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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36 /** \defgroup obd_export PortalRPC export definitions
  37  *
  38  * @{
  39  */
  40
  41 #ifndef __EXPORT_H
  42 #define __EXPORT_H
  43
  44 /** \defgroup export export
  45  *
  46  * @{
  47  */
  48
  49 #include <lprocfs_status.h>
  50 #include <lustre/lustre_idl.h>
  51 #include <lustre_dlm.h>
  52
  53 struct mds_client_data;
  54 struct mdt_client_data;
  55 struct mds_idmap_table;
  56 struct mdt_idmap_table;
  57
  58 /**
  59  * Target-specific export data
  60  */
  61 struct tg_export_data {
  62         /** Protects ted_lcd, ted_reply_* and
  63          * ted_release_* fields below */
  64         struct mutex            ted_lcd_lock;
  65         /** Per-client data for each export */
  66         struct lsd_client_data  *ted_lcd;
  67         /** Offset of record in last_rcvd file */
  68         loff_t                  ted_lr_off;
  69         /** Client index in last_rcvd file */
  70         int                     ted_lr_idx;
  71
  72         /**
  73          * ted_nodemap_lock is used to ensure that the nodemap is not destroyed
  74          * between the time that ted_nodemap is checked for NULL, and a
  75          * reference is taken. Modifications to ted_nodemap require that the
  76          * active_config_lock and the nodemap(s)'s nm_member_list_lock be
  77          * taken, as well as ted_nodemap_lock, so the export can be properly
  78          * added to or removed from the nodemap's member list. When an export
  79          * is added to a nodemap, a reference on that nodemap must be taken.
  80          * That reference can be put only after ted_nodemap no longer refers to
  81          * it.
  82          */
  83         spinlock_t              ted_nodemap_lock;
  84         struct lu_nodemap       *ted_nodemap;
  85         struct list_head        ted_nodemap_member;
  86
  87         /** last version of nodemap config sent to client */
  88         __u64                   ted_nodemap_version;
  89
  90         /* Every reply data fields below are
  91          * protected by ted_lcd_lock */
  92         /** List of reply data */
  93         struct list_head        ted_reply_list;
  94         int                     ted_reply_cnt;
  95         /** Reply data with highest transno is retained */
  96         struct tg_reply_data    *ted_reply_last;
  97         /* Statistics */
  98         int                     ted_reply_max; /* high water mark */
  99         int                     ted_release_xid;
 100         int                     ted_release_tag;
 101 };
 102
 103 /**
 104  * MDT-specific export data
 105  */
 106 struct mdt_export_data {
 107         struct tg_export_data   med_ted;
 108         /** List of all files opened by client on this MDT */
 109         struct list_head        med_open_head;
 110         spinlock_t              med_open_lock; /* med_open_head, mfd_list */
 111 };
 112
 113 struct ec_export_data { /* echo client */
 114         struct list_head        eced_locks;
 115 };
 116
 117 /* In-memory access to client data from OST struct */
 118 /** Filter (oss-side) specific import data */
 119 struct filter_export_data {
 120         struct tg_export_data   fed_ted;
 121         spinlock_t              fed_lock;       /**< protects fed_mod_list */
 122         __u64                   fed_lastid_gen;
 123         long                    fed_dirty;    /* in bytes */
 124         long                    fed_grant;    /* in bytes */
 125         struct list_head        fed_mod_list; /* files being modified */
 126         long                    fed_pending;  /* bytes just being written */
 127         /* count of SOFT_SYNC RPCs, which will be reset after
 128          * ofd_soft_sync_limit number of RPCs, and trigger a sync. */
 129         atomic_t                fed_soft_sync_count;
 130         int                     fed_mod_count;/* items in fed_writing list */
 131         __u32                   fed_group;
 132         __u8                    fed_pagebits; /* log2 of client page size */
 133 };
 134
 135 struct mgs_export_data {
 136         struct list_head        med_clients;    /* mgc fs client via this exp */
 137         spinlock_t              med_lock;       /* protect med_clients */
 138 };
 139
 140 /**
 141  * per-NID statistics structure.
