codeblog code is freedom — patching my itch

Previously: v4.8.

Here are a bunch of security things I’m excited about in the newly released Linux v4.9:

Latent Entropy GCC plugin

Building on her earlier work to bring GCC plugin support to the Linux kernel, Emese Revfy ported PaX’s Latent Entropy GCC plugin to upstream. This plugin is significantly more complex than the others that have already been ported, and performs extensive instrumentation of functions marked with __latent_entropy. These functions have their branches and loops adjusted to mix random values (selected at build time) into a global entropy gathering variable. Since the branch and loop ordering is very specific to boot conditions, CPU quirks, memory layout, etc, this provides some additional uncertainty to the kernel’s entropy pool. Since the entropy actually gathered is hard to measure, no entropy is “credited”, but rather used to mix the existing pool further. Probably the best place to enable this plugin is on small devices without other strong sources of entropy.

vmapped kernel stack and thread_info relocation on x86

Normally, kernel stacks are mapped together in memory. This meant that attackers could use forms of stack exhaustion (or stack buffer overflows) to reach past the end of a stack and start writing over another process’s stack. This is bad, and one way to stop it is to provide guard pages between stacks, which is provided by vmalloced memory. Andy Lutomirski did a bunch of work to move to vmapped kernel stack via CONFIG_VMAP_STACK on x86_64. Now when writing past the end of the stack, the kernel will immediately fault instead of just continuing to blindly write.

Related to this, the kernel was storing thread_info (which contained sensitive values like addr_limit) at the bottom of the kernel stack, which was an easy target for attackers to hit. Between a combination of explicitly moving targets out of thread_info, removing needless fields, and entirely moving thread_info off the stack, Andy Lutomirski and Linus Torvalds created CONFIG_THREAD_INFO_IN_TASK for x86.

CONFIG_DEBUG_RODATA mandatory on arm64

As recently done for x86, Mark Rutland made CONFIG_DEBUG_RODATA mandatory on arm64. This feature controls whether the kernel enforces proper memory protections on its own memory regions (code memory is executable and read-only, read-only data is actually read-only and non-executable, and writable data is non-executable). This protection is a fundamental security primitive for kernel self-protection, so there’s no reason to make the protection optional.

random_page() cleanup

Cleaning up the code around the userspace ASLR implementations makes them easier to reason about. This has been happening for things like the recent consolidation on arch_mmap_rnd() for ET_DYN and during the addition of the entropy sysctl. Both uncovered some awkward uses of get_random_int() (or similar) in and around arch_mmap_rnd() (which is used for mmap (and therefore shared library) and PIE ASLR), as well as in randomize_stack_top() (which is used for stack ASLR). Jason Cooper cleaned things up further by doing away with randomize_range() entirely and replacing it with the saner random_page(), making the per-architecture arch_randomize_brk() (responsible for brk ASLR) much easier to understand.

Edit: missed this next feature when I first posted!

User namespace restrictions

Eric Biederman landed a sysctl to control how many user namespaces can be created (on a per-user and per-namespace basis) via /proc/sys/user/max_user_namespaces. The ability to limit access to user namespaces has been desired for a while, since a large amount of attack surface gets exposed from within a user namespace, due to it allowing various interfaces to operate within the namespace that were originally limited to the true root user before. A system can now disable user namespaces by writing 0 to the sysctl.

That’s it for now! Let me know if there are other fun things to call attention to in v4.10.

© 2016 – 2017, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

My prior post showed my research from earlier in the year at the 2016 Linux Security Summit on kernel security flaw lifetimes. Now that CVE-2016-5195 is public, here are updated graphs and statistics. Due to their rarity, the Critical bug average has now jumped from 3.3 years to 5.2 years. There aren’t many, but, as I mentioned, they still exist, whether you know about them or not. CVE-2016-5195 was sitting on everyone’s machine when I gave my LSS talk, and there are still other flaws on all our Linux machines right now. (And, I should note, this problem is not unique to Linux.) Dealing with knowing that there are always going to be bugs present requires proactive kernel self-protection (to minimize the effects of possible flaws) and vendors dedicated to updating their devices regularly and quickly (to keep the exposure window minimized once a flaw is widely known).

