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 #ifdef CONFIG_FB
  35 extern void fbmem_init(void);
  36 #endif
  37 #if defined(CONFIG_S390_TAPE) && defined(CONFIG_S390_TAPE_CHAR)
  38 extern void tapechar_init(void);
  39 #endif
  40
  41 /*
  42  * Architectures vary in how they handle caching for addresses
  43  * outside of main memory.
  44  *
  45  */
  46 static inline int uncached_access(struct file *file, unsigned long addr)
  47 {
  48 #if defined(__i386__)
  49         /*
  50          * On the PPro and successors, the MTRRs are used to set
  51          * memory types for physical addresses outside main memory,
  52          * so blindly setting PCD or PWT on those pages is wrong.
  53          * For Pentiums and earlier, the surround logic should disable
  54          * caching for the high addresses through the KEN pin, but
  55          * we maintain the tradition of paranoia in this code.
  56          */
  57         if (file->f_flags & O_SYNC)
  58                 return 1;
  59         return !( test_bit(X86_FEATURE_MTRR, boot_cpu_data.x86_capability) ||
  60                   test_bit(X86_FEATURE_K6_MTRR, boot_cpu_data.x86_capability) ||
  61                   test_bit(X86_FEATURE_CYRIX_ARR, boot_cpu_data.x86_capability) ||
  62                   test_bit(X86_FEATURE_CENTAUR_MCR, boot_cpu_data.x86_capability) )
  63           && addr >= __pa(high_memory);
  64 #elif defined(__x86_64__)
  65         /*
  66          * This is broken because it can generate memory type aliases,
  67          * which can cause cache corruptions
  68          * But it is only available for root and we have to be bug-to-bug
  69          * compatible with i386.
  70          */
  71         if (file->f_flags & O_SYNC)
  72                 return 1;
  73         /* same behaviour as i386. PAT always set to cached and MTRRs control the
  74            caching behaviour.
  75            Hopefully a full PAT implementation will fix that soon. */
  76         return 0;
  77 #elif defined(CONFIG_IA64)
  78         /*
  79          * On ia64, we ignore O_SYNC because we cannot tolerate memory attribute aliases.
  80          */
  81         return !(efi_mem_attributes(addr) & EFI_MEMORY_WB);
  82 #elif defined(CONFIG_PPC64)
  83         /* On PPC64, we always do non-cacheable access to the IO hole and
  84          * cacheable elsewhere. Cache paradox can checkstop the CPU and
  85          * the high_memory heuristic below is wrong on machines with memory
  86          * above the IO hole... Ah, and of course, XFree86 doesn't pass
  87          * O_SYNC when mapping us to tap IO space. Surprised ?
  88          */
  89         return !page_is_ram(addr);
  90 #else
  91         /*
  92          * Accessing memory above the top the kernel knows about or through a file pointer
  93          * that was marked O_SYNC will be done non-cached.
