Merge commit 'crater/vendor/OPENPAM'
[dragonfly.git] / sys / dev / disk / md / md.c
1 /*
2  * ----------------------------------------------------------------------------
3  * "THE BEER-WARE LICENSE" (Revision 42):
4  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
5  * can do whatever you want with this stuff. If we meet some day, and you think
6  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
7  * ----------------------------------------------------------------------------
8  *
9  * $FreeBSD: src/sys/dev/md/md.c,v 1.8.2.2 2002/08/19 17:43:34 jdp Exp $
10  * $DragonFly: src/sys/dev/disk/md/md.c,v 1.20 2008/09/07 08:09:39 swildner Exp $
11  *
12  */
13
14 #include "opt_md.h"             /* We have adopted some tasks from MFS */
15
16 #include <sys/param.h>
17 #include <sys/systm.h>
18 #include <sys/buf.h>
19 #include <sys/conf.h>
20 #include <sys/devicestat.h>
21 #include <sys/disk.h>
22 #include <sys/kernel.h>
23 #include <sys/malloc.h>
24 #include <sys/sysctl.h>
25 #include <sys/linker.h>
26 #include <sys/proc.h>
27 #include <sys/buf2.h>
28 #include <sys/thread2.h>
29
30 #ifndef MD_NSECT
31 #define MD_NSECT (10000 * 2)
32 #endif
33
34 MALLOC_DEFINE(M_MD, "MD disk", "Memory Disk");
35 MALLOC_DEFINE(M_MDSECT, "MD sectors", "Memory Disk Sectors");
36
37 static int md_debug;
38 SYSCTL_INT(_debug, OID_AUTO, mddebug, CTLFLAG_RW, &md_debug, 0, "");
39
40 #if defined(MD_ROOT) && defined(MD_ROOT_SIZE)
41 /* Image gets put here: */
42 static u_char mfs_root[MD_ROOT_SIZE*1024] = "MFS Filesystem goes here";
43 static u_char end_mfs_root[] __unused = "MFS Filesystem had better STOP here";
44 #endif
45
46 static int mdrootready;
47
48 static void mdcreate_malloc(void);
49
50 #define CDEV_MAJOR      95
51
52 static d_strategy_t mdstrategy;
53 static d_strategy_t mdstrategy_preload;
54 static d_strategy_t mdstrategy_malloc;
55 static d_open_t mdopen;
56 static d_ioctl_t mdioctl;
57
58 static struct dev_ops md_ops = {
59         { "md", CDEV_MAJOR, D_DISK | D_CANFREE | D_MEMDISK },
60         .d_open =       mdopen,
61         .d_close =      nullclose,
62         .d_read =       physread,
63         .d_write =      physwrite,
64         .d_ioctl =      mdioctl,
65         .d_strategy =   mdstrategy,
66 };
67
68 struct md_s {
69         int unit;
70         struct devstat stats;
71         struct bio_queue_head bio_queue;
72         struct disk disk;
73         cdev_t dev;
74         int busy;
75         enum {MD_MALLOC, MD_PRELOAD} type;
76         unsigned nsect;
77
78         /* MD_MALLOC related fields */
79         unsigned nsecp;
80         u_char **secp;
81
82         /* MD_PRELOAD related fields */
83         u_char *pl_ptr;
84         unsigned pl_len;
85 };
86
87 static int mdunits;
88
89 static int
90 mdopen(struct dev_open_args *ap)
91 {
92         cdev_t dev = ap->a_head.a_dev;
93         struct md_s *sc;
94         struct disk_info info;
95
96         if (md_debug)
97                 kprintf("mdopen(%s %x %x)\n",
98                         devtoname(dev), ap->a_oflags, ap->a_devtype);
99
100         sc = dev->si_drv1;
101         if (sc->unit + 1 == mdunits)
102                 mdcreate_malloc();
103
104         bzero(&info, sizeof(info));
105         info.d_media_blksize = DEV_BSIZE;       /* mandatory */
106         info.d_media_blocks = sc->nsect;
107
108         info.d_secpertrack = 1024;              /* optional */
109         info.d_nheads = 1;
110         info.d_secpercyl = info.d_secpertrack * info.d_nheads;
111         info.d_ncylinders = (u_int)(info.d_media_blocks / info.d_secpercyl);
112         disk_setdiskinfo(&sc->disk, &info);
113
114         return (0);
115 }
116
117 static int
118 mdioctl(struct dev_ioctl_args *ap)
119 {
120         cdev_t dev = ap->a_head.a_dev;
121
122         if (md_debug)
123                 kprintf("mdioctl(%s %lx %p %x)\n",
124                         devtoname(dev), ap->a_cmd, ap->a_data, ap->a_fflag);
125
126         return (ENOIOCTL);
