arch/x86/include/asm/percpu.h

   1 #ifndef _ASM_X86_PERCPU_H
   2 #define _ASM_X86_PERCPU_H
   3
   4 #ifdef CONFIG_X86_64
   5 #define __percpu_seg            gs
   6 #define __percpu_mov_op         movq
   7 #else
   8 #define __percpu_seg            fs
   9 #define __percpu_mov_op         movl
  10 #endif
  11
  12 #ifdef __ASSEMBLY__
  13
  14 /*
  15  * PER_CPU finds an address of a per-cpu variable.
  16  *
  17  * Args:
  18  *    var - variable name
  19  *    reg - 32bit register
  20  *
  21  * The resulting address is stored in the "reg" argument.
  22  *
  23  * Example:
  24  *    PER_CPU(cpu_gdt_descr, %ebx)
  25  */
  26 #ifdef CONFIG_SMP
  27 #define PER_CPU(var, reg)                                               \
  28         __percpu_mov_op %__percpu_seg:this_cpu_off, reg;                \
  29         lea var(reg), reg
  30 #define PER_CPU_VAR(var)        %__percpu_seg:var
  31 #else /* ! SMP */
  32 #define PER_CPU(var, reg)       __percpu_mov_op $var, reg
  33 #define PER_CPU_VAR(var)        var
  34 #endif  /* SMP */
  35
  36 #ifdef CONFIG_X86_64_SMP
  37 #define INIT_PER_CPU_VAR(var)  init_per_cpu__##var
  38 #else
  39 #define INIT_PER_CPU_VAR(var)  var
  40 #endif
  41
  42 #else /* ...!ASSEMBLY */
  43
  44 #include <linux/kernel.h>
  45 #include <linux/stringify.h>
  46
  47 #ifdef CONFIG_SMP
  48 #define __percpu_arg(x)         "%%"__stringify(__percpu_seg)":%P" #x
  49 #define __my_cpu_offset         percpu_read(this_cpu_off)
  50 #else
  51 #define __percpu_arg(x)         "%P" #x
  52 #endif
  53
  54 /*
  55  * Initialized pointers to per-cpu variables needed for the boot
  56  * processor need to use these macros to get the proper address
  57  * offset from __per_cpu_load on SMP.
  58  *
  59  * There also must be an entry in vmlinux_64.lds.S
  60  */
  61 #define DECLARE_INIT_PER_CPU(var) \
  62        extern typeof(var) init_per_cpu_var(var)
  63
  64 #ifdef CONFIG_X86_64_SMP
  65 #define init_per_cpu_var(var)  init_per_cpu__##var
  66 #else
  67 #define init_per_cpu_var(var)  var
  68 #endif
  69
  70 /* For arch-specific code, we can use direct single-insn ops (they
  71  * don't give an lvalue though). */
  72 extern void __bad_percpu_size(void);
  73
  74 #define percpu_to_op(op, var, val)                      \
  75 do {                                                    \
  76         typedef typeof(var) pto_T__;                    \
  77         if (0) {                                        \
  78                 pto_T__ pto_tmp__;                      \
  79                 pto_tmp__ = (val);                      \
  80         }                                               \
  81         switch (sizeof(var)) {                          \
  82         case 1:                                         \
  83                 asm(op "b %1,"__percpu_arg(0)           \
  84                     : "+m" (var)                        \
  85                     : "qi" ((pto_T__)(val)));           \
  86                 break;                                  \
  87         case 2:                                         \
  88                 asm(op "w %1,"__percpu_arg(0)           \
  89                     : "+m" (var)                        \
  90                     : "ri" ((pto_T__)(val)));           \
  91                 break;                                  \
  92         case 4:                                         \
  93                 asm(op "l %1,"__percpu_arg(0)           \
  94                     : "+m" (var)                        \
  95                     : "ri" ((pto_T__)(val)));           \