  94          */
  95         if (file->f_flags & O_SYNC)
  96                 return 1;
  97         return addr >= __pa(high_memory);
  98 #endif
  99 }
 100
 101 #ifndef ARCH_HAS_VALID_PHYS_ADDR_RANGE
 102 static inline int valid_phys_addr_range(unsigned long addr, size_t *count)
 103 {
 104         unsigned long end_mem;
 105
 106         end_mem = __pa(high_memory);
 107         if (addr >= end_mem)
 108                 return 0;
 109
 110         if (*count > end_mem - addr)
 111                 *count = end_mem - addr;
 112
 113         return 1;
 114 }
 115 #endif
 116
 117 static inline int range_is_allowed(unsigned long from, unsigned long to)
 118 {
 119         unsigned long cursor;
 120
 121         cursor = from >> PAGE_SHIFT;
 122         while ((cursor << PAGE_SHIFT) < to) {
 123                 if (!devmem_is_allowed(cursor))
 124                         return 0;
 125                 cursor++;
 126         }
 127         return 1;
 128 }
 129 static ssize_t do_write_mem(void *p, unsigned long realp,
 130                             const char __user * buf, size_t count, loff_t *ppos)
 131 {
 132         ssize_t written;
 133         unsigned long copied;
 134
 135         written = 0;
 136 #if defined(__sparc__) || (defined(__mc68000__) && defined(CONFIG_MMU))
 137         /* we don't have page 0 mapped on sparc and m68k.. */
 138         if (realp < PAGE_SIZE) {
 139                 unsigned long sz = PAGE_SIZE-realp;
 140                 if (sz > count) sz = count;
 141                 /* Hmm. Do something? */
 142                 buf+=sz;
 143                 p+=sz;
 144                 count-=sz;
 145                 written+=sz;
 146         }
 147 #endif
 148         if (!range_is_allowed(realp, realp+count))
 149                 return -EPERM;
 150         copied = copy_from_user(p, buf, count);
 151         if (copied) {
 152                 ssize_t ret = written + (count - copied);
 153
 154                 if (ret)
 155                         return ret;
 156                 return -EFAULT;
 157         }
 158         written += count;
 159         *ppos += written;
 160         return written;
 161 }
 162
 163
 164 /*
 165  * This funcion reads the *physical* memory. The f_pos points directly to the
 166  * memory location.
 167  */
 168 static ssize_t read_mem(struct file * file, char __user * buf,
 169                         size_t count, loff_t *ppos)
 170 {
 171         unsigned long p = *ppos;
 172         ssize_t read;
 173
 174         if (!valid_phys_addr_range(p, &count))
 175                 return -EFAULT;
 176         read = 0;
 177 #if defined(__sparc__) || (defined(__mc68000__) && defined(CONFIG_MMU))
 178         /* we don't have page 0 mapped on sparc and m68k.. */
 179         if (p < PAGE_SIZE) {
 180                 unsigned long sz = PAGE_SIZE-p;
 181                 if (sz > count)
 182                         sz = count;
 183                 if (sz > 0) {
 184                         if (clear_user(buf, sz))
 185                                 return -EFAULT;
 186                         buf += sz;
 187                         p += sz;
 188                         count -= sz;
 189                         read += sz;
 190                 }
 191         }
 192 #endif
 193         if (!range_is_allowed(p, p+count))
 194                 return -EPERM;
 195         if (copy_to_user(buf, __va(p), count))
 196                 return -EFAULT;
 197         read += count;
 198         *ppos += read;
 199         return read;
 200 }
 201
 202 static ssize_t write_mem(struct file * file, const char __user * buf,
 203                          size_t count, loff_t *ppos)
 204 {
 205         unsigned long p = *ppos;
 206
 207         if (!valid_phys_addr_range(p, &count))
 208                 return -EFAULT;
 209         return do_write_mem(__va(p), p, buf, count, ppos);
 210 }
 211
 212 static int mmap_mem(struct file * file, struct vm_area_struct * vma)
 213 {
 214         unsigned long offset = vma->vm_pgoff << PAGE_SHIFT;
 215         int uncached;
 216         unsigned long cursor;
 217
 218         uncached = uncached_access(file, offset);
 219 #ifdef pgprot_noncached
 220         if (uncached)
 221                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 222 #endif
 223
 224         /* Don't try to swap out physical pages.. */
 225         vma->vm_flags |= VM_RESERVED;
 226
 227         /*
 228          * Don't dump addresses that are not real memory to a core file.
 229          */
 230         if (uncached)
 231                 vma->vm_flags |= VM_IO;
 232
 233         cursor = vma->vm_pgoff;
 234         while ((cursor << PAGE_SHIFT) < offset + vma->vm_end-vma->vm_start) {
 235                 if (!devmem_is_allowed(cursor))
 236                         return -EPERM;
 237                 cursor++;
 238         }
 239
 240         if (remap_page_range(vma, vma->vm_start, offset, vma->vm_end-vma->vm_start,
 241                              vma->vm_page_prot))
 242                 return -EAGAIN;
 243         return 0;
 244 }
 245
 246 extern long vread(char *buf, char *addr, unsigned long count);
 247 extern long vwrite(char *buf, char *addr, unsigned long count);
 248
 249 /*
 250  * This function reads the *virtual* memory as seen by the kernel.