So, here are the graphs updated for the 668 CVEs known today:

• Critical: 3 @ 5.2 years average
• High: 44 @ 6.2 years average
• Medium: 404 @ 5.3 years average
• Low: 216 @ 5.5 years average

© 2016, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

In several of my recent presentations, I’ve discussed the lifetime of security flaws in the Linux kernel. Jon Corbet did an analysis in 2010, and found that security bugs appeared to have roughly a 5 year lifetime. As in, the flaw gets introduced in a Linux release, and then goes unnoticed by upstream developers until another release 5 years later, on average. I updated this research for 2011 through 2016, and used the Ubuntu Security Team’s CVE Tracker to assist in the process. The Ubuntu kernel team already does the hard work of trying to identify when flaws were introduced in the kernel, so I didn’t have to re-do this for the 557 kernel CVEs since 2011.

As the README details, the raw CVE data is spread across the active/, retired/, and ignored/ directories. By scanning through the CVE files to find any that contain the line “Patches_linux:”, I can extract the details on when a flaw was introduced and when it was fixed. For example CVE-2016-0728 shows:

Patches_linux:
 break-fix: 3a50597de8635cd05133bd12c95681c82fe7b878 23567fd052a9abb6d67fe8e7a9ccdd9800a540f2

This means that CVE-2016-0728 is believed to have been introduced by commit 3a50597de8635cd05133bd12c95681c82fe7b878 and fixed by commit 23567fd052a9abb6d67fe8e7a9ccdd9800a540f2. If there are multiple lines, then there may be multiple SHAs identified as contributing to the flaw or the fix. And a “-” is just short-hand for the start of Linux git history.

Then for each SHA, I queried git to find its corresponding release, and made a mapping of release version to release date, wrote out the raw data, and rendered graphs. Each vertical line shows a given CVE from when it was introduced to when it was fixed. Red is “Critical”, orange is “High”, blue is “Medium”, and black is “Low”:

And here it is zoomed in to just Critical and High:

The line in the middle is the date from which I started the CVE search (2011). The vertical axis is actually linear time, but it’s labeled with kernel releases (which are pretty regular). The numerical summary is:

• Critical: 2 @ 3.3 years
• High: 34 @ 6.4 years
• Medium: 334 @ 5.2 years
• Low: 186 @ 5.0 years

This comes out to roughly 5 years lifetime again, so not much has changed from Jon’s 2010 analysis.

While we’re getting better at fixing bugs, we’re also adding more bugs. And for many devices that have been built on a given kernel version, there haven’t been frequent (or some times any) security updates, so the bug lifetime for those devices is even longer. To really create a safe kernel, we need to get proactive about self-protection technologies. The systems using a Linux kernel are right now running with security flaws. Those flaws are just not known to the developers yet, but they’re likely known to attackers, as there have been prior boasts/gray-market advertisements for at least CVE-2010-3081 and CVE-2013-2888.

(Edit: see my updated graphs that include CVE-2016-5195.)

© 2016, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

Previously: v4.7. Here are a bunch of security things I’m excited about in Linux v4.8:

SLUB freelist ASLR

Thomas Garnier continued his freelist randomization work by adding SLUB support.

x86_64 KASLR text base offset physical/virtual decoupling

On x86_64, to implement the KASLR text base offset, the physical memory location of the kernel was randomized, which resulted in the virtual address being offset as well. Due to how the kernel’s “-2GB” addressing works (gcc‘s “-mcmodel=kernel“), it wasn’t possible to randomize the physical location beyond the 2GB limit, leaving any additional physical memory unused as a randomization target. In order to decouple the physical and virtual location of the kernel (to make physical address exposures less valuable to attackers), the physical location of the kernel needed to be randomized separately from the virtual location. This required a lot of work for handling very large addresses spanning terabytes of address space. Yinghai Lu, Baoquan He, and I landed a series of patches that ultimately did this (and in the process fixed some other bugs too). This expands the physical offset entropy to roughly $physical_memory_size_of_system / 2MB bits.

x86_64 KASLR memory base offset

Thomas Garnier rolled out KASLR to the kernel’s various statically located memory ranges, randomizing their locations with CONFIG_RANDOMIZE_MEMORY. One of the more notable things randomized is the physical memory mapping, which is a known target for attacks. Also randomized is the vmalloc area, which makes attacks against targets vmalloced during boot (which tend to always end up in the same location on a given system) are now harder to locate. (The vmemmap region randomization accidentally missed the v4.8 window and will appear in v4.9.)

