include/asm-sparc64/system.h

   1 /* $Id: system.h,v 1.69 2002/02/09 19:49:31 davem Exp $ */
   2 #ifndef __SPARC64_SYSTEM_H
   3 #define __SPARC64_SYSTEM_H
   4
   5 #include <asm/ptrace.h>
   6 #include <asm/processor.h>
   7 #include <asm/visasm.h>
   8
   9 #ifndef __ASSEMBLY__
  10
  11 #include <linux/irqflags.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 nor sun4 :) That's just history. */
  30 #define ARCH_SUN4C_SUN4 0
  31 #define ARCH_SUN4 0
  32
  33 /* These are here in an effort to more fully work around Spitfire Errata
  34  * #51.  Essentially, if a memory barrier occurs soon after a mispredicted
  35  * branch, the chip can stop executing instructions until a trap occurs.
  36  * Therefore, if interrupts are disabled, the chip can hang forever.
  37  *
  38  * It used to be believed that the memory barrier had to be right in the
  39  * delay slot, but a case has been traced recently wherein the memory barrier
  40  * was one instruction after the branch delay slot and the chip still hung.
  41  * The offending sequence was the following in sym_wakeup_done() of the
  42  * sym53c8xx_2 driver:
  43  *
  44  *      call    sym_ccb_from_dsa, 0
  45  *       movge  %icc, 0, %l0
  46  *      brz,pn  %o0, .LL1303
  47  *       mov    %o0, %l2
  48  *      membar  #LoadLoad
  49  *
  50  * The branch has to be mispredicted for the bug to occur.  Therefore, we put
  51  * the memory barrier explicitly into a "branch always, predicted taken"
  52  * delay slot to avoid the problem case.
  53  */
  54 #define membar_safe(type) \
  55 do {    __asm__ __volatile__("ba,pt     %%xcc, 1f\n\t" \
  56                              " membar   " type "\n" \
  57                              "1:\n" \
  58                              : : : "memory"); \
  59 } while (0)
  60
  61 #define mb()    \
  62         membar_safe("#LoadLoad | #LoadStore | #StoreStore | #StoreLoad")
  63 #define rmb()   \
  64         membar_safe("#LoadLoad")
  65 #define wmb()   \
  66         membar_safe("#StoreStore")
  67 #define membar_storeload() \
  68         membar_safe("#StoreLoad")
  69 #define membar_storeload_storestore() \
  70         membar_safe("#StoreLoad | #StoreStore")
  71 #define membar_storeload_loadload() \
  72         membar_safe("#StoreLoad | #LoadLoad")
  73 #define membar_storestore_loadstore() \
  74         membar_safe("#StoreStore | #LoadStore")
  75
  76 #endif
  77
  78 #define nop()           __asm__ __volatile__ ("nop")
  79
  80 #define read_barrier_depends()          do { } while(0)
  81 #define set_mb(__var, __value) \
  82         do { __var = __value; membar_storeload_storestore(); } while(0)
  83
  84 #ifdef CONFIG_SMP
  85 #define smp_mb()        mb()
  86 #define smp_rmb()       rmb()
  87 #define smp_wmb()       wmb()
  88 #define smp_read_barrier_depends()      read_barrier_depends()
  89 #else
  90 #define smp_mb()        __asm__ __volatile__("":::"memory")
  91 #define smp_rmb()       __asm__ __volatile__("":::"memory")
  92 #define smp_wmb()       __asm__ __volatile__("":::"memory")
  93 #define smp_read_barrier_depends()      do { } while(0)
  94 #endif
  95
  96 #define flushi(addr)    __asm__ __volatile__ ("flush %0" : : "r" (addr) : "memory")
  97
  98 #define flushw_all()    __asm__ __volatile__("flushw")
  99
 100 /* Performance counter register access. */
 101 #define read_pcr(__p)  __asm__ __volatile__("rd %%pcr, %0" : "=r" (__p))
 102 #define write_pcr(__p) __asm__ __volatile__("wr %0, 0x0, %%pcr" : : "r" (__p))
 103 #define read_pic(__p)  __asm__ __volatile__("rd %%pic, %0" : "=r" (__p))
 104
 105 /* Blackbird errata workaround.  See commentary in
 106  * arch/sparc64/kernel/smp.c:smp_percpu_timer_interrupt()
 107  * for more information.
 108  */
 109 #define reset_pic()                                             \
 110         __asm__ __volatile__("ba,pt     %xcc, 99f\n\t"          \
 111                              ".align    64\n"                   \
 112                           "99:wr        %g0, 0x0, %pic\n\t"     \
 113                              "rd        %pic, %g0")
 114
 115 #ifndef __ASSEMBLY__
 116
 117 extern void sun_do_break(void);
 118 extern int serial_console;
 119 extern int stop_a_enabled;
 120
 121 static __inline__ int con_is_present(void)
 122 {
 123         return serial_console ? 0 : 1;
 124 }
 125
 126 extern void synchronize_user_stack(void);
 127
 128 extern void __flushw_user(void);
 129 #define flushw_user() __flushw_user()
 130
 131 #define flush_user_windows flushw_user
 132 #define flush_register_windows flushw_all
 133
 134 /* Don't hold the runqueue lock over context switch */
 135 #define __ARCH_WANT_UNLOCKED_CTXSW
 136 #define prepare_arch_switch(next)               \
 137 do {                                            \
 138         flushw_all();                           \
 139 } while (0)
 140
 141         /* See what happens when you design the chip correctly?
