On 6/16/23 02:39, Jordy Zomer wrote: > Thanks for the explanation Pawan, a little bit off-topic for this patch but > shall I send a patch to add this to the documentation of array_index_nospec() > and fix other calls to that function where the upper bound is not a constant? :) > Yes, please. We don't want to lose that info. Thanks. > On Fri, Jun 16, 2023 at 5:15 AM Pawan Gupta > <pawan.kumar.gupta@xxxxxxxxxxxxxxx> wrote: >> >> On Fri, Jun 16, 2023 at 12:31:50AM +0100, Phillip Potter wrote: >>> I've now looked at this. It is possible for cdi->capacity to be > 1, as >>> it is set via get_capabilities() -> cdrom_number_of_slots(), if the >>> device is an individual or cartridge changer. >> >> Ohk. Is there an upper limit to cdi->capacity? If not, we are left with >> barrier_nospec(). >> >>> Therefore, I think using CDI_MAX_CAPACITY of 1 is not the correct >>> approach. Jordy's V2 patch is fine therefore, but perhaps using >>> array_index_nospec() with cdi->capacity is still better than a >>> do/while loop from a performance perspective, given it would be cached >>> etc. at that point, so possibly quicker. Thoughts? (I'm no expert on >>> spectre-v1 I'll admit). >> >> array_index_nospec() can only clip the arg correctly if the upper bound >> is correct. Problem with array_index_nospec(arg, cdi->capacity) is >> cdi->capacity is not a constant, so it suffers from the same problem as >> arg i.e. cdi->capacity could also be speculated. Although having to >> control 2 loads makes the attack difficult, but does not rules out >> completely. >> >> barrier_nospec() makes the CPU wait for all previous loads to retire >> before executing following instructions speculatively. This causes the >> conditional branch to resolve correctly. I hope this does not fall into >> a hotpath. -- ~Randy
