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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2014, 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  * lustre/include/lustre_lib.h
  33  *
  34  * Basic Lustre library routines.
  35  */
  36
  37 #ifndef _LUSTRE_LIB_H
  38 #define _LUSTRE_LIB_H
  39
  40 /** \defgroup lib lib
  41  *
  42  * @{
  43  */
  44
  45 #include <libcfs/libcfs.h>
  46 #include <lustre/lustre_idl.h>
  47 #include <lustre_ver.h>
  48 #include <uapi/linux/lustre_cfg.h>
  49
  50 /* target.c */
  51 struct ptlrpc_request;
  52 struct obd_export;
  53 struct lu_target;
  54 struct l_wait_info;
  55 #include <lustre_ha.h>
  56 #include <lustre_net.h>
  57
  58 #define LI_POISON 0x5a5a5a5a
  59 #if BITS_PER_LONG > 32
  60 # define LL_POISON 0x5a5a5a5a5a5a5a5aL
  61 #else
  62 # define LL_POISON 0x5a5a5a5aL
  63 #endif
  64 #define LP_POISON ((void *)LL_POISON)
  65
  66 #ifdef HAVE_SERVER_SUPPORT
  67 int rev_import_init(struct obd_export *exp);
  68 int target_handle_connect(struct ptlrpc_request *req);
  69 int target_handle_disconnect(struct ptlrpc_request *req);
  70 void target_destroy_export(struct obd_export *exp);
  71 int target_handle_ping(struct ptlrpc_request *req);
  72 void target_committed_to_req(struct ptlrpc_request *req);
  73 void target_cancel_recovery_timer(struct obd_device *obd);
  74 void target_stop_recovery_thread(struct obd_device *obd);
  75 void target_cleanup_recovery(struct obd_device *obd);
  76 int target_queue_recovery_request(struct ptlrpc_request *req,
  77                                   struct obd_device *obd);
  78 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
  79                    struct l_wait_info *lwi);
  80 #endif
  81
  82 int target_pack_pool_reply(struct ptlrpc_request *req);
  83 int do_set_info_async(struct obd_import *imp,
  84                       int opcode, int version,
  85                       size_t keylen, void *key,
  86                       size_t vallen, void *val,
  87                       struct ptlrpc_request_set *set);
  88
  89 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id);
  90
  91 /*
  92  * l_wait_event is a flexible sleeping function, permitting simple caller
  93  * configuration of interrupt and timeout sensitivity along with actions to
  94  * be performed in the event of either exception.
  95  *
  96  * The first form of usage looks like this:
  97  *
  98  * struct l_wait_info lwi = LWI_TIMEOUT_INTR(timeout, timeout_handler,
  99  *                                           intr_handler, callback_data);
 100  * rc = l_wait_event(waitq, condition, &lwi);
 101  *
 102  * l_wait_event() makes the current process wait on 'waitq' until 'condition'
 103  * is TRUE or a "killable" signal (SIGTERM, SIKGILL, SIGINT) is pending.  It
 104  * returns 0 to signify 'condition' is TRUE, but if a signal wakes it before
 105  * 'condition' becomes true, it optionally calls the specified 'intr_handler'
 106  * if not NULL, and returns -EINTR.
 107  *
 108  * If a non-zero timeout is specified, signals are ignored until the timeout
 109  * has expired.  At this time, if 'timeout_handler' is not NULL it is called.
 110  * If it returns FALSE l_wait_event() continues to wait as described above with
 111  * signals enabled.  Otherwise it returns -ETIMEDOUT.
 112  *
 113  * LWI_INTR(intr_handler, callback_data) is shorthand for
 114  * LWI_TIMEOUT_INTR(0, NULL, intr_handler, callback_data)
 115  *
 116  * The second form of usage looks like this:
 117  *
 118  * struct l_wait_info lwi = LWI_TIMEOUT(timeout, timeout_handler);
 119  * rc = l_wait_event(waitq, condition, &lwi);
 120  *
 121  * This form is the same as the first except that it COMPLETELY IGNORES
 122  * SIGNALS.  The caller must therefore beware that if 'timeout' is zero, or if
 123  * 'timeout_handler' is not NULL and returns FALSE, then the ONLY thing that
 124  * can unblock the current process is 'condition' becoming TRUE.
