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, 2017, 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 #ifdef HAVE_SCHED_HEADERS
  46 #include <linux/sched/signal.h>
  47 #include <linux/sched/mm.h>
  48 #endif
  49
  50 #include <libcfs/libcfs.h>
  51 #include <uapi/linux/lustre/lustre_idl.h>
  52 #include <uapi/linux/lustre/lustre_ver.h>
  53 #include <uapi/linux/lustre/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 int rev_import_init(struct obd_export *exp);
  73 int target_handle_connect(struct ptlrpc_request *req);
  74 int target_handle_disconnect(struct ptlrpc_request *req);
  75 void target_destroy_export(struct obd_export *exp);
  76 int target_handle_ping(struct ptlrpc_request *req);
  77 void target_committed_to_req(struct ptlrpc_request *req);
  78 void target_cancel_recovery_timer(struct obd_device *obd);
  79 void target_stop_recovery_thread(struct obd_device *obd);
  80 void target_cleanup_recovery(struct obd_device *obd);
  81 int target_queue_recovery_request(struct ptlrpc_request *req,
  82                                   struct obd_device *obd);
  83 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
  84                    struct l_wait_info *lwi);
  85 #endif
  86
  87 int target_pack_pool_reply(struct ptlrpc_request *req);
  88 int do_set_info_async(struct obd_import *imp,
  89                       int opcode, int version,
  90                       size_t keylen, void *key,
  91                       size_t vallen, void *val,
  92                       struct ptlrpc_request_set *set);
  93
  94 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id);
  95
  96 /*
  97  * l_wait_event is a flexible sleeping function, permitting simple caller
  98  * configuration of interrupt and timeout sensitivity along with actions to
  99  * be performed in the event of either exception.
 100  *
 101  * The first form of usage looks like this:
 102  *
 103  * struct l_wait_info lwi = LWI_TIMEOUT_INTR(timeout, timeout_handler,
 104  *                                           intr_handler, callback_data);
 105  * rc = l_wait_event(waitq, condition, &lwi);
 106  *
 107  * l_wait_event() makes the current process wait on 'waitq' until 'condition'
 108  * is TRUE or a "killable" signal (SIGTERM, SIKGILL, SIGINT) is pending.  It
 109  * returns 0 to signify 'condition' is TRUE, but if a signal wakes it before
 110  * 'condition' becomes true, it optionally calls the specified 'intr_handler'
 111  * if not NULL, and returns -EINTR.
 112  *
 113  * If a non-zero timeout is specified, signals are ignored until the timeout
 114  * has expired.  At this time, if 'timeout_handler' is not NULL it is called.
 115  * If it returns FALSE l_wait_event() continues to wait as described above with
 116  * signals enabled.  Otherwise it returns -ETIMEDOUT.
 117  *
 118  * LWI_INTR(intr_handler, callback_data) is shorthand for
 119  * LWI_TIMEOUT_INTR(0, NULL, intr_handler, callback_data)
 120  *
 121  * The second form of usage looks like this:
 122  *
 123  * struct l_wait_info lwi = LWI_TIMEOUT(timeout, timeout_handler);
 124  * rc = l_wait_event(waitq, condition, &lwi);
 125  *
 126  * This form is the same as the first except that it COMPLETELY IGNORES
 127  * SIGNALS.  The caller must therefore beware that if 'timeout' is zero, or if
 128  * 'timeout_handler' is not NULL and returns FALSE, then the ONLY thing that
 129  * can unblock the current process is 'condition' becoming TRUE.
 130  *
 131  * Another form of usage is:
 132  * struct l_wait_info lwi = LWI_TIMEOUT_INTERVAL(timeout, interval,
 133  *                                               timeout_handler);
 134  * rc = l_wait_event(waitq, condition, &lwi);
 135  * This is the same as previous case, but condition is checked once every
 136  * 'interval' jiffies (if non-zero).
