Merge master.kernel.org:/pub/scm/linux/kernel/git/gregkh/aoe-2.6
[sfrench/cifs-2.6.git] / include / asm-x86_64 / system.h
1 #ifndef __ASM_SYSTEM_H
2 #define __ASM_SYSTEM_H
3
4 #include <linux/config.h>
5 #include <linux/kernel.h>
6 #include <asm/segment.h>
7
8 #ifdef __KERNEL__
9
10 #ifdef CONFIG_SMP
11 #define LOCK_PREFIX "lock ; "
12 #else
13 #define LOCK_PREFIX ""
14 #endif
15
16 #define __STR(x) #x
17 #define STR(x) __STR(x)
18
19 #define __SAVE(reg,offset) "movq %%" #reg ",(14-" #offset ")*8(%%rsp)\n\t"
20 #define __RESTORE(reg,offset) "movq (14-" #offset ")*8(%%rsp),%%" #reg "\n\t"
21
22 /* frame pointer must be last for get_wchan */
23 #define SAVE_CONTEXT    "pushq %%rbp ; movq %%rsi,%%rbp\n\t"
24 #define RESTORE_CONTEXT "movq %%rbp,%%rsi ; popq %%rbp\n\t"
25
26 #define __EXTRA_CLOBBER  \
27         ,"rcx","rbx","rdx","r8","r9","r10","r11","r12","r13","r14","r15"
28
29 #define switch_to(prev,next,last) \
30         asm volatile(SAVE_CONTEXT                                                   \
31                      "movq %%rsp,%P[threadrsp](%[prev])\n\t" /* save RSP */       \
32                      "movq %P[threadrsp](%[next]),%%rsp\n\t" /* restore RSP */    \
33                      "call __switch_to\n\t"                                       \
34                      ".globl thread_return\n"                                   \
35                      "thread_return:\n\t"                                           \
36                      "movq %%gs:%P[pda_pcurrent],%%rsi\n\t"                       \
37                      "movq %P[thread_info](%%rsi),%%r8\n\t"                       \
38                      LOCK "btr  %[tif_fork],%P[ti_flags](%%r8)\n\t"               \
39                      "movq %%rax,%%rdi\n\t"                                       \
40                      "jc   ret_from_fork\n\t"                                     \
41                      RESTORE_CONTEXT                                                \
42                      : "=a" (last)                                                \
43                      : [next] "S" (next), [prev] "D" (prev),                      \
44                        [threadrsp] "i" (offsetof(struct task_struct, thread.rsp)), \
45                        [ti_flags] "i" (offsetof(struct thread_info, flags)),\
46                        [tif_fork] "i" (TIF_FORK),                         \
47                        [thread_info] "i" (offsetof(struct task_struct, thread_info)), \
48                        [pda_pcurrent] "i" (offsetof(struct x8664_pda, pcurrent))   \
49                      : "memory", "cc" __EXTRA_CLOBBER)
50     
51 extern void load_gs_index(unsigned); 
52
53 /*
54  * Load a segment. Fall back on loading the zero
55  * segment if something goes wrong..
56  */
57 #define loadsegment(seg,value)  \
58         asm volatile("\n"                       \
59                 "1:\t"                          \
60                 "movl %k0,%%" #seg "\n"         \
61                 "2:\n"                          \
62                 ".section .fixup,\"ax\"\n"      \
63                 "3:\t"                          \
64                 "movl %1,%%" #seg "\n\t"        \
65                 "jmp 2b\n"                      \
66                 ".previous\n"                   \
67                 ".section __ex_table,\"a\"\n\t" \
68                 ".align 8\n\t"                  \
69                 ".quad 1b,3b\n"                 \
70                 ".previous"                     \
71                 : :"r" (value), "r" (0))
72
73 #ifdef __KERNEL__
74 struct alt_instr { 
75         __u8 *instr;            /* original instruction */
76         __u8 *replacement;
77         __u8  cpuid;            /* cpuid bit set for replacement */
78         __u8  instrlen;         /* length of original instruction */
79         __u8  replacementlen;   /* length of new instruction, <= instrlen */ 
80         __u8  pad[5];
81 }; 
82 #endif
83
84 /*
85  * Alternative instructions for different CPU types or capabilities.
