Whamcloud - gitweb
b=16098
[fs/lustre-release.git] / lustre / include / obd_support.h
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see [sun.com URL with a
20  * copy of GPLv2].
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 #ifndef _OBD_SUPPORT
38 #define _OBD_SUPPORT
39
40 #include <libcfs/libcfs.h>
41 #include <lvfs.h>
42 #include <lprocfs_status.h>
43
44 #if defined(__linux__)
45 #include <linux/obd_support.h>
46 #elif defined(__APPLE__)
47 #include <darwin/obd_support.h>
48 #elif defined(__WINNT__)
49 #include <winnt/obd_support.h>
50 #else
51 #error Unsupported operating system.
52 #endif
53
54 /* global variables */
55 extern struct lprocfs_stats *obd_memory;
56 enum {
57         OBD_MEMORY_STAT = 0,
58         OBD_MEMORY_PAGES_STAT = 1,
59         OBD_STATS_NUM,
60 };
61
62 enum {
63         OBD_FAIL_LOC_NOSET      = 0,
64         OBD_FAIL_LOC_ORSET      = 1,
65         OBD_FAIL_LOC_RESET      = 2
66 };
67
68 extern unsigned long obd_fail_loc;
69 extern unsigned int obd_fail_val;
70 extern unsigned int obd_debug_peer_on_timeout;
71 extern unsigned int obd_dump_on_timeout;
72 extern unsigned int obd_dump_on_eviction;
73 /* obd_timeout should only be used for recovery, not for
74    networking / disk / timings affected by load (use Adaptive Timeouts) */
75 extern unsigned int obd_timeout;          /* seconds */
76 extern unsigned int ldlm_timeout;         /* seconds */
77 extern unsigned int obd_sync_filter;
78 extern unsigned int obd_max_dirty_pages;
79 extern atomic_t obd_dirty_pages;
80 extern cfs_waitq_t obd_race_waitq;
81 extern int obd_race_state;
82 extern unsigned int obd_alloc_fail_rate;
83
84 int __obd_fail_check_set(__u32 id, __u32 value, int set);
85 int __obd_fail_timeout_set(__u32 id, __u32 value, int ms, int set);
86
87 /* lvfs.c */
88 int obd_alloc_fail(const void *ptr, const char *name, const char *type,
89                    size_t size, const char *file, int line);
90
91 /* Timeout definitions */
92 #define OBD_TIMEOUT_DEFAULT             100
93 #define LDLM_TIMEOUT_DEFAULT            20
94 #define MDS_LDLM_TIMEOUT_DEFAULT        6
95 /* Time to wait for all clients to reconnect during recovery */
96 /* Should be very conservative; must catch the first reconnect after reboot */
97 #define OBD_RECOVERY_FACTOR (3) /* times obd_timeout */
98 /* Change recovery-small 26b time if you change this */
99 #define PING_INTERVAL max(obd_timeout / 4, 1U)
100 /* Client may skip 1 ping; we must wait at least 2.5. But for multiple
101  * failover targets the client only pings one server at a time, and pings
102  * can be lost on a loaded network. Since eviction has serious consequences,
103  * and there's no urgent need to evict a client just because it's idle, we
104  * should be very conservative here. */
105 #define PING_EVICT_TIMEOUT (PING_INTERVAL * 6)
106 #define DISK_TIMEOUT 50          /* Beyond this we warn about disk speed */
107 #define CONNECTION_SWITCH_MIN 5U /* Connection switching rate limiter */
108  /* Max connect interval for nonresponsive servers; ~50s to avoid building up
109     connect requests in the LND queues, but within obd_timeout so we don't
110     miss the recovery window */
111 #define CONNECTION_SWITCH_MAX min(50U, max(CONNECTION_SWITCH_MIN,obd_timeout))
112 #define CONNECTION_SWITCH_INC 5  /* Connection timeout backoff */
113 #ifndef CRAY_XT3
114 /* In general this should be low to have quick detection of a system
115    running on a backup server. (If it's too low, import_select_connection
116    will increase the timeout anyhow.)  */
117 #define INITIAL_CONNECT_TIMEOUT max(CONNECTION_SWITCH_MIN,obd_timeout/20)
118 #else
119 /* ...but for very large systems (e.g. CRAY) we need to keep the initial
120    connect t.o. high (bz 10803), because they will nearly ALWAYS be doing the
121    connects for the first time (clients "reboot" after every process, so no
122    chance to generate adaptive timeout data. */
123 #define INITIAL_CONNECT_TIMEOUT max(CONNECTION_SWITCH_MIN,obd_timeout/2)
124 #endif
125 #define LONG_UNLINK 300          /* Unlink should happen before now */
126
127
128 #define OBD_FAIL_MDS                     0x100
129 #define OBD_FAIL_MDS_HANDLE_UNPACK       0x101
130 #define OBD_FAIL_MDS_GETATTR_NET         0x102