 137  *
 138  * Subtle synchronization point: this macro does *not* necessary takes
 139  * wait-queue spin-lock before returning, and, hence, following idiom is safe
 140  * ONLY when caller provides some external locking:
 141  *
 142  *             Thread1                            Thread2
 143  *
 144  *   l_wait_event(&obj->wq, ....);                                       (1)
 145  *
 146  *                                    wake_up(&obj->wq):                 (2)
 147  *                                         spin_lock(&q->lock);          (2.1)
 148  *                                         __wake_up_common(q, ...);     (2.2)
 149  *                                         spin_unlock(&q->lock, flags); (2.3)
 150  *
 151  *   OBD_FREE_PTR(obj);                                                  (3)
 152  *
 153  * As l_wait_event() may "short-cut" execution and return without taking
 154  * wait-queue spin-lock, some additional synchronization is necessary to
 155  * guarantee that step (3) can begin only after (2.3) finishes.
 156  *
 157  * XXX nikita: some ptlrpc daemon threads have races of that sort.
 158  *
 159  */
 160 static inline int back_to_sleep(void *arg)
 161 {
 162         return 0;
 163 }
 164
 165 #define LWI_ON_SIGNAL_NOOP ((void (*)(void *))(-1))
 166
 167 struct l_wait_info {
 168         long            lwi_timeout;
 169         long            lwi_interval;
 170         int             lwi_allow_intr;
 171         int  (*lwi_on_timeout)(void *);
 172         void (*lwi_on_signal)(void *);
 173         void  *lwi_cb_data;
 174 };
 175
 176 /* NB: LWI_TIMEOUT ignores signals completely */
 177 #define LWI_TIMEOUT(time, cb, data)             \
 178 ((struct l_wait_info) {                         \
 179         .lwi_timeout    = time,                 \
 180         .lwi_on_timeout = cb,                   \
 181         .lwi_cb_data    = data,                 \
 182         .lwi_interval   = 0,                    \
 183         .lwi_allow_intr = 0                     \
 184 })
 185
 186 #define LWI_TIMEOUT_INTERVAL(time, interval, cb, data)  \
 187 ((struct l_wait_info) {                                 \
 188         .lwi_timeout    = time,                         \
 189         .lwi_on_timeout = cb,                           \
 190         .lwi_cb_data    = data,                         \
 191         .lwi_interval   = interval,                     \
 192         .lwi_allow_intr = 0                             \
 193 })
 194
 195 #define LWI_TIMEOUT_INTR(time, time_cb, sig_cb, data)   \
 196 ((struct l_wait_info) {                                 \
 197         .lwi_timeout    = time,                         \
 198         .lwi_on_timeout = time_cb,                      \
 199         .lwi_on_signal  = sig_cb,                       \
 200         .lwi_cb_data    = data,                         \
 201         .lwi_interval   = 0,                            \
 202         .lwi_allow_intr = 0                             \
 203 })
 204
 205 #define LWI_TIMEOUT_INTR_ALL(time, time_cb, sig_cb, data)       \
 206 ((struct l_wait_info) {                                         \
 207         .lwi_timeout    = time,                                 \
 208         .lwi_on_timeout = time_cb,                              \
 209         .lwi_on_signal  = sig_cb,                               \
 210         .lwi_cb_data    = data,                                 \
 211         .lwi_interval   = 0,                                    \
 212         .lwi_allow_intr = 1                                     \
 213 })
 214
 215 #define LWI_INTR(cb, data)  LWI_TIMEOUT_INTR(0, NULL, cb, data)
 216
 217 #define LUSTRE_FATAL_SIGS                                        \
 218         (sigmask(SIGKILL) | sigmask(SIGINT) | sigmask(SIGTERM) | \
 219          sigmask(SIGQUIT) | sigmask(SIGALRM))
 220
 221 /*
 222  * Wait Queue
 223  */
 224 #if !defined(HAVE___ADD_WAIT_QUEUE_EXCLUSIVE) && !defined(HAVE_WAIT_QUEUE_ENTRY)
 225 static inline void __add_wait_queue_exclusive(wait_queue_head_t *q,
 226                                               wait_queue_t *wait)
 227 {
 228         wait->flags |= WQ_FLAG_EXCLUSIVE;
 229         __add_wait_queue(q, wait);
 230 }
 231 #endif /* HAVE___ADD_WAIT_QUEUE_EXCLUSIVE */
 232
 233 /**
 234  * wait_queue_t of Linux (version < 2.6.34) is a FIFO list for exclusively
 235  * waiting threads, which is not always desirable because all threads will
 236  * be waken up again and again, even user only needs a few of them to be
 237  * active most time. This is not good for performance because cache can
 238  * be polluted by different threads.