86  * 
87  * This allows to use optimized instructions even on generic binary
88  * kernels.
89  * 
90  * length of oldinstr must be longer or equal the length of newinstr
91  * It can be padded with nops as needed.
92  * 
93  * For non barrier like inlines please define new variants
94  * without volatile and memory clobber.
95  */
96 #define alternative(oldinstr, newinstr, feature)        \
97         asm volatile ("661:\n\t" oldinstr "\n662:\n"                 \
98                       ".section .altinstructions,\"a\"\n"            \
99                       "  .align 8\n"                                   \
100                       "  .quad 661b\n"            /* label */          \
101                       "  .quad 663f\n"            /* new instruction */ \
102                       "  .byte %c0\n"             /* feature bit */    \
103                       "  .byte 662b-661b\n"       /* sourcelen */      \
104                       "  .byte 664f-663f\n"       /* replacementlen */ \
105                       ".previous\n"                                     \
106                       ".section .altinstr_replacement,\"ax\"\n"         \
107                       "663:\n\t" newinstr "\n664:\n"   /* replacement */ \
108                       ".previous" :: "i" (feature) : "memory")  
109
110 /*
111  * Alternative inline assembly with input.
112  * 
113  * Peculiarities:
114  * No memory clobber here. 
115  * Argument numbers start with 1.
116  * Best is to use constraints that are fixed size (like (%1) ... "r")
117  * If you use variable sized constraints like "m" or "g" in the 
118  * replacement make sure to pad to the worst case length.
119  */
120 #define alternative_input(oldinstr, newinstr, feature, input...)        \
121         asm volatile ("661:\n\t" oldinstr "\n662:\n"                    \
122                       ".section .altinstructions,\"a\"\n"               \
123                       "  .align 8\n"                                    \
124                       "  .quad 661b\n"            /* label */           \
125                       "  .quad 663f\n"            /* new instruction */ \
126                       "  .byte %c0\n"             /* feature bit */     \
127                       "  .byte 662b-661b\n"       /* sourcelen */       \
128                       "  .byte 664f-663f\n"       /* replacementlen */  \
129                       ".previous\n"                                     \
130                       ".section .altinstr_replacement,\"ax\"\n"         \
131                       "663:\n\t" newinstr "\n664:\n"   /* replacement */ \
132                       ".previous" :: "i" (feature), ##input)
133
134 /* Like alternative_input, but with a single output argument */
135 #define alternative_io(oldinstr, newinstr, feature, output, input...) \
136         asm volatile ("661:\n\t" oldinstr "\n662:\n"                    \
137                       ".section .altinstructions,\"a\"\n"               \
138                       "  .align 8\n"                                    \
139                       "  .quad 661b\n"            /* label */           \
140                       "  .quad 663f\n"            /* new instruction */ \
141                       "  .byte %c[feat]\n"        /* feature bit */     \
142                       "  .byte 662b-661b\n"       /* sourcelen */       \
143                       "  .byte 664f-663f\n"       /* replacementlen */  \
144                       ".previous\n"                                     \
145                       ".section .altinstr_replacement,\"ax\"\n"         \
146                       "663:\n\t" newinstr "\n664:\n"   /* replacement */ \
147                       ".previous" : output : [feat] "i" (feature), ##input)
148
149 /*
150  * Clear and set 'TS' bit respectively
151  */
152 #define clts() __asm__ __volatile__ ("clts")
153
154 static inline unsigned long read_cr0(void)
155
156         unsigned long cr0;
157         asm volatile("movq %%cr0,%0" : "=r" (cr0));
158         return cr0;
159
160
161 static inline void write_cr0(unsigned long val) 
162
163         asm volatile("movq %0,%%cr0" :: "r" (val));
164
165
166 static inline unsigned long read_cr3(void)
167
168         unsigned long cr3;
169         asm("movq %%cr3,%0" : "=r" (cr3));
170         return cr3;
171
172
173 static inline unsigned long read_cr4(void)
174
175         unsigned long cr4;
176         asm("movq %%cr4,%0" : "=r" (cr4));
177         return cr4;
178
179
180 static inline void write_cr4(unsigned long val)
181
182         asm volatile("movq %0,%%cr4" :: "r" (val));
183
184
185 #define stts() write_cr0(8 | read_cr0())
186
187 #define wbinvd() \
188         __asm__ __volatile__ ("wbinvd": : :"memory");
189
190 /*
191  * On SMP systems, when the scheduler does migration-cost autodetection,
192  * it needs a way to flush as much of the CPU's caches as possible.
