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, 2016, 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 <lprocfs_status.h>
  46 #include <lustre/lustre_idl.h>
  47 #include <lustre_dlm.h>
  48
  49 struct mds_client_data;
  50 struct mdt_client_data;
  51 struct mds_idmap_table;
  52 struct mdt_idmap_table;
  53
  54 /**
  55  * Target-specific export data
  56  */
  57 struct tg_export_data {
  58         /** Protects ted_lcd, ted_reply_* and
  59          * ted_release_* fields below */
  60         struct mutex            ted_lcd_lock;
  61         /** Per-client data for each export */
  62         struct lsd_client_data  *ted_lcd;
  63         /** Offset of record in last_rcvd file */
  64         loff_t                  ted_lr_off;
  65         /** Client index in last_rcvd file */
  66         int                     ted_lr_idx;
  67
  68         /**
  69          * ted_nodemap_lock is used to ensure that the nodemap is not destroyed
  70          * between the time that ted_nodemap is checked for NULL, and a
  71          * reference is taken. Modifications to ted_nodemap require that the
  72          * active_config_lock and the nodemap(s)'s nm_member_list_lock be
  73          * taken, as well as ted_nodemap_lock, so the export can be properly
  74          * added to or removed from the nodemap's member list. When an export
  75          * is added to a nodemap, a reference on that nodemap must be taken.
  76          * That reference can be put only after ted_nodemap no longer refers to
  77          * it.
  78          */
  79         spinlock_t              ted_nodemap_lock;
  80         struct lu_nodemap       *ted_nodemap;
  81         struct list_head        ted_nodemap_member;
  82
  83         /** last version of nodemap config sent to client */
  84         __u64                   ted_nodemap_version;
  85
  86         /* Every reply data fields below are
  87          * protected by ted_lcd_lock */
  88         /** List of reply data */
  89         struct list_head        ted_reply_list;
  90         int                     ted_reply_cnt;
  91         /** Reply data with highest transno is retained */
  92         struct tg_reply_data    *ted_reply_last;
  93         /* Statistics */
  94         int                     ted_reply_max; /* high water mark */
  95         int                     ted_release_xid;
  96         int                     ted_release_tag;
  97 };
  98
  99 /**
 100  * MDT-specific export data
 101  */
 102 struct mdt_export_data {
 103         struct tg_export_data   med_ted;
 104         /** List of all files opened by client on this MDT */
 105         struct list_head        med_open_head;
 106         spinlock_t              med_open_lock; /* med_open_head, mfd_list */
 107 };
 108
 109 struct ec_export_data { /* echo client */
 110         struct list_head        eced_locks;
 111 };
 112
 113 /* In-memory access to client data from OST struct */
 114 /** Filter (oss-side) specific import data */
 115 struct filter_export_data {
 116         struct tg_export_data   fed_ted;
 117         spinlock_t              fed_lock;       /**< protects fed_mod_list */
 118         __u64                   fed_lastid_gen;
 119         long                    fed_dirty;    /* in bytes */
 120         long                    fed_grant;    /* in bytes */
 121         struct list_head        fed_mod_list; /* files being modified */
 122         long                    fed_pending;  /* bytes just being written */
 123         /* count of SOFT_SYNC RPCs, which will be reset after
 124          * ofd_soft_sync_limit number of RPCs, and trigger a sync. */
 125         atomic_t                fed_soft_sync_count;
 126         int                     fed_mod_count;/* items in fed_writing list */
 127         __u32                   fed_group;
 128         __u8                    fed_pagebits; /* log2 of client page size */
 129 };
 130
 131 struct mgs_export_data {
 132         struct list_head        med_clients;    /* mgc fs client via this exp */
 133         spinlock_t              med_lock;       /* protect med_clients */
 134 };
 135
 136 /**
 137  * per-NID statistics structure.