  96                 break;                                  \
  97         case 8:                                         \
  98                 asm(op "q %1,"__percpu_arg(0)           \
  99                     : "+m" (var)                        \
 100                     : "re" ((pto_T__)(val)));           \
 101                 break;                                  \
 102         default: __bad_percpu_size();                   \
 103         }                                               \
 104 } while (0)
 105
 106 #define percpu_from_op(op, var, constraint)             \
 107 ({                                                      \
 108         typeof(var) pfo_ret__;                          \
 109         switch (sizeof(var)) {                          \
 110         case 1:                                         \
 111                 asm(op "b "__percpu_arg(1)",%0"         \
 112                     : "=q" (pfo_ret__)                  \
 113                     : constraint);                      \
 114                 break;                                  \
 115         case 2:                                         \
 116                 asm(op "w "__percpu_arg(1)",%0"         \
 117                     : "=r" (pfo_ret__)                  \
 118                     : constraint);                      \
 119                 break;                                  \
 120         case 4:                                         \
 121                 asm(op "l "__percpu_arg(1)",%0"         \
 122                     : "=r" (pfo_ret__)                  \
 123                     : constraint);                      \
 124                 break;                                  \
 125         case 8:                                         \
 126                 asm(op "q "__percpu_arg(1)",%0"         \
 127                     : "=r" (pfo_ret__)                  \
 128                     : constraint);                      \
 129                 break;                                  \
 130         default: __bad_percpu_size();                   \
 131         }                                               \
 132         pfo_ret__;                                      \
 133 })
 134
 135 /*
 136  * percpu_read() makes gcc load the percpu variable every time it is
 137  * accessed while percpu_read_stable() allows the value to be cached.
 138  * percpu_read_stable() is more efficient and can be used if its value
 139  * is guaranteed to be valid across cpus.  The current users include
 140  * get_current() and get_thread_info() both of which are actually
 141  * per-thread variables implemented as per-cpu variables and thus
 142  * stable for the duration of the respective task.
 143  */
 144 #define percpu_read(var)                percpu_from_op("mov", var, "m" (var))
 145 #define percpu_read_stable(var)         percpu_from_op("mov", var, "p" (&(var)))
 146 #define percpu_write(var, val)          percpu_to_op("mov", var, val)
 147 #define percpu_add(var, val)            percpu_to_op("add", var, val)
 148 #define percpu_sub(var, val)            percpu_to_op("sub", var, val)
 149 #define percpu_and(var, val)            percpu_to_op("and", var, val)
 150 #define percpu_or(var, val)             percpu_to_op("or", var, val)
 151 #define percpu_xor(var, val)            percpu_to_op("xor", var, val)
 152
 153 #define __this_cpu_read_1(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
 154 #define __this_cpu_read_2(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
 155 #define __this_cpu_read_4(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
 156
