I tried to figure out how x86 handles this, and gave up after some
time :-) Can you figure it out?
I've had a look, what I found is that on x86
event->attr.exclude_callchain_user is not set from the arch/x86/*
But perf unconditionally uses the NMI specific version of copy_from_user
to prevent perf_callchain_user() to generated a fault:
arch/x86/include/asm/perf_event.h:#define arch_perf_out_copy_user
copy_from_user_nmi
Where copy_from_user_nmi() is defined as follows:
arch/x86/lib/usercopy.c: /*
arch/x86/lib/usercopy.c: * We rely on the nested NMI work to allow
atomic faults from the NMI path; the
arch/x86/lib/usercopy.c: * nested NMI paths are careful to preserve CR2.
arch/x86/lib/usercopy.c: */
arch/x86/lib/usercopy.c: unsigned long
arch/x86/lib/usercopy.c: copy_from_user_nmi(void *to, const void __user
*from, unsigned long n)
arch/x86/lib/usercopy.c: {
arch/x86/lib/usercopy.c: unsigned long ret;
arch/x86/lib/usercopy.c:
arch/x86/lib/usercopy.c: if (__range_not_ok(from, n, TASK_SIZE))
arch/x86/lib/usercopy.c: return 0;
arch/x86/lib/usercopy.c:
arch/x86/lib/usercopy.c: /*
arch/x86/lib/usercopy.c: * Even though this function is
typically called from NMI/IRQ context
arch/x86/lib/usercopy.c: * disable pagefaults so that its
behaviour is consistent even when
arch/x86/lib/usercopy.c: * called form other contexts.
arch/x86/lib/usercopy.c: */
arch/x86/lib/usercopy.c: pagefault_disable();
arch/x86/lib/usercopy.c: ret = __copy_from_user_inatomic(to,
from, n);
arch/x86/lib/usercopy.c: pagefault_enable();
arch/x86/lib/usercopy.c:
arch/x86/lib/usercopy.c: return ret;
arch/x86/lib/usercopy.c: }
arch/x86/lib/usercopy.c: EXPORT_SYMBOL_GPL(copy_from_user_nmi);
And the kernel also falls back (for other archs) to a version very
similar to copy_from_user_nmi() where faults are disabled:
kernel/events/internal.h: #ifndef arch_perf_out_copy_user
kernel/events/internal.h: #define arch_perf_out_copy_user
arch_perf_out_copy_user
kernel/events/internal.h:
kernel/events/internal.h: static inline unsigned long
kernel/events/internal.h: arch_perf_out_copy_user(void *dst, const void
*src, unsigned long n)
kernel/events/internal.h: {
kernel/events/internal.h: unsigned long ret;
kernel/events/internal.h:
kernel/events/internal.h: pagefault_disable();
kernel/events/internal.h: ret = __copy_from_user_inatomic(dst,
src, n);
kernel/events/internal.h: pagefault_enable();
kernel/events/internal.h:
kernel/events/internal.h: return ret;
kernel/events/internal.h: }
kernel/events/internal.h: #endif
kernel/events/internal.h:
kernel/events/internal.h: DEFINE_OUTPUT_COPY(__output_copy_user,
arch_perf_out_copy_user)
I believe that's how faults are prevented from perf.
-eric
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
