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