]> CyberLeo.Net >> Repos - FreeBSD/stable/9.git/blob - sys/geom/geom_dev.c
MFC r238171, r248679:
[FreeBSD/stable/9.git] / sys / geom / geom_dev.c
1 /*-
2  * Copyright (c) 2002 Poul-Henning Kamp
3  * Copyright (c) 2002 Networks Associates Technology, Inc.
4  * All rights reserved.
5  *
6  * This software was developed for the FreeBSD Project by Poul-Henning Kamp
7  * and NAI Labs, the Security Research Division of Network Associates, Inc.
8  * under DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
9  * DARPA CHATS research program.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. The names of the authors may not be used to endorse or promote
20  *    products derived from this software without specific prior written
21  *    permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/malloc.h>
42 #include <sys/kernel.h>
43 #include <sys/conf.h>
44 #include <sys/ctype.h>
45 #include <sys/bio.h>
46 #include <sys/bus.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/proc.h>
50 #include <sys/errno.h>
51 #include <sys/time.h>
52 #include <sys/disk.h>
53 #include <sys/fcntl.h>
54 #include <sys/limits.h>
55 #include <geom/geom.h>
56 #include <geom/geom_int.h>
57 #include <machine/stdarg.h>
58
59 struct g_dev_softc {
60         struct mtx       sc_mtx;
61         struct cdev     *sc_dev;
62         struct cdev     *sc_alias;
63         int              sc_open;
64         int              sc_active;
65 };
66
67 static d_open_t         g_dev_open;
68 static d_close_t        g_dev_close;
69 static d_strategy_t     g_dev_strategy;
70 static d_ioctl_t        g_dev_ioctl;
71
72 static struct cdevsw g_dev_cdevsw = {
73         .d_version =    D_VERSION,
74         .d_open =       g_dev_open,
75         .d_close =      g_dev_close,
76         .d_read =       physread,
77         .d_write =      physwrite,
78         .d_ioctl =      g_dev_ioctl,
79         .d_strategy =   g_dev_strategy,
80         .d_name =       "g_dev",
81         .d_flags =      D_DISK | D_TRACKCLOSE,
82 };
83
84 static g_taste_t g_dev_taste;
85 static g_orphan_t g_dev_orphan;
86 static g_attrchanged_t g_dev_attrchanged;
87
88 static struct g_class g_dev_class       = {
89         .name = "DEV",
90         .version = G_VERSION,
91         .taste = g_dev_taste,
92         .orphan = g_dev_orphan,
93         .attrchanged = g_dev_attrchanged
94 };
95
96 static void
97 g_dev_destroy(void *arg, int flags __unused)
98 {
99         struct g_consumer *cp;
100         struct g_geom *gp;
101         struct g_dev_softc *sc;
102
103         g_topology_assert();
104         cp = arg;
105         gp = cp->geom;
106         sc = cp->private;
107         g_trace(G_T_TOPOLOGY, "g_dev_destroy(%p(%s))", cp, gp->name);
108         if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
109                 g_access(cp, -cp->acr, -cp->acw, -cp->ace);
110         g_detach(cp);
111         g_destroy_consumer(cp);
112         g_destroy_geom(gp);
113         mtx_destroy(&sc->sc_mtx);
114         g_free(sc);
115 }
116
117 void
118 g_dev_print(void)
119 {
120         struct g_geom *gp;
121         char const *p = "";
122
123         LIST_FOREACH(gp, &g_dev_class.geom, geom) {
124                 printf("%s%s", p, gp->name);
125                 p = " ";
126         }
127         printf("\n");
128 }
129
130 static void
131 g_dev_attrchanged(struct g_consumer *cp, const char *attr)
132 {
133         struct g_dev_softc *sc;
134         struct cdev *dev;
135         char buf[SPECNAMELEN + 6];
136
137         sc = cp->private;
138         if (strcmp(attr, "GEOM::media") == 0) {
139                 dev = sc->sc_dev;
140                 snprintf(buf, sizeof(buf), "cdev=%s", dev->si_name);
141                 devctl_notify_f("DEVFS", "CDEV", "MEDIACHANGE", buf, M_WAITOK);
142                 dev = sc->sc_alias;
143                 if (dev != NULL) {
144                         snprintf(buf, sizeof(buf), "cdev=%s", dev->si_name);
145                         devctl_notify_f("DEVFS", "CDEV", "MEDIACHANGE", buf,
146                             M_WAITOK);
147                 }
148                 return;
149         }
150
151         if (strcmp(attr, "GEOM::physpath") != 0)
152                 return;
153
154         if (g_access(cp, 1, 0, 0) == 0) {
155                 char *physpath;
156                 int error, physpath_len;
157
158                 physpath_len = MAXPATHLEN;
159                 physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
160                 error =
161                     g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
162                 g_access(cp, -1, 0, 0);
