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 #if defined(__linux__)
  54 #include <linux/lustre_lib.h>
  55 #elif defined(__APPLE__)
  56 #include <darwin/lustre_lib.h>
  57 #elif defined(__WINNT__)
  58 #include <winnt/lustre_lib.h>
  59 #else
  60 #error Unsupported operating system.
  61 #endif
  62
  63 /* target.c */
  64 struct ptlrpc_request;
  65 struct obd_export;
  66 struct lu_target;
  67 struct l_wait_info;
  68 #include <lustre_ha.h>
  69 #include <lustre_net.h>
  70
  71 #ifdef HAVE_SERVER_SUPPORT
  72 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
  73                           int error);
  74 int target_handle_connect(struct ptlrpc_request *req);
  75 int target_handle_disconnect(struct ptlrpc_request *req);
  76 void target_destroy_export(struct obd_export *exp);
  77 int target_handle_ping(struct ptlrpc_request *req);
  78 void target_committed_to_req(struct ptlrpc_request *req);
  79 void target_cancel_recovery_timer(struct obd_device *obd);
  80 void target_stop_recovery_thread(struct obd_device *obd);
  81 void target_cleanup_recovery(struct obd_device *obd);
  82 int target_queue_recovery_request(struct ptlrpc_request *req,
  83                                   struct obd_device *obd);
  84 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
  85                    struct l_wait_info *lwi);
  86 #endif
  87
  88 int target_pack_pool_reply(struct ptlrpc_request *req);
  89 int do_set_info_async(struct obd_import *imp,
  90                       int opcode, int version,
  91                       obd_count keylen, void *key,
  92                       obd_count vallen, void *val,
  93                       struct ptlrpc_request_set *set);
  94
  95 #define OBD_RECOVERY_MAX_TIME (obd_timeout * 18) /* b13079 */
  96
  97 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id);
  98
  99 /* client.c */
 100
 101 int client_sanobd_setup(struct obd_device *obddev, struct lustre_cfg* lcfg);
 102 struct client_obd *client_conn2cli(struct lustre_handle *conn);
 103
 104 struct md_open_data;
 105 struct obd_client_handle {
 106         struct lustre_handle     och_fh;
 107         struct lu_fid            och_fid;
 108         struct md_open_data     *och_mod;
 109         struct lustre_handle     och_lease_handle; /* open lock for lease */
 110         __u32                    och_magic;
 111         int                      och_flags;
 112 };
 113 #define OBD_CLIENT_HANDLE_MAGIC 0xd15ea5ed
 114
 115 /* statfs_pack.c */
 116 void statfs_pack(struct obd_statfs *osfs, struct kstatfs *sfs);
 117 void statfs_unpack(struct kstatfs *sfs, struct obd_statfs *osfs);
 118
 119 /* Until such time as we get_info the per-stripe maximum from the OST,
 120  * we define this to be 2T - 4k, which is the ext3 maxbytes. */
 121 #define LUSTRE_STRIPE_MAXBYTES 0x1fffffff000ULL
 122
 123 /* Special values for remove LOV EA from disk */
 124 #define LOVEA_DELETE_VALUES(size, count, offset) (size == 0 && count == 0 && \
 125                                                  offset == (typeof(offset))(-1))
 126
 127 #define LMVEA_DELETE_VALUES(count, offset) ((count) == 0 && \
 128                                             (offset) == (typeof(offset))(-1))
 129 /* #define POISON_BULK 0 */
 130
 131 /*
 132  * l_wait_event is a flexible sleeping function, permitting simple caller
 133  * configuration of interrupt and timeout sensitivity along with actions to
 134  * be performed in the event of either exception.
 135  *
 136  * The first form of usage looks like this:
 137  *
 138  * struct l_wait_info lwi = LWI_TIMEOUT_INTR(timeout, timeout_handler,
 139  *                                           intr_handler, callback_data);
 140  * rc = l_wait_event(waitq, condition, &lwi);
 141  *
 142  * l_wait_event() makes the current process wait on 'waitq' until 'condition'
 143  * is TRUE or a "killable" signal (SIGTERM, SIKGILL, SIGINT) is pending.  It
 144  * returns 0 to signify 'condition' is TRUE, but if a signal wakes it before
 145  * 'condition' becomes true, it optionally calls the specified 'intr_handler'
 146  * if not NULL, and returns -EINTR.
