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 #ifdef pgprot_noncached
 212         unsigned long offset = vma->vm_pgoff << PAGE_SHIFT;
 213         int uncached;
 214
 215         uncached = uncached_access(file, offset);
 216         if (uncached)
 217                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 218 #endif
 219
 220         /* Remap-pfn-range will mark the range VM_IO and VM_RESERVED */
 221         if (remap_pfn_range(vma,
 222                             vma->vm_start,
 223                             vma->vm_pgoff,
 224                             vma->vm_end-vma->vm_start,
 225                             vma->vm_page_prot))
 226                 return -EAGAIN;
 227         return 0;
 228 }
 229
 230 extern long vread(char *buf, char *addr, unsigned long count);
 231 extern long vwrite(char *buf, char *addr, unsigned long count);
 232
 233 /*
 234  * This function reads the *virtual* memory as seen by the kernel.
 235  */
 236 static ssize_t read_kmem(struct file *file, char __user *buf,
 237                          size_t count, loff_t *ppos)
 238 {
 239         unsigned long p = *ppos;
 240         ssize_t read = 0;
 241         ssize_t virtr = 0;
 242         char * kbuf; /* k-addr because vread() takes vmlist_lock rwlock */
 243
 244         return -EPERM;
 245
 246         if (p < (unsigned long) high_memory) {
 247                 read = count;
 248                 if (count > (unsigned long) high_memory - p)
 249                         read = (unsigned long) high_memory - p;
 250
 251 #if defined(__sparc__) || (defined(__mc68000__) && defined(CONFIG_MMU))
 252                 /* we don't have page 0 mapped on sparc and m68k.. */
 253                 if (p < PAGE_SIZE && read > 0) {
 254                         size_t tmp = PAGE_SIZE - p;
 255                         if (tmp > read) tmp = read;
 256                         if (clear_user(buf, tmp))
 257                                 return -EFAULT;
 258                         buf += tmp;
 259                         p += tmp;
 260                         read -= tmp;
 261                         count -= tmp;
 262                 }
 263 #endif
 264                 if (copy_to_user(buf, (char *)p, read))
 265                         return -EFAULT;
 266                 p += read;
 267                 buf += read;
 268                 count -= read;
 269         }
 270
 271         if (count > 0) {
 272                 kbuf = (char *)__get_free_page(GFP_KERNEL);
 273                 if (!kbuf)
 274                         return -ENOMEM;
 275                 while (count > 0) {
 276                         int len = count;
 277
 278                         if (len > PAGE_SIZE)
 279                                 len = PAGE_SIZE;
 280                         len = vread(kbuf, (char *)p, len);
 281                         if (!len)
 282                                 break;
 283                         if (copy_to_user(buf, kbuf, len)) {
 284                                 free_page((unsigned long)kbuf);
 285                                 return -EFAULT;
 286                         }
 287                         count -= len;
 288                         buf += len;
 289                         virtr += len;
 290                         p += len;
 291                 }
 292                 free_page((unsigned long)kbuf);
 293         }
 294         *ppos = p;
 295         return virtr + read;
 296 }
 297
 298 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 299 static ssize_t read_port(struct file * file, char __user * buf,
 300                          size_t count, loff_t *ppos)
 301 {
 302         unsigned long i = *ppos;
 303         char __user *tmp = buf;
 304
 305         if (verify_area(VERIFY_WRITE,buf,count))
 306                 return -EFAULT;
 307         while (count-- > 0 && i < 65536) {
 308                 if (__put_user(inb(i),tmp) < 0)
 309                         return -EFAULT;
 310                 i++;
 311                 tmp++;
 312         }
 313         *ppos = i;
 314         return tmp-buf;
 315 }
 316
 317 static ssize_t write_port(struct file * file, const char __user * buf,
 318                           size_t count, loff_t *ppos)
 319 {
 320         unsigned long i = *ppos;
 321         const char __user * tmp = buf;
 322
 323         if (verify_area(VERIFY_READ,buf,count))
 324                 return -EFAULT;
 325         while (count-- > 0 && i < 65536) {
 326                 char c;
 327                 if (__get_user(c, tmp))
 328                         return -EFAULT;
 329                 outb(c,i);
 330                 i++;
 331                 tmp++;
 332         }
 333         *ppos = i;
 334         return tmp-buf;
 335 }
 336 #endif
 337
 338 static ssize_t read_null(struct file * file, char __user * buf,
 339                          size_t count, loff_t *ppos)
 340 {
 341         return 0;
 342 }
 343
 344 static ssize_t write_null(struct file * file, const char __user * buf,
 345                           size_t count, loff_t *ppos)
 346 {
 347         return count;
 348 }
 349
 350 #ifdef CONFIG_MMU
 351 /*
 352  * For fun, we are using the MMU for this.