 125  *
 126  * Another form of usage is:
 127  * struct l_wait_info lwi = LWI_TIMEOUT_INTERVAL(timeout, interval,
 128  *                                               timeout_handler);
 129  * rc = l_wait_event(waitq, condition, &lwi);
 130  * This is the same as previous case, but condition is checked once every
 131  * 'interval' jiffies (if non-zero).
 132  *
 133  * Subtle synchronization point: this macro does *not* necessary takes
 134  * wait-queue spin-lock before returning, and, hence, following idiom is safe
 135  * ONLY when caller provides some external locking:
 136  *
 137  *             Thread1                            Thread2
 138  *
 139  *   l_wait_event(&obj->wq, ....);                                       (1)
 140  *
 141  *                                    wake_up(&obj->wq):                 (2)
 142  *                                         spin_lock(&q->lock);          (2.1)
 143  *                                         __wake_up_common(q, ...);     (2.2)
 144  *                                         spin_unlock(&q->lock, flags); (2.3)
 145  *
 146  *   OBD_FREE_PTR(obj);                                                  (3)
 147  *
 148  * As l_wait_event() may "short-cut" execution and return without taking
 149  * wait-queue spin-lock, some additional synchronization is necessary to
 150  * guarantee that step (3) can begin only after (2.3) finishes.
 151  *
 152  * XXX nikita: some ptlrpc daemon threads have races of that sort.
 153  *
 154  */
 155 static inline int back_to_sleep(void *arg)
 156 {
 157         return 0;
 158 }
 159
 160 #define LWI_ON_SIGNAL_NOOP ((void (*)(void *))(-1))
 161
 162 struct l_wait_info {
 163         cfs_duration_t lwi_timeout;
 164         cfs_duration_t lwi_interval;
 165         int            lwi_allow_intr;
 166         int  (*lwi_on_timeout)(void *);
 167         void (*lwi_on_signal)(void *);
 168         void  *lwi_cb_data;
 169 };
 170
 171 /* NB: LWI_TIMEOUT ignores signals completely */
 172 #define LWI_TIMEOUT(time, cb, data)             \
 173 ((struct l_wait_info) {                         \
 174         .lwi_timeout    = time,                 \
 175         .lwi_on_timeout = cb,                   \
 176         .lwi_cb_data    = data,                 \
 177         .lwi_interval   = 0,                    \
 178         .lwi_allow_intr = 0                     \
 179 })
 180
 181 #define LWI_TIMEOUT_INTERVAL(time, interval, cb, data)  \
 182 ((struct l_wait_info) {                                 \
 183         .lwi_timeout    = time,                         \
 184         .lwi_on_timeout = cb,                           \
 185         .lwi_cb_data    = data,                         \
 186         .lwi_interval   = interval,                     \
 187         .lwi_allow_intr = 0                             \
 188 })
 189
 190 #define LWI_TIMEOUT_INTR(time, time_cb, sig_cb, data)   \
 191 ((struct l_wait_info) {                                 \
 192         .lwi_timeout    = time,                         \
 193         .lwi_on_timeout = time_cb,                      \
 194         .lwi_on_signal  = sig_cb,                       \
 195         .lwi_cb_data    = data,                         \
 196         .lwi_interval   = 0,                            \
 197         .lwi_allow_intr = 0                             \
 198 })
 199
 200 #define LWI_TIMEOUT_INTR_ALL(time, time_cb, sig_cb, data)       \
 201 ((struct l_wait_info) {                                         \
 202         .lwi_timeout    = time,                                 \
 203         .lwi_on_timeout = time_cb,                              \
 204         .lwi_on_signal  = sig_cb,                               \
 205         .lwi_cb_data    = data,                                 \
 206         .lwi_interval   = 0,                                    \
 207         .lwi_allow_intr = 1                                     \
 208 })
 209
 210 #define LWI_INTR(cb, data)  LWI_TIMEOUT_INTR(0, NULL, cb, data)
 211
 212 #define LUSTRE_FATAL_SIGS                                        \
 213         (sigmask(SIGKILL) | sigmask(SIGINT) | sigmask(SIGTERM) | \
 214          sigmask(SIGQUIT) | sigmask(SIGALRM))
 215
 216 /*
 217  * Wait Queue
 218  */
 219 #ifndef HAVE___ADD_WAIT_QUEUE_EXCLUSIVE
 220 static inline void __add_wait_queue_exclusive(wait_queue_head_t *q,
 221                                               wait_queue_t *wait)
 222 {
 223         wait->flags |= WQ_FLAG_EXCLUSIVE;
 224         __add_wait_queue(q, wait);
 225 }
 226 #endif /* HAVE___ADD_WAIT_QUEUE_EXCLUSIVE */
 227
 228 /**
 229  * wait_queue_t of Linux (version < 2.6.34) is a FIFO list for exclusively
 230  * waiting threads, which is not always desirable because all threads will
 231  * be waken up again and again, even user only needs a few of them to be
 232  * active most time. This is not good for performance because cache can
 233  * be polluted by different threads.
