Re: [PATCH] x86/doc: add PTI description

[Randy Dunlap]

On 12/18/2017 02:04 PM, Dave Hansen wrote:
> This got kicked out of the PTI set as the implementation diverged
> from its contents.  I've updated it so it can hopefully rejoin the
> set.
> 
> ---
> 
> From: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> 
> Add some details about how PTI works, what some of the downsides
> are, and how to debug it when things go wrong.
> 
> Also document the kernel parameter: 'nopti'.
> 
> Signed-off-by: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> Cc: Moritz Lipp <moritz.lipp@xxxxxxxxxxxxxx>
> Cc: Daniel Gruss <daniel.gruss@xxxxxxxxxxxxxx>
> Cc: Michael Schwarz <michael.schwarz@xxxxxxxxxxxxxx>
> Cc: Richard Fellner <richard.fellner@xxxxxxxxxxxxxxxxx>
> Cc: Andy Lutomirski <luto@xxxxxxxxxx>
> Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> Cc: Kees Cook <keescook@xxxxxxxxxx>
> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: x86@xxxxxxxxxx
> ---
> 
>  b/Documentation/admin-guide/kernel-parameters.txt |    4 
>  b/Documentation/x86/pti.txt                       |  182 ++++++++++++++++++++++
>  2 files changed, 186 insertions(+)
> 
> diff -puN Documentation/admin-guide/kernel-parameters.txt~kpti-doc Documentation/admin-guide/kernel-parameters.txt
> --- a/Documentation/admin-guide/kernel-parameters.txt~kpti-doc	2017-12-18 13:55:59.635504663 -0800
> +++ b/Documentation/admin-guide/kernel-parameters.txt	2017-12-18 13:55:59.640504663 -0800
> @@ -904,6 +904,10 @@
>  	nopku		[X86] Disable Memory Protection Keys CPU feature found
>  			in some Intel CPUs.
>  
> +	nopti		[X86] Disable Page Table Isolation.  Disabling this
> +			feature removes hardening, but improves performance
> +			of system calls and interrupts.
> +
>  	module.async_probe [KNL]
>  			Enable asynchronous probe on this module.
>  
> diff -puN /dev/null Documentation/x86/pti.txt
> --- /dev/null	2017-12-15 13:48:30.454245127 -0800
> +++ b/Documentation/x86/pti.txt	2017-12-18 13:57:29.433504439 -0800
> @@ -0,0 +1,182 @@
> +Overview
> +========
> +
> +Page Table Isolation (pti, previously known as KAISER[1]) is a
> +countermeasure against attacks on kernel address information.  There
> +are at least three existing, published, approaches using the shared
> +user/kernel mapping and hardware features to defeat KASLR.  One
> +approach referenced in a paper[1] locates the kernel by observing
> +differences in page fault timing between present-but-inaccessable

                                                        inaccessible

> +kernel pages and non-present pages.
> +
> +To avoid leaking address information, we create an new, independent

                                                   a new,

> +copy of the page tables which are used only when running userspace
> +applications.  When the kernel is entered via syscalls, interrupts or
> +exceptions, page tables are switched to the full "kernel" copy.  When
> +the system switches back to user mode, the user copy is used again.

Uh, can userspace on one CPU attempt to observe kernelspace on the processor
that is running kernel code?  I haven't read any of the published papers,
so maybe I'm out in the woods here...