x86_64 KASLR with hibernation

Rafael Wysocki (with Thomas Garnier, Borislav Petkov, Yinghai Lu, Logan Gunthorpe, and myself) worked on a number of fixes to hibernation code that, even without KASLR, were coincidentally exposed by the earlier W^X fix. With that original problem fixed, then memory KASLR exposed more problems. I’m very grateful everyone was able to help out fixing these, especially Rafael and Thomas. It’s a hard place to debug. The bottom line, now, is that hibernation and KASLR are no longer mutually exclusive.

gcc plugin infrastructure

Emese Revfy ported the PaX/Grsecurity gcc plugin infrastructure to upstream. If you want to perform compiler-based magic on kernel builds, now it’s much easier with CONFIG_GCC_PLUGINS! The plugins live in scripts/gcc-plugins/. Current plugins are a short example called “Cyclic Complexity” which just emits the complexity of functions as they’re compiled, and “Sanitizer Coverage” which provides the same functionality as gcc’s recent “-fsanitize-coverage=trace-pc” but back through gcc 4.5. Another notable detail about this work is that it was the first Linux kernel security work funded by Linux Foundation’s Core Infrastructure Initiative. I’m looking forward to more plugins!

If you’re on Debian or Ubuntu, the required gcc plugin headers are available via the gcc-$N-plugin-dev package (and similarly for all cross-compiler packages).

hardened usercopy

Along with work from Rik van Riel, Laura Abbott, Casey Schaufler, and many other folks doing testing on the KSPP mailing list, I ported part of PAX_USERCOPY (the basic runtime bounds checking) to upstream as CONFIG_HARDENED_USERCOPY. One of the interface boundaries between the kernel and user-space are the copy_to_user()/copy_from_user() family of functions. Frequently, the size of a copy is known at compile-time (“built-in constant”), so there’s not much benefit in checking those sizes (hardened usercopy avoids these cases). In the case of dynamic sizes, hardened usercopy checks for 3 areas of memory: slab allocations, stack allocations, and kernel text. Direct kernel text copying is simply disallowed. Stack copying is allowed as long as it is entirely contained by the current stack memory range (and on x86, only if it does not include the saved stack frame and instruction pointers). For slab allocations (e.g. those allocated through kmem_cache_alloc() and the kmalloc()-family of functions), the copy size is compared against the size of the object being copied. For example, if copy_from_user() is writing to a structure that was allocated as size 64, but the copy gets tricked into trying to write 65 bytes, hardened usercopy will catch it and kill the process.

For testing hardened usercopy, lkdtm gained several new tests: USERCOPY_HEAP_SIZE_TO, USERCOPY_HEAP_SIZE_FROM, USERCOPY_STACK_FRAME_TO,
USERCOPY_STACK_FRAME_FROM, USERCOPY_STACK_BEYOND, and USERCOPY_KERNEL. Additionally, USERCOPY_HEAP_FLAG_TO and USERCOPY_HEAP_FLAG_FROM were added to test what will be coming next for hardened usercopy: flagging slab memory as “safe for copy to/from user-space”, effectively whitelisting certainly slab caches, as done by PAX_USERCOPY. This further reduces the scope of what’s allowed to be copied to/from, since most kernel memory is not intended to ever be exposed to user-space. Adding this logic will require some reorganization of usercopy code to add some new APIs, as PAX_USERCOPY’s approach to handling special-cases is to add bounce-copies (copy from slab to stack, then copy to userspace) as needed, which is unlikely to be acceptable upstream.

seccomp reordered after ptrace

By its original design, seccomp filtering happened before ptrace so that seccomp-based ptracers (i.e. SECCOMP_RET_TRACE) could explicitly bypass seccomp filtering and force a desired syscall. Nothing actually used this feature, and as it turns out, it’s not compatible with process launchers that install seccomp filters (e.g. systemd, lxc) since as long as the ptrace and fork syscalls are allowed (and fork is needed for any sensible container environment), a process could spawn a tracer to help bypass a filter by injecting syscalls. After Andy Lutomirski convinced me that ordering ptrace first does not change the attack surface of a running process (unless all syscalls are blacklisted, the entire ptrace attack surface will always be exposed), I rearranged things. Now there is no (expected) way to bypass seccomp filters, and containers with seccomp filters can allow ptrace again.

