[PATCH] arm64: add prctl(PR_PAC_SET_ENABLED_KEYS)
Peter Collingbourne
pcc at google.com
Tue Oct 13 00:03:49 EDT 2020
On Mon, Sep 7, 2020 at 4:03 AM Dave Martin <Dave.Martin at arm.com> wrote:
>
> On Wed, Sep 02, 2020 at 09:31:55PM -0700, Peter Collingbourne wrote:
> > On Tue, Sep 1, 2020 at 9:02 AM Dave Martin <Dave.Martin at arm.com> wrote:
> > >
> > > On Wed, Aug 26, 2020 at 07:55:51PM -0700, Peter Collingbourne wrote:
> > > > On Wed, Aug 26, 2020 at 9:37 AM Dave Martin <Dave.Martin at arm.com> wrote:
> > > > >
> > > > > On Mon, Aug 24, 2020 at 05:23:27PM -0700, Peter Collingbourne wrote:
> > > > > > On Mon, Aug 24, 2020 at 7:49 AM Dave Martin <Dave.Martin at arm.com> wrote:
> > > > > > >
> > > > > > > On Wed, Aug 19, 2020 at 02:25:45PM -0700, Peter Collingbourne wrote:
> > > > > > > > On Wed, Aug 19, 2020 at 3:18 AM Dave Martin <Dave.Martin at arm.com> wrote:
> > > > > > > > >
> > > > > > > > > On Fri, Jul 31, 2020 at 06:11:52PM -0700, Peter Collingbourne wrote:
> > >
> > > [...]
> > >
> > > > > > > > > > diff --git a/arch/arm64/include/asm/asm_pointer_auth.h b/arch/arm64/include/asm/asm_pointer_auth.h
> > > > > > > > > > index 52dead2a8640..d121fa5fed5f 100644
> > > > > > > > > > --- a/arch/arm64/include/asm/asm_pointer_auth.h
> > > > > > > > > > +++ b/arch/arm64/include/asm/asm_pointer_auth.h
> > > > > > > > > > @@ -31,6 +31,14 @@ alternative_else_nop_endif
> > > > > > > > > > ldp \tmp2, \tmp3, [\tmp1, #PTRAUTH_USER_KEY_APDB]
> > > > > > > > > > msr_s SYS_APDBKEYLO_EL1, \tmp2
> > > > > > > > > > msr_s SYS_APDBKEYHI_EL1, \tmp3
> > > > > > > > > > +
> > > > > > > > > > + ldr \tmp2, [\tsk, #THREAD_SCTLR_ENXX_MASK]
> > > > > > > > > > + cbz \tmp2, .Laddr_auth_skip_\@
> > > > > > > > >
> > > > > > > > > I wonder whether it would make sense to simple store the thread's base
> > > > > > > > > SCTLR value (containing the ENxx bits), rather than storing the ENxx
> > > > > > > > > bits separately. There may be reasons outside this snippet why this
> > > > > > > > > isn't such a good idea though -- I haven't gone digging so far.
> > > > > > > >
> > > > > > > > As far as I know (as I learned [1] from the MTE patch series), it can
> > > > > > > > be expensive to access SCTLR_EL1, therefore I arrange to only access
> > > > > > > > SCTLR_EL1 in the hopefully-uncommon case where a process has disabled
> > > > > > > > keys. Detecting the "process has disabled keys" case is quite simple
> > > > > > > > if we only store the disabled keys mask here, not so much if we store
> > > > > > > > the full value of SCTLR_EL1.
> > > > > > >
> > > > > > > My thought was that we would still only write SCTLR_EL1 if needed, but
> > > > > > > we would do the write-if-needed across the whole register in one go.
> > > > > > > This would be easier to extend if we have to twiddle additional
> > > > > > > SCTLR_EL1 bits in the future. If the key enable bits are the only
> > > > > > > affected bits for now then we could of course defer this factoring until
> > > > > > > later. (I vaguely remember a similar discussion in the past, but
> > > > > > > possibly it was about the pauth keys anyway, rather than something
> > > > > > > else.)
> > > > > >
> > > > > > We will also need a per-task SCTLR_EL1.TCF0 setting for MTE. We could
> > > > > > potentially combine the two, so that both
> > > > > > prctl(PR_PAC_SET_ENABLED_KEYS) and prctl(PR_SET_TAGGED_ADDR_CTRL) end
> > > > > > up affecting a common SCTLR_EL1 field in the task_struct, and we do:
> > > > > >
> > > > > > On kernel entry:
> > > > > > - read SCTLR_EL1 from task_struct (or from the system register if we
> > > > > > decide that's cheap enough)
> > > > > > - if EnIA clear, set it, and write the value to the system register.