 147  *
 148  * If a non-zero timeout is specified, signals are ignored until the timeout
 149  * has expired.  At this time, if 'timeout_handler' is not NULL it is called.
 150  * If it returns FALSE l_wait_event() continues to wait as described above with
 151  * signals enabled.  Otherwise it returns -ETIMEDOUT.
 152  *
 153  * LWI_INTR(intr_handler, callback_data) is shorthand for
 154  * LWI_TIMEOUT_INTR(0, NULL, intr_handler, callback_data)
 155  *
 156  * The second form of usage looks like this:
 157  *
 158  * struct l_wait_info lwi = LWI_TIMEOUT(timeout, timeout_handler);
 159  * rc = l_wait_event(waitq, condition, &lwi);
 160  *
 161  * This form is the same as the first except that it COMPLETELY IGNORES
 162  * SIGNALS.  The caller must therefore beware that if 'timeout' is zero, or if
 163  * 'timeout_handler' is not NULL and returns FALSE, then the ONLY thing that
 164  * can unblock the current process is 'condition' becoming TRUE.
 165  *
 166  * Another form of usage is:
 167  * struct l_wait_info lwi = LWI_TIMEOUT_INTERVAL(timeout, interval,
 168  *                                               timeout_handler);
 169  * rc = l_wait_event(waitq, condition, &lwi);
 170  * This is the same as previous case, but condition is checked once every
 171  * 'interval' jiffies (if non-zero).
 172  *
 173  * Subtle synchronization point: this macro does *not* necessary takes
 174  * wait-queue spin-lock before returning, and, hence, following idiom is safe
 175  * ONLY when caller provides some external locking:
 176  *
 177  *             Thread1                            Thread2
 178  *
 179  *   l_wait_event(&obj->wq, ....);                                       (1)
 180  *
 181  *                                    wake_up(&obj->wq):                 (2)
 182  *                                         spin_lock(&q->lock);          (2.1)
 183  *                                         __wake_up_common(q, ...);     (2.2)
 184  *                                         spin_unlock(&q->lock, flags); (2.3)
 185  *
 186  *   OBD_FREE_PTR(obj);                                                  (3)
 187  *
 188  * As l_wait_event() may "short-cut" execution and return without taking
 189  * wait-queue spin-lock, some additional synchronization is necessary to
 190  * guarantee that step (3) can begin only after (2.3) finishes.
 191  *
 192  * XXX nikita: some ptlrpc daemon threads have races of that sort.
 193  *
 194  */
 195 static inline int back_to_sleep(void *arg)
 196 {
 197         return 0;
 198 }
 199
 200 #define LWI_ON_SIGNAL_NOOP ((void (*)(void *))(-1))
 201
 202 struct l_wait_info {
 203         cfs_duration_t lwi_timeout;
 204         cfs_duration_t lwi_interval;
 205         int            lwi_allow_intr;
 206         int  (*lwi_on_timeout)(void *);
 207         void (*lwi_on_signal)(void *);
 208         void  *lwi_cb_data;
 209 };
 210
 211 /* NB: LWI_TIMEOUT ignores signals completely */
 212 #define LWI_TIMEOUT(time, cb, data)             \
 213 ((struct l_wait_info) {                         \
 214         .lwi_timeout    = time,                 \
 215         .lwi_on_timeout = cb,                   \
 216         .lwi_cb_data    = data,                 \
 217         .lwi_interval   = 0,                    \
 218         .lwi_allow_intr = 0                     \
 219 })
 220
 221 #define LWI_TIMEOUT_INTERVAL(time, interval, cb, data)  \
 222 ((struct l_wait_info) {                                 \
 223         .lwi_timeout    = time,                         \
 224         .lwi_on_timeout = cb,                           \
 225         .lwi_cb_data    = data,                         \
 226         .lwi_interval   = interval,                     \
 227         .lwi_allow_intr = 0                             \
 228 })
 229
 230 #define LWI_TIMEOUT_INTR(time, time_cb, sig_cb, data)   \
 231 ((struct l_wait_info) {                                 \
 232         .lwi_timeout    = time,                         \
 233         .lwi_on_timeout = time_cb,                      \
 234         .lwi_on_signal  = sig_cb,                       \
 235         .lwi_cb_data    = data,                         \
 236         .lwi_interval   = 0,                            \
 237         .lwi_allow_intr = 0                             \
 238 })
 239
 240 #define LWI_TIMEOUT_INTR_ALL(time, time_cb, sig_cb, data)       \
 241 ((struct l_wait_info) {                                         \
 242         .lwi_timeout    = time,                                 \
 243         .lwi_on_timeout = time_cb,                              \
 244         .lwi_on_signal  = sig_cb,                               \
 245         .lwi_cb_data    = data,                                 \
 246         .lwi_interval   = 0,                                    \
 247         .lwi_allow_intr = 1                                     \
 248 })
 249
 250 #define LWI_INTR(cb, data)  LWI_TIMEOUT_INTR(0, NULL, cb, data)
 251
 252 #ifdef __KERNEL__
 253
 254 /*
 255  * wait for @condition to become true, but no longer than timeout, specified
 256  * by @info.
