include/asm-x86_64/processor.h

   1 /*
   2  * include/asm-x86_64/processor.h
   3  *
   4  * Copyright (C) 1994 Linus Torvalds
   5  */
   6
   7 #ifndef __ASM_X86_64_PROCESSOR_H
   8 #define __ASM_X86_64_PROCESSOR_H
   9
  10 #include <asm/segment.h>
  11 #include <asm/page.h>
  12 #include <asm/types.h>
  13 #include <asm/sigcontext.h>
  14 #include <asm/cpufeature.h>
  15 #include <linux/config.h>
  16 #include <linux/threads.h>
  17 #include <asm/msr.h>
  18 #include <asm/current.h>
  19 #include <asm/system.h>
  20 #include <asm/mmsegment.h>
  21 #include <asm/percpu.h>
  22 #include <linux/personality.h>
  23
  24 #define TF_MASK         0x00000100
  25 #define IF_MASK         0x00000200
  26 #define IOPL_MASK       0x00003000
  27 #define NT_MASK         0x00004000
  28 #define VM_MASK         0x00020000
  29 #define AC_MASK         0x00040000
  30 #define VIF_MASK        0x00080000      /* virtual interrupt flag */
  31 #define VIP_MASK        0x00100000      /* virtual interrupt pending */
  32 #define ID_MASK         0x00200000
  33
  34 #define desc_empty(desc) \
  35                (!((desc)->a | (desc)->b))
  36
  37 #define desc_equal(desc1, desc2) \
  38                (((desc1)->a == (desc2)->a) && ((desc1)->b == (desc2)->b))
  39
  40 /*
  41  * Default implementation of macro that returns current
  42  * instruction pointer ("program counter").
  43  */
  44 #define current_text_addr() ({ void *pc; asm volatile("leaq 1f(%%rip),%0\n1:":"=r"(pc)); pc; })
  45
  46 /*
  47  *  CPU type and hardware bug flags. Kept separately for each CPU.
  48  */
  49
  50 struct cpuinfo_x86 {
  51         __u8    x86;            /* CPU family */
  52         __u8    x86_vendor;     /* CPU vendor */
  53         __u8    x86_model;
  54         __u8    x86_mask;
  55         int     cpuid_level;    /* Maximum supported CPUID level, -1=no CPUID */
  56         __u32   x86_capability[NCAPINTS];
  57         char    x86_vendor_id[16];
  58         char    x86_model_id[64];
  59         int     x86_cache_size;  /* in KB */
  60         int     x86_clflush_size;
  61         int     x86_cache_alignment;
  62         int     x86_tlbsize;    /* number of 4K pages in DTLB/ITLB combined(in pages)*/
  63         __u8    x86_virt_bits, x86_phys_bits;
  64         __u8    x86_max_cores;  /* cpuid returned max cores value */
  65         __u32   x86_power;
  66         __u32   extended_cpuid_level;   /* Max extended CPUID function supported */
  67         unsigned long loops_per_jiffy;
  68         __u8    apicid;
  69         __u8    booted_cores;   /* number of cores as seen by OS */
  70 } ____cacheline_aligned;
  71
  72 #define X86_VENDOR_INTEL 0
  73 #define X86_VENDOR_CYRIX 1
  74 #define X86_VENDOR_AMD 2
  75 #define X86_VENDOR_UMC 3
  76 #define X86_VENDOR_NEXGEN 4
  77 #define X86_VENDOR_CENTAUR 5
  78 #define X86_VENDOR_RISE 6
  79 #define X86_VENDOR_TRANSMETA 7
  80 #define X86_VENDOR_NUM 8
  81 #define X86_VENDOR_UNKNOWN 0xff
  82
  83 #ifdef CONFIG_SMP
  84 extern struct cpuinfo_x86 cpu_data[];
  85 #define current_cpu_data cpu_data[smp_processor_id()]
  86 #else
  87 #define cpu_data (&boot_cpu_data)
  88 #define current_cpu_data boot_cpu_data
  89 #endif
  90
  91 extern char ignore_irq13;
  92
  93 extern void identify_cpu(struct cpuinfo_x86 *);
  94 extern void print_cpu_info(struct cpuinfo_x86 *);
  95 extern unsigned int init_intel_cacheinfo(struct cpuinfo_x86 *c);
  96
  97 /*
  98  * EFLAGS bits
  99  */