163                 if (error == 0 && strlen(physpath) != 0) {
164                         struct cdev *old_alias_dev;
165                         struct cdev **alias_devp;
166
167                         dev = sc->sc_dev;
168                         old_alias_dev = sc->sc_alias;
169                         alias_devp = (struct cdev **)&sc->sc_alias;
170                         make_dev_physpath_alias(MAKEDEV_WAITOK, alias_devp,
171                             dev, old_alias_dev, physpath);
172                 } else if (sc->sc_alias) {
173                         destroy_dev((struct cdev *)sc->sc_alias);
174                         sc->sc_alias = NULL;
175                 }
176                 g_free(physpath);
177         }
178 }
179
180 struct g_provider *
181 g_dev_getprovider(struct cdev *dev)
182 {
183         struct g_consumer *cp;
184
185         g_topology_assert();
186         if (dev == NULL)
187                 return (NULL);
188         if (dev->si_devsw != &g_dev_cdevsw)
189                 return (NULL);
190         cp = dev->si_drv2;
191         return (cp->provider);
192 }
193
194 static struct g_geom *
195 g_dev_taste(struct g_class *mp, struct g_provider *pp, int insist __unused)
196 {
197         struct g_geom *gp;
198         struct g_consumer *cp;
199         struct g_dev_softc *sc;
200         int error, len;
201         struct cdev *dev, *adev;
202         char buf[64], *val;
203
204         g_trace(G_T_TOPOLOGY, "dev_taste(%s,%s)", mp->name, pp->name);
205         g_topology_assert();
206         gp = g_new_geomf(mp, "%s", pp->name);
207         sc = g_malloc(sizeof(*sc), M_WAITOK | M_ZERO);
208         mtx_init(&sc->sc_mtx, "g_dev", NULL, MTX_DEF);
209         cp = g_new_consumer(gp);
210         cp->private = sc;
211         error = g_attach(cp, pp);
212         KASSERT(error == 0,
213             ("g_dev_taste(%s) failed to g_attach, err=%d", pp->name, error));
214         error = make_dev_p(MAKEDEV_CHECKNAME | MAKEDEV_WAITOK, &dev,
215             &g_dev_cdevsw, NULL, UID_ROOT, GID_OPERATOR, 0640, "%s", gp->name);
216         if (error != 0) {
217                 printf("%s: make_dev_p() failed (gp->name=%s, error=%d)\n",
218                     __func__, gp->name, error);
219                 g_detach(cp);
220                 g_destroy_consumer(cp);
221                 g_destroy_geom(gp);
222                 mtx_destroy(&sc->sc_mtx);
223                 g_free(sc);
224                 return (NULL);
225         }
226         sc->sc_dev = dev;
227
228         /* Search for device alias name and create it if found. */
229         adev = NULL;
230         for (len = MIN(strlen(gp->name), sizeof(buf) - 15); len > 0; len--) {
231                 snprintf(buf, sizeof(buf), "kern.devalias.%s", gp->name);
232                 buf[14 + len] = 0;
233                 val = getenv(buf);
234                 if (val != NULL) {
235                         snprintf(buf, sizeof(buf), "%s%s",
236                             val, gp->name + len);
237                         freeenv(val);
238                         make_dev_alias_p(MAKEDEV_CHECKNAME | MAKEDEV_WAITOK,
239                             &adev, dev, "%s", buf);
240                         break;
241                 }
242         }
243
244         if (pp->flags & G_PF_CANDELETE)
245                 dev->si_flags |= SI_CANDELETE;
246         dev->si_iosize_max = MAXPHYS;
247         dev->si_drv2 = cp;
248         if (adev != NULL) {
249                 if (pp->flags & G_PF_CANDELETE)
250                         adev->si_flags |= SI_CANDELETE;
251                 adev->si_iosize_max = MAXPHYS;
252                 adev->si_drv2 = cp;
253         }
254
255         g_dev_attrchanged(cp, "GEOM::physpath");
256
257         return (gp);
258 }
259
260 static int
261 g_dev_open(struct cdev *dev, int flags, int fmt, struct thread *td)
262 {
263         struct g_consumer *cp;
264         struct g_dev_softc *sc;
265         int error, r, w, e;
266
267         cp = dev->si_drv2;
268         if (cp == NULL)
269                 return(ENXIO);          /* g_dev_taste() not done yet */
270         g_trace(G_T_ACCESS, "g_dev_open(%s, %d, %d, %p)",
271             cp->geom->name, flags, fmt, td);
272
273         r = flags & FREAD ? 1 : 0;
274         w = flags & FWRITE ? 1 : 0;
275 #ifdef notyet
276         e = flags & O_EXCL ? 1 : 0;
277 #else
278         e = 0;
279 #endif
280         if (w) {
281                 /*
282                  * When running in very secure mode, do not allow
283                  * opens for writing of any disks.
