Merge from vendor branch GROFF:
[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 1.8 2005/06/06 22:51:54 corecode Exp $
11  *
12  */
13
14 #include "opt_mfs.h"            /* We have adopted some tasks from MFS */
15 #include "opt_md.h"             /* We have adopted some tasks from MFS */
16
17 #include <sys/param.h>
18 #include <sys/systm.h>
19 #include <sys/buf.h>
20 #include <sys/conf.h>
21 #include <sys/devicestat.h>
22 #include <sys/disk.h>
23 #include <sys/kernel.h>
24 #include <sys/malloc.h>
25 #include <sys/sysctl.h>
26 #include <sys/linker.h>
27 #include <sys/proc.h>
28 #include <sys/buf2.h>
29 #include <sys/thread2.h>
30
31 #ifndef MD_NSECT
32 #define MD_NSECT (10000 * 2)
33 #endif
34
35 MALLOC_DEFINE(M_MD, "MD disk", "Memory Disk");
36 MALLOC_DEFINE(M_MDSECT, "MD sectors", "Memory Disk Sectors");
37
38 static int md_debug;
39 SYSCTL_INT(_debug, OID_AUTO, mddebug, CTLFLAG_RW, &md_debug, 0, "");
40
41 #if defined(MFS_ROOT) && !defined(MD_ROOT)
42 #define MD_ROOT MFS_ROOT
43 #warning "option MFS_ROOT has been superceeded by MD_ROOT"
44 #endif
45
46 #if defined(MFS_ROOT_SIZE) && !defined(MD_ROOT_SIZE)
47 #define MD_ROOT_SIZE MFS_ROOT_SIZE
48 #warning "option MFS_ROOT_SIZE has been superceeded by MD_ROOT_SIZE"
49 #endif
50
51 #if defined(MD_ROOT) && defined(MD_ROOT_SIZE)
52 /* Image gets put here: */
53 static u_char mfs_root[MD_ROOT_SIZE*1024] = "MFS Filesystem goes here";
54 static u_char end_mfs_root[] __unused = "MFS Filesystem had better STOP here";
55 #endif
56
57 static int mdrootready;
58
59 static void mdcreate_malloc(void);
60
61 #define CDEV_MAJOR      95
62
63 static d_strategy_t mdstrategy;
64 static d_strategy_t mdstrategy_preload;
65 static d_strategy_t mdstrategy_malloc;
66 static d_open_t mdopen;
67 static d_ioctl_t mdioctl;
68
69 static struct cdevsw md_cdevsw = {
70         /* name */      "md",
71         /* maj */       CDEV_MAJOR,
72         /* flags */     D_DISK | D_CANFREE | D_MEMDISK,
73         /* port */      NULL,
74         /* clone */     NULL,
75
76         /* open */      mdopen,
77         /* close */     nullclose,
78         /* read */      physread,
79         /* write */     physwrite,
80         /* ioctl */     mdioctl,
81         /* poll */      nopoll,
82         /* mmap */      nommap,
83         /* strategy */  mdstrategy,
84         /* dump */      nodump,
85         /* psize */     nopsize,
86 };
87
88 struct md_s {
89         int unit;
90         struct devstat stats;
91         struct buf_queue_head buf_queue;
92         struct disk disk;
93         dev_t dev;
94         int busy;
95         enum {MD_MALLOC, MD_PRELOAD} type;
96         unsigned nsect;
97
98         /* MD_MALLOC related fields */
99         unsigned nsecp;
100         u_char **secp;
101
102         /* MD_PRELOAD related fields */
103         u_char *pl_ptr;
104         unsigned pl_len;
105 };
106
107 static int mdunits;
108
109 static int
110 mdopen(dev_t dev, int flag, int fmt, struct thread *td)
111 {
112         struct md_s *sc;
113         struct disklabel *dl;
114
115         if (md_debug)
116                 printf("mdopen(%s %x %x %p)\n",
117                         devtoname(dev), flag, fmt, td);
118
119         sc = dev->si_drv1;
120         if (sc->unit + 1 == mdunits)
121                 mdcreate_malloc();
122
123         dl = &sc->disk.d_label;
124         bzero(dl, sizeof(*dl));
125         dl->d_secsize = DEV_BSIZE;
126         dl->d_nsectors = 1024;
127         dl->d_ntracks = 1;
128         dl->d_secpercyl = dl->d_nsectors * dl->d_ntracks;
129         dl->d_secperunit = sc->nsect;
130         dl->d_ncylinders = dl->d_secperunit / dl->d_secpercyl;
131         return (0);
132 }
133
134 static int
135 mdioctl(dev_t dev, u_long cmd, caddr_t addr, int flags, struct thread *td)
136 {
137
138         if (md_debug)
139                 printf("mdioctl(%s %lx %p %x %p)\n",
140                         devtoname(dev), cmd, addr, flags, td);
141
142         return (ENOIOCTL);
143 }
144
145 static void
146 mdstrategy(struct buf *bp)
147 {
148         struct md_s *sc;
149
150         if (md_debug > 1)
151                 printf("mdstrategy(%p) %s %lx, %d, %ld, %p)\n",
152                     bp, devtoname(bp->b_dev), bp->b_flags, bp->b_blkno, 
153                     bp->b_bcount / DEV_BSIZE, bp->b_data);
