On Tue, Oct 10, 2023 at 10:19:38AM +0200, Borislav Petkov wrote: > On Tue, Oct 10, 2023 at 08:37:16AM +0300, Kirill A. Shutemov wrote: > > On machines with 5-level paging, cpu_feature_enabled(X86_FEATURE_LA57) > > got patched. It includes KASAN code, where KASAN_SHADOW_START depends on > > __VIRTUAL_MASK_SHIFT, which is defined with the cpu_feature_enabled(). > > So use boot_cpu_has(X86_FEATURE_LA57). > > > It seems that KASAN gets confused when apply_alternatives() patches the > > It seems? > > > KASAN_SHADOW_START users. A test patch that makes KASAN_SHADOW_START > > static, by replacing __VIRTUAL_MASK_SHIFT with 56, fixes the issue. > > > > During text_poke_early() in apply_alternatives(), KASAN should be > > disabled. KASAN is already disabled in non-_early() text_poke(). > > > > It is unclear why the issue was not reported earlier. Bisecting does not > > help. Older kernels trigger the issue less frequently, but it still > > occurs. In the absence of any other clear offenders, the initial dynamic > > 5-level paging support is to blame. > > This whole thing sounds like it is still not really clear what is > actually happening... somewhere along the line __asan_loadN() gets tripped, this then ends up in kasan_check_range() -> check_region_inline() -> addr_has_metadata(). This latter has: kasan_shadow_to_mem() which is compared against KASAN_SHADOW_START, which includes, as Kirill says __VIRTUAL_MASK_SHIFT. Now, obviously you really don't want boot_cpu_has() in __VIRTUAL_MASK_SHIFT, that would be really bad (Linus recently complained about how horrible the code-gen is around this already, must not make it far worse). Anyway, being half-way through patching X86_FEATURE_LA57 thing *are* inconsistent and I really can't blame things for going sideways. That said, I don't particularly like the patch, I think it should, at the veyr least, cover all of apply_alternatives, not just text_poke_early().
