include/asm-i386/system.h

   1 #ifndef __ASM_SYSTEM_H
   2 #define __ASM_SYSTEM_H
   3
   4 #include <linux/kernel.h>
   5 #include <asm/segment.h>
   6 #include <asm/cpufeature.h>
   7 #include <linux/bitops.h> /* for LOCK_PREFIX */
   8
   9 #ifdef __KERNEL__
  10
  11 struct task_struct;     /* one of the stranger aspects of C forward declarations.. */
  12 extern struct task_struct * FASTCALL(__switch_to(struct task_struct *prev, struct task_struct *next));
  13
  14 /*
  15  * Saving eflags is important. It switches not only IOPL between tasks,
  16  * it also protects other tasks from NT leaking through sysenter etc.
  17  */
  18 #define switch_to(prev,next,last) do {                                  \
  19         unsigned long esi,edi;                                          \
  20         asm volatile("pushfl\n\t"               /* Save flags */        \
  21                      "pushl %%ebp\n\t"                                  \
  22                      "movl %%esp,%0\n\t"        /* save ESP */          \
  23                      "movl %5,%%esp\n\t"        /* restore ESP */       \
  24                      "movl $1f,%1\n\t"          /* save EIP */          \
  25                      "pushl %6\n\t"             /* restore EIP */       \
  26                      "jmp __switch_to\n"                                \
  27                      "1:\t"                                             \
  28                      "popl %%ebp\n\t"                                   \
  29                      "popfl"                                            \
  30                      :"=m" (prev->thread.esp),"=m" (prev->thread.eip),  \
  31                       "=a" (last),"=S" (esi),"=D" (edi)                 \
  32                      :"m" (next->thread.esp),"m" (next->thread.eip),    \
  33                       "2" (prev), "d" (next));                          \
  34 } while (0)
  35
  36 #define _set_base(addr,base) do { unsigned long __pr; \
  37 __asm__ __volatile__ ("movw %%dx,%1\n\t" \
  38         "rorl $16,%%edx\n\t" \
  39         "movb %%dl,%2\n\t" \
  40         "movb %%dh,%3" \
  41         :"=&d" (__pr) \
  42         :"m" (*((addr)+2)), \
  43          "m" (*((addr)+4)), \
  44          "m" (*((addr)+7)), \
  45          "0" (base) \
  46         ); } while(0)
  47
  48 #define _set_limit(addr,limit) do { unsigned long __lr; \
  49 __asm__ __volatile__ ("movw %%dx,%1\n\t" \
  50         "rorl $16,%%edx\n\t" \
  51         "movb %2,%%dh\n\t" \
  52         "andb $0xf0,%%dh\n\t" \
  53         "orb %%dh,%%dl\n\t" \
  54         "movb %%dl,%2" \
  55         :"=&d" (__lr) \
  56         :"m" (*(addr)), \
  57          "m" (*((addr)+6)), \
  58          "0" (limit) \
  59         ); } while(0)
  60
  61 #define set_base(ldt,base) _set_base( ((char *)&(ldt)) , (base) )
  62 #define set_limit(ldt,limit) _set_limit( ((char *)&(ldt)) , ((limit)-1) )
  63
  64 /*
  65  * Load a segment. Fall back on loading the zero
  66  * segment if something goes wrong..
