arch/x86/include/asm/system.h

   1 #ifndef _ASM_X86_SYSTEM_H
   2 #define _ASM_X86_SYSTEM_H
   3
   4 #include <asm/asm.h>
   5 #include <asm/segment.h>
   6 #include <asm/cpufeature.h>
   7 #include <asm/cmpxchg.h>
   8 #include <asm/nops.h>
   9
  10 #include <linux/kernel.h>
  11 #include <linux/irqflags.h>
  12
  13 /* entries in ARCH_DLINFO: */
  14 #if defined(CONFIG_IA32_EMULATION) || !defined(CONFIG_X86_64)
  15 # define AT_VECTOR_SIZE_ARCH 2
  16 #else /* else it's non-compat x86-64 */
  17 # define AT_VECTOR_SIZE_ARCH 1
  18 #endif
  19
  20 struct task_struct; /* one of the stranger aspects of C forward declarations */
  21 struct task_struct *__switch_to(struct task_struct *prev,
  22                                 struct task_struct *next);
  23 struct tss_struct;
  24 void __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
  25                       struct tss_struct *tss);
  26 extern void show_regs_common(void);
  27
  28 #ifdef CONFIG_X86_32
  29
  30 #ifdef CONFIG_CC_STACKPROTECTOR
  31 #define __switch_canary                                                 \
  32         "movl %P[task_canary](%[next]), %%ebx\n\t"                      \
  33         "movl %%ebx, "__percpu_arg([stack_canary])"\n\t"
  34 #define __switch_canary_oparam                                          \
  35         , [stack_canary] "=m" (stack_canary.canary)
  36 #define __switch_canary_iparam                                          \
  37         , [task_canary] "i" (offsetof(struct task_struct, stack_canary))
  38 #else   /* CC_STACKPROTECTOR */
  39 #define __switch_canary
  40 #define __switch_canary_oparam
  41 #define __switch_canary_iparam
  42 #endif  /* CC_STACKPROTECTOR */
  43
  44 /*
  45  * Saving eflags is important. It switches not only IOPL between tasks,
  46  * it also protects other tasks from NT leaking through sysenter etc.
  47  */
  48 #define switch_to(prev, next, last)                                     \
  49 do {                                                                    \
  50         /*                                                              \
  51          * Context-switching clobbers all registers, so we clobber      \
  52          * them explicitly, via unused output variables.                \
  53          * (EAX and EBP is not listed because EBP is saved/restored     \
  54          * explicitly for wchan access and EAX is the return value of   \
  55          * __switch_to())                                               \
  56          */                                                             \
  57         unsigned long ebx, ecx, edx, esi, edi;                          \
  58                                                                         \
  59         asm volatile("pushfl\n\t"               /* save    flags */     \
  60                      "pushl %%ebp\n\t"          /* save    EBP   */     \
  61                      "movl %%esp,%[prev_sp]\n\t"        /* save    ESP   */ \
  62                      "movl %[next_sp],%%esp\n\t"        /* restore ESP   */ \
  63                      "movl $1f,%[prev_ip]\n\t"  /* save    EIP   */     \
  64                      "pushl %[next_ip]\n\t"     /* restore EIP   */     \
  65                      __switch_canary                                    \
  66                      "jmp __switch_to\n"        /* regparm call  */     \
  67                      "1:\t"                                             \
  68                      "popl %%ebp\n\t"           /* restore EBP   */     \
  69                      "popfl\n"                  /* restore flags */     \
  70                                                                         \
  71                      /* output parameters */                            \
  72                      : [prev_sp] "=m" (prev->thread.sp),                \
  73                        [prev_ip] "=m" (prev->thread.ip),                \
  74                        "=a" (last),                                     \
  75                                                                         \
  76                        /* clobbered output registers: */                \
  77                        "=b" (ebx), "=c" (ecx), "=d" (edx),              \
  78                        "=S" (esi), "=D" (edi)                           \
  79                                                                         \
  80                        __switch_canary_oparam                           \
  81                                                                         \
  82                        /* input parameters: */                          \
  83                      : [next_sp]  "m" (next->thread.sp),                \
  84                        [next_ip]  "m" (next->thread.ip),                \
  85                                                                         \
  86                        /* regparm parameters for __switch_to(): */      \
  87                        [prev]     "a" (prev),                           \