131 #define OBD_FAIL_MDS_GETATTR_PACK        0x103
132 #define OBD_FAIL_MDS_READPAGE_NET        0x104
133 #define OBD_FAIL_MDS_READPAGE_PACK       0x105
134 #define OBD_FAIL_MDS_SENDPAGE            0x106
135 #define OBD_FAIL_MDS_REINT_NET           0x107
136 #define OBD_FAIL_MDS_REINT_UNPACK        0x108
137 #define OBD_FAIL_MDS_REINT_SETATTR       0x109
138 #define OBD_FAIL_MDS_REINT_SETATTR_WRITE 0x10a
139 #define OBD_FAIL_MDS_REINT_CREATE        0x10b
140 #define OBD_FAIL_MDS_REINT_CREATE_WRITE  0x10c
141 #define OBD_FAIL_MDS_REINT_UNLINK        0x10d
142 #define OBD_FAIL_MDS_REINT_UNLINK_WRITE  0x10e
143 #define OBD_FAIL_MDS_REINT_LINK          0x10f
144 #define OBD_FAIL_MDS_REINT_LINK_WRITE    0x110
145 #define OBD_FAIL_MDS_REINT_RENAME        0x111
146 #define OBD_FAIL_MDS_REINT_RENAME_WRITE  0x112
147 #define OBD_FAIL_MDS_OPEN_NET            0x113
148 #define OBD_FAIL_MDS_OPEN_PACK           0x114
149 #define OBD_FAIL_MDS_CLOSE_NET           0x115
150 #define OBD_FAIL_MDS_CLOSE_PACK          0x116
151 #define OBD_FAIL_MDS_CONNECT_NET         0x117
152 #define OBD_FAIL_MDS_CONNECT_PACK        0x118
153 #define OBD_FAIL_MDS_REINT_NET_REP       0x119
154 #define OBD_FAIL_MDS_DISCONNECT_NET      0x11a
155 #define OBD_FAIL_MDS_GETSTATUS_NET       0x11b
156 #define OBD_FAIL_MDS_GETSTATUS_PACK      0x11c
157 #define OBD_FAIL_MDS_STATFS_PACK         0x11d
158 #define OBD_FAIL_MDS_STATFS_NET          0x11e
159 #define OBD_FAIL_MDS_GETATTR_NAME_NET    0x11f
160 #define OBD_FAIL_MDS_PIN_NET             0x120
161 #define OBD_FAIL_MDS_UNPIN_NET           0x121
162 #define OBD_FAIL_MDS_ALL_REPLY_NET       0x122
163 #define OBD_FAIL_MDS_ALL_REQUEST_NET     0x123
164 #define OBD_FAIL_MDS_SYNC_NET            0x124
165 #define OBD_FAIL_MDS_SYNC_PACK           0x125
166 #define OBD_FAIL_MDS_DONE_WRITING_NET    0x126
167 #define OBD_FAIL_MDS_DONE_WRITING_PACK   0x127
168 #define OBD_FAIL_MDS_ALLOC_OBDO          0x128
169 #define OBD_FAIL_MDS_PAUSE_OPEN          0x129
170 #define OBD_FAIL_MDS_STATFS_LCW_SLEEP    0x12a
171 #define OBD_FAIL_MDS_OPEN_CREATE         0x12b
172 #define OBD_FAIL_MDS_OST_SETATTR         0x12c
173 #define OBD_FAIL_MDS_QUOTACHECK_NET      0x12d
174 #define OBD_FAIL_MDS_QUOTACTL_NET        0x12e
175 #define OBD_FAIL_MDS_CLIENT_ADD          0x12f
176 #define OBD_FAIL_MDS_GETXATTR_NET        0x130
177 #define OBD_FAIL_MDS_GETXATTR_PACK       0x131
178 #define OBD_FAIL_MDS_SETXATTR_NET        0x132
179 #define OBD_FAIL_MDS_SETXATTR            0x133
180 #define OBD_FAIL_MDS_SETXATTR_WRITE      0x134
181 #define OBD_FAIL_MDS_FS_SETUP            0x135
182 #define OBD_FAIL_MDS_RESEND              0x136
183 #define OBD_FAIL_MDS_IS_SUBDIR_NET       0x137
184 #define OBD_FAIL_MDS_IS_SUBDIR_PACK      0x138
185 #define OBD_FAIL_MDS_SET_INFO_NET        0x139
186 #define OBD_FAIL_MDS_WRITEPAGE_NET       0x13a
187 #define OBD_FAIL_MDS_WRITEPAGE_PACK      0x13b
188 #define OBD_FAIL_MDS_LLOG_CREATE_FAILED  0x13c
189 #define OBD_FAIL_MDS_OSC_PRECREATE       0x13d
190 #define OBD_FAIL_MDS_LOV_SYNC_RACE       0x13e
191 #define OBD_FAIL_MDS_CLOSE_NET_REP       0x13f
192 #define OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT   0x140
193
194 #define OBD_FAIL_OST                     0x200
195 #define OBD_FAIL_OST_CONNECT_NET         0x201
196 #define OBD_FAIL_OST_DISCONNECT_NET      0x202
197 #define OBD_FAIL_OST_GET_INFO_NET        0x203
198 #define OBD_FAIL_OST_CREATE_NET          0x204
199 #define OBD_FAIL_OST_DESTROY_NET         0x205
200 #define OBD_FAIL_OST_GETATTR_NET         0x206
201 #define OBD_FAIL_OST_SETATTR_NET         0x207
202 #define OBD_FAIL_OST_OPEN_NET            0x208
203 #define OBD_FAIL_OST_CLOSE_NET           0x209
204 #define OBD_FAIL_OST_BRW_NET             0x20a
205 #define OBD_FAIL_OST_PUNCH_NET           0x20b
206 #define OBD_FAIL_OST_STATFS_NET          0x20c
207 #define OBD_FAIL_OST_HANDLE_UNPACK       0x20d
208 #define OBD_FAIL_OST_BRW_WRITE_BULK      0x20e
209 #define OBD_FAIL_OST_BRW_READ_BULK       0x20f
210 #define OBD_FAIL_OST_SYNC_NET            0x210
211 #define OBD_FAIL_OST_ALL_REPLY_NET       0x211
212 #define OBD_FAIL_OST_ALL_REQUEST_NET     0x212
213 #define OBD_FAIL_OST_LDLM_REPLY_NET      0x213
214 #define OBD_FAIL_OST_BRW_PAUSE_BULK      0x214
215 #define OBD_FAIL_OST_ENOSPC              0x215