127 }
128
129 static int
130 mdstrategy(struct dev_strategy_args *ap)
131 {
132         cdev_t dev = ap->a_head.a_dev;
133         struct bio *bio = ap->a_bio;
134         struct buf *bp = bio->bio_buf;
135         struct md_s *sc;
136
137         if (md_debug > 1) {
138                 kprintf("mdstrategy(%p) %s %08x, %lld, %d, %p)\n",
139                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
140                     bp->b_bcount, bp->b_data);
141         }
142         bio->bio_driver_info = dev;
143         sc = dev->si_drv1;
144         if (sc->type == MD_MALLOC) {
145                 mdstrategy_malloc(ap);
146         } else {
147                 mdstrategy_preload(ap);
148         }
149         return(0);
150 }
151
152
153 static int
154 mdstrategy_malloc(struct dev_strategy_args *ap)
155 {
156         cdev_t dev = ap->a_head.a_dev;
157         struct bio *bio = ap->a_bio;
158         struct buf *bp = bio->bio_buf;
159         unsigned secno, nsec, secval, uc;
160         u_char *secp, **secpp, *dst;
161         struct md_s *sc;
162         int i;
163
164         if (md_debug > 1)
165                 kprintf("mdstrategy_malloc(%p) %s %08xx, %lld, %d, %p)\n",
166                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
167                     bp->b_bcount, bp->b_data);
168
169         sc = dev->si_drv1;
170
171         crit_enter();
172
173         bioqdisksort(&sc->bio_queue, bio);
174
175         if (sc->busy) {
176                 crit_exit();
177                 return(0);
178         }
179
180         sc->busy++;
181         
182         while (1) {
183                 bio = bioq_first(&sc->bio_queue);
184                 if (bio == NULL) {
185                         crit_exit();
186                         break;
187                 }
188                 crit_exit();
189                 bioq_remove(&sc->bio_queue, bio);
190                 bp = bio->bio_buf;
191
192                 devstat_start_transaction(&sc->stats);
193
194                 switch (bp->b_cmd) {
195                 case BUF_CMD_FREEBLKS:
196                 case BUF_CMD_READ:
197                 case BUF_CMD_WRITE:
198                         break;
199                 default:
200                         panic("md: bad b_cmd %d", bp->b_cmd);
201                 }
202
203                 nsec = bp->b_bcount >> DEV_BSHIFT;
204                 secno = (unsigned)(bio->bio_offset >> DEV_BSHIFT);
205                 dst = bp->b_data;
206                 while (nsec--) {
207                         if (secno < sc->nsecp) {
208                                 secpp = &sc->secp[secno];
209                                 if ((u_int)*secpp > 255) {
210                                         secp = *secpp;
211                                         secval = 0;
212                                 } else {
213                                         secp = 0;
214                                         secval = (u_int) *secpp;
215                                 }
216                         } else {
217                                 secpp = 0;
218                                 secp = 0;
219                                 secval = 0;
220                         }
221                         if (md_debug > 2)
222                                 kprintf("%08x %p %p %d\n", bp->b_flags, secpp, secp, secval);
223
224                         switch (bp->b_cmd) {
225                         case BUF_CMD_FREEBLKS:
226                                 if (secpp) {
227                                         if (secp)
228                                                 FREE(secp, M_MDSECT);
229                                         *secpp = 0;
230                                 }
231                                 break;
232                         case BUF_CMD_READ:
233                                 if (secp) {
234                                         bcopy(secp, dst, DEV_BSIZE);
