lustre/include/linux/lustre_lib.h

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2001 Cluster File Systems, Inc. <braam@clusterfs.com>
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  * Basic Lustre library routines.
  22  *
  23  */
  24
  25 #ifndef _LUSTRE_LIB_H
  26 #define _LUSTRE_LIB_H
  27
  28 #ifndef __KERNEL__
  29 # include <string.h>
  30 # include <sys/types.h>
  31 #else
  32 # include <asm/semaphore.h>
  33 # include <linux/rwsem.h>
  34 # include <linux/sched.h>
  35 # include <linux/signal.h>
  36 # include <linux/types.h>
  37 #endif
  38 #include <libcfs/kp30.h>
  39 #include <linux/lustre_idl.h>
  40 #include <linux/lustre_cfg.h>
  41
  42 #ifndef LP_POISON
  43 #if BITS_PER_LONG > 32
  44 # define LI_POISON ((int)0x5a5a5a5a5a5a5a5a)
  45 # define LL_POISON ((long)0x5a5a5a5a5a5a5a5a)
  46 # define LP_POISON ((void *)(long)0x5a5a5a5a5a5a5a5a)
  47 #else
  48 # define LI_POISON ((int)0x5a5a5a5a)
  49 # define LL_POISON ((long)0x5a5a5a5a)
  50 # define LP_POISON ((void *)(long)0x5a5a5a5a)
  51 #endif
  52 #endif
  53
  54 /* prng.c */
  55 unsigned int ll_rand(void);        /* returns a random 32-bit integer */
  56 void ll_srand(unsigned int, unsigned int);     /* seed the generator */
  57
  58 /* target.c */
  59 struct ptlrpc_request;
  60 struct recovd_data;
  61 struct recovd_obd;
  62 struct obd_export;
  63 #include <linux/lustre_ha.h>
  64 #include <linux/lustre_net.h>
  65 #include <linux/lustre_compat25.h>
  66 #include <linux/lvfs.h>
  67
  68 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler);
  69 int target_handle_disconnect(struct ptlrpc_request *req);
  70 void target_destroy_export(struct obd_export *exp);
  71 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
  72                             struct obd_uuid *cluuid);
  73 int target_handle_ping(struct ptlrpc_request *req);
  74 void target_committed_to_req(struct ptlrpc_request *req);
  75
  76 #ifdef HAVE_QUOTA_SUPPORT
  77 /* quotacheck callback, dqacq/dqrel callback handler */
  78 int target_handle_qc_callback(struct ptlrpc_request *req);
  79 int target_handle_dqacq_callback(struct ptlrpc_request *req);
  80 #else
  81 #define target_handle_dqacq_callback(req) ldlm_callback_reply(req, -ENOTSUPP)
  82 #define target_handle_qc_callback(req) (0)
  83 #endif
  84
  85 void target_cancel_recovery_timer(struct obd_device *obd);
  86
  87 #define OBD_RECOVERY_TIMEOUT (obd_timeout * 5 * HZ / 2) /* *waves hands* */
  88 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler);
  89 void target_abort_recovery(void *data);
  90 void target_cleanup_recovery(struct obd_device *obd);
  91 int target_queue_recovery_request(struct ptlrpc_request *req,
  92                                   struct obd_device *obd);
  93 int target_queue_final_reply(struct ptlrpc_request *req, int rc);
  94 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id);
  95
  96 /* client.c */
  97
  98 int client_sanobd_setup(struct obd_device *obddev, struct lustre_cfg* lcfg);
  99 struct client_obd *client_conn2cli(struct lustre_handle *conn);
 100
 101 struct mdc_open_data;
 102 struct obd_client_handle {
 103         struct lustre_handle och_fh;
 104         struct llog_cookie och_cookie;
 105         struct mdc_open_data *och_mod;
 106         __u32 och_magic;
 107 };
 108 #define OBD_CLIENT_HANDLE_MAGIC 0xd15ea5ed
 109
 110 /* statfs_pack.c */
 111 struct obd_statfs;
 112 struct kstatfs;
 113 void statfs_pack(struct obd_statfs *osfs, struct kstatfs *sfs);
 114 void statfs_unpack(struct kstatfs *sfs, struct obd_statfs *osfs);
 115
 116 /* l_lock.c */