Edit: missed this next feature when I originally posted

NX stack and heap on MIPS

Other architectures have had a non-executable stack and heap for a while now, and now MIPS has caught up, thanks to Paul Barton. The primary reason for the delay was finding a way to cleanly deal with branch delay slot instructions which needed a place to write instructions. Traditionally this was on the stack, but now it’s handled with a per-mm page.

That’s it for v4.8! The merge window is open for v4.9…

© 2016 – 2017, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

Previously: v4.6. Onward to security things I found interesting in Linux v4.7:

KASLR text base offset for MIPS

Matt Redfearn added text base address KASLR to MIPS, similar to what’s available on x86 and arm64. As done with x86, MIPS attempts to gather entropy from various build-time, run-time, and CPU locations in an effort to find reasonable sources during early-boot. MIPS doesn’t yet have anything as strong as x86’s RDRAND (though most have an instruction counter like x86’s RDTSC), but it does have the benefit of being able to use Device Tree (i.e. the “/chosen/kaslr-seed” property) like arm64 does. By my understanding, even without Device Tree, MIPS KASLR entropy should be as strong as pre-RDRAND x86 entropy, which is more than sufficient for what is, similar to x86, not a huge KASLR range anyway: default 8 bits (a span of 16MB with 64KB alignment), though CONFIG_RANDOMIZE_BASE_MAX_OFFSET can be tuned to the device’s memory, giving a maximum of 11 bits on 32-bit, and 15 bits on EVA or 64-bit.

SLAB freelist ASLR

Thomas Garnier added CONFIG_SLAB_FREELIST_RANDOM to make slab allocation layouts less deterministic with a per-boot randomized freelist order. This raises the bar for successful kernel slab attacks. Attackers will need to either find additional bugs to help leak slab layout information or will need to perform more complex grooming during an attack. Thomas wrote a post describing the feature in more detail here: Randomizing the Linux kernel heap freelists. (SLAB is done in v4.7, and SLUB in v4.8.)

eBPF JIT constant blinding

Daniel Borkmann implemented constant blinding in the eBPF JIT subsystem. With strong kernel memory protections (CONFIG_DEBUG_RODATA) in place, and with the segregation of user-space memory execution from kernel (i.e SMEP, PXN, CONFIG_CPU_SW_DOMAIN_PAN), having a place where user-space can inject content into an executable area of kernel memory becomes very high-value to an attacker. The eBPF JIT was exactly such a thing: the use of BPF constants could result in the JIT producing instruction flows that could include attacker-controlled instructions (e.g. by directing execution into the middle of an instruction with a constant that would be interpreted as a native instruction). The eBPF JIT already uses a number of other defensive tricks (e.g. random starting position), but this added randomized blinding to any BPF constants, which makes building a malicious execution path in the eBPF JIT memory much more difficult (and helps block attempts at JIT spraying to bypass other protections).

Elena Reshetova updated a 2012 proof-of-concept attack to succeed against modern kernels to help provide a working example of what needed fixing in the JIT. This serves as a thorough regression test for the protection.

The cBPF JITs that exist in ARM, MIPS, PowerPC, and Sparc still need to be updated to eBPF, but when they do, they’ll gain all these protections immediatley.

Bottom line is that if you enable the (disabled-by-default) bpf_jit_enable sysctl, be sure to set the bpf_jit_harden sysctl to 2 (to perform blinding even for root).

fix brk ASLR weakness on arm64 compat

There have been a few ASLR fixes recently (e.g. ET_DYN, x86 32-bit unlimited stack), and while reviewing some suggested fixes to arm64 brk ASLR code from Jon Medhurst, I noticed that arm64’s brk ASLR entropy was slightly too low (less than 1 bit) for 64-bit and noticeably lower (by 2 bits) for 32-bit compat processes when compared to native 32-bit arm. I simplified the code by using literals for the entropy. Maybe we can add a sysctl some day to control brk ASLR entropy like was done for mmap ASLR entropy.

LoadPin LSM

LSM stacking is well-defined since v4.2, so I finally upstreamed a “small” LSM that implements a protection I wrote for Chrome OS several years back. On systems with a static root of trust that extends to the filesystem level (e.g. Chrome OS’s coreboot+depthcharge boot firmware chaining to dm-verity, or a system booting from read-only media), it’s redundant to sign kernel modules (you’ve already got the modules on read-only media: they can’t change). The kernel just needs to know they’re all coming from the correct location. (And this solves loading known-good firmware too, since there is no convention for signed firmware in the kernel yet.) LoadPin requires that all modules, firmware, etc come from the same mount (and assumes that the first loaded file defines which mount is “correct”, hence load “pinning”).

