Whamcloud - gitweb
LU-641 obdclass: use rwlock for obd_dev_lock
[fs/lustre-release.git] / lustre / obdclass / class_obd.c
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
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
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 (c) 1999, 2010, Oracle and/or its affiliates. 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 #define DEBUG_SUBSYSTEM S_CLASS
38 #ifndef EXPORT_SYMTAB
39 # define EXPORT_SYMTAB
40 #endif
41 #ifndef __KERNEL__
42 # include <liblustre.h>
43 #else
44 # include <asm/atomic.h>
45 #endif
46
47 #include <obd_support.h>
48 #include <obd_class.h>
49 #include <lnet/lnetctl.h>
50 #include <lustre_debug.h>
51 #include <lprocfs_status.h>
52 #include <lustre/lustre_build_version.h>
53 #include <libcfs/list.h>
54 #include "llog_internal.h"
55
56 #ifndef __KERNEL__
57 /* liblustre workaround */
58 cfs_atomic_t libcfs_kmemory = {0};
59 #endif
60
61 struct obd_device *obd_devs[MAX_OBD_DEVICES];
62 cfs_list_t obd_types;
63 cfs_rwlock_t obd_dev_lock = CFS_RW_LOCK_UNLOCKED;
64
65 #ifndef __KERNEL__
66 __u64 obd_max_pages = 0;
67 __u64 obd_max_alloc = 0;
68 __u64 obd_alloc;
69 __u64 obd_pages;
70 #endif
71
72 /* The following are visible and mutable through /proc/sys/lustre/. */
73 unsigned int obd_debug_peer_on_timeout;
74 unsigned int obd_dump_on_timeout;
75 unsigned int obd_dump_on_eviction;
76 unsigned int obd_max_dirty_pages = 256;
77 unsigned int obd_timeout = OBD_TIMEOUT_DEFAULT;   /* seconds */
78 unsigned int ldlm_timeout = LDLM_TIMEOUT_DEFAULT; /* seconds */
79 /* Adaptive timeout defs here instead of ptlrpc module for /proc/sys/ access */
80 unsigned int at_min = 0;
81 unsigned int at_max = 600;
82 unsigned int at_history = 600;
83 int at_early_margin = 5;
84 int at_extra = 30;
85
86 cfs_atomic_t obd_dirty_pages;
87 cfs_atomic_t obd_dirty_transit_pages;
88
89 static inline void obd_data2conn(struct lustre_handle *conn,
90                                  struct obd_ioctl_data *data)
91 {
92         memset(conn, 0, sizeof *conn);
93         conn->cookie = data->ioc_cookie;
94 }
95
96 static inline void obd_conn2data(struct obd_ioctl_data *data,
97                                  struct lustre_handle *conn)
98 {
99         data->ioc_cookie = conn->cookie;
100 }
101
102 int class_resolve_dev_name(__u32 len, const char *name)
103 {
104         int rc;
105         int dev;
106
107         ENTRY;
108         if (!len || !name) {
109                 CERROR("No name passed,!\n");
110                 GOTO(out, rc = -EINVAL);
111         }
112         if (name[len - 1] != 0) {
113                 CERROR("Name not nul terminated!\n");
114                 GOTO(out, rc = -EINVAL);
115         }
116
117         CDEBUG(D_IOCTL, "device name %s\n", name);
118         dev = class_name2dev(name);
119         if (dev == -1) {
120                 CDEBUG(D_IOCTL, "No device for name %s!\n", name);
121                 GOTO(out, rc = -EINVAL);
122         }
123
124         CDEBUG(D_IOCTL, "device name %s, dev %d\n", name, dev);
125         rc = dev;
126
127 out:
128         RETURN(rc);
129 }
130
131 int class_handle_ioctl(unsigned int cmd, unsigned long arg)
132 {
133         char *buf = NULL;
134         struct obd_ioctl_data *data;