 138  * It tracks access patterns to this export on a per-client-NID basis
 139  */
 140 struct nid_stat {
 141         lnet_nid_t               nid;
 142         struct hlist_node        nid_hash;
 143         struct list_head         nid_list;
 144         struct obd_device       *nid_obd;
 145         struct proc_dir_entry   *nid_proc;
 146         struct lprocfs_stats    *nid_stats;
 147         struct lprocfs_stats    *nid_ldlm_stats;
 148         atomic_t                 nid_exp_ref_count; /* for obd_nid_stats_hash
 149                                                        exp_nid_stats */
 150 };
 151
 152 #define nidstat_getref(nidstat)                                                \
 153 do {                                                                           \
 154         atomic_inc(&(nidstat)->nid_exp_ref_count);                         \
 155 } while(0)
 156
 157 #define nidstat_putref(nidstat)                                                \
 158 do {                                                                           \
 159         atomic_dec(&(nidstat)->nid_exp_ref_count);                         \
 160         LASSERTF(atomic_read(&(nidstat)->nid_exp_ref_count) >= 0,          \
 161                  "stat %p nid_exp_ref_count < 0\n", nidstat);                  \
 162 } while(0)
 163
 164 enum obd_option {
 165         OBD_OPT_FORCE =         0x0001,
 166         OBD_OPT_FAILOVER =      0x0002,
 167         OBD_OPT_ABORT_RECOV =   0x0004,
 168 };
 169
 170 /**
 171  * Export structure. Represents target-side of connection in portals.
 172  * Also used in Lustre to connect between layers on the same node when
 173  * there is no network-connection in-between.
 174  * For every connected client there is an export structure on the server
 175  * attached to the same obd device.
 176  */
 177 struct obd_export {
 178         /**
 179          * Export handle, it's id is provided to client on connect
 180          * Subsequent client RPCs contain this handle id to identify
 181          * what export they are talking to.
 182          */
 183         struct portals_handle   exp_handle;
 184         atomic_t                exp_refcount;
 185         /**
 186          * Set of counters below is to track where export references are
 187          * kept. The exp_rpc_count is used for reconnect handling also,
 188          * the cb_count and locks_count are for debug purposes only for now.
 189          * The sum of them should be less than exp_refcount by 3
 190          */
 191         atomic_t                exp_rpc_count; /* RPC references */
 192         atomic_t                exp_cb_count; /* Commit callback references */
 193         /** Number of queued replay requests to be processes */
 194         atomic_t                exp_replay_count;
 195         atomic_t                exp_locks_count; /** Lock references */
 196 #if LUSTRE_TRACKS_LOCK_EXP_REFS
 197         struct list_head        exp_locks_list;
 198         spinlock_t              exp_locks_list_guard;
 199 #endif
 200         /** UUID of client connected to this export */
 201         struct obd_uuid         exp_client_uuid;
 202         /** To link all exports on an obd device */
 203         struct list_head        exp_obd_chain;
 204         /* Unlinked export list */
 205         struct list_head        exp_stale_list;
 206         struct hlist_node       exp_uuid_hash;  /** uuid-export hash*/
 207         struct hlist_node       exp_nid_hash;   /** nid-export hash */
 208         struct hlist_node       exp_gen_hash;   /** last_rcvd clt gen hash */
 209         /**
 210          * All exports eligible for ping evictor are linked into a list
 211          * through this field in "most time since last request on this export"
 212          * order
 213          * protected by obd_dev_lock
 214          */
 215         struct list_head        exp_obd_chain_timed;
 216         /** Obd device of this export */
 217         struct obd_device      *exp_obd;
 218         /**
 219          * "reverse" import to send requests (e.g. from ldlm) back to client
 220          * exp_lock protect its change
 221          */
 222         struct obd_import        *exp_imp_reverse;
 223         struct nid_stat          *exp_nid_stats;
 224         /** Active connetion */
 225         struct ptlrpc_connection *exp_connection;
 226         /** Connection count value from last successful reconnect rpc */
 227         __u32                     exp_conn_cnt;
 228         /** Hash list of all ldlm locks granted on this export */
 229         struct cfs_hash          *exp_lock_hash;
 230         /**
 231          * Hash list for Posix lock deadlock detection, added with
 232          * ldlm_lock::l_exp_flock_hash.
