Re: [PATCH 1/3] mseal sysmap: Add mseal_system_mappings= boot parameter
From: Chris Down
Date: Tue Oct 27 2026 - 12:41:29 EST
Hi Lorenzo,
Thanks for the quick and thorough review!
Lorenzo Stoakes (ARM) writes:
The existing implementation is a bit of a mess, so any work that extends it
has to put in the work to improve things first.
I _hate_ VM_SEALED_SYSMAP. It's a hack. A tasteful way of doing this is
really required for a series touching this to be taken, honestly.
See below for implementation ideas.
That's fair, I'll do the cleanup first in v2.
(Also it's the merge window, so no more revisions until -rc1, when you need
to resend, thanks.)
Sure, I'll wait until after -rc1.
Could we have some actual justification for the use of the feature please?
So far I've seen arm-waving.
Sure. At Meta, we build sealing into the kernel for our VM hosts, and we'd like it on as much of the rest of the fleet as possible. The problem is that the rest of the fleet also runs things which it breaks. Just as one example among many, Reverie (which Hermit our deterministic execution tool is built on) patches the vDSO of the processes it traces by injecting an mprotect()[0].
So right now the hosts which want sealing need their own kernel image, and this option is one of the things which forces that. We're working towards hopefully running one kernel across the fleet, and with a boot parameter we can build it with CONFIG_MSEAL_SYSTEM_MAPPINGS=y, keep sealing on by default, and turn it off only on the hosts which run those tools.
Let me know if you'd like more information. I can also put that in the commit message for v2.
0: https://docs.rs/reverie-core/latest/reverie/vdso/index.html
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index e75344f4e0cd..b1ec9a8a979e 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -4270,6 +4270,13 @@ Kernel parameters
MTD_Region= [MTD] Format:
<name>,<region-number>[,<base>,<size>,<buswidth>,<altbuswidth>]
+ mseal_system_mappings=
+ [KNL,EARLY] Under CONFIG_MSEAL_SYSTEM_MAPPINGS,
+ whether system mappings are sealed for this boot.
+ See Documentation/userspace-api/mseal.rst.
+ Format: <bool>
+ Default: 1 (enabled).
You _have_ to duplicate the list of things this breaks here please.
Will do.
Let's negate things first. Make the boot flag 'no_mseal_sys_mappings'
rather than confusingly implying mseal_system_mappings=1 could enable it
even with a config option not provided.
Then add something like:
static inline bool mseal_system_mappings(void)
{
return IS_ENABLED(CONFIG_MSEAL_SYSTEM_MAPPINGS) &&
!mseal_system_mappings_disabled;
}
And call that in _install_special_mapping().
Makes sense. I'll call it in __install_special_mapping(), since that's where we already have a vma_flags_t, so it can just set VMA_SEALED_BIT without going near the legacy flags at all, if that sounds good? Then VM_SEALED_SYSMAP and all of its users can hopefully go away.
The LLM overlords tell me that only uprobe trampolines would see a delta
from this given the ARCH_SUPPORTS_MSEAL_SYSTEM_MAPPINGS restriction.
This would improve things a lot.
Yes, but there is a wrinkle. Since 07c308eb2bcf ("uprobes/x86: Do not leak trampoline vma mapping on optimization failure"), __arch_uprobe_optimize() unmaps a trampoline it has just created if swbp_optimize() fails, and with the trampoline sealed that would now fail with -EPERM. I can see two ways to deal with that:
One is to seal trampolines too, and just leave a sealed trampoline in place on that path. It's already a WARN_ON_ONCE(), and get_uprobe_trampoline() reuses any trampoline which is in reach, so nothing is really leaked.
Another way would be to keep trampolines unsealed, but that means bringing back some kind of per-mapping opt-in, which is exactly what we're trying to get rid of.
What do you think? Either way I'll Cc Jiri and Oleg on it.
+#define VM_SEALED_SYSMAP (mseal_sysmap_enabled ? VM_SEALED : VM_NONE)
No, you don't use the legacy VMA flags API for new work. Sorry.
See above as to a pointer as to how this should look.
Sure thing.
static bool range_contains_unmapped(unsigned long start, unsigned long end)
{
VMA_ITERATOR(vmi, current->mm, start);
diff --git a/security/Kconfig b/security/Kconfig
index f7bf6cdc6229..d5e83af1bfc6 100644
--- a/security/Kconfig
+++ b/security/Kconfig
@@ -67,6 +67,8 @@ config MSEAL_SYSTEM_MAPPINGS
of writing includes CHECKPOINT_RESTORE, UML, gVisor, rr. Therefore
this config can't be enabled universally.
+ Sealing can be turned off at boot with "mseal_system_mappings=0".
+
'With this configuration option set, sealing can...' right now this reads a
little like it's defaulted on.
Also can we move this Kconfig option out of security and into mm please?
This is explicitly an mm feature, and it should never have been added there
(mm needs to know if somebody sets this default=y or something for
instance).
Sure thing.
Tests as separate commits please.
Happy to do so, but aren't you worried about bisect safety? Or would you suggest another way?
This seems total overkill, you're checking whether a boot parameter works?
It was mostly there so that the existing test wouldn't fail when booted with mseal_system_mappings=0, which meant parsing a bool. With a negated flag it can be a lot simpler, but I'm not sure which of these you'd prefer:
One option is to skip if [vdso] isn't sealed, and otherwise expect all system mappings to be sealed. Then there's parsing at all, and it also stops failing on kernels built without CONFIG_MSEAL_SYSTEM_MAPPINGS, which matters more once it's in run_vmtests.sh. The downside is that it can't notice sealing being broken entirely.
Another is to check /proc/cmdline for no_mseal_sys_mappings and expect the mappings to be unsealed if it's there. That's only a few lines now, but the test still fails on kernels without the option, as it does today.
Do you have a preference between the two?
Thanks a lot for the review!