  88                        [next]     "d" (next)                            \
  89                                                                         \
  90                        __switch_canary_iparam                           \
  91                                                                         \
  92                      : /* reloaded segment registers */                 \
  93                         "memory");                                      \
  94 } while (0)
  95
  96 /*
  97  * disable hlt during certain critical i/o operations
  98  */
  99 #define HAVE_DISABLE_HLT
 100 #else
 101
 102 /* frame pointer must be last for get_wchan */
 103 #define SAVE_CONTEXT    "pushf ; pushq %%rbp ; movq %%rsi,%%rbp\n\t"
 104 #define RESTORE_CONTEXT "movq %%rbp,%%rsi ; popq %%rbp ; popf\t"
 105
 106 #define __EXTRA_CLOBBER  \
 107         , "rcx", "rbx", "rdx", "r8", "r9", "r10", "r11", \
 108           "r12", "r13", "r14", "r15"
 109
 110 #ifdef CONFIG_CC_STACKPROTECTOR
 111 #define __switch_canary                                                   \
 112         "movq %P[task_canary](%%rsi),%%r8\n\t"                            \
 113         "movq %%r8,"__percpu_arg([gs_canary])"\n\t"
 114 #define __switch_canary_oparam                                            \
 115         , [gs_canary] "=m" (irq_stack_union.stack_canary)
 116 #define __switch_canary_iparam                                            \
 117         , [task_canary] "i" (offsetof(struct task_struct, stack_canary))
 118 #else   /* CC_STACKPROTECTOR */
 119 #define __switch_canary
 120 #define __switch_canary_oparam
 121 #define __switch_canary_iparam
 122 #endif  /* CC_STACKPROTECTOR */
 123
 124 /* Save restore flags to clear handle leaking NT */
 125 #define switch_to(prev, next, last) \
 126         asm volatile(SAVE_CONTEXT                                         \
 127              "movq %%rsp,%P[threadrsp](%[prev])\n\t" /* save RSP */       \
 128              "movq %P[threadrsp](%[next]),%%rsp\n\t" /* restore RSP */    \
 129              "call __switch_to\n\t"                                       \
 130              "movq "__percpu_arg([current_task])",%%rsi\n\t"              \
 131              __switch_canary                                              \
 132              "movq %P[thread_info](%%rsi),%%r8\n\t"                       \
 133              "movq %%rax,%%rdi\n\t"                                       \
 134              "testl  %[_tif_fork],%P[ti_flags](%%r8)\n\t"                 \
 135              "jnz   ret_from_fork\n\t"                                    \
 136              RESTORE_CONTEXT                                              \
 137              : "=a" (last)                                                \
 138                __switch_canary_oparam                                     \
 139              : [next] "S" (next), [prev] "D" (prev),                      \
 140                [threadrsp] "i" (offsetof(struct task_struct, thread.sp)), \
 141                [ti_flags] "i" (offsetof(struct thread_info, flags)),      \
 142                [_tif_fork] "i" (_TIF_FORK),                               \
 143                [thread_info] "i" (offsetof(struct task_struct, stack)),   \
 144                [current_task] "m" (current_task)                          \
 145                __switch_canary_iparam                                     \
 146              : "memory", "cc" __EXTRA_CLOBBER)
 147 #endif
 148
 149 #ifdef __KERNEL__
 150
 151 extern void native_load_gs_index(unsigned);
 152
 153 /*
 154  * Load a segment. Fall back on loading the zero
 155  * segment if something goes wrong..
 156  */
 157 #define loadsegment(seg, value)                                         \
 158 do {                                                                    \
 159         unsigned short __val = (value);                                 \
 160                                                                         \
 161         asm volatile("                                          \n"     \
 162                      "1:        movl %k0,%%" #seg "             \n"     \
 163                                                                         \
 164                      ".section .fixup,\"ax\"                    \n"     \
 165                      "2:        xorl %k0,%k0                    \n"     \
 166                      "          jmp 1b                          \n"     \
 167                      ".previous                                 \n"     \
 168                                                                         \
 169                      _ASM_EXTABLE(1b, 2b)                               \
 170                                                                         \
 171                      : "+r" (__val) : : "memory");                      \
 172 } while (0)
 173
 174 /*
 175  * Save a segment register away
 176  */
 177 #define savesegment(seg, value)                         \
 178         asm("mov %%" #seg ",%0":"=r" (value) : : "memory")
 179
 180 /*
 181  * x86_32 user gs accessors.