216 #define OBD_FAIL_OST_EROFS               0x216
217 #define OBD_FAIL_OST_ENOENT              0x217
218 #define OBD_FAIL_OST_QUOTACHECK_NET      0x218
219 #define OBD_FAIL_OST_QUOTACTL_NET        0x219
220 #define OBD_FAIL_OST_CHECKSUM_RECEIVE    0x21a
221 #define OBD_FAIL_OST_CHECKSUM_SEND       0x21b
222 #define OBD_FAIL_OST_BRW_SIZE            0x21c
223 #define OBD_FAIL_OST_DROP_REQ            0x21d
224 #define OBD_FAIL_OST_SETATTR_CREDITS     0x21e
225 #define OBD_FAIL_OST_HOLD_WRITE_RPC      0x21f
226 #define OBD_FAIL_OST_BRW_WRITE_BULK2     0x220
227 #define OBD_FAIL_OST_LLOG_RECOVERY_TIMEOUT 0x221
228 #define OBD_FAIL_OST_CANCEL_COOKIE_TIMEOUT 0x222
229 #define OBD_FAIL_OST_PAUSE_CREATE        0x223
230 #define OBD_FAIL_OST_BRW_PAUSE_PACK      0x224
231 #define OBD_FAIL_OST_CONNECT_NET2        0x225
232
233 #define OBD_FAIL_LDLM                    0x300
234 #define OBD_FAIL_LDLM_NAMESPACE_NEW      0x301
235 #define OBD_FAIL_LDLM_ENQUEUE            0x302
236 #define OBD_FAIL_LDLM_CONVERT            0x303
237 #define OBD_FAIL_LDLM_CANCEL             0x304
238 #define OBD_FAIL_LDLM_BL_CALLBACK        0x305
239 #define OBD_FAIL_LDLM_CP_CALLBACK        0x306
240 #define OBD_FAIL_LDLM_GL_CALLBACK        0x307
241 #define OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR 0x308
242 #define OBD_FAIL_LDLM_ENQUEUE_INTENT_ERR 0x309
243 #define OBD_FAIL_LDLM_CREATE_RESOURCE    0x30a
244 #define OBD_FAIL_LDLM_ENQUEUE_BLOCKED    0x30b
245 #define OBD_FAIL_LDLM_REPLY              0x30c
246 #define OBD_FAIL_LDLM_RECOV_CLIENTS      0x30d
247 #define OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT 0x30e
248 #define OBD_FAIL_LDLM_GLIMPSE            0x30f
249 #define OBD_FAIL_LDLM_CANCEL_RACE        0x310
250 #define OBD_FAIL_LDLM_CANCEL_EVICT_RACE  0x311
251 #define OBD_FAIL_LDLM_PAUSE_CANCEL       0x312
252 #define OBD_FAIL_LDLM_CLOSE_THREAD       0x313
253 #define OBD_FAIL_LDLM_CANCEL_BL_CB_RACE  0x314
254
255 #define OBD_FAIL_OSC                     0x400
256 #define OBD_FAIL_OSC_BRW_READ_BULK       0x401
257 #define OBD_FAIL_OSC_BRW_WRITE_BULK      0x402
258 #define OBD_FAIL_OSC_LOCK_BL_AST         0x403
259 #define OBD_FAIL_OSC_LOCK_CP_AST         0x404
260 #define OBD_FAIL_OSC_MATCH               0x405
261 #define OBD_FAIL_OSC_BRW_PREP_REQ        0x406
262 #define OBD_FAIL_OSC_SHUTDOWN            0x407
263 #define OBD_FAIL_OSC_CHECKSUM_RECEIVE    0x408
264 #define OBD_FAIL_OSC_CHECKSUM_SEND       0x409
265 #define OBD_FAIL_OSC_BRW_PREP_REQ2       0x40a
266 #define OBD_FAIL_OSC_CONNECT_CKSUM       0x40b
267 #define OBD_FAIL_OSC_CKSUM_ADLER_ONLY    0x40c
268 #define OBD_FAIL_OSC_DIO_PAUSE           0x40d
269
270 #define OBD_FAIL_PTLRPC                  0x500
271 #define OBD_FAIL_PTLRPC_ACK              0x501
272 #define OBD_FAIL_PTLRPC_RQBD             0x502
273 #define OBD_FAIL_PTLRPC_BULK_GET_NET     0x503
274 #define OBD_FAIL_PTLRPC_BULK_PUT_NET     0x504
275 #define OBD_FAIL_PTLRPC_DROP_RPC         0x505
276 #define OBD_FAIL_PTLRPC_DELAY_SEND       0x506
277 #define OBD_FAIL_PTLRPC_DELAY_RECOV      0x507
278 #define OBD_FAIL_PTLRPC_CLIENT_BULK_CB   0x508
279 #define OBD_FAIL_PTLRPC_PAUSE_REQ        0x50a
280 #define OBD_FAIL_PTLRPC_PAUSE_REP        0x50c
281
282 #define OBD_FAIL_OBD_PING_NET            0x600
283 #define OBD_FAIL_OBD_LOG_CANCEL_NET      0x601
284 #define OBD_FAIL_OBD_LOGD_NET            0x602
285 #define OBD_FAIL_OBD_QC_CALLBACK_NET     0x603
286 #define OBD_FAIL_OBD_DQACQ               0x604
287
288 #define OBD_FAIL_TGT_REPLY_NET           0x700
289 #define OBD_FAIL_TGT_CONN_RACE           0x701
290 #define OBD_FAIL_TGT_FORCE_RECONNECT     0x702
291 #define OBD_FAIL_TGT_DELAY_CONNECT       0x703
292 #define OBD_FAIL_TGT_DELAY_RECONNECT     0x704
293 #define OBD_FAIL_TGT_DELAY_PRECREATE     0x705
294 #define OBD_FAIL_TGT_TOOMANY_THREADS     0x706
295 #define OBD_FAIL_TGT_REPLAY_DROP         0x707
296
297 #define OBD_FAIL_MDC_REVALIDATE_PAUSE    0x800
298 #define OBD_FAIL_MDC_ENQUEUE_PAUSE       0x801
299 #define OBD_FAIL_MDC_OLD_EXT_FLAGS       0x802
300 #define OBD_FAIL_MDC_GETATTR_ENQUEUE     0x803
301
302 #define OBD_FAIL_MGS                     0x900
303 #define OBD_FAIL_MGS_ALL_REQUEST_NET     0x901
304 #define OBD_FAIL_MGS_ALL_REPLY_NET       0x902
