Re: [PATCH v4 0/3] Hardening perf subsystem

[Peter Zijlstra]

On Mon, Jun 10, 2024 at 10:28:52AM -0700, Kees Cook wrote:
> On Sat, Jun 01, 2024 at 06:56:15PM +0200, Erick Archer wrote:
> > Hi everyone,
> > 
> > This is an effort to get rid of all multiplications from allocation
> > functions in order to prevent integer overflows [1][2].
> 
> I didn't actually see these 3 patches in this thread nor via lore.

He managed to break threading between 0/n and the rest.

> > In the first patch, the "struct amd_uncore_ctx" can be refactored to
> > use a flex array for the "events" member. This way, the allocation/
> > freeing of the memory can be simplified. Then, the struct_size()
> > helper can be used to do the arithmetic calculation for the memory
> > to be allocated.
> 
> I like this patch because it reduces the allocation from 2 to 1. This
> isn't what Peter might see as "churn": this is an improvement in resource
> utilization.

But then he went and used that struct_size() abomination :/

> I prefer this style, as it makes things unambiguous ("this will never
> wrap around") without having to check the associated types and doesn't make
> the resulting binary code different in the "can never overflow" case.
> 
> In this particular case:
> 
> int size = sizeof(*box) + numshared * sizeof(struct intel_uncore_extra_reg);
> 
> "int numshared" comes from struct intel_uncore_type::num_shared_regs,
> which is:
> 
>         unsigned num_shared_regs:8;
> 
> And the struct sizes are:
> 
> $ pahole -C intel_uncore_box gcc-boot/vmlinux | grep size:
>         /* size: 488, cachelines: 8, members: 19 */
> $ pahole -C intel_uncore_extra_reg gcc-boot/vmlinux | grep size:
>         /* size: 96, cachelines: 2, members: 5 */
> 
> So we have:
> 
> s32 size = 488 + u8 * 96
> 
> Max size here is 24968 so it can never overflow an s32, so I can see
> why Peter views this as "churn".
> 
> I still think the patch is a coding style improvement, but okay.

I really detest this thing because it makes what was trivially readable
into something opaque. Get me that type qualifier that traps on overflow
and write plain C. All this __builtin_overflow garbage is just that,
unreadable nonsense.

> This provides __counted_by coverage, and I think this is important to
> gain in ever place we can. Given that this is part of a ring buffer
> implementation that is arbitrarily sized, this is exactly the kind of
> place I'd like to see __counted_by used. This is a runtime robustness
> improvement, so I don't see this a "churn" at all.

Again, mixed in with that other crap. Anyway, remind me wth this
__counted_by thing actually does?

> Peter, for patches 1 and 3, if you'd prefer not to carry them, I could
> put them in the hardening tree to keep them out of your way. It seems
> clear you don't want patch 2 at all.

I prefer to not have struct_size() anywhere at all. Please just write
readable code.

Again, if you really care about that overflow muck, get a useable C type
qualifier or something so we can write readable code.