 234  *
 235  * LIFO list can resolve this problem because we always wakeup the most
 236  * recent active thread by default.
 237  *
 238  * NB: please don't call non-exclusive & exclusive wait on the same
 239  * waitq if add_wait_queue_exclusive_head is used.
 240  */
 241 #define add_wait_queue_exclusive_head(waitq, link)              \
 242 {                                                               \
 243         unsigned long flags;                                    \
 244                                                                 \
 245         spin_lock_irqsave(&((waitq)->lock), flags);             \
 246         __add_wait_queue_exclusive(waitq, link);                \
 247         spin_unlock_irqrestore(&((waitq)->lock), flags);        \
 248 }
 249
 250 /*
 251  * wait for @condition to become true, but no longer than timeout, specified
 252  * by @info.
 253  */
 254 #define __l_wait_event(wq, condition, info, ret, l_add_wait)                   \
 255 do {                                                                           \
 256         wait_queue_t __wait;                                                   \
 257         cfs_duration_t __timeout = info->lwi_timeout;                          \
 258         sigset_t   __blocked;                                              \
 259         int   __allow_intr = info->lwi_allow_intr;                             \
 260                                                                                \
 261         ret = 0;                                                               \
 262         if (condition)                                                         \
 263                 break;                                                         \
 264                                                                                \
 265         init_waitqueue_entry(&__wait, current);                                \
 266         l_add_wait(&wq, &__wait);                                              \
 267                                                                                \
 268         /* Block all signals (just the non-fatal ones if no timeout). */       \
 269         if (info->lwi_on_signal != NULL && (__timeout == 0 || __allow_intr))   \
 270                 __blocked = cfs_block_sigsinv(LUSTRE_FATAL_SIGS);              \
 271         else                                                                   \
 272                 __blocked = cfs_block_sigsinv(0);                              \
 273                                                                                \
 274         for (;;) {                                                             \
 275                 set_current_state(TASK_INTERRUPTIBLE);                         \
 276                                                                                \
 277                 /* To guarantee that the condition check will be done */       \
 278                 /* after setting the thread state as TASK_INTERRUPTIBLE. */    \
 279                 /* Otherwise, out-of-order execution may cause some race. */   \
 280                 /* Consider the following real execution order: */             \
 281                                                                                \
 282                 /* 1. Thread1 checks condition on CPU1, gets false. */         \
 283                 /* 2. Thread2 sets condition on CPU2. */                       \
 284                 /* 3. Thread2 calls wake_up() on CPU2 to wake the threads */   \
 285                 /*    with state TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE. */ \
 286                 /*    But the Thread1's state is TASK_RUNNING at that time. */ \
 287                 /* 4. Thread1 sets its state as TASK_INTERRUPTIBLE on CPU1, */ \
 288                 /*    then schedule. */                                        \
 289                                                                                \
 290                 /* If the '__timeout' variable is zero, the Thread1 will */    \
 291                 /* have no chance to check the condition again. */             \
 292                                                                                \
 293                 /* Generally, the interval between out-of-ordered step1 and */ \
 294                 /* step4 is very tiny, as to above step2 and step3 cannot */   \
 295                 /* happen. On some degree, it can explain why we seldom hit */ \
 296                 /* related trouble. But such race really exists, especially */ \
 297                 /* consider that the step1 and step4 can be interruptible. */  \
 298                 /* So add barrier to avoid Thread1 out-of-order execution. */  \
 299                 smp_mb();                                                      \
 300                                                                                \
 301                 if (condition)                                                 \
 302                         break;                                                 \
 303                                                                                \
 304                 if (__timeout == 0) {                                          \
 305                         schedule();                                            \
 306                 } else {                                                       \
 307                         cfs_duration_t interval = info->lwi_interval?          \
 308                                              min_t(cfs_duration_t,             \
 309                                                  info->lwi_interval,__timeout):\
 310                                              __timeout;                        \
 311                         cfs_duration_t remaining = schedule_timeout(interval); \
 312                         __timeout = cfs_time_sub(__timeout,                    \
