arch/sparc/include/asm/system_64.h

   1 #ifndef __SPARC64_SYSTEM_H
   2 #define __SPARC64_SYSTEM_H
   3
   4 #include <asm/ptrace.h>
   5 #include <asm/processor.h>
   6 #include <asm/visasm.h>
   7
   8 #ifndef __ASSEMBLY__
   9
  10 #include <linux/irqflags.h>
  11 #include <asm-generic/cmpxchg-local.h>
  12
  13 /*
  14  * Sparc (general) CPU types
  15  */
  16 enum sparc_cpu {
  17   sun4        = 0x00,
  18   sun4c       = 0x01,
  19   sun4m       = 0x02,
  20   sun4d       = 0x03,
  21   sun4e       = 0x04,
  22   sun4u       = 0x05, /* V8 ploos ploos */
  23   sun_unknown = 0x06,
  24   ap1000      = 0x07, /* almost a sun4m */
  25 };
  26
  27 #define sparc_cpu_model sun4u
  28
  29 /* This cannot ever be a sun4c :) That's just history. */
  30 #define ARCH_SUN4C 0
  31
  32 extern char reboot_command[];
  33
  34 /* These are here in an effort to more fully work around Spitfire Errata
  35  * #51.  Essentially, if a memory barrier occurs soon after a mispredicted
  36  * branch, the chip can stop executing instructions until a trap occurs.
  37  * Therefore, if interrupts are disabled, the chip can hang forever.
  38  *
  39  * It used to be believed that the memory barrier had to be right in the
  40  * delay slot, but a case has been traced recently wherein the memory barrier
  41  * was one instruction after the branch delay slot and the chip still hung.
  42  * The offending sequence was the following in sym_wakeup_done() of the
  43  * sym53c8xx_2 driver:
  44  *
  45  *      call    sym_ccb_from_dsa, 0
  46  *       movge  %icc, 0, %l0
  47  *      brz,pn  %o0, .LL1303
  48  *       mov    %o0, %l2
  49  *      membar  #LoadLoad
  50  *
  51  * The branch has to be mispredicted for the bug to occur.  Therefore, we put
  52  * the memory barrier explicitly into a "branch always, predicted taken"
  53  * delay slot to avoid the problem case.
  54  */
  55 #define membar_safe(type) \
  56 do {    __asm__ __volatile__("ba,pt     %%xcc, 1f\n\t" \
  57                              " membar   " type "\n" \
  58                              "1:\n" \
  59                              : : : "memory"); \
  60 } while (0)
  61
  62 #define mb()    membar_safe("#StoreLoad")
  63 #define rmb()   __asm__ __volatile__("":::"memory")
  64 #define wmb()   __asm__ __volatile__("":::"memory")
  65
  66 #endif
  67
  68 #define nop()           __asm__ __volatile__ ("nop")
  69
  70 #define read_barrier_depends()          do { } while(0)
  71 #define set_mb(__var, __value) \
  72         do { __var = __value; membar_safe("#StoreLoad"); } while(0)
  73
  74 #ifdef CONFIG_SMP
  75 #define smp_mb()        mb()
  76 #define smp_rmb()       rmb()
  77 #define smp_wmb()       wmb()
  78 #else
  79 #define smp_mb()        __asm__ __volatile__("":::"memory")
  80 #define smp_rmb()       __asm__ __volatile__("":::"memory")
  81 #define smp_wmb()       __asm__ __volatile__("":::"memory")
  82 #endif
  83
  84 #define smp_read_barrier_depends()      do { } while(0)
  85
  86 #define flushi(addr)    __asm__ __volatile__ ("flush %0" : : "r" (addr) : "memory")
  87
  88 #define flushw_all()    __asm__ __volatile__("flushw")
  89
  90 /* Performance counter register access. */
  91 #define read_pcr(__p)  __asm__ __volatile__("rd %%pcr, %0" : "=r" (__p))
  92 #define write_pcr(__p) __asm__ __volatile__("wr %0, 0x0, %%pcr" : : "r" (__p))
  93 #define read_pic(__p)  __asm__ __volatile__("rd %%pic, %0" : "=r" (__p))
  94
  95 /* Blackbird errata workaround.  See commentary in
  96  * arch/sparc64/kernel/smp.c:smp_percpu_timer_interrupt()
  97  * for more information.
