Whamcloud - gitweb
Branch HEAD
[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  *  Copyright (C) 2001, 2002 Cluster File Systems, Inc.
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  */
22
23 #ifndef _OBD_SUPPORT
24 #define _OBD_SUPPORT
25
26 #include <libcfs/libcfs.h>
27 #include <lvfs.h>
28 #include <lprocfs_status.h>
29
30 #if defined(__linux__)
31 #include <linux/obd_support.h>
32 #elif defined(__APPLE__)
33 #include <darwin/obd_support.h>
34 #elif defined(__WINNT__)
35 #include <winnt/obd_support.h>
36 #else
37 #error Unsupported operating system.
38 #endif
39
40 /* global variables */
41 extern struct lprocfs_stats *obd_memory;
42 enum {
43         OBD_MEMORY_STAT = 0,
44         OBD_MEMORY_PAGES_STAT = 1,
45         OBD_STATS_NUM,
46 };
47
48 enum {
49         OBD_FAIL_LOC_NOSET      = 0,
50         OBD_FAIL_LOC_ORSET      = 1,
51         OBD_FAIL_LOC_RESET      = 2
52 };
53
54 extern unsigned long obd_fail_loc;
55 extern unsigned int obd_fail_val;
56 extern unsigned int obd_debug_peer_on_timeout;
57 extern unsigned int obd_dump_on_timeout;
58 extern unsigned int obd_dump_on_eviction;
59 /* obd_timeout should only be used for recovery, not for
60    networking / disk / timings affected by load (use Adaptive Timeouts) */
61 extern unsigned int obd_timeout;          /* seconds */
62 extern unsigned int ldlm_timeout;         /* seconds */
63 extern unsigned int obd_sync_filter;
64 extern unsigned int obd_max_dirty_pages;
65 extern atomic_t obd_dirty_pages;
66 extern cfs_waitq_t obd_race_waitq;
67 extern int obd_race_state;
68 extern unsigned int obd_alloc_fail_rate;
69
70 int __obd_fail_check_set(__u32 id, __u32 value, int set);
71 int __obd_fail_timeout_set(__u32 id, __u32 value, int ms, int set);
72
73 /* lvfs.c */
74 int obd_alloc_fail(const void *ptr, const char *name, const char *type,
75                    size_t size, const char *file, int line);
76
77 /* Timeout definitions */
78 #define OBD_TIMEOUT_DEFAULT             100
79 #define LDLM_TIMEOUT_DEFAULT            20
80 #define MDS_LDLM_TIMEOUT_DEFAULT        6
81 /* Time to wait for all clients to reconnect during recovery */
82 /* Should be very conservative; must catch the first reconnect after reboot */
83 #define OBD_RECOVERY_FACTOR (3) /* times obd_timeout */
84 /* Change recovery-small 26b time if you change this */
85 #define PING_INTERVAL max(obd_timeout / 4, 1U)
86 /* Client may skip 1 ping; we must wait at least 2.5. But for multiple
87  * failover targets the client only pings one server at a time, and pings
88  * can be lost on a loaded network. Since eviction has serious consequences,
89  * and there's no urgent need to evict a client just because it's idle, we
90  * should be very conservative here. */
91 #define PING_EVICT_TIMEOUT (PING_INTERVAL * 6)
92 #define DISK_TIMEOUT 50          /* Beyond this we warn about disk speed */
93 #define CONNECTION_SWITCH_MIN 5U /* Connection switching rate limiter */
94  /* Max connect interval for nonresponsive servers; ~50s to avoid building up
95     connect requests in the LND queues, but within obd_timeout so we don't
96     miss the recovery window */
97 #define CONNECTION_SWITCH_MAX min(50U, max(CONNECTION_SWITCH_MIN,obd_timeout))
98 #define CONNECTION_SWITCH_INC 5  /* Connection timeout backoff */
99 #ifndef CRAY_XT3
100 /* In general this should be low to have quick detection of a system
101    running on a backup server. (If it's too low, import_select_connection
102    will increase the timeout anyhow.)  */
103 #define INITIAL_CONNECT_TIMEOUT max(CONNECTION_SWITCH_MIN,obd_timeout/20)
104 #else
105 /* ...but for very large systems (e.g. CRAY) we need to keep the initial
106    connect t.o. high (bz 10803), because they will nearly ALWAYS be doing the
107    connects for the first time (clients "reboot" after every process, so no
108    chance to generate adaptive timeout data. */
109 #define INITIAL_CONNECT_TIMEOUT max(CONNECTION_SWITCH_MIN,obd_timeout/2)
110 #endif
111 #define LONG_UNLINK 300          /* Unlink should happen before now */
112
113
114 #define OBD_FAIL_MDS                     0x100
115 #define OBD_FAIL_MDS_HANDLE_UNPACK       0x101
116 #define OBD_FAIL_MDS_GETATTR_NET         0x102
117 #define OBD_FAIL_MDS_GETATTR_PACK        0x103
118 #define OBD_FAIL_MDS_READPAGE_NET        0x104
119 #define OBD_FAIL_MDS_READPAGE_PACK       0x105
120 #define OBD_FAIL_MDS_SENDPAGE            0x106
121 #define OBD_FAIL_MDS_REINT_NET           0x107
122 #define OBD_FAIL_MDS_REINT_UNPACK        0x108
123 #define OBD_FAIL_MDS_REINT_SETATTR       0x109
124 #define OBD_FAIL_MDS_REINT_SETATTR_WRITE 0x10a
125 #define OBD_FAIL_MDS_REINT_CREATE        0x10b
126 #define OBD_FAIL_MDS_REINT_CREATE_WRITE  0x10c
127 #define OBD_FAIL_MDS_REINT_UNLINK        0x10d
