drivers/char/mem.c

   1 /*
   2  *  linux/drivers/char/mem.c
   3  *
   4  *  Copyright (C) 1991, 1992  Linus Torvalds
   5  *
   6  *  Added devfs support.
   7  *    Jan-11-1998, C. Scott Ananian <cananian@alumni.princeton.edu>
   8  *  Shared /dev/zero mmaping support, Feb 2000, Kanoj Sarcar <kanoj@sgi.com>
   9  */
  10
  11 #include <linux/config.h>
  12 #include <linux/mm.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/slab.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/mman.h>
  17 #include <linux/random.h>
  18 #include <linux/init.h>
  19 #include <linux/raw.h>
  20 #include <linux/tty.h>
  21 #include <linux/capability.h>
  22 #include <linux/smp_lock.h>
  23 #include <linux/devfs_fs_kernel.h>
  24 #include <linux/ptrace.h>
  25 #include <linux/device.h>
  26
  27 #include <asm/uaccess.h>
  28 #include <asm/io.h>
  29
  30 #ifdef CONFIG_IA64
  31 # include <linux/efi.h>
  32 #endif
  33
  34 #if defined(CONFIG_S390_TAPE) && defined(CONFIG_S390_TAPE_CHAR)
  35 extern void tapechar_init(void);
  36 #endif
  37
  38 /*
  39  * Architectures vary in how they handle caching for addresses
  40  * outside of main memory.
  41  *
  42  */
  43 static inline int uncached_access(struct file *file, unsigned long addr)
  44 {
  45 #if defined(__i386__)
  46         /*
  47          * On the PPro and successors, the MTRRs are used to set
  48          * memory types for physical addresses outside main memory,
  49          * so blindly setting PCD or PWT on those pages is wrong.
  50          * For Pentiums and earlier, the surround logic should disable
  51          * caching for the high addresses through the KEN pin, but
  52          * we maintain the tradition of paranoia in this code.
  53          */
  54         if (file->f_flags & O_SYNC)
  55                 return 1;
  56         return !( test_bit(X86_FEATURE_MTRR, boot_cpu_data.x86_capability) ||
  57                   test_bit(X86_FEATURE_K6_MTRR, boot_cpu_data.x86_capability) ||
  58                   test_bit(X86_FEATURE_CYRIX_ARR, boot_cpu_data.x86_capability) ||
  59                   test_bit(X86_FEATURE_CENTAUR_MCR, boot_cpu_data.x86_capability) )
  60           && addr >= __pa(high_memory);
  61 #elif defined(__x86_64__)
  62         /*
  63          * This is broken because it can generate memory type aliases,
  64          * which can cause cache corruptions
  65          * But it is only available for root and we have to be bug-to-bug
  66          * compatible with i386.
  67          */
  68         if (file->f_flags & O_SYNC)
  69                 return 1;
  70         /* same behaviour as i386. PAT always set to cached and MTRRs control the
  71            caching behaviour.
  72            Hopefully a full PAT implementation will fix that soon. */
  73         return 0;
  74 #elif defined(CONFIG_IA64)
  75         /*
  76          * On ia64, we ignore O_SYNC because we cannot tolerate memory attribute aliases.
  77          */
  78         return !(efi_mem_attributes(addr) & EFI_MEMORY_WB);
  79 #elif defined(CONFIG_PPC64)
  80         /* On PPC64, we always do non-cacheable access to the IO hole and
  81          * cacheable elsewhere. Cache paradox can checkstop the CPU and
  82          * the high_memory heuristic below is wrong on machines with memory
  83          * above the IO hole... Ah, and of course, XFree86 doesn't pass
  84          * O_SYNC when mapping us to tap IO space. Surprised ?
  85          */
  86         return !page_is_ram(addr >> PAGE_SHIFT);
  87 #else
  88         /*
  89          * Accessing memory above the top the kernel knows about or through a file pointer
  90          * that was marked O_SYNC will be done non-cached.
