Merge from vendor branch FILE:
[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 1.15 2006/12/22 23:26:16 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 disklabel *dl;
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         dl = &sc->disk.d_label;
105         bzero(dl, sizeof(*dl));
106         dl->d_secsize = DEV_BSIZE;
107         dl->d_nsectors = 1024;
108         dl->d_ntracks = 1;
109         dl->d_secpercyl = dl->d_nsectors * dl->d_ntracks;
110         dl->d_secperunit = sc->nsect;
111         dl->d_ncylinders = dl->d_secperunit / dl->d_secpercyl;
112         return (0);
113 }
114
115 static int
116 mdioctl(struct dev_ioctl_args *ap)
117 {
118         cdev_t dev = ap->a_head.a_dev;
119
120         if (md_debug)
121                 kprintf("mdioctl(%s %lx %p %x)\n",
122                         devtoname(dev), ap->a_cmd, ap->a_data, ap->a_fflag);
123
124         return (ENOIOCTL);
125 }
126
127 static int
128 mdstrategy(struct dev_strategy_args *ap)
129 {
130         cdev_t dev = ap->a_head.a_dev;
131         struct bio *bio = ap->a_bio;
132         struct buf *bp = bio->bio_buf;
133         struct md_s *sc;
134
135         if (md_debug > 1) {
136                 kprintf("mdstrategy(%p) %s %08x, %lld, %d, %p)\n",
137                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
138                     bp->b_bcount, bp->b_data);
139         }
140         bio->bio_driver_info = dev;
141         sc = dev->si_drv1;
142         if (sc->type == MD_MALLOC) {
143                 mdstrategy_malloc(ap);
144         } else {
145                 mdstrategy_preload(ap);
146         }
147         return(0);
148 }
149
150
151 static int
152 mdstrategy_malloc(struct dev_strategy_args *ap)
153 {
154         cdev_t dev = ap->a_head.a_dev;
155         struct bio *bio = ap->a_bio;
156         struct buf *bp = bio->bio_buf;
157         unsigned secno, nsec, secval, uc;
158         u_char *secp, **secpp, *dst;
159         devstat_trans_flags dop;
160         struct md_s *sc;
161         int i;
162
163         if (md_debug > 1)
164                 kprintf("mdstrategy_malloc(%p) %s %08xx, %lld, %d, %p)\n",
165                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
166                     bp->b_bcount, bp->b_data);
167
168         sc = dev->si_drv1;
169
170         crit_enter();
171
172         bioqdisksort(&sc->bio_queue, bio);
173
174         if (sc->busy) {
175                 crit_exit();
176                 return(0);
177         }
178
179         sc->busy++;
180         
181         while (1) {
182                 bio = bioq_first(&sc->bio_queue);
183                 if (bp)
184                         bioq_remove(&sc->bio_queue, bio);
185                 crit_exit();
186                 if (bio == NULL)
187                         break;
188
189                 devstat_start_transaction(&sc->stats);
190
191                 switch(bp->b_cmd) {
192                 case BUF_CMD_FREEBLKS:
193                         dop = DEVSTAT_NO_DATA;
194                         break;
195                 case BUF_CMD_READ:
196                         dop = DEVSTAT_READ;
197                         break;
198                 case BUF_CMD_WRITE:
199                         dop = DEVSTAT_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)*secpp > 255) {
212                                         secp = *secpp;
213                                         secval = 0;
214                                 } else {
215                                         secp = 0;
216                                         secval = (u_int) *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 *)uc;
254                                 } else {
255                                         if (!secpp) {
256                                                 MALLOC(secpp, u_char **, (secno + nsec + 1) * sizeof(u_char *), M_MD, M_WAITOK);
257                                                 bzero(secpp, (secno + nsec + 1) * sizeof(u_char *));
258                                                 bcopy(sc->secp, secpp, sc->nsecp * sizeof(u_char *));
259                                                 FREE(sc->secp, M_MD);
260                                                 sc->secp = secpp;
261                                                 sc->nsecp = secno + nsec + 1;
262                                                 secpp = &sc->secp[secno];
263                                         }
264                                         if (i == DEV_BSIZE) {
265                                                 if (secp)
266                                                         FREE(secp, M_MDSECT);
267                                                 *secpp = (u_char *)uc;
268                                         } else {
269                                                 if (!secp) 
270                                                         MALLOC(secp, u_char *, DEV_BSIZE, M_MDSECT, M_WAITOK);
271                                                 bcopy(dst, secp, DEV_BSIZE);
272
273                                                 *secpp = secp;
274                                         }
275                                 }
276                                 break;
277                         default:
278                                 panic("md: bad b_cmd %d", bp->b_cmd);
279
280                         }
