lustre/include/lustre_export.h

   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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32 /** \defgroup obd_export PortalRPC export definitions
  33  *
  34  * @{
  35  */
  36
  37 #ifndef __EXPORT_H
  38 #define __EXPORT_H
  39
  40 /** \defgroup export export
  41  *
  42  * @{
  43  */
  44
  45 #include <linux/workqueue.h>
  46
  47 #include <lprocfs_status.h>
  48 #include <uapi/linux/lustre/lustre_idl.h>
  49 #include <lustre_dlm.h>
  50
  51 struct mds_client_data;
  52 struct mdt_client_data;
  53 struct mds_idmap_table;
  54 struct mdt_idmap_table;
  55
  56 /**
  57  * Target-specific export data
  58  */
  59 struct tg_export_data {
  60         /** Protects ted_lcd, ted_reply_* and
  61          * ted_release_* fields below */
  62         struct mutex            ted_lcd_lock;
  63         /** Per-client data for each export */
  64         struct lsd_client_data  *ted_lcd;
  65         /** Offset of record in last_rcvd file */
  66         loff_t                  ted_lr_off;
  67         /** Client index in last_rcvd file */
  68         int                     ted_lr_idx;
  69
  70         /**
  71          * ted_nodemap_lock is used to ensure that the nodemap is not destroyed
  72          * between the time that ted_nodemap is checked for NULL, and a
  73          * reference is taken. Modifications to ted_nodemap require that the
  74          * active_config_lock and the nodemap(s)'s nm_member_list_lock be
  75          * taken, as well as ted_nodemap_lock, so the export can be properly
  76          * added to or removed from the nodemap's member list. When an export
  77          * is added to a nodemap, a reference on that nodemap must be taken.
  78          * That reference can be put only after ted_nodemap no longer refers to
  79          * it.
  80          */
  81         spinlock_t              ted_nodemap_lock;
  82         struct lu_nodemap       *ted_nodemap;
  83         struct list_head        ted_nodemap_member;
  84
  85         /** last version of nodemap config sent to client */
  86         __u64                   ted_nodemap_version;
  87
  88         /* Every reply data fields below are
  89          * protected by ted_lcd_lock */
  90         /** List of reply data */
  91         struct list_head        ted_reply_list;
  92         int                     ted_reply_cnt;
  93         /** Reply data with highest transno is retained */
  94         struct tg_reply_data    *ted_reply_last;
  95         /* Statistics */
  96         int                     ted_reply_max; /* high water mark */
  97         int                     ted_release_xid;
  98         int                     ted_release_tag;
  99         /* grants */
 100         long                    ted_dirty;    /* in bytes */
 101         long                    ted_grant;    /* in bytes */
 102         long                    ted_pending;  /* bytes just being written */
 103         __u8                    ted_pagebits; /* log2 of client page size */
 104
 105         /**
 106          * File Modification Data (FMD) tracking
 107          */
 108         spinlock_t              ted_fmd_lock; /* protects ted_fmd_list */
 109         struct list_head        ted_fmd_list; /* FIDs being modified */
 110         int                     ted_fmd_count;/* items in ted_fmd_list */
 111 };
 112
 113 /**
 114  * MDT-specific export data
 115  */
 116 struct mdt_export_data {
 117         struct tg_export_data   med_ted;
 118         /** List of all files opened by client on this MDT */
 119         struct list_head        med_open_head;
 120         spinlock_t              med_open_lock; /* med_open_head, mfd_list */
 121 };
 122
 123 struct ec_export_data { /* echo client */
 124         struct list_head        eced_locks;
 125 };
 126
 127 /* In-memory access to client data from OST struct */
 128 /** Filter (oss-side) specific import data */
 129 struct filter_export_data {
 130         struct tg_export_data   fed_ted;
 131         __u64                   fed_lastid_gen;
 132         /* count of SOFT_SYNC RPCs, which will be reset after
 133          * ofd_soft_sync_limit number of RPCs, and trigger a sync. */
 134         atomic_t                fed_soft_sync_count;
 135         __u32                   fed_group;
 136 };
 137
 138 struct mgs_export_data {
 139         struct list_head        med_clients;    /* mgc fs client via this exp */
 140         spinlock_t              med_lock;       /* protect med_clients */
 141 };
 142
 143 /**
 144  * per-NID statistics structure.