 117 struct lustre_lock {
 118         int l_depth;
 119         struct task_struct *l_owner;
 120         struct semaphore l_sem;
 121         spinlock_t l_spin;
 122 };
 123
 124 void l_lock_init(struct lustre_lock *);
 125 void l_lock(struct lustre_lock *);
 126 void l_unlock(struct lustre_lock *);
 127 int l_has_lock(struct lustre_lock *);
 128
 129
 130 /*
 131  *   OBD IOCTLS
 132  */
 133 #define OBD_IOCTL_VERSION 0x00010004
 134
 135 struct obd_ioctl_data {
 136         uint32_t ioc_len;
 137         uint32_t ioc_version;
 138
 139         uint64_t ioc_cookie;
 140         uint32_t ioc_conn1;
 141         uint32_t ioc_conn2;
 142
 143         struct obdo ioc_obdo1;
 144         struct obdo ioc_obdo2;
 145
 146         obd_size         ioc_count;
 147         obd_off          ioc_offset;
 148         uint32_t         ioc_dev;
 149         uint32_t         ioc_command;
 150
 151         uint64_t ioc_nid;
 152         uint32_t ioc_nal;
 153         uint32_t ioc_type;
 154
 155         /* buffers the kernel will treat as user pointers */
 156         uint32_t ioc_plen1;
 157         char    *ioc_pbuf1;
 158         uint32_t ioc_plen2;
 159         char    *ioc_pbuf2;
 160
 161         /* inline buffers for various arguments */
 162         uint32_t ioc_inllen1;
 163         char    *ioc_inlbuf1;
 164         uint32_t ioc_inllen2;
 165         char    *ioc_inlbuf2;
 166         uint32_t ioc_inllen3;
 167         char    *ioc_inlbuf3;
 168         uint32_t ioc_inllen4;
 169         char    *ioc_inlbuf4;
 170
 171         char    ioc_bulk[0];
 172 };
 173
 174 struct obd_ioctl_hdr {
 175         uint32_t ioc_len;
 176         uint32_t ioc_version;
 177 };
 178
 179 static inline int obd_ioctl_packlen(struct obd_ioctl_data *data)
 180 {
 181         int len = size_round(sizeof(struct obd_ioctl_data));
 182         len += size_round(data->ioc_inllen1);
 183         len += size_round(data->ioc_inllen2);
 184         len += size_round(data->ioc_inllen3);
 185         len += size_round(data->ioc_inllen4);
 186         return len;
 187 }
 188
 189
 190 static inline int obd_ioctl_is_invalid(struct obd_ioctl_data *data)
 191 {
 192         if (data->ioc_len > (1<<30)) {
 193                 CERROR("OBD ioctl: ioc_len larger than 1<<30\n");
 194                 return 1;
 195         }
 196         if (data->ioc_inllen1 > (1<<30)) {
 197                 CERROR("OBD ioctl: ioc_inllen1 larger than 1<<30\n");
 198                 return 1;
 199         }
 200         if (data->ioc_inllen2 > (1<<30)) {
 201                 CERROR("OBD ioctl: ioc_inllen2 larger than 1<<30\n");
 202                 return 1;
 203         }
 204         if (data->ioc_inllen3 > (1<<30)) {
 205                 CERROR("OBD ioctl: ioc_inllen3 larger than 1<<30\n");
 206                 return 1;
 207         }
 208         if (data->ioc_inllen4 > (1<<30)) {
 209                 CERROR("OBD ioctl: ioc_inllen4 larger than 1<<30\n");
 210                 return 1;
 211         }
 212         if (data->ioc_inlbuf1 && !data->ioc_inllen1) {
 213                 CERROR("OBD ioctl: inlbuf1 pointer but 0 length\n");
 214                 return 1;
 215         }
 216         if (data->ioc_inlbuf2 && !data->ioc_inllen2) {
 217                 CERROR("OBD ioctl: inlbuf2 pointer but 0 length\n");
 218                 return 1;
 219         }
 220         if (data->ioc_inlbuf3 && !data->ioc_inllen3) {
 221                 CERROR("OBD ioctl: inlbuf3 pointer but 0 length\n");
 222                 return 1;
 223         }
 224         if (data->ioc_inlbuf4 && !data->ioc_inllen4) {