135         struct libcfs_debug_ioctl_data *debug_data;
136         struct obd_device *obd = NULL;
137         int err = 0, len = 0;
138         ENTRY;
139
140         /* only for debugging */
141         if (cmd == LIBCFS_IOC_DEBUG_MASK) {
142                 debug_data = (struct libcfs_debug_ioctl_data*)arg;
143                 libcfs_subsystem_debug = debug_data->subs;
144                 libcfs_debug = debug_data->debug;
145                 return 0;
146         }
147
148         CDEBUG(D_IOCTL, "cmd = %x\n", cmd);
149         if (obd_ioctl_getdata(&buf, &len, (void *)arg)) {
150                 CERROR("OBD ioctl: data error\n");
151                 RETURN(-EINVAL);
152         }
153         data = (struct obd_ioctl_data *)buf;
154
155         switch (cmd) {
156         case OBD_IOC_PROCESS_CFG: {
157                 struct lustre_cfg *lcfg;
158
159                 if (!data->ioc_plen1 || !data->ioc_pbuf1) {
160                         CERROR("No config buffer passed!\n");
161                         GOTO(out, err = -EINVAL);
162                 }
163                 OBD_ALLOC(lcfg, data->ioc_plen1);
164                 if (lcfg == NULL)
165                         GOTO(out, err = -ENOMEM);
166                 err = cfs_copy_from_user(lcfg, data->ioc_pbuf1,
167                                          data->ioc_plen1);
168                 if (!err)
169                         err = lustre_cfg_sanity_check(lcfg, data->ioc_plen1);
170                 if (!err)
171                         err = class_process_config(lcfg);
172
173                 OBD_FREE(lcfg, data->ioc_plen1);
174                 GOTO(out, err);
175         }
176
177         case OBD_GET_VERSION:
178                 if (!data->ioc_inlbuf1) {
179                         CERROR("No buffer passed in ioctl\n");
180                         GOTO(out, err = -EINVAL);
181                 }
182
183                 if (strlen(BUILD_VERSION) + 1 > data->ioc_inllen1) {
184                         CERROR("ioctl buffer too small to hold version\n");
185                         GOTO(out, err = -EINVAL);
186                 }
187
188                 memcpy(data->ioc_bulk, BUILD_VERSION,
189                        strlen(BUILD_VERSION) + 1);
190
191                 err = obd_ioctl_popdata((void *)arg, data, len);
192                 if (err)
193                         err = -EFAULT;
194                 GOTO(out, err);
195
196         case OBD_IOC_NAME2DEV: {
197                 /* Resolve a device name.  This does not change the
198                  * currently selected device.
199                  */
200                 int dev;
201
202                 dev = class_resolve_dev_name(data->ioc_inllen1,
203                                              data->ioc_inlbuf1);
204                 data->ioc_dev = dev;
205                 if (dev < 0)
206                         GOTO(out, err = -EINVAL);
207
208                 err = obd_ioctl_popdata((void *)arg, data, sizeof(*data));
209                 if (err)
210                         err = -EFAULT;
211                 GOTO(out, err);
212         }
213
214         case OBD_IOC_UUID2DEV: {
215                 /* Resolve a device uuid.  This does not change the
216                  * currently selected device.