  98  */
  99 #define write_pic(__p)                                          \
 100         __asm__ __volatile__("ba,pt     %%xcc, 99f\n\t"         \
 101                              ".align    64\n"                   \
 102                           "99:wr        %0, 0x0, %%pic\n\t"     \
 103                              "rd        %%pic, %%g0" : : "r" (__p))
 104 #define reset_pic()     write_pic(0)
 105
 106 #ifndef __ASSEMBLY__
 107
 108 extern void sun_do_break(void);
 109 extern int stop_a_enabled;
 110 extern int scons_pwroff;
 111
 112 extern void fault_in_user_windows(void);
 113 extern void synchronize_user_stack(void);
 114
 115 extern void __flushw_user(void);
 116 #define flushw_user() __flushw_user()
 117
 118 #define flush_user_windows flushw_user
 119 #define flush_register_windows flushw_all
 120
 121 /* Don't hold the runqueue lock over context switch */
 122 #define __ARCH_WANT_UNLOCKED_CTXSW
 123 #define prepare_arch_switch(next)               \
 124 do {                                            \
 125         flushw_all();                           \
 126 } while (0)
 127
 128         /* See what happens when you design the chip correctly?
 129          *
 130          * We tell gcc we clobber all non-fixed-usage registers except
 131          * for l0/l1.  It will use one for 'next' and the other to hold
 132          * the output value of 'last'.  'next' is not referenced again
 133          * past the invocation of switch_to in the scheduler, so we need
 134          * not preserve it's value.  Hairy, but it lets us remove 2 loads
 135          * and 2 stores in this critical code path.  -DaveM
 136          */
 137 #define switch_to(prev, next, last)                                     \
 138 do {    if (test_thread_flag(TIF_PERFCTR)) {                            \
 139                 unsigned long __tmp;                                    \
 140                 read_pcr(__tmp);                                        \
 141                 current_thread_info()->pcr_reg = __tmp;                 \
 142                 read_pic(__tmp);                                        \
 143                 current_thread_info()->kernel_cntd0 += (unsigned int)(__tmp);\
 144                 current_thread_info()->kernel_cntd1 += ((__tmp) >> 32); \
 145         }                                                               \
 146         flush_tlb_pending();                                            \
 147         save_and_clear_fpu();                                           \
 148         /* If you are tempted to conditionalize the following */        \
 149         /* so that ASI is only written if it changes, think again. */   \
 150         __asm__ __volatile__("wr %%g0, %0, %%asi"                       \
 151         : : "r" (__thread_flag_byte_ptr(task_thread_info(next))[TI_FLAG_BYTE_CURRENT_DS]));\
 152         trap_block[current_thread_info()->cpu].thread =                 \
 153                 task_thread_info(next);                                 \
 154         __asm__ __volatile__(                                           \
 155         "mov    %%g4, %%g7\n\t"                                         \
 156         "stx    %%i6, [%%sp + 2047 + 0x70]\n\t"                         \
 157         "stx    %%i7, [%%sp + 2047 + 0x78]\n\t"                         \
 158         "rdpr   %%wstate, %%o5\n\t"                                     \
 159         "stx    %%o6, [%%g6 + %6]\n\t"                                  \
 160         "stb    %%o5, [%%g6 + %5]\n\t"                                  \
 161         "rdpr   %%cwp, %%o5\n\t"                                        \
 162         "stb    %%o5, [%%g6 + %8]\n\t"                                  \
 163         "wrpr   %%g0, 15, %%pil\n\t"                                    \
 164         "mov    %4, %%g6\n\t"                                           \
 165         "ldub   [%4 + %8], %%g1\n\t"                                    \
 166         "wrpr   %%g1, %%cwp\n\t"                                        \
 167         "ldx    [%%g6 + %6], %%o6\n\t"                                  \
 168         "ldub   [%%g6 + %5], %%o5\n\t"                                  \
