lustre/include/lustre_lib.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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, 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  * lustre/include/lustre_lib.h
  37  *
  38  * Basic Lustre library routines.
  39  */
  40
  41 #ifndef _LUSTRE_LIB_H
  42 #define _LUSTRE_LIB_H
  43
  44 /** \defgroup lib lib
  45  *
  46  * @{
  47  */
  48
  49 #include <libcfs/libcfs.h>
  50 #include <lustre/lustre_idl.h>
  51 #include <lustre_ver.h>
  52 #include <lustre_cfg.h>
  53 #include <linux/lustre_lib.h>
  54
  55 /* target.c */
  56 struct kstatfs;
  57 struct ptlrpc_request;
  58 struct obd_export;
  59 struct lu_target;
  60 struct l_wait_info;
  61 #include <lustre_ha.h>
  62 #include <lustre_net.h>
  63
  64 #ifdef HAVE_SERVER_SUPPORT
  65 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
  66                           int error);
  67 int target_handle_connect(struct ptlrpc_request *req);
  68 int target_handle_disconnect(struct ptlrpc_request *req);
  69 void target_destroy_export(struct obd_export *exp);
  70 int target_handle_ping(struct ptlrpc_request *req);
  71 void target_committed_to_req(struct ptlrpc_request *req);
  72 void target_cancel_recovery_timer(struct obd_device *obd);
  73 void target_stop_recovery_thread(struct obd_device *obd);
  74 void target_cleanup_recovery(struct obd_device *obd);
  75 int target_queue_recovery_request(struct ptlrpc_request *req,
  76                                   struct obd_device *obd);
  77 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
  78                    struct l_wait_info *lwi);
  79 #endif
  80
  81 int target_pack_pool_reply(struct ptlrpc_request *req);
  82 int do_set_info_async(struct obd_import *imp,
  83                       int opcode, int version,
  84                       obd_count keylen, void *key,
  85                       obd_count vallen, void *val,
  86                       struct ptlrpc_request_set *set);
  87
  88 #define OBD_RECOVERY_MAX_TIME (obd_timeout * 18) /* b13079 */
  89
  90 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id);
  91
  92 /* client.c */
  93
  94 int client_sanobd_setup(struct obd_device *obddev, struct lustre_cfg* lcfg);
  95 struct client_obd *client_conn2cli(struct lustre_handle *conn);
  96
  97 struct md_open_data;
  98 struct obd_client_handle {
  99         struct lustre_handle     och_fh;
 100         struct lu_fid            och_fid;
 101         struct md_open_data     *och_mod;
 102         struct lustre_handle     och_lease_handle; /* open lock for lease */
 103         __u32                    och_magic;
 104         int                      och_flags;
 105 };
 106 #define OBD_CLIENT_HANDLE_MAGIC 0xd15ea5ed
 107
 108 /* statfs_pack.c */
 109 void statfs_pack(struct obd_statfs *osfs, struct kstatfs *sfs);
 110 void statfs_unpack(struct kstatfs *sfs, struct obd_statfs *osfs);
 111
 112 /* Until such time as we get_info the per-stripe maximum from the OST,
 113  * we define this to be 2T - 4k, which is the ext3 maxbytes. */
 114 #define LUSTRE_STRIPE_MAXBYTES 0x1fffffff000ULL
 115
 116 /* Special values for remove LOV EA from disk */
 117 #define LOVEA_DELETE_VALUES(size, count, offset) (size == 0 && count == 0 && \
 118                                                  offset == (typeof(offset))(-1))
 119
 120 #define LMVEA_DELETE_VALUES(count, offset) ((count) == 0 && \
 121                                             (offset) == (typeof(offset))(-1))
 122 /* #define POISON_BULK 0 */
 123
 124 /*
 125  * l_wait_event is a flexible sleeping function, permitting simple caller
 126  * configuration of interrupt and timeout sensitivity along with actions to
 127  * be performed in the event of either exception.
