lustre/include/lustre_import.h

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef __IMPORT_H
  38 #define __IMPORT_H
  39
  40 /** \defgroup export export
  41  *
  42  * @{
  43  */
  44
  45 #include <lustre_handles.h>
  46 #include <lustre/lustre_idl.h>
  47
  48
  49 /* Adaptive Timeout stuff */
  50 #define D_ADAPTTO D_OTHER
  51 #define AT_BINS 4                  /* "bin" means "N seconds of history" */
  52 #define AT_FLG_NOHIST 0x1          /* use last reported value only */
  53
  54 struct adaptive_timeout {
  55         time_t           at_binstart;         /* bin start time */
  56         unsigned int     at_hist[AT_BINS];    /* timeout history bins */
  57         unsigned int     at_flags;
  58         unsigned int     at_current;          /* current timeout value */
  59         unsigned int     at_worst_ever;       /* worst-ever timeout value */
  60         time_t           at_worst_time;       /* worst-ever timeout timestamp */
  61         cfs_spinlock_t   at_lock;
  62 };
  63
  64 enum lustre_imp_state {
  65         LUSTRE_IMP_CLOSED     = 1,
  66         LUSTRE_IMP_NEW        = 2,
  67         LUSTRE_IMP_DISCON     = 3,
  68         LUSTRE_IMP_CONNECTING = 4,
  69         LUSTRE_IMP_REPLAY     = 5,
  70         LUSTRE_IMP_REPLAY_LOCKS = 6,
  71         LUSTRE_IMP_REPLAY_WAIT  = 7,
  72         LUSTRE_IMP_RECOVER    = 8,
  73         LUSTRE_IMP_FULL       = 9,
  74         LUSTRE_IMP_EVICTED    = 10,
  75 };
  76
  77 struct ptlrpc_at_array {
  78         cfs_list_t       *paa_reqs_array; /* array to hold requests */
  79         __u32             paa_size;       /* the size of array */
  80         __u32             paa_count;      /* the total count of reqs */
  81         time_t            paa_deadline;   /* earliest deadline of reqs */
  82         __u32            *paa_reqs_count; /* count of reqs in each entry */
  83 };
  84
  85 static inline char * ptlrpc_import_state_name(enum lustre_imp_state state)
  86 {
  87         static char* import_state_names[] = {
  88                 "<UNKNOWN>", "CLOSED",  "NEW", "DISCONN",
  89                 "CONNECTING", "REPLAY", "REPLAY_LOCKS", "REPLAY_WAIT",
  90                 "RECOVER", "FULL", "EVICTED",
  91         };
  92
  93         LASSERT (state <= LUSTRE_IMP_EVICTED);
  94         return import_state_names[state];
  95 }
  96
  97 enum obd_import_event {
  98         IMP_EVENT_DISCON     = 0x808001,
  99         IMP_EVENT_INACTIVE   = 0x808002,
 100         IMP_EVENT_INVALIDATE = 0x808003,
 101         IMP_EVENT_ACTIVE     = 0x808004,
 102         IMP_EVENT_OCD        = 0x808005,
 103 };
 104
 105 struct obd_import_conn {
 106         cfs_list_t                oic_item;
 107         struct ptlrpc_connection *oic_conn;
 108         struct obd_uuid           oic_uuid;
 109         __u64                     oic_last_attempt; /* jiffies, 64-bit */
 110 };
 111
 112 #define IMP_AT_MAX_PORTALS 8
 113 struct imp_at {
 114         int                     iat_portal[IMP_AT_MAX_PORTALS];
 115         struct adaptive_timeout iat_net_latency;
 116         struct adaptive_timeout iat_service_estimate[IMP_AT_MAX_PORTALS];
 117 };
 118
 119 /* state history */
 120 #define IMP_STATE_HIST_LEN 16
 121 struct import_state_hist {
 122         enum lustre_imp_state ish_state;
 123         time_t                ish_time;
 124 };
 125
 126 struct obd_import {
 127         struct portals_handle     imp_handle;
 128         cfs_atomic_t              imp_refcount;
 129         struct lustre_handle      imp_dlm_handle; /* client's ldlm export */
 130         struct ptlrpc_connection *imp_connection;
 131         struct ptlrpc_client     *imp_client;
 132         cfs_list_t                imp_pinger_chain;
 133         cfs_list_t                imp_zombie_chain; /* queue for destruction */
 134
 135         /* Lists of requests that are retained for replay, waiting for a reply,
 136          * or waiting for recovery to complete, respectively.
