d18d26431a53d41518afd64e9478dfc4a248dcf7
[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,
140                     (long long)bio->bio_offset,
141                     bp->b_bcount, bp->b_data);
142         }
143         bio->bio_driver_info = dev;
144         sc = dev->si_drv1;
145         if (sc->type == MD_MALLOC) {
146                 mdstrategy_malloc(ap);
147         } else {
148                 mdstrategy_preload(ap);
149         }
150         return(0);
151 }
152
153
154 static int
155 mdstrategy_malloc(struct dev_strategy_args *ap)
156 {
157         cdev_t dev = ap->a_head.a_dev;
158         struct bio *bio = ap->a_bio;
159         struct buf *bp = bio->bio_buf;
160         unsigned secno, nsec, secval, uc;
161         u_char *secp, **secpp, *dst;
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,
168                     (long long)bio->bio_offset,
169                     bp->b_bcount, bp->b_data);
170
171         sc = dev->si_drv1;
172
173         crit_enter();
174
175         bioqdisksort(&sc->bio_queue, bio);
176
177         if (sc->busy) {
178                 crit_exit();
179                 return(0);
180         }
181
182         sc->busy++;
183         
184         while (1) {
185                 bio = bioq_first(&sc->bio_queue);
186                 if (bio == NULL) {
187                         crit_exit();
188                         break;
189                 }
190                 crit_exit();
191                 bioq_remove(&sc->bio_queue, bio);
192                 bp = bio->bio_buf;
193
194                 devstat_start_transaction(&sc->stats);
195
196                 switch (bp->b_cmd) {
197                 case BUF_CMD_FREEBLKS:
198                 case BUF_CMD_READ:
199                 case BUF_CMD_WRITE:
200                         break;
201                 default:
202                         panic("md: bad b_cmd %d", bp->b_cmd);
203                 }
204
205                 nsec = bp->b_bcount >> DEV_BSHIFT;
206                 secno = (unsigned)(bio->bio_offset >> DEV_BSHIFT);
207                 dst = bp->b_data;
208                 while (nsec--) {
209                         if (secno < sc->nsecp) {
210                                 secpp = &sc->secp[secno];
211                                 if ((u_int)(uintptr_t)*secpp > 255) {
212                                         secp = *secpp;
213                                         secval = 0;
214                                 } else {
215                                         secp = 0;
216                                         secval = (u_int)(uintptr_t)*secpp;
217                                 }
218                         } else {
219                                 secpp = 0;
220                                 secp = 0;
221                                 secval = 0;
222                         }
223                         if (md_debug > 2)
224                                 kprintf("%08x %p %p %d\n", bp->b_flags, secpp, secp, secval);
225
226                         switch (bp->b_cmd) {
227                         case BUF_CMD_FREEBLKS:
228                                 if (secpp) {
229                                         if (secp)
230                                                 FREE(secp, M_MDSECT);
231                                         *secpp = 0;
232                                 }
233                                 break;
234                         case BUF_CMD_READ:
235                                 if (secp) {
236                                         bcopy(secp, dst, DEV_BSIZE);
237                                 } else if (secval) {
238                                         for (i = 0; i < DEV_BSIZE; i++)
239                                                 dst[i] = secval;
240                                 } else {
241                                         bzero(dst, DEV_BSIZE);
242                                 }
243                                 break;
244                         case BUF_CMD_WRITE:
245                                 uc = dst[0];
246                                 for (i = 1; i < DEV_BSIZE; i++) 
247                                         if (dst[i] != uc)
248                                                 break;
249                                 if (i == DEV_BSIZE && !uc) {
250                                         if (secp)
251                                                 FREE(secp, M_MDSECT);
252                                         if (secpp)
253                                                 *secpp = (u_char *)(uintptr_t)uc;
254                                 } else {
255                                         if (!secpp) {
256                                                 MALLOC(secpp, u_char **, (secno + nsec + 1) * sizeof(u_char *), M_MD, M_WAITOK | M_ZERO);
257                                                 bcopy(sc->secp, secpp, sc->nsecp * sizeof(u_char *));
258                                                 FREE(sc->secp, M_MD);
259                                                 sc->secp = secpp;
260                                                 sc->nsecp = secno + nsec + 1;
261                                                 secpp = &sc->secp[secno];
262                                         }
263                                         if (i == DEV_BSIZE) {
264                                                 if (secp)
265                                                         FREE(secp, M_MDSECT);
266                                                 *secpp = (u_char *)(uintptr_t)uc;
267                                         } else {
268                                                 if (!secp) 
269                                                         MALLOC(secp, u_char *, DEV_BSIZE, M_MDSECT, M_WAITOK);
270                                                 bcopy(dst, secp, DEV_BSIZE);
271
272                                                 *secpp = secp;
273                                         }
