arch/sparc/include/asm/switch_to_32.h

   1 #ifndef __SPARC_SWITCH_TO_H
   2 #define __SPARC_SWITCH_TO_H
   3
   4 #include <asm/smp.h>
   5
   6 extern struct thread_info *current_set[NR_CPUS];
   7
   8 /*
   9  * Flush windows so that the VM switch which follows
  10  * would not pull the stack from under us.
  11  *
  12  * SWITCH_ENTER and SWITH_DO_LAZY_FPU do not work yet (e.g. SMP does not work)
  13  * XXX WTF is the above comment? Found in late teen 2.4.x.
  14  */
  15 #ifdef CONFIG_SMP
  16 #define SWITCH_ENTER(prv) \
  17         do {                    \
  18         if (test_tsk_thread_flag(prv, TIF_USEDFPU)) { \
  19                 put_psr(get_psr() | PSR_EF); \
  20                 fpsave(&(prv)->thread.float_regs[0], &(prv)->thread.fsr, \
  21                        &(prv)->thread.fpqueue[0], &(prv)->thread.fpqdepth); \
  22                 clear_tsk_thread_flag(prv, TIF_USEDFPU); \
  23                 (prv)->thread.kregs->psr &= ~PSR_EF; \
  24         } \
  25         } while(0)
  26
  27 #define SWITCH_DO_LAZY_FPU(next)        /* */
  28 #else
  29 #define SWITCH_ENTER(prv)               /* */
  30 #define SWITCH_DO_LAZY_FPU(nxt) \
  31         do {                    \
  32         if (last_task_used_math != (nxt))               \
  33                 (nxt)->thread.kregs->psr&=~PSR_EF;      \
  34         } while(0)
  35 #endif
  36
  37 #define prepare_arch_switch(next) do { \
  38         __asm__ __volatile__( \
  39         ".globl\tflush_patch_switch\nflush_patch_switch:\n\t" \
  40         "save %sp, -0x40, %sp; save %sp, -0x40, %sp; save %sp, -0x40, %sp\n\t" \
  41         "save %sp, -0x40, %sp; save %sp, -0x40, %sp; save %sp, -0x40, %sp\n\t" \
  42         "save %sp, -0x40, %sp\n\t" \
  43         "restore; restore; restore; restore; restore; restore; restore"); \
  44 } while(0)
  45
  46         /* Much care has gone into this code, do not touch it.
  47          *
  48          * We need to loadup regs l0/l1 for the newly forked child
  49          * case because the trap return path relies on those registers
  50          * holding certain values, gcc is told that they are clobbered.
  51          * Gcc needs registers for 3 values in and 1 value out, so we
  52          * clobber every non-fixed-usage register besides l2/l3/o4/o5.  -DaveM
  53          *
  54          * Hey Dave, that do not touch sign is too much of an incentive
  55          * - Anton & Pete
  56          */
  57 #define switch_to(prev, next, last) do {                                                \
  58         SWITCH_ENTER(prev);                                                             \
  59         SWITCH_DO_LAZY_FPU(next);                                                       \
  60         cpumask_set_cpu(smp_processor_id(), mm_cpumask(next->active_mm));               \
  61         __asm__ __volatile__(                                                           \
  62         "sethi  %%hi(here - 0x8), %%o7\n\t"                                             \
  63         "mov    %%g6, %%g3\n\t"                                                         \
  64         "or     %%o7, %%lo(here - 0x8), %%o7\n\t"                                       \
  65         "rd     %%psr, %%g4\n\t"                                                        \
  66         "std    %%sp, [%%g6 + %4]\n\t"                                                  \
  67         "rd     %%wim, %%g5\n\t"                                                        \
  68         "wr     %%g4, 0x20, %%psr\n\t"                                                  \
  69         "nop\n\t"                                                                       \
  70         "std    %%g4, [%%g6 + %3]\n\t"                                                  \
  71         "ldd    [%2 + %3], %%g4\n\t"                                                    \
  72         "mov    %2, %%g6\n\t"                                                           \
  73         ".globl patchme_store_new_current\n"                                            \
  74 "patchme_store_new_current:\n\t"                                                        \
  75         "st     %2, [%1]\n\t"                                                           \
  76         "wr     %%g4, 0x20, %%psr\n\t"                                                  \
  77         "nop\n\t"                                                                       \
  78         "nop\n\t"                                                                       \
  79         "nop\n\t"       /* LEON needs all 3 nops: load to %sp depends on CWP. */                \
  80         "ldd    [%%g6 + %4], %%sp\n\t"                                                  \
  81         "wr     %%g5, 0x0, %%wim\n\t"                                                   \
  82         "ldd    [%%sp + 0x00], %%l0\n\t"                                                \
  83         "ldd    [%%sp + 0x38], %%i6\n\t"                                                \
  84         "wr     %%g4, 0x0, %%psr\n\t"                                                   \
  85         "nop\n\t"                                                                       \
  86         "nop\n\t"                                                                       \
  87         "jmpl   %%o7 + 0x8, %%g0\n\t"                                                   \
  88         " ld    [%%g3 + %5], %0\n\t"                                                    \
  89         "here:\n"                                                                       \
  90         : "=&r" (last)                                                                  \
  91         : "r" (&(current_set[hard_smp_processor_id()])),        \
  92           "r" (task_thread_info(next)),                         \
  93           "i" (TI_KPSR),                                        \
  94           "i" (TI_KSP),                                         \
  95           "i" (TI_TASK)                                         \
  96         :       "g1", "g2", "g3", "g4", "g5",       "g7",       \
  97           "l0", "l1",       "l3", "l4", "l5", "l6", "l7",       \
  98           "i0", "i1", "i2", "i3", "i4", "i5",                   \
  99           "o0", "o1", "o2", "o3",                   "o7");      \
 100         } while(0)
 101
 102 void fpsave(unsigned long *fpregs, unsigned long *fsr,
 103             void *fpqueue, unsigned long *fpqdepth);
 104 void synchronize_user_stack(void);
 105
 106 #endif /* __SPARC_SWITCH_TO_H */