> +The userspace page tables contain only a minimal amount of kernel
> +data: only what is needed to enter/exit the kernel such as the
> +entry/exit functions themselves and the interrupt descriptor table
> +(IDT).  There are a few unnecessary things that get mapped such as the
> +first C function when entering an interrupt (see comments in pti.c).
> +
> +This approach helps to ensure that side-channel attacks that leverage
> +the paging structures do not function when PTI is enabled.  It can be
> +enabled by setting CONFIG_PAGE_TABLE_ISOLATION=y at compile time.
> +Once enabled at compile-time, it can be disabled at boot with the
> +'nopti' kernel parameter.
> +
> +Page Table Management
> +=====================
> +
> +When PTI is enabled, the kernel manages two sets of page
> +tables.  The first copy is very similar to what would be present
> +for a kernel without PTI.  This includes a complete mapping of
> +userspace that the kernel can use for things like copy_to_user().
> +
> +The userspace copy is used when running userspace and mirrors the
> +mapping of userspace present in the kernel copy.  It maps a only
> +the kernel data needed to enter and exit the kernel.  This data
> +is entirely contained in the 'struct cpu_entry_area' structure
> +which is placed in the fixmap and thus each CPU's copy of the
> +area has a compile-time-fixed virtual address.

Is that last sentence supposed to be a good thing?  Doesn't sound it
to me...

> +For new userspace mappings, the kernel makes the entries in its
> +page tables like normal.  The only difference is when the kernel
> +makes entries in the top (PGD) level.  In addition to setting the
> +entry in the main kernel PGD, a copy of the entry is made in the
> +userspace page tables' PGD.
> +
> +This sharing at the PGD level also inherently shares all the lower
> +layers of the page tables.  This leaves a single, shared set of
> +userspace page tables to manage.  One PTE to lock, one set set of
> +accessed bits, dirty bits, etc...
> +
> +Overhead
> +========
> +
> +Protection against side-channel attacks is important.  But,
> +this protection comes at a cost:
> +
> +1. Increased Memory Use
> +  a. Each process now needs an order-1 PGD instead of order-0.
> +     (Consumes 4k per process).

is that:          4KB more per process ?

> +  b. The pre-allocated second-level (p4d or pud) kernel page
> +     table pages cost ~1MB of additional memory at boot.  This
> +     is not totally wasted because some of these pages would
> +     have been needed eventually for normal kernel page tables
> +     and things in the vmalloc() area like vmemmap[].
> +  c. The 'cpu_entry_area' structure must be 2MB in size and 2MB
> +     aligned so that it can be mapped by setting a single PMD
> +     entry.  This consumes nearly 2MB of RAM once the kernel
> +     is decompressed, but no space in the kernel image itself.
> +
> +2. Runtime Cost
> +  a. CR3 manipulation to switch between the page table copies
> +     must be done at interrupt, syscall, and exception entry
> +     and exit (it can be skipped when the kernel is interrupted,
> +     though.)  Moves to CR3 are on the order of a hundred
> +     cycles, and are required every at entry and every at exit.

                    are required at each entry and at each exit.
(or every, if you prefer)

> +  b. A "trampoline" must be used for SYSCALL entry.  This
> +     trampoline depends on a smaller set of resources than the
> +     non-PTI SYSCALL entry code, so requires mapping fewer
> +     things into the userspace page tables.  The downside is
> +     that stacks must be switched at entry time.
> +  d. Global pages are disabled for all kernel structures not
> +     mapped in both to kernel and userspace page tables.  This

     drop             "to"?  Doesn't make sense to me.

> +     feature of the MMU allows different processes to share TLB
> +     entries mapping the kernel.  Losing the feature means more
> +     TLB misses after a context switch.  The actual loss of
> +     performance is very small, however, never exceeding 1%.
> +  d. Process Context IDentifiers (PCID) is a CPU feature that
> +     allows us to skip flushing the entire TLB when switching page
> +     tables.  This makes switching the page tables (at context
> +     switch, or kernel entry/exit) cheaper.  But, on systems with
> +     PCID support, the context switch code must flush both the user
> +     and kernel entries out of the TLB.  The user PCID TLB flush is
> +     deferred until the exit to userspace, minimizing the cost.
> +  e. The userspace page tables must be populated for each new
> +     process.  Even without PTI, the shared kernel mappings
> +     are created by copying top-level (PGD) entries into each
> +     new process.  But, with PTI, there are now *two* kernel
> +     mappings: one in the kernel page tables that maps everything
> +     and one for the entry/exit structures.  At fork(), we need to
> +     copy both.
> +  f. In addition to the fork()-time copying, there must also
> +     be an update to the userspace PGD any time a set_pgd() is done
> +     on a PGD used to map userspace.  This ensures that the kernel
> +     and userspace copies always map the same userspace
> +     memory.
> +  g. On systems without PCID support, each CR3 write flushes
> +     the entire TLB.  That means that each syscall, interrupt
> +     or exception flushes the TLB.