217                  */
218                 int dev;
219                 struct obd_uuid uuid;
220
221                 if (!data->ioc_inllen1 || !data->ioc_inlbuf1) {
222                         CERROR("No UUID passed!\n");
223                         GOTO(out, err = -EINVAL);
224                 }
225                 if (data->ioc_inlbuf1[data->ioc_inllen1 - 1] != 0) {
226                         CERROR("UUID not NUL terminated!\n");
227                         GOTO(out, err = -EINVAL);
228                 }
229
230                 CDEBUG(D_IOCTL, "device name %s\n", data->ioc_inlbuf1);
231                 obd_str2uuid(&uuid, data->ioc_inlbuf1);
232                 dev = class_uuid2dev(&uuid);
233                 data->ioc_dev = dev;
234                 if (dev == -1) {
235                         CDEBUG(D_IOCTL, "No device for UUID %s!\n",
236                                data->ioc_inlbuf1);
237                         GOTO(out, err = -EINVAL);
238                 }
239
240                 CDEBUG(D_IOCTL, "device name %s, dev %d\n", data->ioc_inlbuf1,
241                        dev);
242                 err = obd_ioctl_popdata((void *)arg, data, sizeof(*data));
243                 if (err)
244                         err = -EFAULT;
245                 GOTO(out, err);
246         }
247
248         case OBD_IOC_CLOSE_UUID: {
249                 CDEBUG(D_IOCTL, "closing all connections to uuid %s (NOOP)\n",
250                        data->ioc_inlbuf1);
251                 GOTO(out, err = 0);
252         }
253
254         case OBD_IOC_GETDEVICE: {
255                 int     index = data->ioc_count;
256                 char    *status, *str;
257
258                 if (!data->ioc_inlbuf1) {
259                         CERROR("No buffer passed in ioctl\n");
260                         GOTO(out, err = -EINVAL);
261                 }
262                 if (data->ioc_inllen1 < 128) {
263                         CERROR("ioctl buffer too small to hold version\n");
264                         GOTO(out, err = -EINVAL);
265                 }
266
267                 obd = class_num2obd(index);
268                 if (!obd)
269                         GOTO(out, err = -ENOENT);
270
271                 if (obd->obd_stopping)
272                         status = "ST";
273                 else if (obd->obd_set_up)
274                         status = "UP";
275                 else if (obd->obd_attached)
276                         status = "AT";
277                 else
278                         status = "--";
279                 str = (char *)data->ioc_bulk;
280                 snprintf(str, len - sizeof(*data), "%3d %s %s %s %s %d",
281                          (int)index, status, obd->obd_type->typ_name,
282                          obd->obd_name, obd->obd_uuid.uuid,
283                          cfs_atomic_read(&obd->obd_refcount));
284                 err = obd_ioctl_popdata((void *)arg, data, len);
285
286                 GOTO(out, err = 0);
287         }
288
289         }
290
291         if (data->ioc_dev == OBD_DEV_BY_DEVNAME) {
292                 if (data->ioc_inllen4 <= 0 || data->ioc_inlbuf4 == NULL)
293                         GOTO(out, err = -EINVAL);
294                 if (strnlen(data->ioc_inlbuf4, MAX_OBD_NAME) >= MAX_OBD_NAME)
295                         GOTO(out, err = -EINVAL);
296                 obd = class_name2obd(data->ioc_inlbuf4);
297         } else if (data->ioc_dev < class_devno_max()) {
298                 obd = class_num2obd(data->ioc_dev);
299         } else {
300                 CERROR("OBD ioctl: No device\n");
301                 GOTO(out, err = -EINVAL);
302         }
303
304         if (obd == NULL) {
305                 CERROR("OBD ioctl : No Device %d\n", data->ioc_dev);
306                 GOTO(out, err = -EINVAL);
307         }
308         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
309
310         if (!obd->obd_set_up || obd->obd_stopping) {
311                 CERROR("OBD ioctl: device not setup %d \n", data->ioc_dev);
