[RFC PATCH v2 0/4] arm64: realm: Support for probing RSI earlier

[Suzuki K Poulose]

On 11/06/2026 12:14, Catalin Marinas wrote:
> On Fri, May 29, 2026 at 01:27:01PM +0100, Suzuki K Poulose wrote:
>> On 28/05/2026 17:06, Catalin Marinas wrote:
>>> On Tue, May 05, 2026 at 04:57:38PM +0100, Suzuki K Poulose wrote:
>>>> This is an updated series, addressing the review comments from AI agent on
>>>> the version 1 [0] of the series, (some of which were documented as short comings).
>>>> See below for the changes.
>>>>
>>>> The Realm Guest linux support is broken without rodata=full (fortunately default
>>>> for arm64), as we detect the RSI support after we have created the Linear map
>>>> with Block/Contiguous mappings. If the boot CPU doesn't support BBML2_NOABORT
>>>> (there are CPUs out there with FEAT_RME and no - useable - BBML2_NOABORT)
>>>> we are then not able to split the page tables down to PTE level if the system
>>>> as such doesn't support BBML2.
>>>>
>>>> See the following link for the discussion.
>>>>
>>>> https://lore.kernel.org/all/20260330161705.3349825-2-ryan.roberts@arm.com/
>>>>
>>>> The available options are :
>>>>    1. Start with PTE level mappings at paging_init() and then "FOLD" the page tables
>>>>       to Block/Cont mappings after we have the full picture available. Looking at the
>>>>       future (with BBML3), this might mean "additional work" for most of the systems
>>>>       at boot. But not bad as splitting them ?
>>>>    2. Hold the secondary CPUs in busy loop with MMU disabled and split the mappings
>>>>       by the boot CPU with MMU off (if Boot CPU can't support BBML2). This is tricky
>>>>       with the page allocations required to add the page-tables.
>>>>    3. Move the detection of Realm support earlier to make a better decision for
>>>>       paging_init(), with an added bonus of earlycon support for Realms without
>>>>       the user having to work out the "top bit" for the Realm.
>>>>
>>>> This series is an attempt to implement (3) (without the earlycon support). We try
>>>> to probe the PSCI conduit early from the DT/ACPI. DT is not flattened at this time.
> [...]
>>> Could we instead add a more informative message in arm64_rsi_init() if
>>> !force_pte_mappings() && !cpu_supports_bbml2_noabort() (before
>>> is_realm_world() becomes true)? Well, it may not print anything if the
>>> early console is not set up yet.
>>
>> That is true, but with some expertise you may be able to enable earlycon
>> and may be we could get some new mechanism for "earlycon"  for Realms.
>>
>> The other way to look at is:
>>
>> When the system doesn't support BBML2 Abort:
>>
>> Creating block/Cont mappings to start with and then splitting it to PTE
>> is quite difficult as we :
>> 1. Need to allocate pages for leaf level tables
>> 2. Hold the other CPUs in tight loop
> 
> Agree, that's not easily possible at runtime.
> 
>> Instead, creating the block/CONT levels from a fully "page level"
>> mappings are easier, as we can:
>>
>> 1. Can easily fold the tables to Block mapping with reclaiming the leaf
>> level pagetables.
>>
>> 2. Avoid the secondary CPUs dance, as they all support BBML2_NOABORT.
>>
>> This shouldn't be that bad as the opposite ?
> 
> I don't think it solves our problem. Aren't we concerned with the
> rodata=off && !BBML2_NOABORT && is_realm_world() case? I don't think
> your second point stands.
> 
> Currently we have:
> 
> rodata=full && BBML2_NOABORT => block mappings irrespective of realms
> 
> rodata=off && BBML2_NOABORT  => block mappings first, can be split later
> 				if is_realm_world()
> 
> rodata=off && !BBML2_NOABORT => block mappings first, serious problem if
> 				is_realm_world()
> 
> It's the last case we need to fix. Starting with page mappings does
> avoid the in-realm failure but the !is_realm_world() case folding to
> block mappings still requires proper BBM.

I see, the case I was missing is : !is_realm_world() and !BBML2_NO_ABORT
and we want Block mapping if rodata=off. Yes, in this case we need the 
secondaries on hold, with proper BBM on the boot CPU too. Again, it is
easier to "collapsing the tables to Block" than the reverse.

Suzuki






>