That’s it for v4.7. Prepare yourself for v4.8 next!

© 2016, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

Previously: v4.5. The v4.6 Linux kernel release included a bunch of stuff, with much more of it under the KSPP umbrella.

seccomp support for parisc

Helge Deller added seccomp support for parisc, which including plumbing support for PTRACE_GETREGSET to get the self-tests working.

x86 32-bit mmap ASLR vs unlimited stack fixed

Hector Marco-Gisbert removed a long-standing limitation to mmap ASLR on 32-bit x86, where setting an unlimited stack (e.g. “ulimit -s unlimited“) would turn off mmap ASLR (which provided a way to bypass ASLR when executing setuid processes). Given that ASLR entropy can now be controlled directly (see the v4.5 post), and that the cases where this created an actual problem are very rare, means that if a system sees collisions between unlimited stack and mmap ASLR, they can just adjust the 32-bit ASLR entropy instead.

x86 execute-only memory

Dave Hansen added Protection Key support for future x86 CPUs and, as part of this, implemented support for “execute only” memory in user-space. On pkeys-supporting CPUs, using mmap(..., PROT_EXEC) (i.e. without PROT_READ) will mean that the memory can be executed but cannot be read (or written). This provides some mitigation against automated ROP gadget finding where an executable is read out of memory to find places that can be used to build a malicious execution path. Using this will require changing some linker behavior (to avoid putting data in executable areas), but seems to otherwise Just Work. I’m looking forward to either emulated QEmu support or access to one of these fancy CPUs.

CONFIG_DEBUG_RODATA enabled by default on arm and arm64, and mandatory on x86

Ard Biesheuvel (arm64) and I (arm) made the poorly-named CONFIG_DEBUG_RODATA enabled by default. This feature controls whether the kernel enforces proper memory protections on its own memory regions (code memory is executable and read-only, read-only data is actually read-only and non-executable, and writable data is non-executable). This protection is a fundamental security primitive for kernel self-protection, so making it on-by-default is required to start any kind of attack surface reduction within the kernel.

On x86 CONFIG_DEBUG_RODATA was already enabled by default, but, at Ingo Molnar’s suggestion, I made it mandatory: CONFIG_DEBUG_RODATA cannot be turned off on x86. I expect we’ll get there with arm and arm64 too, but the protection is still somewhat new on these architectures, so it’s reasonable to continue to leave an “out” for developers that find themselves tripping over it.

arm64 KASLR text base offset

Ard Biesheuvel reworked a ton of arm64 infrastructure to support kernel relocation and, building on that, Kernel Address Space Layout Randomization of the kernel text base offset (and module base offset). As with x86 text base KASLR, this is a probabilistic defense that raises the bar for kernel attacks where finding the KASLR offset must be added to the chain of exploits used for a successful attack. One big difference from x86 is that the entropy for the KASLR must come either from Device Tree (in the “/chosen/kaslr-seed” property) or from UEFI (via EFI_RNG_PROTOCOL), so if you’re building arm64 devices, make sure you have a strong source of early-boot entropy that you can expose through your boot-firmware or boot-loader.

zero-poison after free

Laura Abbott reworked a bunch of the kernel memory management debugging code to add zeroing of freed memory, similar to PaX/Grsecurity’s PAX_MEMORY_SANITIZE feature. This feature means that memory is cleared at free, wiping any sensitive data so it doesn’t have an opportunity to leak in various ways (e.g. accidentally uninitialized structures or padding), and that certain types of use-after-free flaws cannot be exploited since the memory has been wiped. To take things even a step further, the poisoning can be verified at allocation time to make sure that nothing wrote to it between free and allocation (called “sanity checking”), which can catch another small subset of flaws.

To understand the pieces of this, it’s worth describing that the kernel’s higher level allocator, the “page allocator” (e.g. __get_free_pages()) is used by the finer-grained “slab allocator” (e.g. kmem_cache_alloc(), kmalloc()). Poisoning is handled separately in both allocators. The zero-poisoning happens at the page allocator level. Since the slab allocators tend to do their own allocation/freeing, their poisoning happens separately (since on slab free nothing has been freed up to the page allocator).

