Re: [PATCH] x86/efi: page align EFI ROM image ranges

[Stuart Hayes]

On 8/10/2017 9:46 AM, Ingo Molnar wrote:
> 
> * Matt Fleming <matt@xxxxxxxxxxxxxxxxxxx> wrote:
> 
>> On Wed, 02 Aug, at 11:41:38AM, Stuart Hayes wrote:
>>> (Resend because I mistyped the maintainer's email address the first time.)
>>>
>>> The kernel's EFI stub locates and copies EFI ROM images into memory,
>>> which it allocates using the byte-granular EFI allocate_pool
>>> function.  These memory ranges are then added to setup_data, and
>>> later to e820 (in e820__reserve_setup_data()).  The e820 ranges are
>>> parsed to create nosave regions (in
>>> e820__register_nosave_regions()), but when non-page-aligned e820
>>> regions are parsed, a nosave page is added at the beginning and end
>>> of each non-page-aligned region, which results in data not getting
>>> saved or restored during a hibernate/resume.  This can result in
>>> random failures after a hibernate/resume.
>>>  
>>> Round up the allocation size to a whole number of pages, and use EFI
>>> allocate_pages to ensure that the EFI ROM copy regions are
>>> page-aligned.
>>>
>>> On a system with six EFI ROM images, before the patch:
>>>
>>> e820: update [mem 0x64866020-0x6486e05f] usable ==> usable
>>> e820: update [mem 0x6147a020-0x61499c5f] usable ==> usable
>>> e820: update [mem 0x60fff020-0x6105785f] usable ==> usable
>>> e820: update [mem 0x60fa6020-0x60ffe85f] usable ==> usable
>>> e820: update [mem 0x60f4d020-0x60fa585f] usable ==> usable
>>> e820: update [mem 0x60ef4020-0x60f4c85f] usable ==> usable
>>> ...
>>> PM: Registered nosave memory: [mem 0x00000000-0x00000fff]
>>> PM: Registered nosave memory: [mem 0x000a0000-0x000fffff]
>>> PM: Registered nosave memory: [mem 0x60ef4000-0x60ef4fff]
>>> PM: Registered nosave memory: [mem 0x60f4c000-0x60f4cfff]
>>> PM: Registered nosave memory: [mem 0x60f4d000-0x60f4dfff]
>>> PM: Registered nosave memory: [mem 0x60fa5000-0x60fa5fff]
>>> PM: Registered nosave memory: [mem 0x60fa6000-0x60fa6fff]
>>> PM: Registered nosave memory: [mem 0x60ffe000-0x60ffefff]
>>> PM: Registered nosave memory: [mem 0x60fff000-0x60ffffff]
>>> PM: Registered nosave memory: [mem 0x61057000-0x61057fff]
>>> PM: Registered nosave memory: [mem 0x6147a000-0x6147afff]
>>> PM: Registered nosave memory: [mem 0x61499000-0x61499fff]
>>> PM: Registered nosave memory: [mem 0x64866000-0x64866fff]
>>> PM: Registered nosave memory: [mem 0x6486e000-0x6486efff]
>>> PM: Registered nosave memory: [mem 0x6cf6e000-0x6f3ccfff]
>>>
>>> After the patch:
>>>
>>> e820: update [mem 0x64866000-0x6486efff] usable ==> usable
>>> e820: update [mem 0x6147a000-0x61499fff] usable ==> usable
>>> e820: update [mem 0x60fff000-0x61057fff] usable ==> usable
>>> e820: update [mem 0x60fa6000-0x60ffefff] usable ==> usable
>>> e820: update [mem 0x60f4d000-0x60fa5fff] usable ==> usable
>>> e820: update [mem 0x60ef4000-0x60f4cfff] usable ==> usable
>>> ...
>>> PM: Registered nosave memory: [mem 0x00000000-0x00000fff]
>>> PM: Registered nosave memory: [mem 0x000a0000-0x000fffff]
>>> PM: Registered nosave memory: [mem 0x6cf6e000-0x6f3ccfff]
>>>
>>> Signed-off-by: Stuart Hayes <stuart.w.hayes@xxxxxxxxx>
>>> ---
>>>
>>> --- linux-4.13-rc2/arch/x86/boot/compressed/eboot.c.orig	2017-08-01 12:12:04.696049106 -0400
>>> +++ linux-4.13-rc2/arch/x86/boot/compressed/eboot.c	2017-08-01 12:11:33.120182236 -0400
>>> @@ -235,7 +235,12 @@ __setup_efi_pci64(efi_pci_io_protocol_64
>>>  
>>>  	size = pci->romsize + sizeof(*rom);
>>>  
>>> -	status = efi_call_early(allocate_pool, EFI_LOADER_DATA, size, &rom);
>>> +	/*
>>> +	 * Get whole pages because this will be added to e820.
>>> +	 */
>>> +	size = PAGE_ALIGN(size);
>>> +	status = efi_call_early(allocate_pages, EFI_ALLOCATE_ANY_PAGES,
>>> +				EFI_LOADER_DATA, (size >> PAGE_SHIFT), &rom);
>>>  	if (status != EFI_SUCCESS) {
>>>  		efi_printk(sys_table, "Failed to alloc mem for rom\n");
>>>  		return status;
>>>  
>>  
>> Nice catch. The comment could do with a little more information,
>> including the fact that it's the e820 nosave code that expects
>> page-aligned ROM regions.
>>
>> Also, you'll need the same fix for __setup_efi_pci32().
> 
> Just a quick ping: is this fix being worked on?
> 
> Thanks,
> 
> 	Ingo
> 

Yes... I sent a v2 patch earlier today.  Sorry for the delay.

Thanks
Stuart