> > > > > >
> > > > > > On kernel exit:
> > > > > > - read system register SCTLR_EL1
> > > > > > - read SCTLR_EL1 from task_struct
> > > > > > - if they are different, write task's SCTLR_EL1 to the system register.
> > > > > >
> > > > > > But this would require at least an unconditional read of SCTLR_EL1 per
> > > > > > kernel exit. The comment that I referenced says that "accesses" to the
> > > > > > register are expensive. I was operating under the assumption that both
> > > > > > reads and writes are expensive, but if it's actually only writes that
> > > > > > are expensive, we may be able to get away with the above.
> > > > >
> > > > > The kernel is in full control over what's in SCTLR_EL1, so perhaps we
> > > > > could cache what we wrote percpu, if the overhead of reading it is
> > > > > problematic.
> > > >
> > > > Maybe, but that sounds like the kind of complexity that I'd like to avoid.
> > >
> > > Ack, this is more something to explore if the other approaches turn out
> > > to be more costly.
> > >
> > > > I went ahead and did some performance measurements using the following
> > > > program, which issues 10M invalid syscalls and exits. This should give
> > > > us something as close as possible to a measurement of a kernel entry
> > > > and exit on its own.
> > > >
> > > > .globl _start
> > > > _start:
> > > > movz x1, :abs_g1:10000000
> > > > movk x1, :abs_g0_nc:10000000
> > > >
> > > > .Lloop:
> > > > mov x8, #65536 // invalid
> > > > svc #0
> > > > sub x1, x1, #1
> > > > cbnz x1, .Lloop
> > > >
> > > > mov x0, #0
> > > > mov x8, #0x5d // exit
> > > > svc #0
> > > >
> > > > On a Pixel 4 (which according to Wikipedia has a CPU based on
> > > > Cortex-A76/A55) I measured the median of 1000 runs execution time at
> > > > 0.554511s, implying an entry/exit cost of 55.5ns. If I make this
> > > > modification to the kernel (Pixel 4 uses kernel version 4.14 so this
> > > > won't apply cleanly to modern kernels; this is in the same place as
> > > > the ptrauth_keys_install_user macro invocation in a modern kernel):
> > > >
> > > > diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S
> > > > index 303447c765f7..b7c72d767f3e 100644
> > > > --- a/arch/arm64/kernel/entry.S
> > > > +++ b/arch/arm64/kernel/entry.S
> > > > @@ -340,6 +340,7 @@ alternative_else_nop_endif
> > > > #endif
> > > > 3:
> > > > apply_ssbd 0, 5f, x0, x1
> > > > + mrs x0, sctlr_el1
> > > > 5:
> > > > .endif
> > > >
> > > > The median execution time goes to 0.565604s, implying a cost of 1.1ns
> > > > to read the system register. I also measured the cost of reading from
> > > > a percpu variable at kernel exit like so:
> > > >
> > > > diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S
> > > > index 303447c765f7..c5a89785f650 100644
> > > > --- a/arch/arm64/kernel/entry.S
> > > > +++ b/arch/arm64/kernel/entry.S
> > > > @@ -340,6 +340,7 @@ alternative_else_nop_endif
> > > > #endif
> > > > 3:
> > > > apply_ssbd 0, 5f, x0, x1
> > > > + ldr_this_cpu x0, sctlr_el1_cache, x1
> > > > 5:
> > > > .endif
> > > >
> > > > diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c
> > > > index 0b6ca5479f1f..ac9c9a6179d4 100644
> > > > --- a/arch/arm64/kernel/traps.c
> > > > +++ b/arch/arm64/kernel/traps.c
> > > > @@ -52,6 +52,8 @@
> > > > #include <asm/system_misc.h>
> > > > #include <asm/sysreg.h>
> > > >
> > > > +DEFINE_PER_CPU(unsigned long, sctlr_el1_cache);
> > > > +
> > > > static const char *handler[]= {
> > > > "Synchronous Abort",
> > > > "IRQ",
> > > >
> > > > With that the median execution time goes to 0.600186s, implying a cost
> > > > of 4.5ns. So at least on existing hardware, it looks like reading
> > > > SCTLR_EL1 directly is probably our cheapest option if we're going to
> > > > read *something*. Therefore I'll look at implementing this with
> > > > unconditional reads from SCTLR_EL1 in v2.
> > >
> > > Thanks for the investigation!
> > >
> > > I'm not too surprised by this outcome. There is a possibility that
> > > reading SCTLR_EL1 sucks badly on some implementations, but on others it
> > > may be little more than a register rename. We should probably wait for
> > > evidence before panicking about it.