128 #define OBD_FAIL_MDS_REINT_UNLINK_WRITE  0x10e
129 #define OBD_FAIL_MDS_REINT_LINK          0x10f
130 #define OBD_FAIL_MDS_REINT_LINK_WRITE    0x110
131 #define OBD_FAIL_MDS_REINT_RENAME        0x111
132 #define OBD_FAIL_MDS_REINT_RENAME_WRITE  0x112
133 #define OBD_FAIL_MDS_OPEN_NET            0x113
134 #define OBD_FAIL_MDS_OPEN_PACK           0x114
135 #define OBD_FAIL_MDS_CLOSE_NET           0x115
136 #define OBD_FAIL_MDS_CLOSE_PACK          0x116
137 #define OBD_FAIL_MDS_CONNECT_NET         0x117
138 #define OBD_FAIL_MDS_CONNECT_PACK        0x118
139 #define OBD_FAIL_MDS_REINT_NET_REP       0x119
140 #define OBD_FAIL_MDS_DISCONNECT_NET      0x11a
141 #define OBD_FAIL_MDS_GETSTATUS_NET       0x11b
142 #define OBD_FAIL_MDS_GETSTATUS_PACK      0x11c
143 #define OBD_FAIL_MDS_STATFS_PACK         0x11d
144 #define OBD_FAIL_MDS_STATFS_NET          0x11e
145 #define OBD_FAIL_MDS_GETATTR_NAME_NET    0x11f
146 #define OBD_FAIL_MDS_PIN_NET             0x120
147 #define OBD_FAIL_MDS_UNPIN_NET           0x121
148 #define OBD_FAIL_MDS_ALL_REPLY_NET       0x122
149 #define OBD_FAIL_MDS_ALL_REQUEST_NET     0x123
150 #define OBD_FAIL_MDS_SYNC_NET            0x124
151 #define OBD_FAIL_MDS_SYNC_PACK           0x125
152 #define OBD_FAIL_MDS_DONE_WRITING_NET    0x126
153 #define OBD_FAIL_MDS_DONE_WRITING_PACK   0x127
154 #define OBD_FAIL_MDS_ALLOC_OBDO          0x128
155 #define OBD_FAIL_MDS_PAUSE_OPEN          0x129
156 #define OBD_FAIL_MDS_STATFS_LCW_SLEEP    0x12a
157 #define OBD_FAIL_MDS_OPEN_CREATE         0x12b
158 #define OBD_FAIL_MDS_OST_SETATTR         0x12c
159 #define OBD_FAIL_MDS_QUOTACHECK_NET      0x12d
160 #define OBD_FAIL_MDS_QUOTACTL_NET        0x12e
161 #define OBD_FAIL_MDS_CLIENT_ADD          0x12f
162 #define OBD_FAIL_MDS_GETXATTR_NET        0x130
163 #define OBD_FAIL_MDS_GETXATTR_PACK       0x131
164 #define OBD_FAIL_MDS_SETXATTR_NET        0x132
165 #define OBD_FAIL_MDS_SETXATTR            0x133
166 #define OBD_FAIL_MDS_SETXATTR_WRITE      0x134
167 #define OBD_FAIL_MDS_FS_SETUP            0x135
168 #define OBD_FAIL_MDS_RESEND              0x136
169 #define OBD_FAIL_MDS_IS_SUBDIR_NET       0x137
170 #define OBD_FAIL_MDS_IS_SUBDIR_PACK      0x138
171 #define OBD_FAIL_MDS_SET_INFO_NET        0x139
172 #define OBD_FAIL_MDS_WRITEPAGE_NET       0x13a
173 #define OBD_FAIL_MDS_WRITEPAGE_PACK      0x13b
174 #define OBD_FAIL_MDS_LLOG_CREATE_FAILED  0x13c
175 #define OBD_FAIL_MDS_OSC_PRECREATE       0x13d
176 #define OBD_FAIL_MDS_LOV_SYNC_RACE       0x13e
177 #define OBD_FAIL_MDS_CLOSE_NET_REP       0x13f
178 #define OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT   0x140
179
180 #define OBD_FAIL_OST                     0x200
181 #define OBD_FAIL_OST_CONNECT_NET         0x201
182 #define OBD_FAIL_OST_DISCONNECT_NET      0x202
183 #define OBD_FAIL_OST_GET_INFO_NET        0x203
184 #define OBD_FAIL_OST_CREATE_NET          0x204
185 #define OBD_FAIL_OST_DESTROY_NET         0x205
186 #define OBD_FAIL_OST_GETATTR_NET         0x206
187 #define OBD_FAIL_OST_SETATTR_NET         0x207
188 #define OBD_FAIL_OST_OPEN_NET            0x208
189 #define OBD_FAIL_OST_CLOSE_NET           0x209
190 #define OBD_FAIL_OST_BRW_NET             0x20a
191 #define OBD_FAIL_OST_PUNCH_NET           0x20b
192 #define OBD_FAIL_OST_STATFS_NET          0x20c
193 #define OBD_FAIL_OST_HANDLE_UNPACK       0x20d
194 #define OBD_FAIL_OST_BRW_WRITE_BULK      0x20e
195 #define OBD_FAIL_OST_BRW_READ_BULK       0x20f
196 #define OBD_FAIL_OST_SYNC_NET            0x210
197 #define OBD_FAIL_OST_ALL_REPLY_NET       0x211
198 #define OBD_FAIL_OST_ALL_REQUEST_NET     0x212
199 #define OBD_FAIL_OST_LDLM_REPLY_NET      0x213
200 #define OBD_FAIL_OST_BRW_PAUSE_BULK      0x214
201 #define OBD_FAIL_OST_ENOSPC              0x215
202 #define OBD_FAIL_OST_EROFS               0x216
203 #define OBD_FAIL_OST_ENOENT              0x217
204 #define OBD_FAIL_OST_QUOTACHECK_NET      0x218
205 #define OBD_FAIL_OST_QUOTACTL_NET        0x219
206 #define OBD_FAIL_OST_CHECKSUM_RECEIVE    0x21a
207 #define OBD_FAIL_OST_CHECKSUM_SEND       0x21b
208 #define OBD_FAIL_OST_BRW_SIZE            0x21c
209 #define OBD_FAIL_OST_DROP_REQ            0x21d
210 #define OBD_FAIL_OST_SETATTR_CREDITS     0x21e
211 #define OBD_FAIL_OST_HOLD_WRITE_RPC      0x21f
212 #define OBD_FAIL_OST_BRW_WRITE_BULK2     0x220
213 #define OBD_FAIL_OST_LLOG_RECOVERY_TIMEOUT 0x221
214 #define OBD_FAIL_OST_CANCEL_COOKIE_TIMEOUT 0x222
215 #define OBD_FAIL_OST_PAUSE_CREATE        0x223
216 #define OBD_FAIL_OST_BRW_PAUSE_PACK      0x224
217 #define OBD_FAIL_OST_CONNECT_NET2        0x225
218
219 #define OBD_FAIL_LDLM                    0x300
220 #define OBD_FAIL_LDLM_NAMESPACE_NEW      0x301
221 #define OBD_FAIL_LDLM_ENQUEUE            0x302
222 #define OBD_FAIL_LDLM_CONVERT            0x303