 142  * It tracks access patterns to this export on a per-client-NID basis
 143  */
 144 struct nid_stat {
 145         lnet_nid_t               nid;
 146         struct hlist_node        nid_hash;
 147         struct list_head         nid_list;
 148         struct obd_device       *nid_obd;
 149         struct proc_dir_entry   *nid_proc;
 150         struct lprocfs_stats    *nid_stats;
 151         struct lprocfs_stats    *nid_ldlm_stats;
 152         atomic_t                 nid_exp_ref_count; /* for obd_nid_stats_hash
 153                                                        exp_nid_stats */
 154 };
 155
 156 #define nidstat_getref(nidstat)                                                \
 157 do {                                                                           \
 158         atomic_inc(&(nidstat)->nid_exp_ref_count);                         \
 159 } while(0)
 160
 161 #define nidstat_putref(nidstat)                                                \
 162 do {                                                                           \
 163         atomic_dec(&(nidstat)->nid_exp_ref_count);                         \
 164         LASSERTF(atomic_read(&(nidstat)->nid_exp_ref_count) >= 0,          \
 165                  "stat %p nid_exp_ref_count < 0\n", nidstat);                  \
 166 } while(0)
 167
 168 enum obd_option {
 169         OBD_OPT_FORCE =         0x0001,
 170         OBD_OPT_FAILOVER =      0x0002,
 171         OBD_OPT_ABORT_RECOV =   0x0004,
 172 };
 173
 174 /**
 175  * Export structure. Represents target-side of connection in portals.
 176  * Also used in Lustre to connect between layers on the same node when
 177  * there is no network-connection in-between.
 178  * For every connected client there is an export structure on the server
 179  * attached to the same obd device.
 180  */
 181 struct obd_export {
 182         /**
 183          * Export handle, it's id is provided to client on connect
 184          * Subsequent client RPCs contain this handle id to identify
 185          * what export they are talking to.
 186          */
 187         struct portals_handle   exp_handle;
 188         atomic_t                exp_refcount;
 189         /**
 190          * Set of counters below is to track where export references are
 191          * kept. The exp_rpc_count is used for reconnect handling also,
 192          * the cb_count and locks_count are for debug purposes only for now.
 193          * The sum of them should be less than exp_refcount by 3
 194          */
 195         atomic_t                exp_rpc_count; /* RPC references */
 196         atomic_t                exp_cb_count; /* Commit callback references */
 197         /** Number of queued replay requests to be processes */
 198         atomic_t                exp_replay_count;
 199         atomic_t                exp_locks_count; /** Lock references */
 200 #if LUSTRE_TRACKS_LOCK_EXP_REFS
 201         struct list_head        exp_locks_list;
 202         spinlock_t              exp_locks_list_guard;
 203 #endif
 204         /** UUID of client connected to this export */
 205         struct obd_uuid         exp_client_uuid;
 206         /** To link all exports on an obd device */
 207         struct list_head        exp_obd_chain;
 208         /* Unlinked export list */
 209         struct list_head        exp_stale_list;
 210         struct hlist_node       exp_uuid_hash;  /** uuid-export hash*/
 211         struct hlist_node       exp_nid_hash;   /** nid-export hash */
 212         struct hlist_node       exp_gen_hash;   /** last_rcvd clt gen hash */
 213         /**
 214          * All exports eligible for ping evictor are linked into a list
 215          * through this field in "most time since last request on this export"
 216          * order
 217          * protected by obd_dev_lock
 218          */
 219         struct list_head        exp_obd_chain_timed;
 220         /** Obd device of this export */
 221         struct obd_device      *exp_obd;
 222         /**
 223          * "reverse" import to send requests (e.g. from ldlm) back to client
 224          * exp_lock protect its change
 225          */
 226         struct obd_import        *exp_imp_reverse;
 227         struct nid_stat          *exp_nid_stats;
 228         /** Active connetion */
 229         struct ptlrpc_connection *exp_connection;
 230         /** Connection count value from last successful reconnect rpc */
 231         __u32                     exp_conn_cnt;
 232         /** Hash list of all ldlm locks granted on this export */
 233         struct cfs_hash          *exp_lock_hash;
 234         /**
 235          * Hash list for Posix lock deadlock detection, added with
 236          * ldlm_lock::l_exp_flock_hash.