 142          *
 143          * We tell gcc we clobber all non-fixed-usage registers except
 144          * for l0/l1.  It will use one for 'next' and the other to hold
 145          * the output value of 'last'.  'next' is not referenced again
 146          * past the invocation of switch_to in the scheduler, so we need
 147          * not preserve it's value.  Hairy, but it lets us remove 2 loads
 148          * and 2 stores in this critical code path.  -DaveM
 149          */
 150 #define EXTRA_CLOBBER ,"%l1"
 151 #define switch_to(prev, next, last)                                     \
 152 do {    if (test_thread_flag(TIF_PERFCTR)) {                            \
 153                 unsigned long __tmp;                                    \
 154                 read_pcr(__tmp);                                        \
 155                 current_thread_info()->pcr_reg = __tmp;                 \
 156                 read_pic(__tmp);                                        \
 157                 current_thread_info()->kernel_cntd0 += (unsigned int)(__tmp);\
 158                 current_thread_info()->kernel_cntd1 += ((__tmp) >> 32); \
 159         }                                                               \
 160         flush_tlb_pending();                                            \
 161         save_and_clear_fpu();                                           \
 162         /* If you are tempted to conditionalize the following */        \
 163         /* so that ASI is only written if it changes, think again. */   \
 164         __asm__ __volatile__("wr %%g0, %0, %%asi"                       \
 165         : : "r" (__thread_flag_byte_ptr(task_thread_info(next))[TI_FLAG_BYTE_CURRENT_DS]));\
 166         trap_block[current_thread_info()->cpu].thread =                 \
 167                 task_thread_info(next);                                 \
 168         __asm__ __volatile__(                                           \
 169         "mov    %%g4, %%g7\n\t"                                         \
 170         "stx    %%i6, [%%sp + 2047 + 0x70]\n\t"                         \
 171         "stx    %%i7, [%%sp + 2047 + 0x78]\n\t"                         \
 172         "rdpr   %%wstate, %%o5\n\t"                                     \
 173         "stx    %%o6, [%%g6 + %3]\n\t"                                  \
 174         "stb    %%o5, [%%g6 + %2]\n\t"                                  \
 175         "rdpr   %%cwp, %%o5\n\t"                                        \
 176         "stb    %%o5, [%%g6 + %5]\n\t"                                  \
 177         "mov    %1, %%g6\n\t"                                           \
 178         "ldub   [%1 + %5], %%g1\n\t"                                    \
 179         "wrpr   %%g1, %%cwp\n\t"                                        \
 180         "ldx    [%%g6 + %3], %%o6\n\t"                                  \
 181         "ldub   [%%g6 + %2], %%o5\n\t"                                  \
 182         "ldub   [%%g6 + %4], %%o7\n\t"                                  \
 183         "wrpr   %%o5, 0x0, %%wstate\n\t"                                \
 184         "ldx    [%%sp + 2047 + 0x70], %%i6\n\t"                         \
 185         "ldx    [%%sp + 2047 + 0x78], %%i7\n\t"                         \
 186         "ldx    [%%g6 + %6], %%g4\n\t"                                  \
 187         "brz,pt %%o7, 1f\n\t"                                           \
 188         " mov   %%g7, %0\n\t"                                           \
 189         "b,a ret_from_syscall\n\t"                                      \
 190         "1:\n\t"                                                        \
 191         : "=&r" (last)                                                  \
 192         : "0" (task_thread_info(next)),                                 \
 193           "i" (TI_WSTATE), "i" (TI_KSP), "i" (TI_NEW_CHILD),            \
 194           "i" (TI_CWP), "i" (TI_TASK)                                   \
 195         : "cc",                                                         \
 196                 "g1", "g2", "g3",                   "g7",               \
 197                       "l2", "l3", "l4", "l5", "l6", "l7",               \
 198           "i0", "i1", "i2", "i3", "i4", "i5",                           \
 199           "o0", "o1", "o2", "o3", "o4", "o5",       "o7" EXTRA_CLOBBER);\
 200         /* If you fuck with this, update ret_from_syscall code too. */  \
 201         if (test_thread_flag(TIF_PERFCTR)) {                            \
 202                 write_pcr(current_thread_info()->pcr_reg);              \
 203                 reset_pic();                                            \
 204         }                                                               \
 205 } while(0)
 206
 207 /*
 208  * On SMP systems, when the scheduler does migration-cost autodetection,
 209  * it needs a way to flush as much of the CPU's caches as possible.
 210  *
 211  * TODO: fill this in!