  91          */
  92         if (file->f_flags & O_SYNC)
  93                 return 1;
  94         return addr >= __pa(high_memory);
  95 #endif
  96 }
  97
  98 #ifndef ARCH_HAS_VALID_PHYS_ADDR_RANGE
  99 static inline int valid_phys_addr_range(unsigned long addr, size_t *count)
 100 {
 101         unsigned long end_mem;
 102
 103         end_mem = __pa(high_memory);
 104         if (addr >= end_mem)
 105                 return 0;
 106
 107         if (*count > end_mem - addr)
 108                 *count = end_mem - addr;
 109
 110         return 1;
 111 }
 112 #endif
 113
 114 static inline int range_is_allowed(unsigned long from, unsigned long to)
 115 {
 116         unsigned long cursor;
 117
 118         cursor = from >> PAGE_SHIFT;
 119         while ((cursor << PAGE_SHIFT) < to) {
 120                 if (!devmem_is_allowed(cursor))
 121                         return 0;
 122                 cursor++;
 123         }
 124         return 1;
 125 }
 126 static ssize_t do_write_mem(void *p, unsigned long realp,
 127                             const char __user * buf, size_t count, loff_t *ppos)
 128 {
 129         ssize_t written;
 130         unsigned long copied;
 131
 132         written = 0;
 133 #if defined(__sparc__) || (defined(__mc68000__) && defined(CONFIG_MMU))
 134         /* we don't have page 0 mapped on sparc and m68k.. */
 135         if (realp < PAGE_SIZE) {
 136                 unsigned long sz = PAGE_SIZE-realp;
 137                 if (sz > count) sz = count;
 138                 /* Hmm. Do something? */
 139                 buf+=sz;
 140                 p+=sz;
 141                 count-=sz;
 142                 written+=sz;
 143         }
 144 #endif
 145         if (!range_is_allowed(realp, realp+count))
 146                 return -EPERM;
 147         copied = copy_from_user(p, buf, count);
 148         if (copied) {
 149                 ssize_t ret = written + (count - copied);
 150
 151                 if (ret)
 152                         return ret;
 153                 return -EFAULT;
 154         }
 155         written += count;
 156         *ppos += written;
 157         return written;
 158 }
 159
 160
 161 /*
 162  * This funcion reads the *physical* memory. The f_pos points directly to the
 163  * memory location.
 164  */
 165 static ssize_t read_mem(struct file * file, char __user * buf,
 166                         size_t count, loff_t *ppos)
 167 {
 168         unsigned long p = *ppos;
 169         ssize_t read;
 170
 171         if (!valid_phys_addr_range(p, &count))
 172                 return -EFAULT;
 173         read = 0;
 174 #if defined(__sparc__) || (defined(__mc68000__) && defined(CONFIG_MMU))
 175         /* we don't have page 0 mapped on sparc and m68k.. */
 176         if (p < PAGE_SIZE) {
 177                 unsigned long sz = PAGE_SIZE-p;
 178                 if (sz > count)
 179                         sz = count;
 180                 if (sz > 0) {
 181                         if (clear_user(buf, sz))
 182                                 return -EFAULT;
 183                         buf += sz;
 184                         p += sz;
 185                         count -= sz;
 186                         read += sz;
 187                 }
 188         }
 189 #endif
 190         if (!range_is_allowed(p, p+count))
 191                 return -EPERM;
 192         if (copy_to_user(buf, __va(p), count))
 193                 return -EFAULT;
 194         read += count;
 195         *ppos += read;
 196         return read;
 197 }
 198
 199 static ssize_t write_mem(struct file * file, const char __user * buf,
 200                          size_t count, loff_t *ppos)
 201 {
 202         unsigned long p = *ppos;
 203
 204         if (!valid_phys_addr_range(p, &count))
 205                 return -EFAULT;
 206         return do_write_mem(__va(p), p, buf, count, ppos);
 207 }
 208
 209 static int mmap_mem(struct file * file, struct vm_area_struct * vma)
 210 {
 211         unsigned long offset = vma->vm_pgoff << PAGE_SHIFT;
 212         int uncached;
 213
 214         uncached = uncached_access(file, offset);
 215 #ifdef pgprot_noncached
 216         if (uncached)
 217                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 218 #endif
 219
 220         /* Don't try to swap out physical pages.. */
 221         vma->vm_flags |= VM_RESERVED;
 222
 223         /*
 224          * Don't dump addresses that are not real memory to a core file.