 182  */
 183 #ifdef CONFIG_X86_32
 184 #ifdef CONFIG_X86_32_LAZY_GS
 185 #define get_user_gs(regs)       (u16)({unsigned long v; savesegment(gs, v); v;})
 186 #define set_user_gs(regs, v)    loadsegment(gs, (unsigned long)(v))
 187 #define task_user_gs(tsk)       ((tsk)->thread.gs)
 188 #define lazy_save_gs(v)         savesegment(gs, (v))
 189 #define lazy_load_gs(v)         loadsegment(gs, (v))
 190 #else   /* X86_32_LAZY_GS */
 191 #define get_user_gs(regs)       (u16)((regs)->gs)
 192 #define set_user_gs(regs, v)    do { (regs)->gs = (v); } while (0)
 193 #define task_user_gs(tsk)       (task_pt_regs(tsk)->gs)
 194 #define lazy_save_gs(v)         do { } while (0)
 195 #define lazy_load_gs(v)         do { } while (0)
 196 #endif  /* X86_32_LAZY_GS */
 197 #endif  /* X86_32 */
 198
 199 static inline unsigned long get_limit(unsigned long segment)
 200 {
 201         unsigned long __limit;
 202         asm("lsll %1,%0" : "=r" (__limit) : "r" (segment));
 203         return __limit + 1;
 204 }
 205
 206 static inline void native_clts(void)
 207 {
 208         asm volatile("clts");
 209 }
 210
 211 /*
 212  * Volatile isn't enough to prevent the compiler from reordering the
 213  * read/write functions for the control registers and messing everything up.
 214  * A memory clobber would solve the problem, but would prevent reordering of
 215  * all loads stores around it, which can hurt performance. Solution is to
 216  * use a variable and mimic reads and writes to it to enforce serialization
 217  */
 218 static unsigned long __force_order;
 219
 220 static inline unsigned long native_read_cr0(void)
 221 {
 222         unsigned long val;
 223         asm volatile("mov %%cr0,%0\n\t" : "=r" (val), "=m" (__force_order));
 224         return val;
 225 }
 226
 227 static inline void native_write_cr0(unsigned long val)
 228 {
 229         asm volatile("mov %0,%%cr0": : "r" (val), "m" (__force_order));
 230 }
 231
 232 static inline unsigned long native_read_cr2(void)
 233 {
 234         unsigned long val;
 235         asm volatile("mov %%cr2,%0\n\t" : "=r" (val), "=m" (__force_order));
 236         return val;
 237 }
 238
 239 static inline void native_write_cr2(unsigned long val)
 240 {
 241         asm volatile("mov %0,%%cr2": : "r" (val), "m" (__force_order));
 242 }
 243
 244 static inline unsigned long native_read_cr3(void)
 245 {
 246         unsigned long val;
 247         asm volatile("mov %%cr3,%0\n\t" : "=r" (val), "=m" (__force_order));
 248         return val;
 249 }
 250
 251 static inline void native_write_cr3(unsigned long val)
 252 {
 253         asm volatile("mov %0,%%cr3": : "r" (val), "m" (__force_order));
 254 }
 255
 256 static inline unsigned long native_read_cr4(void)
 257 {
 258         unsigned long val;
 259         asm volatile("mov %%cr4,%0\n\t" : "=r" (val), "=m" (__force_order));
 260         return val;
 261 }
 262
 263 static inline unsigned long native_read_cr4_safe(void)
 264 {
 265         unsigned long val;
 266         /* This could fault if %cr4 does not exist. In x86_64, a cr4 always
 267          * exists, so it will never fail. */
 268 #ifdef CONFIG_X86_32
 269         asm volatile("1: mov %%cr4, %0\n"
 270                      "2:\n"
 271                      _ASM_EXTABLE(1b, 2b)
 272                      : "=r" (val), "=m" (__force_order) : "0" (0));
 273 #else
 274         val = native_read_cr4();
 275 #endif
 276         return val;
 277 }
 278
 279 static inline void native_write_cr4(unsigned long val)
 280 {
 281         asm volatile("mov %0,%%cr4": : "r" (val), "m" (__force_order));
 282 }