Only limited performance tuning has been done, so the penalty is rather high at the moment, at about 9% when doing a kernel build workload. Future work will include some exclusion of frequently-freed caches (similar to PAX_MEMORY_SANITIZE), and making the options entirely CONFIG controlled (right now both CONFIGs are needed to build in the code, and a kernel command line is needed to activate it). Performing the sanity checking (mentioned above) adds another roughly 3% penalty. In the general case (and once the performance of the poisoning is improved), the security value of the sanity checking isn’t worth the performance trade-off.

Tests for the features can be found in lkdtm as READ_AFTER_FREE and READ_BUDDY_AFTER_FREE. If you’re feeling especially paranoid and have enabled sanity-checking, WRITE_AFTER_FREE and WRITE_BUDDY_AFTER_FREE can test these as well.

To perform zero-poisoning of page allocations and (currently non-zero) poisoning of slab allocations, build with:

CONFIG_DEBUG_PAGEALLOC=n
CONFIG_PAGE_POISONING=y
CONFIG_PAGE_POISONING_NO_SANITY=y
CONFIG_PAGE_POISONING_ZERO=y
CONFIG_SLUB_DEBUG=y

and enable the page allocator poisoning and slab allocator poisoning at boot with this on the kernel command line:

page_poison=on slub_debug=P

To add sanity-checking, change PAGE_POISONING_NO_SANITY=n, and add “F” to slub_debug as “slub_debug=PF“.

read-only after init

I added the infrastructure to support making certain kernel memory read-only after kernel initialization (inspired by a small part of PaX/Grsecurity’s KERNEXEC functionality). The goal is to continue to reduce the attack surface within the kernel by making even more of the memory, especially function pointer tables, read-only (which depends on CONFIG_DEBUG_RODATA above).

Function pointer tables (and similar structures) are frequently targeted by attackers when redirecting execution. While many are already declared “const” in the kernel source code, making them read-only (and therefore unavailable to attackers) for their entire lifetime, there is a class of variables that get initialized during kernel (and module) start-up (i.e. written to during functions that are marked “__init“) and then never (intentionally) written to again. Some examples are things like the VDSO, vector tables, arch-specific callbacks, etc.

As it turns out, most architectures with kernel memory protection already delay making their data read-only until after __init (see mark_rodata_ro()), so it’s trivial to declare a new data section (“.data..ro_after_init“) and add it to the existing read-only data section (“.rodata“). Kernel structures can be annotated with the new section (via the “__ro_after_init” macro), and they’ll become read-only once boot has finished.

The next step for attack surface reduction infrastructure will be to create a kernel memory region that is passively read-only, but can be made temporarily writable (by a single un-preemptable CPU), for storing sensitive structures that are written to only very rarely. Once this is done, much more of the kernel’s attack surface can be made read-only for the majority of its lifetime.

As people identify places where __ro_after_init can be used, we can grow the protection. A good place to start is to look through the PaX/Grsecurity patch to find uses of __read_only on variables that are only written to during __init functions. The rest are places that will need the temporarily-writable infrastructure (PaX/Grsecurity uses pax_open_kernel()/pax_close_kernel() for these).

That’s it for v4.6, next up will be v4.7!

© 2016 – 2018, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

Previously: v4.4. Some things I found interesting in the Linux kernel v4.5:

CONFIG_IO_STRICT_DEVMEM

The CONFIG_STRICT_DEVMEM setting that has existed for a long time already protects system RAM from being accessible through the /dev/mem device node to root in user-space. Dan Williams added CONFIG_IO_STRICT_DEVMEM to extend this so that if a kernel driver has reserved a device memory region for use, it will become unavailable to /dev/mem also. The reservation in the kernel was to keep other kernel things from using the memory, so this is just common sense to make sure user-space can’t stomp on it either. Everyone should have this enabled. (And if you have a system where you discover you need IO memory access from userspace, you can boot with “iomem=relaxed” to disable this at runtime.)

If you’re looking to create a very bright line between user-space having access to device memory, it’s worth noting that if a device driver is a module, a malicious root user can just unload the module (freeing the kernel memory reservation), fiddle with the device memory, and then reload the driver module. So either just leave out /dev/mem entirely (not currently possible with upstream), build a monolithic kernel (no modules), or otherwise block (un)loading of modules (/proc/sys/kernel/modules_disabled).

ptrace fsuid checking

Jann Horn fixed some corner-cases in how ptrace access checks were handled on special files in /proc. For example, prior to this fix, if a setuid process temporarily dropped privileges to perform actions as a regular user, the ptrace checks would not notice the reduced privilege, possibly allowing a regular user to trick a privileged process into disclosing things out of /proc (ASLR offsets, restricted directories, etc) that they normally would be restricted from seeing.