305 #define OBD_FAIL_MGC_PAUSE_PROCESS_LOG   0x903
306 #define OBD_FAIL_MGS_PAUSE_REQ           0x904
307 #define OBD_FAIL_MGS_PAUSE_TARGET_REG    0x905
308
309 #define OBD_FAIL_QUOTA_QD_COUNT_32BIT    0xA00
310
311 #define OBD_FAIL_LPROC_REMOVE            0xB00
312
313 #define OBD_FAIL_GENERAL_ALLOC           0xC00
314
315 #define OBD_FAIL_SEQ                     0x1000
316 #define OBD_FAIL_SEQ_QUERY_NET           0x1001
317
318 #define OBD_FAIL_FLD                     0x1100
319 #define OBD_FAIL_FLD_QUERY_NET           0x1101
320
321 #define OBD_FAIL_SEC_CTX                 0x1200
322 #define OBD_FAIL_SEC_CTX_INIT_NET        0x1201
323 #define OBD_FAIL_SEC_CTX_INIT_CONT_NET   0x1202
324 #define OBD_FAIL_SEC_CTX_FINI_NET        0x1203
325 #define OBD_FAIL_SEC_CTX_HDL_PAUSE       0x1204
326
327 /* Failure injection control */
328 #define OBD_FAIL_MASK_SYS    0x0000FF00
329 #define OBD_FAIL_MASK_LOC   (0x000000FF | OBD_FAIL_MASK_SYS)
330
331 #define OBD_FAILED_BIT       30
332 /* OBD_FAILED is 0x40000000 */
333 #define OBD_FAILED          (1 << OBD_FAILED_BIT)
334
335 #define OBD_FAIL_ONCE_BIT    31
336 /* OBD_FAIL_ONCE is 0x80000000 */
337 #define OBD_FAIL_ONCE       (1 << OBD_FAIL_ONCE_BIT)
338
339 /* The following flags aren't made to be combined */
340 #define OBD_FAIL_SKIP        0x20000000 /* skip N times then fail */
341 #define OBD_FAIL_SOME        0x10000000 /* only fail N times */
342 #define OBD_FAIL_RAND        0x08000000 /* fail 1/N of the times */
343 #define OBD_FAIL_USR1        0x04000000 /* user flag */
344
345 #define OBD_FAIL_PRECHECK(id) (obd_fail_loc &&                                \
346                               (obd_fail_loc & OBD_FAIL_MASK_LOC) ==           \
347                               ((id) & OBD_FAIL_MASK_LOC))
348
349 static inline int obd_fail_check_set(__u32 id, __u32 value, int set)
350 {
351         int ret = 0;
352         if (unlikely(OBD_FAIL_PRECHECK(id) &&
353             (ret = __obd_fail_check_set(id, value, set)))) {
354                 CERROR("*** obd_fail_loc=%x ***\n", id);
355         }
356         return ret;
357 }
358
359 /* If id hit obd_fail_loc, return 1, otherwise return 0 */
360 #define OBD_FAIL_CHECK(id) \
361         obd_fail_check_set(id, 0, OBD_FAIL_LOC_NOSET)
362
363 /* If id hit obd_fail_loc, obd_fail_loc |= value and return 1,
364  * otherwise return 0 */
365 #define OBD_FAIL_CHECK_ORSET(id, value) \
366         obd_fail_check_set(id, value, OBD_FAIL_LOC_ORSET)
367
368 /* If id hit obd_fail_loc, obd_fail_loc = value and return 1,
369  * otherwise return 0 */
370 #define OBD_FAIL_CHECK_RESET(id, value) \
371         obd_fail_check_set(id, value, OBD_FAIL_LOC_RESET)
372
373
374 static inline int obd_fail_timeout_set(__u32 id, __u32 value, int ms, int set)
375 {
376         if (unlikely(OBD_FAIL_PRECHECK(id)))
377                 return __obd_fail_timeout_set(id, value, ms, set);
378         else
379                 return 0;
380 }
381
382 /* If id hit obd_fail_loc, sleep for seconds or milliseconds */
383 #define OBD_FAIL_TIMEOUT(id, secs) \
384         obd_fail_timeout_set(id, 0, secs * 1000, OBD_FAIL_LOC_NOSET)
385
386 #define OBD_FAIL_TIMEOUT_MS(id, ms) \
387         obd_fail_timeout_set(id, 0, ms, OBD_FAIL_LOC_NOSET)
388
389 /* If id hit obd_fail_loc, obd_fail_loc |= value and
390  * sleep seconds or milliseconds */
391 #define OBD_FAIL_TIMEOUT_ORSET(id, value, secs) \
392         obd_fail_timeout_set(id, value, secs * 1000, OBD_FAIL_LOC_ORSET)
393
394 #define OBD_FAIL_TIMEOUT_MS_ORSET(id, value, ms) \
395         obd_fail_timeout_set(id, value, ms, OBD_FAIL_LOC_ORSET)
396
397 #ifdef __KERNEL__
398 static inline void obd_fail_write(int id, struct super_block *sb)
399 {
400         /* We set FAIL_ONCE because we never "un-fail" a device */
401         if (OBD_FAIL_CHECK_ORSET(id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
402 #ifdef LIBCFS_DEBUG
403                 BDEVNAME_DECLARE_STORAGE(tmp);
404                 CERROR("obd_fail_loc=%x, fail write operation on %s\n",
405                        id, ll_bdevname(sb, tmp));
406 #endif
407                 /* TODO-CMD: fix getting jdev */
408                 __lvfs_set_rdonly(lvfs_sbdev(sb), (lvfs_sbdev_type)0);
409         }
410 }
411 #define OBD_FAIL_WRITE(id, sb) obd_fail_write(id, sb)
