Re: [PATCH v3 2/7] x86/jump_label: Use text_poke_early() during early_init

[Andy Lutomirski <luto@amacapital.net>]

> On Nov 5, 2018, at 6:09 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> 
>> On Fri, Nov 02, 2018 at 04:29:41PM -0700, Nadav Amit wrote:
>> diff --git a/arch/x86/kernel/jump_label.c b/arch/x86/kernel/jump_label.c
>> index aac0c1f7e354..367c1d0c20a3 100644
>> --- a/arch/x86/kernel/jump_label.c
>> +++ b/arch/x86/kernel/jump_label.c
>> @@ -52,7 +52,13 @@ static void __ref __jump_label_transform(struct jump_entry *entry,
>>    jmp.offset = jump_entry_target(entry) -
>>             (jump_entry_code(entry) + JUMP_LABEL_NOP_SIZE);
>> 
>> -    if (early_boot_irqs_disabled)
>> +    /*
>> +     * As long as we are in early boot, we can use text_poke_early(), which
>> +     * is more efficient: the memory was still not marked as read-only (it
>> +     * is only marked after poking_init()). This also prevents us from using
>> +     * text_poke() before poking_init() is called.
>> +     */
>> +    if (!early_boot_done)
>>        poker = text_poke_early;
>> 
>>    if (type == JUMP_LABEL_JMP) {
> 
> It took me a while to untangle init/maze^H^Hin.c... but I think this
> is all we need:
> 
> diff --git a/arch/x86/kernel/jump_label.c b/arch/x86/kernel/jump_label.c
> index aac0c1f7e354..ed5fe274a7d8 100644
> --- a/arch/x86/kernel/jump_label.c
> +++ b/arch/x86/kernel/jump_label.c
> @@ -52,7 +52,12 @@ static void __ref __jump_label_transform(struct jump_entry *entry,
>    jmp.offset = jump_entry_target(entry) -
>             (jump_entry_code(entry) + JUMP_LABEL_NOP_SIZE);
> 
> -    if (early_boot_irqs_disabled)
> +    /*
> +     * As long as we're UP and not yet marked RO, we can use
> +     * text_poke_early; SYSTEM_BOOTING guarantees both, as we switch to
> +     * SYSTEM_SCHEDULING before going either.
> +     */
> +    if (system_state == SYSTEM_BOOTING)
>        poker = text_poke_early;
> 
>    if (type == JUMP_LABEL_JMP) {

Can we move this logic into text_poke() and get rid of text_poke_early()?

FWIW, alternative patching was, at some point, a significant fraction of total boot time in some cases. This was probably mostly due to unnecessary sync_core() calls.  Although I think this was reported on a VM, and sync_core() used to be *extremely* expensive on a VM, but that’s fixed now, and it even got backported, I think.

(Hmm. Maybe we can also make jump label patching work in early boot, too!)