lustre/include/lustre_import.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_import PtlRPC import definitions
  33  * Imports are client-side representation of remote obd target.
  34  *
  35  * @{
  36  */
  37
  38 #ifndef __IMPORT_H
  39 #define __IMPORT_H
  40
  41 /** \defgroup export export
  42  *
  43  * @{
  44  */
  45 #include <linux/atomic.h>
  46 #include <linux/list.h>
  47 #include <linux/mutex.h>
  48 #include <linux/refcount.h>
  49 #include <linux/spinlock.h>
  50 #include <linux/time.h>
  51 #include <linux/types.h>
  52 #include <linux/workqueue.h>
  53 #include <libcfs/libcfs.h>
  54 #include <uapi/linux/lustre/lustre_idl.h>
  55
  56 /**
  57  * Adaptive Timeout stuff
  58  *
  59  * @{
  60  */
  61 #define D_ADAPTTO D_OTHER
  62 #define AT_BINS 4                  /* "bin" means "N seconds of history" */
  63 #define AT_FLG_NOHIST 0x1          /* use last reported value only */
  64
  65 struct adaptive_timeout {
  66         time64_t        at_binstart;         /* bin start time */
  67         unsigned int    at_hist[AT_BINS];    /* timeout history bins */
  68         unsigned int    at_flags;
  69         unsigned int    at_current;          /* current timeout value */
  70         unsigned int    at_worst_ever;       /* worst-ever timeout value */
  71         time64_t        at_worst_time;       /* worst-ever timeout timestamp */
  72         spinlock_t      at_lock;
  73 };
  74
  75 enum lustre_at_flags {
  76         LATF_SKIP       = 0x0,
  77         LATF_STATS      = 0x1,
  78 };
  79
  80 struct ptlrpc_at_array {
  81         struct list_head *paa_reqs_array; /** array to hold requests */
  82         __u32             paa_size;       /** the size of array */
  83         __u32             paa_count;      /** the total count of reqs */
  84         time64_t          paa_deadline;   /** the earliest deadline of reqs */
  85         __u32            *paa_reqs_count; /** the count of reqs in each entry */
  86 };
  87
  88 #define IMP_AT_MAX_PORTALS 8
  89 struct imp_at {
  90         int                     iat_portal[IMP_AT_MAX_PORTALS];
  91         struct adaptive_timeout iat_net_latency;
  92         struct adaptive_timeout iat_service_estimate[IMP_AT_MAX_PORTALS];
  93 };
  94
  95
  96 /** @} */
  97
  98 /** Possible import states */
  99 enum lustre_imp_state {
 100         LUSTRE_IMP_CLOSED     = 1,
 101         LUSTRE_IMP_NEW        = 2,
 102         LUSTRE_IMP_DISCON     = 3,
 103         LUSTRE_IMP_CONNECTING = 4,
 104         LUSTRE_IMP_REPLAY     = 5,
 105         LUSTRE_IMP_REPLAY_LOCKS = 6,
 106         LUSTRE_IMP_REPLAY_WAIT  = 7,
 107         LUSTRE_IMP_RECOVER    = 8,
 108         LUSTRE_IMP_FULL       = 9,
 109         LUSTRE_IMP_EVICTED    = 10,
 110         LUSTRE_IMP_IDLE       = 11,
 111         LUSTRE_IMP_LAST
 112 };
 113
 114 /** Returns test string representation of numeric import state \a state */
 115 static inline char * ptlrpc_import_state_name(enum lustre_imp_state state)
 116 {
 117         static char *import_state_names[] = {
 118                 "<UNKNOWN>", "CLOSED",  "NEW", "DISCONN",
 119                 "CONNECTING", "REPLAY", "REPLAY_LOCKS", "REPLAY_WAIT",
 120                 "RECOVER", "FULL", "EVICTED", "IDLE",
 121         };
 122
 123         LASSERT(state < LUSTRE_IMP_LAST);
 124         return import_state_names[state];
 125 }
 126
 127 /**
 128  * List of import event types
 129  */
 130 enum obd_import_event {
 131         IMP_EVENT_DISCON     = 0x808001,
 132         IMP_EVENT_INACTIVE   = 0x808002,
 133         IMP_EVENT_INVALIDATE = 0x808003,
 134         IMP_EVENT_ACTIVE     = 0x808004,
 135         IMP_EVENT_OCD        = 0x808005,
 136         IMP_EVENT_DEACTIVATE = 0x808006,
 137         IMP_EVENT_ACTIVATE   = 0x808007,
 138 };
 139
 140 /**
 141  * Definition of import connection structure
 142  */
 143 struct obd_import_conn {
 144         /** Item for linking connections together */
 145         struct list_head          oic_item;
 146         /** Pointer to actual PortalRPC connection */
 147         struct ptlrpc_connection *oic_conn;
 148         /** uuid of remote side */
 149         struct obd_uuid           oic_uuid;
 150         /**
 151          * Time (64 bit seconds) of last connection attempt on this connection
 152          */
 153         time64_t                  oic_last_attempt;
 154 };
 155
 156 /* state history */
 157 #define IMP_STATE_HIST_LEN 16
 158 struct import_state_hist {
 159         enum lustre_imp_state   ish_state;
 160         time64_t                ish_time;
 161 };
 162
 163 /**
 164  * Defintion of PortalRPC import structure.
