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, 2014, 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 <linux/signal.h>
  50 #include <libcfs/libcfs.h>
  51 #include <lustre/lustre_idl.h>
  52 #include <lustre_ver.h>
  53 #include <lustre_cfg.h>
  54
  55 /* target.c */
  56 struct ptlrpc_request;
  57 struct obd_export;
  58 struct lu_target;
  59 struct l_wait_info;
  60 #include <lustre_ha.h>
  61 #include <lustre_net.h>
  62
  63 #ifdef HAVE_SERVER_SUPPORT
  64 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
  65                           int error);
  66 int rev_import_init(struct obd_export *exp);
  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                       size_t keylen, void *key,
  85                       size_t vallen, void *val,
  86                       struct ptlrpc_request_set *set);
  87
  88 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id);
  89
  90 /*
  91  * l_wait_event is a flexible sleeping function, permitting simple caller
  92  * configuration of interrupt and timeout sensitivity along with actions to
  93  * be performed in the event of either exception.
  94  *
  95  * The first form of usage looks like this:
  96  *
  97  * struct l_wait_info lwi = LWI_TIMEOUT_INTR(timeout, timeout_handler,
  98  *                                           intr_handler, callback_data);
  99  * rc = l_wait_event(waitq, condition, &lwi);
 100  *
 101  * l_wait_event() makes the current process wait on 'waitq' until 'condition'
 102  * is TRUE or a "killable" signal (SIGTERM, SIKGILL, SIGINT) is pending.  It
 103  * returns 0 to signify 'condition' is TRUE, but if a signal wakes it before
 104  * 'condition' becomes true, it optionally calls the specified 'intr_handler'
 105  * if not NULL, and returns -EINTR.
 106  *
 107  * If a non-zero timeout is specified, signals are ignored until the timeout
 108  * has expired.  At this time, if 'timeout_handler' is not NULL it is called.
 109  * If it returns FALSE l_wait_event() continues to wait as described above with
 110  * signals enabled.  Otherwise it returns -ETIMEDOUT.
 111  *
 112  * LWI_INTR(intr_handler, callback_data) is shorthand for
 113  * LWI_TIMEOUT_INTR(0, NULL, intr_handler, callback_data)
 114  *
 115  * The second form of usage looks like this:
 116  *
 117  * struct l_wait_info lwi = LWI_TIMEOUT(timeout, timeout_handler);
 118  * rc = l_wait_event(waitq, condition, &lwi);
 119  *
 120  * This form is the same as the first except that it COMPLETELY IGNORES
 121  * SIGNALS.  The caller must therefore beware that if 'timeout' is zero, or if
 122  * 'timeout_handler' is not NULL and returns FALSE, then the ONLY thing that
 123  * can unblock the current process is 'condition' becoming TRUE.
 124  *
 125  * Another form of usage is:
 126  * struct l_wait_info lwi = LWI_TIMEOUT_INTERVAL(timeout, interval,
 127  *                                               timeout_handler);
 128  * rc = l_wait_event(waitq, condition, &lwi);
 129  * This is the same as previous case, but condition is checked once every
 130  * 'interval' jiffies (if non-zero).
 131  *
 132  * Subtle synchronization point: this macro does *not* necessary takes
 133  * wait-queue spin-lock before returning, and, hence, following idiom is safe
 134  * ONLY when caller provides some external locking:
 135  *
 136  *             Thread1                            Thread2
 137  *
 138  *   l_wait_event(&obj->wq, ....);                                       (1)
 139  *
 140  *                                    wake_up(&obj->wq):                 (2)
 141  *                                         spin_lock(&q->lock);          (2.1)
 142  *                                         __wake_up_common(q, ...);     (2.2)
 143  *                                         spin_unlock(&q->lock, flags); (2.3)
 144  *
 145  *   OBD_FREE_PTR(obj);                                                  (3)
 146  *
 147  * As l_wait_event() may "short-cut" execution and return without taking
 148  * wait-queue spin-lock, some additional synchronization is necessary to
 149  * guarantee that step (3) can begin only after (2.3) finishes.
 150  *
 151  * XXX nikita: some ptlrpc daemon threads have races of that sort.
