_sem);
 725         list_for_each_entry(tmp, &super->pccs_datasets, pccd_linkage) {
 726                 if (strcmp(tmp->pccd_pathname, pathname) == 0 ||
 727                     (dataset->pccd_rwid != 0 &&
 728                      dataset->pccd_rwid == tmp->pccd_rwid) ||
 729                     (dataset->pccd_roid != 0 &&
 730                      dataset->pccd_roid == tmp->pccd_roid)) {
 731                         found = true;
 732                         break;
 733                 }
 734         }
 735         if (!found)
 736                 list_add(&dataset->pccd_linkage, &super->pccs_datasets);
 737         up_write(&super->pccs_rw_sem);
 738
 739         if (found) {
 740                 pcc_dataset_put(dataset);
 741                 rc = -EEXIST;
 742         }
 743
 744         return rc;
 745 }
 746
 747 struct pcc_dataset *
 748 pcc_dataset_get(struct pcc_super *super, enum lu_pcc_type type, __u32 id)
 749 {
 750         struct pcc_dataset *dataset;
 751         struct pcc_dataset *selected = NULL;
 752
 753         if (id == 0)
 754                 return NULL;
 755
 756         /*
 757          * archive ID (read-write ID) or read-only ID is unique in the list,
 758          * we just return last added one as first priority.
 759          */
 760         down_read(&super->pccs_rw_sem);
 761         list_for_each_entry(dataset, &super->pccs_datasets, pccd_linkage) {
 762                 if (type == LU_PCC_READWRITE && (dataset->pccd_rwid != id ||
 763                     !(dataset->pccd_flags & PCC_DATASET_RWPCC)))
 764                         continue;
 765                 atomic_inc(&dataset->pccd_refcount);
 766                 selected = dataset;
 767                 break;
 768         }
 769         up_read(&super->pccs_rw_sem);
 770         if (selected)
 771                 CDEBUG(D_CACHE, "matched id %u, PCC mode %d\n", id, type);
 772
 773         return selected;
 774 }
 775
 776 void
 777 pcc_dataset_put(struct pcc_dataset *dataset)
 778 {
 779         if (atomic_dec_and_test(&dataset->pccd_refcount)) {
 780                 pcc_dataset_rule_fini(&dataset->pccd_rule);
 781                 path_put(&dataset->pccd_path);
 782                 OBD_FREE_PTR(dataset);
 783         }
 784 }
 785
 786 static int
 787 pcc_dataset_del(struct pcc_super *super, char *pathname)
 788 {
 789         struct list_head *l, *tmp;
 790         struct pcc_dataset *dataset;
 791         int rc = -ENOENT;
 792
 793         down_write(&super->pccs_rw_sem);
 794         list_for_each_safe(l, tmp, &super->pccs_datasets) {
 795                 dataset = list_entry(l, struct pcc_dataset, pccd_linkage);
 796                 if (strcmp(dataset->pccd_pathname, pathname) == 0) {
 797                         list_del_init(&dataset->pccd_linkage);
 798                         pcc_dataset_put(dataset);
 799                         rc = 0;
 800                         break;
 801                 }
 802         }
 803         up_write(&super->pccs_rw_sem);
 804         return rc;
 805 }
 806
 807 static void
 808 pcc_dataset_dump(struct pcc_dataset *dataset, struct seq_file *m)
 809 {
 810         seq_printf(m, "%s:\n", dataset->pccd_pathname);
 811         seq_printf(m, "  rwid: %u\n", dataset->pccd_rwid);
 812         seq_printf(m, "  flags: %x\n", dataset->pccd_flags);
 813         seq_printf(m, "  autocache: %s\n", dataset->pccd_rule.pmr_conds_str);
 814 }
 815
 816 int
 817 pcc_super_dump(struct pcc_super *super, struct seq_file *m)
 818 {
 819         struct pcc_dataset *dataset;
 820
 821         down_read(&super->pccs_rw_sem);
 822         list_for_each_entry(dataset, &super->pccs_datasets, pccd_linkage) {
 823                 pcc_dataset_dump(dataset, m);
 824         }
 825         up_read(&super->pccs_rw_sem);
 826         return 0;
 827 }
 828
 829 static void pcc_remove_datasets(struct pcc_super *super)
 830 {
 831         struct pcc_dataset *dataset, *tmp;
 832
 833         down_write(&super->pccs_rw_sem);
 834         list_for_each_entry_safe(dataset, tmp,
 835                                  &super->pccs_datasets, pccd_linkage) {
 836                 list_del(&dataset->pccd_linkage);
 837                 pcc_dataset_put(dataset);
 838         }
 839         up_write(&super->pccs_rw_sem);
 840 }
 841
 842 void pcc_super_fini(struct pcc_super *super)
 843 {
 844         pcc_remove_datasets(super);
 845         put_cred(super->pccs_cred);
 846 }
 847
 848 static bool pathname_is_valid(const char *pathname)
 849 {
 850         /* Needs to be absolute path */
 851         if (pathname == NULL || strlen(pathname) == 0 ||
 852             strlen(pathname) >= PATH_MAX || pathname[0] != '/')
 853                 return false;
 854         return true;
 855 }
 856
 857 static struct pcc_cmd *
 858 pcc_cmd_parse(char *buffer, unsigned long count)
 859 {
 860         static struct pcc_cmd *cmd;
 861         char *token;
 862         char *val;
 863         int rc = 0;
 864
 865         OBD_ALLOC_PTR(cmd);
 866         if (cmd == NULL)
 867                 GOTO(out, rc = -ENOMEM);
 868
 869         /* clear all setting */
 870         if (strncmp(buffer, "clear", 5) == 0) {
 871                 cmd->pccc_cmd = PCC_CLEAR_ALL;
 872                 GOTO(out, rc = 0);
 873         }
 874
 875         val = buffer;
 876         token = strsep(&val, " ");
 877         if (val == NULL || strlen(val) == 0)
 878                 GOTO(out_free_cmd, rc = -EINVAL);
 879
 880         /* Type of the command */
 881         if (strcmp(token, "add") == 0)
 882                 cmd->pccc_cmd = PCC_ADD_DATASET;
 883         else if (strcmp(token, "del") == 0)
 884                 cmd->pccc_cmd = PCC_DEL_DATASET;
 885         else
 886                 GOTO(out_free_cmd, rc = -EINVAL);
 887
 888         /* Pathname of the dataset */
 889         token = strsep(&val, " ");
 890         if ((val == NULL && cmd->pccc_cmd != PCC_DEL_DATASET) ||
 891             !pathname_is_valid(token))
 892                 GOTO(out_free_cmd, rc = -EINVAL);
 893         cmd->pccc_pathname = token;
 894
 895         if (cmd->pccc_cmd == PCC_ADD_DATASET) {
 896                 /* List of ID */
 897                 LASSERT(val);
 898                 token = val;
 899                 val = strrchr(token, '}');
 900                 if (!val)
 901                         GOTO(out_free_cmd, rc = -EINVAL);
 902
 903                 /* Skip '}' */
 904                 val++;
 905                 if (*val == '\0') {
 906                         val = NULL;
 907                 } else if (*val == ' ') {
 908                         *val = '\0';
 909                         val++;
 910                 } else {