  67  */
  68 #define loadsegment(seg,value)                  \
  69         asm volatile("\n"                       \
  70                 "1:\t"                          \
  71                 "mov %0,%%" #seg "\n"           \
  72                 "2:\n"                          \
  73                 ".section .fixup,\"ax\"\n"      \
  74                 "3:\t"                          \
  75                 "pushl $0\n\t"                  \
  76                 "popl %%" #seg "\n\t"           \
  77                 "jmp 2b\n"                      \
  78                 ".previous\n"                   \
  79                 ".section __ex_table,\"a\"\n\t" \
  80                 ".align 4\n\t"                  \
  81                 ".long 1b,3b\n"                 \
  82                 ".previous"                     \
  83                 : :"rm" (value))
  84
  85 /*
  86  * Save a segment register away
  87  */
  88 #define savesegment(seg, value) \
  89         asm volatile("mov %%" #seg ",%0":"=rm" (value))
  90
  91 #define read_cr0() ({ \
  92         unsigned int __dummy; \
  93         __asm__ __volatile__( \
  94                 "movl %%cr0,%0\n\t" \
  95                 :"=r" (__dummy)); \
  96         __dummy; \
  97 })
  98 #define write_cr0(x) \
  99         __asm__ __volatile__("movl %0,%%cr0": :"r" (x))
 100
 101 #define read_cr2() ({ \
 102         unsigned int __dummy; \
 103         __asm__ __volatile__( \
 104                 "movl %%cr2,%0\n\t" \
 105                 :"=r" (__dummy)); \
 106         __dummy; \
 107 })
 108 #define write_cr2(x) \
 109         __asm__ __volatile__("movl %0,%%cr2": :"r" (x))
 110
 111 #define read_cr3() ({ \
 112         unsigned int __dummy; \
 113         __asm__ ( \
 114                 "movl %%cr3,%0\n\t" \
 115                 :"=r" (__dummy)); \
 116         __dummy; \
 117 })
 118 #define write_cr3(x) \
 119         __asm__ __volatile__("movl %0,%%cr3": :"r" (x))
 120
 121 #define read_cr4() ({ \
 122         unsigned int __dummy; \
 123         __asm__( \
 124                 "movl %%cr4,%0\n\t" \
 125                 :"=r" (__dummy)); \
 126         __dummy; \
 127 })
 128 #define read_cr4_safe() ({                            \
 129         unsigned int __dummy;                         \
 130         /* This could fault if %cr4 does not exist */ \
 131         __asm__("1: movl %%cr4, %0              \n"   \
 132                 "2:                             \n"   \
 133                 ".section __ex_table,\"a\"      \n"   \
 134                 ".long 1b,2b                    \n"   \
 135                 ".previous                      \n"   \
 136                 : "=r" (__dummy): "0" (0));           \
 137         __dummy;                                      \
 138 })
 139 #define write_cr4(x) \
 140         __asm__ __volatile__("movl %0,%%cr4": :"r" (x))
 141
 142 /*
 143  * Clear and set 'TS' bit respectively
 144  */
 145 #define clts() __asm__ __volatile__ ("clts")
 146 #define stts() write_cr0(8 | read_cr0())
 147
 148 #endif  /* __KERNEL__ */
 149
 150 #define wbinvd() \
 151         __asm__ __volatile__ ("wbinvd": : :"memory")
 152
 153 static inline unsigned long get_limit(unsigned long segment)
 154 {
 155         unsigned long __limit;
 156         __asm__("lsll %1,%0"
 157                 :"=r" (__limit):"r" (segment));
 158         return __limit+1;
 159 }
 160
 161 #define nop() __asm__ __volatile__ ("nop")
 162
 163 #define xchg(ptr,v) ((__typeof__(*(ptr)))__xchg((unsigned long)(v),(ptr),sizeof(*(ptr))))
 164
 165 #define tas(ptr) (xchg((ptr),1))
 166
 167 struct __xchg_dummy { unsigned long a[100]; };
 168 #define __xg(x) ((struct __xchg_dummy *)(x))
 169
 170
 171 #ifdef CONFIG_X86_CMPXCHG64
 172
 173 /*
 174  * The semantics of XCHGCMP8B are a bit strange, this is why
 175  * there is a loop and the loading of %%eax and %%edx has to
 176  * be inside. This inlines well in most cases, the cached
 177  * cost is around ~38 cycles. (in the future we might want
 178  * to do an SIMD/3DNOW!/MMX/FPU 64-bit store here, but that
 179  * might have an implicit FPU-save as a cost, so it's not
 180  * clear which path to go.)
 181  *
 182  * cmpxchg8b must be used with the lock prefix here to allow
 183  * the instruction to be executed atomically, see page 3-102
 184  * of the instruction set reference 24319102.pdf. We need
 185  * the reader side to see the coherent 64bit value.
