Merge from vendor branch LIBARCHIVE:
[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 1.16 2007/05/15 00:01:03 dillon 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         devstat_trans_flags dop;
162         struct md_s *sc;
163         int i;
164
165         if (md_debug > 1)
166                 kprintf("mdstrategy_malloc(%p) %s %08xx, %lld, %d, %p)\n",
167                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
168                     bp->b_bcount, bp->b_data);
169
170         sc = dev->si_drv1;
171
172         crit_enter();
173
174         bioqdisksort(&sc->bio_queue, bio);
175
176         if (sc->busy) {
177                 crit_exit();
178                 return(0);
179         }
180
181         sc->busy++;
182         
183         while (1) {
184                 bio = bioq_first(&sc->bio_queue);
185                 if (bp)
186                         bioq_remove(&sc->bio_queue, bio);
187                 crit_exit();
188                 if (bio == NULL)
189                         break;
190
191                 devstat_start_transaction(&sc->stats);
192
193                 switch(bp->b_cmd) {
194                 case BUF_CMD_FREEBLKS:
195                         dop = DEVSTAT_NO_DATA;
196                         break;
197                 case BUF_CMD_READ:
198                         dop = DEVSTAT_READ;
199                         break;
200                 case BUF_CMD_WRITE:
201                         dop = DEVSTAT_WRITE;
202                         break;
203                 default:
204                         panic("md: bad b_cmd %d", bp->b_cmd);
205                 }
206
207                 nsec = bp->b_bcount >> DEV_BSHIFT;
208                 secno = (unsigned)(bio->bio_offset >> DEV_BSHIFT);
209                 dst = bp->b_data;
210                 while (nsec--) {
211                         if (secno < sc->nsecp) {
212                                 secpp = &sc->secp[secno];
213                                 if ((u_int)*secpp > 255) {
214                                         secp = *secpp;
215                                         secval = 0;
216                                 } else {
217                                         secp = 0;
218                                         secval = (u_int) *secpp;
219                                 }
220                         } else {
221                                 secpp = 0;
222                                 secp = 0;
223                                 secval = 0;
224                         }
225                         if (md_debug > 2)
226                                 kprintf("%08x %p %p %d\n", bp->b_flags, secpp, secp, secval);
227
228                         switch(bp->b_cmd) {
229                         case BUF_CMD_FREEBLKS:
230                                 if (secpp) {
231                                         if (secp)
232                                                 FREE(secp, M_MDSECT);
233                                         *secpp = 0;
234                                 }
235                                 break;
236                         case BUF_CMD_READ:
237                                 if (secp) {
238                                         bcopy(secp, dst, DEV_BSIZE);
239                                 } else if (secval) {
240                                         for (i = 0; i < DEV_BSIZE; i++)
241                                                 dst[i] = secval;
242                                 } else {
243                                         bzero(dst, DEV_BSIZE);
244                                 }
245                                 break;
246                         case BUF_CMD_WRITE:
247                                 uc = dst[0];
248                                 for (i = 1; i < DEV_BSIZE; i++) 
249                                         if (dst[i] != uc)
250                                                 break;
251                                 if (i == DEV_BSIZE && !uc) {
252                                         if (secp)
253                                                 FREE(secp, M_MDSECT);
254                                         if (secpp)
255                                                 *secpp = (u_char *)uc;
256                                 } else {
257                                         if (!secpp) {
258                                                 MALLOC(secpp, u_char **, (secno + nsec + 1) * sizeof(u_char *), M_MD, M_WAITOK);
259                                                 bzero(secpp, (secno + nsec + 1) * sizeof(u_char *));
260                                                 bcopy(sc->secp, secpp, sc->nsecp * sizeof(u_char *));
261                                                 FREE(sc->secp, M_MD);
262                                                 sc->secp = secpp;
263                                                 sc->nsecp = secno + nsec + 1;
264                                                 secpp = &sc->secp[secno];
265                                         }
266                                         if (i == DEV_BSIZE) {
267                                                 if (secp)
268                                                         FREE(secp, M_MDSECT);
269                                                 *secpp = (u_char *)uc;
270                                         } else {
271                                                 if (!secp) 
272                                                         MALLOC(secp, u_char *, DEV_BSIZE, M_MDSECT, M_WAITOK);
273                                                 bcopy(dst, secp, DEV_BSIZE);
274
275                                                 *secpp = secp;