 283
 284 #ifdef CONFIG_X86_64
 285 static inline unsigned long native_read_cr8(void)
 286 {
 287         unsigned long cr8;
 288         asm volatile("movq %%cr8,%0" : "=r" (cr8));
 289         return cr8;
 290 }
 291
 292 static inline void native_write_cr8(unsigned long val)
 293 {
 294         asm volatile("movq %0,%%cr8" :: "r" (val) : "memory");
 295 }
 296 #endif
 297
 298 static inline void native_wbinvd(void)
 299 {
 300         asm volatile("wbinvd": : :"memory");
 301 }
 302
 303 #ifdef CONFIG_PARAVIRT
 304 #include <asm/paravirt.h>
 305 #else
 306
 307 static inline unsigned long read_cr0(void)
 308 {
 309         return native_read_cr0();
 310 }
 311
 312 static inline void write_cr0(unsigned long x)
 313 {
 314         native_write_cr0(x);
 315 }
 316
 317 static inline unsigned long read_cr2(void)
 318 {
 319         return native_read_cr2();
 320 }
 321
 322 static inline void write_cr2(unsigned long x)
 323 {
 324         native_write_cr2(x);
 325 }
 326
 327 static inline unsigned long read_cr3(void)
 328 {
 329         return native_read_cr3();
 330 }
 331
 332 static inline void write_cr3(unsigned long x)
 333 {
 334         native_write_cr3(x);
 335 }
 336
 337 static inline unsigned long read_cr4(void)
 338 {
 339         return native_read_cr4();
 340 }
 341
 342 static inline unsigned long read_cr4_safe(void)
 343 {
 344         return native_read_cr4_safe();
 345 }
 346
 347 static inline void write_cr4(unsigned long x)
 348 {
 349         native_write_cr4(x);
 350 }
 351
 352 static inline void wbinvd(void)
 353 {
 354         native_wbinvd();
 355 }
 356
 357 #ifdef CONFIG_X86_64
 358
 359 static inline unsigned long read_cr8(void)
 360 {
 361         return native_read_cr8();
 362 }
 363
 364 static inline void write_cr8(unsigned long x)
 365 {
 366         native_write_cr8(x);
 367 }
 368
 369 static inline void load_gs_index(unsigned selector)
 370 {
 371         native_load_gs_index(selector);
 372 }
 373
 374 #endif
 375
 376 /* Clear the 'TS' bit */
 377 static inline void clts(void)
 378 {
 379         native_clts();
 380 }
 381
 382 #endif/* CONFIG_PARAVIRT */
 383
 384 #define stts() write_cr0(read_cr0() | X86_CR0_TS)
 385
 386 #endif /* __KERNEL__ */
 387
 388 static inline void clflush(volatile void *__p)
 389 {
 390         asm volatile("clflush %0" : "+m" (*(volatile char __force *)__p));
 391 }
 392
 393 #define nop() asm volatile ("nop")
 394
 395 void disable_hlt(void);
 396 void enable_hlt(void);
 397
 398 void cpu_idle_wait(void);
 399
 400 extern unsigned long arch_align_stack(unsigned long sp);
 401 extern void free_init_pages(char *what, unsigned long begin, unsigned long end);
 402
 403 void default_idle(void);
 404
 405 void stop_this_cpu(void *dummy);
 406
 407 /*
 408  * Force strict CPU ordering.
 409  * And yes, this is required on UP too when we're talking
 410  * to devices.
 411  */
 412 #ifdef CONFIG_X86_32
 413 /*
 414  * Some non-Intel clones support out of order store. wmb() ceases to be a
 415  * nop for these.
 416  */
 417 #define mb() alternative("lock; addl $0,0(%%esp)", "mfence", X86_FEATURE_XMM2)
 418 #define rmb() alternative("lock; addl $0,0(%%esp)", "lfence", X86_FEATURE_XMM2)
 419 #define wmb() alternative("lock; addl $0,0(%%esp)", "sfence", X86_FEATURE_XMM)
 420 #else
 421 #define mb()    asm volatile("mfence":::"memory")
 422 #define rmb()   asm volatile("lfence":::"memory")
 423 #define wmb()   asm volatile("sfence" ::: "memory")
 424 #endif
 425
 426 /**
 427  * read_barrier_depends - Flush all pending reads that subsequents reads
 428  * depend on.