 169         "ldub   [%%g6 + %7], %%o7\n\t"                                  \
 170         "wrpr   %%o5, 0x0, %%wstate\n\t"                                \
 171         "ldx    [%%sp + 2047 + 0x70], %%i6\n\t"                         \
 172         "ldx    [%%sp + 2047 + 0x78], %%i7\n\t"                         \
 173         "ldx    [%%g6 + %9], %%g4\n\t"                                  \
 174         "wrpr   %%g0, 14, %%pil\n\t"                                    \
 175         "brz,pt %%o7, switch_to_pc\n\t"                                 \
 176         " mov   %%g7, %0\n\t"                                           \
 177         "sethi  %%hi(ret_from_syscall), %%g1\n\t"                       \
 178         "jmpl   %%g1 + %%lo(ret_from_syscall), %%g0\n\t"                \
 179         " nop\n\t"                                                      \
 180         ".globl switch_to_pc\n\t"                                       \
 181         "switch_to_pc:\n\t"                                             \
 182         : "=&r" (last), "=r" (current), "=r" (current_thread_info_reg), \
 183           "=r" (__local_per_cpu_offset)                                 \
 184         : "0" (task_thread_info(next)),                                 \
 185           "i" (TI_WSTATE), "i" (TI_KSP), "i" (TI_NEW_CHILD),            \
 186           "i" (TI_CWP), "i" (TI_TASK)                                   \
 187         : "cc",                                                         \
 188                 "g1", "g2", "g3",                   "g7",               \
 189                 "l1", "l2", "l3", "l4", "l5", "l6", "l7",               \
 190           "i0", "i1", "i2", "i3", "i4", "i5",                           \
 191           "o0", "o1", "o2", "o3", "o4", "o5",       "o7");              \
 192         /* If you fuck with this, update ret_from_syscall code too. */  \
 193         if (test_thread_flag(TIF_PERFCTR)) {                            \
 194                 write_pcr(current_thread_info()->pcr_reg);              \
 195                 reset_pic();                                            \
 196         }                                                               \
 197 } while(0)
 198
 199 static inline unsigned long xchg32(__volatile__ unsigned int *m, unsigned int val)
 200 {
 201         unsigned long tmp1, tmp2;
 202
 203         __asm__ __volatile__(
 204 "       mov             %0, %1\n"
 205 "1:     lduw            [%4], %2\n"
 206 "       cas             [%4], %2, %0\n"
 207 "       cmp             %2, %0\n"
 208 "       bne,a,pn        %%icc, 1b\n"
 209 "        mov            %1, %0\n"
 210         : "=&r" (val), "=&r" (tmp1), "=&r" (tmp2)
 211         : "0" (val), "r" (m)
 212         : "cc", "memory");
 213         return val;
 214 }
 215
 216 static inline unsigned long xchg64(__volatile__ unsigned long *m, unsigned long val)
 217 {
 218         unsigned long tmp1, tmp2;
 219
 220         __asm__ __volatile__(
 221 "       mov             %0, %1\n"
 222 "1:     ldx             [%4], %2\n"
 223 "       casx            [%4], %2, %0\n"
 224 "       cmp             %2, %0\n"
 225 "       bne,a,pn        %%xcc, 1b\n"
 226 "        mov            %1, %0\n"
 227         : "=&r" (val), "=&r" (tmp1), "=&r" (tmp2)
 228         : "0" (val), "r" (m)
 229         : "cc", "memory");
 230         return val;
 231 }
 232
 233 #define xchg(ptr,x) ((__typeof__(*(ptr)))__xchg((unsigned long)(x),(ptr),sizeof(*(ptr))))
 234
 235 extern void __xchg_called_with_bad_pointer(void);
 236
 237 static inline unsigned long __xchg(unsigned long x, __volatile__ void * ptr,
 238                                        int size)
 239 {
 240         switch (size) {
 241         case 4:
 242                 return xchg32(ptr, x);
 243         case 8:
 244                 return xchg64(ptr, x);
 245         };
 246         __xchg_called_with_bad_pointer();
 247         return x;
 248 }
 249
 250 extern void die_if_kernel(char *str, struct pt_regs *regs) __attribute__ ((noreturn));
 251
 252 /*
 253  * Atomic compare and exchange.  Compare OLD with MEM, if identical,
 254  * store NEW in MEM.  Return the initial value in MEM.  Success is
 255  * indicated by comparing RETURN with OLD.