223 #define OBD_FAIL_LDLM_CANCEL             0x304
224 #define OBD_FAIL_LDLM_BL_CALLBACK        0x305
225 #define OBD_FAIL_LDLM_CP_CALLBACK        0x306
226 #define OBD_FAIL_LDLM_GL_CALLBACK        0x307
227 #define OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR 0x308
228 #define OBD_FAIL_LDLM_ENQUEUE_INTENT_ERR 0x309
229 #define OBD_FAIL_LDLM_CREATE_RESOURCE    0x30a
230 #define OBD_FAIL_LDLM_ENQUEUE_BLOCKED    0x30b
231 #define OBD_FAIL_LDLM_REPLY              0x30c
232 #define OBD_FAIL_LDLM_RECOV_CLIENTS      0x30d
233 #define OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT 0x30e
234 #define OBD_FAIL_LDLM_GLIMPSE            0x30f
235 #define OBD_FAIL_LDLM_CANCEL_RACE        0x310
236 #define OBD_FAIL_LDLM_CANCEL_EVICT_RACE  0x311
237 #define OBD_FAIL_LDLM_PAUSE_CANCEL       0x312
238 #define OBD_FAIL_LDLM_CLOSE_THREAD       0x313
239 #define OBD_FAIL_LDLM_CANCEL_BL_CB_RACE  0x314
240
241 #define OBD_FAIL_OSC                     0x400
242 #define OBD_FAIL_OSC_BRW_READ_BULK       0x401
243 #define OBD_FAIL_OSC_BRW_WRITE_BULK      0x402
244 #define OBD_FAIL_OSC_LOCK_BL_AST         0x403
245 #define OBD_FAIL_OSC_LOCK_CP_AST         0x404
246 #define OBD_FAIL_OSC_MATCH               0x405
247 #define OBD_FAIL_OSC_BRW_PREP_REQ        0x406
248 #define OBD_FAIL_OSC_SHUTDOWN            0x407
249 #define OBD_FAIL_OSC_CHECKSUM_RECEIVE    0x408
250 #define OBD_FAIL_OSC_CHECKSUM_SEND       0x409
251 #define OBD_FAIL_OSC_BRW_PREP_REQ2       0x40a
252 #define OBD_FAIL_OSC_CONNECT_CKSUM       0x40b
253 #define OBD_FAIL_OSC_CKSUM_ADLER_ONLY    0x40c
254 #define OBD_FAIL_OSC_DIO_PAUSE           0x40d
255
256 #define OBD_FAIL_PTLRPC                  0x500
257 #define OBD_FAIL_PTLRPC_ACK              0x501
258 #define OBD_FAIL_PTLRPC_RQBD             0x502
259 #define OBD_FAIL_PTLRPC_BULK_GET_NET     0x503
260 #define OBD_FAIL_PTLRPC_BULK_PUT_NET     0x504
261 #define OBD_FAIL_PTLRPC_DROP_RPC         0x505
262 #define OBD_FAIL_PTLRPC_DELAY_SEND       0x506
263 #define OBD_FAIL_PTLRPC_DELAY_RECOV      0x507
264 #define OBD_FAIL_PTLRPC_CLIENT_BULK_CB   0x508
265 #define OBD_FAIL_PTLRPC_PAUSE_REQ        0x50a
266 #define OBD_FAIL_PTLRPC_PAUSE_REP        0x50c
267
268 #define OBD_FAIL_OBD_PING_NET            0x600
269 #define OBD_FAIL_OBD_LOG_CANCEL_NET      0x601
270 #define OBD_FAIL_OBD_LOGD_NET            0x602
271 #define OBD_FAIL_OBD_QC_CALLBACK_NET     0x603
272 #define OBD_FAIL_OBD_DQACQ               0x604
273
274 #define OBD_FAIL_TGT_REPLY_NET           0x700
275 #define OBD_FAIL_TGT_CONN_RACE           0x701
276 #define OBD_FAIL_TGT_FORCE_RECONNECT     0x702
277 #define OBD_FAIL_TGT_DELAY_CONNECT       0x703
278 #define OBD_FAIL_TGT_DELAY_RECONNECT     0x704
279 #define OBD_FAIL_TGT_DELAY_PRECREATE     0x705
280 #define OBD_FAIL_TGT_TOOMANY_THREADS     0x706
281 #define OBD_FAIL_TGT_REPLAY_DROP         0x707
282
283 #define OBD_FAIL_MDC_REVALIDATE_PAUSE    0x800
284 #define OBD_FAIL_MDC_ENQUEUE_PAUSE       0x801
285 #define OBD_FAIL_MDC_OLD_EXT_FLAGS       0x802
286 #define OBD_FAIL_MDC_GETATTR_ENQUEUE     0x803
287
288 #define OBD_FAIL_MGS                     0x900
289 #define OBD_FAIL_MGS_ALL_REQUEST_NET     0x901
290 #define OBD_FAIL_MGS_ALL_REPLY_NET       0x902
291 #define OBD_FAIL_MGC_PAUSE_PROCESS_LOG   0x903
292 #define OBD_FAIL_MGS_PAUSE_REQ           0x904
293 #define OBD_FAIL_MGS_PAUSE_TARGET_REG    0x905
294
295 #define OBD_FAIL_QUOTA_QD_COUNT_32BIT    0xA00
296
297 #define OBD_FAIL_LPROC_REMOVE            0xB00
298
299 #define OBD_FAIL_GENERAL_ALLOC           0xC00
300
301 #define OBD_FAIL_SEQ                     0x1000
302 #define OBD_FAIL_SEQ_QUERY_NET           0x1001
303
304 #define OBD_FAIL_FLD                     0x1100
305 #define OBD_FAIL_FLD_QUERY_NET           0x1101
306
307 #define OBD_FAIL_SEC_CTX                 0x1200
308 #define OBD_FAIL_SEC_CTX_INIT_NET        0x1201
309 #define OBD_FAIL_SEC_CTX_INIT_CONT_NET   0x1202
310 #define OBD_FAIL_SEC_CTX_FINI_NET        0x1203
311 #define OBD_FAIL_SEC_CTX_HDL_PAUSE       0x1204
312
313 /* Failure injection control */
314 #define OBD_FAIL_MASK_SYS    0x0000FF00
315 #define OBD_FAIL_MASK_LOC   (0x000000FF | OBD_FAIL_MASK_SYS)
316
317 #define OBD_FAILED_BIT       30
318 /* OBD_FAILED is 0x40000000 */
319 #define OBD_FAILED          (1 << OBD_FAILED_BIT)
320
321 #define OBD_FAIL_ONCE_BIT    31
322 /* OBD_FAIL_ONCE is 0x80000000 */
323 #define OBD_FAIL_ONCE       (1 << OBD_FAIL_ONCE_BIT)
324
325 /* The following flags aren't made to be combined */
326 #define OBD_FAIL_SKIP        0x20000000 /* skip N times then fail */
327 #define OBD_FAIL_SOME        0x10000000 /* only fail N times */
328 #define OBD_FAIL_RAND        0x08000000 /* fail 1/N of the times */
329 #define OBD_FAIL_USR1        0x04000000 /* user flag */
