Skip to contentSkip to conversation

Looking at PAX_MEMORY_SANITIZE

kernel-hardening Laura Abbott 3 participants 7 messages Dec 16, 2015 – Dec 17, 2015

lore
Share & tools
lore
#0 lore raw

Hi,

I started looking at PAX_MEMORY_SANITIZE for bringing into the kernel. I thought I would give a short update on what I've found so far for feedback and the like.

PAX_MEMORY_SANITIZE is used for clearing both the SL*B allocators and the buddy allocator on free. Arguably, similar behavior exists already as debug features (SLUB_DEBUG poison, DEBUG_PAGEALLOC for some arches). Given what we're already finding with features like DEBUG_RODATA though, the sanitization really needs to be a separate Kconfig not tied to debugging. I debated trying to make those Kconfigs non-debug but they were tied to other features besides poison/ sanitization.

I've been focusing my efforts on the SL*B allocators. As it stands, the feature is fairly self-contained and mostly just needs some refactoring. I plan on expanding the command line option to give a bit more control on where the sanitization happens. The sanitization currently always happens on the fast path so my thought was to allow the option of sanitizing only on the slow path. The existing PaX code also disables cache merging. It's not clear if this is an additional security measure but the sanitization as written doesn't work with merging. For at least the first version, slab merging will be disabled when sanitization is enabled on a slab.

I'm hoping to post actual patches before I go on vacation for the holidays next week. Early feedback is appreciated as well if I missed anything.

Thanks,
Laura

#1 lore raw
Quoted 2 lines · Laura Abbott

I'm hoping to post actual patches before I go on vacation for the holidays next week. Early feedback is appreciated as well if I missed anything.

Note that SANITIZE conflicted with hibernation last time I test, so that's something you might want to test.

Regards,

Signature · 1 line
Yves-Alexis
#2 lore raw
Quoted 7 lines · Yves-Alexis Perez
Quoted 2 lines · Laura Abbott

I'm hoping to post actual patches before I go on vacation for the holidays next week. Early feedback is appreciated as well if I missed anything.

Note that SANITIZE conflicted with hibernation last time I test, so that's something you might want to test.

That's a good point. I think it's only related to the page allocator which I'm not focusing on for my first version but I'll still make sure to test before I send that out.

Thanks,
Laura

#3 lore raw
Quoted 9 lines · Laura Abbott
Quoted 8 lines · Yves-Alexis Perez
Quoted 4 lines · Laura Abbott

I'm hoping to post actual patches before I go on vacation for the holidays next
week. Early feedback is appreciated as well if I missed anything.

Quoted 8 lines
Quoted 3 lines

Note that SANITIZE conflicted with hibernation last time I test, so that's something you might want to test.

That's a good point. I think it's only related to the page allocator which I'm
not focusing on for my first version but I'll still make sure to test before I send that out.

It's also easy now to have runtime conflict with hibernation (this is what kASLR does to avoid a CONFIG conflict with hibernation, via __setup("kaslr", kaslr_nohibernate_setup)).

-Kees

Signature · 2 lines
Kees Cook
Chrome OS & Brillo Security
#4 lore raw
In reply to Laura Abbott
Quoted 7 lines · Laura Abbott

Hi,

I started looking at PAX_MEMORY_SANITIZE for bringing into the kernel. I thought
I would give a short update on what I've found so far for feedback and the like.

Quoted 12 lines

PAX_MEMORY_SANITIZE is used for clearing both the SL*B allocators and the buddy allocator on free. Arguably, similar behavior exists already as debug features (SLUB_DEBUG poison, DEBUG_PAGEALLOC for some arches). Given what we're
already finding with features like DEBUG_RODATA though, the sanitization really
needs to be a separate Kconfig not tied to debugging. I debated trying to make
those Kconfigs non-debug but they were tied to other features besides poison/
sanitization.

Sounds great! As for CONFIGs, one thing that was request was that we ultimately put all the hardening configs under a common Kconfig heading so they can be found easily. I view this as a late-stage tweak, and mostly we just need to worry about functionality first.

For upstreaming, we'll need a comparison between PAX_MEMORY_SANITIZE and the existing upstream things that overlap with it.

Quoted 8 lines

I've been focusing my efforts on the SL*B allocators. As it stands, the feature
is fairly self-contained and mostly just needs some refactoring. I plan on expanding the command line option to give a bit more control on where the sanitization happens. The sanitization currently always happens on the fast path
so my thought was to allow the option of sanitizing only on the slow path.

Seems like it should happen on all paths to be a true always-on hardening feature, though?

Quoted 7 lines

The existing PaX code also disables cache merging. It's not clear if this is an
additional security measure but the sanitization as written doesn't work with
merging. For at least the first version, slab merging will be disabled when sanitization is enabled on a slab.

That seems fine to me. These trade-offs are to be expected.

Another thing we should call attention to are past bugs and exploits or exploit methods that would have been hampered or eliminated with this feature. e.g. How does this play with classic use-after-free, etc?

I'm hoping to post actual patches before I go on vacation for the holidays

Quoted 3 lines

next
week. Early feedback is appreciated as well if I missed anything.

Fantastic! I look forward to testing it. Which gets to a third piece:
adding an lkdtm test for the feature so it's easy to test the kernel's reaction to the state that is being protected.

Thanks for working on this!

-Kees

Signature · 2 lines
Kees Cook
Chrome OS & Brillo Security
#5 lore raw
Quoted 18 lines · Kees Cook

On Wed, Dec 16, 2015 at 10:46 AM, Laura Abbott <[email protected] <mailto:[email protected]>> wrote:

Hi,

I started looking at PAX_MEMORY_SANITIZE for bringing into the kernel. I thought I would give a short update on what I've found so far for feedback and the like.