193  */
194 static inline void sched_cacheflush(void)
195 {
196         wbinvd();
197 }
198
199 #endif  /* __KERNEL__ */
200
201 #define nop() __asm__ __volatile__ ("nop")
202
203 #define xchg(ptr,v) ((__typeof__(*(ptr)))__xchg((unsigned long)(v),(ptr),sizeof(*(ptr))))
204
205 #define tas(ptr) (xchg((ptr),1))
206
207 #define __xg(x) ((volatile long *)(x))
208
209 static inline void set_64bit(volatile unsigned long *ptr, unsigned long val)
210 {
211         *ptr = val;
212 }
213
214 #define _set_64bit set_64bit
215
216 /*
217  * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
218  * Note 2: xchg has side effect, so that attribute volatile is necessary,
219  *        but generally the primitive is invalid, *ptr is output argument. --ANK
220  */
221 static inline unsigned long __xchg(unsigned long x, volatile void * ptr, int size)
222 {
223         switch (size) {
224                 case 1:
225                         __asm__ __volatile__("xchgb %b0,%1"
226                                 :"=q" (x)
227                                 :"m" (*__xg(ptr)), "0" (x)
228                                 :"memory");
229                         break;
230                 case 2:
231                         __asm__ __volatile__("xchgw %w0,%1"
232                                 :"=r" (x)
233                                 :"m" (*__xg(ptr)), "0" (x)
234                                 :"memory");
235                         break;
236                 case 4:
237                         __asm__ __volatile__("xchgl %k0,%1"
238                                 :"=r" (x)
239                                 :"m" (*__xg(ptr)), "0" (x)
240                                 :"memory");
241                         break;
242                 case 8:
243                         __asm__ __volatile__("xchgq %0,%1"
244                                 :"=r" (x)
245                                 :"m" (*__xg(ptr)), "0" (x)
246                                 :"memory");
247                         break;
248         }
249         return x;
250 }
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 __cmpxchg(volatile void *ptr, unsigned long old,
261                                       unsigned long new, int size)
262 {
263         unsigned long prev;
264         switch (size) {
265         case 1:
266                 __asm__ __volatile__(LOCK_PREFIX "cmpxchgb %b1,%2"
267                                      : "=a"(prev)
268                                      : "q"(new), "m"(*__xg(ptr)), "0"(old)
269                                      : "memory");
270                 return prev;
271         case 2:
272                 __asm__ __volatile__(LOCK_PREFIX "cmpxchgw %w1,%2"
273                                      : "=a"(prev)
274                                      : "r"(new), "m"(*__xg(ptr)), "0"(old)
275                                      : "memory");
276                 return prev;
277         case 4:
278                 __asm__ __volatile__(LOCK_PREFIX "cmpxchgl %k1,%2"
279                                      : "=a"(prev)
280                                      : "r"(new), "m"(*__xg(ptr)), "0"(old)
281                                      : "memory");
282                 return prev;
283         case 8:
284                 __asm__ __volatile__(LOCK_PREFIX "cmpxchgq %1,%2"
285                                      : "=a"(prev)
286                                      : "r"(new), "m"(*__xg(ptr)), "0"(old)
287                                      : "memory");
288                 return prev;
289         }
290         return old;
291 }
292
293 #define cmpxchg(ptr,o,n)\
294         ((__typeof__(*(ptr)))__cmpxchg((ptr),(unsigned long)(o),\
295                                         (unsigned long)(n),sizeof(*(ptr))))
296
297 #ifdef CONFIG_SMP
298 #define smp_mb()        mb()
299 #define smp_rmb()       rmb()
300 #define smp_wmb()       wmb()
301 #define smp_read_barrier_depends()      do {} while(0)
302 #else
303 #define smp_mb()        barrier()
304 #define smp_rmb()       barrier()
305 #define smp_wmb()       barrier()
306 #define smp_read_barrier_depends()      do {} while(0)
307 #endif
308
309     
310 /*
311  * Force strict CPU ordering.
