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  *
  11  */
  12
  13 #include "opt_md.h"             /* We have adopted some tasks from MFS */
  14
  15 #include <sys/param.h>
  16 #include <sys/systm.h>
  17 #include <sys/buf.h>
  18 #include <sys/conf.h>
  19 #include <sys/devicestat.h>
  20 #include <sys/disk.h>
  21 #include <sys/kernel.h>
  22 #include <sys/malloc.h>
  23 #include <sys/sysctl.h>
  24 #include <sys/linker.h>
  25 #include <sys/proc.h>
  26 #include <sys/buf2.h>
  27 #include <sys/thread2.h>
  28 #include <sys/queue.h>
  29 #include <sys/udev.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     "Enable debug output for memory disk devices");
  41
  42 #if defined(MD_ROOT) && defined(MD_ROOT_SIZE)
  43 /* Image gets put here: */
  44 static u_char mfs_root[MD_ROOT_SIZE*1024] = "MFS Filesystem goes here";
  45 static u_char end_mfs_root[] __unused = "MFS Filesystem had better STOP here";
  46 #endif
  47
  48 static int mdrootready;
  49
  50 static d_strategy_t mdstrategy;
  51 static d_strategy_t mdstrategy_preload;
  52 static d_strategy_t mdstrategy_malloc;
  53 static d_open_t mdopen;
  54 static d_close_t mdclose;
  55 static d_ioctl_t mdioctl;
  56
  57 static struct dev_ops md_ops = {
  58         { "md", 0, D_DISK | D_CANFREE | D_MEMDISK | D_TRACKCLOSE},
  59         .d_open =       mdopen,
  60         .d_close =      mdclose,
  61         .d_read =       physread,
  62         .d_write =      physwrite,
  63         .d_ioctl =      mdioctl,
  64         .d_strategy =   mdstrategy,
  65 };
  66
  67 struct md_s {
  68         int unit;
  69         struct devstat stats;
  70         struct bio_queue_head bio_queue;
  71         struct disk disk;
  72         cdev_t dev;
  73         int busy;
  74         enum {                  /* Memory disk type */
  75                 MD_MALLOC,
  76                 MD_PRELOAD
  77         } type;
  78         unsigned nsect;
  79
  80         /* MD_MALLOC related fields */
  81         unsigned nsecp;
  82         u_char **secp;
  83
  84         /* MD_PRELOAD related fields */
  85         u_char *pl_ptr;
  86         unsigned pl_len;
  87         TAILQ_ENTRY(md_s) link;
  88 };
  89 TAILQ_HEAD(mdshead, md_s) mdlist = TAILQ_HEAD_INITIALIZER(mdlist);
  90
  91 static int mdunits;
  92 static int refcnt;
  93
  94 static struct md_s *mdcreate(unsigned);
  95 static void mdcreate_malloc(void);
  96 static int mdinit(module_t, int, void *);
  97 static void md_drvinit(void *);
  98 static int md_drvcleanup(void);
  99
 100 static int
 101 mdinit(module_t mod, int cmd, void *arg)
 102 {
 103     int ret = 0;
 104
 105     switch(cmd) {
 106         case MOD_LOAD:
 107                 TAILQ_INIT(&mdlist);
 108                 md_drvinit(NULL);
 109                 break;
 110         case MOD_UNLOAD:
 111                 ret = md_drvcleanup();
 112                 break;
 113         default:
 114                 ret = EINVAL;
 115                 break;
 116     }
 117
 118     return (ret);
 119 }
 120
 121 static int
 122 mdopen(struct dev_open_args *ap)
 123 {
 124         cdev_t dev = ap->a_head.a_dev;
 125         struct md_s *sc;
 126
 127         if (md_debug)
 128                 kprintf("mdopen(%s %x %x)\n",
 129                         devtoname(dev), ap->a_oflags, ap->a_devtype);
 130
 131         sc = dev->si_drv1;
 132         if (sc->unit + 1 == mdunits)
 133                 mdcreate_malloc();
 134
 135         atomic_add_int(&refcnt, 1);
 136         return (0);
 137 }
 138
 139 static int
 140 mdclose(struct dev_close_args *ap)
 141 {
 142         cdev_t dev = ap->a_head.a_dev;
 143         struct md_s *sc;
 144
 145         if (md_debug)
 146                 kprintf("mdclose(%s %x %x)\n",
 147                         devtoname(dev), ap->a_fflag, ap->a_devtype);
 148         sc = dev->si_drv1;
 149         atomic_add_int(&refcnt, -1);
 150
 151         return (0);
 152 }
 153
 154 static int
 155 mdioctl(struct dev_ioctl_args *ap)
 156 {
 157         cdev_t dev = ap->a_head.a_dev;
 158
 159         if (md_debug)
 160                 kprintf("mdioctl(%s %lx %p %x)\n",
 161                         devtoname(dev), ap->a_cmd, ap->a_data, ap->a_fflag);
 162
 163         return (ENOIOCTL);
 164 }
 165
 166 static int
 167 mdstrategy(struct dev_strategy_args *ap)
 168 {
 169         cdev_t dev = ap->a_head.a_dev;
 170         struct bio *bio = ap->a_bio;
 171         struct buf *bp = bio->bio_buf;
 172         struct md_s *sc;
 173
 174         if (md_debug > 1) {
 175                 kprintf("mdstrategy(%p) %s %08x, %lld, %d, %p)\n",