 239  *
 240  * LIFO list can resolve this problem because we always wakeup the most
 241  * recent active thread by default.
 242  *
 243  * NB: please don't call non-exclusive & exclusive wait on the same
 244  * waitq if add_wait_queue_exclusive_head is used.
 245  */
 246 #define add_wait_queue_exclusive_head(waitq, link)              \
 247 {                                                               \
 248         unsigned long flags;                                    \
 249                                                                 \
 250         spin_lock_irqsave(&((waitq)->lock), flags);             \
 251         __add_wait_queue_exclusive(waitq, link);                \
 252         spin_unlock_irqrestore(&((waitq)->lock), flags);        \
 253 }
 254
 255 /*
 256  * wait for @condition to become true, but no longer than timeout, specified
 257  * by @info.
 258  */
 259 #define __l_wait_event(wq, condition, info, ret, l_add_wait)                   \
 260 do {                                                                           \
 261         wait_queue_entry_t __wait;                                             \
 262         long __timeout = info->lwi_timeout;                                    \
 263         sigset_t __blocked;                                                    \
 264         int   __allow_intr = info->lwi_allow_intr;                             \
 265                                                                                \
 266         ret = 0;                                                               \
 267         if (condition)                                                         \
 268                 break;                                                         \
 269                                                                                \
 270         init_waitqueue_entry(&__wait, current);                                \
 271         l_add_wait(&wq, &__wait);                                              \
 272                                                                                \
 273         /* Block all signals (just the non-fatal ones if no timeout). */       \
 274         if (info->lwi_on_signal != NULL && (__timeout == 0 || __allow_intr))   \
 275                 __blocked = cfs_block_sigsinv(LUSTRE_FATAL_SIGS);              \
 276         else                                                                   \
 277                 __blocked = cfs_block_sigsinv(0);                              \
 278                                                                                \
 279         for (;;) {                                                             \
 280                 set_current_state(TASK_INTERRUPTIBLE);                         \
 281                                                                                \
 282                 /* To guarantee that the condition check will be done */       \
 283                 /* after setting the thread state as TASK_INTERRUPTIBLE. */    \
 284                 /* Otherwise, out-of-order execution may cause some race. */   \
 285                 /* Consider the following real execution order: */             \
 286                                                                                \
 287                 /* 1. Thread1 checks condition on CPU1, gets false. */         \
 288                 /* 2. Thread2 sets condition on CPU2. */                       \
 289                 /* 3. Thread2 calls wake_up() on CPU2 to wake the threads */   \
 290                 /*    with state TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE. */ \
 291                 /*    But the Thread1's state is TASK_RUNNING at that time. */ \
 292                 /* 4. Thread1 sets its state as TASK_INTERRUPTIBLE on CPU1, */ \
 293                 /*    then schedule. */                                        \
 294                                                                                \
 295                 /* If the '__timeout' variable is zero, the Thread1 will */    \
 296                 /* have no chance to check the condition again. */             \
 297                                                                                \
 298                 /* Generally, the interval between out-of-ordered step1 and */ \
 299                 /* step4 is very tiny, as to above step2 and step3 cannot */   \
 300                 /* happen. On some degree, it can explain why we seldom hit */ \
 301                 /* related trouble. But such race really exists, especially */ \
 302                 /* consider that the step1 and step4 can be interruptible. */  \
 303                 /* So add barrier to avoid Thread1 out-of-order execution. */  \
 304                 smp_mb();                                                      \
 305                                                                                \
 306                 if (condition)                                                 \
 307                         break;                                                 \
 308                                                                                \
 309                 if (__timeout == 0) {                                          \
 310                         schedule();                                            \
 311                 } else {                                                       \
 312                         long interval = info->lwi_interval ?                   \
 313                                                 min_t(long, info->lwi_interval,\
 314                                                       __timeout) : __timeout;  \
 315                         long remaining = schedule_timeout(interval);           \
 316                                                                                \
