Merge branch 'x86/core' into tracing/textedit
[safe/jmp/linux-2.6] / arch / x86 / kernel / process.c
index 18c70fe..8c03705 100644 (file)
@@ -2,12 +2,18 @@
 #include <linux/kernel.h>
 #include <linux/mm.h>
 #include <linux/smp.h>
+#include <linux/prctl.h>
 #include <linux/slab.h>
 #include <linux/sched.h>
 #include <linux/module.h>
 #include <linux/pm.h>
 #include <linux/clockchips.h>
+#include <trace/power.h>
 #include <asm/system.h>
+#include <asm/apic.h>
+#include <asm/idle.h>
+#include <asm/uaccess.h>
+#include <asm/i387.h>
 
 unsigned long idle_halt;
 EXPORT_SYMBOL(idle_halt);
@@ -16,6 +22,9 @@ EXPORT_SYMBOL(idle_nomwait);
 
 struct kmem_cache *task_xstate_cachep;
 
+DEFINE_TRACE(power_start);
+DEFINE_TRACE(power_end);
+
 int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
 {
        *dst = *src;
@@ -53,6 +62,192 @@ void arch_task_cache_init(void)
 }
 
 /*
+ * Free current thread data structures etc..
+ */
+void exit_thread(void)
+{
+       struct task_struct *me = current;
+       struct thread_struct *t = &me->thread;
+
+       if (me->thread.io_bitmap_ptr) {
+               struct tss_struct *tss = &per_cpu(init_tss, get_cpu());
+
+               kfree(t->io_bitmap_ptr);
+               t->io_bitmap_ptr = NULL;
+               clear_thread_flag(TIF_IO_BITMAP);
+               /*
+                * Careful, clear this in the TSS too:
+                */
+               memset(tss->io_bitmap, 0xff, t->io_bitmap_max);
+               t->io_bitmap_max = 0;
+               put_cpu();
+       }
+
+       ds_exit_thread(current);
+}
+
+void flush_thread(void)
+{
+       struct task_struct *tsk = current;
+
+#ifdef CONFIG_X86_64
+       if (test_tsk_thread_flag(tsk, TIF_ABI_PENDING)) {
+               clear_tsk_thread_flag(tsk, TIF_ABI_PENDING);
+               if (test_tsk_thread_flag(tsk, TIF_IA32)) {
+                       clear_tsk_thread_flag(tsk, TIF_IA32);
+               } else {
+                       set_tsk_thread_flag(tsk, TIF_IA32);
+                       current_thread_info()->status |= TS_COMPAT;
+               }
+       }
+#endif
+
+       clear_tsk_thread_flag(tsk, TIF_DEBUG);
+
+       tsk->thread.debugreg0 = 0;
+       tsk->thread.debugreg1 = 0;
+       tsk->thread.debugreg2 = 0;
+       tsk->thread.debugreg3 = 0;
+       tsk->thread.debugreg6 = 0;
+       tsk->thread.debugreg7 = 0;
+       memset(tsk->thread.tls_array, 0, sizeof(tsk->thread.tls_array));
+       /*
+        * Forget coprocessor state..
+        */
+       tsk->fpu_counter = 0;
+       clear_fpu(tsk);
+       clear_used_math();
+}
+
+static void hard_disable_TSC(void)
+{
+       write_cr4(read_cr4() | X86_CR4_TSD);
+}
+
+void disable_TSC(void)
+{
+       preempt_disable();
+       if (!test_and_set_thread_flag(TIF_NOTSC))
+               /*
+                * Must flip the CPU state synchronously with
+                * TIF_NOTSC in the current running context.
+                */
+               hard_disable_TSC();
+       preempt_enable();
+}
+
+static void hard_enable_TSC(void)
+{
+       write_cr4(read_cr4() & ~X86_CR4_TSD);
+}
+
+static void enable_TSC(void)
+{
+       preempt_disable();
+       if (test_and_clear_thread_flag(TIF_NOTSC))
+               /*
+                * Must flip the CPU state synchronously with
+                * TIF_NOTSC in the current running context.
+                */
+               hard_enable_TSC();
+       preempt_enable();
+}
+
+int get_tsc_mode(unsigned long adr)
+{
+       unsigned int val;
+
+       if (test_thread_flag(TIF_NOTSC))
+               val = PR_TSC_SIGSEGV;
+       else
+               val = PR_TSC_ENABLE;
+
+       return put_user(val, (unsigned int __user *)adr);
+}
+
+int set_tsc_mode(unsigned int val)
+{
+       if (val == PR_TSC_SIGSEGV)
+               disable_TSC();
+       else if (val == PR_TSC_ENABLE)
+               enable_TSC();
+       else
+               return -EINVAL;
+
+       return 0;
+}
+
+void __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
+                     struct tss_struct *tss)
+{
+       struct thread_struct *prev, *next;
+
+       prev = &prev_p->thread;
+       next = &next_p->thread;
+
+       if (test_tsk_thread_flag(next_p, TIF_DS_AREA_MSR) ||
+           test_tsk_thread_flag(prev_p, TIF_DS_AREA_MSR))
+               ds_switch_to(prev_p, next_p);
+       else if (next->debugctlmsr != prev->debugctlmsr)
+               update_debugctlmsr(next->debugctlmsr);
+
+       if (test_tsk_thread_flag(next_p, TIF_DEBUG)) {
+               set_debugreg(next->debugreg0, 0);
+               set_debugreg(next->debugreg1, 1);
+               set_debugreg(next->debugreg2, 2);
+               set_debugreg(next->debugreg3, 3);
+               /* no 4 and 5 */
+               set_debugreg(next->debugreg6, 6);
+               set_debugreg(next->debugreg7, 7);
+       }
+
+       if (test_tsk_thread_flag(prev_p, TIF_NOTSC) ^
+           test_tsk_thread_flag(next_p, TIF_NOTSC)) {
+               /* prev and next are different */
+               if (test_tsk_thread_flag(next_p, TIF_NOTSC))
+                       hard_disable_TSC();
+               else
+                       hard_enable_TSC();
+       }
+
+       if (test_tsk_thread_flag(next_p, TIF_IO_BITMAP)) {
+               /*
+                * Copy the relevant range of the IO bitmap.
+                * Normally this is 128 bytes or less:
+                */
+               memcpy(tss->io_bitmap, next->io_bitmap_ptr,
+                      max(prev->io_bitmap_max, next->io_bitmap_max));
+       } else if (test_tsk_thread_flag(prev_p, TIF_IO_BITMAP)) {
+               /*
+                * Clear any possible leftover bits:
+                */
+               memset(tss->io_bitmap, 0xff, prev->io_bitmap_max);
+       }
+}
+
+int sys_fork(struct pt_regs *regs)
+{
+       return do_fork(SIGCHLD, regs->sp, regs, 0, NULL, NULL);
+}
+
+/*
+ * This is trivial, and on the face of it looks like it
+ * could equally well be done in user mode.
+ *
+ * Not so, for quite unobvious reasons - register pressure.
+ * In user mode vfork() cannot have a stack frame, and if
+ * done by calling the "clone()" system call directly, you
+ * do not have enough call-clobbered registers to hold all
+ * the information you need.
+ */
+int sys_vfork(struct pt_regs *regs)
+{
+       return do_fork(CLONE_VFORK | CLONE_VM | SIGCHLD, regs->sp, regs, 0,
+                      NULL, NULL);
+}
+
+
+/*
  * Idle related variables and functions
  */
 unsigned long boot_option_idle_override = 0;
@@ -100,6 +295,9 @@ static inline int hlt_use_halt(void)
 void default_idle(void)
 {
        if (hlt_use_halt()) {
+               struct power_trace it;
+
+               trace_power_start(&it, POWER_CSTATE, 1);
                current_thread_info()->status &= ~TS_POLLING;
                /*
                 * TS_POLLING-cleared state must be visible before we
@@ -112,6 +310,7 @@ void default_idle(void)
                else
                        local_irq_enable();
                current_thread_info()->status |= TS_POLLING;
+               trace_power_end(&it);
        } else {
                local_irq_enable();
                /* loop is done by the caller */
@@ -122,6 +321,21 @@ void default_idle(void)
 EXPORT_SYMBOL(default_idle);
 #endif
 
