On 27/01/2026 11:03, Ard Biesheuvel wrote: > On Tue, 27 Jan 2026 at 12:00, Ryan Roberts <[email protected]> wrote: >> >> On 27/01/2026 10:47, Ard Biesheuvel wrote: >>> On Tue, 27 Jan 2026 at 11:39, Ryan Roberts <[email protected]> wrote: >>>> >>>> On 27/01/2026 10:27, Ard Biesheuvel wrote: >>>>> On Tue, 27 Jan 2026 at 11:21, Ryan Roberts <[email protected]> wrote: >>>>>> >>>>>> On 26/01/2026 09:26, Ard Biesheuvel wrote: >>>>>>> From: Ard Biesheuvel <[email protected]> >>>>>>> >>>>>>> Now that the DRAM mapping routines respect existing table mappings and >>>>>>> contiguous block and page mappings, it is no longer needed to fiddle >>>>>>> with the memblock tables to set and clear the NOMAP attribute. Instead, >>>>>>> map the kernel text and rodata alias first, avoiding contiguous >>>>>>> mappings, so that they will not be added later when mapping the >>>>>>> memblocks. >>>>>> >>>>>> Should we do something similar for kfence? Currently we have >>>>>> arm64_kfence_alloc_pool() which marks some memory NOMAP then >>>>>> arm64_kfence_map_pool() which PTE-maps it and clears NOMAP. Presumably >>>>>> we could >>>>>> rationalize into a single function that does it all, prior to mapping >>>>>> the bulk >>>>>> of the linear map? >>>>>> >>>>> >>>>> Yeah, good point - I did not spot that but I will address it in the >>>>> next revision. >>>>> >>>>>>> >>>>>>> Signed-off-by: Ard Biesheuvel <[email protected]> >>>>>>> --- >>>>>>> arch/arm64/mm/mmu.c | 27 ++++++++------------ >>>>>>> 1 file changed, 10 insertions(+), 17 deletions(-) >>>>>>> >>>>>>> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c >>>>>>> index 80587cd47ce7..18415d4743bf 100644 >>>>>>> --- a/arch/arm64/mm/mmu.c >>>>>>> +++ b/arch/arm64/mm/mmu.c >>>>>>> @@ -1149,12 +1149,17 @@ static void __init map_mem(void) >>>>>>> flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS; >>>>>>> >>>>>>> /* >>>>>>> - * Take care not to create a writable alias for the >>>>>>> - * read-only text and rodata sections of the kernel image. >>>>>>> - * So temporarily mark them as NOMAP to skip mappings in >>>>>>> - * the following for-loop >>>>>>> + * Map the linear alias of the [_text, __init_begin) interval >>>>>>> + * as non-executable now, and remove the write permission in >>>>>>> + * mark_linear_text_alias_ro() above (which will be called after >>>>>>> + * alternative patching has completed). This makes the contents >>>>>>> + * of the region accessible to subsystems such as hibernate, >>>>>>> + * but protects it from inadvertent modification or execution. >>>>>>> + * Note that contiguous mappings cannot be remapped in this way, >>>>>>> + * so we should avoid them here. >>>>>>> */ >>>>>>> - memblock_mark_nomap(kernel_start, kernel_end - kernel_start); >>>>>>> + __map_memblock(kernel_start, kernel_end, PAGE_KERNEL, >>>>>>> + flags | NO_CONT_MAPPINGS); >>>>>> >>>>>> So the reason to disallow cont mappings is because we need to modify the >>>>>> permissions later? It _is_ safe to change permissions on a live >>>>>> contiguous >>>>>> mapping in this way. That was clarified in the architecture a couple of >>>>>> years >>>>>> back and we rely on it for contpte_wrprotect_ptes(); see comment there. >>>>>> >>>>> >>>>> OK, good to know - I was hoping to get your take on this ... >>>>> >>>>>> I think we could relax this? >>>>>> >>>>> >>>>> OK, I suppose that means that we can drop the NO_CONT_MAPPINGS here, >>>>> but we still need to map the kernel text/rodata alias initially to >>>>> ensure that no block mappings are created that would need to broken >>>>> down, right? >>>> >>>> Yes, but... >>>> >>>> I think your intent is that the multiple __map_memblock() calls are just >>>> controlling the allowed leaf mapping sizes. >>> >>> Indeed. >>> >>>> It becomes problematic if the 2 >>>> calls use different permissions... which they do. >>>> >>>> PAGE_KERNEL vs pgprot_tagged(PAGE_KERNEL). >>>> >>>> Is it possible that the text/rodata section ends up tagged, which is not >>>> intended? >>>> >>> >>> >>> OK so toggling the R/O attribute on a live contiguous mapping is >>> permitted (provided that ultimately, the entire contiguous region is >>> updated accordingly) but the same doesn't apply to MT_NORMAL vs >>> MT_NORMAL_TAGGED, right? >> >> This is the rule: >> >> RJQQTC >> For a TLB lookup in a contiguous region mapped by translation table entries >> that >> have consistent values for the Contiguous bit, but have the OA, attributes, >> or >> permissions misprogrammed, that TLB lookup is permitted to produce an OA, >> access >> permissions, and memory attributes that are consistent with any one of the >> programmed translation table values. >> >> So, yes, it is fine to modify the attributes (and I assume the memory type >> counts as attributes) on the live mapping... >> > > OK. > >>> >>> So let's use just the same prot and flags for the initial mapping, and >>> use MT_NORMAL_TAGGED when we remap the text alias R/O. >> >> ...but that wasn't my point. I belive the intent is that all of the linear >> map >> uses MT_NORMAL_TAGGED except for the kernel text and rodata, which uses >> MT_NORMAL. In some configs (e.g. force_pte_mapping() == true), your change >> will >> result in all the linear map, including text and rodata being >> MT_NORMAL_TAGGED. >> > > I didn't think that would matter tbh. But if it does, are we fine as > long as we change it back to MT_NORMAL by the time we remap the region > read-only?
I have no idea. I worry MT_NORMAL_TAGGED on the text might imply a performance cost? But not sure about functional... Will ask around.