 225                 CERROR("OBD ioctl: inlbuf4 pointer but 0 length\n");
 226                 return 1;
 227         }
 228         if (data->ioc_pbuf1 && !data->ioc_plen1) {
 229                 CERROR("OBD ioctl: pbuf1 pointer but 0 length\n");
 230                 return 1;
 231         }
 232         if (data->ioc_pbuf2 && !data->ioc_plen2) {
 233                 CERROR("OBD ioctl: pbuf2 pointer but 0 length\n");
 234                 return 1;
 235         }
 236         if (data->ioc_plen1 && !data->ioc_pbuf1) {
 237                 CERROR("OBD ioctl: plen1 set but NULL pointer\n");
 238                 return 1;
 239         }
 240         if (data->ioc_plen2 && !data->ioc_pbuf2) {
 241                 CERROR("OBD ioctl: plen2 set but NULL pointer\n");
 242                 return 1;
 243         }
 244         if (obd_ioctl_packlen(data) > data->ioc_len) {
 245                 CERROR("OBD ioctl: packlen exceeds ioc_len (%d > %d)\n",
 246                        obd_ioctl_packlen(data), data->ioc_len);
 247                 return 1;
 248         }
 249         return 0;
 250 }
 251
 252 #ifndef __KERNEL__
 253 static inline int obd_ioctl_pack(struct obd_ioctl_data *data, char **pbuf,
 254                                  int max)
 255 {
 256         char *ptr;
 257         struct obd_ioctl_data *overlay;
 258         data->ioc_len = obd_ioctl_packlen(data);
 259         data->ioc_version = OBD_IOCTL_VERSION;
 260
 261         if (*pbuf && data->ioc_len > max)
 262                 return 1;
 263         if (*pbuf == NULL) {
 264                 *pbuf = malloc(data->ioc_len);
 265         }
 266         if (!*pbuf)
 267                 return 1;
 268         overlay = (struct obd_ioctl_data *)*pbuf;
 269         memcpy(*pbuf, data, sizeof(*data));
 270
 271         ptr = overlay->ioc_bulk;
 272         if (data->ioc_inlbuf1)
 273                 LOGL(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
 274         if (data->ioc_inlbuf2)
 275                 LOGL(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
 276         if (data->ioc_inlbuf3)
 277                 LOGL(data->ioc_inlbuf3, data->ioc_inllen3, ptr);
 278         if (data->ioc_inlbuf4)
 279                 LOGL(data->ioc_inlbuf4, data->ioc_inllen4, ptr);
 280         if (obd_ioctl_is_invalid(overlay))
 281                 return 1;
 282
 283         return 0;
 284 }
 285
 286 static inline int obd_ioctl_unpack(struct obd_ioctl_data *data, char *pbuf,
 287                                    int max)
 288 {
 289         char *ptr;
 290         struct obd_ioctl_data *overlay;
 291
 292         if (!pbuf)
 293                 return 1;
 294         overlay = (struct obd_ioctl_data *)pbuf;
 295
 296         /* Preserve the caller's buffer pointers */
 297         overlay->ioc_inlbuf1 = data->ioc_inlbuf1;
 298         overlay->ioc_inlbuf2 = data->ioc_inlbuf2;
 299         overlay->ioc_inlbuf3 = data->ioc_inlbuf3;
 300         overlay->ioc_inlbuf4 = data->ioc_inlbuf4;
 301
 302         memcpy(data, pbuf, sizeof(*data));
 303
 304         ptr = overlay->ioc_bulk;
 305         if (data->ioc_inlbuf1)
 306                 LOGU(data->ioc_inlbuf1, data->ioc_inllen1, ptr);
 307         if (data->ioc_inlbuf2)
 308                 LOGU(data->ioc_inlbuf2, data->ioc_inllen2, ptr);
 309         if (data->ioc_inlbuf3)
 310                 LOGU(data->ioc_inlbuf3, data->ioc_inllen3, ptr);
 311         if (data->ioc_inlbuf4)
 312                 LOGU(data->ioc_inlbuf4, data->ioc_inllen4, ptr);
 313
 314         return 0;
 315 }
 316 #endif
 317