ASLR entropy sysctl

Daniel Cashman standardized the way architectures declare their maximum user-space ASLR entropy (CONFIG_ARCH_MMAP_RND_BITS_MAX) and then created a sysctl (/proc/sys/vm/mmap_rnd_bits) so that system owners could crank up entropy. For example, the default entropy on 32-bit ARM was 8 bits, but the maximum could be as much as 16. If your 64-bit kernel is built with CONFIG_COMPAT, there’s a compat version of the sysctl as well, for controlling the ASLR entropy of 32-bit processes: /proc/sys/vm/mmap_rnd_compat_bits.

Here’s how to crank your entropy to the max, without regard to what architecture you’re on:

for i in "" "compat_"; do f=/proc/sys/vm/mmap_rnd_${i}bits; n=$(cat $f); while echo $n > $f ; do n=$(( n + 1 )); done; done

strict sysctl writes

Two years ago I added a sysctl for treating sysctl writes more like regular files (i.e. what’s written first is what appears at the start), rather than like a ring-buffer (what’s written last is what appears first). At the time it wasn’t clear what might break if this was enabled, so a WARN was added to the kernel. Since only one such string showed up in searches over the last two years, the strict writing mode was made the default. The setting remains available as /proc/sys/kernel/sysctl_writes_strict.

seccomp UM support

Mickaël Salaün added seccomp support (and selftests) for user-mode Linux. Moar architectures!

seccomp NNP vs TSYNC fix

Jann Horn noticed and fixed a problem where if a seccomp filter was already in place on a process (after being installed by a privileged process like systemd, a container launcher, etc) then the setting of the “no new privs” flag could be bypassed when adding filters with the SECCOMP_FILTER_FLAG_TSYNC flag set. Bypassing NNP meant it might be possible to trick a buggy setuid program into doing things as root after a seccomp filter forced a privilege drop to fail (generally referred to as the “sendmail setuid flaw”). With NNP set, a setuid program can’t be run in the first place.

That’s it! Next I’ll cover v4.6…

Edit: Added notes about “iomem=…”

© 2016, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

Previously: v4.3. Continuing with interesting security things in the Linux kernel, here’s v4.4. As before, if you think there’s stuff I missed that should get some attention, please let me know.

seccomp Checkpoint/Restore-In-Userspace

Tycho Andersen added a way to extract and restore seccomp filters from running processes via PTRACE_SECCOMP_GET_FILTER under CONFIG_CHECKPOINT_RESTORE. This is a continuation of his work (that I failed to mention in my prior post) from v4.3, which introduced a way to suspend and resume seccomp filters. As I mentioned at the time (and for which he continues to quote me) “this feature gives me the creeps.” :)

x86 W^X detection

Stephen Smalley noticed that there was still a range of kernel memory (just past the end of the kernel code itself) that was incorrectly marked writable and executable, defeating the point of CONFIG_DEBUG_RODATA which seeks to eliminate these kinds of memory ranges. He corrected this in v4.3 and added CONFIG_DEBUG_WX in v4.4 which performs a scan of memory at boot time and yells loudly if unexpected memory protection are found. To nobody’s delight, it was shortly discovered the UEFI leaves chunks of memory in this state too, which posed an ugly-to-solve problem (which Matt Fleming addressed in v4.6).

x86_64 vsyscall CONFIG

I introduced a way to control the mode of the x86_64 vsyscall with a build-time CONFIG selection, though the choice I really care about is CONFIG_LEGACY_VSYSCALL_NONE, to force the vsyscall memory region off by default. The vsyscall memory region was always mapped into process memory at a fixed location, and it originally posed a security risk as a ROP gadget execution target. The vsyscall emulation mode was added to mitigate the problem, but it still left fixed-position static memory content in all processes, which could still pose a security risk. The good news is that glibc since version 2.15 doesn’t need vsyscall at all, so it can just be removed entirely. Any kernel built this way that discovered they needed to support a pre-2.15 glibc could still re-enable it at the kernel command line with “vsyscall=emulate”.

That’s it for v4.4. Tune in tomorrow for v4.5!