 128  *
 129  * The first form of usage looks like this:
 130  *
 131  * struct l_wait_info lwi = LWI_TIMEOUT_INTR(timeout, timeout_handler,
 132  *                                           intr_handler, callback_data);
 133  * rc = l_wait_event(waitq, condition, &lwi);
 134  *
 135  * l_wait_event() makes the current process wait on 'waitq' until 'condition'
 136  * is TRUE or a "killable" signal (SIGTERM, SIKGILL, SIGINT) is pending.  It
 137  * returns 0 to signify 'condition' is TRUE, but if a signal wakes it before
 138  * 'condition' becomes true, it optionally calls the specified 'intr_handler'
 139  * if not NULL, and returns -EINTR.
 140  *
 141  * If a non-zero timeout is specified, signals are ignored until the timeout
 142  * has expired.  At this time, if 'timeout_handler' is not NULL it is called.
 143  * If it returns FALSE l_wait_event() continues to wait as described above with
 144  * signals enabled.  Otherwise it returns -ETIMEDOUT.
 145  *
 146  * LWI_INTR(intr_handler, callback_data) is shorthand for
 147  * LWI_TIMEOUT_INTR(0, NULL, intr_handler, callback_data)
 148  *
 149  * The second form of usage looks like this:
 150  *
 151  * struct l_wait_info lwi = LWI_TIMEOUT(timeout, timeout_handler);
 152  * rc = l_wait_event(waitq, condition, &lwi);
 153  *
 154  * This form is the same as the first except that it COMPLETELY IGNORES
 155  * SIGNALS.  The caller must therefore beware that if 'timeout' is zero, or if
 156  * 'timeout_handler' is not NULL and returns FALSE, then the ONLY thing that
 157  * can unblock the current process is 'condition' becoming TRUE.
 158  *
 159  * Another form of usage is:
 160  * struct l_wait_info lwi = LWI_TIMEOUT_INTERVAL(timeout, interval,
 161  *                                               timeout_handler);
 162  * rc = l_wait_event(waitq, condition, &lwi);
 163  * This is the same as previous case, but condition is checked once every
 164  * 'interval' jiffies (if non-zero).
 165  *
 166  * Subtle synchronization point: this macro does *not* necessary takes
 167  * wait-queue spin-lock before returning, and, hence, following idiom is safe
 168  * ONLY when caller provides some external locking:
 169  *
 170  *             Thread1                            Thread2
 171  *
 172  *   l_wait_event(&obj->wq, ....);                                       (1)
 173  *
 174  *                                    wake_up(&obj->wq):                 (2)
 175  *                                         spin_lock(&q->lock);          (2.1)
 176  *                                         __wake_up_common(q, ...);     (2.2)
 177  *                                         spin_unlock(&q->lock, flags); (2.3)
 178  *
 179  *   OBD_FREE_PTR(obj);                                                  (3)
 180  *
 181  * As l_wait_event() may "short-cut" execution and return without taking
 182  * wait-queue spin-lock, some additional synchronization is necessary to
 183  * guarantee that step (3) can begin only after (2.3) finishes.
 184  *
 185  * XXX nikita: some ptlrpc daemon threads have races of that sort.
