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 #ifdef HAVE_REFCOUNT_T
  49 #include <linux/refcount.h>
  50 #else
  51 #include <libcfs/linux/linux-refcount.h>
  52 #endif
  53 #include <linux/spinlock.h>
  54 #include <linux/time.h>
  55 #include <linux/types.h>
  56 #include <linux/workqueue.h>
  57 #include <libcfs/libcfs.h>
  58 #include <uapi/linux/lustre/lustre_idl.h>
  59
  60 /**
  61  * Adaptive Timeout stuff
  62  *
  63  * @{
  64  */
  65 #define D_ADAPTTO D_OTHER
  66 #define AT_BINS 4                  /* "bin" means "N seconds of history" */
  67 #define AT_FLG_NOHIST 0x1          /* use last reported value only */
  68
  69 struct adaptive_timeout {
  70         time64_t        at_binstart;         /* bin start time */
  71         unsigned int    at_hist[AT_BINS];    /* timeout history bins */
  72         unsigned int    at_flags;
  73         unsigned int    at_current;          /* current timeout value */
  74         unsigned int    at_worst_ever;       /* worst-ever timeout value */
  75         time64_t        at_worst_time;       /* worst-ever timeout timestamp */
  76         spinlock_t      at_lock;
  77 };
  78
  79 enum lustre_at_flags {
  80         LATF_SKIP       = 0x0,
  81         LATF_STATS      = 0x1,
  82 };
  83
  84 struct ptlrpc_at_array {
  85         struct list_head *paa_reqs_array; /** array to hold requests */
  86         __u32             paa_size;       /** the size of array */
  87         __u32             paa_count;      /** the total count of reqs */
  88         time64_t          paa_deadline;   /** the earliest deadline of reqs */
  89         __u32            *paa_reqs_count; /** the count of reqs in each entry */
  90 };
  91
  92 #define IMP_AT_MAX_PORTALS 8
  93 struct imp_at {
  94         int                     iat_portal[IMP_AT_MAX_PORTALS];
  95         struct adaptive_timeout iat_net_latency;
  96         struct adaptive_timeout iat_service_estimate[IMP_AT_MAX_PORTALS];
  97 };
  98
  99
 100 /** @} */
 101
 102 /** Possible import states */
 103 enum lustre_imp_state {
 104         LUSTRE_IMP_CLOSED     = 1,
 105         LUSTRE_IMP_NEW        = 2,
 106         LUSTRE_IMP_DISCON     = 3,
 107         LUSTRE_IMP_CONNECTING = 4,
 108         LUSTRE_IMP_REPLAY     = 5,
 109         LUSTRE_IMP_REPLAY_LOCKS = 6,
 110         LUSTRE_IMP_REPLAY_WAIT  = 7,
 111         LUSTRE_IMP_RECOVER    = 8,
 112         LUSTRE_IMP_FULL       = 9,
 113         LUSTRE_IMP_EVICTED    = 10,
 114         LUSTRE_IMP_IDLE       = 11,
 115         LUSTRE_IMP_LAST
 116 };
 117
 118 /** Returns test string representation of numeric import state \a state */
 119 static inline char * ptlrpc_import_state_name(enum lustre_imp_state state)
 120 {
 121         static char *import_state_names[] = {
 122                 "<UNKNOWN>", "CLOSED",  "NEW", "DISCONN",
 123                 "CONNECTING", "REPLAY", "REPLAY_LOCKS", "REPLAY_WAIT",
 124                 "RECOVER", "FULL", "EVICTED", "IDLE",
 125         };
 126
 127         LASSERT(state < LUSTRE_IMP_LAST);
 128         return import_state_names[state];
 129 }
 130
 131 /**
 132  * List of import event types
 133  */
 134 enum obd_import_event {
 135         IMP_EVENT_DISCON     = 0x808001,
 136         IMP_EVENT_INACTIVE   = 0x808002,
 137         IMP_EVENT_INVALIDATE = 0x808003,
 138         IMP_EVENT_ACTIVE     = 0x808004,
 139         IMP_EVENT_OCD        = 0x808005,
 140         IMP_EVENT_DEACTIVATE = 0x808006,
 141         IMP_EVENT_ACTIVATE   = 0x808007,
 142 };
 143
 144 /**
 145  * Definition of import connection structure
 146  */
 147 struct obd_import_conn {
 148         /** Item for linking connections together */
 149         struct list_head          oic_item;
 150         /** Pointer to actual PortalRPC connection */
 151         struct ptlrpc_connection *oic_conn;
 152         /** uuid of remote side */
 153         struct obd_uuid           oic_uuid;
 154         /**
 155          * Time (64 bit seconds) of last connection attempt on this connection
 156          */
 157         time64_t                  oic_last_attempt;
 158 };
 159
 160 /* state history */
 161 #define IMP_STATE_HIST_LEN 16
 162 struct import_state_hist {
 163         enum lustre_imp_state   ish_state;
 164         time64_t                ish_time;
 165 };
 166
 167 /**
 168  * Defintion of PortalRPC import structure.
