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         /* grants */
  98         long                    ted_dirty;    /* in bytes */
  99         long                    ted_grant;    /* in bytes */
 100         long                    ted_pending;  /* bytes just being written */
 101         __u8                    ted_pagebits; /* log2 of client page size */
 102 };
 103
 104 /**
 105  * MDT-specific export data
 106  */
 107 struct mdt_export_data {
 108         struct tg_export_data   med_ted;
 109         /** List of all files opened by client on this MDT */
 110         struct list_head        med_open_head;
 111         spinlock_t              med_open_lock; /* med_open_head, mfd_list */
 112 };
 113
 114 struct ec_export_data { /* echo client */
 115         struct list_head        eced_locks;
 116 };
 117
 118 /* In-memory access to client data from OST struct */
 119 /** Filter (oss-side) specific import data */
 120 struct filter_export_data {
 121         struct tg_export_data   fed_ted;
 122         spinlock_t              fed_lock;       /**< protects fed_mod_list */
 123         __u64                   fed_lastid_gen;
 124         struct list_head        fed_mod_list; /* files being modified */
 125         /* count of SOFT_SYNC RPCs, which will be reset after
 126          * ofd_soft_sync_limit number of RPCs, and trigger a sync. */
 127         atomic_t                fed_soft_sync_count;
 128         int                     fed_mod_count;/* items in fed_writing list */
 129         __u32                   fed_group;
 130 };
 131
 132 struct mgs_export_data {
 133         struct list_head        med_clients;    /* mgc fs client via this exp */
 134         spinlock_t              med_lock;       /* protect med_clients */
 135 };
 136
 137 /**
 138  * per-NID statistics structure.
 139  * It tracks access patterns to this export on a per-client-NID basis
 140  */
 141 struct nid_stat {
 142         lnet_nid_t               nid;
 143         struct hlist_node        nid_hash;
 144         struct list_head         nid_list;
 145         struct obd_device       *nid_obd;
 146         struct proc_dir_entry   *nid_proc;
 147         struct lprocfs_stats    *nid_stats;
 148         struct lprocfs_stats    *nid_ldlm_stats;
 149         atomic_t                 nid_exp_ref_count; /* for obd_nid_stats_hash
 150                                                        exp_nid_stats */
 151 };
 152
 153 #define nidstat_getref(nidstat)                                                \
 154 do {                                                                           \
 155         atomic_inc(&(nidstat)->nid_exp_ref_count);                         \
 156 } while(0)
 157
 158 #define nidstat_putref(nidstat)                                                \
 159 do {                                                                           \
 160         atomic_dec(&(nidstat)->nid_exp_ref_count);                         \
 161         LASSERTF(atomic_read(&(nidstat)->nid_exp_ref_count) >= 0,          \
 162                  "stat %p nid_exp_ref_count < 0\n", nidstat);                  \
 163 } while(0)
 164
 165 enum obd_option {
 166         OBD_OPT_FORCE =         0x0001,
 167         OBD_OPT_FAILOVER =      0x0002,
 168         OBD_OPT_ABORT_RECOV =   0x0004,
 169 };
 170
 171 /**
 172  * Export structure. Represents target-side of connection in portals.
 173  * Also used in Lustre to connect between layers on the same node when
 174  * there is no network-connection in-between.
 175  * For every connected client there is an export structure on the server
 176  * attached to the same obd device.
 177  */
 178 struct obd_export {
 179         /**
 180          * Export handle, it's id is provided to client on connect
 181          * Subsequent client RPCs contain this handle id to identify
 182          * what export they are talking to.
 183          */
 184         struct portals_handle   exp_handle;
 185         atomic_t                exp_refcount;
 186         /**
 187          * Set of counters below is to track where export references are
 188          * kept. The exp_rpc_count is used for reconnect handling also,
 189          * the cb_count and locks_count are for debug purposes only for now.