235                                 } else if (secval) {
236                                         for (i = 0; i < DEV_BSIZE; i++)
237                                                 dst[i] = secval;
238                                 } else {
239                                         bzero(dst, DEV_BSIZE);
240                                 }
241                                 break;
242                         case BUF_CMD_WRITE:
243                                 uc = dst[0];
244                                 for (i = 1; i < DEV_BSIZE; i++) 
245                                         if (dst[i] != uc)
246                                                 break;
247                                 if (i == DEV_BSIZE && !uc) {
248                                         if (secp)
249                                                 FREE(secp, M_MDSECT);
250                                         if (secpp)
251                                                 *secpp = (u_char *)uc;
252                                 } else {
253                                         if (!secpp) {
254                                                 MALLOC(secpp, u_char **, (secno + nsec + 1) * sizeof(u_char *), M_MD, M_WAITOK | M_ZERO);
255                                                 bcopy(sc->secp, secpp, sc->nsecp * sizeof(u_char *));
256                                                 FREE(sc->secp, M_MD);
257                                                 sc->secp = secpp;
258                                                 sc->nsecp = secno + nsec + 1;
259                                                 secpp = &sc->secp[secno];
260                                         }
261                                         if (i == DEV_BSIZE) {
262                                                 if (secp)
263                                                         FREE(secp, M_MDSECT);
264                                                 *secpp = (u_char *)uc;
265                                         } else {
266                                                 if (!secp) 
267                                                         MALLOC(secp, u_char *, DEV_BSIZE, M_MDSECT, M_WAITOK);
268                                                 bcopy(dst, secp, DEV_BSIZE);
269
270                                                 *secpp = secp;
271                                         }
272                                 }
273                                 break;
274                         default:
275                                 panic("md: bad b_cmd %d", bp->b_cmd);
276
277                         }
278                         secno++;
279                         dst += DEV_BSIZE;
280                 }
281                 bp->b_resid = 0;
282                 devstat_end_transaction_buf(&sc->stats, bp);
283                 biodone(bio);
284                 crit_enter();
285         }
286         sc->busy = 0;
287         return(0);
288 }
289
290
291 static int
292 mdstrategy_preload(struct dev_strategy_args *ap)
293 {
294         cdev_t dev = ap->a_head.a_dev;
295         struct bio *bio = ap->a_bio;
296         struct buf *bp = bio->bio_buf;
297         struct md_s *sc;
298
299         if (md_debug > 1)
300                 kprintf("mdstrategy_preload(%p) %s %08x, %lld, %d, %p)\n",
301                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
302                     bp->b_bcount, bp->b_data);
303
304         sc = dev->si_drv1;
305
306         crit_enter();
307
308         bioqdisksort(&sc->bio_queue, bio);
309
310         if (sc->busy) {
311                 crit_exit();
312                 return(0);
313         }
314
315         sc->busy++;
316         
317         while (1) {
318                 bio = bioq_first(&sc->bio_queue);
319                 if (bio)
320                         bioq_remove(&sc->bio_queue, bio);
321                 crit_exit();