 353  */
 354 static inline size_t read_zero_pagealigned(char __user * buf, size_t size)
 355 {
 356         struct mm_struct *mm;
 357         struct vm_area_struct * vma;
 358         unsigned long addr=(unsigned long)buf;
 359
 360         mm = current->mm;
 361         /* Oops, this was forgotten before. -ben */
 362         down_read(&mm->mmap_sem);
 363
 364         /* For private mappings, just map in zero pages. */
 365         for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
 366                 unsigned long count;
 367
 368                 if (vma->vm_start > addr || (vma->vm_flags & VM_WRITE) == 0)
 369                         goto out_up;
 370                 if (vma->vm_flags & (VM_SHARED | VM_HUGETLB))
 371                         break;
 372                 count = vma->vm_end - addr;
 373                 if (count > size)
 374                         count = size;
 375
 376                 zap_page_range(vma, addr, count, NULL);
 377                 zeromap_page_range(vma, addr, count, PAGE_COPY);
 378
 379                 size -= count;
 380                 buf += count;
 381                 addr += count;
 382                 if (size == 0)
 383                         goto out_up;
 384         }
 385
 386         up_read(&mm->mmap_sem);
 387
 388         /* The shared case is hard. Let's do the conventional zeroing. */
 389         do {
 390                 unsigned long unwritten = clear_user(buf, PAGE_SIZE);
 391                 if (unwritten)
 392                         return size + unwritten - PAGE_SIZE;
 393                 cond_resched();
 394                 buf += PAGE_SIZE;
 395                 size -= PAGE_SIZE;
 396         } while (size);
 397
 398         return size;
 399 out_up:
 400         up_read(&mm->mmap_sem);
 401         return size;
 402 }
 403
 404 static ssize_t read_zero(struct file * file, char __user * buf,
 405                          size_t count, loff_t *ppos)
 406 {
 407         unsigned long left, unwritten, written = 0;
 408
 409         if (!count)
 410                 return 0;
 411
 412         if (!access_ok(VERIFY_WRITE, buf, count))
 413                 return -EFAULT;
 414
 415         left = count;
 416
 417         /* do we want to be clever? Arbitrary cut-off */
 418         if (count >= PAGE_SIZE*4) {
 419                 unsigned long partial;
 420
 421                 /* How much left of the page? */
 422                 partial = (PAGE_SIZE-1) & -(unsigned long) buf;
 423                 unwritten = clear_user(buf, partial);
 424                 written = partial - unwritten;
 425                 if (unwritten)
 426                         goto out;
 427                 left -= partial;
 428                 buf += partial;
 429                 unwritten = read_zero_pagealigned(buf, left & PAGE_MASK);
 430                 written += (left & PAGE_MASK) - unwritten;
 431                 if (unwritten)
 432                         goto out;
 433                 buf += left & PAGE_MASK;
 434                 left &= ~PAGE_MASK;
 435         }
 436         unwritten = clear_user(buf, left);
 437         written += left - unwritten;
 438 out:
 439         return written ? written : -EFAULT;
 440 }
 441
 442 static int mmap_zero(struct file * file, struct vm_area_struct * vma)
 443 {
 444         if (vma->vm_flags & VM_SHARED)
 445                 return shmem_zero_setup(vma);
 446         if (zeromap_page_range(vma, vma->vm_start, vma->vm_end - vma->vm_start, vma->vm_page_prot))
 447                 return -EAGAIN;
 448         return 0;
 449 }
 450 #else /* CONFIG_MMU */
 451 static ssize_t read_zero(struct file * file, char * buf,
 452                          size_t count, loff_t *ppos)
 453 {
 454         size_t todo = count;
 455
 456         while (todo) {
 457                 size_t chunk = todo;
 458
 459                 if (chunk > 4096)
 460                         chunk = 4096;   /* Just for latency reasons */