276                                         }
277                                 }
278                                 break;
279                         default:
280                                 panic("md: bad b_cmd %d", bp->b_cmd);
281
282                         }
283                         secno++;
284                         dst += DEV_BSIZE;
285                 }
286                 bp->b_resid = 0;
287                 devstat_end_transaction_buf(&sc->stats, bp);
288                 biodone(bio);
289                 crit_enter();
290         }
291         sc->busy = 0;
292         return(0);
293 }
294
295
296 static int
297 mdstrategy_preload(struct dev_strategy_args *ap)
298 {
299         cdev_t dev = ap->a_head.a_dev;
300         struct bio *bio = ap->a_bio;
301         struct buf *bp = bio->bio_buf;
302         devstat_trans_flags dop;
303         struct md_s *sc;
304
305         if (md_debug > 1)
306                 kprintf("mdstrategy_preload(%p) %s %08x, %lld, %d, %p)\n",
307                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
308                     bp->b_bcount, bp->b_data);
309
310         sc = dev->si_drv1;
311
312         crit_enter();
313
314         bioqdisksort(&sc->bio_queue, bio);
315
316         if (sc->busy) {
317                 crit_exit();
318                 return(0);
319         }
320
321         sc->busy++;
322         
323         while (1) {
324                 bio = bioq_first(&sc->bio_queue);
325                 if (bio)
326                         bioq_remove(&sc->bio_queue, bio);
327                 crit_exit();
328                 if (bio == NULL)
329                         break;
330
331                 devstat_start_transaction(&sc->stats);
332
333                 switch(bp->b_cmd) {
334                 case BUF_CMD_FREEBLKS:
335                         dop = DEVSTAT_NO_DATA;
336                         break;
337                 case BUF_CMD_READ:
338                         dop = DEVSTAT_READ;
339                         bcopy(sc->pl_ptr + bio->bio_offset, 
340                                bp->b_data, bp->b_bcount);
341                         break;
342                 case BUF_CMD_WRITE:
343                         dop = DEVSTAT_WRITE;
344                         bcopy(bp->b_data, sc->pl_ptr + bio->bio_offset,
345                               bp->b_bcount);
346                         break;
347                 default:
348                         panic("md: bad cmd %d\n", bp->b_cmd);
349                 }
350                 bp->b_resid = 0;
351                 devstat_end_transaction_buf(&sc->stats, bp);
352                 biodone(bio);
353                 crit_enter();
354         }
355         sc->busy = 0;
356         return(0);
357 }
358
359 static struct md_s *
360 mdcreate(void)
361 {
362         struct md_s *sc;
363
364         MALLOC(sc, struct md_s *,sizeof(*sc), M_MD, M_WAITOK);
365         bzero(sc, sizeof(*sc));
366         sc->unit = mdunits++;
367         bioq_init(&sc->bio_queue);
368         devstat_add_entry(&sc->stats, "md", sc->unit, DEV_BSIZE,
369                 DEVSTAT_NO_ORDERED_TAGS, 
370                 DEVSTAT_TYPE_DIRECT | DEVSTAT_TYPE_IF_OTHER,
371                 DEVSTAT_PRIORITY_OTHER);
372         sc->dev = disk_create(sc->unit, &sc->disk, &md_ops);
373         sc->dev->si_drv1 = sc;
374         return (sc);
375 }
376
377 static void
378 mdcreate_preload(u_char *image, unsigned length)
379 {
380         struct md_s *sc;
381
382         sc = mdcreate();
383         sc->type = MD_PRELOAD;
384         sc->nsect = length / DEV_BSIZE;
385         sc->pl_ptr = image;
386         sc->pl_len = length;
387
388         if (sc->unit == 0) 
389                 mdrootready = 1;
390 }
391
392 static void
393 mdcreate_malloc(void)
394 {
395         struct md_s *sc;
396
397         sc = mdcreate();
398         sc->type = MD_MALLOC;
399
400         sc->nsect = MD_NSECT;   /* for now */
401         MALLOC(sc->secp, u_char **, sizeof(u_char *), M_MD, M_WAITOK);
402         bzero(sc->secp, sizeof(u_char *));
403         sc->nsecp = 1;
404         kprintf("md%d: Malloc disk\n", sc->unit);
405 }
406
407 static void
408 md_drvinit(void *unused)
409 {
410
411         caddr_t mod;
412         caddr_t c;
413         u_char *ptr, *name, *type;
414         unsigned len;
415
416 #ifdef MD_ROOT_SIZE
417         mdcreate_preload(mfs_root, MD_ROOT_SIZE*1024);
418 #endif
419         mod = NULL;
420         while ((mod = preload_search_next_name(mod)) != NULL) {
421                 name = (char *)preload_search_info(mod, MODINFO_NAME);
422                 type = (char *)preload_search_info(mod, MODINFO_TYPE);
423                 if (name == NULL)
424                         continue;
425                 if (type == NULL)
426                         continue;
427                 if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
428                         continue;
429                 c = preload_search_info(mod, MODINFO_ADDR);
430                 ptr = *(u_char **)c;
431                 c = preload_search_info(mod, MODINFO_SIZE);
432                 len = *(unsigned *)c;
433                 kprintf("md%d: Preloaded image <%s> %d bytes at %p\n",
434                    mdunits, name, len, ptr);
435                 mdcreate_preload(ptr, len);
436         } 
437         mdcreate_malloc();
438 }
439
440 SYSINIT(mddev,SI_SUB_DRIVERS,SI_ORDER_MIDDLE+CDEV_MAJOR, md_drvinit,NULL)
441
442 #ifdef MD_ROOT
443 static void
444 md_takeroot(void *junk)
445 {
446         if (mdrootready)
447                 rootdevnames[0] = "ufs:/dev/md0c";
448 }
449
450 SYSINIT(md_root, SI_SUB_MOUNT_ROOT, SI_ORDER_FIRST, md_takeroot, NULL);
451 #endif