x86, fpu: fix CONFIG_PREEMPT=y corruption of application's FPU stack
[safe/jmp/linux-2.6] / arch / x86 / kernel / process_32.c
index 968371a..6d54833 100644 (file)
@@ -23,7 +23,6 @@
 #include <linux/slab.h>
 #include <linux/vmalloc.h>
 #include <linux/user.h>
-#include <linux/a.out.h>
 #include <linux/interrupt.h>
 #include <linux/utsname.h>
 #include <linux/delay.h>
@@ -37,6 +36,7 @@
 #include <linux/personality.h>
 #include <linux/tick.h>
 #include <linux/percpu.h>
+#include <linux/prctl.h>
 
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
@@ -46,7 +46,6 @@
 #include <asm/processor.h>
 #include <asm/i387.h>
 #include <asm/desc.h>
-#include <asm/vm86.h>
 #ifdef CONFIG_MATH_EMULATION
 #include <asm/math_emu.h>
 #endif
@@ -83,7 +82,6 @@ unsigned long thread_saved_pc(struct task_struct *tsk)
  */
 void (*pm_idle)(void);
 EXPORT_SYMBOL(pm_idle);
-static DEFINE_PER_CPU(unsigned int, cpu_idle_state);
 
 void disable_hlt(void)
 {
@@ -113,22 +111,13 @@ void default_idle(void)
                 */
                smp_mb();
 
-               local_irq_disable();
-               if (!need_resched()) {
-                       ktime_t t0, t1;
-                       u64 t0n, t1n;
-
-                       t0 = ktime_get();
-                       t0n = ktime_to_ns(t0);
+               if (!need_resched())
                        safe_halt();    /* enables interrupts racelessly */
-                       local_irq_disable();
-                       t1 = ktime_get();
-                       t1n = ktime_to_ns(t1);
-                       sched_clock_idle_wakeup_event(t1n - t0n);
-               }
-               local_irq_enable();
+               else
+                       local_irq_enable();
                current_thread_info()->status |= TS_POLLING;
        } else {
+               local_irq_enable();
                /* loop is done by the caller */
                cpu_relax();
        }
@@ -137,16 +126,6 @@ void default_idle(void)
 EXPORT_SYMBOL(default_idle);
 #endif
 
-/*
- * On SMP it's slightly faster (but much more power-consuming!)
- * to poll the ->work.need_resched flag instead of waiting for the
- * cross-CPU IPI to arrive. Use this option with caution.
- */
-static void poll_idle(void)
-{
-       cpu_relax();
-}
-
 #ifdef CONFIG_HOTPLUG_CPU
 #include <asm/nmi.h>
 /* We don't actually take CPU down, just spin without interrupts. */
@@ -191,9 +170,6 @@ void cpu_idle(void)
                while (!need_resched()) {
                        void (*idle)(void);
 
-                       if (__get_cpu_var(cpu_idle_state))
-                               __get_cpu_var(cpu_idle_state) = 0;
-
                        check_pgt_cache();
                        rmb();
                        idle = pm_idle;
@@ -207,6 +183,7 @@ void cpu_idle(void)
                        if (cpu_is_offline(cpu))
                                play_dead();
 
+                       local_irq_disable();
                        __get_cpu_var(irq_stat).idle_timestamp = jiffies;
                        idle();
                }
@@ -217,122 +194,6 @@ void cpu_idle(void)
        }
 }
 