 233          */
 234         struct cfs_hash        *exp_flock_hash;
 235         struct list_head        exp_outstanding_replies;
 236         struct list_head        exp_uncommitted_replies;
 237         spinlock_t              exp_uncommitted_replies_lock;
 238         /** Last committed transno for this export */
 239         __u64                   exp_last_committed;
 240         /** When was last request received */
 241         cfs_time_t              exp_last_request_time;
 242         /** On replay all requests waiting for replay are linked here */
 243         struct list_head        exp_req_replay_queue;
 244         /**
 245          * protects exp_flags, exp_outstanding_replies and the change
 246          * of exp_imp_reverse
 247          */
 248         spinlock_t                exp_lock;
 249         /** Compatibility flags for this export are embedded into
 250          *  exp_connect_data */
 251         struct obd_connect_data   exp_connect_data;
 252         enum obd_option           exp_flags;
 253         unsigned long             exp_failed:1,
 254                                   exp_in_recovery:1,
 255                                   exp_disconnected:1,
 256                                   exp_connecting:1,
 257                                   /** VBR: export missed recovery */
 258                                   exp_delayed:1,
 259                                   /** VBR: failed version checking */
 260                                   exp_vbr_failed:1,
 261                                   exp_req_replay_needed:1,
 262                                   exp_lock_replay_needed:1,
 263                                   exp_need_sync:1,
 264                                   exp_flvr_changed:1,
 265                                   exp_flvr_adapt:1,
 266                                   exp_libclient:1, /* liblustre client? */
 267                                   /* if to swap nidtbl entries for 2.2 clients.
 268                                    * Only used by the MGS to fix LU-1644. */
 269                                   exp_need_mne_swab:1,
 270                                   /* The export already got final replay ping
 271                                    * request. */
 272                                   exp_replay_done:1;
 273         /* also protected by exp_lock */
 274         enum lustre_sec_part      exp_sp_peer;
 275         struct sptlrpc_flavor     exp_flvr;             /* current */
 276         struct sptlrpc_flavor     exp_flvr_old[2];      /* about-to-expire */
 277         time64_t                  exp_flvr_expire[2];   /* seconds */
 278
 279         /** protects exp_hp_rpcs */
 280         spinlock_t              exp_rpc_lock;
 281         struct list_head        exp_hp_rpcs;    /* (potential) HP RPCs */
 282         struct list_head        exp_reg_rpcs;  /* RPC being handled */
 283
 284         /** blocking dlm lock list, protected by exp_bl_list_lock */
 285         struct list_head        exp_bl_list;
 286         spinlock_t              exp_bl_list_lock;
 287
 288         /** Target specific data */
 289         union {
 290                 struct tg_export_data     eu_target_data;
 291                 struct mdt_export_data    eu_mdt_data;
 292                 struct filter_export_data eu_filter_data;
 293                 struct ec_export_data     eu_ec_data;
 294                 struct mgs_export_data    eu_mgs_data;
 295         } u;
 296
 297         struct adaptive_timeout    exp_bl_lock_at;
 298
 299         /** highest XID received by export client that has no
 300          * unreceived lower-numbered XID
 301          */
 302         __u64                     exp_last_xid;
 303 };
 304
 305 #define exp_target_data u.eu_target_data
 306 #define exp_mdt_data    u.eu_mdt_data
 307 #define exp_filter_data u.eu_filter_data
