]> CyberLeo.Net >> Repos - FreeBSD/releng/9.3.git/blob - sys/geom/geom_dev.c
Anticipate when we will announce 9.3-RELEASE.
[FreeBSD/releng/9.3.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 <sys/sysctl.h>
56 #include <geom/geom.h>
57 #include <geom/geom_int.h>
58 #include <machine/stdarg.h>
59
60 struct g_dev_softc {
61         struct mtx       sc_mtx;
62         struct cdev     *sc_dev;
63         struct cdev     *sc_alias;
64         int              sc_open;
65         int              sc_active;
66 };
67
68 static d_open_t         g_dev_open;
69 static d_close_t        g_dev_close;
70 static d_strategy_t     g_dev_strategy;
71 static d_ioctl_t        g_dev_ioctl;
72
73 static struct cdevsw g_dev_cdevsw = {
74         .d_version =    D_VERSION,
75         .d_open =       g_dev_open,
76         .d_close =      g_dev_close,
77         .d_read =       physread,
78         .d_write =      physwrite,
79         .d_ioctl =      g_dev_ioctl,
80         .d_strategy =   g_dev_strategy,
81         .d_name =       "g_dev",
82         .d_flags =      D_DISK | D_TRACKCLOSE,
83 };
84
85 static g_taste_t g_dev_taste;
86 static g_orphan_t g_dev_orphan;
87 static g_attrchanged_t g_dev_attrchanged;
88
89 static struct g_class g_dev_class       = {
90         .name = "DEV",
91         .version = G_VERSION,
92         .taste = g_dev_taste,
93         .orphan = g_dev_orphan,
94         .attrchanged = g_dev_attrchanged
95 };
96
97 /*
98  * We target 262144 (8 x 32768) sectors by default as this significantly
99  * increases the throughput on commonly used SSD's with a marginal
100  * increase in non-interruptible request latency.
101  */
102 static uint64_t g_dev_del_max_sectors = 262144;
103 SYSCTL_DECL(_kern_geom);
104 SYSCTL_NODE(_kern_geom, OID_AUTO, dev, CTLFLAG_RW, 0, "GEOM_DEV stuff");
105 SYSCTL_QUAD(_kern_geom_dev, OID_AUTO, delete_max_sectors, CTLFLAG_RW,
106     &g_dev_del_max_sectors, 0, "Maximum number of sectors in a single "
107     "delete request sent to the provider. Larger requests are chunked "
108     "so they can be interrupted. (0 = disable chunking)");
109
110 static void
111 g_dev_destroy(void *arg, int flags __unused)
112 {
113         struct g_consumer *cp;
114         struct g_geom *gp;
115         struct g_dev_softc *sc;
116
117         g_topology_assert();
118         cp = arg;
119         gp = cp->geom;
120         sc = cp->private;
121         g_trace(G_T_TOPOLOGY, "g_dev_destroy(%p(%s))", cp, gp->name);
122         if (cp->acr > 0 || cp->acw > 0 || cp->ace > 0)
123                 g_access(cp, -cp->acr, -cp->acw, -cp->ace);
124         g_detach(cp);
125         g_destroy_consumer(cp);
126         g_destroy_geom(gp);
127         mtx_destroy(&sc->sc_mtx);
128         g_free(sc);
129 }
130
131 void
132 g_dev_print(void)
133 {
134         struct g_geom *gp;
135         char const *p = "";
136
137         LIST_FOREACH(gp, &g_dev_class.geom, geom) {
138                 printf("%s%s", p, gp->name);
139                 p = " ";
140         }
141         printf("\n");
142 }
143
144 static void
145 g_dev_attrchanged(struct g_consumer *cp, const char *attr)
146 {
147         struct g_dev_softc *sc;
148         struct cdev *dev;
149         char buf[SPECNAMELEN + 6];
150
151         sc = cp->private;
152         if (strcmp(attr, "GEOM::media") == 0) {
153                 dev = sc->sc_dev;
154                 snprintf(buf, sizeof(buf), "cdev=%s", dev->si_name);
155                 devctl_notify_f("DEVFS", "CDEV", "MEDIACHANGE", buf, M_WAITOK);
156                 dev = sc->sc_alias;
157                 if (dev != NULL) {
158                         snprintf(buf, sizeof(buf), "cdev=%s", dev->si_name);
159                         devctl_notify_f("DEVFS", "CDEV", "MEDIACHANGE", buf,
160                             M_WAITOK);
161                 }
162                 return;
163         }
164
165         if (strcmp(attr, "GEOM::physpath") != 0)
166                 return;
167
168         if (g_access(cp, 1, 0, 0) == 0) {
169                 char *physpath;
170                 int error, physpath_len;
171
172                 physpath_len = MAXPATHLEN;
173                 physpath = g_malloc(physpath_len, M_WAITOK|M_ZERO);
174                 error =
175                     g_io_getattr("GEOM::physpath", cp, &physpath_len, physpath);
176                 g_access(cp, -1, 0, 0);
177                 if (error == 0 && strlen(physpath) != 0) {
178                         struct cdev *old_alias_dev;
179                         struct cdev **alias_devp;