 318 #include <linux/obd_support.h>
 319
 320 /* buffer MUST be at least the size of obd_ioctl_hdr */
 321 static inline int obd_ioctl_getdata(char **buf, int *len, void *arg)
 322 {
 323         struct obd_ioctl_hdr hdr;
 324         struct obd_ioctl_data *data;
 325         int err;
 326         int offset = 0;
 327         ENTRY;
 328
 329         err = copy_from_user(&hdr, (void *)arg, sizeof(hdr));
 330         if (err)
 331                 RETURN(err);
 332
 333         if (hdr.ioc_version != OBD_IOCTL_VERSION) {
 334                 CERROR("Version mismatch kernel vs application\n");
 335                 RETURN(-EINVAL);
 336         }
 337
 338         if (hdr.ioc_len > OBD_MAX_IOCTL_BUFFER) {
 339                 CERROR("User buffer len %d exceeds %d max buffer\n",
 340                        hdr.ioc_len, OBD_MAX_IOCTL_BUFFER);
 341                 RETURN(-EINVAL);
 342         }
 343
 344         if (hdr.ioc_len < sizeof(struct obd_ioctl_data)) {
 345                 CERROR("user buffer too small for ioctl (%d)\n", hdr.ioc_len);
 346                 RETURN(-EINVAL);
 347         }
 348
 349         /* XXX allocate this more intelligently, using kmalloc when
 350          * appropriate */
 351         OBD_VMALLOC(*buf, hdr.ioc_len);
 352         if (*buf == NULL) {
 353                 CERROR("Cannot allocate control buffer of len %d\n",
 354                        hdr.ioc_len);
 355                 RETURN(-EINVAL);
 356         }
 357         *len = hdr.ioc_len;
 358         data = (struct obd_ioctl_data *)*buf;
 359
 360         err = copy_from_user(*buf, (void *)arg, hdr.ioc_len);
 361         if (err) {
 362                 OBD_VFREE(*buf, hdr.ioc_len);
 363                 RETURN(err);
 364         }
 365
 366         if (obd_ioctl_is_invalid(data)) {
 367                 CERROR("ioctl not correctly formatted\n");
 368                 OBD_VFREE(*buf, hdr.ioc_len);
 369                 RETURN(-EINVAL);
 370         }
 371
 372         if (data->ioc_inllen1) {
 373                 data->ioc_inlbuf1 = &data->ioc_bulk[0];
 374                 offset += size_round(data->ioc_inllen1);
 375         }
 376
 377         if (data->ioc_inllen2) {
 378                 data->ioc_inlbuf2 = &data->ioc_bulk[0] + offset;
 379                 offset += size_round(data->ioc_inllen2);
 380         }
 381
 382         if (data->ioc_inllen3) {
 383                 data->ioc_inlbuf3 = &data->ioc_bulk[0] + offset;
 384                 offset += size_round(data->ioc_inllen3);
 385         }
 386
 387         if (data->ioc_inllen4) {
 388                 data->ioc_inlbuf4 = &data->ioc_bulk[0] + offset;
 389         }
 390
 391         RETURN(0);
 392 }
 393
 394 static inline void obd_ioctl_freedata(char *buf, int len)
 395 {
 396         ENTRY;
 397
 398         OBD_VFREE(buf, len);
 399         EXIT;
 400         return;
 401 }
 402
 403 #define OBD_IOC_CREATE                 _IOR ('f', 101, long)
 404 #define OBD_IOC_DESTROY                _IOW ('f', 104, long)
 405 #define OBD_IOC_PREALLOCATE            _IOWR('f', 105, long)
 406
 407 #define OBD_IOC_SETATTR                _IOW ('f', 107, long)
 408 #define OBD_IOC_GETATTR                _IOR ('f', 108, long)
 409 #define OBD_IOC_READ                   _IOWR('f', 109, long)
 410 #define OBD_IOC_WRITE                  _IOWR('f', 110, long)
 411
 412
 413 #define OBD_IOC_STATFS                 _IOWR('f', 113, long)
 414 #define OBD_IOC_SYNC                   _IOW ('f', 114, long)
 415 #define OBD_IOC_READ2                  _IOWR('f', 115, long)
 416 #define OBD_IOC_FORMAT                 _IOWR('f', 116, long)
