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