281                         secno++;
282                         dst += DEV_BSIZE;
283                 }
284                 bp->b_resid = 0;
285                 devstat_end_transaction_buf(&sc->stats, bp);
286                 biodone(bio);
287                 crit_enter();
288         }
289         sc->busy = 0;
290         return(0);
291 }
292
293
294 static int
295 mdstrategy_preload(struct dev_strategy_args *ap)
296 {
297         cdev_t dev = ap->a_head.a_dev;
298         struct bio *bio = ap->a_bio;
299         struct buf *bp = bio->bio_buf;
300         devstat_trans_flags dop;
301         struct md_s *sc;
302
303         if (md_debug > 1)
304                 kprintf("mdstrategy_preload(%p) %s %08x, %lld, %d, %p)\n",
305                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
306                     bp->b_bcount, bp->b_data);
307
308         sc = dev->si_drv1;
309
310         crit_enter();
311
312         bioqdisksort(&sc->bio_queue, bio);
313
314         if (sc->busy) {
315                 crit_exit();
316                 return(0);
317         }
318
319         sc->busy++;
320         
321         while (1) {
322                 bio = bioq_first(&sc->bio_queue);
323                 if (bio)
324                         bioq_remove(&sc->bio_queue, bio);
325                 crit_exit();
326                 if (bio == NULL)
327                         break;
328
329                 devstat_start_transaction(&sc->stats);
330
331                 switch(bp->b_cmd) {
332                 case BUF_CMD_FREEBLKS:
333                         dop = DEVSTAT_NO_DATA;
334                         break;
335                 case BUF_CMD_READ:
336                         dop = DEVSTAT_READ;
337                         bcopy(sc->pl_ptr + bio->bio_offset, 
338                                bp->b_data, bp->b_bcount);
339                         break;
340                 case BUF_CMD_WRITE:
341                         dop = DEVSTAT_WRITE;
342                         bcopy(bp->b_data, sc->pl_ptr + bio->bio_offset,
343                               bp->b_bcount);
344                         break;
345                 default:
346                         panic("md: bad cmd %d\n", bp->b_cmd);
347                 }
348                 bp->b_resid = 0;
349                 devstat_end_transaction_buf(&sc->stats, bp);
350                 biodone(bio);
351                 crit_enter();
352         }
353         sc->busy = 0;
354         return(0);
355 }
356
357 static struct md_s *
358 mdcreate(void)
359 {
360         struct md_s *sc;
361
362         MALLOC(sc, struct md_s *,sizeof(*sc), M_MD, M_WAITOK);
363         bzero(sc, sizeof(*sc));
364         sc->unit = mdunits++;
365         bioq_init(&sc->bio_queue);
366         devstat_add_entry(&sc->stats, "md", sc->unit, DEV_BSIZE,
367                 DEVSTAT_NO_ORDERED_TAGS, 
368                 DEVSTAT_TYPE_DIRECT | DEVSTAT_TYPE_IF_OTHER,
369                 DEVSTAT_PRIORITY_OTHER);
370         sc->dev = disk_create(sc->unit, &sc->disk, 0, &md_ops);
371         sc->dev->si_drv1 = sc;
372         return (sc);
373 }
374
375 static void
376 mdcreate_preload(u_char *image, unsigned length)
377 {
378         struct md_s *sc;
379
380         sc = mdcreate();
381         sc->type = MD_PRELOAD;
382         sc->nsect = length / DEV_BSIZE;
383         sc->pl_ptr = image;
384         sc->pl_len = length;
385
386         if (sc->unit == 0) 
387                 mdrootready = 1;
388 }
389
390 static void
391 mdcreate_malloc(void)
392 {
393         struct md_s *sc;
394
395         sc = mdcreate();
396         sc->type = MD_MALLOC;
397
398         sc->nsect = MD_NSECT;   /* for now */
399         MALLOC(sc->secp, u_char **, sizeof(u_char *), M_MD, M_WAITOK);
400         bzero(sc->secp, sizeof(u_char *));
401         sc->nsecp = 1;
402         kprintf("md%d: Malloc disk\n", sc->unit);
403 }
404
405 static void
406 md_drvinit(void *unused)
407 {
408
409         caddr_t mod;
410         caddr_t c;
411         u_char *ptr, *name, *type;
412         unsigned len;
413
414 #ifdef MD_ROOT_SIZE
415         mdcreate_preload(mfs_root, MD_ROOT_SIZE*1024);
416 #endif
417         mod = NULL;
418         while ((mod = preload_search_next_name(mod)) != NULL) {
419                 name = (char *)preload_search_info(mod, MODINFO_NAME);
420                 type = (char *)preload_search_info(mod, MODINFO_TYPE);
421                 if (name == NULL)
422                         continue;
423                 if (type == NULL)
424                         continue;
425                 if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
426                         continue;
427                 c = preload_search_info(mod, MODINFO_ADDR);
428                 ptr = *(u_char **)c;
429                 c = preload_search_info(mod, MODINFO_SIZE);
430                 len = *(unsigned *)c;
431                 kprintf("md%d: Preloaded image <%s> %d bytes at %p\n",
432                    mdunits, name, len, ptr);
433                 mdcreate_preload(ptr, len);
434         } 
435         mdcreate_malloc();
436 }
437
438 SYSINIT(mddev,SI_SUB_DRIVERS,SI_ORDER_MIDDLE+CDEV_MAJOR, md_drvinit,NULL)
439
440 #ifdef MD_ROOT
441 static void
442 md_takeroot(void *junk)
443 {
444         if (mdrootready)
445                 rootdevnames[0] = "ufs:/dev/md0c";
446 }
447
448 SYSINIT(md_root, SI_SUB_MOUNT_ROOT, SI_ORDER_FIRST, md_takeroot, NULL);
449 #endif