 212  */
 213 static inline void sched_cacheflush(void)
 214 {
 215 }
 216
 217 static inline unsigned long xchg32(__volatile__ unsigned int *m, unsigned int val)
 218 {
 219         unsigned long tmp1, tmp2;
 220
 221         __asm__ __volatile__(
 222 "       membar          #StoreLoad | #LoadLoad\n"
 223 "       mov             %0, %1\n"
 224 "1:     lduw            [%4], %2\n"
 225 "       cas             [%4], %2, %0\n"
 226 "       cmp             %2, %0\n"
 227 "       bne,a,pn        %%icc, 1b\n"
 228 "        mov            %1, %0\n"
 229 "       membar          #StoreLoad | #StoreStore\n"
 230         : "=&r" (val), "=&r" (tmp1), "=&r" (tmp2)
 231         : "0" (val), "r" (m)
 232         : "cc", "memory");
 233         return val;
 234 }
 235
 236 static inline unsigned long xchg64(__volatile__ unsigned long *m, unsigned long val)
 237 {
 238         unsigned long tmp1, tmp2;
 239
 240         __asm__ __volatile__(
 241 "       membar          #StoreLoad | #LoadLoad\n"
 242 "       mov             %0, %1\n"
 243 "1:     ldx             [%4], %2\n"
 244 "       casx            [%4], %2, %0\n"
 245 "       cmp             %2, %0\n"
 246 "       bne,a,pn        %%xcc, 1b\n"
 247 "        mov            %1, %0\n"
 248 "       membar          #StoreLoad | #StoreStore\n"
 249         : "=&r" (val), "=&r" (tmp1), "=&r" (tmp2)
 250         : "0" (val), "r" (m)
 251         : "cc", "memory");
 252         return val;
 253 }
 254
 255 #define xchg(ptr,x) ((__typeof__(*(ptr)))__xchg((unsigned long)(x),(ptr),sizeof(*(ptr))))
 256
 257 extern void __xchg_called_with_bad_pointer(void);
 258
 259 static __inline__ unsigned long __xchg(unsigned long x, __volatile__ void * ptr,
 260                                        int size)
 261 {
 262         switch (size) {
 263         case 4:
 264                 return xchg32(ptr, x);
 265         case 8:
 266                 return xchg64(ptr, x);
 267         };
 268         __xchg_called_with_bad_pointer();
 269         return x;
 270 }
 271
 272 extern void die_if_kernel(char *str, struct pt_regs *regs) __attribute__ ((noreturn));
 273
 274 /*
 275  * Atomic compare and exchange.  Compare OLD with MEM, if identical,
 276  * store NEW in MEM.  Return the initial value in MEM.  Success is
 277  * indicated by comparing RETURN with OLD.
 278  */
 279
 280 #define __HAVE_ARCH_CMPXCHG 1
 281
 282 static __inline__ unsigned long
 283 __cmpxchg_u32(volatile int *m, int old, int new)
 284 {
 285         __asm__ __volatile__("membar #StoreLoad | #LoadLoad\n"
 286                              "cas [%2], %3, %0\n\t"
 287                              "membar #StoreLoad | #StoreStore"
 288                              : "=&r" (new)
 289                              : "0" (new), "r" (m), "r" (old)
 290                              : "memory");
 291
 292         return new;
 293 }
 294
 295 static __inline__ unsigned long
 296 __cmpxchg_u64(volatile long *m, unsigned long old, unsigned long new)
 297 {
 298         __asm__ __volatile__("membar #StoreLoad | #LoadLoad\n"
 299                              "casx [%2], %3, %0\n\t"
 300                              "membar #StoreLoad | #StoreStore"
 301                              : "=&r" (new)
 302                              : "0" (new), "r" (m), "r" (old)
 303                              : "memory");
 304
 305         return new;
 306 }
 307
 308 /* This function doesn't exist, so you'll get a linker error
 309    if something tries to do an invalid cmpxchg().  */
 310 extern void __cmpxchg_called_with_bad_pointer(void);
 311
 312 static __inline__ unsigned long
 313 __cmpxchg(volatile void *ptr, unsigned long old, unsigned long new, int size)
 314 {
 315         switch (size) {
 316                 case 4:
 317                         return __cmpxchg_u32(ptr, old, new);
 318                 case 8:
 319                         return __cmpxchg_u64(ptr, old, new);
 320         }
 321         __cmpxchg_called_with_bad_pointer();
 322         return old;
 323 }
 324
 325 #define cmpxchg(ptr,o,n)                                                 \
 326   ({                                                                     \
 327      __typeof__(*(ptr)) _o_ = (o);                                       \
 328      __typeof__(*(ptr)) _n_ = (n);                                       \
 329      (__typeof__(*(ptr))) __cmpxchg((ptr), (unsigned long)_o_,           \
 330                                     (unsigned long)_n_, sizeof(*(ptr))); \
 331   })
 332
 333 #endif /* !(__ASSEMBLY__) */
 334
 335 #define arch_align_stack(x) (x)
 336
 337 #endif /* !(__SPARC64_SYSTEM_H) */