 256  */
 257
 258 #define __HAVE_ARCH_CMPXCHG 1
 259
 260 static inline unsigned long
 261 __cmpxchg_u32(volatile int *m, int old, int new)
 262 {
 263         __asm__ __volatile__("cas [%2], %3, %0"
 264                              : "=&r" (new)
 265                              : "0" (new), "r" (m), "r" (old)
 266                              : "memory");
 267
 268         return new;
 269 }
 270
 271 static inline unsigned long
 272 __cmpxchg_u64(volatile long *m, unsigned long old, unsigned long new)
 273 {
 274         __asm__ __volatile__("casx [%2], %3, %0"
 275                              : "=&r" (new)
 276                              : "0" (new), "r" (m), "r" (old)
 277                              : "memory");
 278
 279         return new;
 280 }
 281
 282 /* This function doesn't exist, so you'll get a linker error
 283    if something tries to do an invalid cmpxchg().  */
 284 extern void __cmpxchg_called_with_bad_pointer(void);
 285
 286 static inline unsigned long
 287 __cmpxchg(volatile void *ptr, unsigned long old, unsigned long new, int size)
 288 {
 289         switch (size) {
 290                 case 4:
 291                         return __cmpxchg_u32(ptr, old, new);
 292                 case 8:
 293                         return __cmpxchg_u64(ptr, old, new);
 294         }
 295         __cmpxchg_called_with_bad_pointer();
 296         return old;
 297 }
 298
 299 #define cmpxchg(ptr,o,n)                                                 \
 300   ({                                                                     \
 301      __typeof__(*(ptr)) _o_ = (o);                                       \
 302      __typeof__(*(ptr)) _n_ = (n);                                       \
 303      (__typeof__(*(ptr))) __cmpxchg((ptr), (unsigned long)_o_,           \
 304                                     (unsigned long)_n_, sizeof(*(ptr))); \
 305   })
 306
 307 /*
 308  * cmpxchg_local and cmpxchg64_local are atomic wrt current CPU. Always make
 309  * them available.
 310  */
 311
 312 static inline unsigned long __cmpxchg_local(volatile void *ptr,
 313                                       unsigned long old,
 314                                       unsigned long new, int size)
 315 {
 316         switch (size) {
 317         case 4:
 318         case 8: return __cmpxchg(ptr, old, new, size);
 319         default:
 320                 return __cmpxchg_local_generic(ptr, old, new, size);
 321         }
 322
 323         return old;
 324 }
 325
 326 #define cmpxchg_local(ptr, o, n)                                        \
 327         ((__typeof__(*(ptr)))__cmpxchg_local((ptr), (unsigned long)(o), \
 328                         (unsigned long)(n), sizeof(*(ptr))))
 329 #define cmpxchg64_local(ptr, o, n)                                      \
 330   ({                                                                    \
 331         BUILD_BUG_ON(sizeof(*(ptr)) != 8);                              \
 332         cmpxchg_local((ptr), (o), (n));                                 \
 333   })
 334
 335 #endif /* !(__ASSEMBLY__) */
 336
 337 #define arch_align_stack(x) (x)
 338
 339 #endif /* !(__SPARC64_SYSTEM_H) */