 911                         GOTO(out_free_cmd, rc = -EINVAL);
 912                 }
 913
 914                 rc = pcc_id_parse(cmd, token);
 915                 if (rc)
 916                         GOTO(out_free_cmd, rc);
 917
 918                 rc = pcc_parse_value_pairs(cmd, val);
 919                 if (rc)
 920                         GOTO(out_cmd_fini, rc = -EINVAL);
 921         }
 922         goto out;
 923 out_cmd_fini:
 924         pcc_cmd_fini(cmd);
 925 out_free_cmd:
 926         OBD_FREE_PTR(cmd);
 927 out:
 928         if (rc)
 929                 cmd = ERR_PTR(rc);
 930         return cmd;
 931 }
 932
 933 int pcc_cmd_handle(char *buffer, unsigned long count,
 934                    struct pcc_super *super)
 935 {
 936         int rc = 0;
 937         struct pcc_cmd *cmd;
 938
 939         cmd = pcc_cmd_parse(buffer, count);
 940         if (IS_ERR(cmd))
 941                 return PTR_ERR(cmd);
 942
 943         switch (cmd->pccc_cmd) {
 944         case PCC_ADD_DATASET:
 945                 rc = pcc_dataset_add(super, cmd);
 946                 break;
 947         case PCC_DEL_DATASET:
 948                 rc = pcc_dataset_del(super, cmd->pccc_pathname);
 949                 break;
 950         case PCC_CLEAR_ALL:
 951                 pcc_remove_datasets(super);
 952                 break;
 953         default:
 954                 rc = -EINVAL;
 955                 break;
 956         }
 957
 958         pcc_cmd_fini(cmd);
 959         OBD_FREE_PTR(cmd);
 960         return rc;
 961 }
 962
 963 static inline void pcc_inode_lock(struct inode *inode)
 964 {
 965         mutex_lock(&ll_i2info(inode)->lli_pcc_lock);
 966 }
 967
 968 static inline void pcc_inode_unlock(struct inode *inode)
 969 {
 970         mutex_unlock(&ll_i2info(inode)->lli_pcc_lock);
 971 }
 972
 973 static void pcc_inode_init(struct pcc_inode *pcci, struct ll_inode_info *lli)
 974 {
 975         pcci->pcci_lli = lli;
 976         lli->lli_pcc_inode = pcci;
 977         atomic_set(&pcci->pcci_refcount, 0);
 978         pcci->pcci_type = LU_PCC_NONE;
 979         pcci->pcci_layout_gen = CL_LAYOUT_GEN_NONE;
 980         atomic_set(&pcci->pcci_active_ios, 0);
 981         init_waitqueue_head(&pcci->pcci_waitq);
 982 }
 983
 984 static void pcc_inode_fini(struct pcc_inode *pcci)
 985 {
 986         struct ll_inode_info *lli = pcci->pcci_lli;
 987
 988         path_put(&pcci->pcci_path);
 989         pcci->pcci_type = LU_PCC_NONE;
 990         OBD_SLAB_FREE_PTR(pcci, pcc_inode_slab);
 991         lli->lli_pcc_inode = NULL;
 992 }
 993
 994 static void pcc_inode_get(struct pcc_inode *pcci)
 995 {
 996         atomic_inc(&pcci->pcci_refcount);
 997 }
 998
 999 static void pcc_inode_put(struct pcc_inode *pcci)
1000 {
1001         if (atomic_dec_and_test(&pcci->pcci_refcount))
1002                 pcc_inode_fini(pcci);
1003 }
1004
1005 void pcc_inode_free(struct inode *inode)
1006 {
1007         struct pcc_inode *pcci = ll_i2pcci(inode);
1008
1009         if (pcci) {
1010                 WARN_ON(atomic_read(&pcci->pcci_refcount) > 1);
1011                 pcc_inode_put(pcci);
1012         }
1013 }
1014
1015 /*
1016  * TODO:
1017  * As Andreas suggested, we'd better use new layout to
1018  * reduce overhead:
1019  * (fid->f_oid >> 16 & oxFFFF)/FID
1020  */
1021 #define MAX_PCC_DATABASE_PATH (6 * 5 + FID_NOBRACE_LEN + 1)
1022 static int pcc_fid2dataset_path(char *buf, int sz, struct lu_fid *fid)
1023 {
1024         return snprintf(buf, sz, "%04x/%04x/%04x/%04x/%04x/%04x/"
1025                         DFID_NOBRACE,
1026                         (fid)->f_oid       & 0xFFFF,
1027                         (fid)->f_oid >> 16 & 0xFFFF,
1028                         (unsigned int)((fid)->f_seq       & 0xFFFF),
1029                         (unsigned int)((fid)->f_seq >> 16 & 0xFFFF),
1030                         (unsigned int)((fid)->f_seq >> 32 & 0xFFFF),
1031                         (unsigned int)((fid)->f_seq >> 48 & 0xFFFF),
1032                         PFID(fid));
1033 }
1034
1035 static inline const struct cred *pcc_super_cred(struct super_block *sb)
1036 {
1037         return ll_s2sbi(sb)->ll_pcc_super.pccs_cred;
1038 }
1039
1040 void pcc_file_init(struct pcc_file *pccf)
1041 {
1042         pccf->pccf_file = NULL;
1043         pccf->pccf_type = LU_PCC_NONE;
1044 }
1045
1046 static inline bool pcc_open_attach_enabled(struct pcc_dataset *dataset)
1047 {
1048         return dataset->pccd_flags & PCC_DATASET_OPEN_ATTACH;
1049 }
1050
1051 static const char pcc_xattr_layout[] = XATTR_USER_PREFIX "PCC.layout";
1052
1053 static int pcc_layout_xattr_set(struct pcc_inode *pcci, __u32 gen)
1054 {
1055         struct dentry *pcc_dentry = pcci->pcci_path.dentry;
1056         struct ll_inode_info *lli = pcci->pcci_lli;
1057         int rc;
1058
1059         ENTRY;
1060
1061         if (!(lli->lli_pcc_state & PCC_STATE_FL_OPEN_ATTACH))
1062                 RETURN(0);
1063
1064 #ifndef HAVE_VFS_SETXATTR
1065         if (!pcc_dentry->d_inode->i_op->setxattr)
1066                 RETURN(-ENOTSUPP);
1067
1068         rc = pcc_dentry->d_inode->i_op->setxattr(pcc_dentry, pcc_xattr_layout,
1069                                                  &gen, sizeof(gen), 0);
1070 #else
1071         rc = __vfs_setxattr(pcc_dentry, pcc_dentry->d_inode, pcc_xattr_layout,
1072                             &gen, sizeof(gen), 0);
1073 #endif
1074         RETURN(rc);
1075 }
1076
1077 static int pcc_get_layout_info(struct inode *inode, struct cl_layout *clt)
1078 {
1079         struct lu_env *env;
1080         struct ll_inode_info *lli = ll_i2info(inode);
1081         __u16 refcheck;
1082         int rc;
1083
1084         ENTRY;
1085
1086         if (!lli->lli_clob)
1087                 RETURN(-EINVAL);
1088
1089         env = cl_env_get(&refcheck);
1090         if (IS_ERR(env))
1091                 RETURN(PTR_ERR(env));
1092
1093         rc = cl_object_layout_get(env, lli->lli_clob, clt);
1094         if (rc)
1095                 CDEBUG(D_INODE, "Cannot get layout for "DFID"\n",
1096                        PFID(ll_inode2fid(inode)));
1097
1098         cl_env_put(env, &refcheck);
1099         RETURN(rc);
1100 }
1101
1102 static int pcc_fid2dataset_fullpath(char *buf, int sz, struct lu_fid *fid,
1103                                     struct pcc_dataset *dataset)
1104 {
1105         return snprintf(buf, sz, "%s/%04x/%04x/%04x/%04x/%04x/%04x/"
1106                         DFID_NOBRACE,
1107                         dataset->pccd_pathname,
1108                         (fid)->f_oid       & 0xFFFF,
1109                         (fid)->f_oid >> 16 & 0xFFFF,
1110                         (unsigned int)((fid)->f_seq       & 0xFFFF),