412
413 /* The idea here is to synchronise two threads to force a race. The
414  * first thread that calls this with a matching fail_loc is put to
415  * sleep. The next thread that calls with the same fail_loc wakes up
416  * the first and continues. */
417 static inline void obd_race(__u32 id)
418 {
419         if (OBD_FAIL_PRECHECK(id)) {
420                 if (unlikely(__obd_fail_check_set(id, 0, OBD_FAIL_LOC_NOSET))) {
421                         obd_race_state = 0;
422                         CERROR("obd_race id %x sleeping\n", id);
423                         OBD_SLEEP_ON(obd_race_waitq, obd_race_state != 0);
424                         CERROR("obd_fail_race id %x awake\n", id);
425                 } else {
426                         CERROR("obd_fail_race id %x waking\n", id);
427                         obd_race_state = 1;
428                         wake_up(&obd_race_waitq);
429                 }
430         }
431 }
432 #define OBD_RACE(id) obd_race(id)
433 #else
434 /* sigh.  an expedient fix until OBD_RACE is fixed up */
435 #define OBD_RACE(foo) do {} while(0)
436 #endif
437
438 #define fixme() CDEBUG(D_OTHER, "FIXME\n");
439
440 extern atomic_t libcfs_kmemory;
441
442 #ifdef LPROCFS
443 #define obd_memory_add(size)                                                  \
444         lprocfs_counter_add(obd_memory, OBD_MEMORY_STAT, (long)(size))
445 #define obd_memory_sub(size)                                                  \
446         lprocfs_counter_sub(obd_memory, OBD_MEMORY_STAT, (long)(size))
447 #define obd_memory_sum()                                                      \
448         lprocfs_stats_collector(obd_memory, OBD_MEMORY_STAT,                  \
449                                 LPROCFS_FIELDS_FLAGS_SUM)
450 #define obd_pages_add(order)                                                  \
451         lprocfs_counter_add(obd_memory, OBD_MEMORY_PAGES_STAT,                \
452                             (long)(1 << (order)))
453 #define obd_pages_sub(order)                                                  \
454         lprocfs_counter_sub(obd_memory, OBD_MEMORY_PAGES_STAT,                \
455                             (long)(1 << (order)))
456 #define obd_pages_sum()                                                       \
457         lprocfs_stats_collector(obd_memory, OBD_MEMORY_PAGES_STAT,            \
458                                 LPROCFS_FIELDS_FLAGS_SUM)
459
460 extern void obd_update_maxusage(void);
461 extern __u64 obd_memory_max(void);
462 extern __u64 obd_pages_max(void);
463
464 #else
465
466 extern __u64 obd_alloc;
467 extern __u64 obd_pages;
468
469 extern __u64 obd_max_alloc;
470 extern __u64 obd_max_pages;
471
472 static inline void obd_memory_add(long size)
473 {
474         obd_alloc += size;
475         if (obd_alloc > obd_max_alloc)
476                 obd_max_alloc = obd_alloc;
477 }
478
479 static inline void obd_memory_sub(long size)
480 {
481         obd_alloc -= size;
482 }
483
484 static inline void obd_pages_add(int order)
485 {
486         obd_pages += 1<< order;
487         if (obd_pages > obd_max_pages)
488                 obd_max_pages = obd_pages;
489 }
490
491 static inline void obd_pages_sub(int order)
492 {
493         obd_pages -= 1<< order;
494 }
495
496 #define obd_memory_sum() (obd_alloc)
497 #define obd_pages_sum()  (obd_pages)
498
499 #define obd_memory_max() (obd_max_alloc)
500 #define obd_pages_max() (obd_max_pages)
501
502 #endif
503
504 #define OBD_DEBUG_MEMUSAGE (1)
505
506 #if OBD_DEBUG_MEMUSAGE
507 #define OBD_ALLOC_POST(ptr, size, name)                                 \
508                 obd_memory_add(size);                                   \
509                 CDEBUG(D_MALLOC, name " '" #ptr "': %d at %p.\n",       \
510                        (int)(size), ptr)
511
512 #define OBD_FREE_PRE(ptr, size, name)                                   \
513         LASSERT(ptr);                                                   \
514         obd_memory_sub(size);                                           \
515         CDEBUG(D_MALLOC, name " '" #ptr "': %d at %p.\n",               \
516                (int)(size), ptr);                                       \
517         POISON(ptr, 0x5a, size)
518
519 #else /* !OBD_DEBUG_MEMUSAGE */