 157 #define __this_cpu_write_1(pcp, val)    percpu_to_op("mov", (pcp), val)
 158 #define __this_cpu_write_2(pcp, val)    percpu_to_op("mov", (pcp), val)
 159 #define __this_cpu_write_4(pcp, val)    percpu_to_op("mov", (pcp), val)
 160 #define __this_cpu_add_1(pcp, val)      percpu_to_op("add", (pcp), val)
 161 #define __this_cpu_add_2(pcp, val)      percpu_to_op("add", (pcp), val)
 162 #define __this_cpu_add_4(pcp, val)      percpu_to_op("add", (pcp), val)
 163 #define __this_cpu_and_1(pcp, val)      percpu_to_op("and", (pcp), val)
 164 #define __this_cpu_and_2(pcp, val)      percpu_to_op("and", (pcp), val)
 165 #define __this_cpu_and_4(pcp, val)      percpu_to_op("and", (pcp), val)
 166 #define __this_cpu_or_1(pcp, val)       percpu_to_op("or", (pcp), val)
 167 #define __this_cpu_or_2(pcp, val)       percpu_to_op("or", (pcp), val)
 168 #define __this_cpu_or_4(pcp, val)       percpu_to_op("or", (pcp), val)
 169 #define __this_cpu_xor_1(pcp, val)      percpu_to_op("xor", (pcp), val)
 170 #define __this_cpu_xor_2(pcp, val)      percpu_to_op("xor", (pcp), val)
 171 #define __this_cpu_xor_4(pcp, val)      percpu_to_op("xor", (pcp), val)
 172
 173 #define this_cpu_read_1(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
 174 #define this_cpu_read_2(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
 175 #define this_cpu_read_4(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
 176 #define this_cpu_write_1(pcp, val)      percpu_to_op("mov", (pcp), val)
 177 #define this_cpu_write_2(pcp, val)      percpu_to_op("mov", (pcp), val)
 178 #define this_cpu_write_4(pcp, val)      percpu_to_op("mov", (pcp), val)
 179 #define this_cpu_add_1(pcp, val)        percpu_to_op("add", (pcp), val)
 180 #define this_cpu_add_2(pcp, val)        percpu_to_op("add", (pcp), val)
 181 #define this_cpu_add_4(pcp, val)        percpu_to_op("add", (pcp), val)
 182 #define this_cpu_and_1(pcp, val)        percpu_to_op("and", (pcp), val)
 183 #define this_cpu_and_2(pcp, val)        percpu_to_op("and", (pcp), val)
 184 #define this_cpu_and_4(pcp, val)        percpu_to_op("and", (pcp), val)
 185 #define this_cpu_or_1(pcp, val)         percpu_to_op("or", (pcp), val)
 186 #define this_cpu_or_2(pcp, val)         percpu_to_op("or", (pcp), val)
 187 #define this_cpu_or_4(pcp, val)         percpu_to_op("or", (pcp), val)
 188 #define this_cpu_xor_1(pcp, val)        percpu_to_op("xor", (pcp), val)
 189 #define this_cpu_xor_2(pcp, val)        percpu_to_op("xor", (pcp), val)
 190 #define this_cpu_xor_4(pcp, val)        percpu_to_op("xor", (pcp), val)
 191
 192 #define irqsafe_cpu_add_1(pcp, val)     percpu_to_op("add", (pcp), val)
 193 #define irqsafe_cpu_add_2(pcp, val)     percpu_to_op("add", (pcp), val)
 194 #define irqsafe_cpu_add_4(pcp, val)     percpu_to_op("add", (pcp), val)
 195 #define irqsafe_cpu_and_1(pcp, val)     percpu_to_op("and", (pcp), val)
 196 #define irqsafe_cpu_and_2(pcp, val)     percpu_to_op("and", (pcp), val)
 197 #define irqsafe_cpu_and_4(pcp, val)     percpu_to_op("and", (pcp), val)
 198 #define irqsafe_cpu_or_1(pcp, val)      percpu_to_op("or", (pcp), val)
 199 #define irqsafe_cpu_or_2(pcp, val)      percpu_to_op("or", (pcp), val)
 200 #define irqsafe_cpu_or_4(pcp, val)      percpu_to_op("or", (pcp), val)
 201 #define irqsafe_cpu_xor_1(pcp, val)     percpu_to_op("xor", (pcp), val)
 202 #define irqsafe_cpu_xor_2(pcp, val)     percpu_to_op("xor", (pcp), val)
 203 #define irqsafe_cpu_xor_4(pcp, val)     percpu_to_op("xor", (pcp), val)
 204
 205 /*
 206  * Per cpu atomic 64 bit operations are only available under 64 bit.
 207  * 32 bit must fall back to generic operations.