1111                         (unsigned int)((fid)->f_seq >> 16 & 0xFFFF),
1112                         (unsigned int)((fid)->f_seq >> 32 & 0xFFFF),
1113                         (unsigned int)((fid)->f_seq >> 48 & 0xFFFF),
1114                         PFID(fid));
1115 }
1116
1117 /* Must be called with pcci->pcci_lock held */
1118 static void pcc_inode_attach_init(struct pcc_dataset *dataset,
1119                                   struct pcc_inode *pcci,
1120                                   struct dentry *dentry,
1121                                   enum lu_pcc_type type)
1122 {
1123         pcci->pcci_path.mnt = mntget(dataset->pccd_path.mnt);
1124         pcci->pcci_path.dentry = dentry;
1125         LASSERT(atomic_read(&pcci->pcci_refcount) == 0);
1126         atomic_set(&pcci->pcci_refcount, 1);
1127         pcci->pcci_type = type;
1128         pcci->pcci_attr_valid = false;
1129
1130         if (pcc_open_attach_enabled(dataset)) {
1131                 struct ll_inode_info *lli = pcci->pcci_lli;
1132
1133                 lli->lli_pcc_state |= PCC_STATE_FL_OPEN_ATTACH;
1134         }
1135 }
1136
1137 static inline void pcc_layout_gen_set(struct pcc_inode *pcci,
1138                                       __u32 gen)
1139 {
1140         pcci->pcci_layout_gen = gen;
1141 }
1142
1143 static inline bool pcc_inode_has_layout(struct pcc_inode *pcci)
1144 {
1145         return pcci->pcci_layout_gen != CL_LAYOUT_GEN_NONE;
1146 }
1147
1148 static int pcc_try_dataset_attach(struct inode *inode, __u32 gen,
1149                                   enum lu_pcc_type type,
1150                                   struct pcc_dataset *dataset,
1151                                   bool *cached)
1152 {
1153         struct ll_inode_info *lli = ll_i2info(inode);
1154         struct pcc_inode *pcci = lli->lli_pcc_inode;
1155         const struct cred *old_cred;
1156         struct dentry *pcc_dentry;
1157         struct path path;
1158         char *pathname;
1159         __u32 pcc_gen;
1160         int rc;
1161
1162         ENTRY;
1163
1164         if (type == LU_PCC_READWRITE &&
1165             !(dataset->pccd_flags & PCC_DATASET_RWPCC))
1166                 RETURN(0);
1167
1168         OBD_ALLOC(pathname, PATH_MAX);
1169         if (pathname == NULL)
1170                 RETURN(-ENOMEM);
1171
1172         pcc_fid2dataset_fullpath(pathname, PATH_MAX, &lli->lli_fid, dataset);
1173
1174         old_cred = override_creds(pcc_super_cred(inode->i_sb));
1175         rc = kern_path(pathname, LOOKUP_FOLLOW, &path);
1176         if (rc)
1177                 /* ignore this error */
1178                 GOTO(out, rc = 0);
1179
1180         pcc_dentry = path.dentry;
1181 #ifndef HAVE_VFS_SETXATTR
1182         if (!pcc_dentry->d_inode->i_op->getxattr)
1183                 /* ignore this error */
1184                 GOTO(out_put_path, rc = 0);
1185
1186         rc = pcc_dentry->d_inode->i_op->getxattr(pcc_dentry, pcc_xattr_layout,
1187                                                  &pcc_gen, sizeof(pcc_gen));
1188 #else
1189         rc = __vfs_getxattr(pcc_dentry, pcc_dentry->d_inode, pcc_xattr_layout,
1190                             &pcc_gen, sizeof(pcc_gen));
1191 #endif
1192
1193         if (rc < 0)
1194                 /* ignore this error */
1195                 GOTO(out_put_path, rc = 0);
1196
1197         rc = 0;
1198         /* The file is still valid cached in PCC, attach it immediately. */
1199         if (pcc_gen == gen) {
1200                 CDEBUG(D_CACHE, DFID" L.Gen (%d) consistent, auto attached.\n",
1201                        PFID(&lli->lli_fid), gen);
1202                 if (!pcci) {
1203                         OBD_SLAB_ALLOC_PTR_GFP(pcci, pcc_inode_slab, GFP_NOFS);
1204                         if (pcci == NULL)
1205                                 GOTO(out_put_path, rc = -ENOMEM);
1206
1207                         pcc_inode_init(pcci, lli);
1208                         dget(pcc_dentry);
1209                         pcc_inode_attach_init(dataset, pcci, pcc_dentry, type);
1210                 } else {
1211                         /*
1212                          * This happened when a file was once attached into
1213                          * PCC, and some processes keep this file opened
1214                          * (pcci->refcount > 1) and corresponding PCC file
1215                          * without any I/O activity, and then this file was
1216                          * detached by the manual detach command or the
1217                          * revocation of the layout lock (i.e. cached LRU lock
1218                          * shrinking).
1219                          */
1220                         pcc_inode_get(pcci);
1221                         pcci->pcci_type = type;
1222                 }
1223                 pcc_layout_gen_set(pcci, gen);
1224                 *cached = true;
1225         }
1226 out_put_path:
1227         path_put(&path);
1228 out:
1229         revert_creds(old_cred);
1230         OBD_FREE(pathname, PATH_MAX);
1231         RETURN(rc);
1232 }
1233
1234 static int pcc_try_datasets_attach(struct inode *inode, __u32 gen,
1235                                    enum lu_pcc_type type, bool *cached)
1236 {
1237         struct pcc_dataset *dataset, *tmp;
1238         struct pcc_super *super = &ll_i2sbi(inode)->ll_pcc_super;
1239         int rc = 0;
1240
1241         ENTRY;
1242
1243         down_read(&super->pccs_rw_sem);
1244         list_for_each_entry_safe(dataset, tmp,
1245                                  &super->pccs_datasets, pccd_linkage) {
1246                 if (!pcc_open_attach_enabled(dataset))
1247                         continue;
1248                 rc = pcc_try_dataset_attach(inode, gen, type, dataset, cached);
1249                 if (rc < 0 || (!rc && *cached))
1250                         break;
1251         }
1252         up_read(&super->pccs_rw_sem);
1253
1254         RETURN(rc);
1255 }
1256
1257 static int pcc_try_open_attach(struct inode *inode, bool *cached)
1258 {
1259         struct pcc_super *super = &ll_i2sbi(inode)->ll_pcc_super;
1260         struct cl_layout clt = {
1261                 .cl_layout_gen = 0,
1262                 .cl_is_released = false,
1263         };
1264         int rc;
1265
1266         ENTRY;
1267
1268         /*
1269          * Quick check whether there is PCC device.
1270          */
1271         if (list_empty(&super->pccs_datasets))
1272                 RETURN(0);
1273
1274         /*
1275          * The file layout lock was cancelled. And this open does not
1276          * obtain valid layout lock from MDT (i.e. the file is being
1277          * HSM restoring).