 152  *
 153  */
 154 static inline int back_to_sleep(void *arg)
 155 {
 156         return 0;
 157 }
 158
 159 #define LWI_ON_SIGNAL_NOOP ((void (*)(void *))(-1))
 160
 161 struct l_wait_info {
 162         cfs_duration_t lwi_timeout;
 163         cfs_duration_t lwi_interval;
 164         int            lwi_allow_intr;
 165         int  (*lwi_on_timeout)(void *);
 166         void (*lwi_on_signal)(void *);
 167         void  *lwi_cb_data;
 168 };
 169
 170 /* NB: LWI_TIMEOUT ignores signals completely */
 171 #define LWI_TIMEOUT(time, cb, data)             \
 172 ((struct l_wait_info) {                         \
 173         .lwi_timeout    = time,                 \
 174         .lwi_on_timeout = cb,                   \
 175         .lwi_cb_data    = data,                 \
 176         .lwi_interval   = 0,                    \
 177         .lwi_allow_intr = 0                     \
 178 })
 179
 180 #define LWI_TIMEOUT_INTERVAL(time, interval, cb, data)  \
 181 ((struct l_wait_info) {                                 \
 182         .lwi_timeout    = time,                         \
 183         .lwi_on_timeout = cb,                           \
 184         .lwi_cb_data    = data,                         \
 185         .lwi_interval   = interval,                     \
 186         .lwi_allow_intr = 0                             \
 187 })
 188
 189 #define LWI_TIMEOUT_INTR(time, time_cb, sig_cb, data)   \
 190 ((struct l_wait_info) {                                 \
 191         .lwi_timeout    = time,                         \
 192         .lwi_on_timeout = time_cb,                      \
 193         .lwi_on_signal  = sig_cb,                       \
 194         .lwi_cb_data    = data,                         \
 195         .lwi_interval   = 0,                            \
 196         .lwi_allow_intr = 0                             \
 197 })
 198
 199 #define LWI_TIMEOUT_INTR_ALL(time, time_cb, sig_cb, data)       \
 200 ((struct l_wait_info) {                                         \
 201         .lwi_timeout    = time,                                 \
 202         .lwi_on_timeout = time_cb,                              \
 203         .lwi_on_signal  = sig_cb,                               \
 204         .lwi_cb_data    = data,                                 \
 205         .lwi_interval   = 0,                                    \
 206         .lwi_allow_intr = 1                                     \
 207 })
 208
 209 #define LWI_INTR(cb, data)  LWI_TIMEOUT_INTR(0, NULL, cb, data)
 210
 211 #define LUSTRE_FATAL_SIGS                                        \
 212         (sigmask(SIGKILL) | sigmask(SIGINT) | sigmask(SIGTERM) | \
 213          sigmask(SIGQUIT) | sigmask(SIGALRM))
 214
 215 /*
 216  * Wait Queue
 217  */
 218 #ifndef HAVE___ADD_WAIT_QUEUE_EXCLUSIVE
 219 static inline void __add_wait_queue_exclusive(wait_queue_head_t *q,
 220                                               wait_queue_t *wait)
 221 {
 222         wait->flags |= WQ_FLAG_EXCLUSIVE;
 223         __add_wait_queue(q, wait);
 224 }
 225 #endif /* HAVE___ADD_WAIT_QUEUE_EXCLUSIVE */
 226
 227 /**
 228  * wait_queue_t of Linux (version < 2.6.34) is a FIFO list for exclusively
 229  * waiting threads, which is not always desirable because all threads will
 230  * be waken up again and again, even user only needs a few of them to be
 231  * active most time. This is not good for performance because cache can
 232  * be polluted by different threads.
 233  *
 234  * LIFO list can resolve this problem because we always wakeup the most
 235  * recent active thread by default.
 236  *
 237  * NB: please don't call non-exclusive & exclusive wait on the same
 238  * waitq if add_wait_queue_exclusive_head is used.
 239  */
 240 #define add_wait_queue_exclusive_head(waitq, link)              \
 241 {                                                               \
 242         unsigned long flags;                                    \
 243                                                                 \
 244         spin_lock_irqsave(&((waitq)->lock), flags);             \
 245         __add_wait_queue_exclusive(waitq, link);                \
 246         spin_unlock_irqrestore(&((waitq)->lock), flags);        \
 247 }
 248
 249 /*
 250  * wait for @condition to become true, but no longer than timeout, specified
 251  * by @info.