180
181                         dev = sc->sc_dev;
182                         old_alias_dev = sc->sc_alias;
183                         alias_devp = (struct cdev **)&sc->sc_alias;
184                         make_dev_physpath_alias(MAKEDEV_WAITOK, alias_devp,
185                             dev, old_alias_dev, physpath);
186                 } else if (sc->sc_alias) {
187                         destroy_dev((struct cdev *)sc->sc_alias);
188                         sc->sc_alias = NULL;
189                 }
190                 g_free(physpath);
191         }
192 }
193
194 struct g_provider *
195 g_dev_getprovider(struct cdev *dev)
196 {
197         struct g_consumer *cp;
198
199         g_topology_assert();
200         if (dev == NULL)
201                 return (NULL);
202         if (dev->si_devsw != &g_dev_cdevsw)
203                 return (NULL);
204         cp = dev->si_drv2;
205         return (cp->provider);
206 }
207
208 static struct g_geom *
209 g_dev_taste(struct g_class *mp, struct g_provider *pp, int insist __unused)
210 {
211         struct g_geom *gp;
212         struct g_consumer *cp;
213         struct g_dev_softc *sc;
214         int error, len;
215         struct cdev *dev, *adev;
216         char buf[64], *val;
217
218         g_trace(G_T_TOPOLOGY, "dev_taste(%s,%s)", mp->name, pp->name);
219         g_topology_assert();
220         gp = g_new_geomf(mp, "%s", pp->name);
221         sc = g_malloc(sizeof(*sc), M_WAITOK | M_ZERO);
222         mtx_init(&sc->sc_mtx, "g_dev", NULL, MTX_DEF);
223         cp = g_new_consumer(gp);
224         cp->private = sc;
225         error = g_attach(cp, pp);
226         KASSERT(error == 0,
227             ("g_dev_taste(%s) failed to g_attach, err=%d", pp->name, error));
228         error = make_dev_p(MAKEDEV_CHECKNAME | MAKEDEV_WAITOK, &dev,
229             &g_dev_cdevsw, NULL, UID_ROOT, GID_OPERATOR, 0640, "%s", gp->name);
230         if (error != 0) {
231                 printf("%s: make_dev_p() failed (gp->name=%s, error=%d)\n",
232                     __func__, gp->name, error);
233                 g_detach(cp);
234                 g_destroy_consumer(cp);
235                 g_destroy_geom(gp);
236                 mtx_destroy(&sc->sc_mtx);
237                 g_free(sc);
238                 return (NULL);
239         }
240         dev->si_flags |= SI_UNMAPPED;
241         sc->sc_dev = dev;
242
243         /* Search for device alias name and create it if found. */
244         adev = NULL;
245         for (len = MIN(strlen(gp->name), sizeof(buf) - 15); len > 0; len--) {
246                 snprintf(buf, sizeof(buf), "kern.devalias.%s", gp->name);
247                 buf[14 + len] = 0;
248                 val = getenv(buf);
249                 if (val != NULL) {
250                         snprintf(buf, sizeof(buf), "%s%s",
251                             val, gp->name + len);
252                         freeenv(val);
253                         make_dev_alias_p(MAKEDEV_CHECKNAME | MAKEDEV_WAITOK,
254                             &adev, dev, "%s", buf);
255                         adev->si_flags |= SI_UNMAPPED;
256                         break;
257                 }
258         }
259
260         if (pp->flags & G_PF_CANDELETE)
261                 dev->si_flags |= SI_CANDELETE;
262         dev->si_iosize_max = MAXPHYS;
263         dev->si_drv2 = cp;
264         if (adev != NULL) {
265                 if (pp->flags & G_PF_CANDELETE)
266                         adev->si_flags |= SI_CANDELETE;
267                 adev->si_iosize_max = MAXPHYS;
268                 adev->si_drv2 = cp;
269         }
270
271         g_dev_attrchanged(cp, "GEOM::physpath");
272
273         return (gp);
274 }
275
276 static int
277 g_dev_open(struct cdev *dev, int flags, int fmt, struct thread *td)
278 {
279         struct g_consumer *cp;
280         struct g_dev_softc *sc;
281         int error, r, w, e;
282
283         cp = dev->si_drv2;
284         if (cp == NULL)
285                 return(ENXIO);          /* g_dev_taste() not done yet */
286         g_trace(G_T_ACCESS, "g_dev_open(%s, %d, %d, %p)",
287             cp->geom->name, flags, fmt, td);
288
289         r = flags & FREAD ? 1 : 0;
290         w = flags & FWRITE ? 1 : 0;
291 #ifdef notyet
292         e = flags & O_EXCL ? 1 : 0;
293 #else
294         e = 0;
295 #endif
296         if (w) {
297                 /*
298                  * When running in very secure mode, do not allow
299                  * opens for writing of any disks.