 165  * Imports are representing client-side view to remote target.
 166  */
 167 struct obd_import {
 168         /** Reference counter */
 169         refcount_t                imp_refcount;
 170         struct lustre_handle      imp_dlm_handle; /* client's ldlm export */
 171         /** Currently active connection */
 172         struct ptlrpc_connection *imp_connection;
 173         /** PortalRPC client structure for this import */
 174         struct ptlrpc_client     *imp_client;
 175         /** List element for linking into pinger chain */
 176         struct list_head          imp_pinger_chain;
 177         /** work struct for destruction of import */
 178         struct work_struct        imp_zombie_work;
 179
 180         /**
 181          * Lists of requests that are retained for replay, waiting for a reply,
 182          * or waiting for recovery to complete, respectively.
 183          * @{
 184          */
 185         struct list_head        imp_replay_list;
 186         struct list_head        imp_sending_list;
 187         struct list_head        imp_delayed_list;
 188         /** @} */
 189
 190         /**
 191          * List of requests that are retained for committed open replay. Once
 192          * open is committed, open replay request will be moved from the
 193          * imp_replay_list into the imp_committed_list.
 194          * The imp_replay_cursor is for accelerating searching during replay.
 195          * @{
 196          */
 197         struct list_head        imp_committed_list;
 198         struct list_head        *imp_replay_cursor;
 199         /** @} */
 200
 201         /** List of not replied requests */
 202         struct list_head        imp_unreplied_list;
 203         /** Known maximal replied XID */
 204         __u64                   imp_known_replied_xid;
 205
 206         /** obd device for this import */
 207         struct obd_device       *imp_obd;
 208
 209         /**
 210          * some seciruty-related fields
 211          * @{
 212          */
 213         struct ptlrpc_sec        *imp_sec;
 214         rwlock_t                  imp_sec_lock;
 215         time64_t                imp_sec_expire;
 216         pid_t                     imp_sec_refpid;
 217         /** @} */
 218
 219         /** Wait queue for those who need to wait for recovery completion */
 220         wait_queue_head_t         imp_recovery_waitq;
 221
 222         /** Number of requests currently in-flight */
 223         atomic_t                  imp_inflight;
 224         /** Number of requests currently unregistering */
 225         atomic_t                  imp_unregistering;
 226         /** Number of replay requests inflight */
 227         atomic_t                  imp_replay_inflight;
 228         /** Number of currently happening import invalidations */
 229         atomic_t                  imp_inval_count;
 230         /** Numbner of request timeouts */
 231         atomic_t                  imp_timeouts;
 232         /** Current import state */
 233         enum lustre_imp_state     imp_state;
 234         /** Last replay state */
 235         enum lustre_imp_state     imp_replay_state;
 236         /** History of import states */
 237         struct import_state_hist  imp_state_hist[IMP_STATE_HIST_LEN];
 238         int                       imp_state_hist_idx;
 239         /** Current import generation. Incremented on every reconnect */
 240         int                       imp_generation;
 241         /** Idle connection initiated at this generation */
 242         int                       imp_initiated_at;
 243         /** Incremented every time we send reconnection request */
 244         __u32                     imp_conn_cnt;
 245        /**
 246         * \see ptlrpc_free_committed remembers imp_generation value here
 247         * after a check to save on unnecessary replay list iterations
 248         */
 249         int                       imp_last_generation_checked;
 250         /** Last tranno we replayed */
 251         __u64                     imp_last_replay_transno;
 252         /** Last transno committed on remote side */
 253         __u64                     imp_peer_committed_transno;
 254         /**
 255          * \see ptlrpc_free_committed remembers last_transno since its last
 256          * check here and if last_transno did not change since last run of
 257          * ptlrpc_free_committed and import generation is the same, we can
 258          * skip looking for requests to remove from replay list as optimisation
 259          */
 260         __u64                     imp_last_transno_checked;
 261         /**
 262          * Remote export handle. This is how remote side knows what export
 263          * we are talking to. Filled from response to connect request
 264          */
 265         struct lustre_handle      imp_remote_handle;
 266         /** When to perform next ping. time in jiffies. */
 267         time64_t                imp_next_ping;
 268         /** When we last successfully connected. time in 64bit jiffies */
 269         time64_t                imp_last_success_conn;
 270
 271         /** List of all possible connection for import. */
 272         struct list_head        imp_conn_list;
 273         /**
 274          * Current connection. \a imp_connection is imp_conn_current->oic_conn