 100 #define X86_EFLAGS_CF   0x00000001 /* Carry Flag */
 101 #define X86_EFLAGS_PF   0x00000004 /* Parity Flag */
 102 #define X86_EFLAGS_AF   0x00000010 /* Auxillary carry Flag */
 103 #define X86_EFLAGS_ZF   0x00000040 /* Zero Flag */
 104 #define X86_EFLAGS_SF   0x00000080 /* Sign Flag */
 105 #define X86_EFLAGS_TF   0x00000100 /* Trap Flag */
 106 #define X86_EFLAGS_IF   0x00000200 /* Interrupt Flag */
 107 #define X86_EFLAGS_DF   0x00000400 /* Direction Flag */
 108 #define X86_EFLAGS_OF   0x00000800 /* Overflow Flag */
 109 #define X86_EFLAGS_IOPL 0x00003000 /* IOPL mask */
 110 #define X86_EFLAGS_NT   0x00004000 /* Nested Task */
 111 #define X86_EFLAGS_RF   0x00010000 /* Resume Flag */
 112 #define X86_EFLAGS_VM   0x00020000 /* Virtual Mode */
 113 #define X86_EFLAGS_AC   0x00040000 /* Alignment Check */
 114 #define X86_EFLAGS_VIF  0x00080000 /* Virtual Interrupt Flag */
 115 #define X86_EFLAGS_VIP  0x00100000 /* Virtual Interrupt Pending */
 116 #define X86_EFLAGS_ID   0x00200000 /* CPUID detection flag */
 117
 118 /*
 119  * Intel CPU features in CR4
 120  */
 121 #define X86_CR4_VME             0x0001  /* enable vm86 extensions */
 122 #define X86_CR4_PVI             0x0002  /* virtual interrupts flag enable */
 123 #define X86_CR4_TSD             0x0004  /* disable time stamp at ipl 3 */
 124 #define X86_CR4_DE              0x0008  /* enable debugging extensions */
 125 #define X86_CR4_PSE             0x0010  /* enable page size extensions */
 126 #define X86_CR4_PAE             0x0020  /* enable physical address extensions */
 127 #define X86_CR4_MCE             0x0040  /* Machine check enable */
 128 #define X86_CR4_PGE             0x0080  /* enable global pages */
 129 #define X86_CR4_PCE             0x0100  /* enable performance counters at ipl 3 */
 130 #define X86_CR4_OSFXSR          0x0200  /* enable fast FPU save and restore */
 131 #define X86_CR4_OSXMMEXCPT      0x0400  /* enable unmasked SSE exceptions */
 132
 133 /*
 134  * Save the cr4 feature set we're using (ie
 135  * Pentium 4MB enable and PPro Global page
 136  * enable), so that any CPU's that boot up
 137  * after us can get the correct flags.
 138  */
 139 extern unsigned long mmu_cr4_features;
 140
 141 static inline void set_in_cr4 (unsigned long mask)
 142 {
 143         mmu_cr4_features |= mask;
 144         __asm__("movq %%cr4,%%rax\n\t"
 145                 "orq %0,%%rax\n\t"
 146                 "movq %%rax,%%cr4\n"
 147                 : : "irg" (mask)
 148                 :"ax");
 149 }
 150
 151 static inline void clear_in_cr4 (unsigned long mask)
 152 {
 153         mmu_cr4_features &= ~mask;
 154         __asm__("movq %%cr4,%%rax\n\t"
 155                 "andq %0,%%rax\n\t"
 156                 "movq %%rax,%%cr4\n"
 157                 : : "irg" (~mask)
 158                 :"ax");
 159 }
 160
 161
 162 /*
 163  * User space process size. 47bits minus one guard page.
 164  */
 165 #define TASK_SIZE64     (0x800000000000UL - 4096)
 166
 167 /* This decides where the kernel will search for a free chunk of vm
 168  * space during mmap's.
 169  */
 170 #define IA32_PAGE_OFFSET ((current->personality & ADDR_LIMIT_3GB) ? 0xc0000000 : 0xFFFFe000)
 171
 172 #define TASK_SIZE               (test_thread_flag(TIF_IA32) ? IA32_PAGE_OFFSET : TASK_SIZE64)
 173 #define TASK_SIZE_OF(child)     ((test_tsk_thread_flag(child, TIF_IA32)) ? IA32_PAGE_OFFSET : TASK_SIZE64)
 174
 175 #define TASK_UNMAPPED_BASE      PAGE_ALIGN(TASK_SIZE/3)
 176
 177 /*
 178  * Size of io_bitmap.