 186  */
 187 static inline void __set_64bit (unsigned long long * ptr,
 188                 unsigned int low, unsigned int high)
 189 {
 190         __asm__ __volatile__ (
 191                 "\n1:\t"
 192                 "movl (%0), %%eax\n\t"
 193                 "movl 4(%0), %%edx\n\t"
 194                 "lock cmpxchg8b (%0)\n\t"
 195                 "jnz 1b"
 196                 : /* no outputs */
 197                 :       "D"(ptr),
 198                         "b"(low),
 199                         "c"(high)
 200                 :       "ax","dx","memory");
 201 }
 202
 203 static inline void __set_64bit_constant (unsigned long long *ptr,
 204                                                  unsigned long long value)
 205 {
 206         __set_64bit(ptr,(unsigned int)(value), (unsigned int)((value)>>32ULL));
 207 }
 208 #define ll_low(x)       *(((unsigned int*)&(x))+0)
 209 #define ll_high(x)      *(((unsigned int*)&(x))+1)
 210
 211 static inline void __set_64bit_var (unsigned long long *ptr,
 212                          unsigned long long value)
 213 {
 214         __set_64bit(ptr,ll_low(value), ll_high(value));
 215 }
 216
 217 #define set_64bit(ptr,value) \
 218 (__builtin_constant_p(value) ? \
 219  __set_64bit_constant(ptr, value) : \
 220  __set_64bit_var(ptr, value) )
 221
 222 #define _set_64bit(ptr,value) \
 223 (__builtin_constant_p(value) ? \
 224  __set_64bit(ptr, (unsigned int)(value), (unsigned int)((value)>>32ULL) ) : \
 225  __set_64bit(ptr, ll_low(value), ll_high(value)) )
 226
 227 #endif
 228
 229 /*
 230  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
 231  * Note 2: xchg has side effect, so that attribute volatile is necessary,
 232  *        but generally the primitive is invalid, *ptr is output argument. --ANK
 233  */
 234 static inline unsigned long __xchg(unsigned long x, volatile void * ptr, int size)
 235 {
 236         switch (size) {
 237                 case 1:
 238                         __asm__ __volatile__("xchgb %b0,%1"
 239                                 :"=q" (x)
 240                                 :"m" (*__xg(ptr)), "0" (x)
 241                                 :"memory");
 242                         break;
 243                 case 2:
 244                         __asm__ __volatile__("xchgw %w0,%1"
 245                                 :"=r" (x)
 246                                 :"m" (*__xg(ptr)), "0" (x)
 247                                 :"memory");
 248                         break;
 249                 case 4:
 250                         __asm__ __volatile__("xchgl %0,%1"
 251                                 :"=r" (x)
 252                                 :"m" (*__xg(ptr)), "0" (x)
 253                                 :"memory");
 254                         break;
 255         }
 256         return x;
 257 }
 258
 259 /*
 260  * Atomic compare and exchange.  Compare OLD with MEM, if identical,
 261  * store NEW in MEM.  Return the initial value in MEM.  Success is
 262  * indicated by comparing RETURN with OLD.
 263  */
 264
 265 #ifdef CONFIG_X86_CMPXCHG
 266 #define __HAVE_ARCH_CMPXCHG 1
 267 #define cmpxchg(ptr,o,n)\
 268         ((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
 269                                         (unsigned long)(n),sizeof(*(ptr))))
 270 #define sync_cmpxchg(ptr,o,n)\
 271         ((__typeof__(*(ptr)))__sync_cmpxchg((ptr),(unsigned long)(o),\
 272                                         (unsigned long)(n),sizeof(*(ptr))))
 273 #endif
 274
 275 static inline unsigned long __cmpxchg(volatile void *ptr, unsigned long old,
 276                                       unsigned long new, int size)
 277 {
 278         unsigned long prev;
 279         switch (size) {
 280         case 1:
 281                 __asm__ __volatile__(LOCK_PREFIX "cmpxchgb %b1,%2"
 282                                      : "=a"(prev)
 283                                      : "q"(new), "m"(*__xg(ptr)), "0"(old)
 284                                      : "memory");
 285                 return prev;
 286         case 2:
 287                 __asm__ __volatile__(LOCK_PREFIX "cmpxchgw %w1,%2"
 288                                      : "=a"(prev)
 289                                      : "r"(new), "m"(*__xg(ptr)), "0"(old)
 290                                      : "memory");
 291                 return prev;
 292         case 4:
 293                 __asm__ __volatile__(LOCK_PREFIX "cmpxchgl %1,%2"
 294                                      : "=a"(prev)
 295                                      : "r"(new), "m"(*__xg(ptr)), "0"(old)
 296                                      : "memory");
 297                 return prev;
 298         }
 299         return old;
 300 }
 301
 302 /*
 303  * Always use locked operations when touching memory shared with a
 304  * hypervisor, since the system may be SMP even if the guest kernel
 305  * isn't.