300                  */
301                 error = securelevel_ge(td->td_ucred, 2);
302                 if (error)
303                         return (error);
304         }
305         g_topology_lock();
306         error = g_access(cp, r, w, e);
307         g_topology_unlock();
308         if (error == 0) {
309                 sc = cp->private;
310                 mtx_lock(&sc->sc_mtx);
311                 if (sc->sc_open == 0 && sc->sc_active != 0)
312                         wakeup(&sc->sc_active);
313                 sc->sc_open += r + w + e;
314                 mtx_unlock(&sc->sc_mtx);
315         }
316         return(error);
317 }
318
319 static int
320 g_dev_close(struct cdev *dev, int flags, int fmt, struct thread *td)
321 {
322         struct g_consumer *cp;
323         struct g_dev_softc *sc;
324         int error, r, w, e;
325
326         cp = dev->si_drv2;
327         if (cp == NULL)
328                 return(ENXIO);
329         g_trace(G_T_ACCESS, "g_dev_close(%s, %d, %d, %p)",
330             cp->geom->name, flags, fmt, td);
331         
332         r = flags & FREAD ? -1 : 0;
333         w = flags & FWRITE ? -1 : 0;
334 #ifdef notyet
335         e = flags & O_EXCL ? -1 : 0;
336 #else
337         e = 0;
338 #endif
339         sc = cp->private;
340         mtx_lock(&sc->sc_mtx);
341         sc->sc_open += r + w + e;
342         while (sc->sc_open == 0 && sc->sc_active != 0)
343                 msleep(&sc->sc_active, &sc->sc_mtx, 0, "PRIBIO", 0);
344         mtx_unlock(&sc->sc_mtx);
345         g_topology_lock();
346         error = g_access(cp, r, w, e);
347         g_topology_unlock();
348         return (error);
349 }
350
351 /*
352  * XXX: Until we have unmessed the ioctl situation, there is a race against
353  * XXX: a concurrent orphanization.  We cannot close it by holding topology
354  * XXX: since that would prevent us from doing our job, and stalling events
355  * XXX: will break (actually: stall) the BSD disklabel hacks.
356  */
357 static int
358 g_dev_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td)
359 {
360         struct g_consumer *cp;
361         struct g_provider *pp;
362         struct g_kerneldump kd;
363         off_t offset, length, chunk;
364         int i, error;
365         u_int u;
366
367         cp = dev->si_drv2;
368         pp = cp->provider;
369
370         error = 0;
371         KASSERT(cp->acr || cp->acw,
372             ("Consumer with zero access count in g_dev_ioctl"));
373
374         i = IOCPARM_LEN(cmd);
375         switch (cmd) {
376         case DIOCGSECTORSIZE:
377                 *(u_int *)data = cp->provider->sectorsize;
378                 if (*(u_int *)data == 0)
379                         error = ENOENT;
380                 break;
381         case DIOCGMEDIASIZE:
382                 *(off_t *)data = cp->provider->mediasize;
383                 if (*(off_t *)data == 0)
384                         error = ENOENT;
385                 break;
386         case DIOCGFWSECTORS:
387                 error = g_io_getattr("GEOM::fwsectors", cp, &i, data);
388                 if (error == 0 && *(u_int *)data == 0)
389                         error = ENOENT;
390                 break;
391         case DIOCGFWHEADS:
392                 error = g_io_getattr("GEOM::fwheads", cp, &i, data);
393                 if (error == 0 && *(u_int *)data == 0)
394                         error = ENOENT;
395                 break;
396         case DIOCGFRONTSTUFF:
397                 error = g_io_getattr("GEOM::frontstuff", cp, &i, data);
398                 break;
399         case DIOCSKERNELDUMP:
400                 u = *((u_int *)data);
401                 if (!u) {
402                         set_dumper(NULL);
403                         error = 0;
404                         break;
405                 }
406                 kd.offset = 0;
407                 kd.length = OFF_MAX;
408                 i = sizeof kd;
409                 error = g_io_getattr("GEOM::kerneldump", cp, &i, &kd);
410                 if (!error) {
411                         error = set_dumper(&kd.di);
412                         if (!error)