 190          * The sum of them should be less than exp_refcount by 3
 191          */
 192         atomic_t                exp_rpc_count; /* RPC references */
 193         atomic_t                exp_cb_count; /* Commit callback references */
 194         /** Number of queued replay requests to be processes */
 195         atomic_t                exp_replay_count;
 196         atomic_t                exp_locks_count; /** Lock references */
 197 #if LUSTRE_TRACKS_LOCK_EXP_REFS
 198         struct list_head        exp_locks_list;
 199         spinlock_t              exp_locks_list_guard;
 200 #endif
 201         /** UUID of client connected to this export */
 202         struct obd_uuid         exp_client_uuid;
 203         /** To link all exports on an obd device */
 204         struct list_head        exp_obd_chain;
 205         /* Unlinked export list */
 206         struct list_head        exp_stale_list;
 207         struct hlist_node       exp_uuid_hash;  /** uuid-export hash*/
 208         struct hlist_node       exp_nid_hash;   /** nid-export hash */
 209         struct hlist_node       exp_gen_hash;   /** last_rcvd clt gen hash */
 210         /**
 211          * All exports eligible for ping evictor are linked into a list
 212          * through this field in "most time since last request on this export"
 213          * order
 214          * protected by obd_dev_lock
 215          */
 216         struct list_head        exp_obd_chain_timed;
 217         /** Obd device of this export */
 218         struct obd_device      *exp_obd;
 219         /**
 220          * "reverse" import to send requests (e.g. from ldlm) back to client
 221          * exp_lock protect its change
 222          */
 223         struct obd_import        *exp_imp_reverse;
 224         struct nid_stat          *exp_nid_stats;
 225         /** Active connetion */
 226         struct ptlrpc_connection *exp_connection;
 227         /** Connection count value from last successful reconnect rpc */
 228         __u32                     exp_conn_cnt;
 229         /** Hash list of all ldlm locks granted on this export */
 230         struct cfs_hash          *exp_lock_hash;
 231         /**
 232          * Hash list for Posix lock deadlock detection, added with
 233          * ldlm_lock::l_exp_flock_hash.
 234          */
 235         struct cfs_hash        *exp_flock_hash;
 236         struct list_head        exp_outstanding_replies;
 237         struct list_head        exp_uncommitted_replies;
 238         spinlock_t              exp_uncommitted_replies_lock;
 239         /** Last committed transno for this export */
 240         __u64                   exp_last_committed;
 241         /** When was last request received */
 242         cfs_time_t              exp_last_request_time;
 243         /** On replay all requests waiting for replay are linked here */
 244         struct list_head        exp_req_replay_queue;
 245         /**
 246          * protects exp_flags, exp_outstanding_replies and the change
 247          * of exp_imp_reverse
 248          */
 249         spinlock_t                exp_lock;
 250         /** Compatibility flags for this export are embedded into
 251          *  exp_connect_data */
 252         struct obd_connect_data   exp_connect_data;
 253         enum obd_option           exp_flags;
 254         unsigned long             exp_failed:1,
 255                                   exp_in_recovery:1,
 256                                   exp_disconnected:1,
 257                                   exp_connecting:1,
 258                                   /** VBR: export missed recovery */
 259                                   exp_delayed:1,
 260                                   /** VBR: failed version checking */
 261                                   exp_vbr_failed:1,
 262                                   exp_req_replay_needed:1,
 263                                   exp_lock_replay_needed:1,
 264                                   exp_need_sync:1,
 265                                   exp_flvr_changed:1,
 266                                   exp_flvr_adapt:1,
 267                                   exp_libclient:1, /* liblustre client? */
 268                                   /* if to swap nidtbl entries for 2.2 clients.
 269                                    * Only used by the MGS to fix LU-1644. */
 270                                   exp_need_mne_swab:1,
 271                                   /* The export already got final replay ping
 272                                    * request. */
 273                                   exp_replay_done:1;
 274         /* also protected by exp_lock */
 275         enum lustre_sec_part      exp_sp_peer;
 276         struct sptlrpc_flavor     exp_flvr;             /* current */
 277         struct sptlrpc_flavor     exp_flvr_old[2];      /* about-to-expire */
 278         time64_t                  exp_flvr_expire[2];   /* seconds */
 279
 280         /** protects exp_hp_rpcs */
 281         spinlock_t              exp_rpc_lock;
 282         struct list_head        exp_hp_rpcs;    /* (potential) HP RPCs */