-static void do_nothing(void *unused)
-{
-}
-
-void cpu_idle_wait(void)
-{
-       unsigned int cpu, this_cpu = get_cpu();
-       cpumask_t map, tmp = current->cpus_allowed;
-
-       set_cpus_allowed(current, cpumask_of_cpu(this_cpu));
-       put_cpu();
-
-       cpus_clear(map);
-       for_each_online_cpu(cpu) {
-               per_cpu(cpu_idle_state, cpu) = 1;
-               cpu_set(cpu, map);
-       }
-
-       __get_cpu_var(cpu_idle_state) = 0;
-
-       wmb();
-       do {
-               ssleep(1);
-               for_each_online_cpu(cpu) {
-                       if (cpu_isset(cpu, map) && !per_cpu(cpu_idle_state, cpu))
-                               cpu_clear(cpu, map);
-               }
-               cpus_and(map, map, cpu_online_map);
-               /*
-                * We waited 1 sec, if a CPU still did not call idle
-                * it may be because it is in idle and not waking up
-                * because it has nothing to do.
-                * Give all the remaining CPUS a kick.
-                */
-               smp_call_function_mask(map, do_nothing, 0, 0);
-       } while (!cpus_empty(map));
-
-       set_cpus_allowed(current, tmp);
-}
-EXPORT_SYMBOL_GPL(cpu_idle_wait);
-
-/*
- * This uses new MONITOR/MWAIT instructions on P4 processors with PNI,
- * which can obviate IPI to trigger checking of need_resched.
- * We execute MONITOR against need_resched and enter optimized wait state
- * through MWAIT. Whenever someone changes need_resched, we would be woken
- * up from MWAIT (without an IPI).
- *
- * New with Core Duo processors, MWAIT can take some hints based on CPU
- * capability.
- */
-void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
-{
-       if (!need_resched()) {
-               __monitor((void *)&current_thread_info()->flags, 0, 0);
-               smp_mb();
-               if (!need_resched())
-                       __mwait(ax, cx);
-       }
-}
-
-/* Default MONITOR/MWAIT with no hints, used for default C1 state */
-static void mwait_idle(void)
-{
-       local_irq_enable();
-       mwait_idle_with_hints(0, 0);
-}
-
-static int __cpuinit mwait_usable(const struct cpuinfo_x86 *c)
-{
-       if (force_mwait)
-               return 1;
-       /* Any C1 states supported? */
-       return c->cpuid_level >= 5 && ((cpuid_edx(5) >> 4) & 0xf) > 0;
-}
-
-void __cpuinit select_idle_routine(const struct cpuinfo_x86 *c)
-{
-       static int selected;
-
-       if (selected)
-               return;
-#ifdef CONFIG_X86_SMP
-       if (pm_idle == poll_idle && smp_num_siblings > 1) {
-               printk(KERN_WARNING "WARNING: polling idle and HT enabled,"
-                       " performance may degrade.\n");
-       }
-#endif
-       if (cpu_has(c, X86_FEATURE_MWAIT) && mwait_usable(c)) {
-               /*
-                * Skip, if setup has overridden idle.
-                * One CPU supports mwait => All CPUs supports mwait
-                */
-               if (!pm_idle) {
-                       printk(KERN_INFO "using mwait in idle threads.\n");
-                       pm_idle = mwait_idle;
-               }
-       }
-       selected = 1;
-}
-
-static int __init idle_setup(char *str)
-{
-       if (!strcmp(str, "poll")) {
-               printk("using polling idle threads.\n");
-               pm_idle = poll_idle;
-       } else if (!strcmp(str, "mwait"))
-               force_mwait = 1;
-       else
-               return -1;
-
-       boot_option_idle_override = 1;
-       return 0;
-}
-early_param("idle", idle_setup);
-
 void __show_registers(struct pt_regs *regs, int all)
 {
        unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L;
@@ -358,7 +219,7 @@ void __show_registers(struct pt_regs *regs, int all)
                        init_utsname()->version);
 
        printk("EIP: %04x:[<%08lx>] EFLAGS: %08lx CPU: %d\n",
-                       0xffff & regs->cs, regs->ip, regs->flags,
+                       (u16)regs->cs, regs->ip, regs->flags,
                        smp_processor_id());
        print_symbol("EIP is at %s\n", regs->ip);
 
@@ -367,8 +228,7 @@ void __show_registers(struct pt_regs *regs, int all)
        printk("ESI: %08lx EDI: %08lx EBP: %08lx ESP: %08lx\n",
                regs->si, regs->di, regs->bp, sp);
        printk(" DS: %04x ES: %04x FS: %04x GS: %04x SS: %04x\n",
-              regs->ds & 0xffff, regs->es & 0xffff,
-              regs->fs & 0xffff, gs, ss);
+              (u16)regs->ds, (u16)regs->es, (u16)regs->fs, gs, ss);
 
        if (!all)
                return;
@@ -539,60 +399,30 @@ int copy_thread(int nr, unsigned long clone_flags, unsigned long sp,
        return err;
 }
 
