Re: paravirt & xen & SMP

[Jeremy Fitzhardinge <jeremy@goop.org>]

Gerd Hoffmann wrote:
> Well, after reading the code a bit more it seems the intention is to
> call the function in a row with others, have stuff queued up (in case
> lazy_mode == PARAVIRT_LAZY_CPU is set) and update everything in one go
> then.  I don't think that makes sense for that function as it isn't
> performance-critical.  It is called on boot and cpu bringup (well, not
> yet) only, right?
>   

Yes, that's right.  While its not performance critical in itself, the
idea was that if it were in a stream of other deferred cpu-state
updates, then it should happen in the appropriate order.   But in
practice this is mainly to make a context switch a single hypercall, so
the most important ones are stack-switch followed by 3x
update_descriptor (for TLS).

> We have a simliar issue for xen_write_gdt_entry() btw, it calls
> mc_flush() which doesn't work too.  I've fixed it that way:
>
> @@ -761,7 +762,13 @@ static asmlinkage void __init xen_start_
>         /* set up PDA descriptor */
>         pack_descriptor(&low, &high, (unsigned)&boot_pda,
> sizeof(boot_pda)-1,
>                         0x80 | DESCTYPE_S | 0x02, 0);
> +#if 0
>         xen_write_gdt_entry(cpu_gdt_table, GDT_ENTRY_PDA, low, high);
> +#else
> +       if (HYPERVISOR_update_descriptor(virt_to_machine(cpu_gdt_table +
> GDT_ENTRY_PDA).maddr,
> +                                        (u64)high << 32 | low))
> +               BUG();
> +#endif
>
>         /* set up %gs and init Xen parts of the PDA */
>         asm volatile("mov %0, %%gs" : : "r" (__KERNEL_PDA) : "memory");
>
> [ gna, thunderbird isn't great for sending patches inline, you should
>   get the idea though ... ]
>   

Yeah, that's ugly, but I guess it will do for now.

(You can inline patches without damage in thunderbird if you paste them
as "preformat")

    J