520
521 #define OBD_ALLOC_POST(ptr, size, name) ((void)0)
522 #define OBD_FREE_PRE(ptr, size, name)   ((void)0)
523
524 #endif /* !OBD_DEBUG_MEMUSAGE */
525
526 #ifdef RANDOM_FAIL_ALLOC
527 #define HAS_FAIL_ALLOC_FLAG OBD_FAIL_CHECK(OBD_FAIL_GENERAL_ALLOC)
528 #else
529 #define HAS_FAIL_ALLOC_FLAG 0
530 #endif
531
532 #define OBD_ALLOC_FAIL_BITS 24
533 #define OBD_ALLOC_FAIL_MASK ((1 << OBD_ALLOC_FAIL_BITS) - 1)
534 #define OBD_ALLOC_FAIL_MULT (OBD_ALLOC_FAIL_MASK / 100)
535
536 #if defined(LUSTRE_UTILS) /* this version is for utils only */
537 #define OBD_ALLOC_GFP(ptr, size, gfp_mask)                                    \
538 do {                                                                          \
539         (ptr) = cfs_alloc(size, (gfp_mask));                                  \
540         if (unlikely((ptr) == NULL)) {                                        \
541                 CERROR("kmalloc of '" #ptr "' (%d bytes) failed at %s:%d\n",  \
542                        (int)(size), __FILE__, __LINE__);                      \
543         } else {                                                              \
544                 memset(ptr, 0, size);                                         \
545                 CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p\n",          \
546                        (int)(size), ptr);                                     \
547         }                                                                     \
548 } while (0)
549 #else /* this version is for the kernel and liblustre */
550 #define OBD_FREE_RTN0(ptr)                                                    \
551 ({                                                                            \
552         cfs_free(ptr);                                                        \
553         (ptr) = NULL;                                                         \
554         0;                                                                    \
555 })
556 #define OBD_ALLOC_GFP(ptr, size, gfp_mask)                                    \
557 do {                                                                          \
558         (ptr) = cfs_alloc(size, (gfp_mask));                                  \
559         if (likely((ptr) != NULL &&                                           \
560                    (!HAS_FAIL_ALLOC_FLAG || obd_alloc_fail_rate == 0 ||       \
561                     !obd_alloc_fail(ptr, #ptr, "km", size,                    \
562                                     __FILE__, __LINE__) ||                    \
563                     OBD_FREE_RTN0(ptr)))){                                    \
564                 memset(ptr, 0, size);                                         \
565                 OBD_ALLOC_POST(ptr, size, "kmalloced");                       \
566         }                                                                     \
567 } while (0)
568 #endif
569
570 #ifndef OBD_ALLOC_MASK
571 # define OBD_ALLOC_MASK CFS_ALLOC_IO
572 #endif
573
574 #define OBD_ALLOC(ptr, size) OBD_ALLOC_GFP(ptr, size, OBD_ALLOC_MASK)
575 #define OBD_ALLOC_WAIT(ptr, size) OBD_ALLOC_GFP(ptr, size, CFS_ALLOC_STD)
576 #define OBD_ALLOC_PTR(ptr) OBD_ALLOC(ptr, sizeof *(ptr))
577 #define OBD_ALLOC_PTR_WAIT(ptr) OBD_ALLOC_WAIT(ptr, sizeof *(ptr))
578
579 #ifdef __arch_um__
580 # define OBD_VMALLOC(ptr, size) OBD_ALLOC(ptr, size)
581 #else
582 # define OBD_VMALLOC(ptr, size)                                               \
583 do {                                                                          \
584         (ptr) = cfs_alloc_large(size);                                        \
585         if (unlikely((ptr) == NULL)) {                                        \
586                 CERROR("vmalloc of '" #ptr "' (%d bytes) failed\n",           \
587                        (int)(size));                                          \
588                 CERROR(LPU64" total bytes allocated by Lustre, %d by LNET\n", \
589                        obd_memory_sum(), atomic_read(&libcfs_kmemory));       \
590         } else {                                                              \
591                 memset(ptr, 0, size);                                         \