 176                     bp, devtoname(dev), bp->b_flags,
 177                     (long long)bio->bio_offset,
 178                     bp->b_bcount, bp->b_data);
 179         }
 180         bio->bio_driver_info = dev;
 181         sc = dev->si_drv1;
 182         if (sc->type == MD_MALLOC) {
 183                 mdstrategy_malloc(ap);
 184         } else {
 185                 mdstrategy_preload(ap);
 186         }
 187         return(0);
 188 }
 189
 190
 191 static int
 192 mdstrategy_malloc(struct dev_strategy_args *ap)
 193 {
 194         cdev_t dev = ap->a_head.a_dev;
 195         struct bio *bio = ap->a_bio;
 196         struct buf *bp = bio->bio_buf;
 197         unsigned secno, nsec, secval, uc;
 198         u_char *secp, **secpp, *dst;
 199         struct md_s *sc;
 200         int i;
 201
 202         if (md_debug > 1)
 203                 kprintf("mdstrategy_malloc(%p) %s %08xx, %lld, %d, %p)\n",
 204                     bp, devtoname(dev), bp->b_flags,
 205                     (long long)bio->bio_offset,
 206                     bp->b_bcount, bp->b_data);
 207
 208         sc = dev->si_drv1;
 209
 210         crit_enter();
 211
 212         bioqdisksort(&sc->bio_queue, bio);
 213
 214         if (sc->busy) {
 215                 crit_exit();
 216                 return(0);
 217         }
 218
 219         sc->busy++;
 220
 221         while (1) {
 222                 bio = bioq_first(&sc->bio_queue);
 223                 if (bio == NULL) {
 224                         crit_exit();
 225                         break;
 226                 }
 227                 crit_exit();
 228                 bioq_remove(&sc->bio_queue, bio);
 229                 bp = bio->bio_buf;
 230
 231                 devstat_start_transaction(&sc->stats);
 232
 233                 switch (bp->b_cmd) {
 234                 case BUF_CMD_FREEBLKS:
 235                 case BUF_CMD_READ:
 236                 case BUF_CMD_WRITE:
 237                         break;
 238                 default:
 239                         panic("md: bad b_cmd %d", bp->b_cmd);
 240                 }
 241
 242                 nsec = bp->b_bcount >> DEV_BSHIFT;
 243                 secno = (unsigned)(bio->bio_offset >> DEV_BSHIFT);
 244                 dst = bp->b_data;
 245                 while (nsec--) {
 246                         if (secno < sc->nsecp) {
 247                                 secpp = &sc->secp[secno];
 248                                 if ((u_int)(uintptr_t)*secpp > 255) {
 249                                         secp = *secpp;
 250                                         secval = 0;
 251                                 } else {
 252                                         secp = NULL;
 253                                         secval = (u_int)(uintptr_t)*secpp;
 254                                 }
 255                         } else {
 256                                 secpp = NULL;
 257                                 secp = NULL;
 258                                 secval = 0;
 259                         }
 260                         if (md_debug > 2)
 261                                 kprintf("%08x %p %p %d\n", bp->b_flags, secpp, secp, secval);
 262
 263                         switch (bp->b_cmd) {
 264                         case BUF_CMD_FREEBLKS:
 265                                 if (secpp) {
 266                                         if (secp)
 267                                                 kfree(secp, M_MDSECT);
 268                                         *secpp = NULL;
 269                                 }
 270                                 break;
 271                         case BUF_CMD_READ:
 272                                 if (secp) {
 273                                         bcopy(secp, dst, DEV_BSIZE);
 274                                 } else if (secval) {
 275                                         for (i = 0; i < DEV_BSIZE; i++)
 276                                                 dst[i] = secval;
 277                                 } else {
 278                                         bzero(dst, DEV_BSIZE);
 279                                 }
 280                                 break;
 281                         case BUF_CMD_WRITE:
 282                                 uc = dst[0];
 283                                 for (i = 1; i < DEV_BSIZE; i++)
 284                                         if (dst[i] != uc)
 285                                                 break;
 286                                 if (i == DEV_BSIZE && !uc) {
 287                                         if (secp)
