]> CyberLeo.Net >> Repos - FreeBSD/releng/9.2.git/blob - sys/geom/geom_dev.c
- Copy stable/9 to releng/9.2 as part of the 9.2-RELEASE cycle.
[FreeBSD/releng/9.2.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 | D_UNMAPPED_IO,
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         sc->sc_dev = dev;
241
242         /* Search for device alias name and create it if found. */
243         adev = NULL;
244         for (len = MIN(strlen(gp->name), sizeof(buf) - 15); len > 0; len--) {
245                 snprintf(buf, sizeof(buf), "kern.devalias.%s", gp->name);
246                 buf[14 + len] = 0;
247                 val = getenv(buf);
248                 if (val != NULL) {
249                         snprintf(buf, sizeof(buf), "%s%s",
250                             val, gp->name + len);
251                         freeenv(val);
252                         make_dev_alias_p(MAKEDEV_CHECKNAME | MAKEDEV_WAITOK,
253                             &adev, dev, "%s", buf);
254                         break;
255                 }
256         }
257
258         if (pp->flags & G_PF_CANDELETE)
259                 dev->si_flags |= SI_CANDELETE;
260         dev->si_iosize_max = MAXPHYS;
261         dev->si_drv2 = cp;
262         if (adev != NULL) {
263                 if (pp->flags & G_PF_CANDELETE)
264                         adev->si_flags |= SI_CANDELETE;
265                 adev->si_iosize_max = MAXPHYS;
266                 adev->si_drv2 = cp;
267         }
268
269         g_dev_attrchanged(cp, "GEOM::physpath");
270
271         return (gp);
272 }
273
274 static int
275 g_dev_open(struct cdev *dev, int flags, int fmt, struct thread *td)
276 {
277         struct g_consumer *cp;
278         struct g_dev_softc *sc;
279         int error, r, w, e;
280
281         cp = dev->si_drv2;
282         if (cp == NULL)
283                 return(ENXIO);          /* g_dev_taste() not done yet */
284         g_trace(G_T_ACCESS, "g_dev_open(%s, %d, %d, %p)",
285             cp->geom->name, flags, fmt, td);
286
287         r = flags & FREAD ? 1 : 0;
288         w = flags & FWRITE ? 1 : 0;
289 #ifdef notyet
290         e = flags & O_EXCL ? 1 : 0;
291 #else
292         e = 0;
293 #endif
294         if (w) {
295                 /*
296                  * When running in very secure mode, do not allow
297                  * opens for writing of any disks.
298                  */
299                 error = securelevel_ge(td->td_ucred, 2);
300                 if (error)
301                         return (error);
302         }
303         g_topology_lock();
304         error = g_access(cp, r, w, e);
305         g_topology_unlock();
306         if (error == 0) {
307                 sc = cp->private;
308                 mtx_lock(&sc->sc_mtx);
309                 if (sc->sc_open == 0 && sc->sc_active != 0)
310                         wakeup(&sc->sc_active);
311                 sc->sc_open += r + w + e;
312                 mtx_unlock(&sc->sc_mtx);
313         }
314         return(error);
315 }
316
317 static int
318 g_dev_close(struct cdev *dev, int flags, int fmt, struct thread *td)
319 {
320         struct g_consumer *cp;
321         struct g_dev_softc *sc;
322         int error, r, w, e;
323
324         cp = dev->si_drv2;
325         if (cp == NULL)
326                 return(ENXIO);
327         g_trace(G_T_ACCESS, "g_dev_close(%s, %d, %d, %p)",
328             cp->geom->name, flags, fmt, td);
329         
330         r = flags & FREAD ? -1 : 0;
331         w = flags & FWRITE ? -1 : 0;
332 #ifdef notyet
333         e = flags & O_EXCL ? -1 : 0;
334 #else
335         e = 0;
336 #endif
337         sc = cp->private;
338         mtx_lock(&sc->sc_mtx);
339         sc->sc_open += r + w + e;
340         while (sc->sc_open == 0 && sc->sc_active != 0)
341                 msleep(&sc->sc_active, &sc->sc_mtx, 0, "PRIBIO", 0);
342         mtx_unlock(&sc->sc_mtx);
343         g_topology_lock();
344         error = g_access(cp, r, w, e);
345         g_topology_unlock();
346         return (error);
347 }
348
349 /*
350  * XXX: Until we have unmessed the ioctl situation, there is a race against
351  * XXX: a concurrent orphanization.  We cannot close it by holding topology
352  * XXX: since that would prevent us from doing our job, and stalling events
353  * XXX: will break (actually: stall) the BSD disklabel hacks.