592                 OBD_ALLOC_POST(ptr, size, "vmalloced");                       \
593         }                                                                     \
594 } while(0)
595 #endif
596
597 #ifdef CONFIG_DEBUG_SLAB
598 #define POISON(ptr, c, s) do {} while (0)
599 #define POISON_PTR(ptr)  ((void)0)
600 #else
601 #define POISON(ptr, c, s) memset(ptr, c, s)
602 #define POISON_PTR(ptr)  (ptr) = (void *)0xdeadbeef
603 #endif
604
605 #ifdef POISON_BULK
606 #define POISON_PAGE(page, val) do { memset(kmap(page), val, CFS_PAGE_SIZE);   \
607                                     kunmap(page); } while (0)
608 #else
609 #define POISON_PAGE(page, val) do { } while (0)
610 #endif
611
612 #ifdef __KERNEL__
613 #define OBD_FREE(ptr, size)                                                   \
614 do {                                                                          \
615         OBD_FREE_PRE(ptr, size, "kfreed");                                    \
616         cfs_free(ptr);                                                        \
617         POISON_PTR(ptr);                                                      \
618 } while(0)
619
620
621 #ifdef HAVE_RCU
622 # ifdef HAVE_CALL_RCU_PARAM
623 #  define my_call_rcu(rcu, cb)            call_rcu(rcu, cb, rcu)
624 # else
625 #  define my_call_rcu(rcu, cb)            call_rcu(rcu, cb)
626 # endif
627 #else
628 # define my_call_rcu(rcu, cb)             (cb)(rcu)
629 #endif
630
631 #define OBD_FREE_RCU_CB(ptr, size, handle, free_cb)                           \
632 do {                                                                          \
633         struct portals_handle *__h = (handle);                                \
634         LASSERT(handle);                                                      \
635         __h->h_ptr = (ptr);                                                   \
636         __h->h_size = (size);                                                 \
637         __h->h_free_cb = (void (*)(void *, size_t))(free_cb);                 \
638         my_call_rcu(&__h->h_rcu, class_handle_free_cb);                       \
639         POISON_PTR(ptr);                                                      \
640 } while(0)
641 #define OBD_FREE_RCU(ptr, size, handle) OBD_FREE_RCU_CB(ptr, size, handle, NULL)
642
643 #else
644 #define OBD_FREE(ptr, size) ((void)(size), free((ptr)))
645 #define OBD_FREE_RCU(ptr, size, handle) (OBD_FREE(ptr, size))
646 #define OBD_FREE_RCU_CB(ptr, size, handle, cb)     ((*(cb))(ptr, size))
647 #endif /* ifdef __KERNEL__ */
648
649 #ifdef __arch_um__
650 # define OBD_VFREE(ptr, size) OBD_FREE(ptr, size)
651 #else
652 # define OBD_VFREE(ptr, size)                                                 \
653 do {                                                                          \
654         OBD_FREE_PRE(ptr, size, "vfreed");                                    \
655         cfs_free_large(ptr);                                                  \
656         POISON_PTR(ptr);                                                      \
657 } while(0)
658 #endif
659
660 /* we memset() the slab object to 0 when allocation succeeds, so DO NOT
661  * HAVE A CTOR THAT DOES ANYTHING.  its work will be cleared here.  we'd
662  * love to assert on that, but slab.c keeps kmem_cache_s all to itself. */
663 #define OBD_SLAB_FREE_RTN0(ptr, slab)                                         \
664 ({                                                                            \
665         cfs_mem_cache_free((slab), (ptr));                                    \
666         (ptr) = NULL;                                                         \
667         0;                                                                    \
668 })
669 #define OBD_SLAB_ALLOC(ptr, slab, type, size)                                 \
670 do {                                                                          \
671         LASSERT(!in_interrupt());                                             \
672         (ptr) = cfs_mem_cache_alloc(slab, (type));                            \
673         if (likely((ptr) != NULL &&                                           \
674                    (!HAS_FAIL_ALLOC_FLAG || obd_alloc_fail_rate == 0 ||       \
675                     !obd_alloc_fail(ptr, #ptr, "slab-", size,                 \
