First Support on Ginger and OMAP TI
[linux-ginger.git] / arch / sparc / include / asm / system_64.h
blob25e848f0cad783281e5d6c251e02ed7f45d6ebcb
1 #ifndef __SPARC64_SYSTEM_H
2 #define __SPARC64_SYSTEM_H
4 #include <asm/ptrace.h>
5 #include <asm/processor.h>
6 #include <asm/visasm.h>
8 #ifndef __ASSEMBLY__
10 #include <linux/irqflags.h>
11 #include <asm-generic/cmpxchg-local.h>
14 * Sparc (general) CPU types
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 */
27 #define sparc_cpu_model sun4u
29 /* This cannot ever be a sun4c :) That's just history. */
30 #define ARCH_SUN4C 0
32 extern const char *sparc_cpu_type;
33 extern const char *sparc_fpu_type;
34 extern const char *sparc_pmu_type;
36 extern char reboot_command[];
38 /* These are here in an effort to more fully work around Spitfire Errata
39 * #51. Essentially, if a memory barrier occurs soon after a mispredicted
40 * branch, the chip can stop executing instructions until a trap occurs.
41 * Therefore, if interrupts are disabled, the chip can hang forever.
43 * It used to be believed that the memory barrier had to be right in the
44 * delay slot, but a case has been traced recently wherein the memory barrier
45 * was one instruction after the branch delay slot and the chip still hung.
46 * The offending sequence was the following in sym_wakeup_done() of the
47 * sym53c8xx_2 driver:
49 * call sym_ccb_from_dsa, 0
50 * movge %icc, 0, %l0
51 * brz,pn %o0, .LL1303
52 * mov %o0, %l2
53 * membar #LoadLoad
55 * The branch has to be mispredicted for the bug to occur. Therefore, we put
56 * the memory barrier explicitly into a "branch always, predicted taken"
57 * delay slot to avoid the problem case.
59 #define membar_safe(type) \
60 do { __asm__ __volatile__("ba,pt %%xcc, 1f\n\t" \
61 " membar " type "\n" \
62 "1:\n" \
63 : : : "memory"); \
64 } while (0)
66 #define mb() membar_safe("#StoreLoad")
67 #define rmb() __asm__ __volatile__("":::"memory")
68 #define wmb() __asm__ __volatile__("":::"memory")
70 #endif
72 #define nop() __asm__ __volatile__ ("nop")
74 #define read_barrier_depends() do { } while(0)
75 #define set_mb(__var, __value) \
76 do { __var = __value; membar_safe("#StoreLoad"); } while(0)
78 #ifdef CONFIG_SMP
79 #define smp_mb() mb()
80 #define smp_rmb() rmb()
81 #define smp_wmb() wmb()
82 #else
83 #define smp_mb() __asm__ __volatile__("":::"memory")
84 #define smp_rmb() __asm__ __volatile__("":::"memory")
85 #define smp_wmb() __asm__ __volatile__("":::"memory")
86 #endif
88 #define smp_read_barrier_depends() do { } while(0)
90 #define flushi(addr) __asm__ __volatile__ ("flush %0" : : "r" (addr) : "memory")
92 #define flushw_all() __asm__ __volatile__("flushw")
94 /* Performance counter register access. */
95 #define read_pcr(__p) __asm__ __volatile__("rd %%pcr, %0" : "=r" (__p))
96 #define write_pcr(__p) __asm__ __volatile__("wr %0, 0x0, %%pcr" : : "r" (__p))
97 #define read_pic(__p) __asm__ __volatile__("rd %%pic, %0" : "=r" (__p))
99 /* Blackbird errata workaround. See commentary in
100 * arch/sparc64/kernel/smp.c:smp_percpu_timer_interrupt()
101 * for more information.
103 #define write_pic(__p) \
104 __asm__ __volatile__("ba,pt %%xcc, 99f\n\t" \
105 ".align 64\n" \
106 "99:wr %0, 0x0, %%pic\n\t" \
107 "rd %%pic, %%g0" : : "r" (__p))
108 #define reset_pic() write_pic(0)
110 #ifndef __ASSEMBLY__
112 extern void sun_do_break(void);
113 extern int stop_a_enabled;
114 extern int scons_pwroff;
116 extern void fault_in_user_windows(void);
117 extern void synchronize_user_stack(void);
119 extern void __flushw_user(void);
120 #define flushw_user() __flushw_user()
122 #define flush_user_windows flushw_user
123 #define flush_register_windows flushw_all
125 /* Don't hold the runqueue lock over context switch */
126 #define __ARCH_WANT_UNLOCKED_CTXSW
127 #define prepare_arch_switch(next) \
128 do { \
129 flushw_all(); \
130 } while (0)
132 /* See what happens when you design the chip correctly?
