Replace the the buffer cache's B_READ, B_WRITE, B_FORMAT, and B_FREEBUF
[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 1.11 2006/04/30 17:22:16 dillon 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 bio_queue_head bio_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(dev_t dev, struct bio *bio)
147 {
148         struct buf *bp = bio->bio_buf;
149         struct md_s *sc;
150
151         if (md_debug > 1) {
152                 printf("mdstrategy(%p) %s %08x, %lld, %d, %p)\n",
153                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
154                     bp->b_bcount, bp->b_data);
155         }
156         bio->bio_driver_info = dev;
157         sc = dev->si_drv1;
158         if (sc->type == MD_MALLOC) {
159                 mdstrategy_malloc(dev, bio);
160         } else {
161                 mdstrategy_preload(dev, bio);
162         }
163 }
164
165
166 static void
167 mdstrategy_malloc(dev_t dev, struct bio *bio)
168 {
169         struct buf *bp = bio->bio_buf;
170         unsigned secno, nsec, secval, uc;
171         u_char *secp, **secpp, *dst;
172         devstat_trans_flags dop;
173         struct md_s *sc;
174         int i;
175
176         if (md_debug > 1)
177                 printf("mdstrategy_malloc(%p) %s %08xx, %lld, %d, %p)\n",
178                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
179                     bp->b_bcount, bp->b_data);
180
181         sc = dev->si_drv1;
182
183         crit_enter();
184
185         bioqdisksort(&sc->bio_queue, bio);
186
187         if (sc->busy) {
188                 crit_exit();
189                 return;
190         }
191
192         sc->busy++;
193         
194         while (1) {
195                 bio = bioq_first(&sc->bio_queue);
196                 if (bp)
197                         bioq_remove(&sc->bio_queue, bio);
198                 crit_exit();
199                 if (bio == NULL)
200                         break;
201
202                 devstat_start_transaction(&sc->stats);
203
204                 switch(bp->b_cmd) {
205                 case BUF_CMD_FREEBLKS:
206                         dop = DEVSTAT_NO_DATA;
207                         break;
208                 case BUF_CMD_READ:
209                         dop = DEVSTAT_READ;
210                         break;
211                 case BUF_CMD_WRITE:
212                         dop = DEVSTAT_WRITE;
213                         break;
214                 default:
215                         panic("md: bad b_cmd %d", bp->b_cmd);
216                 }
217
218                 nsec = bp->b_bcount >> DEV_BSHIFT;
219                 secno = (unsigned)(bio->bio_offset >> DEV_BSHIFT);
220                 dst = bp->b_data;
221                 while (nsec--) {
222                         if (secno < sc->nsecp) {
223                                 secpp = &sc->secp[secno];
224                                 if ((u_int)*secpp > 255) {
225                                         secp = *secpp;
226                                         secval = 0;
227                                 } else {
228                                         secp = 0;
229                                         secval = (u_int) *secpp;
230                                 }
231                         } else {
232                                 secpp = 0;
233                                 secp = 0;
234                                 secval = 0;
235                         }
236                         if (md_debug > 2)
237                                 printf("%08x %p %p %d\n", bp->b_flags, secpp, secp, secval);
238
239                         switch(bp->b_cmd) {
240                         case BUF_CMD_FREEBLKS:
241                                 if (secpp) {
242                                         if (secp)
243                                                 FREE(secp, M_MDSECT);
244                                         *secpp = 0;
245                                 }
246                                 break;
247                         case BUF_CMD_READ:
248                                 if (secp) {
249                                         bcopy(secp, dst, DEV_BSIZE);
250                                 } else if (secval) {
251                                         for (i = 0; i < DEV_BSIZE; i++)
252                                                 dst[i] = secval;
253                                 } else {
254                                         bzero(dst, DEV_BSIZE);
255                                 }
256                                 break;
257                         case BUF_CMD_WRITE:
258                                 uc = dst[0];
259                                 for (i = 1; i < DEV_BSIZE; i++) 
260                                         if (dst[i] != uc)
261                                                 break;
262                                 if (i == DEV_BSIZE && !uc) {
263                                         if (secp)
264                                                 FREE(secp, M_MDSECT);
265                                         if (secpp)
266                                                 *secpp = (u_char *)uc;
267                                 } else {
268                                         if (!secpp) {
269                                                 MALLOC(secpp, u_char **, (secno + nsec + 1) * sizeof(u_char *), M_MD, M_WAITOK);
270                                                 bzero(secpp, (secno + nsec + 1) * sizeof(u_char *));
271                                                 bcopy(sc->secp, secpp, sc->nsecp * sizeof(u_char *));
272                                                 FREE(sc->secp, M_MD);
273                                                 sc->secp = secpp;
274                                                 sc->nsecp = secno + nsec + 1;
275                                                 secpp = &sc->secp[secno];
276                                         }
277                                         if (i == DEV_BSIZE) {
278                                                 if (secp)