676                                     __FILE__, __LINE__) ||                    \
677                     OBD_SLAB_FREE_RTN0(ptr, slab)))) {                        \
678                 memset(ptr, 0, size);                                         \
679                 OBD_ALLOC_POST(ptr, size, "slab-alloced");                    \
680         }                                                                     \
681 } while(0)
682
683 #define OBD_FREE_PTR(ptr) OBD_FREE(ptr, sizeof *(ptr))
684
685 #define OBD_SLAB_FREE(ptr, slab, size)                                        \
686 do {                                                                          \
687         OBD_FREE_PRE(ptr, size, "slab-freed");                                \
688         cfs_mem_cache_free(slab, ptr);                                        \
689         POISON_PTR(ptr);                                                      \
690 } while(0)
691
692 #define OBD_SLAB_ALLOC_PTR(ptr, slab)                                         \
693         OBD_SLAB_ALLOC((ptr), (slab), CFS_ALLOC_STD, sizeof *(ptr))
694 #define OBD_SLAB_FREE_PTR(ptr, slab)                                          \
695         OBD_SLAB_FREE((ptr), (slab), sizeof *(ptr))
696
697 #define KEY_IS(str) \
698         (keylen >= (sizeof(str)-1) && memcmp(key, str, (sizeof(str)-1)) == 0)
699
700 /* Wrapper for contiguous page frame allocation */
701 #define OBD_PAGES_ALLOC(ptr, order, gfp_mask)                                 \
702 do {                                                                          \
703         (ptr) = cfs_alloc_pages(gfp_mask, order);                             \
704         if (unlikely((ptr) == NULL)) {                                        \
705                 CERROR("alloc_pages of '" #ptr "' %d page(s) / "LPU64" bytes "\
706                        "failed\n", (int)(1 << (order)),                       \
707                        (__u64)((1 << (order)) << CFS_PAGE_SHIFT));            \
708                 CERROR(LPU64" total bytes and "LPU64" total pages "           \
709                        "("LPU64" bytes) allocated by Lustre, "                \
710                        "%d total bytes by LNET\n",                            \
711                        obd_memory_sum(),                                      \
712                        obd_pages_sum() << CFS_PAGE_SHIFT,                     \
713                        obd_pages_sum(),                                       \
714                        atomic_read(&libcfs_kmemory));                         \
715         } else {                                                              \
716                 obd_pages_add(order);                                         \
717                 CDEBUG(D_MALLOC, "alloc_pages '" #ptr "': %d page(s) / "      \
718                        LPU64" bytes at %p.\n",                                \
719                        (int)(1 << (order)),                                   \
720                        (__u64)((1 << (order)) << CFS_PAGE_SHIFT), ptr);       \
721         }                                                                     \
722 } while (0)
723
724 #define OBD_PAGE_ALLOC(ptr, gfp_mask)                                         \
725         OBD_PAGES_ALLOC(ptr, 0, gfp_mask)
726
727 #define OBD_PAGES_FREE(ptr, order)                                            \
728 do {                                                                          \
729         LASSERT(ptr);                                                         \
730         obd_pages_sub(order);                                                 \
731         CDEBUG(D_MALLOC, "free_pages '" #ptr "': %d page(s) / "LPU64" bytes " \
732                "at %p.\n",                                                    \
733                (int)(1 << (order)), (__u64)((1 << (order)) << CFS_PAGE_SHIFT),\
734                ptr);                                                          \
735         __cfs_free_pages(ptr, order);                                         \
736         (ptr) = (void *)0xdeadbeef;                                           \
737 } while (0)
738
739 #define OBD_PAGE_FREE(ptr) OBD_PAGES_FREE(ptr, 0)
740
741 #endif