Does this imply that PTI is best suited to CPUs that have PCID?

> +Possible Future Work
> +====================
> +1. We can be more careful about not actually writing to CR3
> +   unless its value is actually changed.
> +2. Continue to compress the userspace-mapped data to be mapped
> +   together.  Currently, it is using four entries, but could
> +   be further minimized.
> +3. Allow PTI to enabled/disabled at runtime in addition to the

                to be

> +   boot-time switching.
> +
> +Testing
> +========
> +
> +To test stability of PTI, the following test procedure is recommended,
> +ideally doing all of these in parallel:
> +
> +1. Set CONFIG_DEBUG_ENTRY=y
> +2. Run several copies of all of the tools/testing/selftests/x86/ tests
> +   (excluding MPX and protection_keys) in a loop on multiple CPUs for
> +   several minutes.  These tests frequently uncover corner cases in the
> +   kernel entry code.  In general, old kernels might cause these tests
> +   themselves to crash, but they should never crash the kernel.
> +3. Run the 'perf' tool in a mode (top or record) that generates many
> +   frequent performance monitoring non-maskable interrupts (see "NMI"
> +   in /proc/interrupts).  This exercises the NMI entry/exit code which
> +   is known to trigger bugs in code paths that did not expect to be
> +   interrupted, including nested NMIs.  Using "-c" boosts the rate of
> +   NMIs, and using two -c with separate counters encourages nested NMIs
> +   and less deterministic behavior.
> +
> +	while true; do perf record -c 10000 -e instructions,cycles -a sleep 10; done
> +
> +4. Launch a KVM virtual machine.
> +
> +Debugging
> +=========
> +
> +Bugs in PTI cause a few different signatures of crashes
> +that are worth noting here.
> +
> + * Failures of the selftests/x86 code.  Usually a bug in one of the
> +   more obscure corners of entry_64.S
> + * Crashes in early boot, especially around CPU bringup.  Bugs
> +   in the trampoline code or mappings cause these.
> + * Crashes at the first interrupt.  Caused by bugs in entry_64.S,
> +   like screwing up a page table switch.  Also caused by
> +   incorrectly mapping the IRQ handler entry code.
> + * Crashes at the first NMI.  The NMI code is separate from main
> +   interrupt handlers and can have bugs that do not affect
> +   normal interrupts.  Also caused by incorrectly mapping NMI
> +   code.  NMIs that interrupt the entry code must be very
> +   careful and can be the cause of crashes that show up when
> +   running perf.
> + * Kernel crashes at the first exit to userspace.  entry_64.S
> +   bugs, or failing to map some of the exit code.
> + * Crashes at first interrupt that interrupts userspace. The paths
> +   in entry_64.S that return to userspace are sometimes separate
> +   from the ones that return to the kernel.
> + * Double faults: overflowing the kernel stack because of page
> +   faults upon page faults.  Caused by touching non-pti-mapped
> +   data in the entry code, or forgetting to switch to kernel
> +   CR3 before calling into C functions which are not pti-mapped.
> + * Userspace segfaults early in boot, sometimes manifesting
> +   as mount(8) failing to mount the rootfs.  These have
> +   tended to be TLB invalidation issues.  Usually invalidating
> +   the wrong PCID, or otherwise missing an invalidation.
> +
> +1. https://gruss.cc/files/kaiser.pdf
> _
> 


Good stuff.  Thanks.
-- 
~Randy