1278          */
1279         if (ll_layout_version_get(ll_i2info(inode)) == CL_LAYOUT_GEN_NONE)
1280                 RETURN(0);
1281
1282         rc = pcc_get_layout_info(inode, &clt);
1283         if (rc)
1284                 RETURN(rc);
1285
1286         if (clt.cl_is_released)
1287                 rc = pcc_try_datasets_attach(inode, clt.cl_layout_gen,
1288                                              LU_PCC_READWRITE, cached);
1289
1290         RETURN(rc);
1291 }
1292
1293 int pcc_file_open(struct inode *inode, struct file *file)
1294 {
1295         struct pcc_inode *pcci;
1296         struct ll_inode_info *lli = ll_i2info(inode);
1297         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1298         struct pcc_file *pccf = &fd->fd_pcc_file;
1299         struct file *pcc_file;
1300         struct path *path;
1301         struct qstr *dname;
1302         bool cached = false;
1303         int rc = 0;
1304
1305         ENTRY;
1306
1307         if (!S_ISREG(inode->i_mode))
1308                 RETURN(0);
1309
1310         pcc_inode_lock(inode);
1311         pcci = ll_i2pcci(inode);
1312
1313         if (lli->lli_pcc_state & PCC_STATE_FL_ATTACHING)
1314                 GOTO(out_unlock, rc = 0);
1315
1316         if (!pcci || !pcc_inode_has_layout(pcci)) {
1317                 rc = pcc_try_open_attach(inode, &cached);
1318                 if (rc < 0 || !cached)
1319                         GOTO(out_unlock, rc);
1320
1321                 if (!pcci)
1322                         pcci = ll_i2pcci(inode);
1323         }
1324
1325         pcc_inode_get(pcci);
1326         WARN_ON(pccf->pccf_file);
1327
1328         path = &pcci->pcci_path;
1329         dname = &path->dentry->d_name;
1330         CDEBUG(D_CACHE, "opening pcc file '%.*s'\n", dname->len,
1331                dname->name);
1332
1333 #ifdef HAVE_DENTRY_OPEN_USE_PATH
1334         pcc_file = dentry_open(path, file->f_flags,
1335                                pcc_super_cred(inode->i_sb));
1336 #else
1337         pcc_file = dentry_open(path->dentry, path->mnt, file->f_flags,
1338                                pcc_super_cred(inode->i_sb));
1339 #endif
1340         if (IS_ERR_OR_NULL(pcc_file)) {
1341                 rc = pcc_file == NULL ? -EINVAL : PTR_ERR(pcc_file);
1342                 pcc_inode_put(pcci);
1343         } else {
1344                 pccf->pccf_file = pcc_file;
1345                 pccf->pccf_type = pcci->pcci_type;
1346         }
1347
1348 out_unlock:
1349         pcc_inode_unlock(inode);
1350         RETURN(rc);
1351 }
1352
1353 void pcc_file_release(struct inode *inode, struct file *file)
1354 {
1355         struct pcc_inode *pcci;
1356         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1357         struct pcc_file *pccf;
1358         struct path *path;
1359         struct qstr *dname;
1360
1361         ENTRY;
1362
1363         if (!S_ISREG(inode->i_mode) || fd == NULL)
1364                 RETURN_EXIT;
1365
1366         pccf = &fd->fd_pcc_file;
1367         pcc_inode_lock(inode);
1368         if (pccf->pccf_file == NULL)
1369                 goto out;
1370
1371         pcci = ll_i2pcci(inode);
1372         LASSERT(pcci);
1373         path = &pcci->pcci_path;
1374         dname = &path->dentry->d_name;
1375         CDEBUG(D_CACHE, "releasing pcc file \"%.*s\"\n", dname->len,
1376                dname->name);
1377         pcc_inode_put(pcci);
1378         fput(pccf->pccf_file);
1379         pccf->pccf_file = NULL;
1380 out:
1381         pcc_inode_unlock(inode);
1382         RETURN_EXIT;
1383 }
1384
1385 static void pcc_io_init(struct inode *inode, bool *cached)
1386 {
1387         struct pcc_inode *pcci;
1388
1389         pcc_inode_lock(inode);
1390         pcci = ll_i2pcci(inode);
1391         if (pcci && pcc_inode_has_layout(pcci)) {
1392                 LASSERT(atomic_read(&pcci->pcci_refcount) > 0);
1393                 atomic_inc(&pcci->pcci_active_ios);
1394                 *cached = true;
1395         } else {
1396                 *cached = false;
1397         }
1398         pcc_inode_unlock(inode);
1399 }
1400
1401 static void pcc_io_fini(struct inode *inode)
1402 {
1403         struct pcc_inode *pcci = ll_i2pcci(inode);
1404
1405         LASSERT(pcci && atomic_read(&pcci->pcci_active_ios) > 0);
1406         if (atomic_dec_and_test(&pcci->pcci_active_ios))
1407                 wake_up_all(&pcci->pcci_waitq);
1408 }
1409
1410
1411 static ssize_t
1412 __pcc_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1413 {
1414         struct file *file = iocb->ki_filp;
1415
1416 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
1417         return file->f_op->read_iter(iocb, iter);
1418 #else
1419         struct iovec iov;
1420         struct iov_iter i;
1421         ssize_t bytes = 0;
1422
1423         iov_for_each(iov, i, *iter) {
1424                 ssize_t res;
1425
1426                 res = file->f_op->aio_read(iocb, &iov, 1, iocb->ki_pos);
1427                 if (-EIOCBQUEUED == res)
1428                         res = wait_on_sync_kiocb(iocb);
1429                 if (res <= 0) {
1430                         if (bytes == 0)
1431                                 bytes = res;
1432                         break;
1433                 }
1434
1435                 bytes += res;
1436                 if (res < iov.iov_len)
1437                         break;
1438         }
1439
1440         if (bytes > 0)
1441                 iov_iter_advance(iter, bytes);
1442         return bytes;
1443 #endif
1444 }
1445
1446 ssize_t pcc_file_read_iter(struct kiocb *iocb,
1447                            struct iov_iter *iter, bool *cached)
1448 {
1449         struct file *file = iocb->ki_filp;
1450         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1451         struct pcc_file *pccf = &fd->fd_pcc_file;
1452         struct inode *inode = file_inode(file);
1453         ssize_t result;
1454
1455         ENTRY;
1456
1457         if (pccf->pccf_file == NULL) {
1458                 *cached = false;
1459                 RETURN(0);
1460         }
1461
1462         pcc_io_init(inode, cached);
1463         if (!*cached)
1464                 RETURN(0);
1465
1466         iocb->ki_filp = pccf->pccf_file;
1467         /* generic_file_aio_read does not support ext4-dax,
1468          * __pcc_file_read_iter uses ->aio_read hook directly
1469          * to add support for ext4-dax.
1470          */
1471         result = __pcc_file_read_iter(iocb, iter);
1472         iocb->ki_filp = file;
1473
1474         pcc_io_fini(inode);
1475         RETURN(result);
1476 }
1477
1478 static ssize_t
1479 __pcc_file_write_iter(struct kiocb *iocb, struct iov_iter *iter)
1480 {
1481         struct file *file = iocb->ki_filp;
1482
1483 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
1484         return file->f_op->write_iter(iocb, iter);
1485 #else
1486         struct iovec iov;
1487         struct iov_iter i;
1488         ssize_t bytes = 0;
1489
1490         iov_for_each(iov, i, *iter) {
1491                 ssize_t res;
1492
1493                 res = file->f_op->aio_write(iocb, &iov, 1, iocb->ki_pos);
1494                 if (-EIOCBQUEUED == res)
1495                         res = wait_on_sync_kiocb(iocb);
1496                 if (res <= 0) {
1497                         if (bytes == 0)
1498                                 bytes = res;
1499                         break;
1500                 }
1501
1502                 bytes += res;
1503                 if (res < iov.iov_len)
1504                         break;
1505         }
1506
1507         if (bytes > 0)
1508                 iov_iter_advance(iter, bytes);
1509         return bytes;
1510 #endif
1511 }
1512
1513 ssize_t pcc_file_write_iter(struct kiocb *iocb,
1514                             struct iov_iter *iter, bool *cached)
1515 {
1516         struct file *file = iocb->ki_filp;
1517         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1518         struct pcc_file *pccf = &fd->fd_pcc_file;
1519         struct inode *inode = file_inode(file);
1520         ssize_t result;
1521
1522         ENTRY;
1523
1524         if (pccf->pccf_file == NULL) {
1525                 *cached = false;
1526                 RETURN(0);
1527         }
1528
1529         if (pccf->pccf_type != LU_PCC_READWRITE) {
1530                 *cached = false;
1531                 RETURN(-EAGAIN);
1532         }
1533
1534         pcc_io_init(inode, cached);
1535         if (!*cached)
1536                 RETURN(0);
1537
1538         if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_PCC_FAKE_ERROR))
1539                 GOTO(out, result = -ENOSPC);
1540
1541         iocb->ki_filp = pccf->pccf_file;
1542
1543         /* Since __pcc_file_write_iter makes write calls via
1544          * the normal vfs interface to the local PCC file system,
1545          * the inode lock is not needed.