 308 #define exp_ec_data     u.eu_ec_data
 309
 310 static inline __u64 *exp_connect_flags_ptr(struct obd_export *exp)
 311 {
 312         return &exp->exp_connect_data.ocd_connect_flags;
 313 }
 314
 315 static inline __u64 exp_connect_flags(struct obd_export *exp)
 316 {
 317         return *exp_connect_flags_ptr(exp);
 318 }
 319
 320 static inline int exp_max_brw_size(struct obd_export *exp)
 321 {
 322         LASSERT(exp != NULL);
 323         if (exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE)
 324                 return exp->exp_connect_data.ocd_brw_size;
 325
 326         return ONE_MB_BRW_SIZE;
 327 }
 328
 329 static inline int exp_connect_multibulk(struct obd_export *exp)
 330 {
 331         return exp_max_brw_size(exp) > ONE_MB_BRW_SIZE;
 332 }
 333
 334 static inline int exp_expired(struct obd_export *exp, cfs_duration_t age)
 335 {
 336         LASSERT(exp->exp_delayed);
 337         return cfs_time_before(cfs_time_add(exp->exp_last_request_time, age),
 338                                cfs_time_current_sec());
 339 }
 340
 341 static inline int exp_connect_cancelset(struct obd_export *exp)
 342 {
 343         LASSERT(exp != NULL);
 344         return !!(exp_connect_flags(exp) & OBD_CONNECT_CANCELSET);
 345 }
 346
 347 static inline int exp_connect_lru_resize(struct obd_export *exp)
 348 {
 349         LASSERT(exp != NULL);
 350         return !!(exp_connect_flags(exp) & OBD_CONNECT_LRU_RESIZE);
 351 }
 352
 353 static inline int exp_connect_vbr(struct obd_export *exp)
 354 {
 355         LASSERT(exp != NULL);
 356         LASSERT(exp->exp_connection);
 357         return !!(exp_connect_flags(exp) & OBD_CONNECT_VBR);
 358 }
 359
 360 static inline int exp_connect_umask(struct obd_export *exp)
 361 {
 362         return !!(exp_connect_flags(exp) & OBD_CONNECT_UMASK);
 363 }
 364
 365 static inline int imp_connect_lru_resize(struct obd_import *imp)
 366 {
 367         struct obd_connect_data *ocd;
 368
 369         LASSERT(imp != NULL);
 370         ocd = &imp->imp_connect_data;
 371         return !!(ocd->ocd_connect_flags & OBD_CONNECT_LRU_RESIZE);
 372 }
 373
 374 static inline int exp_connect_layout(struct obd_export *exp)
 375 {
 376         return !!(exp_connect_flags(exp) & OBD_CONNECT_LAYOUTLOCK);
 377 }
 378
 379 static inline bool exp_connect_lvb_type(struct obd_export *exp)
 380 {
 381         LASSERT(exp != NULL);
 382         if (exp_connect_flags(exp) & OBD_CONNECT_LVB_TYPE)
 383                 return true;
 384         else
 385                 return false;
 386 }
 387
 388 static inline bool imp_connect_lvb_type(struct obd_import *imp)
 389 {
 390         struct obd_connect_data *ocd;
 391
 392         LASSERT(imp != NULL);
 393         ocd = &imp->imp_connect_data;
 394         if (ocd->ocd_connect_flags & OBD_CONNECT_LVB_TYPE)
 395                 return true;
 396         else
 397                 return false;
 398 }
 399
 400 static inline bool imp_connect_disp_stripe(struct obd_import *imp)
 401 {
 402         struct obd_connect_data *ocd;
 403
 404         LASSERT(imp != NULL);
 405         ocd = &imp->imp_connect_data;
 406         return ocd->ocd_connect_flags & OBD_CONNECT_DISP_STRIPE;
 407 }
 408
 409 static inline __u64 exp_connect_ibits(struct obd_export *exp)
 410 {
 411         struct obd_connect_data *ocd;
 412
 413         ocd = &exp->exp_connect_data;
 414         return ocd->ocd_ibits_known;
 415 }
 416
 417 extern struct obd_export *class_conn2export(struct lustre_handle *conn);
 418 extern struct obd_device *class_conn2obd(struct lustre_handle *conn);
 419
 420 #define KKUC_CT_DATA_MAGIC      0x092013cea
 421 struct kkuc_ct_data {
 422         __u32           kcd_magic;
 423         struct obd_uuid kcd_uuid;
 424         __u32           kcd_archive;
 425 };
 426
 427 /** @} export */
 428
 429 #endif /* __EXPORT_H */
 430 /** @} obd_export */