413                                 dev->si_flags |= SI_DUMPDEV;
414                 }
415                 break;
416         case DIOCGFLUSH:
417                 error = g_io_flush(cp);
418                 break;
419         case DIOCGDELETE:
420                 offset = ((off_t *)data)[0];
421                 length = ((off_t *)data)[1];
422                 if ((offset % cp->provider->sectorsize) != 0 ||
423                     (length % cp->provider->sectorsize) != 0 || length <= 0) {
424                         printf("%s: offset=%jd length=%jd\n", __func__, offset,
425                             length);
426                         error = EINVAL;
427                         break;
428                 }
429                 while (length > 0) { 
430                         chunk = length;
431                         if (g_dev_del_max_sectors != 0 && chunk >
432                             g_dev_del_max_sectors * cp->provider->sectorsize) {
433                                 chunk = g_dev_del_max_sectors *
434                                     cp->provider->sectorsize;
435                         }
436                         error = g_delete_data(cp, offset, chunk);
437                         length -= chunk;
438                         offset += chunk;
439                         if (error)
440                                 break;
441                         /*
442                          * Since the request size can be large, the service
443                          * time can be is likewise.  We make this ioctl
444                          * interruptible by checking for signals for each bio.
445                          */
446                         if (SIGPENDING(td))
447                                 break;
448                 }
449                 break;
450         case DIOCGIDENT:
451                 error = g_io_getattr("GEOM::ident", cp, &i, data);
452                 break;
453         case DIOCGPROVIDERNAME:
454                 if (pp == NULL)
455                         return (ENOENT);
456                 strlcpy(data, pp->name, i);
457                 break;
458         case DIOCGSTRIPESIZE:
459                 *(off_t *)data = cp->provider->stripesize;
460                 break;
461         case DIOCGSTRIPEOFFSET:
462                 *(off_t *)data = cp->provider->stripeoffset;
463                 break;
464         case DIOCGPHYSPATH:
465                 error = g_io_getattr("GEOM::physpath", cp, &i, data);
466                 if (error == 0 && *(char *)data == '\0')
467                         error = ENOENT;
468                 break;
469         default:
470                 if (cp->provider->geom->ioctl != NULL) {
471                         error = cp->provider->geom->ioctl(cp->provider, cmd, data, fflag, td);
472                 } else {
473                         error = ENOIOCTL;
474                 }
475         }
476
477         return (error);
478 }
479
480 static void
481 g_dev_done(struct bio *bp2)
482 {
483         struct g_consumer *cp;
484         struct g_dev_softc *sc;
485         struct bio *bp;
486         int destroy;
487
488         cp = bp2->bio_from;
489         sc = cp->private;
490         bp = bp2->bio_parent;
491         bp->bio_error = bp2->bio_error;
492         bp->bio_completed = bp2->bio_completed;
493         bp->bio_resid = bp->bio_length - bp2->bio_completed;
494         if (bp2->bio_error != 0) {
495                 g_trace(G_T_BIO, "g_dev_done(%p) had error %d",
496                     bp2, bp2->bio_error);
497                 bp->bio_flags |= BIO_ERROR;
498         } else {
499                 g_trace(G_T_BIO, "g_dev_done(%p/%p) resid %ld completed %jd",
500                     bp2, bp, bp2->bio_resid, (intmax_t)bp2->bio_completed);
501         }
502         g_destroy_bio(bp2);
503         destroy = 0;
504         mtx_lock(&sc->sc_mtx);
505         if ((--sc->sc_active) == 0) {
506                 if (sc->sc_open == 0)
507                         wakeup(&sc->sc_active);
508                 if (sc->sc_dev == NULL)
509                         destroy = 1;