284                  */
285                 error = securelevel_ge(td->td_ucred, 2);
286                 if (error)
287                         return (error);
288         }
289         g_topology_lock();
290         error = g_access(cp, r, w, e);
291         g_topology_unlock();
292         if (error == 0) {
293                 sc = cp->private;
294                 mtx_lock(&sc->sc_mtx);
295                 if (sc->sc_open == 0 && sc->sc_active != 0)
296                         wakeup(&sc->sc_active);
297                 sc->sc_open += r + w + e;
298                 mtx_unlock(&sc->sc_mtx);
299         }
300         return(error);
301 }
302
303 static int
304 g_dev_close(struct cdev *dev, int flags, int fmt, struct thread *td)
305 {
306         struct g_consumer *cp;
307         struct g_dev_softc *sc;
308         int error, r, w, e;
309
310         cp = dev->si_drv2;
311         if (cp == NULL)
312                 return(ENXIO);
313         g_trace(G_T_ACCESS, "g_dev_close(%s, %d, %d, %p)",
314             cp->geom->name, flags, fmt, td);
315         
316         r = flags & FREAD ? -1 : 0;
317         w = flags & FWRITE ? -1 : 0;
318 #ifdef notyet
319         e = flags & O_EXCL ? -1 : 0;
320 #else
321         e = 0;
322 #endif
323         sc = cp->private;
324         mtx_lock(&sc->sc_mtx);
325         sc->sc_open += r + w + e;
326         while (sc->sc_open == 0 && sc->sc_active != 0)
327                 msleep(&sc->sc_active, &sc->sc_mtx, 0, "PRIBIO", 0);
328         mtx_unlock(&sc->sc_mtx);
329         g_topology_lock();
330         error = g_access(cp, r, w, e);
331         g_topology_unlock();
332         return (error);
333 }
334
335 /*
336  * XXX: Until we have unmessed the ioctl situation, there is a race against
337  * XXX: a concurrent orphanization.  We cannot close it by holding topology
338  * XXX: since that would prevent us from doing our job, and stalling events
339  * XXX: will break (actually: stall) the BSD disklabel hacks.
340  */
341 static int
342 g_dev_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td)
343 {
344         struct g_consumer *cp;
345         struct g_provider *pp;
346         struct g_kerneldump kd;
347         off_t offset, length, chunk;
348         int i, error;
349         u_int u;
350
351         cp = dev->si_drv2;
352         pp = cp->provider;
353
354         error = 0;
355         KASSERT(cp->acr || cp->acw,
356             ("Consumer with zero access count in g_dev_ioctl"));
357
358         i = IOCPARM_LEN(cmd);
359         switch (cmd) {
360         case DIOCGSECTORSIZE:
361                 *(u_int *)data = cp->provider->sectorsize;
362                 if (*(u_int *)data == 0)
363                         error = ENOENT;
364                 break;
365         case DIOCGMEDIASIZE:
366                 *(off_t *)data = cp->provider->mediasize;
367                 if (*(off_t *)data == 0)
368                         error = ENOENT;
369                 break;
370         case DIOCGFWSECTORS:
371                 error = g_io_getattr("GEOM::fwsectors", cp, &i, data);
372                 if (error == 0 && *(u_int *)data == 0)
373                         error = ENOENT;
374                 break;
375         case DIOCGFWHEADS:
376                 error = g_io_getattr("GEOM::fwheads", cp, &i, data);
377                 if (error == 0 && *(u_int *)data == 0)
378                         error = ENOENT;
379                 break;
380         case DIOCGFRONTSTUFF:
381                 error = g_io_getattr("GEOM::frontstuff", cp, &i, data);
382                 break;
383         case DIOCSKERNELDUMP:
384                 u = *((u_int *)data);
385                 if (!u) {
386                         set_dumper(NULL);
387                         error = 0;
388                         break;
389                 }
390                 kd.offset = 0;
391                 kd.length = OFF_MAX;
392                 i = sizeof kd;
393                 error = g_io_getattr("GEOM::kerneldump", cp, &i, &kd);
394                 if (!error) {