 225          */
 226         if (uncached)
 227                 vma->vm_flags |= VM_IO;
 228
 229         if (remap_page_range(vma, vma->vm_start, offset, vma->vm_end-vma->vm_start,
 230                              vma->vm_page_prot))
 231                 return -EAGAIN;
 232         return 0;
 233 }
 234
 235 extern long vread(char *buf, char *addr, unsigned long count);
 236 extern long vwrite(char *buf, char *addr, unsigned long count);
 237
 238 /*
 239  * This function reads the *virtual* memory as seen by the kernel.
 240  */
 241 static ssize_t read_kmem(struct file *file, char __user *buf,
 242                          size_t count, loff_t *ppos)
 243 {
 244         unsigned long p = *ppos;
 245         ssize_t read = 0;
 246         ssize_t virtr = 0;
 247         char * kbuf; /* k-addr because vread() takes vmlist_lock rwlock */
 248
 249         return -EPERM;
 250
 251         if (p < (unsigned long) high_memory) {
 252                 read = count;
 253                 if (count > (unsigned long) high_memory - p)
 254                         read = (unsigned long) high_memory - p;
 255
 256 #if defined(__sparc__) || (defined(__mc68000__) && defined(CONFIG_MMU))
 257                 /* we don't have page 0 mapped on sparc and m68k.. */
 258                 if (p < PAGE_SIZE && read > 0) {
 259                         size_t tmp = PAGE_SIZE - p;
 260                         if (tmp > read) tmp = read;
 261                         if (clear_user(buf, tmp))
 262                                 return -EFAULT;
 263                         buf += tmp;
 264                         p += tmp;
 265                         read -= tmp;
 266                         count -= tmp;
 267                 }
 268 #endif
 269                 if (copy_to_user(buf, (char *)p, read))
 270                         return -EFAULT;
 271                 p += read;
 272                 buf += read;
 273                 count -= read;
 274         }
 275
 276         if (count > 0) {
 277                 kbuf = (char *)__get_free_page(GFP_KERNEL);
 278                 if (!kbuf)
 279                         return -ENOMEM;
 280                 while (count > 0) {
 281                         int len = count;
 282
 283                         if (len > PAGE_SIZE)
 284                                 len = PAGE_SIZE;
 285                         len = vread(kbuf, (char *)p, len);
 286                         if (!len)
 287                                 break;
 288                         if (copy_to_user(buf, kbuf, len)) {
 289                                 free_page((unsigned long)kbuf);
 290                                 return -EFAULT;
 291                         }
 292                         count -= len;
 293                         buf += len;
 294                         virtr += len;
 295                         p += len;
 296                 }
 297                 free_page((unsigned long)kbuf);
 298         }
 299         *ppos = p;
 300         return virtr + read;
 301 }
 302
 303 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 304 static ssize_t read_port(struct file * file, char __user * buf,
 305                          size_t count, loff_t *ppos)
 306 {
 307         unsigned long i = *ppos;
 308         char __user *tmp = buf;
 309
 310         if (verify_area(VERIFY_WRITE,buf,count))
 311                 return -EFAULT;
 312         while (count-- > 0 && i < 65536) {
 313                 if (__put_user(inb(i),tmp) < 0)
 314                         return -EFAULT;
 315                 i++;
 316                 tmp++;
 317         }
 318         *ppos = i;
 319         return tmp-buf;
 320 }
 321
 322 static ssize_t write_port(struct file * file, const char __user * buf,
 323                           size_t count, loff_t *ppos)
 324 {
 325         unsigned long i = *ppos;
 326         const char __user * tmp = buf;
 327
 328         if (verify_area(VERIFY_READ,buf,count))
 329                 return -EFAULT;
 330         while (count-- > 0 && i < 65536) {
 331                 char c;
 332                 if (__get_user(c, tmp))
 333                         return -EFAULT;
 334                 outb(c,i);
 335                 i++;
 336                 tmp++;
 337         }
 338         *ppos = i;
 339         return tmp-buf;
 340 }
 341 #endif
 342
 343 static ssize_t read_null(struct file * file, char __user * buf,
 344                          size_t count, loff_t *ppos)
 345 {
 346         return 0;
 347 }
 348
 349 static ssize_t write_null(struct file * file, const char __user * buf,
 350                           size_t count, loff_t *ppos)
 351 {
 352         return count;
 353 }
 354
 355 #ifdef CONFIG_MMU
 356 /*
 357  * For fun, we are using the MMU for this.