 252  */
 253 #define __l_wait_event(wq, condition, info, ret, l_add_wait)                   \
 254 do {                                                                           \
 255         wait_queue_t __wait;                                                   \
 256         cfs_duration_t __timeout = info->lwi_timeout;                          \
 257         sigset_t   __blocked;                                              \
 258         int   __allow_intr = info->lwi_allow_intr;                             \
 259                                                                                \
 260         ret = 0;                                                               \
 261         if (condition)                                                         \
 262                 break;                                                         \
 263                                                                                \
 264         init_waitqueue_entry(&__wait, current);                                \
 265         l_add_wait(&wq, &__wait);                                              \
 266                                                                                \
 267         /* Block all signals (just the non-fatal ones if no timeout). */       \
 268         if (info->lwi_on_signal != NULL && (__timeout == 0 || __allow_intr))   \
 269                 __blocked = cfs_block_sigsinv(LUSTRE_FATAL_SIGS);              \
 270         else                                                                   \
 271                 __blocked = cfs_block_sigsinv(0);                              \
 272                                                                                \
 273         for (;;) {                                                             \
 274                 set_current_state(TASK_INTERRUPTIBLE);                         \
 275                                                                                \
 276                 if (condition)                                                 \
 277                         break;                                                 \
 278                                                                                \
 279                 if (__timeout == 0) {                                          \
 280                         schedule();                                            \
 281                 } else {                                                       \
 282                         cfs_duration_t interval = info->lwi_interval?          \
 283                                              min_t(cfs_duration_t,             \
 284                                                  info->lwi_interval,__timeout):\
 285                                              __timeout;                        \
 286                         cfs_duration_t remaining = schedule_timeout(interval); \
 287                         __timeout = cfs_time_sub(__timeout,                    \
 288                                             cfs_time_sub(interval, remaining));\
 289                         if (__timeout == 0) {                                  \
 290                                 if (info->lwi_on_timeout == NULL ||            \
 291                                     info->lwi_on_timeout(info->lwi_cb_data)) { \
 292                                         ret = -ETIMEDOUT;                      \
 293                                         break;                                 \
 294                                 }                                              \
 295                                 /* Take signals after the timeout expires. */  \
 296                                 if (info->lwi_on_signal != NULL)               \
 297                                     (void)cfs_block_sigsinv(LUSTRE_FATAL_SIGS);\
 298                         }                                                      \
 299                 }                                                              \
 300                                                                                \
 301                 if (condition)                                                 \
 302                         break;                                                 \
 303                 if (signal_pending(current)) {                                 \
 304                         if (info->lwi_on_signal != NULL &&                     \
 305                             (__timeout == 0 || __allow_intr)) {                \
 306                                 if (info->lwi_on_signal != LWI_ON_SIGNAL_NOOP) \
 307                                         info->lwi_on_signal(info->lwi_cb_data);\
 308                                 ret = -EINTR;                                  \
 309                                 break;                                         \
 310                         }                                                      \
 311                         /* We have to do this here because some signals */     \
 312                         /* are not blockable - ie from strace(1).       */     \
 313                         /* In these cases we want to schedule_timeout() */     \
 314                         /* again, because we don't want that to return  */     \
 315                         /* -EINTR when the RPC actually succeeded.      */     \
 316                         /* the recalc_sigpending() below will deliver the */   \
 317                         /* signal properly.                             */     \
 318                         cfs_clear_sigpending();                                \
 319                 }                                                              \
 320         }                                                                      \
 321                                                                                \
 322         cfs_restore_sigs(__blocked);                                           \
 323                                                                                \
 324         set_current_state(TASK_RUNNING);                                       \
 325         remove_wait_queue(&wq, &__wait);                                       \
 326 } while (0)
 327
 328
 329 #define l_wait_event(wq, condition, info)                       \
 330 ({                                                              \
 331         int                 __ret;                              \
 332         struct l_wait_info *__info = (info);                    \
 333                                                                 \
 334         __l_wait_event(wq, condition, __info,                   \
 335                        __ret, add_wait_queue);                  \
 336         __ret;                                                  \
 337 })
 338
 339 #define l_wait_event_exclusive(wq, condition, info)             \
 340 ({                                                              \
 341         int                 __ret;                              \
 342         struct l_wait_info *__info = (info);                    \
 343                                                                 \
 344         __l_wait_event(wq, condition, __info,                   \
 345                        __ret, add_wait_queue_exclusive);        \
 346         __ret;                                                  \
 347 })
 348
 349 #define l_wait_event_exclusive_head(wq, condition, info)        \
 350 ({                                                              \
 351         int                 __ret;                              \
 352         struct l_wait_info *__info = (info);                    \
 353                                                                 \
 354         __l_wait_event(wq, condition, __info,                   \
 355                        __ret, add_wait_queue_exclusive_head);   \
 356         __ret;                                                  \
 357 })
 358
 359 #define l_wait_condition(wq, condition)                         \
 360 ({                                                              \
 361         struct l_wait_info lwi = { 0 };                         \
 362         l_wait_event(wq, condition, &lwi);                      \
 363 })
 364
 365 #define l_wait_condition_exclusive(wq, condition)               \
 366 ({                                                              \
 367         struct l_wait_info lwi = { 0 };                         \
 368         l_wait_event_exclusive(wq, condition, &lwi);            \
 369 })
 370
 371 #define l_wait_condition_exclusive_head(wq, condition)          \
 372 ({                                                              \
 373         struct l_wait_info lwi = { 0 };                         \
 374         l_wait_event_exclusive_head(wq, condition, &lwi);       \
 375 })
 376
 377 /** @} lib */
 378
 379 #endif /* _LUSTRE_LIB_H */