 179  */
 180 #define IO_BITMAP_BITS  65536
 181 #define IO_BITMAP_BYTES (IO_BITMAP_BITS/8)
 182 #define IO_BITMAP_LONGS (IO_BITMAP_BYTES/sizeof(long))
 183 #define IO_BITMAP_OFFSET offsetof(struct tss_struct,io_bitmap)
 184 #define INVALID_IO_BITMAP_OFFSET 0x8000
 185
 186 struct i387_fxsave_struct {
 187         u16     cwd;
 188         u16     swd;
 189         u16     twd;
 190         u16     fop;
 191         u64     rip;
 192         u64     rdp;
 193         u32     mxcsr;
 194         u32     mxcsr_mask;
 195         u32     st_space[32];   /* 8*16 bytes for each FP-reg = 128 bytes */
 196         u32     xmm_space[64];  /* 16*16 bytes for each XMM-reg = 128 bytes */
 197         u32     padding[24];
 198 } __attribute__ ((aligned (16)));
 199
 200 union i387_union {
 201         struct i387_fxsave_struct       fxsave;
 202 };
 203
 204 struct tss_struct {
 205         u32 reserved1;
 206         u64 rsp0;
 207         u64 rsp1;
 208         u64 rsp2;
 209         u64 reserved2;
 210         u64 ist[7];
 211         u32 reserved3;
 212         u32 reserved4;
 213         u16 reserved5;
 214         u16 io_bitmap_base;
 215         /*
 216          * The extra 1 is there because the CPU will access an
 217          * additional byte beyond the end of the IO permission
 218          * bitmap. The extra byte must be all 1 bits, and must
 219          * be within the limit. Thus we have:
 220          *
 221          * 128 bytes, the bitmap itself, for ports 0..0x3ff
 222          * 8 bytes, for an extra "long" of ~0UL
 223          */
 224         unsigned long io_bitmap[IO_BITMAP_LONGS + 1];
 225 } __attribute__((packed)) ____cacheline_aligned;
 226
 227 extern struct cpuinfo_x86 boot_cpu_data;
 228 DECLARE_PER_CPU(struct tss_struct,init_tss);
 229
 230 #ifdef CONFIG_X86_VSMP
 231 #define ARCH_MIN_TASKALIGN      (1 << INTERNODE_CACHE_SHIFT)
 232 #define ARCH_MIN_MMSTRUCT_ALIGN (1 << INTERNODE_CACHE_SHIFT)
 233 #else
 234 #define ARCH_MIN_TASKALIGN      16
 235 #define ARCH_MIN_MMSTRUCT_ALIGN 0
 236 #endif
 237
 238 struct thread_struct {
 239         unsigned long   rsp0;
 240         unsigned long   rsp;
 241         unsigned long   userrsp;        /* Copy from PDA */
 242         unsigned long   fs;
 243         unsigned long   gs;
 244         unsigned short  es, ds, fsindex, gsindex;
 245 /* Hardware debugging registers */
 246         unsigned long   debugreg0;
 247         unsigned long   debugreg1;
 248         unsigned long   debugreg2;
 249         unsigned long   debugreg3;
 250         unsigned long   debugreg6;
 251         unsigned long   debugreg7;
 252 /* fault info */
 253         unsigned long   cr2, trap_no, error_code;
 254 /* floating point info */
 255         union i387_union        i387  __attribute__((aligned(16)));
 256 /* IO permissions. the bitmap could be moved into the GDT, that would make
 257    switch faster for a limited number of ioperm using tasks. -AK */
 258         int             ioperm;
 259         unsigned long   *io_bitmap_ptr;
 260         unsigned io_bitmap_max;
 261 /* cached TLS descriptors. */
 262         u64 tls_array[GDT_ENTRY_TLS_ENTRIES];
 263 } __attribute__((aligned(16)));
 264
 265 #define INIT_THREAD  { \
 266         .rsp0 = (unsigned long)&init_stack + sizeof(init_stack) \
 267 }
 268
 269 #define INIT_TSS  { \
 270         .rsp0 = (unsigned long)&init_stack + sizeof(init_stack) \
 271 }
 272
 273 #define INIT_MMAP \
 274 { &init_mm, 0, 0, NULL, PAGE_SHARED, VM_READ | VM_WRITE | VM_EXEC, 1, NULL, NULL }
 275
 276 #define STACKFAULT_STACK 1