 358  */
 359 static inline size_t read_zero_pagealigned(char __user * buf, size_t size)
 360 {
 361         struct mm_struct *mm;
 362         struct vm_area_struct * vma;
 363         unsigned long addr=(unsigned long)buf;
 364
 365         mm = current->mm;
 366         /* Oops, this was forgotten before. -ben */
 367         down_read(&mm->mmap_sem);
 368
 369         /* For private mappings, just map in zero pages. */
 370         for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
 371                 unsigned long count;
 372
 373                 if (vma->vm_start > addr || (vma->vm_flags & VM_WRITE) == 0)
 374                         goto out_up;
 375                 if (vma->vm_flags & (VM_SHARED | VM_HUGETLB))
 376                         break;
 377                 count = vma->vm_end - addr;
 378                 if (count > size)
 379                         count = size;
 380
 381                 zap_page_range(vma, addr, count, NULL);
 382                 zeromap_page_range(vma, addr, count, PAGE_COPY);
 383
 384                 size -= count;
 385                 buf += count;
 386                 addr += count;
 387                 if (size == 0)
 388                         goto out_up;
 389         }
 390
 391         up_read(&mm->mmap_sem);
 392
 393         /* The shared case is hard. Let's do the conventional zeroing. */
 394         do {
 395                 unsigned long unwritten = clear_user(buf, PAGE_SIZE);
 396                 if (unwritten)
 397                         return size + unwritten - PAGE_SIZE;
 398                 cond_resched();
 399                 buf += PAGE_SIZE;
 400                 size -= PAGE_SIZE;
 401         } while (size);
 402
 403         return size;
 404 out_up:
 405         up_read(&mm->mmap_sem);
 406         return size;
 407 }
 408
 409 static ssize_t read_zero(struct file * file, char __user * buf,
 410                          size_t count, loff_t *ppos)
 411 {
 412         unsigned long left, unwritten, written = 0;
 413
 414         if (!count)
 415                 return 0;
 416
 417         if (!access_ok(VERIFY_WRITE, buf, count))
 418                 return -EFAULT;
 419
 420         left = count;
 421
 422         /* do we want to be clever? Arbitrary cut-off */
 423         if (count >= PAGE_SIZE*4) {
 424                 unsigned long partial;
 425
 426                 /* How much left of the page? */
 427                 partial = (PAGE_SIZE-1) & -(unsigned long) buf;
 428                 unwritten = clear_user(buf, partial);
 429                 written = partial - unwritten;
 430                 if (unwritten)
 431                         goto out;
 432                 left -= partial;
 433                 buf += partial;
 434                 unwritten = read_zero_pagealigned(buf, left & PAGE_MASK);
 435                 written += (left & PAGE_MASK) - unwritten;
 436                 if (unwritten)
 437                         goto out;
 438                 buf += left & PAGE_MASK;
 439                 left &= ~PAGE_MASK;
 440         }
 441         unwritten = clear_user(buf, left);
 442         written += left - unwritten;
 443 out:
 444         return written ? written : -EFAULT;
 445 }
 446
 447 static int mmap_zero(struct file * file, struct vm_area_struct * vma)
 448 {
 449         if (vma->vm_flags & VM_SHARED)
 450                 return shmem_zero_setup(vma);
 451         if (zeromap_page_range(vma, vma->vm_start, vma->vm_end - vma->vm_start, vma->vm_page_prot))
 452                 return -EAGAIN;
 453         return 0;
 454 }
 455 #else /* CONFIG_MMU */
 456 static ssize_t read_zero(struct file * file, char * buf,
 457                          size_t count, loff_t *ppos)
 458 {
 459         size_t todo = count;
 460
 461         while (todo) {
 462                 size_t chunk = todo;
 463
 464                 if (chunk > 4096)
 465                         chunk = 4096;   /* Just for latency reasons */