312  * And yes, this is required on UP too when we're talking
313  * to devices.
314  */
315 #define mb()    asm volatile("mfence":::"memory")
316 #define rmb()   asm volatile("lfence":::"memory")
317
318 #ifdef CONFIG_UNORDERED_IO
319 #define wmb()   asm volatile("sfence" ::: "memory")
320 #else
321 #define wmb()   asm volatile("" ::: "memory")
322 #endif
323 #define read_barrier_depends()  do {} while(0)
324 #define set_mb(var, value) do { (void) xchg(&var, value); } while (0)
325 #define set_wmb(var, value) do { var = value; wmb(); } while (0)
326
327 #define warn_if_not_ulong(x) do { unsigned long foo; (void) (&(x) == &foo); } while (0)
328
329 /* interrupt control.. */
330 #define local_save_flags(x)     do { warn_if_not_ulong(x); __asm__ __volatile__("# save_flags \n\t pushfq ; popq %q0":"=g" (x): /* no input */ :"memory"); } while (0)
331 #define local_irq_restore(x)    __asm__ __volatile__("# restore_flags \n\t pushq %0 ; popfq": /* no output */ :"g" (x):"memory", "cc")
332
333 #ifdef CONFIG_X86_VSMP
334 /* Interrupt control for VSMP  architecture */
335 #define local_irq_disable()     do { unsigned long flags; local_save_flags(flags); local_irq_restore((flags & ~(1 << 9)) | (1 << 18)); } while (0)
336 #define local_irq_enable()      do { unsigned long flags; local_save_flags(flags); local_irq_restore((flags | (1 << 9)) & ~(1 << 18)); } while (0)
337
338 #define irqs_disabled()                                 \
339 ({                                                      \
340         unsigned long flags;                            \
341         local_save_flags(flags);                        \
342         (flags & (1<<18)) || !(flags & (1<<9));         \
343 })
344
345 /* For spinlocks etc */
346 #define local_irq_save(x)       do { local_save_flags(x); local_irq_restore((x & ~(1 << 9)) | (1 << 18)); } while (0)
347 #else  /* CONFIG_X86_VSMP */
348 #define local_irq_disable()     __asm__ __volatile__("cli": : :"memory")
349 #define local_irq_enable()      __asm__ __volatile__("sti": : :"memory")
350
351 #define irqs_disabled()                 \
352 ({                                      \
353         unsigned long flags;            \
354         local_save_flags(flags);        \
355         !(flags & (1<<9));              \
356 })
357
358 /* For spinlocks etc */
359 #define local_irq_save(x)       do { warn_if_not_ulong(x); __asm__ __volatile__("# local_irq_save \n\t pushfq ; popq %0 ; cli":"=g" (x): /* no input */ :"memory"); } while (0)
360 #endif
361
362 /* used in the idle loop; sti takes one instruction cycle to complete */
363 #define safe_halt()             __asm__ __volatile__("sti; hlt": : :"memory")
364 /* used when interrupts are already enabled or to shutdown the processor */
365 #define halt()                  __asm__ __volatile__("hlt": : :"memory")
366
367 void cpu_idle_wait(void);
368
369 extern unsigned long arch_align_stack(unsigned long sp);
370
371 #endif