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