330
331 #define OBD_FAIL_PRECHECK(id) (obd_fail_loc &&                                \
332                               (obd_fail_loc & OBD_FAIL_MASK_LOC) ==           \
333                               ((id) & OBD_FAIL_MASK_LOC))
334
335 static inline int obd_fail_check_set(__u32 id, __u32 value, int set)
336 {
337         int ret = 0;
338         if (unlikely(OBD_FAIL_PRECHECK(id) &&
339             (ret = __obd_fail_check_set(id, value, set)))) {
340                 CERROR("*** obd_fail_loc=%x ***\n", id);
341         }
342         return ret;
343 }
344
345 /* If id hit obd_fail_loc, return 1, otherwise return 0 */
346 #define OBD_FAIL_CHECK(id) \
347         obd_fail_check_set(id, 0, OBD_FAIL_LOC_NOSET)
348
349 /* If id hit obd_fail_loc, obd_fail_loc |= value and return 1,
350  * otherwise return 0 */
351 #define OBD_FAIL_CHECK_ORSET(id, value) \
352         obd_fail_check_set(id, value, OBD_FAIL_LOC_ORSET)
353
354 /* If id hit obd_fail_loc, obd_fail_loc = value and return 1,
355  * otherwise return 0 */
356 #define OBD_FAIL_CHECK_RESET(id, value) \
357         obd_fail_check_set(id, value, OBD_FAIL_LOC_RESET)
358
359
360 static inline int obd_fail_timeout_set(__u32 id, __u32 value, int ms, int set)
361 {
362         if (unlikely(OBD_FAIL_PRECHECK(id)))
363                 return __obd_fail_timeout_set(id, value, ms, set);
364         else
365                 return 0;
366 }
367
368 /* If id hit obd_fail_loc, sleep for seconds or milliseconds */
369 #define OBD_FAIL_TIMEOUT(id, secs) \
370         obd_fail_timeout_set(id, 0, secs * 1000, OBD_FAIL_LOC_NOSET)
371
372 #define OBD_FAIL_TIMEOUT_MS(id, ms) \
373         obd_fail_timeout_set(id, 0, ms, OBD_FAIL_LOC_NOSET)
374
375 /* If id hit obd_fail_loc, obd_fail_loc |= value and
376  * sleep seconds or milliseconds */
377 #define OBD_FAIL_TIMEOUT_ORSET(id, value, secs) \
378         obd_fail_timeout_set(id, value, secs * 1000, OBD_FAIL_LOC_ORSET)
379
380 #define OBD_FAIL_TIMEOUT_MS_ORSET(id, value, ms) \
381         obd_fail_timeout_set(id, value, ms, OBD_FAIL_LOC_ORSET)
382
383 #ifdef __KERNEL__
384 static inline void obd_fail_write(int id, struct super_block *sb)
385 {
386         /* We set FAIL_ONCE because we never "un-fail" a device */
387         if (OBD_FAIL_CHECK_ORSET(id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
388 #ifdef LIBCFS_DEBUG
389                 BDEVNAME_DECLARE_STORAGE(tmp);
390                 CERROR("obd_fail_loc=%x, fail write operation on %s\n",
391                        id, ll_bdevname(sb, tmp));
392 #endif
393                 /* TODO-CMD: fix getting jdev */
394                 __lvfs_set_rdonly(lvfs_sbdev(sb), (lvfs_sbdev_type)0);
395         }
396 }
397 #define OBD_FAIL_WRITE(id, sb) obd_fail_write(id, sb)
398
399 /* The idea here is to synchronise two threads to force a race. The
400  * first thread that calls this with a matching fail_loc is put to
401  * sleep. The next thread that calls with the same fail_loc wakes up
402  * the first and continues. */
403 static inline void obd_race(__u32 id)
404 {
405         if (OBD_FAIL_PRECHECK(id)) {
406                 if (unlikely(__obd_fail_check_set(id, 0, OBD_FAIL_LOC_NOSET))) {
407                         obd_race_state = 0;
408                         CERROR("obd_race id %x sleeping\n", id);
409                         OBD_SLEEP_ON(obd_race_waitq, obd_race_state != 0);
410                         CERROR("obd_fail_race id %x awake\n", id);
411                 } else {
412                         CERROR("obd_fail_race id %x waking\n", id);
413                         obd_race_state = 1;
414                         wake_up(&obd_race_waitq);
415                 }
416         }
417 }
418 #define OBD_RACE(id) obd_race(id)
419 #else
420 /* sigh.  an expedient fix until OBD_RACE is fixed up */
421 #define OBD_RACE(foo) do {} while(0)
422 #endif
423
424 #define fixme() CDEBUG(D_OTHER, "FIXME\n");
425
426 extern atomic_t libcfs_kmemory;
427
428 #ifdef LPROCFS
429 #define obd_memory_add(size)                                                  \
430         lprocfs_counter_add(obd_memory, OBD_MEMORY_STAT, (long)(size))
431 #define obd_memory_sub(size)                                                  \
432         lprocfs_counter_sub(obd_memory, OBD_MEMORY_STAT, (long)(size))
433 #define obd_memory_sum()                                                      \
434         lprocfs_stats_collector(obd_memory, OBD_MEMORY_STAT,                  \
435                                 LPROCFS_FIELDS_FLAGS_SUM)
436 #define obd_pages_add(order)                                                  \
437         lprocfs_counter_add(obd_memory, OBD_MEMORY_PAGES_STAT,                \
438                             (long)(1 << (order)))
