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