-/*
- * fill in the user structure for a core dump..
- */
-void dump_thread(struct pt_regs * regs, struct user * dump)
+void
+start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
 {
-       u16 gs;
-
-/* changed the size calculations - should hopefully work better. lbt */
-       dump->magic = CMAGIC;
-       dump->start_code = 0;
-       dump->start_stack = regs->sp & ~(PAGE_SIZE - 1);
-       dump->u_tsize = ((unsigned long) current->mm->end_code) >> PAGE_SHIFT;
-       dump->u_dsize = ((unsigned long) (current->mm->brk + (PAGE_SIZE-1))) >> PAGE_SHIFT;
-       dump->u_dsize -= dump->u_tsize;
-       dump->u_ssize = 0;
-       dump->u_debugreg[0] = current->thread.debugreg0;
-       dump->u_debugreg[1] = current->thread.debugreg1;
-       dump->u_debugreg[2] = current->thread.debugreg2;
-       dump->u_debugreg[3] = current->thread.debugreg3;
-       dump->u_debugreg[4] = 0;
-       dump->u_debugreg[5] = 0;
-       dump->u_debugreg[6] = current->thread.debugreg6;
-       dump->u_debugreg[7] = current->thread.debugreg7;
-
-       if (dump->start_stack < TASK_SIZE)
-               dump->u_ssize = ((unsigned long) (TASK_SIZE - dump->start_stack)) >> PAGE_SHIFT;
-
-       dump->regs.bx = regs->bx;
-       dump->regs.cx = regs->cx;
-       dump->regs.dx = regs->dx;
-       dump->regs.si = regs->si;
-       dump->regs.di = regs->di;
-       dump->regs.bp = regs->bp;
-       dump->regs.ax = regs->ax;
-       dump->regs.ds = (u16)regs->ds;
-       dump->regs.es = (u16)regs->es;
-       dump->regs.fs = (u16)regs->fs;
-       savesegment(gs,gs);
-       dump->regs.orig_ax = regs->orig_ax;
-       dump->regs.ip = regs->ip;
-       dump->regs.cs = (u16)regs->cs;
-       dump->regs.flags = regs->flags;
-       dump->regs.sp = regs->sp;
-       dump->regs.ss = (u16)regs->ss;
-
-       dump->u_fpvalid = dump_fpu (regs, &dump->i387);
+       __asm__("movl %0, %%gs" :: "r"(0));
+       regs->fs                = 0;
+       set_fs(USER_DS);
+       regs->ds                = __USER_DS;
+       regs->es                = __USER_DS;
+       regs->ss                = __USER_DS;
+       regs->cs                = __USER_CS;
+       regs->ip                = new_ip;
+       regs->sp                = new_sp;
+       /*
+        * Free the old FP and other extended state
+        */
+       free_thread_xstate(current);
 }
-EXPORT_SYMBOL(dump_thread);
+EXPORT_SYMBOL_GPL(start_thread);
 
-#ifdef CONFIG_SECCOMP
 static void hard_disable_TSC(void)
 {
        write_cr4(read_cr4() | X86_CR4_TSD);
 }
+
 void disable_TSC(void)
 {
        preempt_disable();
@@ -604,11 +434,47 @@ void disable_TSC(void)
                hard_disable_TSC();
        preempt_enable();
 }
+
 static void hard_enable_TSC(void)
 {
        write_cr4(read_cr4() & ~X86_CR4_TSD);
 }
-#endif /* CONFIG_SECCOMP */
+
+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;
+}
 
 static noinline void
 __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
@@ -625,12 +491,12 @@ __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
                /* we clear debugctl to make sure DS
                 * is not in use when we change it */
                debugctl = 0;
-               wrmsrl(MSR_IA32_DEBUGCTLMSR, 0);
+               update_debugctlmsr(0);
                wrmsr(MSR_IA32_DS_AREA, next->ds_area_msr, 0);
        }
 
        if (next->debugctlmsr != debugctl)
-               wrmsr(MSR_IA32_DEBUGCTLMSR, next->debugctlmsr, 0);
+               update_debugctlmsr(next->debugctlmsr);
 
        if (test_tsk_thread_flag(next_p, TIF_DEBUG)) {
                set_debugreg(next->debugreg0, 0);
@@ -642,7 +508,6 @@ __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
                set_debugreg(next->debugreg7, 7);
        }
 
-#ifdef CONFIG_SECCOMP
        if (test_tsk_thread_flag(prev_p, TIF_NOTSC) ^
            test_tsk_thread_flag(next_p, TIF_NOTSC)) {
                /* prev and next are different */
@@ -651,13 +516,14 @@ __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
                else
                        hard_enable_TSC();
        }
-#endif
 
+#ifdef X86_BTS
        if (test_tsk_thread_flag(prev_p, TIF_BTS_TRACE_TS))
                ptrace_bts_take_timestamp(prev_p, BTS_TASK_DEPARTS);
 
        if (test_tsk_thread_flag(next_p, TIF_BTS_TRACE_TS))
                ptrace_bts_take_timestamp(next_p, BTS_TASK_ARRIVES);
+#endif
 
 
        if (!test_tsk_thread_flag(next_p, TIF_IO_BITMAP)) {
@@ -731,7 +597,7 @@ struct task_struct * __switch_to(struct task_struct *prev_p, struct task_struct
 
        /* we're going to use this soon, after a few expensive things */
        if (next_p->fpu_counter > 5)
-               prefetch(&next->i387.fxsave);
+               prefetch(next->xstate);
 
        /*
         * Reload esp0.
@@ -783,8 +649,11 @@ struct task_struct * __switch_to(struct task_struct *prev_p, struct task_struct
        /* If the task has used fpu the last 5 timeslices, just do a full
         * restore of the math state immediately to avoid the trap; the
         * chances of needing FPU soon are obviously high now
+        *
+        * tsk_used_math() checks prevent calling math_state_restore(),
+        * which can sleep in the case of !tsk_used_math()
         */
-       if (next_p->fpu_counter > 5)
+       if (tsk_used_math(next_p) && next_p->fpu_counter > 5)
                math_state_restore();
 
        /*