134 * We tell gcc we clobber all non-fixed-usage registers except
135 * for l0/l1. It will use one for 'next' and the other to hold
136 * the output value of 'last'. 'next' is not referenced again
137 * past the invocation of switch_to in the scheduler, so we need
138 * not preserve it's value. Hairy, but it lets us remove 2 loads
139 * and 2 stores in this critical code path. -DaveM
141 #define switch_to(prev, next, last) \
142 do { if (test_thread_flag(TIF_PERFCTR)) { \
143 unsigned long __tmp; \
144 read_pcr(__tmp); \
145 current_thread_info()->pcr_reg = __tmp; \
146 read_pic(__tmp); \
147 current_thread_info()->kernel_cntd0 += (unsigned int)(__tmp);\
148 current_thread_info()->kernel_cntd1 += ((__tmp) >> 32); \
150 flush_tlb_pending(); \
151 save_and_clear_fpu(); \
152 /* If you are tempted to conditionalize the following */ \
153 /* so that ASI is only written if it changes, think again. */ \
154 __asm__ __volatile__("wr %%g0, %0, %%asi" \
155 : : "r" (__thread_flag_byte_ptr(task_thread_info(next))[TI_FLAG_BYTE_CURRENT_DS]));\
156 trap_block[current_thread_info()->cpu].thread = \
157 task_thread_info(next); \
158 __asm__ __volatile__( \
159 "mov %%g4, %%g7\n\t" \
160 "stx %%i6, [%%sp + 2047 + 0x70]\n\t" \
161 "stx %%i7, [%%sp + 2047 + 0x78]\n\t" \
162 "rdpr %%wstate, %%o5\n\t" \
163 "stx %%o6, [%%g6 + %6]\n\t" \
164 "stb %%o5, [%%g6 + %5]\n\t" \
165 "rdpr %%cwp, %%o5\n\t" \
166 "stb %%o5, [%%g6 + %8]\n\t" \
167 "wrpr %%g0, 15, %%pil\n\t" \
168 "mov %4, %%g6\n\t" \
169 "ldub [%4 + %8], %%g1\n\t" \
170 "wrpr %%g1, %%cwp\n\t" \
171 "ldx [%%g6 + %6], %%o6\n\t" \
172 "ldub [%%g6 + %5], %%o5\n\t" \
173 "ldub [%%g6 + %7], %%o7\n\t" \
174 "wrpr %%o5, 0x0, %%wstate\n\t" \
175 "ldx [%%sp + 2047 + 0x70], %%i6\n\t" \
176 "ldx [%%sp + 2047 + 0x78], %%i7\n\t" \
177 "ldx [%%g6 + %9], %%g4\n\t" \
178 "wrpr %%g0, 14, %%pil\n\t" \
179 "brz,pt %%o7, switch_to_pc\n\t" \
180 " mov %%g7, %0\n\t" \
181 "sethi %%hi(ret_from_syscall), %%g1\n\t" \
182 "jmpl %%g1 + %%lo(ret_from_syscall), %%g0\n\t" \
183 " nop\n\t" \
184 ".globl switch_to_pc\n\t" \
185 "switch_to_pc:\n\t" \
186 : "=&r" (last), "=r" (current), "=r" (current_thread_info_reg), \
187 "=r" (__local_per_cpu_offset) \
188 : "0" (task_thread_info(next)), \
189 "i" (TI_WSTATE), "i" (TI_KSP), "i" (TI_NEW_CHILD), \
190 "i" (TI_CWP), "i" (TI_TASK) \
191 : "cc", \
192 "g1", "g2", "g3", "g7", \
193 "l1", "l2", "l3", "l4", "l5", "l6", "l7", \
194 "i0", "i1", "i2", "i3", "i4", "i5", \
195 "o0", "o1", "o2", "o3", "o4", "o5", "o7"); \
196 /* If you fuck with this, update ret_from_syscall code too. */ \
197 if (test_thread_flag(TIF_PERFCTR)) { \
198 write_pcr(current_thread_info()->pcr_reg); \
199 reset_pic(); \
201 } while(0)
203 static inline unsigned long xchg32(__volatile__ unsigned int *m, unsigned int val)
205 unsigned long tmp1, tmp2;
207 __asm__ __volatile__(
208 " mov %0, %1\n"
209 "1: lduw [%4], %2\n"
210 " cas [%4], %2, %0\n"
211 " cmp %2, %0\n"
212 " bne,a,pn %%icc, 1b\n"
213 " mov %1, %0\n"
214 : "=&r" (val), "=&r" (tmp1), "=&r" (tmp2)
215 : "0" (val), "r" (m)