 208  */
 209 #ifdef CONFIG_X86_64
 210 #define __this_cpu_read_8(pcp)          percpu_from_op("mov", (pcp), "m"(pcp))
 211 #define __this_cpu_write_8(pcp, val)    percpu_to_op("mov", (pcp), val)
 212 #define __this_cpu_add_8(pcp, val)      percpu_to_op("add", (pcp), val)
 213 #define __this_cpu_and_8(pcp, val)      percpu_to_op("and", (pcp), val)
 214 #define __this_cpu_or_8(pcp, val)       percpu_to_op("or", (pcp), val)
 215 #define __this_cpu_xor_8(pcp, val)      percpu_to_op("xor", (pcp), val)
 216
 217 #define this_cpu_read_8(pcp)            percpu_from_op("mov", (pcp), "m"(pcp))
 218 #define this_cpu_write_8(pcp, val)      percpu_to_op("mov", (pcp), val)
 219 #define this_cpu_add_8(pcp, val)        percpu_to_op("add", (pcp), val)
 220 #define this_cpu_and_8(pcp, val)        percpu_to_op("and", (pcp), val)
 221 #define this_cpu_or_8(pcp, val)         percpu_to_op("or", (pcp), val)
 222 #define this_cpu_xor_8(pcp, val)        percpu_to_op("xor", (pcp), val)
 223
 224 #define irqsafe_cpu_add_8(pcp, val)     percpu_to_op("add", (pcp), val)
 225 #define irqsafe_cpu_and_8(pcp, val)     percpu_to_op("and", (pcp), val)
 226 #define irqsafe_cpu_or_8(pcp, val)      percpu_to_op("or", (pcp), val)
 227 #define irqsafe_cpu_xor_8(pcp, val)     percpu_to_op("xor", (pcp), val)
 228
 229 #endif
 230
 231 /* This is not atomic against other CPUs -- CPU preemption needs to be off */
 232 #define x86_test_and_clear_bit_percpu(bit, var)                         \
 233 ({                                                                      \
 234         int old__;                                                      \
 235         asm volatile("btr %2,"__percpu_arg(1)"\n\tsbbl %0,%0"           \
 236                      : "=r" (old__), "+m" (var)                         \
 237                      : "dIr" (bit));                                    \
 238         old__;                                                          \
 239 })
 240
 241 #include <asm-generic/percpu.h>
 242
 243 /* We can use this directly for local CPU (faster). */
 244 DECLARE_PER_CPU(unsigned long, this_cpu_off);
 245
 246 #endif /* !__ASSEMBLY__ */
 247
 248 #ifdef CONFIG_SMP
 249
 250 /*
 251  * Define the "EARLY_PER_CPU" macros.  These are used for some per_cpu
 252  * variables that are initialized and accessed before there are per_cpu
 253  * areas allocated.
 254  */
 255
 256 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)                  \
 257         DEFINE_PER_CPU(_type, _name) = _initvalue;                      \
 258         __typeof__(_type) _name##_early_map[NR_CPUS] __initdata =       \
 259                                 { [0 ... NR_CPUS-1] = _initvalue };     \
 260         __typeof__(_type) *_name##_early_ptr __refdata = _name##_early_map
 261
 262 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)                      \
 263         EXPORT_PER_CPU_SYMBOL(_name)
 264
 265 #define DECLARE_EARLY_PER_CPU(_type, _name)                     \
 266         DECLARE_PER_CPU(_type, _name);                          \
 267         extern __typeof__(_type) *_name##_early_ptr;            \
 268         extern __typeof__(_type)  _name##_early_map[]
 269
 270 #define early_per_cpu_ptr(_name) (_name##_early_ptr)
 271 #define early_per_cpu_map(_name, _idx) (_name##_early_map[_idx])
 272 #define early_per_cpu(_name, _cpu)                              \
 273         *(early_per_cpu_ptr(_name) ?                            \
 274                 &early_per_cpu_ptr(_name)[_cpu] :               \
 275                 &per_cpu(_name, _cpu))
 276
 277 #else   /* !CONFIG_SMP */
 278 #define DEFINE_EARLY_PER_CPU(_type, _name, _initvalue)          \
 279         DEFINE_PER_CPU(_type, _name) = _initvalue
 280
 281 #define EXPORT_EARLY_PER_CPU_SYMBOL(_name)                      \
 282         EXPORT_PER_CPU_SYMBOL(_name)
 283
 284 #define DECLARE_EARLY_PER_CPU(_type, _name)                     \
 285         DECLARE_PER_CPU(_type, _name)
 286
 287 #define early_per_cpu(_name, _cpu) per_cpu(_name, _cpu)
 288 #define early_per_cpu_ptr(_name) NULL
 289 /* no early_per_cpu_map() */
 290
 291 #endif  /* !CONFIG_SMP */
 292
 293 #endif /* _ASM_X86_PERCPU_H */