[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [PATCH for-4.15] x86/ucode: Fix microcode payload size for Fam19 processors
On 10/02/2021 08:37, Jan Beulich wrote: > On 09.02.2021 18:39, Andrew Cooper wrote: >> On 09/02/2021 17:17, Ian Jackson wrote: >>> Jan Beulich writes ("Re: [PATCH for-4.15] x86/ucode: Fix microcode payload >>> size for Fam19 processors"): >>>> On 09.02.2021 16:33, Andrew Cooper wrote: >>>>> The original limit provided wasn't accurate. Blobs are in fact rather >>>>> larger. >>>>> >>>>> Fixes: fe36a173d1 ("x86/amd: Initial support for Fam19h processors") >>>>> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx> >>>> Acked-by: Jan Beulich <jbeulich@xxxxxxxx> >>>> >>>>> --- a/xen/arch/x86/cpu/microcode/amd.c >>>>> +++ b/xen/arch/x86/cpu/microcode/amd.c >>>>> @@ -111,7 +111,7 @@ static bool verify_patch_size(uint32_t patch_size) >>>>> #define F15H_MPB_MAX_SIZE 4096 >>>>> #define F16H_MPB_MAX_SIZE 3458 >>>>> #define F17H_MPB_MAX_SIZE 3200 >>>>> -#define F19H_MPB_MAX_SIZE 4800 >>>>> +#define F19H_MPB_MAX_SIZE 5568 >>>> How certain is it that there's not going to be another increase? >>>> And in comparison, how bad would it be if we pulled this upper >>>> limit to something that's at least slightly less of an "odd" >>>> number, e.g. 0x1800, and thus provide some headroom? >>> 5568 does seem really an excessively magic number... >> It reads better in hex - 0x15c0. It is exactly the header, >> match/patch-ram, and the digital signature of a fixed algorithm. > And I realize it's less odd than Fam16's 3458 (0xd82). This particular value is under investigation. Firmware packages for Fam16's have the blob size at 0xd60. >> Its far simpler than Intel's format which contains multiple embedded >> blobs, and support for minor platform variations within the same blob. >> >> There are a lot of problems with how we do patch verification on AMD >> right now, but its all a consequence of the header not containing a >> length field. >> >> This number wont change now. Zen3 processors are out in the world. > Given history on earlier families, where in some cases sizes > also vary by model, how can this fact guarantee anything? There is one example where it changed, and it got shorter. Making it larger would involve someone re-laying out the core so more silicon area could be used for patch ram space, which is increasingly unlikely with newer models in the same family. When 4.16 comes, I do have plans to try and make us more robust to changes in debug builds, particularly given the lack of suitable documentation on the matter. ~Andrew
|
Lists.xenproject.org is hosted with RackSpace, monitoring our |