 251  */
 252 static ssize_t read_kmem(struct file *file, char __user *buf,
 253                          size_t count, loff_t *ppos)
 254 {
 255         unsigned long p = *ppos;
 256         ssize_t read = 0;
 257         ssize_t virtr = 0;
 258         char * kbuf; /* k-addr because vread() takes vmlist_lock rwlock */
 259
 260         return -EPERM;
 261
 262         if (p < (unsigned long) high_memory) {
 263                 read = count;
 264                 if (count > (unsigned long) high_memory - p)
 265                         read = (unsigned long) high_memory - p;
 266
 267 #if defined(__sparc__) || (defined(__mc68000__) && defined(CONFIG_MMU))
 268                 /* we don't have page 0 mapped on sparc and m68k.. */
 269                 if (p < PAGE_SIZE && read > 0) {
 270                         size_t tmp = PAGE_SIZE - p;
 271                         if (tmp > read) tmp = read;
 272                         if (clear_user(buf, tmp))
 273                                 return -EFAULT;
 274                         buf += tmp;
 275                         p += tmp;
 276                         read -= tmp;
 277                         count -= tmp;
 278                 }
 279 #endif
 280                 if (copy_to_user(buf, (char *)p, read))
 281                         return -EFAULT;
 282                 p += read;
 283                 buf += read;
 284                 count -= read;
 285         }
 286
 287         if (count > 0) {
 288                 kbuf = (char *)__get_free_page(GFP_KERNEL);
 289                 if (!kbuf)
 290                         return -ENOMEM;
 291                 while (count > 0) {
 292                         int len = count;
 293
 294                         if (len > PAGE_SIZE)
 295                                 len = PAGE_SIZE;
 296                         len = vread(kbuf, (char *)p, len);
 297                         if (!len)
 298                                 break;
 299                         if (copy_to_user(buf, kbuf, len)) {
 300                                 free_page((unsigned long)kbuf);
 301                                 return -EFAULT;
 302                         }
 303                         count -= len;
 304                         buf += len;
 305                         virtr += len;
 306                         p += len;
 307                 }
 308                 free_page((unsigned long)kbuf);
 309         }
 310         *ppos = p;
 311         return virtr + read;
 312 }
 313
 314 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 315 static ssize_t read_port(struct file * file, char __user * buf,
 316                          size_t count, loff_t *ppos)
 317 {
 318         unsigned long i = *ppos;
 319         char __user *tmp = buf;
 320
 321         if (verify_area(VERIFY_WRITE,buf,count))
 322                 return -EFAULT;
 323         while (count-- > 0 && i < 65536) {
 324                 if (__put_user(inb(i),tmp) < 0)
 325                         return -EFAULT;
 326                 i++;
 327                 tmp++;
 328         }
 329         *ppos = i;
 330         return tmp-buf;
 331 }
 332
 333 static ssize_t write_port(struct file * file, const char __user * buf,
 334                           size_t count, loff_t *ppos)
 335 {
 336         unsigned long i = *ppos;
 337         const char __user * tmp = buf;
 338
 339         if (verify_area(VERIFY_READ,buf,count))
 340                 return -EFAULT;
 341         while (count-- > 0 && i < 65536) {
 342                 char c;
 343                 if (__get_user(c, tmp))
 344                         return -EFAULT;
 345                 outb(c,i);
 346                 i++;
 347                 tmp++;
 348         }
 349         *ppos = i;
 350         return tmp-buf;
 351 }
 352 #endif
 353
 354 static ssize_t read_null(struct file * file, char __user * buf,
 355                          size_t count, loff_t *ppos)
 356 {
 357         return 0;
 358 }
 359
 360 static ssize_t write_null(struct file * file, const char __user * buf,
 361                           size_t count, loff_t *ppos)
 362 {
 363         return count;
 364 }
 365
 366 #ifdef CONFIG_MMU
 367 /*
 368  * For fun, we are using the MMU for this.