 137          */
 138         cfs_list_t                imp_replay_list;
 139         cfs_list_t                imp_sending_list;
 140         cfs_list_t                imp_delayed_list;
 141
 142         struct obd_device        *imp_obd;
 143         struct ptlrpc_sec        *imp_sec;
 144         cfs_semaphore_t           imp_sec_mutex;
 145         cfs_time_t                imp_sec_expire;
 146         cfs_waitq_t               imp_recovery_waitq;
 147
 148         cfs_atomic_t              imp_inflight;
 149         cfs_atomic_t              imp_unregistering;
 150         cfs_atomic_t              imp_replay_inflight;
 151         cfs_atomic_t              imp_inval_count;  /* in-progress invalidations */
 152         cfs_atomic_t              imp_timeouts;
 153         enum lustre_imp_state     imp_state;
 154         struct import_state_hist  imp_state_hist[IMP_STATE_HIST_LEN];
 155         int                       imp_state_hist_idx;
 156         int                       imp_generation;
 157         __u32                     imp_conn_cnt;
 158         int                       imp_last_generation_checked;
 159         __u64                     imp_last_replay_transno;
 160         __u64                     imp_peer_committed_transno;
 161         __u64                     imp_last_transno_checked;
 162         struct lustre_handle      imp_remote_handle;
 163         cfs_time_t                imp_next_ping;   /* jiffies */
 164         __u64                     imp_last_success_conn;   /* jiffies, 64-bit */
 165
 166         /* all available obd_import_conn linked here */
 167         cfs_list_t                imp_conn_list;
 168         struct obd_import_conn   *imp_conn_current;
 169
 170         /* Protects flags, level, generation, conn_cnt, *_list */
 171         cfs_spinlock_t            imp_lock;
 172
 173         /* flags */
 174         unsigned long             imp_no_timeout:1,       /* timeouts are disabled */
 175                                   imp_invalid:1,          /* evicted */
 176                                   imp_deactive:1,         /* administratively disabled */
 177                                   imp_replayable:1,       /* try to recover the import */
 178                                   imp_dlm_fake:1,         /* don't run recovery (timeout instead) */
 179                                   imp_server_timeout:1,   /* use 1/2 timeout on MDS' OSCs */
 180                                   imp_delayed_recovery:1, /* VBR: imp in delayed recovery */
 181                                   imp_no_lock_replay:1,   /* VBR: if gap was found then no lock replays */
 182                                   imp_vbr_failed:1,       /* recovery by versions was failed */
 183                                   imp_force_verify:1,     /* force an immidiate ping */
 184                                   imp_pingable:1,         /* pingable */
 185                                   imp_resend_replay:1,    /* resend for replay */
 186                                   imp_force_reconnect:1;  /* import must be reconnected instead of chouse new connection */
 187         __u32                     imp_connect_op;
 188         struct obd_connect_data   imp_connect_data;
 189         __u64                     imp_connect_flags_orig;
 190         int                       imp_connect_error;
 191
 192         __u32                     imp_msg_magic;
 193         __u32                     imp_msghdr_flags;       /* adjusted based on server capability */
 194
 195         struct ptlrpc_request_pool *imp_rq_pool;          /* emergency request pool */
 196
 197         struct imp_at             imp_at;                 /* adaptive timeout data */
 198         time_t                    imp_last_reply_time;    /* for health check */
 199 };
 200
 201 typedef void (*obd_import_callback)(struct obd_import *imp, void *closure,
 202                                     int event, void *event_arg, void *cb_data);
 203
 204 struct obd_import_observer {
 205         cfs_list_t           oio_chain;
 206         obd_import_callback  oio_cb;
 207         void                *oio_cb_data;
 208 };
 209
 210 void class_observe_import(struct obd_import *imp, obd_import_callback cb,
 211                           void *cb_data);
 212 void class_unobserve_import(struct obd_import *imp, obd_import_callback cb,
 213                             void *cb_data);
 214 void class_notify_import_observers(struct obd_import *imp, int event,
 215                                    void *event_arg);
 216
 217 /* import.c */
 218 static inline unsigned int at_est2timeout(unsigned int val)
 219 {
 220         /* add an arbitrary minimum: 125% +5 sec */
 221         return (val + (val >> 2) + 5);
 222 }
 223
 224 static inline unsigned int at_timeout2est(unsigned int val)
 225 {
 226         /* restore estimate value from timeout: e=4/5(t-5) */
 227         LASSERT(val);
 228         return (max((val << 2) / 5, 5U) - 4);
 229 }
 230
 231 static inline void at_init(struct adaptive_timeout *at, int val, int flags) {
 232         memset(at, 0, sizeof(*at));
 233         at->at_current = val;
 234         at->at_worst_ever = val;
 235         at->at_worst_time = cfs_time_current_sec();
 236         at->at_flags = flags;
 237         cfs_spin_lock_init(&at->at_lock);
 238 }
 239 static inline int at_get(struct adaptive_timeout *at) {
 240         return at->at_current;
 241 }
 242 int at_measured(struct adaptive_timeout *at, unsigned int val);
 243 int import_at_get_index(struct obd_import *imp, int portal);
 244 extern unsigned int at_max;
 245 #define AT_OFF (at_max == 0)
 246
 247 /* genops.c */
 248 struct obd_export;
 249 extern struct obd_import *class_exp2cliimp(struct obd_export *);
 250 extern struct obd_import *class_conn2cliimp(struct lustre_handle *);
 251
 252 /** @} import */
 253
 254 #endif /* __IMPORT_H */