 306  */
 307 static inline unsigned long __sync_cmpxchg(volatile void *ptr,
 308                                             unsigned long old,
 309                                             unsigned long new, int size)
 310 {
 311         unsigned long prev;
 312         switch (size) {
 313         case 1:
 314                 __asm__ __volatile__("lock; cmpxchgb %b1,%2"
 315                                      : "=a"(prev)
 316                                      : "q"(new), "m"(*__xg(ptr)), "0"(old)
 317                                      : "memory");
 318                 return prev;
 319         case 2:
 320                 __asm__ __volatile__("lock; cmpxchgw %w1,%2"
 321                                      : "=a"(prev)
 322                                      : "r"(new), "m"(*__xg(ptr)), "0"(old)
 323                                      : "memory");
 324                 return prev;
 325         case 4:
 326                 __asm__ __volatile__("lock; cmpxchgl %1,%2"
 327                                      : "=a"(prev)
 328                                      : "r"(new), "m"(*__xg(ptr)), "0"(old)
 329                                      : "memory");
 330                 return prev;
 331         }
 332         return old;
 333 }
 334
 335 #ifndef CONFIG_X86_CMPXCHG
 336 /*
 337  * Building a kernel capable running on 80386. It may be necessary to
 338  * simulate the cmpxchg on the 80386 CPU. For that purpose we define
 339  * a function for each of the sizes we support.
 340  */
 341
 342 extern unsigned long cmpxchg_386_u8(volatile void *, u8, u8);
 343 extern unsigned long cmpxchg_386_u16(volatile void *, u16, u16);
 344 extern unsigned long cmpxchg_386_u32(volatile void *, u32, u32);
 345
 346 static inline unsigned long cmpxchg_386(volatile void *ptr, unsigned long old,
 347                                       unsigned long new, int size)
 348 {
 349         switch (size) {
 350         case 1:
 351                 return cmpxchg_386_u8(ptr, old, new);
 352         case 2:
 353                 return cmpxchg_386_u16(ptr, old, new);
 354         case 4:
 355                 return cmpxchg_386_u32(ptr, old, new);
 356         }
 357         return old;
 358 }
 359
 360 #define cmpxchg(ptr,o,n)                                                \
 361 ({                                                                      \
 362         __typeof__(*(ptr)) __ret;                                       \
 363         if (likely(boot_cpu_data.x86 > 3))                              \
 364                 __ret = __cmpxchg((ptr), (unsigned long)(o),            \
 365                                         (unsigned long)(n), sizeof(*(ptr))); \
 366         else                                                            \
 367                 __ret = cmpxchg_386((ptr), (unsigned long)(o),          \
 368                                         (unsigned long)(n), sizeof(*(ptr))); \
 369         __ret;                                                          \
 370 })
 371 #endif
 372
 373 #ifdef CONFIG_X86_CMPXCHG64
 374
 375 static inline unsigned long long __cmpxchg64(volatile void *ptr, unsigned long long old,
 376                                       unsigned long long new)
 377 {
 378         unsigned long long prev;
 379         __asm__ __volatile__(LOCK_PREFIX "cmpxchg8b %3"
 380                              : "=A"(prev)
 381                              : "b"((unsigned long)new),
 382                                "c"((unsigned long)(new >> 32)),
 383                                "m"(*__xg(ptr)),
 384                                "0"(old)
 385                              : "memory");
 386         return prev;
 387 }
 388
 389 #define cmpxchg64(ptr,o,n)\
 390         ((__typeof__(*(ptr)))__cmpxchg64((ptr),(unsigned long long)(o),\
 391                                         (unsigned long long)(n)))
 392
 393 #endif
 394
 395 /*
 396  * Force strict CPU ordering.
 397  * And yes, this is required on UP too when we're talking
 398  * to devices.
 399  *
 400  * For now, "wmb()" doesn't actually do anything, as all
 401  * Intel CPU's follow what Intel calls a *Processor Order*,
 402  * in which all writes are seen in the program order even
 403  * outside the CPU.
 404  *
 405  * I expect future Intel CPU's to have a weaker ordering,
 406  * but I'd also expect them to finally get their act together
 407  * and add some real memory barriers if so.
 408  *
 409  * Some non intel clones support out of order store. wmb() ceases to be a
 410  * nop for these.
 411  */
 412
 413
 414 /*
 415  * Actually only lfence would be needed for mb() because all stores done
 416  * by the kernel should be already ordered. But keep a full barrier for now.