279                                                         FREE(secp, M_MDSECT);
280                                                 *secpp = (u_char *)uc;
281                                         } else {
282                                                 if (!secp) 
283                                                         MALLOC(secp, u_char *, DEV_BSIZE, M_MDSECT, M_WAITOK);
284                                                 bcopy(dst, secp, DEV_BSIZE);
285
286                                                 *secpp = secp;
287                                         }
288                                 }
289                                 break;
290                         default:
291                                 panic("md: bad b_cmd %d", bp->b_cmd);
292
293                         }
294                         secno++;
295                         dst += DEV_BSIZE;
296                 }
297                 bp->b_resid = 0;
298                 devstat_end_transaction_buf(&sc->stats, bp);
299                 biodone(bio);
300                 crit_enter();
301         }
302         sc->busy = 0;
303 }
304
305
306 static void
307 mdstrategy_preload(dev_t dev, struct bio *bio)
308 {
309         struct buf *bp = bio->bio_buf;
310         devstat_trans_flags dop;
311         struct md_s *sc;
312
313         if (md_debug > 1)
314                 printf("mdstrategy_preload(%p) %s %08x, %lld, %d, %p)\n",
315                     bp, devtoname(dev), bp->b_flags, bio->bio_offset, 
316                     bp->b_bcount, bp->b_data);
317
318         sc = dev->si_drv1;
319
320         crit_enter();
321
322         bioqdisksort(&sc->bio_queue, bio);
323
324         if (sc->busy) {
325                 crit_exit();
326                 return;
327         }
328
329         sc->busy++;
330         
331         while (1) {
332                 bio = bioq_first(&sc->bio_queue);
333                 if (bio)
334                         bioq_remove(&sc->bio_queue, bio);
335                 crit_exit();
336                 if (bio == NULL)
337                         break;
338
339                 devstat_start_transaction(&sc->stats);
340
341                 switch(bp->b_cmd) {
342                 case BUF_CMD_FREEBLKS:
343                         dop = DEVSTAT_NO_DATA;
344                         break;
345                 case BUF_CMD_READ:
346                         dop = DEVSTAT_READ;
347                         bcopy(sc->pl_ptr + bio->bio_offset, 
348                                bp->b_data, bp->b_bcount);
349                         break;
350                 case BUF_CMD_WRITE:
351                         dop = DEVSTAT_WRITE;
352                         bcopy(bp->b_data, sc->pl_ptr + bio->bio_offset,
353                               bp->b_bcount);
354                         break;
355                 default:
356                         panic("md: bad cmd %d\n", bp->b_cmd);
357                 }
358                 bp->b_resid = 0;
359                 devstat_end_transaction_buf(&sc->stats, bp);
360                 biodone(bio);
361                 crit_enter();
362         }
363         sc->busy = 0;
364 }
365
366 static struct md_s *
367 mdcreate(void)
368 {
369         struct md_s *sc;
370
371         MALLOC(sc, struct md_s *,sizeof(*sc), M_MD, M_WAITOK);
372         bzero(sc, sizeof(*sc));
373         sc->unit = mdunits++;
374         bioq_init(&sc->bio_queue);
375         devstat_add_entry(&sc->stats, "md", sc->unit, DEV_BSIZE,
376                 DEVSTAT_NO_ORDERED_TAGS, 
377                 DEVSTAT_TYPE_DIRECT | DEVSTAT_TYPE_IF_OTHER,
378                 DEVSTAT_PRIORITY_OTHER);
379         sc->dev = disk_create(sc->unit, &sc->disk, 0, &md_cdevsw);
380         sc->dev->si_drv1 = sc;
381         return (sc);
382 }
383
384 static void
385 mdcreate_preload(u_char *image, unsigned length)
386 {
387         struct md_s *sc;
388
389         sc = mdcreate();
390         sc->type = MD_PRELOAD;
391         sc->nsect = length / DEV_BSIZE;
392         sc->pl_ptr = image;
393         sc->pl_len = length;
394
395         if (sc->unit == 0) 
396                 mdrootready = 1;
397 }
398
399 static void
400 mdcreate_malloc(void)
401 {
402         struct md_s *sc;
403
404         sc = mdcreate();
405         sc->type = MD_MALLOC;
406
407         sc->nsect = MD_NSECT;   /* for now */
408         MALLOC(sc->secp, u_char **, sizeof(u_char *), M_MD, M_WAITOK);
409         bzero(sc->secp, sizeof(u_char *));
410         sc->nsecp = 1;
411         printf("md%d: Malloc disk\n", sc->unit);
412 }
413
414 static void
415 md_drvinit(void *unused)
416 {
417
418         caddr_t mod;
419         caddr_t c;
420         u_char *ptr, *name, *type;
421         unsigned len;
422
423 #ifdef MD_ROOT_SIZE
424         mdcreate_preload(mfs_root, MD_ROOT_SIZE*1024);
425 #endif
426         mod = NULL;
427         while ((mod = preload_search_next_name(mod)) != NULL) {
428                 name = (char *)preload_search_info(mod, MODINFO_NAME);
429                 type = (char *)preload_search_info(mod, MODINFO_TYPE);
430                 if (name == NULL)
431                         continue;
432                 if (type == NULL)
433                         continue;
434                 if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
435                         continue;
436                 c = preload_search_info(mod, MODINFO_ADDR);
437                 ptr = *(u_char **)c;
438                 c = preload_search_info(mod, MODINFO_SIZE);
439                 len = *(unsigned *)c;
440                 printf("md%d: Preloaded image <%s> %d bytes at %p\n",
441                    mdunits, name, len, ptr);
442                 mdcreate_preload(ptr, len);
443         } 
444         mdcreate_malloc();
445 }
446
447 SYSINIT(mddev,SI_SUB_DRIVERS,SI_ORDER_MIDDLE+CDEV_MAJOR, md_drvinit,NULL)
448
449 #ifdef MD_ROOT
450 static void
451 md_takeroot(void *junk)
452 {
453         if (mdrootready)
454                 rootdevnames[0] = "ufs:/dev/md0c";
455 }
456
457 SYSINIT(md_root, SI_SUB_MOUNT_ROOT, SI_ORDER_FIRST, md_takeroot, NULL);
458 #endif