274                                 }
275                                 break;
276                         default:
277                                 panic("md: bad b_cmd %d", bp->b_cmd);
278
279                         }
280                         secno++;
281                         dst += DEV_BSIZE;
282                 }
283                 bp->b_resid = 0;
284                 devstat_end_transaction_buf(&sc->stats, bp);
285                 biodone(bio);
286                 crit_enter();
287         }
288         sc->busy = 0;
289         return(0);
290 }
291
292
293 static int
294 mdstrategy_preload(struct dev_strategy_args *ap)
295 {
296         cdev_t dev = ap->a_head.a_dev;
297         struct bio *bio = ap->a_bio;
298         struct buf *bp = bio->bio_buf;
299         struct md_s *sc;
300
301         if (md_debug > 1)
302                 kprintf("mdstrategy_preload(%p) %s %08x, %lld, %d, %p)\n",
303                     bp, devtoname(dev), bp->b_flags,
304                     (long long)bio->bio_offset,
305                     bp->b_bcount, bp->b_data);
306
307         sc = dev->si_drv1;
308
309         crit_enter();
310
311         bioqdisksort(&sc->bio_queue, bio);
312
313         if (sc->busy) {
314                 crit_exit();
315                 return(0);
316         }
317
318         sc->busy++;
319         
320         while (1) {
321                 bio = bioq_first(&sc->bio_queue);
322                 if (bio)
323                         bioq_remove(&sc->bio_queue, bio);
324                 crit_exit();
325                 if (bio == NULL)
326                         break;
327
328                 devstat_start_transaction(&sc->stats);
329
330                 switch (bp->b_cmd) {
331                 case BUF_CMD_FREEBLKS:
332                         break;
333                 case BUF_CMD_READ:
334                         bcopy(sc->pl_ptr + bio->bio_offset, 
335                                bp->b_data, bp->b_bcount);
336                         break;
337                 case BUF_CMD_WRITE:
338                         bcopy(bp->b_data, sc->pl_ptr + bio->bio_offset,
339                               bp->b_bcount);
340                         break;
341                 default:
342                         panic("md: bad cmd %d\n", bp->b_cmd);
343                 }
344                 bp->b_resid = 0;
345                 devstat_end_transaction_buf(&sc->stats, bp);
346                 biodone(bio);
347                 crit_enter();
348         }
349         sc->busy = 0;
350         return(0);
351 }
352
353 static struct md_s *
354 mdcreate(void)
355 {
356         struct md_s *sc;
357         struct disk_info info;
358
359         MALLOC(sc, struct md_s *,sizeof(*sc), M_MD, M_WAITOK | M_ZERO);
360         sc->unit = mdunits++;
361         bioq_init(&sc->bio_queue);
362         devstat_add_entry(&sc->stats, "md", sc->unit, DEV_BSIZE,
363                 DEVSTAT_NO_ORDERED_TAGS, 
364                 DEVSTAT_TYPE_DIRECT | DEVSTAT_TYPE_IF_OTHER,
365                 DEVSTAT_PRIORITY_OTHER);
366         sc->dev = disk_create(sc->unit, &sc->disk, &md_ops);
367         sc->dev->si_drv1 = sc;
368         sc->dev->si_iosize_max = DFLTPHYS;
369
370         bzero(&info, sizeof(info));
371         info.d_media_blksize = DEV_BSIZE;       /* mandatory */
372         info.d_media_blocks = sc->nsect;
373
374         info.d_secpertrack = 1024;              /* optional */
375         info.d_nheads = 1;
376         info.d_secpercyl = info.d_secpertrack * info.d_nheads;
377         info.d_ncylinders = (u_int)(info.d_media_blocks / info.d_secpercyl);
378         disk_setdiskinfo(&sc->disk, &info);
379
380         return (sc);
381 }
382
383 static void
384 mdcreate_preload(u_char *image, unsigned length)
385 {
386         struct md_s *sc;
387
388         sc = mdcreate();
389         sc->type = MD_PRELOAD;
390         sc->nsect = length / DEV_BSIZE;
391         sc->pl_ptr = image;
392         sc->pl_len = length;
393
394         if (sc->unit == 0) 
395                 mdrootready = 1;
396 }
397
398 static void
399 mdcreate_malloc(void)
400 {
401         struct md_s *sc;
402
403         sc = mdcreate();
404         sc->type = MD_MALLOC;
405
406         sc->nsect = MD_NSECT;   /* for now */
407         MALLOC(sc->secp, u_char **, sizeof(u_char *), M_MD, M_WAITOK | M_ZERO);
408         sc->nsecp = 1;
409         kprintf("md%d: Malloc disk\n", sc->unit);
410 }
411
412 static void
413 md_drvinit(void *unused)
414 {
415
416         caddr_t mod;
417         caddr_t c;
418         u_char *ptr, *name, *type;
419         unsigned len;
420
421 #ifdef MD_ROOT_SIZE
422         mdcreate_preload(mfs_root, MD_ROOT_SIZE*1024);
423 #endif
424         mod = NULL;
425         while ((mod = preload_search_next_name(mod)) != NULL) {
426                 name = (char *)preload_search_info(mod, MODINFO_NAME);
427                 type = (char *)preload_search_info(mod, MODINFO_TYPE);
428                 if (name == NULL)
429                         continue;
430                 if (type == NULL)
431                         continue;
432                 if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
433                         continue;
434                 c = preload_search_info(mod, MODINFO_ADDR);
435                 ptr = *(u_char **)c;
436                 c = preload_search_info(mod, MODINFO_SIZE);
437                 len = *(unsigned *)c;
438                 kprintf("md%d: Preloaded image <%s> %d bytes at %p\n",
439                    mdunits, name, len, ptr);
440                 mdcreate_preload(ptr, len);
441         } 
442         mdcreate_malloc();
443 }
444
445 SYSINIT(mddev,SI_SUB_DRIVERS,SI_ORDER_MIDDLE+CDEV_MAJOR, md_drvinit,NULL)
446
447 #ifdef MD_ROOT
448 static void
449 md_takeroot(void *junk)
450 {
451         if (mdrootready)
452                 rootdevnames[0] = "ufs:/dev/md0c";
453 }
454
455 SYSINIT(md_root, SI_SUB_MOUNT_ROOT, SI_ORDER_FIRST, md_takeroot, NULL);
456 #endif