© 2016, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

When I gave my State of the Kernel Self-Protection Project presentation at the 2016 Linux Security Summit, I included some slides covering some quick bullet points on things I found of interest in recent Linux kernel releases. Since there wasn’t a lot of time to talk about them all, I figured I’d make some short blog posts here about the stuff I was paying attention to, along with links to more information. This certainly isn’t everything security-related or generally of interest, but they’re the things I thought needed to be pointed out. If there’s something security-related you think I should cover from v4.3, please mention it in the comments. I’m sure I haven’t caught everything. :)

A note on timing and context: the momentum for starting the Kernel Self Protection Project got rolling well before it was officially announced on November 5th last year. To that end, I included stuff from v4.3 (which was developed in the months leading up to November) under the umbrella of the project, since the goals of KSPP aren’t unique to the project nor must the goals be met by people that are explicitly participating in it. Additionally, not everything I think worth mentioning here technically falls under the “kernel self-protection” ideal anyway — some things are just really interesting userspace-facing features.

So, to that end, here are things I found interesting in v4.3:

CONFIG_CPU_SW_DOMAIN_PAN

Russell King implemented this feature for ARM which provides emulated segregation of user-space memory when running in kernel mode, by using the ARM Domain access control feature. This is similar to a combination of Privileged eXecute Never (PXN, in later ARMv7 CPUs) and Privileged Access Never (PAN, coming in future ARMv8.1 CPUs): the kernel cannot execute user-space memory, and cannot read/write user-space memory unless it was explicitly prepared to do so. This stops a huge set of common kernel exploitation methods, where either a malicious executable payload has been built in user-space memory and the kernel was redirected to run it, or where malicious data structures have been built in user-space memory and the kernel was tricked into dereferencing the memory, ultimately leading to a redirection of execution flow.

This raises the bar for attackers since they can no longer trivially build code or structures in user-space where they control the memory layout, locations, etc. Instead, an attacker must find areas in kernel memory that are writable (and in the case of code, executable), where they can discover the location as well. For an attacker, there are vastly fewer places where this is possible in kernel memory as opposed to user-space memory. And as we continue to reduce the attack surface of the kernel, these opportunities will continue to shrink.

While hardware support for this kind of segregation exists in s390 (natively separate memory spaces), ARM (PXN and PAN as mentioned above), and very recent x86 (SMEP since Ivy-Bridge, SMAP since Skylake), ARM is the first upstream architecture to provide this emulation for existing hardware. Everyone running ARMv7 CPUs with this kernel feature enabled suddenly gains the protection. Similar emulation protections (PAX_MEMORY_UDEREF) have been available in PaX/Grsecurity for a while, and I’m delighted to see a form of this land in upstream finally.

To test this kernel protection, the ACCESS_USERSPACE and EXEC_USERSPACE triggers for lkdtm have existed since Linux v3.13, when they were introduced in anticipation of the x86 SMEP and SMAP features.

Ambient Capabilities

Andy Lutomirski (with Christoph Lameter and Serge Hallyn) implemented a way for processes to pass capabilities across exec() in a sensible manner. Until Ambient Capabilities, any capabilities available to a process would only be passed to a child process if the new executable was correctly marked with filesystem capability bits. This turns out to be a real headache for anyone trying to build an even marginally complex “least privilege” execution environment. The case that Chrome OS ran into was having a network service daemon responsible for calling out to helper tools that would perform various networking operations. Keeping the daemon not running as root and retaining the needed capabilities in children required conflicting or crazy filesystem capabilities organized across all the binaries in the expected tree of privileged processes. (For example you may need to set filesystem capabilities on bash!) By being able to explicitly pass capabilities at runtime (instead of based on filesystem markings), this becomes much easier.

For more details, the commit message is well-written, almost twice as long as than the code changes, and contains a test case. If that isn’t enough, there is a self-test available in tools/testing/selftests/capabilities/ too.

PowerPC and Tile support for seccomp filter

Michael Ellerman added support for seccomp to PowerPC, and Chris Metcalf added support to Tile. As the seccomp maintainer, I get excited when an architecture adds support, so here we are with two. Also included were updates to the seccomp self-tests (in tools/testing/selftests/seccomp), to help make sure everything continues working correctly.

That’s it for v4.3. If I missed stuff you found interesting, please let me know! I’m going to try to get more per-version posts out in time to catch up to v4.8, which appears to be tentatively scheduled for release this coming weekend. Next: v4.4.

© 2016, Kees Cook. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.