395                         error = set_dumper(&kd.di);
396                         if (!error)
397                                 dev->si_flags |= SI_DUMPDEV;
398                 }
399                 break;
400         case DIOCGFLUSH:
401                 error = g_io_flush(cp);
402                 break;
403         case DIOCGDELETE:
404                 offset = ((off_t *)data)[0];
405                 length = ((off_t *)data)[1];
406                 if ((offset % cp->provider->sectorsize) != 0 ||
407                     (length % cp->provider->sectorsize) != 0 || length <= 0) {
408                         printf("%s: offset=%jd length=%jd\n", __func__, offset,
409                             length);
410                         error = EINVAL;
411                         break;
412                 }
413                 while (length > 0) { 
414                         chunk = length;
415                         if (chunk > 65536 * cp->provider->sectorsize)
416                                 chunk = 65536 * cp->provider->sectorsize;
417                         error = g_delete_data(cp, offset, chunk);
418                         length -= chunk;
419                         offset += chunk;
420                         if (error)
421                                 break;
422                         /*
423                          * Since the request size is unbounded, the service
424                          * time is likewise.  We make this ioctl interruptible
425                          * by checking for signals for each bio.
426                          */
427                         if (SIGPENDING(td))
428                                 break;
429                 }
430                 break;
431         case DIOCGIDENT:
432                 error = g_io_getattr("GEOM::ident", cp, &i, data);
433                 break;
434         case DIOCGPROVIDERNAME:
435                 if (pp == NULL)
436                         return (ENOENT);
437                 strlcpy(data, pp->name, i);
438                 break;
439         case DIOCGSTRIPESIZE:
440                 *(off_t *)data = cp->provider->stripesize;
441                 break;
442         case DIOCGSTRIPEOFFSET:
443                 *(off_t *)data = cp->provider->stripeoffset;
444                 break;
445         case DIOCGPHYSPATH:
446                 error = g_io_getattr("GEOM::physpath", cp, &i, data);
447                 if (error == 0 && *(char *)data == '\0')
448                         error = ENOENT;
449                 break;
450         default:
451                 if (cp->provider->geom->ioctl != NULL) {
452                         error = cp->provider->geom->ioctl(cp->provider, cmd, data, fflag, td);
453                 } else {
454                         error = ENOIOCTL;
455                 }
456         }
457
458         return (error);
459 }
460
461 static void
462 g_dev_done(struct bio *bp2)
463 {
464         struct g_consumer *cp;
465         struct g_dev_softc *sc;
466         struct bio *bp;
467         int destroy;
468
469         cp = bp2->bio_from;
470         sc = cp->private;
471         bp = bp2->bio_parent;
472         bp->bio_error = bp2->bio_error;
473         if (bp->bio_error != 0) {
474                 g_trace(G_T_BIO, "g_dev_done(%p) had error %d",
475                     bp2, bp->bio_error);
476                 bp->bio_flags |= BIO_ERROR;
477         } else {
478                 g_trace(G_T_BIO, "g_dev_done(%p/%p) resid %ld completed %jd",
479                     bp2, bp, bp->bio_resid, (intmax_t)bp2->bio_completed);
480         }
481         bp->bio_resid = bp->bio_length - bp2->bio_completed;
482         bp->bio_completed = bp2->bio_completed;
483         g_destroy_bio(bp2);
484         destroy = 0;
485         mtx_lock(&sc->sc_mtx);
486         if ((--sc->sc_active) == 0) {
487                 if (sc->sc_open == 0)
488                         wakeup(&sc->sc_active);
489                 if (sc->sc_dev == NULL)
490                         destroy = 1;
491         }