154
155         sc = bp->b_dev->si_drv1;
156         if (sc->type == MD_MALLOC) {
157                 mdstrategy_malloc(bp);
158         } else {
159                 mdstrategy_preload(bp);
160         }
161         return;
162 }
163
164
165 static void
166 mdstrategy_malloc(struct buf *bp)
167 {
168         int i;
169         struct md_s *sc;
170         devstat_trans_flags dop;
171         u_char *secp, **secpp, *dst;
172         unsigned secno, nsec, secval, uc;
173
174         if (md_debug > 1)
175                 printf("mdstrategy_malloc(%p) %s %lx, %d, %ld, %p)\n",
176                     bp, devtoname(bp->b_dev), bp->b_flags, bp->b_blkno, 
177                     bp->b_bcount / DEV_BSIZE, bp->b_data);
178
179         sc = bp->b_dev->si_drv1;
180
181         crit_enter();
182
183         bufqdisksort(&sc->buf_queue, bp);
184
185         if (sc->busy) {
186                 crit_exit();
187                 return;
188         }
189
190         sc->busy++;
191         
192         while (1) {
193                 bp = bufq_first(&sc->buf_queue);
194                 if (bp)
195                         bufq_remove(&sc->buf_queue, bp);
196                 crit_exit();
197                 if (!bp)
198                         break;
199
200                 devstat_start_transaction(&sc->stats);
201
202                 if (bp->b_flags & B_FREEBUF) 
203                         dop = DEVSTAT_NO_DATA;
204                 else if (bp->b_flags & B_READ)
205                         dop = DEVSTAT_READ;
206                 else
207                         dop = DEVSTAT_WRITE;
208
209                 nsec = bp->b_bcount / DEV_BSIZE;
210                 secno = bp->b_pblkno;
211                 dst = bp->b_data;
212                 while (nsec--) {
213
214                         if (secno < sc->nsecp) {
215                                 secpp = &sc->secp[secno];
216                                 if ((u_int)*secpp > 255) {
217                                         secp = *secpp;
218                                         secval = 0;
219                                 } else {
220                                         secp = 0;
221                                         secval = (u_int) *secpp;
222                                 }
223                         } else {
224                                 secpp = 0;
225                                 secp = 0;
226                                 secval = 0;
227                         }
228                         if (md_debug > 2)
229                                 printf("%lx %p %p %d\n", bp->b_flags, secpp, secp, secval);
230
231                         if (bp->b_flags & B_FREEBUF) {
232                                 if (secpp) {
233                                         if (secp)
234                                                 FREE(secp, M_MDSECT);
235                                         *secpp = 0;
236                                 }
237                         } else if (bp->b_flags & B_READ) {
238                                 if (secp) {
239                                         bcopy(secp, dst, DEV_BSIZE);
240                                 } else if (secval) {
241                                         for (i = 0; i < DEV_BSIZE; i++)
242                                                 dst[i] = secval;
243                                 } else {
244                                         bzero(dst, DEV_BSIZE);
245                                 }
246                         } else {
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                         }
279                         secno++;
280                         dst += DEV_BSIZE;
281                 }
282                 bp->b_resid = 0;
283                 devstat_end_transaction_buf(&sc->stats, bp);
284                 biodone(bp);
285                 crit_enter();
286         }
287         sc->busy = 0;
288         return;
289 }
290
291
292 static void
293 mdstrategy_preload(struct buf *bp)
294 {
295         struct md_s *sc;
296         devstat_trans_flags dop;
297
298         if (md_debug > 1)
299                 printf("mdstrategy_preload(%p) %s %lx, %d, %ld, %p)\n",
300                     bp, devtoname(bp->b_dev), bp->b_flags, bp->b_blkno, 
301                     bp->b_bcount / DEV_BSIZE, bp->b_data);