439 #define obd_pages_sub(order)                                                  \
440         lprocfs_counter_sub(obd_memory, OBD_MEMORY_PAGES_STAT,                \
441                             (long)(1 << (order)))
442 #define obd_pages_sum()                                                       \
443         lprocfs_stats_collector(obd_memory, OBD_MEMORY_PAGES_STAT,            \
444                                 LPROCFS_FIELDS_FLAGS_SUM)
445
446 extern void obd_update_maxusage(void);
447 extern __u64 obd_memory_max(void);
448 extern __u64 obd_pages_max(void);
449
450 #else
451
452 extern __u64 obd_alloc;
453 extern __u64 obd_pages;
454
455 extern __u64 obd_max_alloc;
456 extern __u64 obd_max_pages;
457
458 static inline void obd_memory_add(long size)
459 {
460         obd_alloc += size;
461         if (obd_alloc > obd_max_alloc)
462                 obd_max_alloc = obd_alloc;
463 }
464
465 static inline void obd_memory_sub(long size)
466 {
467         obd_alloc -= size;
468 }
469
470 static inline void obd_pages_add(int order)
471 {
472         obd_pages += 1<< order;
473         if (obd_pages > obd_max_pages)
474                 obd_max_pages = obd_pages;
475 }
476
477 static inline void obd_pages_sub(int order)
478 {
479         obd_pages -= 1<< order;
480 }
481
482 #define obd_memory_sum() (obd_alloc)
483 #define obd_pages_sum()  (obd_pages)
484
485 #define obd_memory_max() (obd_max_alloc)
486 #define obd_pages_max() (obd_max_pages)
487
488 #endif
489
490 #if defined (CONFIG_DEBUG_MEMORY) && defined(__KERNEL__)
491
492 #define OBD_MT_WRONG_SIZE    (1 << 0)
493 #define OBD_MT_ALREADY_FREED (1 << 1)
494 #define OBD_MT_LOC_LEN       128
495
496 struct obd_mem_track {
497         struct hlist_node mt_hash;
498         char              mt_loc[OBD_MT_LOC_LEN];
499         int               mt_flags;
500         void             *mt_ptr;
501         int               mt_size;
502 };
503
504 void lvfs_memdbg_show(void);
505 void lvfs_memdbg_insert(struct obd_mem_track *mt);
506 void lvfs_memdbg_remove(struct obd_mem_track *mt);
507 struct obd_mem_track *lvfs_memdbg_find(void *ptr);
508
509 int lvfs_memdbg_check_insert(struct obd_mem_track *mt);
510 struct obd_mem_track *lvfs_memdbg_check_remove(void *ptr);
511
512 static inline struct obd_mem_track *
513 __new_mem_track(void *ptr, int size,
514                 char *file, int line)
515 {
516         struct obd_mem_track *mt;
517
518         mt = kmalloc(sizeof(*mt), GFP_KERNEL);
519         if (unlikely(!mt))
520                 return NULL;
521
522         snprintf(mt->mt_loc, sizeof(mt->mt_loc) - 1,
523                  "%s:%d", file, line);
524
525         mt->mt_size = size;
526         mt->mt_ptr = ptr;
527         mt->mt_flags = 0;
528         return mt;
529 }
530
531 static inline void
532 __free_mem_track(struct obd_mem_track *mt)
533 {
534         kfree(mt);
535 }
536
537 static inline int
538 __get_mem_track(void *ptr, int size,
539                 char *file, int line)
540 {
541         struct obd_mem_track *mt;
542
543         mt = __new_mem_track(ptr, size, file, line);
544         if (unlikely(!mt)) {
545                 CWARN("Can't allocate new memory track\n");
546                 return 0;
547         }
548
549         if (!lvfs_memdbg_check_insert(mt))
550                 __free_mem_track(mt);
551
552         return 1;
553 }
554
555 static inline int
556 __put_mem_track(void *ptr, int size,
557                 char *file, int line)
558 {
559         struct obd_mem_track *mt;
560
561         if (unlikely(!(mt = lvfs_memdbg_check_remove(ptr)))) {
562                 CWARN("Ptr 0x%p is not allocated. Attempt to free "
563                       "not allocated memory at %s:%d\n", ptr,
564                       file, line);
565                 LBUG();
566                 return 0;
567         } else {
568                 if (unlikely(mt->mt_size != size)) {
569                         if (!(mt->mt_flags & OBD_MT_ALREADY_FREED)) {
570                                 mt->mt_flags |= (OBD_MT_WRONG_SIZE |
571                                                  OBD_MT_ALREADY_FREED);
572
573                                 CWARN("Freeing memory chunk (at 0x%p) of "
574                                       "different size than allocated "
575                                       "(%d != %d) at %s:%d, allocated at %s\n",
576                                       ptr, mt->mt_size, size, file, line,
577                                       mt->mt_loc);
578                         }
579                 } else {
580                         __free_mem_track(mt);
581                 }
582                 return 1;
583         }
584 }
585
586 #define get_mem_track(ptr, size, file, line)                                         \