 417  */
 418
 419 #define mb() alternative("lock; addl $0,0(%%esp)", "mfence", X86_FEATURE_XMM2)
 420 #define rmb() alternative("lock; addl $0,0(%%esp)", "lfence", X86_FEATURE_XMM2)
 421
 422 /**
 423  * read_barrier_depends - Flush all pending reads that subsequents reads
 424  * depend on.
 425  *
 426  * No data-dependent reads from memory-like regions are ever reordered
 427  * over this barrier.  All reads preceding this primitive are guaranteed
 428  * to access memory (but not necessarily other CPUs' caches) before any
 429  * reads following this primitive that depend on the data return by
 430  * any of the preceding reads.  This primitive is much lighter weight than
 431  * rmb() on most CPUs, and is never heavier weight than is
 432  * rmb().
 433  *
 434  * These ordering constraints are respected by both the local CPU
 435  * and the compiler.
 436  *
 437  * Ordering is not guaranteed by anything other than these primitives,
 438  * not even by data dependencies.  See the documentation for
 439  * memory_barrier() for examples and URLs to more information.
 440  *
 441  * For example, the following code would force ordering (the initial
 442  * value of "a" is zero, "b" is one, and "p" is "&a"):
 443  *
 444  * <programlisting>
 445  *      CPU 0                           CPU 1
 446  *
 447  *      b = 2;
 448  *      memory_barrier();
 449  *      p = &b;                         q = p;
 450  *                                      read_barrier_depends();
 451  *                                      d = *q;
 452  * </programlisting>
 453  *
 454  * because the read of "*q" depends on the read of "p" and these
 455  * two reads are separated by a read_barrier_depends().  However,
 456  * the following code, with the same initial values for "a" and "b":
 457  *
 458  * <programlisting>
 459  *      CPU 0                           CPU 1
 460  *
 461  *      a = 2;
 462  *      memory_barrier();
 463  *      b = 3;                          y = b;
 464  *                                      read_barrier_depends();
 465  *                                      x = a;
 466  * </programlisting>
 467  *
 468  * does not enforce ordering, since there is no data dependency between
 469  * the read of "a" and the read of "b".  Therefore, on some CPUs, such
 470  * as Alpha, "y" could be set to 3 and "x" to 0.  Use rmb()
 471  * in cases like this where there are no data dependencies.
 472  **/
 473
 474 #define read_barrier_depends()  do { } while(0)
 475
 476 #ifdef CONFIG_X86_OOSTORE
 477 /* Actually there are no OOO store capable CPUs for now that do SSE,
 478    but make it already an possibility. */
 479 #define wmb() alternative("lock; addl $0,0(%%esp)", "sfence", X86_FEATURE_XMM)
 480 #else
 481 #define wmb()   __asm__ __volatile__ ("": : :"memory")
 482 #endif
 483
 484 #ifdef CONFIG_SMP
 485 #define smp_mb()        mb()
 486 #define smp_rmb()       rmb()
 487 #define smp_wmb()       wmb()
 488 #define smp_read_barrier_depends()      read_barrier_depends()
 489 #define set_mb(var, value) do { (void) xchg(&var, value); } while (0)
 490 #else
 491 #define smp_mb()        barrier()
 492 #define smp_rmb()       barrier()
 493 #define smp_wmb()       barrier()
 494 #define smp_read_barrier_depends()      do { } while(0)
 495 #define set_mb(var, value) do { var = value; barrier(); } while (0)
 496 #endif
 497
 498 #include <linux/irqflags.h>
 499
 500 /*
 501  * disable hlt during certain critical i/o operations
 502  */
 503 #define HAVE_DISABLE_HLT
 504 void disable_hlt(void);
 505 void enable_hlt(void);
 506
 507 extern int es7000_plat;
 508 void cpu_idle_wait(void);
 509
 510 /*
 511  * On SMP systems, when the scheduler does migration-cost autodetection,
 512  * it needs a way to flush as much of the CPU's caches as possible:
 513  */
 514 static inline void sched_cacheflush(void)
 515 {
 516         wbinvd();
 517 }
 518
 519 extern unsigned long arch_align_stack(unsigned long sp);
 520 extern void free_init_pages(char *what, unsigned long begin, unsigned long end);
 521
 522 void default_idle(void);
 523
 524 #endif