 417 #define OBD_IOC_PARTITION              _IOWR('f', 117, long)
 418 #define OBD_IOC_COPY                   _IOWR('f', 120, long)
 419 #define OBD_IOC_MIGR                   _IOWR('f', 121, long)
 420 #define OBD_IOC_PUNCH                  _IOWR('f', 122, long)
 421
 422 #define OBD_IOC_MODULE_DEBUG           _IOWR('f', 124, long)
 423 #define OBD_IOC_BRW_READ               _IOWR('f', 125, long)
 424 #define OBD_IOC_BRW_WRITE              _IOWR('f', 126, long)
 425 #define OBD_IOC_NAME2DEV               _IOWR('f', 127, long)
 426 #define OBD_IOC_UUID2DEV               _IOWR('f', 130, long)
 427 #define OBD_IOC_GETNAME                _IOR ('f', 131, long)
 428
 429 #define OBD_IOC_LOV_GET_CONFIG         _IOWR('f', 132, long)
 430 #define OBD_IOC_CLIENT_RECOVER         _IOW ('f', 133, long)
 431
 432 #define OBD_IOC_DEC_FS_USE_COUNT       _IO  ('f', 139      )
 433 #define OBD_IOC_NO_TRANSNO             _IOW ('f', 140, long)
 434 #define OBD_IOC_SET_READONLY           _IOW ('f', 141, long)
 435 #define OBD_IOC_ABORT_RECOVERY         _IOR ('f', 142, long)
 436
 437 #define OBD_GET_VERSION                _IOWR ('f', 144, long)
 438
 439 #define OBD_IOC_CLOSE_UUID             _IOWR ('f', 147, long)
 440
 441 #define OBD_IOC_LOV_SETSTRIPE          _IOW ('f', 154, long)
 442 #define OBD_IOC_LOV_GETSTRIPE          _IOW ('f', 155, long)
 443 #define OBD_IOC_LOV_SETEA              _IOW ('f', 156, long)
 444
 445 #define OBD_IOC_QUOTACHECK             _IOW ('f', 160, int)
 446 #define OBD_IOC_POLL_QUOTACHECK        _IOR ('f', 161, struct if_quotacheck *)
 447 #define OBD_IOC_QUOTACTL               _IOWR('f', 162, struct if_quotactl *)
 448
 449 #define OBD_IOC_MOUNTOPT               _IOWR('f', 170, long)
 450
 451 #define OBD_IOC_RECORD                 _IOWR('f', 180, long)
 452 #define OBD_IOC_ENDRECORD              _IOWR('f', 181, long)
 453 #define OBD_IOC_PARSE                  _IOWR('f', 182, long)
 454 #define OBD_IOC_DORECORD               _IOWR('f', 183, long)
 455 #define OBD_IOC_PROCESS_CFG            _IOWR('f', 184, long)
 456 #define OBD_IOC_DUMP_LOG               _IOWR('f', 185, long)
 457 #define OBD_IOC_CLEAR_LOG              _IOWR('f', 186, long)
 458 #define OBD_IOC_PARAM                  _IOW ('f', 187, long)
 459
 460 #define OBD_IOC_CATLOGLIST             _IOWR('f', 190, long)
 461 #define OBD_IOC_LLOG_INFO              _IOWR('f', 191, long)
 462 #define OBD_IOC_LLOG_PRINT             _IOWR('f', 192, long)
 463 #define OBD_IOC_LLOG_CANCEL            _IOWR('f', 193, long)
 464 #define OBD_IOC_LLOG_REMOVE            _IOWR('f', 194, long)
 465 #define OBD_IOC_LLOG_CHECK             _IOWR('f', 195, long)
 466 #define OBD_IOC_LLOG_CATINFO           _IOWR('f', 196, long)
 467
 468 #define ECHO_IOC_GET_STRIPE            _IOWR('f', 200, long)
 469 #define ECHO_IOC_SET_STRIPE            _IOWR('f', 201, long)
 470 #define ECHO_IOC_ENQUEUE               _IOWR('f', 202, long)
 471 #define ECHO_IOC_CANCEL                _IOWR('f', 203, long)
 472
 473 /* XXX _IOWR('f', 250, long) has been defined in
 474  * lnet/include/libcfs/kp30.h for debug, don't use it
 475  */
 476
 477 /* Until such time as we get_info the per-stripe maximum from the OST,
 478  * we define this to be 2T - 4k, which is the ext3 maxbytes. */
 479 #define LUSTRE_STRIPE_MAXBYTES 0x1fffffff000ULL
 480
 481 #define POISON_BULK 0
 482
 483 /*
 484  * l_wait_event is a flexible sleeping function, permitting simple caller
 485  * configuration of interrupt and timeout sensitivity along with actions to
 486  * be performed in the event of either exception.