322                 if (bio == NULL)
323                         break;
324
325                 devstat_start_transaction(&sc->stats);
326
327                 switch (bp->b_cmd) {
328                 case BUF_CMD_FREEBLKS:
329                         break;
330                 case BUF_CMD_READ:
331                         bcopy(sc->pl_ptr + bio->bio_offset, 
332                                bp->b_data, bp->b_bcount);
333                         break;
334                 case BUF_CMD_WRITE:
335                         bcopy(bp->b_data, sc->pl_ptr + bio->bio_offset,
336                               bp->b_bcount);
337                         break;
338                 default:
339                         panic("md: bad cmd %d\n", bp->b_cmd);
340                 }
341                 bp->b_resid = 0;
342                 devstat_end_transaction_buf(&sc->stats, bp);
343                 biodone(bio);
344                 crit_enter();
345         }
346         sc->busy = 0;
347         return(0);
348 }
349
350 static struct md_s *
351 mdcreate(void)
352 {
353         struct md_s *sc;
354
355         MALLOC(sc, struct md_s *,sizeof(*sc), M_MD, M_WAITOK | M_ZERO);
356         sc->unit = mdunits++;
357         bioq_init(&sc->bio_queue);
358         devstat_add_entry(&sc->stats, "md", sc->unit, DEV_BSIZE,
359                 DEVSTAT_NO_ORDERED_TAGS, 
360                 DEVSTAT_TYPE_DIRECT | DEVSTAT_TYPE_IF_OTHER,
361                 DEVSTAT_PRIORITY_OTHER);
362         sc->dev = disk_create(sc->unit, &sc->disk, &md_ops);
363         sc->dev->si_drv1 = sc;
364         sc->dev->si_iosize_max = DFLTPHYS;
365
366         return (sc);
367 }
368
369 static void
370 mdcreate_preload(u_char *image, unsigned length)
371 {
372         struct md_s *sc;
373
374         sc = mdcreate();
375         sc->type = MD_PRELOAD;
376         sc->nsect = length / DEV_BSIZE;
377         sc->pl_ptr = image;
378         sc->pl_len = length;
379
380         if (sc->unit == 0) 
381                 mdrootready = 1;
382 }
383
384 static void
385 mdcreate_malloc(void)
386 {
387         struct md_s *sc;
388
389         sc = mdcreate();
390         sc->type = MD_MALLOC;
391
392         sc->nsect = MD_NSECT;   /* for now */
393         MALLOC(sc->secp, u_char **, sizeof(u_char *), M_MD, M_WAITOK | M_ZERO);
394         sc->nsecp = 1;
395         kprintf("md%d: Malloc disk\n", sc->unit);
396 }
397
398 static void
399 md_drvinit(void *unused)
400 {
401
402         caddr_t mod;
403         caddr_t c;
404         u_char *ptr, *name, *type;
405         unsigned len;
406
407 #ifdef MD_ROOT_SIZE
408         mdcreate_preload(mfs_root, MD_ROOT_SIZE*1024);
409 #endif
410         mod = NULL;
411         while ((mod = preload_search_next_name(mod)) != NULL) {
412                 name = (char *)preload_search_info(mod, MODINFO_NAME);
413                 type = (char *)preload_search_info(mod, MODINFO_TYPE);
414                 if (name == NULL)
415                         continue;
416                 if (type == NULL)
417                         continue;
418                 if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
419                         continue;
420                 c = preload_search_info(mod, MODINFO_ADDR);
421                 ptr = *(u_char **)c;
422                 c = preload_search_info(mod, MODINFO_SIZE);
423                 len = *(unsigned *)c;
424                 kprintf("md%d: Preloaded image <%s> %d bytes at %p\n",
425                    mdunits, name, len, ptr);
426                 mdcreate_preload(ptr, len);
427         } 
428         mdcreate_malloc();
429 }
430
431 SYSINIT(mddev,SI_SUB_DRIVERS,SI_ORDER_MIDDLE+CDEV_MAJOR, md_drvinit,NULL)
432
433 #ifdef MD_ROOT
434 static void
435 md_takeroot(void *junk)
436 {
437         if (mdrootready)
438                 rootdevnames[0] = "ufs:/dev/md0c";
439 }
440
441 SYSINIT(md_root, SI_SUB_MOUNT_ROOT, SI_ORDER_FIRST, md_takeroot, NULL);
442 #endif