492         mtx_unlock(&sc->sc_mtx);
493         if (destroy)
494                 g_post_event(g_dev_destroy, cp, M_WAITOK, NULL);
495         biodone(bp);
496 }
497
498 static void
499 g_dev_strategy(struct bio *bp)
500 {
501         struct g_consumer *cp;
502         struct bio *bp2;
503         struct cdev *dev;
504         struct g_dev_softc *sc;
505
506         KASSERT(bp->bio_cmd == BIO_READ ||
507                 bp->bio_cmd == BIO_WRITE ||
508                 bp->bio_cmd == BIO_DELETE,
509                 ("Wrong bio_cmd bio=%p cmd=%d", bp, bp->bio_cmd));
510         dev = bp->bio_dev;
511         cp = dev->si_drv2;
512         sc = cp->private;
513         KASSERT(cp->acr || cp->acw,
514             ("Consumer with zero access count in g_dev_strategy"));
515 #ifdef INVARIANTS
516         if ((bp->bio_offset % cp->provider->sectorsize) != 0 ||
517             (bp->bio_bcount % cp->provider->sectorsize) != 0) {
518                 bp->bio_resid = bp->bio_bcount;
519                 biofinish(bp, NULL, EINVAL);
520                 return;
521         }
522 #endif
523         mtx_lock(&sc->sc_mtx);
524         KASSERT(sc->sc_open > 0, ("Closed device in g_dev_strategy"));
525         sc->sc_active++;
526         mtx_unlock(&sc->sc_mtx);
527
528         for (;;) {
529                 /*
530                  * XXX: This is not an ideal solution, but I belive it to
531                  * XXX: deadlock safe, all things considered.
532                  */
533                 bp2 = g_clone_bio(bp);
534                 if (bp2 != NULL)
535                         break;
536                 pause("gdstrat", hz / 10);
537         }
538         KASSERT(bp2 != NULL, ("XXX: ENOMEM in a bad place"));
539         bp2->bio_done = g_dev_done;
540         g_trace(G_T_BIO,
541             "g_dev_strategy(%p/%p) offset %jd length %jd data %p cmd %d",
542             bp, bp2, (intmax_t)bp->bio_offset, (intmax_t)bp2->bio_length,
543             bp2->bio_data, bp2->bio_cmd);
544         g_io_request(bp2, cp);
545         KASSERT(cp->acr || cp->acw,
546             ("g_dev_strategy raced with g_dev_close and lost"));
547
548 }
549
550 /*
551  * g_dev_callback()
552  *
553  * Called by devfs when asynchronous device destruction is completed.
554  * - Mark that we have no attached device any more.
555  * - If there are no outstanding requests, schedule geom destruction.
556  *   Otherwise destruction will be scheduled later by g_dev_done().
557  */
558
559 static void
560 g_dev_callback(void *arg)
561 {
562         struct g_consumer *cp;
563         struct g_dev_softc *sc;
564         int destroy;
565
566         cp = arg;
567         sc = cp->private;
568         g_trace(G_T_TOPOLOGY, "g_dev_callback(%p(%s))", cp, cp->geom->name);
569
570         mtx_lock(&sc->sc_mtx);
571         sc->sc_dev = NULL;
572         sc->sc_alias = NULL;
573         destroy = (sc->sc_active == 0);
574         mtx_unlock(&sc->sc_mtx);
575         if (destroy)
576                 g_post_event(g_dev_destroy, cp, M_WAITOK, NULL);
577 }
578
579 /*
580  * g_dev_orphan()
581  *
582  * Called from below when the provider orphaned us.
583  * - Clear any dump settings.
584  * - Request asynchronous device destruction to prevent any more requests
585  *   from coming in.  The provider is already marked with an error, so
586  *   anything which comes in in the interrim will be returned immediately.
587  */
588
589 static void
590 g_dev_orphan(struct g_consumer *cp)
591 {
592         struct cdev *dev;
593         struct g_dev_softc *sc;
594
595         g_topology_assert();
596         sc = cp->private;
597         dev = sc->sc_dev;
598         g_trace(G_T_TOPOLOGY, "g_dev_orphan(%p(%s))", cp, cp->geom->name);
599
600         /* Reset any dump-area set on this device */
601         if (dev->si_flags & SI_DUMPDEV)
602                 set_dumper(NULL);
603
604         /* Destroy the struct cdev *so we get no more requests */
605         destroy_dev_sched_cb(dev, g_dev_callback, cp);
606 }
607
608 DECLARE_GEOM_CLASS(g_dev_class, g_dev);