 487  *
 488  * The first form of usage looks like this:
 489  *
 490  * struct l_wait_info lwi = LWI_TIMEOUT_INTR(timeout, timeout_handler,
 491  *                                           intr_handler, callback_data);
 492  * rc = l_wait_event(waitq, condition, &lwi);
 493  *
 494  * l_wait_event() makes the current process wait on 'waitq' until 'condition'
 495  * is TRUE or a "killable" signal (SIGTERM, SIKGILL, SIGINT) is pending.  It
 496  * returns 0 to signify 'condition' is TRUE, but if a signal wakes it before
 497  * 'condition' becomes true, it optionally calls the specified 'intr_handler'
 498  * if not NULL, and returns -EINTR.
 499  *
 500  * If a non-zero timeout is specified, signals are ignored until the timeout
 501  * has expired.  At this time, if 'timeout_handler' is not NULL it is called.
 502  * If it returns FALSE l_wait_event() continues to wait as described above with
 503  * signals enabled.  Otherwise it returns -ETIMEDOUT.
 504  *
 505  * LWI_INTR(intr_handler, callback_data) is shorthand for
 506  * LWI_TIMEOUT_INTR(0, NULL, intr_handler, callback_data)
 507  *
 508  * The second form of usage looks like this:
 509  *
 510  * struct l_wait_info lwi = LWI_TIMEOUT(timeout, timeout_handler);
 511  * rc = l_wait_event(waitq, condition, &lwi);
 512  *
 513  * This form is the same as the first except that it COMPLETELY IGNORES
 514  * SIGNALS.  The caller must therefore beware that if 'timeout' is zero, or if
 515  * 'timeout_handler' is not NULL and returns FALSE, then the ONLY thing that
 516  * can unblock the current process is 'condition' becoming TRUE.
 517  *
 518  * Another form of usage is:
 519  * struct l_wait_info lwi = LWI_TIMEOUT_INTERVAL(timeout, interval,
 520  *                                               timeout_handler);
 521  * rc = l_wait_event(waitq, condition, &lwi);
 522  * This is the same as previous case, but condition is checked once every
 523  * 'interval' jiffies (if non-zero).
 524  *
 525  * Subtle synchronization point: this macro does *not* necessary takes
 526  * wait-queue spin-lock before returning, and, hence, following idiom is safe
 527  * ONLY when caller provides some external locking:
 528  *
 529  *             Thread1                            Thread2
 530  *
 531  *   l_wait_event(&obj->wq, ....);                                       (1)
 532  *
 533  *                                    wake_up(&obj->wq):                 (2)
 534  *                                         spin_lock(&q->lock);          (2.1)
 535  *                                         __wake_up_common(q, ...);     (2.2)
 536  *                                         spin_unlock(&q->lock, flags); (2.3)
 537  *
 538  *   OBD_FREE_PTR(obj);                                                  (3)
 539  *
 540  * As l_wait_event() may "short-cut" execution and return without taking
 541  * wait-queue spin-lock, some additional synchronization is necessary to
 542  * guarantee that step (3) can begin only after (2.3) finishes.
 543  *
 544  * XXX nikita: some ptlrpc daemon threads have races of that sort.
 545  *
 546  */
 547
 548 #define LWI_ON_SIGNAL_NOOP ((void (*)(void *))(-1))
 549
 550 struct l_wait_info {
 551         long   lwi_timeout;
 552         long   lwi_interval;
 553         int  (*lwi_on_timeout)(void *);
 554         void (*lwi_on_signal)(void *);
 555         void  *lwi_cb_data;
 556 };
 557
 558 /* NB: LWI_TIMEOUT ignores signals completely */
 559 #define LWI_TIMEOUT(time, cb, data)             \
 560 ((struct l_wait_info) {                         \
 561         .lwi_timeout    = time,                 \
 562         .lwi_on_timeout = cb,                   \
 563         .lwi_cb_data    = data,                 \
 564         .lwi_interval   = 0                     \
 565 })
 566
 567 #define LWI_TIMEOUT_INTERVAL(time, interval, cb, data)  \
 568 ((struct l_wait_info) {                                 \
 569         .lwi_timeout    = time,                         \
 570         .lwi_on_timeout = cb,                           \
 571         .lwi_cb_data    = data,                         \
 572         .lwi_interval   = interval                      \
 573 })
 574
 575
 576 #define LWI_TIMEOUT_INTR(time, time_cb, sig_cb, data)                          \
 577 ((struct l_wait_info) {                                                        \
 578         .lwi_timeout    = time,                                                \
 579         .lwi_on_timeout = time_cb,                                             \
 580         .lwi_on_signal = (sig_cb == NULL) ? LWI_ON_SIGNAL_NOOP : sig_cb,       \
 581         .lwi_cb_data    = data,                                                \
 582         .lwi_interval    = 0                                                   \
 583 })
 584
 585 #define LWI_INTR(cb, data)  LWI_TIMEOUT_INTR(0, NULL, cb, data)
 586
 587 #define LUSTRE_FATAL_SIGS (sigmask(SIGKILL) | sigmask(SIGINT) |                \
 588                            sigmask(SIGTERM) | sigmask(SIGQUIT) |               \
 589                            sigmask(SIGALRM))
 590
 591 #ifdef __KERNEL__
 592 static inline sigset_t l_w_e_set_sigs(int sigs)
 593 {
 594         sigset_t old;
 595         unsigned long irqflags;
 596
 597         SIGNAL_MASK_LOCK(current, irqflags);
 598         old = current->blocked;
 599         siginitsetinv(&current->blocked, sigs);
 600         RECALC_SIGPENDING;
 601         SIGNAL_MASK_UNLOCK(current, irqflags);
 602
 603         return old;
 604 }
 605
 606 /*
 607  * wait for @condition to become true, but no longer than timeout, specified
 608  * by @info.
 609  */
 610 #define __l_wait_event(wq, condition, info, ret, excl)                         \