354  */
355 static int
356 g_dev_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td)
357 {
358         struct g_consumer *cp;
359         struct g_provider *pp;
360         struct g_kerneldump kd;
361         off_t offset, length, chunk;
362         int i, error;
363         u_int u;
364
365         cp = dev->si_drv2;
366         pp = cp->provider;
367
368         error = 0;
369         KASSERT(cp->acr || cp->acw,
370             ("Consumer with zero access count in g_dev_ioctl"));
371
372         i = IOCPARM_LEN(cmd);
373         switch (cmd) {
374         case DIOCGSECTORSIZE:
375                 *(u_int *)data = cp->provider->sectorsize;
376                 if (*(u_int *)data == 0)
377                         error = ENOENT;
378                 break;
379         case DIOCGMEDIASIZE:
380                 *(off_t *)data = cp->provider->mediasize;
381                 if (*(off_t *)data == 0)
382                         error = ENOENT;
383                 break;
384         case DIOCGFWSECTORS:
385                 error = g_io_getattr("GEOM::fwsectors", cp, &i, data);
386                 if (error == 0 && *(u_int *)data == 0)
387                         error = ENOENT;
388                 break;
389         case DIOCGFWHEADS:
390                 error = g_io_getattr("GEOM::fwheads", cp, &i, data);
391                 if (error == 0 && *(u_int *)data == 0)
392                         error = ENOENT;
393                 break;
394         case DIOCGFRONTSTUFF:
395                 error = g_io_getattr("GEOM::frontstuff", cp, &i, data);
396                 break;
397         case DIOCSKERNELDUMP:
398                 u = *((u_int *)data);
399                 if (!u) {
400                         set_dumper(NULL);
401                         error = 0;
402                         break;
403                 }
404                 kd.offset = 0;
405                 kd.length = OFF_MAX;
406                 i = sizeof kd;
407                 error = g_io_getattr("GEOM::kerneldump", cp, &i, &kd);
408                 if (!error) {
409                         error = set_dumper(&kd.di);
410                         if (!error)
411                                 dev->si_flags |= SI_DUMPDEV;
412                 }
413                 break;
414         case DIOCGFLUSH:
415                 error = g_io_flush(cp);
416                 break;
417         case DIOCGDELETE:
418                 offset = ((off_t *)data)[0];
419                 length = ((off_t *)data)[1];
420                 if ((offset % cp->provider->sectorsize) != 0 ||
421                     (length % cp->provider->sectorsize) != 0 || length <= 0) {
422                         printf("%s: offset=%jd length=%jd\n", __func__, offset,
423                             length);
424                         error = EINVAL;
425                         break;
426                 }
427                 while (length > 0) { 
428                         chunk = length;
429                         if (g_dev_del_max_sectors != 0 && chunk >
430                             g_dev_del_max_sectors * cp->provider->sectorsize) {
431                                 chunk = g_dev_del_max_sectors *
432                                     cp->provider->sectorsize;
433                         }
434                         error = g_delete_data(cp, offset, chunk);
435                         length -= chunk;
436                         offset += chunk;
437                         if (error)
438                                 break;
439                         /*
440                          * Since the request size can be large, the service
441                          * time can be is likewise.  We make this ioctl
442                          * interruptible by checking for signals for each bio.
443                          */
444                         if (SIGPENDING(td))
445                                 break;
446                 }
447                 break;
448         case DIOCGIDENT:
449                 error = g_io_getattr("GEOM::ident", cp, &i, data);
450                 break;
451         case DIOCGPROVIDERNAME:
452                 if (pp == NULL)
453                         return (ENOENT);
454                 strlcpy(data, pp->name, i);
455                 break;
456         case DIOCGSTRIPESIZE:
457                 *(off_t *)data = cp->provider->stripesize;
458                 break;
459         case DIOCGSTRIPEOFFSET:
460                 *(off_t *)data = cp->provider->stripeoffset;
461                 break;
462         case DIOCGPHYSPATH:
463                 error = g_io_getattr("GEOM::physpath", cp, &i, data);
464                 if (error == 0 && *(char *)data == '\0')
465                         error = ENOENT;
466                 break;
467         default:
468                 if (cp->provider->geom->ioctl != NULL) {
469                         error = cp->provider->geom->ioctl(cp->provider, cmd, data, fflag, td);
470                 } else {
471                         error = ENOIOCTL;
472                 }
473         }
474
475         return (error);
476 }
477
478 static void
479 g_dev_done(struct bio *bp2)
480 {
481         struct g_consumer *cp;