587         __get_mem_track((ptr), (size), (file), (line))
588
589 #define put_mem_track(ptr, size, file, line)                                         \
590         __put_mem_track((ptr), (size), (file), (line))
591
592 #else /* !CONFIG_DEBUG_MEMORY */
593
594 #define get_mem_track(ptr, size, file, line)                                         \
595         do {} while (0)
596
597 #define put_mem_track(ptr, size, file, line)                                         \
598         do {} while (0)
599 #endif /* !CONFIG_DEBUG_MEMORY */
600
601 #define OBD_DEBUG_MEMUSAGE (1)
602
603 #if OBD_DEBUG_MEMUSAGE
604 #define OBD_ALLOC_POST(ptr, size, name)                                 \
605                 obd_memory_add(size);                                   \
606                 get_mem_track((ptr), (size), __FILE__, __LINE__);       \
607                 CDEBUG(D_MALLOC, name " '" #ptr "': %d at %p.\n",       \
608                        (int)(size), ptr)
609
610 #define OBD_FREE_PRE(ptr, size, name)                                   \
611         LASSERT(ptr);                                                   \
612         put_mem_track((ptr), (size), __FILE__, __LINE__);               \
613         obd_memory_sub(size);                                           \
614         CDEBUG(D_MALLOC, name " '" #ptr "': %d at %p.\n",               \
615                (int)(size), ptr);                                       \
616         POISON(ptr, 0x5a, size)
617
618 #else /* !OBD_DEBUG_MEMUSAGE */
619
620 #define OBD_ALLOC_POST(ptr, size, name) ((void)0)
621 #define OBD_FREE_PRE(ptr, size, name)   ((void)0)
622
623 #endif /* !OBD_DEBUG_MEMUSAGE */
624
625 #ifdef RANDOM_FAIL_ALLOC
626 #define HAS_FAIL_ALLOC_FLAG OBD_FAIL_CHECK(OBD_FAIL_GENERAL_ALLOC)
627 #else
628 #define HAS_FAIL_ALLOC_FLAG 0
629 #endif
630
631 #define OBD_ALLOC_FAIL_BITS 24
632 #define OBD_ALLOC_FAIL_MASK ((1 << OBD_ALLOC_FAIL_BITS) - 1)
633 #define OBD_ALLOC_FAIL_MULT (OBD_ALLOC_FAIL_MASK / 100)
634
635 #if defined(LUSTRE_UTILS) /* this version is for utils only */
636 #define OBD_ALLOC_GFP(ptr, size, gfp_mask)                                    \
637 do {                                                                          \
638         (ptr) = cfs_alloc(size, (gfp_mask));                                  \
639         if (unlikely((ptr) == NULL)) {                                        \
640                 CERROR("kmalloc of '" #ptr "' (%d bytes) failed at %s:%d\n",  \
641                        (int)(size), __FILE__, __LINE__);                      \
642         } else {                                                              \
643                 memset(ptr, 0, size);                                         \
644                 CDEBUG(D_MALLOC, "kmalloced '" #ptr "': %d at %p\n",          \
645                        (int)(size), ptr);                                     \
646         }                                                                     \
647 } while (0)
648 #else /* this version is for the kernel and liblustre */
649 #define OBD_FREE_RTN0(ptr)                                                    \
650 ({                                                                            \
651         cfs_free(ptr);                                                        \
652         (ptr) = NULL;                                                         \
653         0;                                                                    \
654 })
655 #define OBD_ALLOC_GFP(ptr, size, gfp_mask)                                    \
656 do {                                                                          \
657         (ptr) = cfs_alloc(size, (gfp_mask));                                  \
658         if (likely((ptr) != NULL &&                                           \
659                    (!HAS_FAIL_ALLOC_FLAG || obd_alloc_fail_rate == 0 ||       \
660                     !obd_alloc_fail(ptr, #ptr, "km", size,                    \
661                                     __FILE__, __LINE__) ||                    \
662                     OBD_FREE_RTN0(ptr)))){                                    \
663                 memset(ptr, 0, size);                                         \
664                 OBD_ALLOC_POST(ptr, size, "kmalloced");                       \
665         }                                                                     \
666 } while (0)
667 #endif
668
669 #ifndef OBD_ALLOC_MASK
670 # define OBD_ALLOC_MASK CFS_ALLOC_IO
671 #endif
672
673 #define OBD_ALLOC(ptr, size) OBD_ALLOC_GFP(ptr, size, OBD_ALLOC_MASK)
674 #define OBD_ALLOC_WAIT(ptr, size) OBD_ALLOC_GFP(ptr, size, CFS_ALLOC_STD)
675 #define OBD_ALLOC_PTR(ptr) OBD_ALLOC(ptr, sizeof *(ptr))
676 #define OBD_ALLOC_PTR_WAIT(ptr) OBD_ALLOC_WAIT(ptr, sizeof *(ptr))