 429  *
 430  * No data-dependent reads from memory-like regions are ever reordered
 431  * over this barrier.  All reads preceding this primitive are guaranteed
 432  * to access memory (but not necessarily other CPUs' caches) before any
 433  * reads following this primitive that depend on the data return by
 434  * any of the preceding reads.  This primitive is much lighter weight than
 435  * rmb() on most CPUs, and is never heavier weight than is
 436  * rmb().
 437  *
 438  * These ordering constraints are respected by both the local CPU
 439  * and the compiler.
 440  *
 441  * Ordering is not guaranteed by anything other than these primitives,
 442  * not even by data dependencies.  See the documentation for
 443  * memory_barrier() for examples and URLs to more information.
 444  *
 445  * For example, the following code would force ordering (the initial
 446  * value of "a" is zero, "b" is one, and "p" is "&a"):
 447  *
 448  * <programlisting>
 449  *      CPU 0                           CPU 1
 450  *
 451  *      b = 2;
 452  *      memory_barrier();
 453  *      p = &b;                         q = p;
 454  *                                      read_barrier_depends();
 455  *                                      d = *q;
 456  * </programlisting>
 457  *
 458  * because the read of "*q" depends on the read of "p" and these
 459  * two reads are separated by a read_barrier_depends().  However,
 460  * the following code, with the same initial values for "a" and "b":
 461  *
 462  * <programlisting>
 463  *      CPU 0                           CPU 1
 464  *
 465  *      a = 2;
 466  *      memory_barrier();
 467  *      b = 3;                          y = b;
 468  *                                      read_barrier_depends();
 469  *                                      x = a;
 470  * </programlisting>
 471  *
 472  * does not enforce ordering, since there is no data dependency between
 473  * the read of "a" and the read of "b".  Therefore, on some CPUs, such
 474  * as Alpha, "y" could be set to 3 and "x" to 0.  Use rmb()
 475  * in cases like this where there are no data dependencies.
 476  **/
 477
 478 #define read_barrier_depends()  do { } while (0)
 479
 480 #ifdef CONFIG_SMP
 481 #define smp_mb()        mb()
 482 #ifdef CONFIG_X86_PPRO_FENCE
 483 # define smp_rmb()      rmb()
 484 #else
 485 # define smp_rmb()      barrier()
 486 #endif
 487 #ifdef CONFIG_X86_OOSTORE
 488 # define smp_wmb()      wmb()
 489 #else
 490 # define smp_wmb()      barrier()
 491 #endif
 492 #define smp_read_barrier_depends()      read_barrier_depends()
 493 #define set_mb(var, value) do { (void)xchg(&var, value); } while (0)
 494 #else
 495 #define smp_mb()        barrier()
 496 #define smp_rmb()       barrier()
 497 #define smp_wmb()       barrier()
 498 #define smp_read_barrier_depends()      do { } while (0)
 499 #define set_mb(var, value) do { var = value; barrier(); } while (0)
 500 #endif
 501
 502 /*
 503  * Stop RDTSC speculation. This is needed when you need to use RDTSC
 504  * (or get_cycles or vread that possibly accesses the TSC) in a defined
 505  * code region.
 506  *
 507  * (Could use an alternative three way for this if there was one.)
 508  */
 509 static __always_inline void rdtsc_barrier(void)
 510 {
 511         alternative(ASM_NOP3, "mfence", X86_FEATURE_MFENCE_RDTSC);
 512         alternative(ASM_NOP3, "lfence", X86_FEATURE_LFENCE_RDTSC);
 513 }
 514
 515 /*
 516  * We handle most unaligned accesses in hardware.  On the other hand
 517  * unaligned DMA can be quite expensive on some Nehalem processors.
 518  *
 519  * Based on this we disable the IP header alignment in network drivers.
 520  */
 521 #define NET_IP_ALIGN    0
 522 #endif /* _ASM_X86_SYSTEM_H */