 237          */
 238         struct cfs_hash        *exp_flock_hash;
 239         struct list_head        exp_outstanding_replies;
 240         struct list_head        exp_uncommitted_replies;
 241         spinlock_t              exp_uncommitted_replies_lock;
 242         /** Last committed transno for this export */
 243         __u64                   exp_last_committed;
 244         /** When was last request received */
 245         cfs_time_t              exp_last_request_time;
 246         /** On replay all requests waiting for replay are linked here */
 247         struct list_head        exp_req_replay_queue;
 248         /**
 249          * protects exp_flags, exp_outstanding_replies and the change
 250          * of exp_imp_reverse
 251          */
 252         spinlock_t                exp_lock;
 253         /** Compatibility flags for this export are embedded into
 254          *  exp_connect_data */
 255         struct obd_connect_data   exp_connect_data;
 256         enum obd_option           exp_flags;
 257         unsigned long             exp_failed:1,
 258                                   exp_in_recovery:1,
 259                                   exp_disconnected:1,
 260                                   exp_connecting:1,
 261                                   /** VBR: export missed recovery */
 262                                   exp_delayed:1,
 263                                   /** VBR: failed version checking */
 264                                   exp_vbr_failed:1,
 265                                   exp_req_replay_needed:1,
 266                                   exp_lock_replay_needed:1,
 267                                   exp_need_sync:1,
 268                                   exp_flvr_changed:1,
 269                                   exp_flvr_adapt:1,
 270                                   exp_libclient:1, /* liblustre client? */
 271                                   /* if to swap nidtbl entries for 2.2 clients.
 272                                    * Only used by the MGS to fix LU-1644. */
 273                                   exp_need_mne_swab:1,
 274                                   /* The export already got final replay ping
 275                                    * request. */
 276                                   exp_replay_done:1;
 277         /* also protected by exp_lock */
 278         enum lustre_sec_part      exp_sp_peer;
 279         struct sptlrpc_flavor     exp_flvr;             /* current */
 280         struct sptlrpc_flavor     exp_flvr_old[2];      /* about-to-expire */
 281         cfs_time_t                exp_flvr_expire[2];   /* seconds */
 282
 283         /** protects exp_hp_rpcs */
 284         spinlock_t              exp_rpc_lock;
 285         struct list_head        exp_hp_rpcs;    /* (potential) HP RPCs */
 286         struct list_head        exp_reg_rpcs;  /* RPC being handled */
 287
 288         /** blocking dlm lock list, protected by exp_bl_list_lock */
 289         struct list_head        exp_bl_list;
 290         spinlock_t              exp_bl_list_lock;
 291
 292         /** Target specific data */
 293         union {
 294                 struct tg_export_data     eu_target_data;
 295                 struct mdt_export_data    eu_mdt_data;
 296                 struct filter_export_data eu_filter_data;
 297                 struct ec_export_data     eu_ec_data;
 298                 struct mgs_export_data    eu_mgs_data;
 299         } u;
 300
 301         struct adaptive_timeout    exp_bl_lock_at;
 302
 303         /** highest XID received by export client that has no
 304          * unreceived lower-numbered XID
 305          */
 306         __u64                     exp_last_xid;
 307 };
 308
 309 #define exp_target_data u.eu_target_data
 310 #define exp_mdt_data    u.eu_mdt_data
 311 #define exp_filter_data u.eu_filter_data