 169  * Imports are representing client-side view to remote target.
 170  */
 171 struct obd_import {
 172         /** Reference counter */
 173         refcount_t                imp_refcount;
 174         struct lustre_handle      imp_dlm_handle; /* client's ldlm export */
 175         /** Currently active connection */
 176         struct ptlrpc_connection *imp_connection;
 177         /** PortalRPC client structure for this import */
 178         struct ptlrpc_client     *imp_client;
 179         /** List element for linking into pinger chain */
 180         struct list_head          imp_pinger_chain;
 181         /** work struct for destruction of import */
 182         struct work_struct        imp_zombie_work;
 183
 184         /**
 185          * Lists of requests that are retained for replay, waiting for a reply,
 186          * or waiting for recovery to complete, respectively.
 187          * @{
 188          */
 189         struct list_head        imp_replay_list;
 190         struct list_head        imp_sending_list;
 191         struct list_head        imp_delayed_list;
 192         /** @} */
 193
 194         /**
 195          * List of requests that are retained for committed open replay. Once
 196          * open is committed, open replay request will be moved from the
 197          * imp_replay_list into the imp_committed_list.
 198          * The imp_replay_cursor is for accelerating searching during replay.
 199          * @{
 200          */
 201         struct list_head        imp_committed_list;
 202         struct list_head        *imp_replay_cursor;
 203         /** @} */
 204
 205         /** List of not replied requests */
 206         struct list_head        imp_unreplied_list;
 207         /** Known maximal replied XID */
 208         __u64                   imp_known_replied_xid;
 209
 210         /** obd device for this import */
 211         struct obd_device       *imp_obd;
 212
 213         /**
 214          * some seciruty-related fields
 215          * @{
 216          */
 217         struct ptlrpc_sec        *imp_sec;
 218         rwlock_t                  imp_sec_lock;
 219         time64_t                imp_sec_expire;
 220         pid_t                     imp_sec_refpid;
 221         /** @} */
 222
 223         /** Wait queue for those who need to wait for recovery completion */
 224         wait_queue_head_t         imp_recovery_waitq;
 225
 226         /** Number of requests currently in-flight */
 227         atomic_t                  imp_inflight;
 228         /** Number of requests currently unregistering */
 229         atomic_t                  imp_unregistering;
 230         /** Number of replay requests inflight */
 231         atomic_t                  imp_replay_inflight;
 232         /** Number of currently happening import invalidations */
 233         atomic_t                  imp_inval_count;
 234         /** Numbner of request timeouts */
 235         atomic_t                  imp_timeouts;
 236         /** Current import state */
 237         enum lustre_imp_state     imp_state;
 238         /** Last replay state */
 239         enum lustre_imp_state     imp_replay_state;
 240         /** History of import states */
 241         struct import_state_hist  imp_state_hist[IMP_STATE_HIST_LEN];
 242         int                       imp_state_hist_idx;
 243         /** Current import generation. Incremented on every reconnect */
 244         int                       imp_generation;
 245         /** Idle connection initiated at this generation */
 246         int                       imp_initiated_at;
 247         /** Incremented every time we send reconnection request */
 248         __u32                     imp_conn_cnt;
 249        /**
 250         * \see ptlrpc_free_committed remembers imp_generation value here
 251         * after a check to save on unnecessary replay list iterations
 252         */
 253         int                       imp_last_generation_checked;
 254         /** Last tranno we replayed */
 255         __u64                     imp_last_replay_transno;
 256         /** Last transno committed on remote side */
 257         __u64                     imp_peer_committed_transno;
 258         /**
 259          * \see ptlrpc_free_committed remembers last_transno since its last
 260          * check here and if last_transno did not change since last run of
 261          * ptlrpc_free_committed and import generation is the same, we can
 262          * skip looking for requests to remove from replay list as optimisation
 263          */
 264         __u64                     imp_last_transno_checked;
 265         /**
 266          * Remote export handle. This is how remote side knows what export
 267          * we are talking to. Filled from response to connect request
 268          */
 269         struct lustre_handle      imp_remote_handle;
 270         /** When to perform next ping. time in jiffies. */
 271         time64_t                imp_next_ping;
 272         /** When we last successfully connected. time in 64bit jiffies */
 273         time64_t                imp_last_success_conn;
 274
 275         /** List of all possible connection for import. */
 276         struct list_head        imp_conn_list;
 277         /**
 278          * Current connection. \a imp_connection is imp_conn_current->oic_conn
