My previous patch erroneously aggregated the fpu-state size computation for XSAVES and other XSAVE methods. It turns out that the IA32_XSS can make a difference and that the size has to be read from cpuid(0xD, 0x1) in one case, and cpuid(0xD, 0x0) in the other.
First patch is a plain revert, to make the new proposal easier to review: [PATCH gnumach 1/2] Revert "fpu: Fix cpuid feature detection" The new proposal: [PATCH gnumach 2/2] fpu: Fix cpuid feature detection I am happy to squash them toghether if you prefer. Regards, Olivier