 257  */
 258 #define __l_wait_event(wq, condition, info, ret, l_add_wait)                   \
 259 do {                                                                           \
 260         wait_queue_t __wait;                                                   \
 261         cfs_duration_t __timeout = info->lwi_timeout;                          \
 262         sigset_t   __blocked;                                              \
 263         int   __allow_intr = info->lwi_allow_intr;                             \
 264                                                                                \
 265         ret = 0;                                                               \
 266         if (condition)                                                         \
 267                 break;                                                         \
 268                                                                                \
 269         init_waitqueue_entry_current(&__wait);                                 \
 270         l_add_wait(&wq, &__wait);                                              \
 271                                                                                \
 272         /* Block all signals (just the non-fatal ones if no timeout). */       \
 273         if (info->lwi_on_signal != NULL && (__timeout == 0 || __allow_intr))   \
 274                 __blocked = cfs_block_sigsinv(LUSTRE_FATAL_SIGS);              \
 275         else                                                                   \
 276                 __blocked = cfs_block_sigsinv(0);                              \
 277                                                                                \
 278         for (;;) {                                                             \
 279                 unsigned       __wstate;                                       \
 280                                                                                \
 281                 __wstate = info->lwi_on_signal != NULL &&                      \
 282                            (__timeout == 0 || __allow_intr) ?                  \
 283                         TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;             \
 284                                                                                \
 285                 set_current_state(TASK_INTERRUPTIBLE);                         \
 286                                                                                \
 287                 if (condition)                                                 \
 288                         break;                                                 \
 289                                                                                \
 290                 if (__timeout == 0) {                                          \
 291                         waitq_wait(&__wait, __wstate);                         \
 292                 } else {                                                       \
 293                         cfs_duration_t interval = info->lwi_interval?          \
 294                                              min_t(cfs_duration_t,             \
 295                                                  info->lwi_interval,__timeout):\
 296                                              __timeout;                        \
 297                         cfs_duration_t remaining = waitq_timedwait(&__wait,    \
 298                                                    __wstate,                   \
 299                                                    interval);                  \
 300                         __timeout = cfs_time_sub(__timeout,                    \
 301                                             cfs_time_sub(interval, remaining));\
 302                         if (__timeout == 0) {                                  \
 303                                 if (info->lwi_on_timeout == NULL ||            \
 304                                     info->lwi_on_timeout(info->lwi_cb_data)) { \
 305                                         ret = -ETIMEDOUT;                      \
 306                                         break;                                 \
 307                                 }                                              \
 308                                 /* Take signals after the timeout expires. */  \
 309                                 if (info->lwi_on_signal != NULL)               \
 310                                     (void)cfs_block_sigsinv(LUSTRE_FATAL_SIGS);\
 311                         }                                                      \
 312                 }                                                              \
 313                                                                                \
 314                 if (condition)                                                 \
 315                         break;                                                 \
 316                 if (cfs_signal_pending()) {                                    \
 317                         if (info->lwi_on_signal != NULL &&                     \
 318                             (__timeout == 0 || __allow_intr)) {                \
 319                                 if (info->lwi_on_signal != LWI_ON_SIGNAL_NOOP) \
 320                                         info->lwi_on_signal(info->lwi_cb_data);\
 321                                 ret = -EINTR;                                  \
 322                                 break;                                         \
 323                         }                                                      \
 324                         /* We have to do this here because some signals */     \
 325                         /* are not blockable - ie from strace(1).       */     \
 326                         /* In these cases we want to schedule_timeout() */     \
 327                         /* again, because we don't want that to return  */     \
 328                         /* -EINTR when the RPC actually succeeded.      */     \
 329                         /* the recalc_sigpending() below will deliver the */   \
 330                         /* signal properly.                             */     \
 331                         cfs_clear_sigpending();                                \
 332                 }                                                              \
 333         }                                                                      \
 334                                                                                \