1546          */
1547         result = __pcc_file_write_iter(iocb, iter);
1548         iocb->ki_filp = file;
1549 out:
1550         pcc_io_fini(inode);
1551         RETURN(result);
1552 }
1553
1554 int pcc_inode_setattr(struct inode *inode, struct iattr *attr,
1555                       bool *cached)
1556 {
1557         int rc;
1558         const struct cred *old_cred;
1559         struct iattr attr2 = *attr;
1560         struct dentry *pcc_dentry;
1561         struct pcc_inode *pcci;
1562
1563         ENTRY;
1564
1565         if (!S_ISREG(inode->i_mode)) {
1566                 *cached = false;
1567                 RETURN(0);
1568         }
1569
1570         pcc_io_init(inode, cached);
1571         if (!*cached)
1572                 RETURN(0);
1573
1574         attr2.ia_valid = attr->ia_valid & (ATTR_SIZE | ATTR_ATIME |
1575                          ATTR_ATIME_SET | ATTR_MTIME | ATTR_MTIME_SET |
1576                          ATTR_CTIME | ATTR_UID | ATTR_GID);
1577         pcci = ll_i2pcci(inode);
1578         pcc_dentry = pcci->pcci_path.dentry;
1579         inode_lock(pcc_dentry->d_inode);
1580         old_cred = override_creds(pcc_super_cred(inode->i_sb));
1581         rc = pcc_dentry->d_inode->i_op->setattr(pcc_dentry, &attr2);
1582         revert_creds(old_cred);
1583         inode_unlock(pcc_dentry->d_inode);
1584
1585         pcc_io_fini(inode);
1586         RETURN(rc);
1587 }
1588
1589 int pcc_inode_getattr(struct inode *inode, bool *cached)
1590 {
1591         struct ll_inode_info *lli = ll_i2info(inode);
1592         const struct cred *old_cred;
1593         struct kstat stat;
1594         s64 atime;
1595         s64 mtime;
1596         s64 ctime;
1597         int rc;
1598
1599         ENTRY;
1600
1601         if (!S_ISREG(inode->i_mode)) {
1602                 *cached = false;
1603                 RETURN(0);
1604         }
1605
1606         pcc_io_init(inode, cached);
1607         if (!*cached)
1608                 RETURN(0);
1609
1610         old_cred = override_creds(pcc_super_cred(inode->i_sb));
1611         rc = ll_vfs_getattr(&ll_i2pcci(inode)->pcci_path, &stat);
1612         revert_creds(old_cred);
1613         if (rc)
1614                 GOTO(out, rc);
1615
1616         ll_inode_size_lock(inode);
1617         if (inode->i_atime.tv_sec < lli->lli_atime ||
1618             lli->lli_update_atime) {
1619                 inode->i_atime.tv_sec = lli->lli_atime;
1620                 lli->lli_update_atime = 0;
1621         }
1622         inode->i_mtime.tv_sec = lli->lli_mtime;
1623         inode->i_ctime.tv_sec = lli->lli_ctime;
1624
1625         atime = inode->i_atime.tv_sec;
1626         mtime = inode->i_mtime.tv_sec;
1627         ctime = inode->i_ctime.tv_sec;
1628
1629         if (atime < stat.atime.tv_sec)
1630                 atime = stat.atime.tv_sec;
1631
1632         if (ctime < stat.ctime.tv_sec)
1633                 ctime = stat.ctime.tv_sec;
1634
1635         if (mtime < stat.mtime.tv_sec)
1636                 mtime = stat.mtime.tv_sec;
1637
1638         i_size_write(inode, stat.size);
1639         inode->i_blocks = stat.blocks;
1640
1641         inode->i_atime.tv_sec = atime;
1642         inode->i_mtime.tv_sec = mtime;
1643         inode->i_ctime.tv_sec = ctime;
1644
1645         ll_inode_size_unlock(inode);
1646 out:
1647         pcc_io_fini(inode);
1648         RETURN(rc);
1649 }
1650
1651 ssize_t pcc_file_splice_read(struct file *in_file, loff_t *ppos,
1652                              struct pipe_inode_info *pipe,
1653                              size_t count, unsigned int flags,
1654                              bool *cached)
1655 {
1656         struct inode *inode = file_inode(in_file);
1657         struct ll_file_data *fd = LUSTRE_FPRIVATE(in_file);
1658         struct file *pcc_file = fd->fd_pcc_file.pccf_file;
1659         ssize_t result;
1660
1661         ENTRY;
1662
1663         *cached = false;
1664         if (!pcc_file)
1665                 RETURN(0);
1666
1667         if (!file_inode(pcc_file)->i_fop->splice_read)
1668                 RETURN(-ENOTSUPP);
1669
1670         pcc_io_init(inode, cached);
1671         if (!*cached)
1672                 RETURN(0);
1673
1674         result = file_inode(pcc_file)->i_fop->splice_read(pcc_file,
1675                                                           ppos, pipe, count,
1676                                                           flags);
1677
1678         pcc_io_fini(inode);
1679         RETURN(result);
1680 }
1681
1682 int pcc_fsync(struct file *file, loff_t start, loff_t end,
1683               int datasync, bool *cached)
1684 {
1685         struct inode *inode = file_inode(file);
1686         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1687         struct file *pcc_file = fd->fd_pcc_file.pccf_file;
1688         int rc;
1689
1690         ENTRY;
1691
1692         if (!pcc_file) {
1693                 *cached = false;
1694                 RETURN(0);
1695         }
1696
1697         pcc_io_init(inode, cached);
1698         if (!*cached)
1699                 RETURN(0);
1700
1701 #ifdef HAVE_FILE_FSYNC_4ARGS
1702         rc = file_inode(pcc_file)->i_fop->fsync(pcc_file,
1703                                                 start, end, datasync);
1704 #elif defined(HAVE_FILE_FSYNC_2ARGS)
1705         rc = file_inode(pcc_file)->i_fop->fsync(pcc_file, datasync);
1706 #else
1707         rc = file_inode(pcc_file)->i_fop->fsync(pcc_file,
1708                                 file_dentry(dentry), datasync);
1709 #endif
1710
1711         pcc_io_fini(inode);
1712         RETURN(rc);
1713 }
1714
1715 int pcc_file_mmap(struct file *file, struct vm_area_struct *vma,
1716                   bool *cached)
1717 {
1718         struct inode *inode = file_inode(file);
1719         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1720         struct file *pcc_file = fd->fd_pcc_file.pccf_file;
1721         struct pcc_inode *pcci;
1722         int rc = 0;
1723
1724         ENTRY;
1725
1726         if (!pcc_file || !file_inode(pcc_file)->i_fop->mmap) {
1727                 *cached = false;
1728                 RETURN(0);
1729         }
1730
1731         pcc_inode_lock(inode);
1732         pcci = ll_i2pcci(inode);
1733         if (pcci && pcc_inode_has_layout(pcci)) {
1734                 LASSERT(atomic_read(&pcci->pcci_refcount) > 1);
1735                 *cached = true;
1736                 vma->vm_file = pcc_file;
1737                 rc = file_inode(pcc_file)->i_fop->mmap(pcc_file, vma);
1738                 vma->vm_file = file;
1739                 /* Save the vm ops of backend PCC */
1740                 vma->vm_private_data = (void *)vma->vm_ops;
1741         } else {
1742                 *cached = false;
1743         }
1744         pcc_inode_unlock(inode);
1745
1746         RETURN(rc);
1747 }
1748
1749 void pcc_vm_open(struct vm_area_struct *vma)
1750 {
1751         struct pcc_inode *pcci;
1752         struct file *file = vma->vm_file;
1753         struct inode *inode = file_inode(file);
1754         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1755         struct file *pcc_file = fd->fd_pcc_file.pccf_file;
1756         struct vm_operations_struct *pcc_vm_ops = vma->vm_private_data;
1757
1758         ENTRY;
1759
1760         if (!pcc_file || !pcc_vm_ops || !pcc_vm_ops->open)
1761                 RETURN_EXIT;
1762
1763         pcc_inode_lock(inode);
1764         pcci = ll_i2pcci(inode);
1765         if (pcci && pcc_inode_has_layout(pcci)) {
1766                 vma->vm_file = pcc_file;
1767                 pcc_vm_ops->open(vma);
1768                 vma->vm_file = file;
1769         }
1770         pcc_inode_unlock(inode);
1771         EXIT;
1772 }
1773
1774 void pcc_vm_close(struct vm_area_struct *vma)
1775 {
1776         struct file *file = vma->vm_file;
1777         struct inode *inode = file_inode(file);
1778         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1779         struct file *pcc_file = fd->fd_pcc_file.pccf_file;
1780         struct vm_operations_struct *pcc_vm_ops = vma->vm_private_data;
1781
1782         ENTRY;
1783
1784         if (!pcc_file || !pcc_vm_ops || !pcc_vm_ops->close)
1785                 RETURN_EXIT;
1786
1787         pcc_inode_lock(inode);
1788         /* Layout lock maybe revoked here */
1789         vma->vm_file = pcc_file;
1790         pcc_vm_ops->close(vma);
1791         vma->vm_file = file;
1792         pcc_inode_unlock(inode);
1793         EXIT;
1794 }
1795
1796 int pcc_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
1797                      bool *cached)
1798 {
1799         struct page *page = vmf->page;
1800         struct mm_struct *mm = vma->vm_mm;
1801         struct file *file = vma->vm_file;
1802         struct inode *inode = file_inode(file);
1803         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1804         struct file *pcc_file = fd->fd_pcc_file.pccf_file;
1805         struct vm_operations_struct *pcc_vm_ops = vma->vm_private_data;
1806         int rc;
1807
1808         ENTRY;
1809
1810         if (!pcc_file || !pcc_vm_ops) {
1811                 *cached = false;
1812                 RETURN(0);
1813         }
1814
1815         if (!pcc_vm_ops->page_mkwrite &&
1816             page->mapping == pcc_file->f_mapping) {
1817                 CDEBUG(D_MMAP,
1818                        "%s: PCC backend fs not support ->page_mkwrite()\n",
1819                        ll_i2sbi(inode)->ll_fsname);
1820                 pcc_ioctl_detach(inode);
1821                 up_read(&mm->mmap_sem);
1822                 RETURN(VM_FAULT_RETRY | VM_FAULT_NOPAGE);
1823         }
1824         /* Pause to allow for a race with concurrent detach */
1825         OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_PCC_MKWRITE_PAUSE, cfs_fail_val);
1826
1827         pcc_io_init(inode, cached);
1828         if (!*cached) {
1829                 /* This happens when the file is detached from PCC after got
1830                  * the fault page via ->fault() on the inode of the PCC copy.