 611 do {                                                                           \
 612         wait_queue_t  __wait;                                                  \
 613         unsigned long __timeout = info->lwi_timeout;                           \
 614         unsigned long __irqflags;                                              \
 615         sigset_t      __blocked;                                               \
 616                                                                                \
 617         ret = 0;                                                               \
 618         if (condition)                                                         \
 619                 break;                                                         \
 620                                                                                \
 621         init_waitqueue_entry(&__wait, current);                                \
 622         if (excl)                                                              \
 623                 add_wait_queue_exclusive(&wq, &__wait);                        \
 624         else                                                                   \
 625                 add_wait_queue(&wq, &__wait);                                  \
 626                                                                                \
 627         /* Block all signals (just the non-fatal ones if no timeout). */       \
 628         if (info->lwi_on_signal != NULL && __timeout == 0)                     \
 629                 __blocked = l_w_e_set_sigs(LUSTRE_FATAL_SIGS);                 \
 630         else                                                                   \
 631                 __blocked = l_w_e_set_sigs(0);                                 \
 632                                                                                \
 633         for (;;) {                                                             \
 634                 set_current_state(TASK_INTERRUPTIBLE);                         \
 635                                                                                \
 636                 if (condition)                                                 \
 637                         break;                                                 \
 638                                                                                \
 639                 if (__timeout == 0) {                                          \
 640                         schedule();                                            \
 641                 } else {                                                       \
 642                         unsigned long interval = info->lwi_interval?           \
 643                                              min_t(unsigned long,              \
 644                                                  info->lwi_interval,__timeout):\
 645                                              __timeout;                        \
 646                         __timeout -= interval - schedule_timeout(interval);    \
 647                         if (__timeout == 0) {                                  \
 648                                 if (info->lwi_on_timeout == NULL ||            \
 649                                     info->lwi_on_timeout(info->lwi_cb_data)) { \
 650                                         ret = -ETIMEDOUT;                      \
 651                                         break;                                 \
 652                                 }                                              \
 653                                 /* Take signals after the timeout expires. */  \
 654                                 if (info->lwi_on_signal != NULL)               \
 655                                     (void)l_w_e_set_sigs(LUSTRE_FATAL_SIGS);   \
 656                         }                                                      \
 657                 }                                                              \
 658                                                                                \
 659                 if (condition)                                                 \
 660                         break;                                                 \
 661                                                                                \
 662                 if (signal_pending(current)) {                                 \
 663                         if (info->lwi_on_signal != NULL && __timeout == 0) {   \
 664                                 if (info->lwi_on_signal != LWI_ON_SIGNAL_NOOP) \
 665                                         info->lwi_on_signal(info->lwi_cb_data);\
 666                                 ret = -EINTR;                                  \
 667                                 break;                                         \
 668                         }                                                      \
 669                         /* We have to do this here because some signals */     \
 670                         /* are not blockable - ie from strace(1).       */     \
 671                         /* In these cases we want to schedule_timeout() */     \
 672                         /* again, because we don't want that to return  */     \
 673                         /* -EINTR when the RPC actually succeeded.      */     \
 674                         /* the RECALC_SIGPENDING below will deliver the */     \
 675                         /* signal properly.                             */     \
 676                         SIGNAL_MASK_LOCK(current, __irqflags);                 \
 677                         CLEAR_SIGPENDING;                                      \