 461                 if (clear_user(buf, chunk))
 462                         return -EFAULT;
 463                 buf += chunk;
 464                 todo -= chunk;
 465                 cond_resched();
 466         }
 467         return count;
 468 }
 469
 470 static int mmap_zero(struct file * file, struct vm_area_struct * vma)
 471 {
 472         return -ENOSYS;
 473 }
 474 #endif /* CONFIG_MMU */
 475
 476 static ssize_t write_full(struct file * file, const char __user * buf,
 477                           size_t count, loff_t *ppos)
 478 {
 479         return -ENOSPC;
 480 }
 481
 482 /*
 483  * Special lseek() function for /dev/null and /dev/zero.  Most notably, you
 484  * can fopen() both devices with "a" now.  This was previously impossible.
 485  * -- SRB.
 486  */
 487
 488 static loff_t null_lseek(struct file * file, loff_t offset, int orig)
 489 {
 490         return file->f_pos = 0;
 491 }
 492
 493 /*
 494  * The memory devices use the full 32/64 bits of the offset, and so we cannot
 495  * check against negative addresses: they are ok. The return value is weird,
 496  * though, in that case (0).
 497  *
 498  * also note that seeking relative to the "end of file" isn't supported:
 499  * it has no meaning, so it returns -EINVAL.
 500  */
 501 static loff_t memory_lseek(struct file * file, loff_t offset, int orig)
 502 {
 503         loff_t ret;
 504
 505         down(&file->f_dentry->d_inode->i_sem);
 506         switch (orig) {
 507                 case 0:
 508                         file->f_pos = offset;
 509                         ret = file->f_pos;
 510                         force_successful_syscall_return();
 511                         break;
 512                 case 1:
 513                         file->f_pos += offset;
 514                         ret = file->f_pos;
 515                         force_successful_syscall_return();
 516                         break;
 517                 default:
 518                         ret = -EINVAL;
 519         }
 520         up(&file->f_dentry->d_inode->i_sem);
 521         return ret;
 522 }
 523
 524 static int open_port(struct inode * inode, struct file * filp)
 525 {
 526         return capable(CAP_SYS_RAWIO) ? 0 : -EPERM;
 527 }
 528
 529 #define mmap_kmem       mmap_mem
 530 #define zero_lseek      null_lseek
 531 #define full_lseek      null_lseek
 532 #define write_zero      write_null
 533 #define read_full       read_zero
 534 #define open_mem        open_port
 535 #define open_kmem       open_mem
 536
 537 static struct file_operations mem_fops = {
 538         .llseek         = memory_lseek,
 539         .read           = read_mem,
 540         .write          = write_mem,
 541         .mmap           = mmap_mem,
 542         .open           = open_mem,
 543 };
 544
 545 static struct file_operations kmem_fops = {
 546         .llseek         = memory_lseek,
 547         .read           = read_kmem,
 548         .mmap           = mmap_kmem,
 549         .open           = open_kmem,
 550 };
 551
 552 static struct file_operations null_fops = {
 553         .llseek         = null_lseek,
 554         .read           = read_null,
 555         .write          = write_null,
 556 };
 557
 558 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 559 static struct file_operations port_fops = {
 560         .llseek         = memory_lseek,
 561         .read           = read_port,
 562         .write          = write_port,
 563         .open           = open_port,
 564 };
 565 #endif
 566
 567 static struct file_operations zero_fops = {
 568         .llseek         = zero_lseek,
 569         .read           = read_zero,
 570         .write          = write_zero,
 571         .mmap           = mmap_zero,
 572 };
 573
 574 static struct file_operations full_fops = {
 575         .llseek         = full_lseek,
 576         .read           = read_full,
 577         .write          = write_full,
 578 };
 579
 580 static ssize_t kmsg_write(struct file * file, const char __user * buf,