510         }
511         mtx_unlock(&sc->sc_mtx);
512         if (destroy)
513                 g_post_event(g_dev_destroy, cp, M_WAITOK, NULL);
514         biodone(bp);
515 }
516
517 static void
518 g_dev_strategy(struct bio *bp)
519 {
520         struct g_consumer *cp;
521         struct bio *bp2;
522         struct cdev *dev;
523         struct g_dev_softc *sc;
524
525         KASSERT(bp->bio_cmd == BIO_READ ||
526                 bp->bio_cmd == BIO_WRITE ||
527                 bp->bio_cmd == BIO_DELETE ||
528                 bp->bio_cmd == BIO_FLUSH,
529                 ("Wrong bio_cmd bio=%p cmd=%d", bp, bp->bio_cmd));
530         dev = bp->bio_dev;
531         cp = dev->si_drv2;
532         sc = cp->private;
533         KASSERT(cp->acr || cp->acw,
534             ("Consumer with zero access count in g_dev_strategy"));
535 #ifdef INVARIANTS
536         if ((bp->bio_offset % cp->provider->sectorsize) != 0 ||
537             (bp->bio_bcount % cp->provider->sectorsize) != 0) {
538                 bp->bio_resid = bp->bio_bcount;
539                 biofinish(bp, NULL, EINVAL);
540                 return;
541         }
542 #endif
543         mtx_lock(&sc->sc_mtx);
544         KASSERT(sc->sc_open > 0, ("Closed device in g_dev_strategy"));
545         sc->sc_active++;
546         mtx_unlock(&sc->sc_mtx);
547
548         for (;;) {
549                 /*
550                  * XXX: This is not an ideal solution, but I belive it to
551                  * XXX: deadlock safe, all things considered.
552                  */
553                 bp2 = g_clone_bio(bp);
554                 if (bp2 != NULL)
555                         break;
556                 pause("gdstrat", hz / 10);
557         }
558         KASSERT(bp2 != NULL, ("XXX: ENOMEM in a bad place"));
559         bp2->bio_done = g_dev_done;
560         g_trace(G_T_BIO,
561             "g_dev_strategy(%p/%p) offset %jd length %jd data %p cmd %d",
562             bp, bp2, (intmax_t)bp->bio_offset, (intmax_t)bp2->bio_length,
563             bp2->bio_data, bp2->bio_cmd);
564         g_io_request(bp2, cp);
565         KASSERT(cp->acr || cp->acw,
566             ("g_dev_strategy raced with g_dev_close and lost"));
567
568 }
569
570 /*
571  * g_dev_callback()
572  *
573  * Called by devfs when asynchronous device destruction is completed.
574  * - Mark that we have no attached device any more.
575  * - If there are no outstanding requests, schedule geom destruction.
576  *   Otherwise destruction will be scheduled later by g_dev_done().
577  */
578
579 static void
580 g_dev_callback(void *arg)
581 {
582         struct g_consumer *cp;
583         struct g_dev_softc *sc;
584         int destroy;
585
586         cp = arg;
587         sc = cp->private;
588         g_trace(G_T_TOPOLOGY, "g_dev_callback(%p(%s))", cp, cp->geom->name);
589
590         mtx_lock(&sc->sc_mtx);
591         sc->sc_dev = NULL;
592         sc->sc_alias = NULL;
593         destroy = (sc->sc_active == 0);
594         mtx_unlock(&sc->sc_mtx);
595         if (destroy)
596                 g_post_event(g_dev_destroy, cp, M_WAITOK, NULL);
597 }
598
599 /*
600  * g_dev_orphan()
601  *
602  * Called from below when the provider orphaned us.
603  * - Clear any dump settings.
604  * - Request asynchronous device destruction to prevent any more requests
605  *   from coming in.  The provider is already marked with an error, so
606  *   anything which comes in in the interrim will be returned immediately.
607  */
608
609 static void
610 g_dev_orphan(struct g_consumer *cp)
611 {
612         struct cdev *dev;
613         struct g_dev_softc *sc;
614
615         g_topology_assert();
616         sc = cp->private;
617         dev = sc->sc_dev;
618         g_trace(G_T_TOPOLOGY, "g_dev_orphan(%p(%s))", cp, cp->geom->name);
619
620         /* Reset any dump-area set on this device */
621         if (dev->si_flags & SI_DUMPDEV)
622                 set_dumper(NULL);
623
624         /* Destroy the struct cdev *so we get no more requests */
625         destroy_dev_sched_cb(dev, g_dev_callback, cp);
626 }
627
628 DECLARE_GEOM_CLASS(g_dev_class, g_dev);