312                 GOTO(out, err = -EINVAL);
313         }
314
315         switch(cmd) {
316         case OBD_IOC_NO_TRANSNO: {
317                 if (!obd->obd_attached) {
318                         CERROR("Device %d not attached\n", obd->obd_minor);
319                         GOTO(out, err = -ENODEV);
320                 }
321                 CDEBUG(D_HA, "%s: disabling committed-transno notification\n",
322                        obd->obd_name);
323                 obd->obd_no_transno = 1;
324                 GOTO(out, err = 0);
325         }
326
327         default: {
328                 err = obd_iocontrol(cmd, obd->obd_self_export, len, data, NULL);
329                 if (err)
330                         GOTO(out, err);
331
332                 err = obd_ioctl_popdata((void *)arg, data, len);
333                 if (err)
334                         err = -EFAULT;
335                 GOTO(out, err);
336         }
337         }
338
339  out:
340         if (buf)
341                 obd_ioctl_freedata(buf, len);
342         RETURN(err);
343 } /* class_handle_ioctl */
344
345
346
347 #ifdef __KERNEL__
348 extern cfs_psdev_t obd_psdev;
349 #else
350 void *obd_psdev = NULL;
351 #endif
352
353 EXPORT_SYMBOL(obd_devs);
354 EXPORT_SYMBOL(obd_debug_peer_on_timeout);
355 EXPORT_SYMBOL(obd_dump_on_timeout);
356 EXPORT_SYMBOL(obd_dump_on_eviction);
357 EXPORT_SYMBOL(obd_timeout);
358 EXPORT_SYMBOL(ldlm_timeout);
359 EXPORT_SYMBOL(obd_max_dirty_pages);
360 EXPORT_SYMBOL(obd_dirty_pages);
361 EXPORT_SYMBOL(obd_dirty_transit_pages);
362 EXPORT_SYMBOL(at_min);
363 EXPORT_SYMBOL(at_max);
364 EXPORT_SYMBOL(at_extra);
365 EXPORT_SYMBOL(at_early_margin);
366 EXPORT_SYMBOL(at_history);
367 EXPORT_SYMBOL(ptlrpc_put_connection_superhack);
368
369 EXPORT_SYMBOL(proc_lustre_root);
370
371 /* uuid.c */
372 EXPORT_SYMBOL(class_uuid_unparse);
373 EXPORT_SYMBOL(lustre_uuid_to_peer);
374
375 EXPORT_SYMBOL(class_handle_hash);
376 EXPORT_SYMBOL(class_handle_unhash);
377 EXPORT_SYMBOL(class_handle_hash_back);
378 EXPORT_SYMBOL(class_handle2object);
379 EXPORT_SYMBOL(class_handle_free_cb);
380
381 /* obd_config.c */
382 EXPORT_SYMBOL(class_incref);
383 EXPORT_SYMBOL(class_decref);
384 EXPORT_SYMBOL(class_get_profile);
385 EXPORT_SYMBOL(class_del_profile);
386 EXPORT_SYMBOL(class_del_profiles);
387 EXPORT_SYMBOL(class_process_config);
388 EXPORT_SYMBOL(class_process_proc_param);
389 EXPORT_SYMBOL(class_config_parse_llog);
390 EXPORT_SYMBOL(class_config_dump_llog);
391 EXPORT_SYMBOL(class_attach);
392 EXPORT_SYMBOL(class_setup);
393 EXPORT_SYMBOL(class_cleanup);
394 EXPORT_SYMBOL(class_detach);
395 EXPORT_SYMBOL(class_manual_cleanup);
396
397 /* mea.c */
398 EXPORT_SYMBOL(mea_name2idx);
399 EXPORT_SYMBOL(raw_name2idx);
400
401 #define OBD_INIT_CHECK
402 #ifdef OBD_INIT_CHECK
403 int obd_init_checks(void)
404 {
405         __u64 u64val, div64val;
406         char buf[64];
407         int len, ret = 0;
408
409         CDEBUG(D_INFO, "LPU64=%s, LPD64=%s, LPX64=%s\n", LPU64, LPD64, LPX64);
410
411         CDEBUG(D_INFO, "OBD_OBJECT_EOF = "LPX64"\n", (__u64)OBD_OBJECT_EOF);
412
413         u64val = OBD_OBJECT_EOF;
414         CDEBUG(D_INFO, "u64val OBD_OBJECT_EOF = "LPX64"\n", u64val);
415         if (u64val != OBD_OBJECT_EOF) {
416                 CERROR("__u64 "LPX64"(%d) != 0xffffffffffffffff\n",
417                        u64val, (int)sizeof(u64val));
418                 ret = -EINVAL;
419         }
420         len = snprintf(buf, sizeof(buf), LPX64, u64val);
421         if (len != 18) {
422                 CWARN("LPX64 wrong length! strlen(%s)=%d != 18\n", buf, len);
423                 ret = -EINVAL;
424         }
425
426         div64val = OBD_OBJECT_EOF;
427         CDEBUG(D_INFO, "u64val OBD_OBJECT_EOF = "LPX64"\n", u64val);
428         if (u64val != OBD_OBJECT_EOF) {
429                 CERROR("__u64 "LPX64"(%d) != 0xffffffffffffffff\n",
430                        u64val, (int)sizeof(u64val));