 288                                                 kfree(secp, M_MDSECT);
 289                                         if (secpp)
 290                                                 *secpp = (u_char *)(uintptr_t)uc;
 291                                 } else {
 292                                         if (!secpp) {
 293                                                 secpp = kmalloc((secno + nsec + 1) * sizeof(u_char *),
 294                                                                 M_MD,
 295                                                                 M_WAITOK | M_ZERO);
 296                                                 bcopy(sc->secp, secpp, sc->nsecp * sizeof(u_char *));
 297                                                 kfree(sc->secp, M_MD);
 298                                                 sc->secp = secpp;
 299                                                 sc->nsecp = secno + nsec + 1;
 300                                                 secpp = &sc->secp[secno];
 301                                         }
 302                                         if (i == DEV_BSIZE) {
 303                                                 if (secp)
 304                                                         kfree(secp, M_MDSECT);
 305                                                 *secpp = (u_char *)(uintptr_t)uc;
 306                                         } else {
 307                                                 if (!secp)
 308                                                         secp = kmalloc(DEV_BSIZE,
 309                                                                        M_MDSECT,
 310                                                                        M_WAITOK);
 311                                                 bcopy(dst, secp, DEV_BSIZE);
 312
 313                                                 *secpp = secp;
 314                                         }
 315                                 }
 316                                 break;
 317                         default:
 318                                 panic("md: bad b_cmd %d", bp->b_cmd);
 319
 320                         }
 321                         secno++;
 322                         dst += DEV_BSIZE;
 323                 }
 324                 bp->b_resid = 0;
 325                 devstat_end_transaction_buf(&sc->stats, bp);
 326                 biodone(bio);
 327                 crit_enter();
 328         }
 329         sc->busy = 0;
 330         return(0);
 331 }
 332
 333
 334 static int
 335 mdstrategy_preload(struct dev_strategy_args *ap)
 336 {
 337         cdev_t dev = ap->a_head.a_dev;
 338         struct bio *bio = ap->a_bio;
 339         struct buf *bp = bio->bio_buf;
 340         struct md_s *sc;
 341
 342         if (md_debug > 1)
 343                 kprintf("mdstrategy_preload(%p) %s %08x, %lld, %d, %p)\n",
 344                     bp, devtoname(dev), bp->b_flags,
 345                     (long long)bio->bio_offset,
 346                     bp->b_bcount, bp->b_data);
 347
 348         sc = dev->si_drv1;
 349
 350         crit_enter();
 351
 352         bioqdisksort(&sc->bio_queue, bio);
 353
 354         if (sc->busy) {
 355                 crit_exit();
 356                 return(0);
 357         }
 358
 359         sc->busy++;
 360
 361         while (1) {
 362                 bio = bioq_first(&sc->bio_queue);
 363                 if (bio)
 364                         bioq_remove(&sc->bio_queue, bio);
 365                 crit_exit();
 366                 if (bio == NULL)
 367                         break;
 368
 369                 devstat_start_transaction(&sc->stats);
 370
 371                 switch (bp->b_cmd) {
 372                 case BUF_CMD_FREEBLKS:
 373                         break;
 374                 case BUF_CMD_READ:
 375                         bcopy(sc->pl_ptr + bio->bio_offset,
 376                                bp->b_data, bp->b_bcount);
 377                         break;
 378                 case BUF_CMD_WRITE:
 379                         bcopy(bp->b_data, sc->pl_ptr + bio->bio_offset,
 380                               bp->b_bcount);
 381                         break;
 382                 default:
 383                         panic("md: bad cmd %d\n", bp->b_cmd);
 384                 }
 385                 bp->b_resid = 0;
 386                 devstat_end_transaction_buf(&sc->stats, bp);
 387                 biodone(bio);
 388                 crit_enter();
 389         }
 390         sc->busy = 0;
 391         return(0);
 392 }
 393
 394 static struct md_s *
 395 mdcreate(unsigned length)