677
678 #ifdef __arch_um__
679 # define OBD_VMALLOC(ptr, size) OBD_ALLOC(ptr, size)
680 #else
681 # define OBD_VMALLOC(ptr, size)                                               \
682 do {                                                                          \
683         (ptr) = cfs_alloc_large(size);                                        \
684         if (unlikely((ptr) == NULL)) {                                        \
685                 CERROR("vmalloc of '" #ptr "' (%d bytes) failed\n",           \
686                        (int)(size));                                          \
687                 CERROR(LPU64" total bytes allocated by Lustre, %d by LNET\n", \
688                        obd_memory_sum(), atomic_read(&libcfs_kmemory));       \
689         } else {                                                              \
690                 memset(ptr, 0, size);                                         \
691                 OBD_ALLOC_POST(ptr, size, "vmalloced");                       \
692         }                                                                     \
693 } while(0)
694 #endif
695
696 #ifdef CONFIG_DEBUG_SLAB
697 #define POISON(ptr, c, s) do {} while (0)
698 #define POISON_PTR(ptr)  ((void)0)
699 #else
700 #define POISON(ptr, c, s) memset(ptr, c, s)
701 #define POISON_PTR(ptr)  (ptr) = (void *)0xdeadbeef
702 #endif
703
704 #ifdef POISON_BULK
705 #define POISON_PAGE(page, val) do { memset(kmap(page), val, CFS_PAGE_SIZE);   \
706                                     kunmap(page); } while (0)
707 #else
708 #define POISON_PAGE(page, val) do { } while (0)
709 #endif
710
711 #ifdef __KERNEL__
712 #define OBD_FREE(ptr, size)                                                   \
713 do {                                                                          \
714         OBD_FREE_PRE(ptr, size, "kfreed");                                    \
715         cfs_free(ptr);                                                        \
716         POISON_PTR(ptr);                                                      \
717 } while(0)
718
719
720 #ifdef HAVE_RCU
721 # ifdef HAVE_CALL_RCU_PARAM
722 #  define my_call_rcu(rcu, cb)            call_rcu(rcu, cb, rcu)
723 # else
724 #  define my_call_rcu(rcu, cb)            call_rcu(rcu, cb)
725 # endif
726 #else
727 # define my_call_rcu(rcu, cb)             (cb)(rcu)
728 #endif
729
730 #define OBD_FREE_RCU_CB(ptr, size, handle, free_cb)                           \
731 do {                                                                          \
732         struct portals_handle *__h = (handle);                                \
733         LASSERT(handle);                                                      \
734         __h->h_ptr = (ptr);                                                   \
735         __h->h_size = (size);                                                 \
736         __h->h_free_cb = (void (*)(void *, size_t))(free_cb);                 \
737         my_call_rcu(&__h->h_rcu, class_handle_free_cb);                       \
738         POISON_PTR(ptr);                                                      \
739 } while(0)
740 #define OBD_FREE_RCU(ptr, size, handle) OBD_FREE_RCU_CB(ptr, size, handle, NULL)
741
742 #else
743 #define OBD_FREE(ptr, size) ((void)(size), free((ptr)))
744 #define OBD_FREE_RCU(ptr, size, handle) (OBD_FREE(ptr, size))
745 #define OBD_FREE_RCU_CB(ptr, size, handle, cb)     ((*(cb))(ptr, size))
746 #endif /* ifdef __KERNEL__ */
747
748 #ifdef __arch_um__
749 # define OBD_VFREE(ptr, size) OBD_FREE(ptr, size)
750 #else
751 # define OBD_VFREE(ptr, size)                                                 \
752 do {                                                                          \
753         OBD_FREE_PRE(ptr, size, "vfreed");                                    \
754         cfs_free_large(ptr);                                                  \
755         POISON_PTR(ptr);                                                      \
756 } while(0)
757 #endif
758
759 /* we memset() the slab object to 0 when allocation succeeds, so DO NOT
760  * HAVE A CTOR THAT DOES ANYTHING.  its work will be cleared here.  we'd
761  * love to assert on that, but slab.c keeps kmem_cache_s all to itself. */
762 #define OBD_SLAB_FREE_RTN0(ptr, slab)                                         \
763 ({                                                                            \
764         cfs_mem_cache_free((slab), (ptr));                                    \
765         (ptr) = NULL;                                                         \
766         0;                                                                    \
767 })
768 #define OBD_SLAB_ALLOC(ptr, slab, type, size)                                 \