 186  *
 187  */
 188 static inline int back_to_sleep(void *arg)
 189 {
 190         return 0;
 191 }
 192
 193 #define LWI_ON_SIGNAL_NOOP ((void (*)(void *))(-1))
 194
 195 struct l_wait_info {
 196         cfs_duration_t lwi_timeout;
 197         cfs_duration_t lwi_interval;
 198         int            lwi_allow_intr;
 199         int  (*lwi_on_timeout)(void *);
 200         void (*lwi_on_signal)(void *);
 201         void  *lwi_cb_data;
 202 };
 203
 204 /* NB: LWI_TIMEOUT ignores signals completely */
 205 #define LWI_TIMEOUT(time, cb, data)             \
 206 ((struct l_wait_info) {                         \
 207         .lwi_timeout    = time,                 \
 208         .lwi_on_timeout = cb,                   \
 209         .lwi_cb_data    = data,                 \
 210         .lwi_interval   = 0,                    \
 211         .lwi_allow_intr = 0                     \
 212 })
 213
 214 #define LWI_TIMEOUT_INTERVAL(time, interval, cb, data)  \
 215 ((struct l_wait_info) {                                 \
 216         .lwi_timeout    = time,                         \
 217         .lwi_on_timeout = cb,                           \
 218         .lwi_cb_data    = data,                         \
 219         .lwi_interval   = interval,                     \
 220         .lwi_allow_intr = 0                             \
 221 })
 222
 223 #define LWI_TIMEOUT_INTR(time, time_cb, sig_cb, data)   \
 224 ((struct l_wait_info) {                                 \
 225         .lwi_timeout    = time,                         \
 226         .lwi_on_timeout = time_cb,                      \
 227         .lwi_on_signal  = sig_cb,                       \
 228         .lwi_cb_data    = data,                         \
 229         .lwi_interval   = 0,                            \
 230         .lwi_allow_intr = 0                             \
 231 })
 232
 233 #define LWI_TIMEOUT_INTR_ALL(time, time_cb, sig_cb, data)       \
 234 ((struct l_wait_info) {                                         \
 235         .lwi_timeout    = time,                                 \
 236         .lwi_on_timeout = time_cb,                              \
 237         .lwi_on_signal  = sig_cb,                               \
 238         .lwi_cb_data    = data,                                 \
 239         .lwi_interval   = 0,                                    \
 240         .lwi_allow_intr = 1                                     \
 241 })
 242
 243 #define LWI_INTR(cb, data)  LWI_TIMEOUT_INTR(0, NULL, cb, data)
 244
 245
 246 /*
 247  * wait for @condition to become true, but no longer than timeout, specified
 248  * by @info.
 249  */
 250 #define __l_wait_event(wq, condition, info, ret, l_add_wait)                   \
 251 do {                                                                           \
 252         wait_queue_t __wait;                                                   \
 253         cfs_duration_t __timeout = info->lwi_timeout;                          \
 254         sigset_t   __blocked;                                              \
 255         int   __allow_intr = info->lwi_allow_intr;                             \
 256                                                                                \
 257         ret = 0;                                                               \
 258         if (condition)                                                         \
 259                 break;                                                         \
 260                                                                                \
 261         init_waitqueue_entry_current(&__wait);                                 \
 262         l_add_wait(&wq, &__wait);                                              \
 263                                                                                \
 264         /* Block all signals (just the non-fatal ones if no timeout). */       \
 265         if (info->lwi_on_signal != NULL && (__timeout == 0 || __allow_intr))   \
 266                 __blocked = cfs_block_sigsinv(LUSTRE_FATAL_SIGS);              \
 267         else                                                                   \
 268                 __blocked = cfs_block_sigsinv(0);                              \
 269                                                                                \
 270         for (;;) {                                                             \
 271                 unsigned       __wstate;                                       \
 272                                                                                \
 273                 __wstate = info->lwi_on_signal != NULL &&                      \
 274                            (__timeout == 0 || __allow_intr) ?                  \
 275                         TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;             \
 276                                                                                \
 277                 set_current_state(TASK_INTERRUPTIBLE);                         \
 278                                                                                \
 279                 if (condition)                                                 \
 280                         break;                                                 \
 281                                                                                \
 282                 if (__timeout == 0) {                                          \
 283                         waitq_wait(&__wait, __wstate);                         \
 284                 } else {                                                       \
 285                         cfs_duration_t interval = info->lwi_interval?          \
 286                                              min_t(cfs_duration_t,             \
 287                                                  info->lwi_interval,__timeout):\
 288                                              __timeout;                        \
 289                         cfs_duration_t remaining = waitq_timedwait(&__wait,    \
 290                                                    __wstate,                   \