 396 {
 397         struct md_s *sc;
 398         struct disk_info info;
 399
 400         sc = kmalloc(sizeof(*sc), M_MD, M_WAITOK | M_ZERO);
 401         sc->unit = mdunits++;
 402         bioq_init(&sc->bio_queue);
 403         devstat_add_entry(&sc->stats, "md", sc->unit, DEV_BSIZE,
 404                 DEVSTAT_NO_ORDERED_TAGS,
 405                 DEVSTAT_TYPE_DIRECT | DEVSTAT_TYPE_IF_OTHER,
 406                 DEVSTAT_PRIORITY_OTHER);
 407         sc->dev = disk_create(sc->unit, &sc->disk, &md_ops);
 408         sc->dev->si_drv1 = sc;
 409         sc->dev->si_iosize_max = DFLTPHYS;
 410         disk_setdisktype(&sc->disk, "memory");
 411
 412         bzero(&info, sizeof(info));
 413         info.d_media_blksize = DEV_BSIZE;       /* mandatory */
 414         info.d_media_blocks = length / DEV_BSIZE;
 415
 416         info.d_secpertrack = 1024;              /* optional */
 417         info.d_nheads = 1;
 418         info.d_secpercyl = info.d_secpertrack * info.d_nheads;
 419         info.d_ncylinders = (u_int)(info.d_media_blocks / info.d_secpercyl);
 420         disk_setdiskinfo(&sc->disk, &info);
 421         TAILQ_INSERT_HEAD(&mdlist, sc, link);
 422
 423         return (sc);
 424 }
 425
 426
 427 static void
 428 mdcreate_preload(u_char *image, unsigned length)
 429 {
 430         struct md_s *sc;
 431
 432         sc = mdcreate(length);
 433         sc->type = MD_PRELOAD;
 434         sc->nsect = length / DEV_BSIZE;
 435         sc->pl_ptr = image;
 436         sc->pl_len = length;
 437
 438         if (sc->unit == 0)
 439                 mdrootready = 1;
 440 }
 441
 442 static void
 443 mdcreate_malloc(void)
 444 {
 445         struct md_s *sc;
 446
 447         sc = mdcreate(MD_NSECT*DEV_BSIZE);
 448         sc->type = MD_MALLOC;
 449
 450         sc->nsect = MD_NSECT;   /* for now */
 451         sc->secp = kmalloc(sizeof(u_char *), M_MD, M_WAITOK | M_ZERO);
 452         sc->nsecp = 1;
 453         kprintf("md%d: Malloc disk\n", sc->unit);
 454 }
 455
 456 static int
 457 md_drvcleanup(void)
 458 {
 459
 460         int secno;
 461         struct md_s *sc, *sc_temp;
 462
 463         if (atomic_fetchadd_int(&refcnt, 0) != 0)
 464                 return EBUSY;
 465
 466         /*
 467          * Go through all the md devices, freeing up all the
 468          * memory allocated for sectors, and the md_s struct
 469          * itself.
 470          */
 471         TAILQ_FOREACH_MUTABLE(sc, &mdlist, link, sc_temp) {
 472                 for (secno = 0; secno < sc->nsecp; secno++) {
 473                         if ((u_int)(uintptr_t)sc->secp[secno] > 255)
 474                                 kfree(sc->secp[secno], M_MDSECT);
 475                 }
 476
 477                 if (sc->dev != NULL)
 478                         disk_destroy(&sc->disk);
 479
 480                 devstat_remove_entry(&sc->stats);
 481                 TAILQ_REMOVE(&mdlist, sc, link);
 482
 483                 kfree(sc->secp, M_MD);
 484                 kfree(sc, M_MD);
 485         }
 486
 487         return 0;
 488
 489 }
 490
 491 static void
 492 md_drvinit(void *unused)
 493 {
 494
 495         caddr_t mod;
 496         caddr_t c;
 497         u_char *ptr, *name, *type;
 498         unsigned len;
 499
 500 #ifdef MD_ROOT_SIZE
 501         mdcreate_preload(mfs_root, MD_ROOT_SIZE*1024);
 502 #endif
 503         mod = NULL;
 504         while ((mod = preload_search_next_name(mod)) != NULL) {
 505                 name = (char *)preload_search_info(mod, MODINFO_NAME);
 506                 type = (char *)preload_search_info(mod, MODINFO_TYPE);
 507                 if (name == NULL)
 508                         continue;
 509                 if (type == NULL)
 510                         continue;
 511                 if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
 512                         continue;
 513                 c = preload_search_info(mod, MODINFO_ADDR);
 514                 ptr = *(u_char **)c;
 515                 c = preload_search_info(mod, MODINFO_SIZE);
 516                 len = *(unsigned *)c;
 517                 kprintf("md%d: Preloaded image <%s> %d bytes at %p\n",
 518                    mdunits, name, len, ptr);
 519                 mdcreate_preload(ptr, len);
 520         }
 521         mdcreate_malloc();
 522 }
 523
 524 DEV_MODULE(md, mdinit, NULL);
 525
 526 #ifdef MD_ROOT
 527 static void
 528 md_takeroot(void *junk)
 529 {
 530         if (mdrootready)
 531                 rootdevnames[0] = "ufs:/dev/md0s0";
 532 }
 533
 534 SYSINIT(md_root, SI_SUB_MOUNT_ROOT, SI_ORDER_FIRST, md_takeroot, NULL);
 535 #endif