 369  */
 370 static inline size_t read_zero_pagealigned(char __user * buf, size_t size)
 371 {
 372         struct mm_struct *mm;
 373         struct vm_area_struct * vma;
 374         unsigned long addr=(unsigned long)buf;
 375
 376         mm = current->mm;
 377         /* Oops, this was forgotten before. -ben */
 378         down_read(&mm->mmap_sem);
 379
 380         /* For private mappings, just map in zero pages. */
 381         for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
 382                 unsigned long count;
 383
 384                 if (vma->vm_start > addr || (vma->vm_flags & VM_WRITE) == 0)
 385                         goto out_up;
 386                 if (vma->vm_flags & VM_SHARED)
 387                         break;
 388                 count = vma->vm_end - addr;
 389                 if (count > size)
 390                         count = size;
 391
 392                 zap_page_range(vma, addr, count, NULL);
 393                 zeromap_page_range(vma, addr, count, PAGE_COPY);
 394
 395                 size -= count;
 396                 buf += count;
 397                 addr += count;
 398                 if (size == 0)
 399                         goto out_up;
 400         }
 401
 402         up_read(&mm->mmap_sem);
 403
 404         /* The shared case is hard. Let's do the conventional zeroing. */
 405         do {
 406                 unsigned long unwritten = clear_user(buf, PAGE_SIZE);
 407                 if (unwritten)
 408                         return size + unwritten - PAGE_SIZE;
 409                 cond_resched();
 410                 buf += PAGE_SIZE;
 411                 size -= PAGE_SIZE;
 412         } while (size);
 413
 414         return size;
 415 out_up:
 416         up_read(&mm->mmap_sem);
 417         return size;
 418 }
 419
 420 static ssize_t read_zero(struct file * file, char __user * buf,
 421                          size_t count, loff_t *ppos)
 422 {
 423         unsigned long left, unwritten, written = 0;
 424
 425         if (!count)
 426                 return 0;
 427
 428         if (!access_ok(VERIFY_WRITE, buf, count))
 429                 return -EFAULT;
 430
 431         left = count;
 432
 433         /* do we want to be clever? Arbitrary cut-off */
 434         if (count >= PAGE_SIZE*4) {
 435                 unsigned long partial;
 436
 437                 /* How much left of the page? */
 438                 partial = (PAGE_SIZE-1) & -(unsigned long) buf;
 439                 unwritten = clear_user(buf, partial);
 440                 written = partial - unwritten;
 441                 if (unwritten)
 442                         goto out;
 443                 left -= partial;
 444                 buf += partial;
 445                 unwritten = read_zero_pagealigned(buf, left & PAGE_MASK);
 446                 written += (left & PAGE_MASK) - unwritten;
 447                 if (unwritten)
 448                         goto out;
 449                 buf += left & PAGE_MASK;
 450                 left &= ~PAGE_MASK;
 451         }
 452         unwritten = clear_user(buf, left);
 453         written += left - unwritten;
 454 out:
 455         return written ? written : -EFAULT;
 456 }
 457
 458 static int mmap_zero(struct file * file, struct vm_area_struct * vma)
 459 {
 460         if (vma->vm_flags & VM_SHARED)
 461                 return shmem_zero_setup(vma);
 462         if (zeromap_page_range(vma, vma->vm_start, vma->vm_end - vma->vm_start, vma->vm_page_prot))
 463                 return -EAGAIN;
 464         return 0;
 465 }
 466 #else /* CONFIG_MMU */
 467 static ssize_t read_zero(struct file * file, char * buf,
 468                          size_t count, loff_t *ppos)
 469 {
 470         size_t todo = count;
 471
 472         while (todo) {
 473                 size_t chunk = todo;
 474
 475                 if (chunk > 4096)
 476                         chunk = 4096;   /* Just for latency reasons */
 477                 if (clear_user(buf, chunk))
 478                         return -EFAULT;
 479                 buf += chunk;
 480                 todo -= chunk;
 481                 cond_resched();
 482         }
 483         return count;
 484 }
 485
 486 static int mmap_zero(struct file * file, struct vm_area_struct * vma)
 487 {
 488         return -ENOSYS;
 489 }
 490 #endif /* CONFIG_MMU */
 491
 492 static ssize_t write_full(struct file * file, const char __user * buf,
 493                           size_t count, loff_t *ppos)
 494 {
 495         return -ENOSPC;
 496 }
 497
 498 /*
 499  * Special lseek() function for /dev/null and /dev/zero.  Most notably, you
 500  * can fopen() both devices with "a" now.  This was previously impossible.