 581                           size_t count, loff_t *ppos)
 582 {
 583         char *tmp;
 584         int ret;
 585
 586         tmp = kmalloc(count + 1, GFP_KERNEL);
 587         if (tmp == NULL)
 588                 return -ENOMEM;
 589         ret = -EFAULT;
 590         if (!copy_from_user(tmp, buf, count)) {
 591                 tmp[count] = 0;
 592                 ret = printk("%s", tmp);
 593         }
 594         kfree(tmp);
 595         return ret;
 596 }
 597
 598 static struct file_operations kmsg_fops = {
 599         .write =        kmsg_write,
 600 };
 601
 602 static int memory_open(struct inode * inode, struct file * filp)
 603 {
 604         switch (iminor(inode)) {
 605                 case 1:
 606                         filp->f_op = &mem_fops;
 607                         break;
 608                 case 2:
 609                         filp->f_op = &kmem_fops;
 610                         break;
 611                 case 3:
 612                         filp->f_op = &null_fops;
 613                         break;
 614 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 615                 case 4:
 616                         filp->f_op = &port_fops;
 617                         break;
 618 #endif
 619                 case 5:
 620                         filp->f_op = &zero_fops;
 621                         break;
 622                 case 7:
 623                         filp->f_op = &full_fops;
 624                         break;
 625                 case 8:
 626                         filp->f_op = &random_fops;
 627                         break;
 628                 case 9:
 629                         filp->f_op = &urandom_fops;
 630                         break;
 631                 case 11:
 632                         filp->f_op = &kmsg_fops;
 633                         break;
 634                 default:
 635                         return -ENXIO;
 636         }
 637         if (filp->f_op && filp->f_op->open)
 638                 return filp->f_op->open(inode,filp);
 639         return 0;
 640 }
 641
 642 static struct file_operations memory_fops = {
 643         .open           = memory_open,  /* just a selector for the real open */
 644 };
 645
 646 static const struct {
 647         unsigned int            minor;
 648         char                    *name;
 649         umode_t                 mode;
 650         struct file_operations  *fops;
 651 } devlist[] = { /* list of minor devices */
 652         {1, "mem",     S_IRUSR | S_IWUSR | S_IRGRP, &mem_fops},
 653         {3, "null",    S_IRUGO | S_IWUGO,           &null_fops},
 654 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 655         {4, "port",    S_IRUSR | S_IWUSR | S_IRGRP, &port_fops},
 656 #endif
 657         {5, "zero",    S_IRUGO | S_IWUGO,           &zero_fops},
 658         {7, "full",    S_IRUGO | S_IWUGO,           &full_fops},
 659         {8, "random",  S_IRUGO | S_IWUSR,           &random_fops},
 660         {9, "urandom", S_IRUGO | S_IWUSR,           &urandom_fops},
 661         {11,"kmsg",    S_IRUGO | S_IWUSR,           &kmsg_fops},
 662 };
 663
 664 static struct class_simple *mem_class;
 665
 666 static int __init chr_dev_init(void)
 667 {
 668         int i;
 669
 670         if (register_chrdev(MEM_MAJOR,"mem",&memory_fops))
 671                 printk("unable to get major %d for memory devs\n", MEM_MAJOR);
 672
 673         mem_class = class_simple_create(THIS_MODULE, "mem");
 674         for (i = 0; i < ARRAY_SIZE(devlist); i++) {
 675                 class_simple_device_add(mem_class,
 676                                         MKDEV(MEM_MAJOR, devlist[i].minor),
 677                                         NULL, devlist[i].name);
 678                 devfs_mk_cdev(MKDEV(MEM_MAJOR, devlist[i].minor),
 679                                 S_IFCHR | devlist[i].mode, devlist[i].name);
 680         }
 681
 682         return 0;
 683 }
 684
 685 fs_initcall(chr_dev_init);