+void stop_this_cpu(void *dummy)
+{
+       local_irq_disable();
+       /*
+        * Remove this CPU:
+        */
+       cpu_clear(smp_processor_id(), cpu_online_map);
+       disable_local_APIC();
+
+       for (;;) {
+               if (hlt_works(smp_processor_id()))
+                       halt();
+       }
+}
+
 static void do_nothing(void *unused)
 {
 }
@@ -154,24 +368,37 @@ EXPORT_SYMBOL_GPL(cpu_idle_wait);
  */
 void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
 {
+       struct power_trace it;
+
+       trace_power_start(&it, POWER_CSTATE, (ax>>4)+1);
        if (!need_resched()) {
+               if (cpu_has(&current_cpu_data, X86_FEATURE_CLFLUSH_MONITOR))
+                       clflush((void *)&current_thread_info()->flags);
+
                __monitor((void *)&current_thread_info()->flags, 0, 0);
                smp_mb();
                if (!need_resched())
                        __mwait(ax, cx);
        }
+       trace_power_end(&it);
 }
 
 /* Default MONITOR/MWAIT with no hints, used for default C1 state */
 static void mwait_idle(void)
 {
+       struct power_trace it;
        if (!need_resched()) {
+               trace_power_start(&it, POWER_CSTATE, 1);
+               if (cpu_has(&current_cpu_data, X86_FEATURE_CLFLUSH_MONITOR))
+                       clflush((void *)&current_thread_info()->flags);
+
                __monitor((void *)&current_thread_info()->flags, 0, 0);
                smp_mb();
                if (!need_resched())
                        __sti_mwait(0, 0);
                else
                        local_irq_enable();
+               trace_power_end(&it);
        } else
                local_irq_enable();
 }
@@ -183,9 +410,13 @@ static void mwait_idle(void)
  */
 static void poll_idle(void)
 {
+       struct power_trace it;
+
+       trace_power_start(&it, POWER_CSTATE, 0);
        local_irq_enable();
        while (!need_resched())
                cpu_relax();
+       trace_power_end(&it);
 }
 
 /*
@@ -311,7 +542,7 @@ static void c1e_idle(void)
 
 void __cpuinit select_idle_routine(const struct cpuinfo_x86 *c)
 {
-#ifdef CONFIG_X86_SMP
+#ifdef CONFIG_SMP
        if (pm_idle == poll_idle && smp_num_siblings > 1) {
                printk(KERN_WARNING "WARNING: polling idle and HT enabled,"
                        " performance may degrade.\n");