 312 #define exp_ec_data     u.eu_ec_data
 313
 314 static inline __u64 *exp_connect_flags_ptr(struct obd_export *exp)
 315 {
 316         return &exp->exp_connect_data.ocd_connect_flags;
 317 }
 318
 319 static inline __u64 exp_connect_flags(struct obd_export *exp)
 320 {
 321         return *exp_connect_flags_ptr(exp);
 322 }
 323
 324 static inline int exp_max_brw_size(struct obd_export *exp)
 325 {
 326         LASSERT(exp != NULL);
 327         if (exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE)
 328                 return exp->exp_connect_data.ocd_brw_size;
 329
 330         return ONE_MB_BRW_SIZE;
 331 }
 332
 333 static inline int exp_connect_multibulk(struct obd_export *exp)
 334 {
 335         return exp_max_brw_size(exp) > ONE_MB_BRW_SIZE;
 336 }
 337
 338 static inline int exp_expired(struct obd_export *exp, cfs_duration_t age)
 339 {
 340         LASSERT(exp->exp_delayed);
 341         return cfs_time_before(cfs_time_add(exp->exp_last_request_time, age),
 342                                cfs_time_current_sec());
 343 }
 344
 345 static inline int exp_connect_cancelset(struct obd_export *exp)
 346 {
 347         LASSERT(exp != NULL);
 348         return !!(exp_connect_flags(exp) & OBD_CONNECT_CANCELSET);
 349 }
 350
 351 static inline int exp_connect_lru_resize(struct obd_export *exp)
 352 {
 353         LASSERT(exp != NULL);
 354         return !!(exp_connect_flags(exp) & OBD_CONNECT_LRU_RESIZE);
 355 }
 356
 357 static inline int exp_connect_vbr(struct obd_export *exp)
 358 {
 359         LASSERT(exp != NULL);
 360         LASSERT(exp->exp_connection);
 361         return !!(exp_connect_flags(exp) & OBD_CONNECT_VBR);
 362 }
 363
 364 static inline int exp_connect_umask(struct obd_export *exp)
 365 {
 366         return !!(exp_connect_flags(exp) & OBD_CONNECT_UMASK);
 367 }
 368
 369 static inline int imp_connect_lru_resize(struct obd_import *imp)
 370 {
 371         struct obd_connect_data *ocd;
 372
 373         LASSERT(imp != NULL);
 374         ocd = &imp->imp_connect_data;
 375         return !!(ocd->ocd_connect_flags & OBD_CONNECT_LRU_RESIZE);
 376 }
 377
 378 static inline int exp_connect_layout(struct obd_export *exp)
 379 {
 380         return !!(exp_connect_flags(exp) & OBD_CONNECT_LAYOUTLOCK);
 381 }
 382
 383 static inline bool exp_connect_lvb_type(struct obd_export *exp)
 384 {
 385         LASSERT(exp != NULL);
 386         if (exp_connect_flags(exp) & OBD_CONNECT_LVB_TYPE)
 387                 return true;
 388         else
 389                 return false;
 390 }
 391
 392 static inline bool imp_connect_lvb_type(struct obd_import *imp)
 393 {
 394         struct obd_connect_data *ocd;
 395
 396         LASSERT(imp != NULL);
 397         ocd = &imp->imp_connect_data;
 398         if (ocd->ocd_connect_flags & OBD_CONNECT_LVB_TYPE)
 399                 return true;
 400         else
 401                 return false;
 402 }
 403
 404 static inline bool imp_connect_disp_stripe(struct obd_import *imp)
 405 {
 406         struct obd_connect_data *ocd;
 407
 408         LASSERT(imp != NULL);
 409         ocd = &imp->imp_connect_data;
 410         return ocd->ocd_connect_flags & OBD_CONNECT_DISP_STRIPE;
 411 }
 412
 413 static inline __u64 exp_connect_ibits(struct obd_export *exp)
 414 {
 415         struct obd_connect_data *ocd;
 416
 417         ocd = &exp->exp_connect_data;
 418         return ocd->ocd_ibits_known;
 419 }
 420
 421 extern struct obd_export *class_conn2export(struct lustre_handle *conn);
 422 extern struct obd_device *class_conn2obd(struct lustre_handle *conn);
 423
 424 #define KKUC_CT_DATA_MAGIC      0x092013cea
 425 struct kkuc_ct_data {
 426         __u32           kcd_magic;
 427         struct obd_uuid kcd_uuid;
 428         __u32           kcd_archive;
 429 };
 430
 431 /** @} export */
 432
 433 #endif /* __EXPORT_H */
 434 /** @} obd_export */