431                 ret = -EOVERFLOW;
432         }
433         if (u64val >> 8 != OBD_OBJECT_EOF >> 8) {
434                 CERROR("__u64 "LPX64"(%d) != 0xffffffffffffffff\n",
435                        u64val, (int)sizeof(u64val));
436                 return -EOVERFLOW;
437         }
438         if (do_div(div64val, 256) != (u64val & 255)) {
439                 CERROR("do_div("LPX64",256) != "LPU64"\n", u64val, u64val &255);
440                 return -EOVERFLOW;
441         }
442         if (u64val >> 8 != div64val) {
443                 CERROR("do_div("LPX64",256) "LPU64" != "LPU64"\n",
444                        u64val, div64val, u64val >> 8);
445                 return -EOVERFLOW;
446         }
447         len = snprintf(buf, sizeof(buf), LPX64, u64val);
448         if (len != 18) {
449                 CWARN("LPX64 wrong length! strlen(%s)=%d != 18\n", buf, len);
450                 ret = -EINVAL;
451         }
452         len = snprintf(buf, sizeof(buf), LPU64, u64val);
453         if (len != 20) {
454                 CWARN("LPU64 wrong length! strlen(%s)=%d != 20\n", buf, len);
455                 ret = -EINVAL;
456         }
457         len = snprintf(buf, sizeof(buf), LPD64, u64val);
458         if (len != 2) {
459                 CWARN("LPD64 wrong length! strlen(%s)=%d != 2\n", buf, len);
460                 ret = -EINVAL;
461         }
462         if ((u64val & ~CFS_PAGE_MASK) >= CFS_PAGE_SIZE) {
463                 CWARN("mask failed: u64val "LPU64" >= "LPU64"\n", u64val,
464                       (__u64)CFS_PAGE_SIZE);
465                 ret = -EINVAL;
466         }
467
468         return ret;
469 }
470 #else
471 #define obd_init_checks() do {} while(0)
472 #endif
473
474 extern cfs_spinlock_t obd_types_lock;
475 extern int class_procfs_init(void);
476 extern int class_procfs_clean(void);
477
478 #ifdef __KERNEL__
479 static int __init init_obdclass(void)
480 #else
481 int init_obdclass(void)
482 #endif
483 {
484         int i, err;
485 #ifdef __KERNEL__
486         int lustre_register_fs(void);
487
488         for (i = CAPA_SITE_CLIENT; i < CAPA_SITE_MAX; i++)
489                 CFS_INIT_LIST_HEAD(&capa_list[i]);
490 #endif
491
492 #ifdef CRAY_XT3
493         LCONSOLE_INFO("Lustre: Build Version: "BUILD_VERSION"\n");
494 #else
495         LCONSOLE_INFO("OBD class driver, http://wiki.whamcloud.com/\n");
496         LCONSOLE_INFO("        Lustre Version: "LUSTRE_VERSION_STRING"\n");
497         LCONSOLE_INFO("        Build Version: "BUILD_VERSION"\n");
498 #endif
499
500         cfs_spin_lock_init(&obd_types_lock);
501         obd_zombie_impexp_init();
502 #ifdef LPROCFS
503         obd_memory = lprocfs_alloc_stats(OBD_STATS_NUM,
504                                          LPROCFS_STATS_FLAG_NONE);
505         if (obd_memory == NULL) {
506                 CERROR("kmalloc of 'obd_memory' failed\n");
507                 RETURN(-ENOMEM);
508         }
509
510         lprocfs_counter_init(obd_memory, OBD_MEMORY_STAT,
511                              LPROCFS_CNTR_AVGMINMAX,
512                              "memused", "bytes");
513         lprocfs_counter_init(obd_memory, OBD_MEMORY_PAGES_STAT,
514                              LPROCFS_CNTR_AVGMINMAX,
515                              "pagesused", "pages");
516 #endif
517         err = obd_init_checks();
518         if (err == -EOVERFLOW)
519                 return err;
520
521         class_init_uuidlist();
522         err = class_handle_init();
523         if (err)
524                 return err;
525
526         CFS_INIT_LIST_HEAD(&obd_types);
527
528         err = cfs_psdev_register(&obd_psdev);
529         if (err) {
530                 CERROR("cannot register %d err %d\n", OBD_DEV_MINOR, err);
531                 return err;
532         }
533
534         /* This struct is already zeroed for us (static global) */
535         for (i = 0; i < class_devno_max(); i++)
536                 obd_devs[i] = NULL;
537
538         /* Default the dirty page cache cap to 1/2 of system memory.