 501  * -- SRB.
 502  */
 503
 504 static loff_t null_lseek(struct file * file, loff_t offset, int orig)
 505 {
 506         return file->f_pos = 0;
 507 }
 508
 509 /*
 510  * The memory devices use the full 32/64 bits of the offset, and so we cannot
 511  * check against negative addresses: they are ok. The return value is weird,
 512  * though, in that case (0).
 513  *
 514  * also note that seeking relative to the "end of file" isn't supported:
 515  * it has no meaning, so it returns -EINVAL.
 516  */
 517 static loff_t memory_lseek(struct file * file, loff_t offset, int orig)
 518 {
 519         loff_t ret;
 520
 521         down(&file->f_dentry->d_inode->i_sem);
 522         switch (orig) {
 523                 case 0:
 524                         file->f_pos = offset;
 525                         ret = file->f_pos;
 526                         force_successful_syscall_return();
 527                         break;
 528                 case 1:
 529                         file->f_pos += offset;
 530                         ret = file->f_pos;
 531                         force_successful_syscall_return();
 532                         break;
 533                 default:
 534                         ret = -EINVAL;
 535         }
 536         up(&file->f_dentry->d_inode->i_sem);
 537         return ret;
 538 }
 539
 540 static int open_port(struct inode * inode, struct file * filp)
 541 {
 542         return capable(CAP_SYS_RAWIO) ? 0 : -EPERM;
 543 }
 544
 545 #define mmap_kmem       mmap_mem
 546 #define zero_lseek      null_lseek
 547 #define full_lseek      null_lseek
 548 #define write_zero      write_null
 549 #define read_full       read_zero
 550 #define open_mem        open_port
 551 #define open_kmem       open_mem
 552
 553 static struct file_operations mem_fops = {
 554         .llseek         = memory_lseek,
 555         .read           = read_mem,
 556         .write          = write_mem,
 557         .mmap           = mmap_mem,
 558         .open           = open_mem,
 559 };
 560
 561 static struct file_operations kmem_fops = {
 562         .llseek         = memory_lseek,
 563         .read           = read_kmem,
 564         .mmap           = mmap_kmem,
 565         .open           = open_kmem,
 566 };
 567
 568 static struct file_operations null_fops = {
 569         .llseek         = null_lseek,
 570         .read           = read_null,
 571         .write          = write_null,
 572 };
 573
 574 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 575 static struct file_operations port_fops = {
 576         .llseek         = memory_lseek,
 577         .read           = read_port,
 578         .write          = write_port,
 579         .open           = open_port,
 580 };
 581 #endif
 582
 583 static struct file_operations zero_fops = {
 584         .llseek         = zero_lseek,
 585         .read           = read_zero,
 586         .write          = write_zero,
 587         .mmap           = mmap_zero,
 588 };
 589
 590 static struct file_operations full_fops = {
 591         .llseek         = full_lseek,
 592         .read           = read_full,
 593         .write          = write_full,
 594 };
 595
 596 static ssize_t kmsg_write(struct file * file, const char __user * buf,
 597                           size_t count, loff_t *ppos)
 598 {
 599         char *tmp;
 600         int ret;
 601
 602         tmp = kmalloc(count + 1, GFP_KERNEL);
 603         if (tmp == NULL)
 604                 return -ENOMEM;
 605         ret = -EFAULT;
 606         if (!copy_from_user(tmp, buf, count)) {
 607                 tmp[count] = 0;
 608                 ret = printk("%s", tmp);
 609         }