 145  * It tracks access patterns to this export on a per-client-NID basis
 146  */
 147 struct nid_stat {
 148         lnet_nid_t               nid;
 149         struct hlist_node        nid_hash;
 150         struct list_head         nid_list;
 151         struct obd_device       *nid_obd;
 152         struct proc_dir_entry   *nid_proc;
 153         struct lprocfs_stats    *nid_stats;
 154         struct lprocfs_stats    *nid_ldlm_stats;
 155         atomic_t                 nid_exp_ref_count; /* for obd_nid_stats_hash
 156                                                        exp_nid_stats */
 157 };
 158
 159 #define nidstat_getref(nidstat)                                                \
 160 do {                                                                           \
 161         atomic_inc(&(nidstat)->nid_exp_ref_count);                         \
 162 } while(0)
 163
 164 #define nidstat_putref(nidstat)                                                \
 165 do {                                                                           \
 166         atomic_dec(&(nidstat)->nid_exp_ref_count);                         \
 167         LASSERTF(atomic_read(&(nidstat)->nid_exp_ref_count) >= 0,          \
 168                  "stat %p nid_exp_ref_count < 0\n", nidstat);                  \
 169 } while(0)
 170
 171 enum obd_option {
 172         OBD_OPT_FORCE =         0x0001,
 173         OBD_OPT_FAILOVER =      0x0002,
 174         OBD_OPT_ABORT_RECOV =   0x0004,
 175 };
 176
 177 /**
 178  * Export structure. Represents target-side of connection in portals.
 179  * Also used in Lustre to connect between layers on the same node when
 180  * there is no network-connection in-between.
 181  * For every connected client there is an export structure on the server
 182  * attached to the same obd device.
 183  */
 184 struct obd_export {
 185         /**
 186          * Export handle, it's id is provided to client on connect
 187          * Subsequent client RPCs contain this handle id to identify
 188          * what export they are talking to.
 189          */
 190         struct portals_handle   exp_handle;
 191         atomic_t                exp_refcount;
 192         /**
 193          * Set of counters below is to track where export references are
 194          * kept. The exp_rpc_count is used for reconnect handling also,
 195          * the cb_count and locks_count are for debug purposes only for now.
 196          * The sum of them should be less than exp_refcount by 3
 197          */
 198         atomic_t                exp_rpc_count; /* RPC references */
 199         atomic_t                exp_cb_count; /* Commit callback references */
 200         /** Number of queued replay requests to be processes */
 201         atomic_t                exp_replay_count;
 202         atomic_t                exp_locks_count; /** Lock references */
 203 #if LUSTRE_TRACKS_LOCK_EXP_REFS
 204         struct list_head        exp_locks_list;
 205         spinlock_t              exp_locks_list_guard;
 206 #endif
 207         /** UUID of client connected to this export */
 208         struct obd_uuid         exp_client_uuid;
 209         /** To link all exports on an obd device */
 210         struct list_head        exp_obd_chain;
 211         /** work_struct for destruction of export */
 212         struct work_struct      exp_zombie_work;
 213         /* Unlinked export list */
 214         struct list_head        exp_stale_list;
 215         struct hlist_node       exp_uuid_hash;  /** uuid-export hash*/
 216         struct hlist_node       exp_nid_hash;   /** nid-export hash */
 217         struct hlist_node       exp_gen_hash;   /** last_rcvd clt gen hash */
 218         /**
 219          * All exports eligible for ping evictor are linked into a list
 220          * through this field in "most time since last request on this export"
 221          * order
 222          * protected by obd_dev_lock
 223          */
 224         struct list_head        exp_obd_chain_timed;
 225         /** Obd device of this export */
 226         struct obd_device      *exp_obd;
 227         /**
 228          * "reverse" import to send requests (e.g. from ldlm) back to client
 229          * exp_lock protect its change
 230          */
 231         struct obd_import        *exp_imp_reverse;
 232         struct nid_stat          *exp_nid_stats;
 233         /** Active connetion */
 234         struct ptlrpc_connection *exp_connection;
 235         /** Connection count value from last successful reconnect rpc */
 236         __u32                     exp_conn_cnt;
 237         /** Hash list of all ldlm locks granted on this export */
 238         struct cfs_hash          *exp_lock_hash;
 239         /**
 240          * Hash list for Posix lock deadlock detection, added with
 241          * ldlm_lock::l_exp_flock_hash.
 242          */
 243         struct cfs_hash        *exp_flock_hash;
 244         struct list_head        exp_outstanding_replies;
 245         struct list_head        exp_uncommitted_replies;
 246         spinlock_t              exp_uncommitted_replies_lock;
 247         /** Last committed transno for this export */
 248         __u64                   exp_last_committed;
 249         /** When was last request received */
 250         time64_t                exp_last_request_time;
 251         /** On replay all requests waiting for replay are linked here */
 252         struct list_head        exp_req_replay_queue;
 253         /**
 254          * protects exp_flags, exp_outstanding_replies and the change
 255          * of exp_imp_reverse
 256          */
 257         spinlock_t              exp_lock;
 258         /** Compatibility flags for this export are embedded into
 259          *  exp_connect_data */
 260         struct obd_connect_data exp_connect_data;
 261         enum obd_option         exp_flags;
 262         unsigned long           exp_failed:1,
 263                                 exp_in_recovery:1,
 264                                 exp_disconnected:1,
 265                                 exp_connecting:1,
 266                                 /** VBR: export missed recovery */
 267                                 exp_delayed:1,
 268                                 /** VBR: failed version checking */
 269                                 exp_vbr_failed:1,
 270                                 exp_req_replay_needed:1,
 271                                 exp_lock_replay_needed:1,
 272                                 exp_need_sync:1,
 273                                 exp_flvr_changed:1,
 274                                 exp_flvr_adapt:1,
 275                                 /* if to swap nidtbl entries for 2.2 clients.