1831                  * Here it can not simply fall back to normal Lustre I/O path.
1832                  * The reason is that the address space of fault page used by
1833                  * ->page_mkwrite() is still the one of PCC inode. In the
1834                  * normal Lustre ->page_mkwrite() I/O path, it will be wrongly
1835                  * handled as the address space of the fault page is not
1836                  * consistent with the one of the Lustre inode (though the
1837                  * fault page was truncated).
1838                  * As the file is detached from PCC, the fault page must
1839                  * be released frist, and retry the mmap write (->fault() and
1840                  * ->page_mkwrite).
1841                  * We use an ugly and tricky method by returning
1842                  * VM_FAULT_NOPAGE | VM_FAULT_RETRY to the caller
1843                  * __do_page_fault and retry the memory fault handling.
1844                  */
1845                 if (page->mapping == pcc_file->f_mapping) {
1846                         *cached = true;
1847                         up_read(&mm->mmap_sem);
1848                         RETURN(VM_FAULT_RETRY | VM_FAULT_NOPAGE);
1849                 }
1850
1851                 RETURN(0);
1852         }
1853
1854         /*
1855          * This fault injection can also be used to simulate -ENOSPC and
1856          * -EDQUOT failure of underlying PCC backend fs.
1857          */
1858         if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_PCC_DETACH_MKWRITE)) {
1859                 pcc_io_fini(inode);
1860                 pcc_ioctl_detach(inode);
1861                 up_read(&mm->mmap_sem);
1862                 RETURN(VM_FAULT_RETRY | VM_FAULT_NOPAGE);
1863         }
1864
1865         vma->vm_file = pcc_file;
1866 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
1867         rc = pcc_vm_ops->page_mkwrite(vmf);
1868 #else
1869         rc = pcc_vm_ops->page_mkwrite(vma, vmf);
1870 #endif
1871         vma->vm_file = file;
1872
1873         pcc_io_fini(inode);
1874         RETURN(rc);
1875 }
1876
1877 int pcc_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
1878               bool *cached)
1879 {
1880         struct file *file = vma->vm_file;
1881         struct inode *inode = file_inode(file);
1882         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1883         struct file *pcc_file = fd->fd_pcc_file.pccf_file;
1884         struct vm_operations_struct *pcc_vm_ops = vma->vm_private_data;
1885         int rc;
1886
1887         ENTRY;
1888
1889         if (!pcc_file || !pcc_vm_ops || !pcc_vm_ops->fault) {
1890                 *cached = false;
1891                 RETURN(0);
1892         }
1893
1894         pcc_io_init(inode, cached);
1895         if (!*cached)
1896                 RETURN(0);
1897
1898         vma->vm_file = pcc_file;
1899 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
1900         rc = pcc_vm_ops->fault(vmf);
1901 #else
1902         rc = pcc_vm_ops->fault(vma, vmf);
1903 #endif
1904         vma->vm_file = file;
1905
1906         pcc_io_fini(inode);
1907         RETURN(rc);
1908 }
1909
1910 static void pcc_layout_wait(struct pcc_inode *pcci)
1911 {
1912         struct l_wait_info lwi = { 0 };
1913
1914         while (atomic_read(&pcci->pcci_active_ios) > 0) {
1915                 CDEBUG(D_CACHE, "Waiting for IO completion: %d\n",
1916                        atomic_read(&pcci->pcci_active_ios));
1917                 l_wait_event(pcci->pcci_waitq,
1918                              atomic_read(&pcci->pcci_active_ios) == 0, &lwi);
1919         }
1920 }
1921
1922 static void __pcc_layout_invalidate(struct pcc_inode *pcci)
1923 {
1924         pcci->pcci_type = LU_PCC_NONE;
1925         pcc_layout_gen_set(pcci, CL_LAYOUT_GEN_NONE);
1926         pcc_layout_wait(pcci);
1927 }
1928
1929 void pcc_layout_invalidate(struct inode *inode)
1930 {
1931         struct pcc_inode *pcci;
1932
1933         pcc_inode_lock(inode);
1934         pcci = ll_i2pcci(inode);
1935         if (pcci && pcc_inode_has_layout(pcci)) {
1936                 LASSERT(atomic_read(&pcci->pcci_refcount) > 0);
1937                 __pcc_layout_invalidate(pcci);
1938
1939                 CDEBUG(D_CACHE, "Invalidate "DFID" layout gen %d\n",
1940                        PFID(&ll_i2info(inode)->lli_fid), pcci->pcci_layout_gen);
1941
1942                 pcc_inode_put(pcci);
1943         }
1944         pcc_inode_unlock(inode);
1945 }
1946
1947 static int pcc_inode_remove(struct inode *inode, struct dentry *pcc_dentry)
1948 {
1949         int rc;
1950
1951         rc = ll_vfs_unlink(pcc_dentry->d_parent->d_inode, pcc_dentry);
1952         if (rc)
1953                 CWARN("%s: failed to unlink PCC file %.*s, rc = %d\n",
1954                       ll_i2sbi(inode)->ll_fsname, pcc_dentry->d_name.len,
1955                       pcc_dentry->d_name.name, rc);
1956
1957         return rc;
1958 }
1959
1960 /* Create directory under base if directory does not exist */
1961 static struct dentry *
1962 pcc_mkdir(struct dentry *base, const char *name, umode_t mode)
1963 {
1964         int rc;
1965         struct dentry *dentry;
1966         struct inode *dir = base->d_inode;
1967
1968         inode_lock(dir);
1969         dentry = lookup_one_len(name, base, strlen(name));
1970         if (IS_ERR(dentry))
1971                 goto out;
1972
1973         if (d_is_positive(dentry))
1974                 goto out;
1975
1976         rc = vfs_mkdir(dir, dentry, mode);
1977         if (rc) {
1978                 dput(dentry);
1979                 dentry = ERR_PTR(rc);
1980                 goto out;
1981         }
1982 out:
1983         inode_unlock(dir);
1984         return dentry;
1985 }
1986
1987 static struct dentry *
1988 pcc_mkdir_p(struct dentry *root, char *path, umode_t mode)
1989 {
1990         char *ptr, *entry_name;
1991         struct dentry *parent;
1992         struct dentry *child = ERR_PTR(-EINVAL);
1993
1994         ptr = path;
1995         while (*ptr == '/')
1996                 ptr++;
1997
1998         entry_name = ptr;
1999         parent = dget(root);
2000         while ((ptr = strchr(ptr, '/')) != NULL) {
2001                 *ptr = '\0';
2002                 child = pcc_mkdir(parent, entry_name, mode);
2003                 *ptr = '/';
2004                 dput(parent);
2005                 if (IS_ERR(child))
2006                         break;
2007
2008                 parent = child;
2009                 ptr++;
2010                 entry_name = ptr;
2011         }
2012
2013         return child;
2014 }
2015
2016 /* Create file under base. If file already exist, return failure */
2017 static struct dentry *
2018 pcc_create(struct dentry *base, const char *name, umode_t mode)
2019 {
2020         int rc;
2021         struct dentry *dentry;
2022         struct inode *dir = base->d_inode;
2023
2024         inode_lock(dir);
2025         dentry = lookup_one_len(name, base, strlen(name));
2026         if (IS_ERR(dentry))
