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