 275          */
 276         struct obd_import_conn   *imp_conn_current;
 277
 278         /** Protects flags, level, generation, conn_cnt, *_list */
 279         spinlock_t                imp_lock;
 280
 281         /* flags */
 282         unsigned long             imp_invalid:1,    /* evicted */
 283                                   /* administratively disabled */
 284                                   imp_deactive:1,
 285                                   /* try to recover the import */
 286                                   imp_replayable:1,
 287                                   /* don't run recovery (timeout instead) */
 288                                   imp_dlm_fake:1,
 289                                   /* use 1/2 timeout on MDS' OSCs */
 290                                   imp_server_timeout:1,
 291                                   /* VBR: imp in delayed recovery */
 292                                   imp_delayed_recovery:1,
 293                                   /* recovery by versions was failed */
 294                                   imp_vbr_failed:1,
 295                                   /* force an immidiate ping */
 296                                   imp_force_verify:1,
 297                                   /* force a scheduled ping */
 298                                   imp_force_next_verify:1,
 299                                   /* pingable */
 300                                   imp_pingable:1,
 301                                   /* resend for replay */
 302                                   imp_resend_replay:1,
 303                                   /* disable normal recovery, for test only. */
 304                                   imp_no_pinger_recover:1,
 305                                   /* import must be reconnected instead of
 306                                    * chouse new connection */
 307                                   imp_force_reconnect:1,
 308                                   /* import has tried to connect with server */
 309                                   imp_connect_tried:1,
 310                                   /* connected but not FULL yet */
 311                                   imp_connected:1,
 312                                   /* grant shrink disabled */
 313                                   imp_grant_shrink_disabled:1,
 314                                   /* to supress LCONSOLE() at conn.restore */
 315                                   imp_was_idle:1;
 316         u32                       imp_connect_op;
 317         u32                       imp_idle_timeout;
 318         u32                       imp_idle_debug;
 319         struct obd_connect_data   imp_connect_data;
 320         __u64                     imp_connect_flags_orig;
 321         __u64                     imp_connect_flags2_orig;
 322         int                       imp_connect_error;
 323
 324         enum lustre_msg_magic   imp_msg_magic;
 325                                 /* adjusted based on server capability */
 326         enum lustre_msghdr      imp_msghdr_flags;
 327
 328                                 /* adaptive timeout data */
 329         struct imp_at           imp_at;
 330         time64_t                imp_last_reply_time;    /* for health check */
 331 };
 332
 333 /* import.c */
 334 static inline unsigned int at_est2timeout(unsigned int val)
 335 {
 336         /* add an arbitrary minimum: 125% +5 sec */
 337         return (val + (val >> 2) + 5);
 338 }
 339
 340 static inline timeout_t at_timeout2est(timeout_t timeout)
 341 {
 342         /* restore estimate value from timeout: e=4/5(t-5) */
 343         LASSERT(timeout > 0);
 344         return max((timeout << 2) / 5, 5) - 4;
 345 }
 346
 347 static inline void at_reset_nolock(struct adaptive_timeout *at, int val)
 348 {
 349         at->at_current = val;
 350         at->at_worst_ever = val;
 351         at->at_worst_time = ktime_get_real_seconds();
 352 }
 353
 354 static inline void at_reset(struct adaptive_timeout *at, int val)
 355 {
 356         spin_lock(&at->at_lock);
 357         at_reset_nolock(at, val);
 358         spin_unlock(&at->at_lock);
 359 }
 360
 361 static inline void at_init(struct adaptive_timeout *at, int val, int flags) {
 362         memset(at, 0, sizeof(*at));
 363         spin_lock_init(&at->at_lock);
 364         at->at_flags = flags;
 365         at_reset(at, val);
 366 }
 367
 368 static inline void at_reinit(struct adaptive_timeout *at, int val, int flags)
 369 {
 370         spin_lock(&at->at_lock);
 371         at->at_binstart = 0;
 372         memset(at->at_hist, 0, sizeof(at->at_hist));
 373         at->at_flags = flags;
 374         at_reset_nolock(at, val);
 375         spin_unlock(&at->at_lock);
 376 }
 377
 378 extern unsigned int at_min;
 379 static inline int at_get(struct adaptive_timeout *at) {
 380         return (at->at_current > at_min) ? at->at_current : at_min;
 381 }
 382 int at_measured(struct adaptive_timeout *at, unsigned int val);
 383 int import_at_get_index(struct obd_import *imp, int portal);
 384 extern unsigned int at_max;
 385 #define AT_OFF (at_max == 0)
 386
 387 /* genops.c */
 388 struct obd_export;
 389 extern struct obd_import *class_exp2cliimp(struct obd_export *);
 390
 391 /** @} import */
 392
 393 #endif /* __IMPORT_H */
 394
 395 /** @} obd_import */