2027                 goto out;
2028
2029         if (d_is_positive(dentry))
2030                 goto out;
2031
2032         rc = vfs_create(dir, dentry, mode, LL_VFS_CREATE_FALSE);
2033         if (rc) {
2034                 dput(dentry);
2035                 dentry = ERR_PTR(rc);
2036                 goto out;
2037         }
2038 out:
2039         inode_unlock(dir);
2040         return dentry;
2041 }
2042
2043 static int __pcc_inode_create(struct pcc_dataset *dataset,
2044                               struct lu_fid *fid,
2045                               struct dentry **dentry)
2046 {
2047         char *path;
2048         struct dentry *base;
2049         struct dentry *child;
2050         int rc = 0;
2051
2052         OBD_ALLOC(path, MAX_PCC_DATABASE_PATH);
2053         if (path == NULL)
2054                 return -ENOMEM;
2055
2056         pcc_fid2dataset_path(path, MAX_PCC_DATABASE_PATH, fid);
2057
2058         base = pcc_mkdir_p(dataset->pccd_path.dentry, path, 0);
2059         if (IS_ERR(base)) {
2060                 rc = PTR_ERR(base);
2061                 GOTO(out, rc);
2062         }
2063
2064         snprintf(path, MAX_PCC_DATABASE_PATH, DFID_NOBRACE, PFID(fid));
2065         child = pcc_create(base, path, 0);
2066         if (IS_ERR(child)) {
2067                 rc = PTR_ERR(child);
2068                 GOTO(out_base, rc);
2069         }
2070         *dentry = child;
2071
2072 out_base:
2073         dput(base);
2074 out:
2075         OBD_FREE(path, MAX_PCC_DATABASE_PATH);
2076         return rc;
2077 }
2078
2079 /* TODO: Set the project ID for PCC copy */
2080 int pcc_inode_store_ugpid(struct dentry *dentry, kuid_t uid, kgid_t gid)
2081 {
2082         struct inode *inode = dentry->d_inode;
2083         struct iattr attr;
2084         int rc;
2085
2086         ENTRY;
2087
2088         attr.ia_valid = ATTR_UID | ATTR_GID;
2089         attr.ia_uid = uid;
2090         attr.ia_gid = gid;
2091
2092         inode_lock(inode);
2093         rc = notify_change(dentry, &attr, NULL);
2094         inode_unlock(inode);
2095
2096         RETURN(rc);
2097 }
2098
2099 int pcc_inode_create(struct super_block *sb, struct pcc_dataset *dataset,
2100                      struct lu_fid *fid, struct dentry **pcc_dentry)
2101 {
2102         const struct cred *old_cred;
2103         int rc;
2104
2105         old_cred = override_creds(pcc_super_cred(sb));
2106         rc = __pcc_inode_create(dataset, fid, pcc_dentry);
2107         revert_creds(old_cred);
2108         return rc;
2109 }
2110
2111 int pcc_inode_create_fini(struct pcc_dataset *dataset, struct inode *inode,
2112                           struct dentry *pcc_dentry)
2113 {
2114         const struct cred *old_cred;
2115         struct pcc_inode *pcci;
2116         int rc = 0;
2117
2118         ENTRY;
2119
2120         old_cred = override_creds(pcc_super_cred(inode->i_sb));
2121         pcc_inode_lock(inode);
2122         LASSERT(ll_i2pcci(inode) == NULL);
2123         OBD_SLAB_ALLOC_PTR_GFP(pcci, pcc_inode_slab, GFP_NOFS);
2124         if (pcci == NULL)
2125                 GOTO(out_put, rc = -ENOMEM);
2126
2127         rc = pcc_inode_store_ugpid(pcc_dentry, old_cred->suid,
2128                                    old_cred->sgid);
2129         if (rc)
2130                 GOTO(out_put, rc);
2131
2132         pcc_inode_init(pcci, ll_i2info(inode));
2133         pcc_inode_attach_init(dataset, pcci, pcc_dentry, LU_PCC_READWRITE);
2134
2135         rc = pcc_layout_xattr_set(pcci, 0);
2136         if (rc) {
2137                 (void) pcc_inode_remove(inode, pcci->pcci_path.dentry);
2138                 pcc_inode_put(pcci);
2139                 GOTO(out_unlock, rc);
2140         }
2141
2142         /* Set the layout generation of newly created file with 0 */
2143         pcc_layout_gen_set(pcci, 0);
2144
2145 out_put:
2146         if (rc) {
2147                 (void) pcc_inode_remove(inode, pcc_dentry);
2148                 dput(pcc_dentry);
2149
2150                 if (pcci)
2151                         OBD_SLAB_FREE_PTR(pcci, pcc_inode_slab);
2152         }
2153 out_unlock:
2154         pcc_inode_unlock(inode);
2155         revert_creds(old_cred);
2156         RETURN(rc);
2157 }
2158
2159 static int pcc_filp_write(struct file *filp, const void *buf, ssize_t count,
2160                           loff_t *offset)
2161 {
2162         while (count > 0) {
2163                 ssize_t size;
2164
2165                 size = vfs_write(filp, (const void __user *)buf, count, offset);
2166                 if (size < 0)
2167                         return size;
2168                 count -= size;
2169                 buf += size;
2170         }
2171         return 0;
2172 }
2173
2174 static int pcc_copy_data(struct file *src, struct file *dst)
2175 {
2176         int rc = 0;
2177         ssize_t rc2;
2178         mm_segment_t oldfs;
2179         loff_t pos, offset = 0;
2180         size_t buf_len = 1048576;
2181         void *buf;
2182
2183         ENTRY;
2184
2185         OBD_ALLOC_LARGE(buf, buf_len);
2186         if (buf == NULL)
2187                 RETURN(-ENOMEM);
2188
2189         oldfs = get_fs();
2190         set_fs(KERNEL_DS);
2191         while (1) {
2192                 pos = offset;
2193                 rc2 = vfs_read(src, (void __user *)buf, buf_len, &pos);
2194                 if (rc2 < 0)
2195                         GOTO(out_fs, rc = rc2);
2196                 else if (rc2 == 0)
2197                         break;
2198
2199                 pos = offset;
2200                 rc = pcc_filp_write(dst, buf, rc2, &pos);
2201                 if (rc < 0)
2202                         GOTO(out_fs, rc);
2203                 offset += rc2;
2204         }
2205
2206 out_fs:
2207         set_fs(oldfs);
2208         OBD_FREE_LARGE(buf, buf_len);
2209         RETURN(rc);
2210 }
2211
2212 static int pcc_attach_allowed_check(struct inode *inode)
2213 {
2214         struct ll_inode_info *lli = ll_i2info(inode);
2215         struct pcc_inode *pcci;
2216         int rc = 0;
2217
2218         ENTRY;
2219
2220         pcc_inode_lock(inode);
2221         if (lli->lli_pcc_state & PCC_STATE_FL_ATTACHING)
2222                 GOTO(out_unlock, rc = -EBUSY);
2223
2224         pcci = ll_i2pcci(inode);
2225         if (pcci && pcc_inode_has_layout(pcci))
2226                 GOTO(out_unlock, rc = -EEXIST);
2227
2228         lli->lli_pcc_state |= PCC_STATE_FL_ATTACHING;
2229 out_unlock:
2230         pcc_inode_unlock(inode);
2231         RETURN(rc);
2232 }
2233
2234 int pcc_readwrite_attach(struct file *file, struct inode *inode,
2235                          __u32 archive_id)
2236 {
2237         struct pcc_dataset *dataset;
2238         struct ll_inode_info *lli = ll_i2info(inode);
2239         struct pcc_inode *pcci;
2240         const struct cred *old_cred;
2241         struct dentry *dentry;
2242         struct file *pcc_filp;
2243         struct path path;
2244         int rc;
2245
2246         ENTRY;
2247
2248         rc = pcc_attach_allowed_check(inode);
2249         if (rc)