 610         kfree(tmp);
 611         return ret;
 612 }
 613
 614 static struct file_operations kmsg_fops = {
 615         .write =        kmsg_write,
 616 };
 617
 618 static int memory_open(struct inode * inode, struct file * filp)
 619 {
 620         switch (iminor(inode)) {
 621                 case 1:
 622                         filp->f_op = &mem_fops;
 623                         break;
 624                 case 2:
 625                         filp->f_op = &kmem_fops;
 626                         break;
 627                 case 3:
 628                         filp->f_op = &null_fops;
 629                         break;
 630 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 631                 case 4:
 632                         filp->f_op = &port_fops;
 633                         break;
 634 #endif
 635                 case 5:
 636                         filp->f_op = &zero_fops;
 637                         break;
 638                 case 7:
 639                         filp->f_op = &full_fops;
 640                         break;
 641                 case 8:
 642                         filp->f_op = &random_fops;
 643                         break;
 644                 case 9:
 645                         filp->f_op = &urandom_fops;
 646                         break;
 647                 case 11:
 648                         filp->f_op = &kmsg_fops;
 649                         break;
 650                 default:
 651                         return -ENXIO;
 652         }
 653         if (filp->f_op && filp->f_op->open)
 654                 return filp->f_op->open(inode,filp);
 655         return 0;
 656 }
 657
 658 static struct file_operations memory_fops = {
 659         .open           = memory_open,  /* just a selector for the real open */
 660 };
 661
 662 static const struct {
 663         unsigned int            minor;
 664         char                    *name;
 665         umode_t                 mode;
 666         struct file_operations  *fops;
 667 } devlist[] = { /* list of minor devices */
 668         {1, "mem",     S_IRUSR | S_IWUSR | S_IRGRP, &mem_fops},
 669         {2, "kmem",    S_IRUSR | S_IWUSR | S_IRGRP, &kmem_fops},
 670         {3, "null",    S_IRUGO | S_IWUGO,           &null_fops},
 671 #if defined(CONFIG_ISA) || !defined(__mc68000__)
 672         {4, "port",    S_IRUSR | S_IWUSR | S_IRGRP, &port_fops},
 673 #endif
 674         {5, "zero",    S_IRUGO | S_IWUGO,           &zero_fops},
 675         {7, "full",    S_IRUGO | S_IWUGO,           &full_fops},
 676         {8, "random",  S_IRUGO | S_IWUSR,           &random_fops},
 677         {9, "urandom", S_IRUGO | S_IWUSR,           &urandom_fops},
 678         {11,"kmsg",    S_IRUGO | S_IWUSR,           &kmsg_fops},
 679 };
 680
 681 static struct class_simple *mem_class;
 682
 683 static int __init chr_dev_init(void)
 684 {
 685         int i;
 686
 687         if (register_chrdev(MEM_MAJOR,"mem",&memory_fops))
 688                 printk("unable to get major %d for memory devs\n", MEM_MAJOR);
 689
 690         mem_class = class_simple_create(THIS_MODULE, "mem");
 691         for (i = 0; i < ARRAY_SIZE(devlist); i++) {
 692                 class_simple_device_add(mem_class,
 693                                         MKDEV(MEM_MAJOR, devlist[i].minor),
 694                                         NULL, devlist[i].name);
 695                 devfs_mk_cdev(MKDEV(MEM_MAJOR, devlist[i].minor),
 696                                 S_IFCHR | devlist[i].mode, devlist[i].name);
 697         }
 698
 699 #if defined (CONFIG_FB)
 700         fbmem_init();
 701 #endif
 702         return 0;
 703 }
 704
 705 fs_initcall(chr_dev_init);