Re: [PATCH v2 1/3] arm64: mm: Fix rodata=full block mapping support for realm guests

[Catalin Marinas <catalin.marinas@arm.com>]

On Tue, Apr 07, 2026 at 10:57:35AM +0100, Suzuki K Poulose wrote:
> On 02/04/2026 21:43, Catalin Marinas wrote:
> > On Mon, Mar 30, 2026 at 05:17:02PM +0100, Ryan Roberts wrote:
> > >   int split_kernel_leaf_mapping(unsigned long start, unsigned long end)
> > >   {
> > >   	int ret;
> > > -	/*
> > > -	 * !BBML2_NOABORT systems should not be trying to change permissions on
> > > -	 * anything that is not pte-mapped in the first place. Just return early
> > > -	 * and let the permission change code raise a warning if not already
> > > -	 * pte-mapped.
> > > -	 */
> > > -	if (!system_supports_bbml2_noabort())
> > > -		return 0;
> > > -
> > >   	/*
> > >   	 * If the region is within a pte-mapped area, there is no need to try to
> > >   	 * split. Additionally, CONFIG_DEBUG_PAGEALLOC and CONFIG_KFENCE may
> > >   	 * change permissions from atomic context so for those cases (which are
> > >   	 * always pte-mapped), we must not go any further because taking the
> > > -	 * mutex below may sleep.
> > > +	 * mutex below may sleep. Do not call force_pte_mapping() here because
> > > +	 * it could return a confusing result if called from a secondary cpu
> > > +	 * prior to finalizing caps. Instead, linear_map_requires_bbml2 gives us
> > > +	 * what we need.
> > >   	 */
> > > -	if (force_pte_mapping() || is_kfence_address((void *)start))
> > > +	if (!linear_map_requires_bbml2 || is_kfence_address((void *)start))
> > >   		return 0;
> > > +	if (!system_supports_bbml2_noabort()) {
> > > +		/*
> > > +		 * !BBML2_NOABORT systems should not be trying to change
> > > +		 * permissions on anything that is not pte-mapped in the first
> > > +		 * place. Just return early and let the permission change code
> > > +		 * raise a warning if not already pte-mapped.
> > > +		 */
> > > +		if (system_capabilities_finalized())
> > > +			return 0;
> > > +
> > > +		/*
> > > +		 * Boot-time: split_kernel_leaf_mapping_locked() allocates from
> > > +		 * page allocator. Can't split until it's available.
> > > +		 */
> > > +		if (WARN_ON(!page_alloc_available))
> > > +			return -EBUSY;
> > > +
> > > +		/*
> > > +		 * Boot-time: Started secondary cpus but don't know if they
> > > +		 * support BBML2_NOABORT yet. Can't allow splitting in this
> > > +		 * window in case they don't.
> > > +		 */
> > > +		if (WARN_ON(num_online_cpus() > 1))
> > > +			return -EBUSY;
> > > +	}
> > 
> > I think sashiko is over cautions here
> > (https://sashiko.dev/#/patchset/20260330161705.3349825-1-ryan.roberts@arm.com)
> > but it has a somewhat valid point from the perspective of
> > num_online_cpus() semantics. We have have num_online_cpus() == 1 while
> > having a secondary CPU just booted and with its MMU enabled. I don't
> > think we can have any asynchronous tasks running at that point to
> > trigger a spit though. Even async_init() is called after smp_init().
> > 
> > An option may be to attempt cpus_read_trylock() as this lock is taken by
> > _cpu_up(). If it fails, return -EBUSY, otherwise check num_online_cpus()
> > and unlock (and return -EBUSY if secondaries already started).
> > 
> > Another thing I couldn't get my head around - IIUC is_realm_world()
> > won't return true for map_mem() yet (if in a realm).
> 
> That is correct. map_mem() comes from paginig_init(), which gets called
> before arm64_rsi_init(). Realm check was delayed until psci_xx_init().
> We had a version which parsed the DT for PSCI conduit early enough
> to be able to make the SMC calls to detect the Realm. But there
> were concerns around it.

Ah, yes, I remember.

Does it mean that commit 42be24a4178f ("arm64: Enable memory encrypt for
Realms") was broken without rodata=full w.r.t. the linear map? Commit
a166563e7ec3 ("arm64: mm: support large block mapping when rodata=full")
introduced force_pte_mapping() but it just copied the logic in the
existing can_set_direct_map(). Looking at the linear_map_requires_bbml2
assignment, we get (!is_realm_world() && is_realm_world()) and it
cancels out, no effect on it but we don't get pte mappings either (even
if we don't have BBML2).

I think we need at least some safety checks:

1. BBML2_NOABORT support on the boot CPU - continue with the existing
   logic (as per Ryan's series)

2. !system_supports_bbml2_noabort() - split in
   linear_map_maybe_split_to_ptes(). This does not currently happen
   because linear_map_requires_bbml2 may be false in the absence of
   rodata=full. Not sure how to fix this without some variable telling
   us how the linear map was mapped. The requires_bbml2 flag doesn't

3. Panic in arm64_rsi_init() if !BBML2_NOABORT on the boot CPU _and_ we
   have block mappings already. People can avoid it with rodata=full

4. If (3) is a common case, a better alternative is to rewrite the
   linear map sometime after arm64_rsi_init() but before we call
   split_kernel_leaf_mapping().

-- 
Catalin