PAX_MEMORY_SANITIZE is used for clearing both the SL*B allocators and the buddy allocator on free. Arguably, similar behavior exists already as debug features (SLUB_DEBUG poison, DEBUG_PAGEALLOC for some arches). Given what we're already finding with features like DEBUG_RODATA though, the sanitization really needs to be a separate Kconfig not tied to debugging. I debated trying to make those Kconfigs non-debug but they were tied to other features besides poison/ sanitization.

Quoted 4 lines

Sounds great! As for CONFIGs, one thing that was request was that we ultimately put all the hardening configs under a common Kconfig heading so they can be found easily. I view this as a late-stage tweak, and mostly we just need to worry about functionality first.

Sure. I was going to put the Kconfig somewhere I think is reasonable and see if there are better suggestions.

Quoted 2 lines

For upstreaming, we'll need a comparison between PAX_MEMORY_SANITIZE and the existing upstream things that overlap with it.

What kind of comparison are you thinking of? Just a list of what differs featurewise or performance characteristics or something else?

Quoted 9 lines

I've been focusing my efforts on the SL*B allocators. As it stands, the feature is fairly self-contained and mostly just needs some refactoring. I plan on expanding the command line option to give a bit more control on where the sanitization happens. The sanitization currently always happens on the fast path so my thought was to allow the option of sanitizing only on the slow path.

Seems like it should happen on all paths to be a true always-on hardening feature, though?

The existing PaX code has an option to skip caches marked with SLAB_NO_SANITIZE. I was thinking skipping on the fast path might be a similar expansion of that. There is still a boot commandline option for full sanitization.

I'm going to profile before I send anything out to get a better idea if this option would make a difference.

Quoted 7 lines

The existing PaX code also disables cache merging. It's not clear if this is an additional security measure but the sanitization as written doesn't work with merging. For at least the first version, slab merging will be disabled when sanitization is enabled on a slab.

Quoted 5 lines

That seems fine to me. These trade-offs are to be expected.

Another thing we should call attention to are past bugs and exploits or exploit methods that would have been hampered or eliminated with this feature. e.g. How does this play with classic use-after-free, etc?

Good point. I'll find some examples.

Quoted 8 lines

I'm hoping to post actual patches before I go on vacation for the holidays next week. Early feedback is appreciated as well if I missed anything.

Fantastic! I look forward to testing it. Which gets to a third piece: adding an lkdtm test for the feature so it's easy to test the kernel's reaction to the state that is being protected.

I'll look into this as well.

Thanks,
Laura

#6 lore raw
Quoted 41 lines · Laura Abbott
Quoted 28 lines · Kees Cook

On Wed, Dec 16, 2015 at 10:46 AM, Laura Abbott <[email protected] <mailto:[email protected]>> wrote:

Hi,

I started looking at PAX_MEMORY_SANITIZE for bringing into the kernel.
I thought
I would give a short update on what I've found so far for feedback and
the like.

PAX_MEMORY_SANITIZE is used for clearing both the SL*B allocators and
the
buddy allocator on free. Arguably, similar behavior exists already as
debug
features (SLUB_DEBUG poison, DEBUG_PAGEALLOC for some arches). Given
what we're
already finding with features like DEBUG_RODATA though, the
sanitization really
needs to be a separate Kconfig not tied to debugging. I debated trying
to make
those Kconfigs non-debug but they were tied to other features besides
poison/
sanitization.

Quoted 7 lines

Sounds great! As for CONFIGs, one thing that was request was that we ultimately put all
the hardening configs under a common Kconfig heading so they can be found easily. I view
this as a late-stage tweak, and mostly we just need to worry about functionality first.

Sure. I was going to put the Kconfig somewhere I think is reasonable and see if
there are better suggestions.

Quoted 8 lines
Quoted 3 lines

For upstreaming, we'll need a comparison between PAX_MEMORY_SANITIZE and the existing
upstream things that overlap with it.

What kind of comparison are you thinking of? Just a list of what differs featurewise
or performance characteristics or something else?

Yeah, I meant features, but I'm sure we'll need to produce performance details too.

Quoted 61 lines
Quoted 14 lines

I've been focusing my efforts on the SL*B allocators. As it stands,
the feature
is fairly self-contained and mostly just needs some refactoring. I
plan on
expanding the command line option to give a bit more control on where
the
sanitization happens. The sanitization currently always happens on the
fast path
so my thought was to allow the option of sanitizing only on the slow
path.

Quoted 2 lines

Seems like it should happen on all paths to be a true always-on hardening feature, though?

The existing PaX code has an option to skip caches marked with SLAB_NO_SANITIZE.
I was thinking skipping on the fast path might be a similar expansion of that.
There is still a boot commandline option for full sanitization.

I'm going to profile before I send anything out to get a better idea if this option would make a difference.

Quoted 10 lines

The existing PaX code also disables cache merging. It's not clear if
this is an
additional security measure but the sanitization as written doesn't
work with
merging. For at least the first version, slab merging will be disabled
when
sanitization is enabled on a slab.

Quoted 7 lines

That seems fine to me. These trade-offs are to be expected.

Another thing we should call attention to are past bugs and exploits or exploit methods
that would have been hampered or eliminated with this feature. e.g. How does this
play with classic use-after-free, etc?

Good point. I'll find some examples.

Quoted 11 lines

I'm hoping to post actual patches before I go on vacation for the
holidays next
week. Early feedback is appreciated as well if I missed anything.

Fantastic! I look forward to testing it. Which gets to a third piece:
adding an lkdtm
test for the feature so it's easy to test the kernel's reaction to the state that is
being protected.

Quoted 4 lines

I'll look into this as well.

Thanks,
Laura

Thanks!

-Kees

Signature · 2 lines
Kees Cook
Chrome OS & Brillo Security