 678                         SIGNAL_MASK_UNLOCK(current, __irqflags);               \
 679                 }                                                              \
 680         }                                                                      \
 681                                                                                \
 682         SIGNAL_MASK_LOCK(current, __irqflags);                                 \
 683         current->blocked = __blocked;                                          \
 684         RECALC_SIGPENDING;                                                     \
 685         SIGNAL_MASK_UNLOCK(current, __irqflags);                               \
 686                                                                                \
 687         current->state = TASK_RUNNING;                                         \
 688         remove_wait_queue(&wq, &__wait);                                       \
 689 } while(0)
 690
 691 #else /* !__KERNEL__ */
 692 #define __l_wait_event(wq, condition, info, ret, excl)                  \
 693 do {                                                                    \
 694         long __timeout = info->lwi_timeout;                             \
 695         long __now;                                                     \
 696         long __then = 0;                                                \
 697         int  __timed_out = 0;                                           \
 698                                                                         \
 699         ret = 0;                                                        \
 700         if (condition)                                                  \
 701                 break;                                                  \
 702                                                                         \
 703         if (__timeout == 0)                                             \
 704                 __timeout = 1000000000;                                 \
 705         else                                                            \
 706                 __then = time(NULL);                                    \
 707                                                                         \
 708         while (!(condition)) {                                          \
 709                 if (liblustre_wait_event(info->lwi_interval?:__timeout) || \
 710                     (info->lwi_interval && info->lwi_interval < __timeout)) {\
 711                         if (__timeout != 0 && info->lwi_timeout != 0) { \
 712                                 __now = time(NULL);                     \
 713                                 __timeout -= __now - __then;            \
 714                                 if (__timeout < 0)                      \
 715                                         __timeout = 0;                  \
 716                                 __then = __now;                         \
 717                         }                                               \
 718                         continue;                                       \
 719                 }                                                       \
 720                                                                         \
 721                 if (info->lwi_timeout != 0 && !__timed_out) {           \
 722                         __timed_out = 1;                                \
 723                         if (info->lwi_on_timeout == NULL ||             \
 724                             info->lwi_on_timeout(info->lwi_cb_data)) {  \
 725                                 ret = -ETIMEDOUT;                       \
 726                                 break;                                  \
 727                         }                                               \
 728                 }                                                       \
 729         }                                                               \
 730 } while (0)
 731
 732 #endif /* __KERNEL__ */
 733
 734 #define l_wait_event(wq, condition, info)                       \
 735 ({                                                              \
 736         int                 __ret;                              \
 737         struct l_wait_info *__info = (info);                    \
 738                                                                 \
 739         __l_wait_event(wq, condition, __info, __ret, 0);        \
 740         __ret;                                                  \
 741 })
 742
 743 #define l_wait_event_exclusive(wq, condition, info)             \
 744 ({                                                              \
 745         int                 __ret;                              \
 746         struct l_wait_info *__info = (info);                    \
 747                                                                 \
 748         __l_wait_event(wq, condition, __info, __ret, 1);        \
 749         __ret;                                                  \
 750 })
 751
 752 #ifdef __KERNEL__
 753 /* initialize ost_lvb according to inode */
 754 static inline void inode_init_lvb(struct inode *inode, struct ost_lvb *lvb)
 755 {
 756         lvb->lvb_size = inode->i_size;
 757         lvb->lvb_blocks = inode->i_blocks;
 758         lvb->lvb_mtime = LTIME_S(inode->i_mtime);
 759         lvb->lvb_atime = LTIME_S(inode->i_atime);
 760         lvb->lvb_ctime = LTIME_S(inode->i_ctime);
 761 }
 762 #else
 763 /* defined in liblustre/llite_lib.h */
 764 #endif
 765
 766 #ifdef __KERNEL__
 767 #define LIBLUSTRE_CLIENT (0)
 768 #else
 769 #define LIBLUSTRE_CLIENT (1)
 770 #endif
 771
 772 #endif /* _LUSTRE_LIB_H */
 773