539          * For clients with less memory, a larger fraction is needed
540          * for other purposes (mostly for BGL). */
541         if (cfs_num_physpages <= 512 << (20 - CFS_PAGE_SHIFT))
542                 obd_max_dirty_pages = cfs_num_physpages / 4;
543         else
544                 obd_max_dirty_pages = cfs_num_physpages / 2;
545
546         err = obd_init_caches();
547         if (err)
548                 return err;
549 #ifdef __KERNEL__
550         err = class_procfs_init();
551         if (err)
552                 return err;
553 #endif
554
555         err = lu_global_init();
556         if (err)
557                 return err;
558
559 #ifdef __KERNEL__
560         err = lustre_register_fs();
561 #endif
562
563         return err;
564 }
565
566 /* liblustre doesn't call cleanup_obdclass, apparently.  we carry on in this
567  * ifdef to the end of the file to cover module and versioning goo.*/
568 #ifdef __KERNEL__
569 static void cleanup_obdclass(void)
570 {
571         int i;
572         int lustre_unregister_fs(void);
573         __u64 memory_leaked, pages_leaked;
574         __u64 memory_max, pages_max;
575         ENTRY;
576
577         lustre_unregister_fs();
578
579         cfs_psdev_deregister(&obd_psdev);
580         for (i = 0; i < class_devno_max(); i++) {
581                 struct obd_device *obd = class_num2obd(i);
582                 if (obd && obd->obd_set_up &&
583                     OBT(obd) && OBP(obd, detach)) {
584                         /* XXX should this call generic detach otherwise? */
585                         LASSERT(obd->obd_magic == OBD_DEVICE_MAGIC);
586                         OBP(obd, detach)(obd);
587                 }
588         }
589         lu_global_fini();
590
591         obd_cleanup_caches();
592         obd_sysctl_clean();
593
594         class_procfs_clean();
595
596         class_handle_cleanup();
597         class_exit_uuidlist();
598         obd_zombie_impexp_stop();
599
600         memory_leaked = obd_memory_sum();
601         pages_leaked = obd_pages_sum();
602
603         memory_max = obd_memory_max();
604         pages_max = obd_pages_max();
605
606         lprocfs_free_stats(&obd_memory);
607         CDEBUG((memory_leaked) ? D_ERROR : D_INFO,
608                "obd_memory max: "LPU64", leaked: "LPU64"\n",
609                memory_max, memory_leaked);
610         CDEBUG((pages_leaked) ? D_ERROR : D_INFO,
611                "obd_memory_pages max: "LPU64", leaked: "LPU64"\n",
612                pages_max, pages_leaked);
613
614         EXIT;
615 }
616
617 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
618 MODULE_DESCRIPTION("Lustre Class Driver Build Version: " BUILD_VERSION);
619 MODULE_LICENSE("GPL");
620
621 cfs_module(obdclass, LUSTRE_VERSION_STRING, init_obdclass, cleanup_obdclass);
622 #endif