 466                 if (clear_user(buf, chunk))
 467                         return -EFAULT;
 468                 buf += chunk;
 469                 todo -= chunk;
 470                 cond_resched();
 471         }
 472         return count;
 473 }
 474
 475 static int mmap_zero(struct file * file, struct vm_area_struct * vma)
 476 {
 477         return -ENOSYS;
 478 }
 479 #endif /* CONFIG_MMU */
 480
 481 static ssize_t write_full(struct file * file, const char __user * buf,
 482                           size_t count, loff_t *ppos)
 483 {
 484         return -ENOSPC;
 485 }
 486
 487 /*
 488  * Special lseek() function for /dev/null and /dev/zero.  Most notably, you
 489  * can fopen() both devices with "a" now.  This was previously impossible.
 490  * -- SRB.
 491  */
 492
 493 static loff_t null_lseek(struct file * file, loff_t offset, int orig)
 494 {
 495         return file->f_pos = 0;
 496 }
 497
 498 /*
 499  * The memory devices use the full 32/64 bits of the offset, and so we cannot
 500  * check against negative addresses: they are ok. The return value is weird,
 501  * though, in that case (0).
 502  *
 503  * also note that seeking relative to the "end of file" isn't supported:
 504  * it has no meaning, so it returns -EINVAL.
 505  */
 506 static loff_t memory_lseek(struct file * file, loff_t offset, int orig)
 507 {
 508         loff_t ret;
 509
 510         down(&file->f_dentry->d_inode->i_sem);
 511         switch (orig) {
 512                 case 0:
 513                         file->f_pos = offset;
 514                         ret = file->f_pos;
 515                         force_successful_syscall_return();
 516                         break;
 517                 case 1:
 518                         file->f_pos += offset;
 519                         ret = file->f_pos;
 520                         force_successful_syscall_return();
 521                         break;
 522                 default:
 523                         ret = -EINVAL;
 524         }
 525         up(&file->f_dentry->d_inode->i_sem);
 526         return ret;
 527 }
 528
 529 static int open_port(struct inode * inode, struct file * filp)
 530 {
 531         return capable(CAP_SYS_RAWIO) ? 0 : -EPERM;
 532 }
 533
 534 #define mmap_kmem       mmap_mem
 535 #define zero_lseek      null_lseek
 536 #define full_lseek      null_lseek
 537 #define write_zero      write_null
 538 #define read_full       read_zero
 539 #define open_mem        open_port
 540 #define open_kmem       open_mem
 541
 542 static struct file_operations mem_fops = {
 543         .llseek         = memory_lseek,
 544         .read           = read_mem,
 545         .write          = write_mem,
 546         .mmap           = mmap_mem,
 547         .open           = open_mem,
 548 };
 549
 550 static struct file_operations kmem_fops = {
 551         .llseek         = memory_lseek,
 552         .read           = read_kmem,
 553         .mmap           = mmap_kmem,
 554         .open           = open_kmem,
 555 };
 556
 557 static struct file_operations null_fops = {
 558         .llseek         = null_lseek,
 559         .read           = read_null,
 560         .write          = write_null,
 561 };
 562
 563 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 564 static struct file_operations port_fops = {
 565         .llseek         = memory_lseek,
 566         .read           = read_port,
 567         .write          = write_port,
 568         .open           = open_port,
 569 };
 570 #endif
 571
 572 static struct file_operations zero_fops = {
 573         .llseek         = zero_lseek,
 574         .read           = read_zero,
 575         .write          = write_zero,
 576         .mmap           = mmap_zero,
 577 };
 578
 579 static struct file_operations full_fops = {
 580         .llseek         = full_lseek,
 581         .read           = read_full,
 582         .write          = write_full,
 583 };
 584
 585 static ssize_t kmsg_write(struct file * file, const char __user * buf,