769 do {                                                                          \
770         LASSERT(!in_interrupt());                                             \
771         (ptr) = cfs_mem_cache_alloc(slab, (type));                            \
772         if (likely((ptr) != NULL &&                                           \
773                    (!HAS_FAIL_ALLOC_FLAG || obd_alloc_fail_rate == 0 ||       \
774                     !obd_alloc_fail(ptr, #ptr, "slab-", size,                 \
775                                     __FILE__, __LINE__) ||                    \
776                     OBD_SLAB_FREE_RTN0(ptr, slab)))) {                        \
777                 memset(ptr, 0, size);                                         \
778                 OBD_ALLOC_POST(ptr, size, "slab-alloced");                    \
779         }                                                                     \
780 } while(0)
781
782 #define OBD_FREE_PTR(ptr) OBD_FREE(ptr, sizeof *(ptr))
783
784 #define OBD_SLAB_FREE(ptr, slab, size)                                        \
785 do {                                                                          \
786         OBD_FREE_PRE(ptr, size, "slab-freed");                                \
787         cfs_mem_cache_free(slab, ptr);                                        \
788         POISON_PTR(ptr);                                                      \
789 } while(0)
790
791 #define OBD_SLAB_ALLOC_PTR(ptr, slab)                                         \
792         OBD_SLAB_ALLOC((ptr), (slab), CFS_ALLOC_STD, sizeof *(ptr))
793 #define OBD_SLAB_FREE_PTR(ptr, slab)                                          \
794         OBD_SLAB_FREE((ptr), (slab), sizeof *(ptr))
795
796 #define KEY_IS(str) \
797         (keylen >= (sizeof(str)-1) && memcmp(key, str, (sizeof(str)-1)) == 0)
798
799 /* Wrapper for contiguous page frame allocation */
800 #define OBD_PAGES_ALLOC(ptr, order, gfp_mask)                                 \
801 do {                                                                          \
802         (ptr) = cfs_alloc_pages(gfp_mask, order);                             \
803         if (unlikely((ptr) == NULL)) {                                        \
804                 CERROR("alloc_pages of '" #ptr "' %d page(s) / "LPU64" bytes "\
805                        "failed\n", (int)(1 << (order)),                       \
806                        (__u64)((1 << (order)) << CFS_PAGE_SHIFT));            \
807                 CERROR(LPU64" total bytes and "LPU64" total pages "           \
808                        "("LPU64" bytes) allocated by Lustre, "                \
809                        "%d total bytes by LNET\n",                            \
810                        obd_memory_sum(),                                      \
811                        obd_pages_sum() << CFS_PAGE_SHIFT,                     \
812                        obd_pages_sum(),                                       \
813                        atomic_read(&libcfs_kmemory));                         \
814         } else {                                                              \
815                 obd_pages_add(order);                                         \
816                 CDEBUG(D_MALLOC, "alloc_pages '" #ptr "': %d page(s) / "      \
817                        LPU64" bytes at %p.\n",                                \
818                        (int)(1 << (order)),                                   \
819                        (__u64)((1 << (order)) << CFS_PAGE_SHIFT), ptr);       \
820         }                                                                     \
821 } while (0)
822
823 #define OBD_PAGE_ALLOC(ptr, gfp_mask)                                         \
824         OBD_PAGES_ALLOC(ptr, 0, gfp_mask)
825
826 #define OBD_PAGES_FREE(ptr, order)                                            \
827 do {                                                                          \
828         LASSERT(ptr);                                                         \
829         obd_pages_sub(order);                                                 \
830         CDEBUG(D_MALLOC, "free_pages '" #ptr "': %d page(s) / "LPU64" bytes " \
831                "at %p.\n",                                                    \
832                (int)(1 << (order)), (__u64)((1 << (order)) << CFS_PAGE_SHIFT),\
833                ptr);                                                          \
834         __cfs_free_pages(ptr, order);                                         \
835         (ptr) = (void *)0xdeadbeef;                                           \
836 } while (0)
837
838 #define OBD_PAGE_FREE(ptr) OBD_PAGES_FREE(ptr, 0)
839
840 #endif