302
303         sc = bp->b_dev->si_drv1;
304
305         crit_enter();
306
307         bufqdisksort(&sc->buf_queue, bp);
308
309         if (sc->busy) {
310                 crit_exit();
311                 return;
312         }
313
314         sc->busy++;
315         
316         while (1) {
317                 bp = bufq_first(&sc->buf_queue);
318                 if (bp)
319                         bufq_remove(&sc->buf_queue, bp);
320                 crit_exit();
321                 if (!bp)
322                         break;
323
324                 devstat_start_transaction(&sc->stats);
325
326                 if (bp->b_flags & B_FREEBUF) {
327                         dop = DEVSTAT_NO_DATA;
328                 } else if (bp->b_flags & B_READ) {
329                         dop = DEVSTAT_READ;
330                         bcopy(sc->pl_ptr + (bp->b_pblkno << DEV_BSHIFT), bp->b_data, bp->b_bcount);
331                 } else {
332                         dop = DEVSTAT_WRITE;
333                         bcopy(bp->b_data, sc->pl_ptr + (bp->b_pblkno << DEV_BSHIFT), bp->b_bcount);
334                 }
335                 bp->b_resid = 0;
336                 devstat_end_transaction_buf(&sc->stats, bp);
337                 biodone(bp);
338                 crit_enter();
339         }
340         sc->busy = 0;
341         return;
342 }
343
344 static struct md_s *
345 mdcreate(void)
346 {
347         struct md_s *sc;
348
349         MALLOC(sc, struct md_s *,sizeof(*sc), M_MD, M_WAITOK);
350         bzero(sc, sizeof(*sc));
351         sc->unit = mdunits++;
352         bufq_init(&sc->buf_queue);
353         devstat_add_entry(&sc->stats, "md", sc->unit, DEV_BSIZE,
354                 DEVSTAT_NO_ORDERED_TAGS, 
355                 DEVSTAT_TYPE_DIRECT | DEVSTAT_TYPE_IF_OTHER,
356                 DEVSTAT_PRIORITY_OTHER);
357         sc->dev = disk_create(sc->unit, &sc->disk, 0, &md_cdevsw);
358         sc->dev->si_drv1 = sc;
359         return (sc);
360 }
361
362 static void
363 mdcreate_preload(u_char *image, unsigned length)
364 {
365         struct md_s *sc;
366
367         sc = mdcreate();
368         sc->type = MD_PRELOAD;
369         sc->nsect = length / DEV_BSIZE;
370         sc->pl_ptr = image;
371         sc->pl_len = length;
372
373         if (sc->unit == 0) 
374                 mdrootready = 1;
375 }
376
377 static void
378 mdcreate_malloc(void)
379 {
380         struct md_s *sc;
381
382         sc = mdcreate();
383         sc->type = MD_MALLOC;
384
385         sc->nsect = MD_NSECT;   /* for now */
386         MALLOC(sc->secp, u_char **, sizeof(u_char *), M_MD, M_WAITOK);
387         bzero(sc->secp, sizeof(u_char *));
388         sc->nsecp = 1;
389         printf("md%d: Malloc disk\n", sc->unit);
390 }
391
392 static void
393 md_drvinit(void *unused)
394 {
395
396         caddr_t mod;
397         caddr_t c;
398         u_char *ptr, *name, *type;
399         unsigned len;
400
401 #ifdef MD_ROOT_SIZE
402         mdcreate_preload(mfs_root, MD_ROOT_SIZE*1024);
403 #endif
404         mod = NULL;
405         while ((mod = preload_search_next_name(mod)) != NULL) {
406                 name = (char *)preload_search_info(mod, MODINFO_NAME);
407                 type = (char *)preload_search_info(mod, MODINFO_TYPE);
408                 if (name == NULL)
409                         continue;
410                 if (type == NULL)
411                         continue;
412                 if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
413                         continue;
414                 c = preload_search_info(mod, MODINFO_ADDR);
415                 ptr = *(u_char **)c;
416                 c = preload_search_info(mod, MODINFO_SIZE);
417                 len = *(unsigned *)c;
418                 printf("md%d: Preloaded image <%s> %d bytes at %p\n",
419                    mdunits, name, len, ptr);
420                 mdcreate_preload(ptr, len);
421         } 
422         mdcreate_malloc();
423 }
424
425 SYSINIT(mddev,SI_SUB_DRIVERS,SI_ORDER_MIDDLE+CDEV_MAJOR, md_drvinit,NULL)
426
427 #ifdef MD_ROOT
428 static void
429 md_takeroot(void *junk)
430 {
431         if (mdrootready)
432                 rootdevnames[0] = "ufs:/dev/md0c";
433 }
434
435 SYSINIT(md_root, SI_SUB_MOUNT_ROOT, SI_ORDER_FIRST, md_takeroot, NULL);
436 #endif