 276                                  * Only used by the MGS to fix LU-1644. */
 277                                 exp_need_mne_swab:1,
 278                                 /* The export already got final replay ping
 279                                  * request. */
 280                                 exp_replay_done:1;
 281         /* also protected by exp_lock */
 282         enum lustre_sec_part    exp_sp_peer;
 283         struct sptlrpc_flavor   exp_flvr;               /* current */
 284         struct sptlrpc_flavor   exp_flvr_old[2];        /* about-to-expire */
 285         time64_t                exp_flvr_expire[2];     /* seconds */
 286
 287         /** protects exp_hp_rpcs */
 288         spinlock_t              exp_rpc_lock;
 289         struct list_head        exp_hp_rpcs;    /* (potential) HP RPCs */
 290         struct list_head        exp_reg_rpcs;  /* RPC being handled */
 291
 292         /** blocking dlm lock list, protected by exp_bl_list_lock */
 293         struct list_head        exp_bl_list;
 294         spinlock_t              exp_bl_list_lock;
 295
 296         /** Target specific data */
 297         union {
 298                 struct tg_export_data     eu_target_data;
 299                 struct mdt_export_data    eu_mdt_data;
 300                 struct filter_export_data eu_filter_data;
 301                 struct ec_export_data     eu_ec_data;
 302                 struct mgs_export_data    eu_mgs_data;
 303         } u;
 304
 305         struct adaptive_timeout    exp_bl_lock_at;
 306
 307         /** highest XID received by export client that has no
 308          * unreceived lower-numbered XID
 309          */
 310         __u64                     exp_last_xid;
 311 };
 312
 313 #define exp_target_data u.eu_target_data
 314 #define exp_mdt_data    u.eu_mdt_data
 315 #define exp_filter_data u.eu_filter_data
 316 #define exp_ec_data     u.eu_ec_data
 317
 318 static inline __u64 *exp_connect_flags_ptr(struct obd_export *exp)
 319 {
 320         return &exp->exp_connect_data.ocd_connect_flags;
 321 }
 322
 323 static inline __u64 exp_connect_flags(struct obd_export *exp)
 324 {
 325         return *exp_connect_flags_ptr(exp);
 326 }
 327
 328 static inline __u64 *exp_connect_flags2_ptr(struct obd_export *exp)
 329 {
 330         return &exp->exp_connect_data.ocd_connect_flags2;
 331 }
 332
 333 static inline __u64 exp_connect_flags2(struct obd_export *exp)
 334 {
 335         if (exp_connect_flags(exp) & OBD_CONNECT_FLAGS2)
 336                 return *exp_connect_flags2_ptr(exp);
 337         return 0;
 338 }
 339
 340 static inline int exp_max_brw_size(struct obd_export *exp)
 341 {
 342         LASSERT(exp != NULL);
 343         if (exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE)
 344                 return exp->exp_connect_data.ocd_brw_size;
 345
 346         return ONE_MB_BRW_SIZE;
 347 }
 348
 349 static inline int exp_connect_multibulk(struct obd_export *exp)
 350 {
 351         return exp_max_brw_size(exp) > ONE_MB_BRW_SIZE;
 352 }
 353
 354 static inline int exp_connect_cancelset(struct obd_export *exp)
 355 {
 356         LASSERT(exp != NULL);
 357         return !!(exp_connect_flags(exp) & OBD_CONNECT_CANCELSET);
 358 }
 359
 360 static inline int exp_connect_lru_resize(struct obd_export *exp)
 361 {
 362         LASSERT(exp != NULL);
 363         return !!(exp_connect_flags(exp) & OBD_CONNECT_LRU_RESIZE);
 364 }
 365
 366 static inline int exp_connect_vbr(struct obd_export *exp)
 367 {
 368         LASSERT(exp != NULL);
 369         LASSERT(exp->exp_connection);
 370         return !!(exp_connect_flags(exp) & OBD_CONNECT_VBR);
 371 }
 372
 373 static inline int exp_connect_umask(struct obd_export *exp)
 374 {
