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