 283         struct list_head        exp_reg_rpcs;  /* RPC being handled */
 284
 285         /** blocking dlm lock list, protected by exp_bl_list_lock */
 286         struct list_head        exp_bl_list;
 287         spinlock_t              exp_bl_list_lock;
 288
 289         /** Target specific data */
 290         union {
 291                 struct tg_export_data     eu_target_data;
 292                 struct mdt_export_data    eu_mdt_data;
 293                 struct filter_export_data eu_filter_data;
 294                 struct ec_export_data     eu_ec_data;
 295                 struct mgs_export_data    eu_mgs_data;
 296         } u;
 297
 298         struct adaptive_timeout    exp_bl_lock_at;
 299
 300         /** highest XID received by export client that has no
 301          * unreceived lower-numbered XID
 302          */
 303         __u64                     exp_last_xid;
 304 };
 305
 306 #define exp_target_data u.eu_target_data
 307 #define exp_mdt_data    u.eu_mdt_data
 308 #define exp_filter_data u.eu_filter_data
 309 #define exp_ec_data     u.eu_ec_data
 310
 311 static inline __u64 *exp_connect_flags_ptr(struct obd_export *exp)
 312 {
 313         return &exp->exp_connect_data.ocd_connect_flags;
 314 }
 315
 316 static inline __u64 exp_connect_flags(struct obd_export *exp)
 317 {
 318         return *exp_connect_flags_ptr(exp);
 319 }
 320
 321 static inline int exp_max_brw_size(struct obd_export *exp)
 322 {
 323         LASSERT(exp != NULL);
 324         if (exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE)
 325                 return exp->exp_connect_data.ocd_brw_size;
 326
 327         return ONE_MB_BRW_SIZE;
 328 }
 329
 330 static inline int exp_connect_multibulk(struct obd_export *exp)
 331 {
 332         return exp_max_brw_size(exp) > ONE_MB_BRW_SIZE;
 333 }
 334
 335 static inline int exp_expired(struct obd_export *exp, cfs_duration_t age)
 336 {
 337         LASSERT(exp->exp_delayed);
 338         return cfs_time_before(cfs_time_add(exp->exp_last_request_time, age),
 339                                cfs_time_current_sec());
 340 }
 341
 342 static inline int exp_connect_cancelset(struct obd_export *exp)
 343 {
 344         LASSERT(exp != NULL);
 345         return !!(exp_connect_flags(exp) & OBD_CONNECT_CANCELSET);
 346 }
 347
 348 static inline int exp_connect_lru_resize(struct obd_export *exp)
 349 {
 350         LASSERT(exp != NULL);
 351         return !!(exp_connect_flags(exp) & OBD_CONNECT_LRU_RESIZE);
 352 }
 353
 354 static inline int exp_connect_vbr(struct obd_export *exp)
 355 {
 356         LASSERT(exp != NULL);
 357         LASSERT(exp->exp_connection);
 358         return !!(exp_connect_flags(exp) & OBD_CONNECT_VBR);
 359 }
 360
 361 static inline int exp_connect_umask(struct obd_export *exp)
 362 {
 363         return !!(exp_connect_flags(exp) & OBD_CONNECT_UMASK);
 364 }
 365
 366 static inline int imp_connect_lru_resize(struct obd_import *imp)
 367 {
 368         struct obd_connect_data *ocd;
 369
 370         LASSERT(imp != NULL);
 371         ocd = &imp->imp_connect_data;
 372         return !!(ocd->ocd_connect_flags & OBD_CONNECT_LRU_RESIZE);
 373 }
 374
 375 static inline int exp_connect_layout(struct obd_export *exp)
 376 {
 377         return !!(exp_connect_flags(exp) & OBD_CONNECT_LAYOUTLOCK);
 378 }
 379
 380 static inline bool exp_connect_lvb_type(struct obd_export *exp)
 381 {
 382         LASSERT(exp != NULL);
 383         if (exp_connect_flags(exp) & OBD_CONNECT_LVB_TYPE)
 384                 return true;
 385         else
 386                 return false;
 387 }
 388
 389 static inline bool imp_connect_lvb_type(struct obd_import *imp)
 390 {
 391         struct obd_connect_data *ocd;
 392
 393         LASSERT(imp != NULL);
 394         ocd = &imp->imp_connect_data;
 395         if (ocd->ocd_connect_flags & OBD_CONNECT_LVB_TYPE)
 396                 return true;
 397         else
 398                 return false;
 399 }
 400
 401 static inline bool imp_connect_disp_stripe(struct obd_import *imp)
 402 {
 403         struct obd_connect_data *ocd;
 404
 405         LASSERT(imp != NULL);
 406         ocd = &imp->imp_connect_data;
 407         return ocd->ocd_connect_flags & OBD_CONNECT_DISP_STRIPE;
 408 }
 409
 410 static inline __u64 exp_connect_ibits(struct obd_export *exp)
 411 {
 412         struct obd_connect_data *ocd;
 413
 414         ocd = &exp->exp_connect_data;
 415         return ocd->ocd_ibits_known;
 416 }
 417
 418 static inline int exp_connect_large_acl(struct obd_export *exp)
 419 {
 420         return !!(exp_connect_flags(exp) & OBD_CONNECT_LARGE_ACL);
 421 }
 422
 423 extern struct obd_export *class_conn2export(struct lustre_handle *conn);
 424 extern struct obd_device *class_conn2obd(struct lustre_handle *conn);
 425
 426 #define KKUC_CT_DATA_MAGIC      0x092013cea
 427 struct kkuc_ct_data {
 428         __u32           kcd_magic;
 429         struct obd_uuid kcd_uuid;
 430         __u32           kcd_archive;
 431 };
 432
 433 /** @} export */
 434
 435 #endif /* __EXPORT_H */
 436 /** @} obd_export */