Re: unified percpu stuff

[Brian Gerst <brgerst@gmail.com>]

On Mon, Jan 26, 2009 at 8:12 PM, Jeremy Fitzhardinge <jeremy@goop.org> wrote:
> I'm really pleased to see the unified percpu stuff in the kernel, but
> unfortunately its breaking Xen at the moment.
> It looks like this is just a matter of initializing %gs properly in
> xen_start_kernel.  Is there any problem with me doing a load_gs_base(0)
> somewhere early in xen_start_kernel (arch/x86/xen/enlighten.c)?

Some of the changes I did made the assumption that the percpu state is
set up early in head_xx.S, which apparently is skipped for xen.  Is
there documentation anywhere for the xen bootstrap process?

Try this patch (untested):

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index bef941f..b90d061 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -1647,6 +1647,8 @@ asmlinkage void __init xen_start_kernel(void)
 	have_vcpu_info_placement = 0;
 #endif

+	switch_to_new_gdt();
+
 	xen_smp_init();

 	/* Get mfn list */
diff --git a/arch/x86/xen/smp.c b/arch/x86/xen/smp.c
index 7735e3d..00d9265 100644
--- a/arch/x86/xen/smp.c
+++ b/arch/x86/xen/smp.c
@@ -235,6 +235,8 @@ cpu_initialize_context(unsigned int cpu, struct
task_struct *idle)
 	ctxt->user_regs.ss = __KERNEL_DS;
 #ifdef CONFIG_X86_32
 	ctxt->user_regs.fs = __KERNEL_PERCPU;
+#else
+	ctxt->gs_base_kernel = per_cpu_offset(cpu);
 #endif
 	ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 	ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */

--
Brian Gerst