Re: [PATCH 2/4] x86/sev: Allocate request in TSC_INFO_REQ on stack

[Alexey Kardashevskiy <aik@amd.com>]

On 6/5/25 02:03, Dave Hansen wrote:
> On 5/5/25 07:12, Alexey Kardashevskiy wrote:
>> Allocate a 88 byte request structure on stack and skip needless
>> kzalloc/kfree.
> 
> Could you maybe take a closer look at _all_ of these rather than poking
> at them one at a time?
> 
> snp_guest_request_ioctl, for example, looks to be ~32 bytes. Why fix
> 'struct snp_guest_req' and leave an even worse offender?

snp_guest_request_ioctl is allocated on the stack in snp_guest_ioctl(), it calls, say, get_report() which allocates snp_guest_req on the stack too. Do I miss something?

> Or, maybe just be done with it and convert them all over to __free().
> Yeah, some of them don't need to be kmalloc(), but kmalloc()s are cheap
> and consistency is nice, like in the attached patch.

I'd rather not. cheap != free, also hurts to read all these __free - I know it is cheap to kmalloc() and initialize pointers on the stack with NULL but also useless.

More to the oint - it helps (at least me) to see from declarations what structure must be page aligned page size (or any other special allocation requirements) allocation for aesgcm_encrypt() to not barf later on and what does not.

> It also wouldn't be awful to mix stack and kmalloc() allocations,
> especially when the freeing semantics are the same for stack and
> __free()-annotated allocations.

If anything, I'd rather merge snp_msg_alloc() into snp_msg_init() and skip on allocating the snp_msg_desc struct.

For now I want the patch to be painfully simple to review and make the code a little easier to read.

> But it would be really nice to completely eliminate the goto mess.

I understand it is 2025 but it is not exactly mess. Thanks for the review, I am planning to follow up on this, just probably not exactly with __free-cation of everything.


-- 
Alexey