 335         cfs_restore_sigs(__blocked);                                           \
 336                                                                                \
 337         set_current_state(TASK_RUNNING);                                       \
 338         remove_wait_queue(&wq, &__wait);                                       \
 339 } while (0)
 340
 341 #else /* !__KERNEL__ */
 342
 343 #define __l_wait_event(wq, condition, info, ret, l_add_wait)            \
 344 do {                                                                    \
 345         long __timeout = info->lwi_timeout;                             \
 346         long __now;                                                     \
 347         long __then = 0;                                                \
 348         int  __timed_out = 0;                                           \
 349         int  __interval = obd_timeout;                                  \
 350                                                                         \
 351         ret = 0;                                                        \
 352         if (condition)                                                  \
 353                 break;                                                  \
 354                                                                         \
 355         if (__timeout != 0)                                             \
 356                 __then = time(NULL);                                    \
 357                                                                         \
 358         if (__timeout && __timeout < __interval)                        \
 359                 __interval = __timeout;                                 \
 360         if (info->lwi_interval && info->lwi_interval < __interval)      \
 361                 __interval = info->lwi_interval;                        \
 362                                                                         \
 363         while (!(condition)) {                                          \
 364                 liblustre_wait_event(__interval);                       \
 365                 if (condition)                                          \
 366                         break;                                          \
 367                                                                         \
 368                 if (!__timed_out && info->lwi_timeout != 0) {           \
 369                         __now = time(NULL);                             \
 370                         __timeout -= __now - __then;                    \
 371                         __then = __now;                                 \
 372                                                                         \
 373                         if (__timeout > 0)                              \
 374                                 continue;                               \
 375                                                                         \
 376                         __timeout = 0;                                  \
 377                         __timed_out = 1;                                \
 378                         if (info->lwi_on_timeout == NULL ||             \
 379                             info->lwi_on_timeout(info->lwi_cb_data)) {  \
 380                                 ret = -ETIMEDOUT;                       \
 381                                 break;                                  \
 382                         }                                               \
 383                 }                                                       \
 384         }                                                               \
 385 } while (0)
 386
 387 #endif /* __KERNEL__ */
 388
 389
 390 #define l_wait_event(wq, condition, info)                       \
 391 ({                                                              \
 392         int                 __ret;                              \
 393         struct l_wait_info *__info = (info);                    \
 394                                                                 \
 395         __l_wait_event(wq, condition, __info,                   \
 396                        __ret, add_wait_queue);                  \
 397         __ret;                                                  \
 398 })
 399
 400 #define l_wait_event_exclusive(wq, condition, info)             \
 401 ({                                                              \
 402         int                 __ret;                              \
 403         struct l_wait_info *__info = (info);                    \
 404                                                                 \
 405         __l_wait_event(wq, condition, __info,                   \
 406                        __ret, add_wait_queue_exclusive);        \
 407         __ret;                                                  \
 408 })
 409
 410 #define l_wait_event_exclusive_head(wq, condition, info)        \
 411 ({                                                              \
 412         int                 __ret;                              \
 413         struct l_wait_info *__info = (info);                    \
 414                                                                 \
 415         __l_wait_event(wq, condition, __info,                   \
 416                        __ret, add_wait_queue_exclusive_head);   \
 417         __ret;                                                  \
 418 })
 419
 420 #define l_wait_condition(wq, condition)                         \
 421 ({                                                              \
 422         struct l_wait_info lwi = { 0 };                         \
 423         l_wait_event(wq, condition, &lwi);                      \
 424 })
 425
 426 #define l_wait_condition_exclusive(wq, condition)               \
 427 ({                                                              \
 428         struct l_wait_info lwi = { 0 };                         \
 429         l_wait_event_exclusive(wq, condition, &lwi);            \
 430 })
 431
 432 #define l_wait_condition_exclusive_head(wq, condition)          \
 433 ({                                                              \
 434         struct l_wait_info lwi = { 0 };                         \
 435         l_wait_event_exclusive_head(wq, condition, &lwi);       \
 436 })
 437
 438 #ifdef __KERNEL__
 439 #define LIBLUSTRE_CLIENT (0)
 440 #else
 441 #define LIBLUSTRE_CLIENT (1)
 442 #endif
 443
 444 /** @} lib */
 445
 446 #endif /* _LUSTRE_LIB_H */