 277 #define DOUBLEFAULT_STACK 2
 278 #define NMI_STACK 3
 279 #define DEBUG_STACK 4
 280 #define MCE_STACK 5
 281 #define N_EXCEPTION_STACKS 5  /* hw limit: 7 */
 282
 283 #define start_thread(regs,new_rip,new_rsp) do { \
 284         asm volatile("movl %0,%%fs; movl %0,%%es; movl %0,%%ds": :"r" (0));      \
 285         load_gs_index(0);                                                       \
 286         (regs)->rip = (new_rip);                                                 \
 287         (regs)->rsp = (new_rsp);                                                 \
 288         write_pda(oldrsp, (new_rsp));                                            \
 289         (regs)->cs = __USER_CS;                                                  \
 290         (regs)->ss = __USER_DS;                                                  \
 291         (regs)->eflags = 0x200;                                                  \
 292         set_fs(USER_DS);                                                         \
 293 } while(0)
 294
 295 #define get_debugreg(var, register)                             \
 296                 __asm__("movq %%db" #register ", %0"            \
 297                         :"=r" (var))
 298 #define set_debugreg(value, register)                   \
 299                 __asm__("movq %0,%%db" #register                \
 300                         : /* no output */                       \
 301                         :"r" (value))
 302
 303 struct task_struct;
 304 struct mm_struct;
 305
 306 /* Free all resources held by a thread. */
 307 extern void release_thread(struct task_struct *);
 308
 309 /* Prepare to copy thread state - unlazy all lazy status */
 310 extern void prepare_to_copy(struct task_struct *tsk);
 311
 312 /*
 313  * create a kernel thread without removing it from tasklists
 314  */
 315 extern long kernel_thread(int (*fn)(void *), void * arg, unsigned long flags);
 316
 317 /*
 318  * Return saved PC of a blocked thread.
 319  * What is this good for? it will be always the scheduler or ret_from_fork.
 320  */
 321 #define thread_saved_pc(t) (*(unsigned long *)((t)->thread.rsp - 8))
 322
 323 extern unsigned long get_wchan(struct task_struct *p);
 324 #define task_pt_regs(tsk) ((struct pt_regs *)(tsk)->thread.rsp0 - 1)
 325 #define KSTK_EIP(tsk) (task_pt_regs(tsk)->rip)
 326 #define KSTK_ESP(tsk) -1 /* sorry. doesn't work for syscall. */
 327
 328
 329 struct microcode_header {
 330         unsigned int hdrver;
 331         unsigned int rev;
 332         unsigned int date;
 333         unsigned int sig;
 334         unsigned int cksum;
 335         unsigned int ldrver;
 336         unsigned int pf;
 337         unsigned int datasize;
 338         unsigned int totalsize;
 339         unsigned int reserved[3];
 340 };
 341
 342 struct microcode {
 343         struct microcode_header hdr;
 344         unsigned int bits[0];
 345 };
 346
 347 typedef struct microcode microcode_t;
 348 typedef struct microcode_header microcode_header_t;
 349
 350 /* microcode format is extended from prescott processors */
 351 struct extended_signature {
 352         unsigned int sig;
 353         unsigned int pf;
 354         unsigned int cksum;
 355 };
 356
 357 struct extended_sigtable {
 358         unsigned int count;
 359         unsigned int cksum;
 360         unsigned int reserved[3];
 361         struct extended_signature sigs[0];
 362 };
 363
 364 /* '6' because it used to be for P6 only (but now covers Pentium 4 as well) */
 365 #define MICROCODE_IOCFREE       _IO('6',0)
 366
 367
 368 #define ASM_NOP1 K8_NOP1