216 : "cc", "memory");
217 return val;
220 static inline unsigned long xchg64(__volatile__ unsigned long *m, unsigned long val)
222 unsigned long tmp1, tmp2;
224 __asm__ __volatile__(
225 " mov %0, %1\n"
226 "1: ldx [%4], %2\n"
227 " casx [%4], %2, %0\n"
228 " cmp %2, %0\n"
229 " bne,a,pn %%xcc, 1b\n"
230 " mov %1, %0\n"
231 : "=&r" (val), "=&r" (tmp1), "=&r" (tmp2)
232 : "0" (val), "r" (m)
233 : "cc", "memory");
234 return val;
237 #define xchg(ptr,x) ((__typeof__(*(ptr)))__xchg((unsigned long)(x),(ptr),sizeof(*(ptr))))
239 extern void __xchg_called_with_bad_pointer(void);
241 static inline unsigned long __xchg(unsigned long x, __volatile__ void * ptr,
242 int size)
244 switch (size) {
245 case 4:
246 return xchg32(ptr, x);
247 case 8:
248 return xchg64(ptr, x);
250 __xchg_called_with_bad_pointer();
251 return x;
254 extern void die_if_kernel(char *str, struct pt_regs *regs) __attribute__ ((noreturn));
257 * Atomic compare and exchange. Compare OLD with MEM, if identical,
258 * store NEW in MEM. Return the initial value in MEM. Success is
259 * indicated by comparing RETURN with OLD.
262 #define __HAVE_ARCH_CMPXCHG 1
264 static inline unsigned long
265 __cmpxchg_u32(volatile int *m, int old, int new)
267 __asm__ __volatile__("cas [%2], %3, %0"
268 : "=&r" (new)
269 : "0" (new), "r" (m), "r" (old)
270 : "memory");
272 return new;
275 static inline unsigned long
276 __cmpxchg_u64(volatile long *m, unsigned long old, unsigned long new)
278 __asm__ __volatile__("casx [%2], %3, %0"
279 : "=&r" (new)
280 : "0" (new), "r" (m), "r" (old)
281 : "memory");
283 return new;
286 /* This function doesn't exist, so you'll get a linker error
287 if something tries to do an invalid cmpxchg(). */
288 extern void __cmpxchg_called_with_bad_pointer(void);
290 static inline unsigned long
291 __cmpxchg(volatile void *ptr, unsigned long old, unsigned long new, int size)
293 switch (size) {
294 case 4:
295 return __cmpxchg_u32(ptr, old, new);
296 case 8:
297 return __cmpxchg_u64(ptr, old, new);
299 __cmpxchg_called_with_bad_pointer();
300 return old;
303 #define cmpxchg(ptr,o,n) \
304 ({ \
305 __typeof__(*(ptr)) _o_ = (o); \
306 __typeof__(*(ptr)) _n_ = (n); \
307 (__typeof__(*(ptr))) __cmpxchg((ptr), (unsigned long)_o_, \
308 (unsigned long)_n_, sizeof(*(ptr))); \
312 * cmpxchg_local and cmpxchg64_local are atomic wrt current CPU. Always make
313 * them available.
316 static inline unsigned long __cmpxchg_local(volatile void *ptr,
317 unsigned long old,
318 unsigned long new, int size)
320 switch (size) {
321 case 4:
322 case 8: return __cmpxchg(ptr, old, new, size);
323 default:
324 return __cmpxchg_local_generic(ptr, old, new, size);
327 return old;
330 #define cmpxchg_local(ptr, o, n) \
331 ((__typeof__(*(ptr)))__cmpxchg_local((ptr), (unsigned long)(o), \
332 (unsigned long)(n), sizeof(*(ptr))))
333 #define cmpxchg64_local(ptr, o, n) \
334 ({ \
335 BUILD_BUG_ON(sizeof(*(ptr)) != 8); \
336 cmpxchg_local((ptr), (o), (n)); \
339 #endif /* !(__ASSEMBLY__) */
341 #define arch_align_stack(x) (x)
343 #endif /* !(__SPARC64_SYSTEM_H) */