482         struct g_dev_softc *sc;
483         struct bio *bp;
484         int destroy;
485
486         cp = bp2->bio_from;
487         sc = cp->private;
488         bp = bp2->bio_parent;
489         bp->bio_error = bp2->bio_error;
490         if (bp->bio_error != 0) {
491                 g_trace(G_T_BIO, "g_dev_done(%p) had error %d",
492                     bp2, bp->bio_error);
493                 bp->bio_flags |= BIO_ERROR;
494         } else {
495                 g_trace(G_T_BIO, "g_dev_done(%p/%p) resid %ld completed %jd",
496                     bp2, bp, bp->bio_resid, (intmax_t)bp2->bio_completed);
497         }
498         bp->bio_resid = bp->bio_length - bp2->bio_completed;
499         bp->bio_completed = bp2->bio_completed;
500         g_destroy_bio(bp2);
501         destroy = 0;
502         mtx_lock(&sc->sc_mtx);
503         if ((--sc->sc_active) == 0) {
504                 if (sc->sc_open == 0)
505                         wakeup(&sc->sc_active);
506                 if (sc->sc_dev == NULL)
507                         destroy = 1;
508         }
509         mtx_unlock(&sc->sc_mtx);
510         if (destroy)
511                 g_post_event(g_dev_destroy, cp, M_WAITOK, NULL);
512         biodone(bp);
513 }
514
515 static void
516 g_dev_strategy(struct bio *bp)
517 {
518         struct g_consumer *cp;
519         struct bio *bp2;
520         struct cdev *dev;
521         struct g_dev_softc *sc;
522
523         KASSERT(bp->bio_cmd == BIO_READ ||
524                 bp->bio_cmd == BIO_WRITE ||
525                 bp->bio_cmd == BIO_DELETE ||
526                 bp->bio_cmd == BIO_FLUSH,
527                 ("Wrong bio_cmd bio=%p cmd=%d", bp, bp->bio_cmd));
528         dev = bp->bio_dev;
529         cp = dev->si_drv2;
530         sc = cp->private;
531         KASSERT(cp->acr || cp->acw,
532             ("Consumer with zero access count in g_dev_strategy"));
533 #ifdef INVARIANTS
534         if ((bp->bio_offset % cp->provider->sectorsize) != 0 ||
535             (bp->bio_bcount % cp->provider->sectorsize) != 0) {
536                 bp->bio_resid = bp->bio_bcount;
537                 biofinish(bp, NULL, EINVAL);
538                 return;
539         }
540 #endif
541         mtx_lock(&sc->sc_mtx);
542         KASSERT(sc->sc_open > 0, ("Closed device in g_dev_strategy"));
543         sc->sc_active++;
544         mtx_unlock(&sc->sc_mtx);
545
546         for (;;) {
547                 /*
548                  * XXX: This is not an ideal solution, but I belive it to
549                  * XXX: deadlock safe, all things considered.
550                  */
551                 bp2 = g_clone_bio(bp);
552                 if (bp2 != NULL)
553                         break;
554                 pause("gdstrat", hz / 10);
555         }
556         KASSERT(bp2 != NULL, ("XXX: ENOMEM in a bad place"));
557         bp2->bio_done = g_dev_done;
558         g_trace(G_T_BIO,
559             "g_dev_strategy(%p/%p) offset %jd length %jd data %p cmd %d",
560             bp, bp2, (intmax_t)bp->bio_offset, (intmax_t)bp2->bio_length,
561             bp2->bio_data, bp2->bio_cmd);
562         g_io_request(bp2, cp);
563         KASSERT(cp->acr || cp->acw,
564             ("g_dev_strategy raced with g_dev_close and lost"));
565
566 }
567
568 /*
569  * g_dev_callback()
570  *
571  * Called by devfs when asynchronous device destruction is completed.
572  * - Mark that we have no attached device any more.
573  * - If there are no outstanding requests, schedule geom destruction.
574  *   Otherwise destruction will be scheduled later by g_dev_done().
575  */
576
577 static void
578 g_dev_callback(void *arg)
579 {
580         struct g_consumer *cp;
581         struct g_dev_softc *sc;
582         int destroy;
583
584         cp = arg;
585         sc = cp->private;
586         g_trace(G_T_TOPOLOGY, "g_dev_callback(%p(%s))", cp, cp->geom->name);
587
588         mtx_lock(&sc->sc_mtx);
589         sc->sc_dev = NULL;
590         sc->sc_alias = NULL;
591         destroy = (sc->sc_active == 0);
592         mtx_unlock(&sc->sc_mtx);
593         if (destroy)
594                 g_post_event(g_dev_destroy, cp, M_WAITOK, NULL);
595 }
596
597 /*
598  * g_dev_orphan()
599  *
600  * Called from below when the provider orphaned us.
601  * - Clear any dump settings.
602  * - Request asynchronous device destruction to prevent any more requests
603  *   from coming in.  The provider is already marked with an error, so
604  *   anything which comes in in the interrim will be returned immediately.
605  */
606
607 static void
608 g_dev_orphan(struct g_consumer *cp)
609 {
610         struct cdev *dev;
611         struct g_dev_softc *sc;
612
613         g_topology_assert();
614         sc = cp->private;
615         dev = sc->sc_dev;
616         g_trace(G_T_TOPOLOGY, "g_dev_orphan(%p(%s))", cp, cp->geom->name);
617
618         /* Reset any dump-area set on this device */
619         if (dev->si_flags & SI_DUMPDEV)
620                 set_dumper(NULL);
621
622         /* Destroy the struct cdev *so we get no more requests */
623         destroy_dev_sched_cb(dev, g_dev_callback, cp);
624 }
625
626 DECLARE_GEOM_CLASS(g_dev_class, g_dev);