 313                                             cfs_time_sub(interval, remaining));\
 314                         if (__timeout == 0) {                                  \
 315                                 if (info->lwi_on_timeout == NULL ||            \
 316                                     info->lwi_on_timeout(info->lwi_cb_data)) { \
 317                                         ret = -ETIMEDOUT;                      \
 318                                         break;                                 \
 319                                 }                                              \
 320                                 /* Take signals after the timeout expires. */  \
 321                                 if (info->lwi_on_signal != NULL)               \
 322                                     (void)cfs_block_sigsinv(LUSTRE_FATAL_SIGS);\
 323                         }                                                      \
 324                 }                                                              \
 325                                                                                \
 326                 if (condition)                                                 \
 327                         break;                                                 \
 328                 if (signal_pending(current)) {                                 \
 329                         if (info->lwi_on_signal != NULL &&                     \
 330                             (__timeout == 0 || __allow_intr)) {                \
 331                                 if (info->lwi_on_signal != LWI_ON_SIGNAL_NOOP) \
 332                                         info->lwi_on_signal(info->lwi_cb_data);\
 333                                 ret = -EINTR;                                  \
 334                                 break;                                         \
 335                         }                                                      \
 336                         /* We have to do this here because some signals */     \
 337                         /* are not blockable - ie from strace(1).       */     \
 338                         /* In these cases we want to schedule_timeout() */     \
 339                         /* again, because we don't want that to return  */     \
 340                         /* -EINTR when the RPC actually succeeded.      */     \
 341                         /* the recalc_sigpending() below will deliver the */   \
 342                         /* signal properly.                             */     \
 343                         cfs_clear_sigpending();                                \
 344                 }                                                              \
 345         }                                                                      \
 346                                                                                \
 347         cfs_restore_sigs(__blocked);                                           \
 348                                                                                \
 349         set_current_state(TASK_RUNNING);                                       \
 350         remove_wait_queue(&wq, &__wait);                                       \
 351 } while (0)
 352
 353
 354 #define l_wait_event(wq, condition, info)                       \
 355 ({                                                              \
 356         int                 __ret;                              \
 357         struct l_wait_info *__info = (info);                    \
 358                                                                 \
 359         __l_wait_event(wq, condition, __info,                   \
 360                        __ret, add_wait_queue);                  \
 361         __ret;                                                  \
 362 })
 363
 364 #define l_wait_event_exclusive(wq, condition, info)             \
 365 ({                                                              \
 366         int                 __ret;                              \
 367         struct l_wait_info *__info = (info);                    \
 368                                                                 \
 369         __l_wait_event(wq, condition, __info,                   \
 370                        __ret, add_wait_queue_exclusive);        \
 371         __ret;                                                  \
 372 })
 373
 374 #define l_wait_event_exclusive_head(wq, condition, info)        \
 375 ({                                                              \
 376         int                 __ret;                              \
 377         struct l_wait_info *__info = (info);                    \
 378                                                                 \
 379         __l_wait_event(wq, condition, __info,                   \
 380                        __ret, add_wait_queue_exclusive_head);   \
 381         __ret;                                                  \
 382 })
 383
 384 #define l_wait_condition(wq, condition)                         \
 385 ({                                                              \
 386         struct l_wait_info lwi = { 0 };                         \
 387         l_wait_event(wq, condition, &lwi);                      \
 388 })
 389
 390 #define l_wait_condition_exclusive(wq, condition)               \
 391 ({                                                              \
 392         struct l_wait_info lwi = { 0 };                         \
 393         l_wait_event_exclusive(wq, condition, &lwi);            \
 394 })
 395
 396 #define l_wait_condition_exclusive_head(wq, condition)          \
 397 ({                                                              \
 398         struct l_wait_info lwi = { 0 };                         \
 399         l_wait_event_exclusive_head(wq, condition, &lwi);       \
 400 })
 401
 402 /** @} lib */
 403
 404 #endif /* _LUSTRE_LIB_H */