[PATCH 1/2] boot: ignore early NMIs

[Fernando Luis Vázquez Cao]

On 03/08/2012 01:41 PM, Eric W. Biederman wrote:
> Fernando Luis Vázquez Cao<fernando at oss.ntt.co.jp>  writes:
>
>> Subject: [PATCH] boot: ignore early NMIs
>>
>> From: Fernando Luis Vazquez Cao<fernando at oss.ntt.co.jp>
>>
>> NMIs very early in the boot process are rarely critical (usually
>> it just means that there was a spurious bit flip somewhere in the
>> hardware, or that this is a kdump kernel and we received an NMI
>> generated in the previous context), so the current behavior of
>> halting the system when one occurs is probably a bit over the top.
>>
>> This patch changes the early IDT so that NMIs are ignored and the
>> kernel can, hopefully, continue executing other code. Harsher
>> measures (panic, etc) are defered to the final NMI handler, which
>> can actually make an informed decision.
>>
>> This issue presented itself in our environment as seemingly
>> random hangs in kdump.
>>
>> Signed-off-by: Fernando Luis Vazquez Cao<fernando at oss.ntt.co.jp>
>> ---
>>
>> diff -urNp linux-3.3-rc6-orig/arch/x86/kernel/head64.c linux-3.3-rc6/arch/x86/kernel/head64.c
>> --- linux-3.3-rc6-orig/arch/x86/kernel/head64.c	2012-03-07 15:49:01.834241787 +0900
>> +++ linux-3.3-rc6/arch/x86/kernel/head64.c	2012-03-07 18:39:03.173732875 +0900
>> @@ -71,7 +71,7 @@ void __init x86_64_start_kernel(char * r
>>   				(__START_KERNEL&  PGDIR_MASK)));
>>   	BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses)<= MODULES_END);
>>
>> -	/* clear bss before set_intr_gate with early_idt_handler */
>> +	/* clear bss before set_intr_gate with early_idt_handlers */
>>   	clear_bss();
>>
>>   	/* Make NULL pointers segfault */
>> @@ -79,13 +79,8 @@ void __init x86_64_start_kernel(char * r
>>
>>   	max_pfn_mapped = KERNEL_IMAGE_SIZE>>  PAGE_SHIFT;
>>
>> -	for (i = 0; i<  NUM_EXCEPTION_VECTORS; i++) {
>> -#ifdef CONFIG_EARLY_PRINTK
>> +	for (i = 0; i<  NUM_EXCEPTION_VECTORS; i++)
>>   		set_intr_gate(i,&early_idt_handlers[i]);
>> -#else
>> -		set_intr_gate(i, early_idt_handler);
>> -#endif
>> -	}
>>   	load_idt((const struct desc_ptr *)&idt_descr);
>>
>>   	if (console_loglevel == 10)
>> diff -urNp linux-3.3-rc6-orig/arch/x86/kernel/head_64.S linux-3.3-rc6/arch/x86/kernel/head_64.S
>> --- linux-3.3-rc6-orig/arch/x86/kernel/head_64.S	2012-03-07 15:49:01.838241839 +0900
>> +++ linux-3.3-rc6/arch/x86/kernel/head_64.S	2012-03-07 18:41:21.811516876 +0900
>> @@ -270,18 +270,29 @@ bad_address:
>>   	jmp bad_address
>>
>>   	.section ".init.text","ax"
>> -#ifdef CONFIG_EARLY_PRINTK
>>   	.globl early_idt_handlers
>>   early_idt_handlers:
>> -	i = 0
>> +	vector = 0
>>   	.rept NUM_EXCEPTION_VECTORS
>> -	movl $i, %esi
>> -	jmp early_idt_handler
>> -	i = i + 1
>> +	/*
>> +	 * NMIs (vector 2) this early in the boot process are rarely critical
>> +	 * (usually it just means that there was a spurious bit flip somewhere
>> +	 * in the hardware, or that this is a kdump kernel and we received an
>> +	 * NMI generated in the previous context), so we ignore them here and
>> +	 * try to continue (see early_nmi_handler implementation below).
>> +	 * Harsher measures (panic, etc) are defered to the final NMI handler,
>> +	 * which can actually make an informed decision.
>> +	 */
>> +	.if vector == 2
>> +	jmp early_nmi_handler
> Is just a jump and not a move followed by a jump still 10 bytes?
> I hate to say it but I think this fails miserably for any exception
> after a nmi.

Thank you for the heads up! Actually, it was working for the
exceptions after the nmi but with a corrupted esi (vector
number). My original intention was to fill the empty space
with nops but forgot to actually implement it... Sorry about
that. Will fix for the next iteration.

> I expect the simplest solution is to modify early_idt_handler to test
> for vector == 2.

That is precisely what I did on a previous version but that would
involve using registers which need to be saved and restored and
I wanted to avoid using the stack in the NMI path. We would also
need to add a "pushq rsi " in early_idt_handlers which implies
modifying "early_idt_handlers" definition in "segment.h".

If you are OK with it I would like to go with the approach in
the two patches I sent.


> Doing something less brittle than:
>> extern const char early_idt_handlers[NUM_EXCEPTION_VECTORS][10];
> in segment.h might be a good idea as well.

Yes, I agree. I will give it some thought.