 369 #define ASM_NOP2 K8_NOP2
 370 #define ASM_NOP3 K8_NOP3
 371 #define ASM_NOP4 K8_NOP4
 372 #define ASM_NOP5 K8_NOP5
 373 #define ASM_NOP6 K8_NOP6
 374 #define ASM_NOP7 K8_NOP7
 375 #define ASM_NOP8 K8_NOP8
 376
 377 /* Opteron nops */
 378 #define K8_NOP1 ".byte 0x90\n"
 379 #define K8_NOP2 ".byte 0x66,0x90\n"
 380 #define K8_NOP3 ".byte 0x66,0x66,0x90\n"
 381 #define K8_NOP4 ".byte 0x66,0x66,0x66,0x90\n"
 382 #define K8_NOP5 K8_NOP3 K8_NOP2
 383 #define K8_NOP6 K8_NOP3 K8_NOP3
 384 #define K8_NOP7 K8_NOP4 K8_NOP3
 385 #define K8_NOP8 K8_NOP4 K8_NOP4
 386
 387 #define ASM_NOP_MAX 8
 388
 389 /* REP NOP (PAUSE) is a good thing to insert into busy-wait loops. */
 390 static inline void rep_nop(void)
 391 {
 392         __asm__ __volatile__("rep;nop": : :"memory");
 393 }
 394
 395 /* Stop speculative execution */
 396 static inline void sync_core(void)
 397 {
 398         int tmp;
 399         asm volatile("cpuid" : "=a" (tmp) : "0" (1) : "ebx","ecx","edx","memory");
 400 }
 401
 402 #define cpu_has_fpu 1
 403
 404 #define ARCH_HAS_PREFETCH
 405 static inline void prefetch(void *x)
 406 {
 407         asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
 408 }
 409
 410 #define ARCH_HAS_PREFETCHW 1
 411 static inline void prefetchw(void *x)
 412 {
 413         alternative_input("prefetcht0 (%1)",
 414                           "prefetchw (%1)",
 415                           X86_FEATURE_3DNOW,
 416                           "r" (x));
 417 }
 418
 419 #define ARCH_HAS_SPINLOCK_PREFETCH 1
 420
 421 #define spin_lock_prefetch(x)  prefetchw(x)
 422
 423 #define cpu_relax()   rep_nop()
 424
 425 /*
 426  *      NSC/Cyrix CPU configuration register indexes
 427  */
 428 #define CX86_CCR0 0xc0
 429 #define CX86_CCR1 0xc1
 430 #define CX86_CCR2 0xc2
 431 #define CX86_CCR3 0xc3
 432 #define CX86_CCR4 0xe8
 433 #define CX86_CCR5 0xe9
 434 #define CX86_CCR6 0xea
 435 #define CX86_CCR7 0xeb
 436 #define CX86_DIR0 0xfe
 437 #define CX86_DIR1 0xff
 438 #define CX86_ARR_BASE 0xc4
 439 #define CX86_RCR_BASE 0xdc
 440
 441 /*
 442  *      NSC/Cyrix CPU indexed register access macros
 443  */
 444
 445 #define getCx86(reg) ({ outb((reg), 0x22); inb(0x23); })
 446
 447 #define setCx86(reg, data) do { \
 448         outb((reg), 0x22); \
 449         outb((data), 0x23); \
 450 } while (0)
 451
 452 static inline void serialize_cpu(void)
 453 {
 454         __asm__ __volatile__ ("cpuid" : : : "ax", "bx", "cx", "dx");
 455 }
 456
 457 static inline void __monitor(const void *eax, unsigned long ecx,
 458                 unsigned long edx)
 459 {
 460         /* "monitor %eax,%ecx,%edx;" */
 461         asm volatile(
 462                 ".byte 0x0f,0x01,0xc8;"
 463                 : :"a" (eax), "c" (ecx), "d"(edx));
 464 }
 465
 466 static inline void __mwait(unsigned long eax, unsigned long ecx)
 467 {
 468         /* "mwait %eax,%ecx;" */
 469         asm volatile(
 470                 ".byte 0x0f,0x01,0xc9;"
 471                 : :"a" (eax), "c" (ecx));
 472 }
 473
 474 #define stack_current() \
 475 ({                                                              \
 476         struct thread_info *ti;                                 \
 477         asm("andq %%rsp,%0; ":"=r" (ti) : "0" (CURRENT_MASK));  \
 478         ti->task;                                       \
 479 })
 480
 481 #define cache_line_size() (boot_cpu_data.x86_cache_alignment)
 482
 483 extern unsigned long boot_option_idle_override;
 484 /* Boot loader type from the setup header */
 485 extern int bootloader_type;
 486
 487 #endif /* __ASM_X86_64_PROCESSOR_H */