 375         return !!(exp_connect_flags(exp) & OBD_CONNECT_UMASK);
 376 }
 377
 378 static inline int imp_connect_lru_resize(struct obd_import *imp)
 379 {
 380         struct obd_connect_data *ocd;
 381
 382         LASSERT(imp != NULL);
 383         ocd = &imp->imp_connect_data;
 384         return !!(ocd->ocd_connect_flags & OBD_CONNECT_LRU_RESIZE);
 385 }
 386
 387 static inline int exp_connect_layout(struct obd_export *exp)
 388 {
 389         return !!(exp_connect_flags(exp) & OBD_CONNECT_LAYOUTLOCK);
 390 }
 391
 392 static inline bool exp_connect_lvb_type(struct obd_export *exp)
 393 {
 394         LASSERT(exp != NULL);
 395         if (exp_connect_flags(exp) & OBD_CONNECT_LVB_TYPE)
 396                 return true;
 397         else
 398                 return false;
 399 }
 400
 401 static inline bool imp_connect_lvb_type(struct obd_import *imp)
 402 {
 403         struct obd_connect_data *ocd;
 404
 405         LASSERT(imp != NULL);
 406         ocd = &imp->imp_connect_data;
 407         if (ocd->ocd_connect_flags & OBD_CONNECT_LVB_TYPE)
 408                 return true;
 409         else
 410                 return false;
 411 }
 412
 413 static inline bool imp_connect_disp_stripe(struct obd_import *imp)
 414 {
 415         struct obd_connect_data *ocd;
 416
 417         LASSERT(imp != NULL);
 418         ocd = &imp->imp_connect_data;
 419         return ocd->ocd_connect_flags & OBD_CONNECT_DISP_STRIPE;
 420 }
 421
 422 static inline bool imp_connect_shortio(struct obd_import *imp)
 423 {
 424         struct obd_connect_data *ocd = &imp->imp_connect_data;
 425
 426         return ocd->ocd_connect_flags & OBD_CONNECT_SHORTIO;
 427 }
 428
 429 static inline __u64 exp_connect_ibits(struct obd_export *exp)
 430 {
 431         struct obd_connect_data *ocd;
 432
 433         ocd = &exp->exp_connect_data;
 434         return ocd->ocd_ibits_known;
 435 }
 436
 437 static inline int exp_connect_large_acl(struct obd_export *exp)
 438 {
 439         return !!(exp_connect_flags(exp) & OBD_CONNECT_LARGE_ACL);
 440 }
 441
 442 static inline int exp_connect_lockahead_old(struct obd_export *exp)
 443 {
 444         return !!(exp_connect_flags(exp) & OBD_CONNECT_LOCKAHEAD_OLD);
 445 }
 446
 447 static inline int exp_connect_lockahead(struct obd_export *exp)
 448 {
 449         return !!(exp_connect_flags2(exp) & OBD_CONNECT2_LOCKAHEAD);
 450 }
 451
 452 static inline int exp_connect_overstriping(struct obd_export *exp)
 453 {
 454         return !!(exp_connect_flags2(exp) & OBD_CONNECT2_OVERSTRIPING);
 455 }
 456
 457 static inline int exp_connect_flr(struct obd_export *exp)
 458 {
 459         return !!(exp_connect_flags2(exp) & OBD_CONNECT2_FLR);
 460 }
 461
 462 static inline int exp_connect_lock_convert(struct obd_export *exp)
 463 {
 464         return !!(exp_connect_flags2(exp) & OBD_CONNECT2_LOCK_CONVERT);
 465 }
 466
 467 extern struct obd_export *class_conn2export(struct lustre_handle *conn);
 468
 469 static inline int exp_connect_archive_id_array(struct obd_export *exp)
 470 {
 471         return !!(exp_connect_flags2(exp) & OBD_CONNECT2_ARCHIVE_ID_ARRAY);
 472 }
 473
 474 static inline int exp_connect_sepol(struct obd_export *exp)
 475 {
 476         return !!(exp_connect_flags2(exp) & OBD_CONNECT2_SELINUX_POLICY);
 477 }
 478
 479 enum {
 480         /* archive_ids in array format */
 481         KKUC_CT_DATA_ARRAY_MAGIC        = 0x092013cea,
 482         /* archive_ids in bitmap format */
 483         KKUC_CT_DATA_BITMAP_MAGIC       = 0x082018cea,
 484 };
 485
 486
 487 struct kkuc_ct_data {
 488         __u32           kcd_magic;
 489         __u32           kcd_nr_archives;
 490         __u32           kcd_archives[0];
 491 };
 492
 493 /** @} export */
 494
 495 #endif /* __EXPORT_H */
 496 /** @} obd_export */