 586                           size_t count, loff_t *ppos)
 587 {
 588         char *tmp;
 589         int ret;
 590
 591         tmp = kmalloc(count + 1, GFP_KERNEL);
 592         if (tmp == NULL)
 593                 return -ENOMEM;
 594         ret = -EFAULT;
 595         if (!copy_from_user(tmp, buf, count)) {
 596                 tmp[count] = 0;
 597                 ret = printk("%s", tmp);
 598         }
 599         kfree(tmp);
 600         return ret;
 601 }
 602
 603 static struct file_operations kmsg_fops = {
 604         .write =        kmsg_write,
 605 };
 606
 607 static int memory_open(struct inode * inode, struct file * filp)
 608 {
 609         switch (iminor(inode)) {
 610                 case 1:
 611                         filp->f_op = &mem_fops;
 612                         break;
 613                 case 2:
 614                         filp->f_op = &kmem_fops;
 615                         break;
 616                 case 3:
 617                         filp->f_op = &null_fops;
 618                         break;
 619 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 620                 case 4:
 621                         filp->f_op = &port_fops;
 622                         break;
 623 #endif
 624                 case 5:
 625                         filp->f_op = &zero_fops;
 626                         break;
 627                 case 7:
 628                         filp->f_op = &full_fops;
 629                         break;
 630                 case 8:
 631                         filp->f_op = &random_fops;
 632                         break;
 633                 case 9:
 634                         filp->f_op = &urandom_fops;
 635                         break;
 636                 case 11:
 637                         filp->f_op = &kmsg_fops;
 638                         break;
 639                 default:
 640                         return -ENXIO;
 641         }
 642         if (filp->f_op && filp->f_op->open)
 643                 return filp->f_op->open(inode,filp);
 644         return 0;
 645 }
 646
 647 static struct file_operations memory_fops = {
 648         .open           = memory_open,  /* just a selector for the real open */
 649 };
 650
 651 static const struct {
 652         unsigned int            minor;
 653         char                    *name;
 654         umode_t                 mode;
 655         struct file_operations  *fops;
 656 } devlist[] = { /* list of minor devices */
 657         {1, "mem",     S_IRUSR | S_IWUSR | S_IRGRP, &mem_fops},
 658         {3, "null",    S_IRUGO | S_IWUGO,           &null_fops},
 659 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 660         {4, "port",    S_IRUSR | S_IWUSR | S_IRGRP, &port_fops},
 661 #endif
 662         {5, "zero",    S_IRUGO | S_IWUGO,           &zero_fops},
 663         {7, "full",    S_IRUGO | S_IWUGO,           &full_fops},
 664         {8, "random",  S_IRUGO | S_IWUSR,           &random_fops},
 665         {9, "urandom", S_IRUGO | S_IWUSR,           &urandom_fops},
 666         {11,"kmsg",    S_IRUGO | S_IWUSR,           &kmsg_fops},
 667 };
 668
 669 static struct class_simple *mem_class;
 670
 671 static int __init chr_dev_init(void)
 672 {
 673         int i;
 674
 675         if (register_chrdev(MEM_MAJOR,"mem",&memory_fops))
 676                 printk("unable to get major %d for memory devs\n", MEM_MAJOR);
 677
 678         mem_class = class_simple_create(THIS_MODULE, "mem");
 679         for (i = 0; i < ARRAY_SIZE(devlist); i++) {
 680                 class_simple_device_add(mem_class,
 681                                         MKDEV(MEM_MAJOR, devlist[i].minor),
 682                                         NULL, devlist[i].name);
 683                 devfs_mk_cdev(MKDEV(MEM_MAJOR, devlist[i].minor),
 684                                 S_IFCHR | devlist[i].mode, devlist[i].name);
 685         }
 686
 687         return 0;
 688 }
 689
 690 fs_initcall(chr_dev_init);