2250                 RETURN(rc);
2251
2252         dataset = pcc_dataset_get(&ll_i2sbi(inode)->ll_pcc_super,
2253                                   LU_PCC_READWRITE, archive_id);
2254         if (dataset == NULL)
2255                 RETURN(-ENOENT);
2256
2257         old_cred = override_creds(pcc_super_cred(inode->i_sb));
2258         rc = __pcc_inode_create(dataset, &lli->lli_fid, &dentry);
2259         if (rc) {
2260                 revert_creds(old_cred);
2261                 GOTO(out_dataset_put, rc);
2262         }
2263
2264         path.mnt = dataset->pccd_path.mnt;
2265         path.dentry = dentry;
2266 #ifdef HAVE_DENTRY_OPEN_USE_PATH
2267         pcc_filp = dentry_open(&path, O_TRUNC | O_WRONLY | O_LARGEFILE,
2268                                current_cred());
2269 #else
2270         pcc_filp = dentry_open(path.dentry, path.mnt,
2271                                O_TRUNC | O_WRONLY | O_LARGEFILE,
2272                                current_cred());
2273 #endif
2274         if (IS_ERR_OR_NULL(pcc_filp)) {
2275                 rc = pcc_filp == NULL ? -EINVAL : PTR_ERR(pcc_filp);
2276                 revert_creds(old_cred);
2277                 GOTO(out_dentry, rc);
2278         }
2279
2280         rc = pcc_inode_store_ugpid(dentry, old_cred->uid, old_cred->gid);
2281         revert_creds(old_cred);
2282         if (rc)
2283                 GOTO(out_fput, rc);
2284
2285         rc = pcc_copy_data(file, pcc_filp);
2286         if (rc)
2287                 GOTO(out_fput, rc);
2288
2289         /* Pause to allow for a race with concurrent HSM remove */
2290         OBD_FAIL_TIMEOUT(OBD_FAIL_LLITE_PCC_ATTACH_PAUSE, cfs_fail_val);
2291
2292         pcc_inode_lock(inode);
2293         pcci = ll_i2pcci(inode);
2294         LASSERT(!pcci);
2295         OBD_SLAB_ALLOC_PTR_GFP(pcci, pcc_inode_slab, GFP_NOFS);
2296         if (pcci == NULL)
2297                 GOTO(out_unlock, rc = -ENOMEM);
2298
2299         pcc_inode_init(pcci, lli);
2300         pcc_inode_attach_init(dataset, pcci, dentry, LU_PCC_READWRITE);
2301 out_unlock:
2302         pcc_inode_unlock(inode);
2303 out_fput:
2304         fput(pcc_filp);
2305 out_dentry:
2306         if (rc) {
2307                 old_cred = override_creds(pcc_super_cred(inode->i_sb));
2308                 (void) pcc_inode_remove(inode, dentry);
2309                 revert_creds(old_cred);
2310                 dput(dentry);
2311         }
2312 out_dataset_put:
2313         pcc_dataset_put(dataset);
2314         RETURN(rc);
2315 }
2316
2317 int pcc_readwrite_attach_fini(struct file *file, struct inode *inode,
2318                               __u32 gen, bool lease_broken, int rc,
2319                               bool attached)
2320 {
2321         struct ll_inode_info *lli = ll_i2info(inode);
2322         const struct cred *old_cred;
2323         struct pcc_inode *pcci;
2324         __u32 gen2;
2325
2326         ENTRY;
2327
2328         old_cred = override_creds(pcc_super_cred(inode->i_sb));
2329         pcc_inode_lock(inode);
2330         pcci = ll_i2pcci(inode);
2331         lli->lli_pcc_state &= ~PCC_STATE_FL_ATTACHING;
2332         if (rc || lease_broken) {
2333                 if (attached && pcci)
2334                         pcc_inode_put(pcci);
2335
2336                 GOTO(out_unlock, rc);
2337         }
2338
2339         /* PCC inode may be released due to layout lock revocatioin */
2340         if (!pcci)
2341                 GOTO(out_unlock, rc = -ESTALE);
2342
2343         LASSERT(attached);
2344         rc = pcc_layout_xattr_set(pcci, gen);
2345         if (rc)
2346                 GOTO(out_put, rc);
2347
2348         rc = ll_layout_refresh(inode, &gen2);
2349         if (!rc) {
2350                 if (gen2 == gen) {
2351                         pcc_layout_gen_set(pcci, gen);
2352                 } else {
2353                         CDEBUG(D_CACHE,
2354                                DFID" layout changed from %d to %d.\n",
2355                                PFID(ll_inode2fid(inode)), gen, gen2);
2356                         GOTO(out_put, rc = -ESTALE);
2357                 }
2358         }
2359
2360 out_put:
2361         if (rc) {
2362                 (void) pcc_inode_remove(inode, pcci->pcci_path.dentry);
2363                 pcc_inode_put(pcci);
2364         }
2365 out_unlock:
2366         pcc_inode_unlock(inode);
2367         revert_creds(old_cred);
2368         RETURN(rc);
2369 }
2370
2371 int pcc_ioctl_detach(struct inode *inode)
2372 {
2373         struct ll_inode_info *lli = ll_i2info(inode);
2374         struct pcc_inode *pcci;
2375         int rc = 0;
2376
2377         ENTRY;
2378
2379         pcc_inode_lock(inode);
2380         pcci = lli->lli_pcc_inode;
2381         if (!pcci || lli->lli_pcc_state & PCC_STATE_FL_ATTACHING ||
2382             !pcc_inode_has_layout(pcci))
2383                 GOTO(out_unlock, rc = 0);
2384
2385         __pcc_layout_invalidate(pcci);
2386         pcc_inode_put(pcci);
2387
2388 out_unlock:
2389         pcc_inode_unlock(inode);
2390         RETURN(rc);
2391 }
2392
2393 int pcc_ioctl_state(struct file *file, struct inode *inode,
2394                     struct lu_pcc_state *state)
2395 {
2396         int rc = 0;
2397         int count;
2398         char *buf;
2399         char *path;
2400         int buf_len = sizeof(state->pccs_path);
2401         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
2402         struct pcc_file *pccf = &fd->fd_pcc_file;
2403         struct pcc_inode *pcci;
2404
2405         ENTRY;
2406
2407         if (buf_len <= 0)
2408                 RETURN(-EINVAL);
2409
2410         OBD_ALLOC(buf, buf_len);
2411         if (buf == NULL)
2412                 RETURN(-ENOMEM);
2413
2414         pcc_inode_lock(inode);
2415         pcci = ll_i2pcci(inode);
2416         if (pcci == NULL) {
2417                 state->pccs_type = LU_PCC_NONE;
2418                 GOTO(out_unlock, rc = 0);
2419         }
2420
2421         count = atomic_read(&pcci->pcci_refcount);
2422         if (count == 0) {
2423                 state->pccs_type = LU_PCC_NONE;
2424                 state->pccs_open_count = 0;
2425                 GOTO(out_unlock, rc = 0);
2426         }
2427
2428         if (pcc_inode_has_layout(pcci))
2429                 count--;
2430         if (pccf->pccf_file != NULL)
2431                 count--;
2432         state->pccs_type = pcci->pcci_type;
2433         state->pccs_open_count = count;
2434         state->pccs_flags = ll_i2info(inode)->lli_pcc_state;
2435 #ifdef HAVE_DENTRY_PATH_RAW
2436         path = dentry_path_raw(pcci->pcci_path.dentry, buf, buf_len);
2437         if (IS_ERR(path))
2438                 GOTO(out_unlock, rc = PTR_ERR(path));
2439 #else
2440         path = "UNKNOWN";
2441 #endif
2442
2443         if (strlcpy(state->pccs_path, path, buf_len) >= buf_len)
2444                 GOTO(out_unlock, rc = -ENAMETOOLONG);
2445
2446 out_unlock:
2447         pcc_inode_unlock(inode);
2448         OBD_FREE(buf, buf_len);
2449         RETURN(rc);
2450 }