 291                                                    interval);                  \
 292                         __timeout = cfs_time_sub(__timeout,                    \
 293                                             cfs_time_sub(interval, remaining));\
 294                         if (__timeout == 0) {                                  \
 295                                 if (info->lwi_on_timeout == NULL ||            \
 296                                     info->lwi_on_timeout(info->lwi_cb_data)) { \
 297                                         ret = -ETIMEDOUT;                      \
 298                                         break;                                 \
 299                                 }                                              \
 300                                 /* Take signals after the timeout expires. */  \
 301                                 if (info->lwi_on_signal != NULL)               \
 302                                     (void)cfs_block_sigsinv(LUSTRE_FATAL_SIGS);\
 303                         }                                                      \
 304                 }                                                              \
 305                                                                                \
 306                 if (condition)                                                 \
 307                         break;                                                 \
 308                 if (cfs_signal_pending()) {                                    \
 309                         if (info->lwi_on_signal != NULL &&                     \
 310                             (__timeout == 0 || __allow_intr)) {                \
 311                                 if (info->lwi_on_signal != LWI_ON_SIGNAL_NOOP) \
 312                                         info->lwi_on_signal(info->lwi_cb_data);\
 313                                 ret = -EINTR;                                  \
 314                                 break;                                         \
 315                         }                                                      \
 316                         /* We have to do this here because some signals */     \
 317                         /* are not blockable - ie from strace(1).       */     \
 318                         /* In these cases we want to schedule_timeout() */     \
 319                         /* again, because we don't want that to return  */     \
 320                         /* -EINTR when the RPC actually succeeded.      */     \
 321                         /* the recalc_sigpending() below will deliver the */   \
 322                         /* signal properly.                             */     \
 323                         cfs_clear_sigpending();                                \
 324                 }                                                              \
 325         }                                                                      \
 326                                                                                \
 327         cfs_restore_sigs(__blocked);                                           \
 328                                                                                \
 329         set_current_state(TASK_RUNNING);                                       \
 330         remove_wait_queue(&wq, &__wait);                                       \
 331 } while (0)
 332
 333
 334
 335 #define l_wait_event(wq, condition, info)                       \
 336 ({                                                              \
 337         int                 __ret;                              \
 338         struct l_wait_info *__info = (info);                    \
 339                                                                 \
 340         __l_wait_event(wq, condition, __info,                   \
 341                        __ret, add_wait_queue);                  \
 342         __ret;                                                  \
 343 })
 344
 345 #define l_wait_event_exclusive(wq, condition, info)             \
 346 ({                                                              \
 347         int                 __ret;                              \
 348         struct l_wait_info *__info = (info);                    \
 349                                                                 \
 350         __l_wait_event(wq, condition, __info,                   \
 351                        __ret, add_wait_queue_exclusive);        \
 352         __ret;                                                  \
 353 })
 354
 355 #define l_wait_event_exclusive_head(wq, condition, info)        \
 356 ({                                                              \
 357         int                 __ret;                              \
 358         struct l_wait_info *__info = (info);                    \
 359                                                                 \
 360         __l_wait_event(wq, condition, __info,                   \
 361                        __ret, add_wait_queue_exclusive_head);   \
 362         __ret;                                                  \
 363 })
 364
 365 #define l_wait_condition(wq, condition)                         \
 366 ({                                                              \
 367         struct l_wait_info lwi = { 0 };                         \
 368         l_wait_event(wq, condition, &lwi);                      \
 369 })
 370
 371 #define l_wait_condition_exclusive(wq, condition)               \
 372 ({                                                              \
 373         struct l_wait_info lwi = { 0 };                         \
 374         l_wait_event_exclusive(wq, condition, &lwi);            \
 375 })
 376
 377 #define l_wait_condition_exclusive_head(wq, condition)          \
 378 ({                                                              \
 379         struct l_wait_info lwi = { 0 };                         \
 380         l_wait_event_exclusive_head(wq, condition, &lwi);       \
 381 })
 382
 383 /** @} lib */
 384
 385 #endif /* _LUSTRE_LIB_H */