 317                         __timeout -= interval - remaining;                     \
 318                         if (__timeout == 0) {                                  \
 319                                 if (info->lwi_on_timeout == NULL ||            \
 320                                     info->lwi_on_timeout(info->lwi_cb_data)) { \
 321                                         ret = -ETIMEDOUT;                      \
 322                                         break;                                 \
 323                                 }                                              \
 324                                 /* Take signals after the timeout expires. */  \
 325                                 if (info->lwi_on_signal != NULL)               \
 326                                     (void)cfs_block_sigsinv(LUSTRE_FATAL_SIGS);\
 327                         }                                                      \
 328                 }                                                              \
 329                                                                                \
 330                 if (condition)                                                 \
 331                         break;                                                 \
 332                 if (signal_pending(current)) {                                 \
 333                         if (info->lwi_on_signal != NULL &&                     \
 334                             (__timeout == 0 || __allow_intr)) {                \
 335                                 if (info->lwi_on_signal != LWI_ON_SIGNAL_NOOP) \
 336                                         info->lwi_on_signal(info->lwi_cb_data);\
 337                                 ret = -EINTR;                                  \
 338                                 break;                                         \
 339                         }                                                      \
 340                         /* We have to do this here because some signals */     \
 341                         /* are not blockable - ie from strace(1).       */     \
 342                         /* In these cases we want to schedule_timeout() */     \
 343                         /* again, because we don't want that to return  */     \
 344                         /* -EINTR when the RPC actually succeeded.      */     \
 345                         /* the recalc_sigpending() below will deliver the */   \
 346                         /* signal properly.                             */     \
 347                         cfs_clear_sigpending();                                \
 348                 }                                                              \
 349         }                                                                      \
 350                                                                                \
 351         cfs_restore_sigs(__blocked);                                           \
 352                                                                                \
 353         set_current_state(TASK_RUNNING);                                       \
 354         remove_wait_queue(&wq, &__wait);                                       \
 355 } while (0)
 356
 357
 358 #define l_wait_event(wq, condition, info)                       \
 359 ({                                                              \
 360         int                 __ret;                              \
 361         struct l_wait_info *__info = (info);                    \
 362                                                                 \
 363         __l_wait_event(wq, condition, __info,                   \
 364                        __ret, add_wait_queue);                  \
 365         __ret;                                                  \
 366 })
 367
 368 #define l_wait_event_exclusive(wq, condition, info)             \
 369 ({                                                              \
 370         int                 __ret;                              \
 371         struct l_wait_info *__info = (info);                    \
 372                                                                 \
 373         __l_wait_event(wq, condition, __info,                   \
 374                        __ret, add_wait_queue_exclusive);        \
 375         __ret;                                                  \
 376 })
 377
 378 #define l_wait_event_exclusive_head(wq, condition, info)        \
 379 ({                                                              \
 380         int                 __ret;                              \
 381         struct l_wait_info *__info = (info);                    \
 382                                                                 \
 383         __l_wait_event(wq, condition, __info,                   \
 384                        __ret, add_wait_queue_exclusive_head);   \
 385         __ret;                                                  \
 386 })
 387
 388 #define l_wait_condition(wq, condition)                         \
 389 ({                                                              \
 390         struct l_wait_info lwi = { 0 };                         \
 391         l_wait_event(wq, condition, &lwi);                      \
 392 })
 393
 394 #define l_wait_condition_exclusive(wq, condition)               \
 395 ({                                                              \
 396         struct l_wait_info lwi = { 0 };                         \
 397         l_wait_event_exclusive(wq, condition, &lwi);            \
 398 })
 399
 400 #define l_wait_condition_exclusive_head(wq, condition)          \
 401 ({                                                              \
 402         struct l_wait_info lwi = { 0 };                         \
 403         l_wait_event_exclusive_head(wq, condition, &lwi);       \
 404 })
 405
 406 /** @} lib */
 407
 408 #endif /* _LUSTRE_LIB_H */