 279          */
 280         struct obd_import_conn   *imp_conn_current;
 281
 282         /** Protects flags, level, generation, conn_cnt, *_list */
 283         spinlock_t                imp_lock;
 284
 285         /* flags */
 286         unsigned long             imp_invalid:1,    /* evicted */
 287                                   /* administratively disabled */
 288                                   imp_deactive:1,
 289                                   /* try to recover the import */
 290                                   imp_replayable:1,
 291                                   /* don't run recovery (timeout instead) */
 292                                   imp_dlm_fake:1,
 293                                   /* use 1/2 timeout on MDS' OSCs */
 294                                   imp_server_timeout:1,
 295                                   /* VBR: imp in delayed recovery */
 296                                   imp_delayed_recovery:1,
 297                                   /* recovery by versions was failed */
 298                                   imp_vbr_failed:1,
 299                                   /* force an immidiate ping */
 300                                   imp_force_verify:1,
 301                                   /* force a scheduled ping */
 302                                   imp_force_next_verify:1,
 303                                   /* pingable */
 304                                   imp_pingable:1,
 305                                   /* resend for replay */
 306                                   imp_resend_replay:1,
 307                                   /* disable normal recovery, for test only. */
 308                                   imp_no_pinger_recover:1,
 309                                   /* import must be reconnected instead of
 310                                    * chouse new connection */
 311                                   imp_force_reconnect:1,
 312                                   /* import has tried to connect with server */
 313                                   imp_connect_tried:1,
 314                                   /* connected but not FULL yet */
 315                                   imp_connected:1,
 316                                   /* grant shrink disabled */
 317                                   imp_grant_shrink_disabled:1,
 318                                   /* to supress LCONSOLE() at conn.restore */
 319                                   imp_was_idle:1;
 320         u32                       imp_connect_op;
 321         u32                       imp_idle_timeout;
 322         u32                       imp_idle_debug;
 323         struct obd_connect_data   imp_connect_data;
 324         __u64                     imp_connect_flags_orig;
 325         __u64                     imp_connect_flags2_orig;
 326         int                       imp_connect_error;
 327
 328         enum lustre_msg_magic   imp_msg_magic;
 329                                 /* adjusted based on server capability */
 330         enum lustre_msghdr      imp_msghdr_flags;
 331
 332                                 /* adaptive timeout data */
 333         struct imp_at           imp_at;
 334         time64_t                imp_last_reply_time;    /* for health check */
 335 };
 336
 337 /* import.c */
 338 static inline unsigned int at_est2timeout(unsigned int val)
 339 {
 340         /* add an arbitrary minimum: 125% +5 sec */
 341         return (val + (val >> 2) + 5);
 342 }
 343
 344 static inline unsigned int at_timeout2est(unsigned int val)
 345 {
 346         /* restore estimate value from timeout: e=4/5(t-5) */
 347         LASSERT(val);
 348         return (max((val << 2) / 5, 5U) - 4);
 349 }
 350
 351 static inline void at_reset_nolock(struct adaptive_timeout *at, int val)
 352 {
 353         at->at_current = val;
 354         at->at_worst_ever = val;
 355         at->at_worst_time = ktime_get_real_seconds();
 356 }
 357
 358 static inline void at_reset(struct adaptive_timeout *at, int val)
 359 {
 360         spin_lock(&at->at_lock);
 361         at_reset_nolock(at, val);
 362         spin_unlock(&at->at_lock);
 363 }
 364
 365 static inline void at_init(struct adaptive_timeout *at, int val, int flags) {
 366         memset(at, 0, sizeof(*at));
 367         spin_lock_init(&at->at_lock);
 368         at->at_flags = flags;
 369         at_reset(at, val);
 370 }
 371
 372 static inline void at_reinit(struct adaptive_timeout *at, int val, int flags)
 373 {
 374         spin_lock(&at->at_lock);
 375         at->at_binstart = 0;
 376         memset(at->at_hist, 0, sizeof(at->at_hist));
 377         at->at_flags = flags;
 378         at_reset_nolock(at, val);
 379         spin_unlock(&at->at_lock);
 380 }
 381
 382 extern unsigned int at_min;
 383 static inline int at_get(struct adaptive_timeout *at) {
 384         return (at->at_current > at_min) ? at->at_current : at_min;
 385 }
 386 int at_measured(struct adaptive_timeout *at, unsigned int val);
 387 int import_at_get_index(struct obd_import *imp, int portal);
 388 extern unsigned int at_max;
 389 #define AT_OFF (at_max == 0)
 390
 391 /* genops.c */
 392 struct obd_export;
 393 extern struct obd_import *class_exp2cliimp(struct obd_export *);
 394
 395 /** @} import */
 396
 397 #endif /* __IMPORT_H */
 398
 399 /** @} obd_import */