> > >
> > > So long as this SCTLR_EL1 switching is completely disabled via
> > > alternatives on hardware to which it isn't applicable, I expect that
> > > should be a reasonable compromise.
> >
> > One disadvantage of disabling SCTLR_EL1 switching on inapplicable
> > hardware is that it complicates the mechanism for enabling and
> > disabling deprecated instruction support (see
> > arch/arm64/kernel/armv8_deprecated.c), which is currently implemented
> > by using on_each_cpu to set or clear certain bits in SCTLR_EL1 on each
> > CPU. This is incompatible with an entry/exit that assumes full control
> > over SCTLR_EL1.
> >
> > I would propose the following instruction sequences in entry.S, which
> > are justified by further experiments run on a Pixel 4 showing that:
> > 1) an unconditional SCTLR_EL1 write on kernel entry/exit costs 3.0ns,
> > whereas conditional writes cost 4.5ns
> > 2) an SCTLR_EL1 read on kernel entry is 0.3ns faster than a load from
> > task_struct
> > 3) an unconditional SCTLR_EL1 write on kernel exit costs 0.7ns.
> >
> > diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S
> > index 55af8b504b65..9e0acd5e7527 100644
> > --- a/arch/arm64/kernel/entry.S
> > +++ b/arch/arm64/kernel/entry.S
> > @@ -186,6 +186,12 @@ alternative_cb_end
> >
> > ptrauth_keys_install_kernel tsk, x20, x22, x23
> >
> > +alternative_if ARM64_HAS_ADDRESS_AUTH
> > + mrs x22, sctlr_el1
> > + orr x22, x22, SCTLR_ELx_ENIA
> > + msr sctlr_el1, x22
> > +alternative_else_nop_endif
> > +
> > scs_load tsk, x20
> > .else
> > add x21, sp, #S_FRAME_SIZE
> > @@ -297,6 +303,9 @@ alternative_else_nop_endif
> > /* No kernel C function calls after this as user keys are set. */
> > ptrauth_keys_install_user tsk, x0, x1, x2
> >
> > + ldr x0, [tsk, THREAD_SCTLR]
> > + msr sctlr_el1, x0
> > +
> > apply_ssbd 0, x0, x1
> > .endif
> >
> >
> > With these instruction sequences I would propose to reimplement
> > toggling deprecated instruction support as follows:
> > 1) Stop the machine using stop_machine.
> > 2) While the machine is stopped, adjust the deprecated instruction
> > support feature bits in each task_struct.
> >
> > This will make toggling deprecated instruction support more expensive,
> > but that seems like the right tradeoff since I imagine that these
> > features are rarely toggled, and making it more efficient to toggle
> > them would mean more instructions in the entry/exit hot path.
> >
> > This seems like it will work well if SCTLR_EL1 is set on kernel exit.
> > The deprecated feature bits are CP15BEN and SED, and each of them
> > affects features that are only implemented at EL0, so it seems like we
> > can wait until kernel exit to update SCTLR_EL1. But if we don't set
> > SCTLR_EL1 on exit on certain hardware, we will need to have both
> > mechanisms, use the existing mechanism on hardware that doesn't set
> > SCTLR_EL1 and the new mechanism on hardware that does set SCTLR_EL1.
> > Which I wouldn't mind too much, but I'm not sure that it's worth it to
> > save 0.7ns (I'm not claiming that my numbers are universal, just that
> > we should probably favor simpler approaches if we don't have evidence
> > that they are significantly more expensive than more complex ones).
> >
> > Peter
>
> This approach does look like close to the simplest option for the entry/
> exit path -- I'll let others comment on whether the actual overhead
> numbers sound acceptable.
>
> I think there are more places where the kernel modifies SCTLR than just
> for the deprecated instruction emulation though. Do you think your
> approach will work with all of them, and can they all tolerate a
> stop_machine()?
You're right, there are other places. Looking at the places written in
assembly, I think all of them do not expect the value to be retained.
But there are more places where we were setting this register in C
that I hadn't checked. For example, in cpufeature.c we have:
static void cpu_emulate_effective_ctr(const struct
arm64_cpu_capabilities *__unused)
{
/*
* If the CPU exposes raw CTR_EL0.IDC = 0, while effectively
* CTR_EL0.IDC = 1 (from CLIDR values), we need to trap accesses
* to the CTR_EL0 on this CPU and emulate it with the real/safe
* value.
*/
if (!(read_cpuid_cachetype() & BIT(CTR_IDC_SHIFT)))
sysreg_clear_set(sctlr_el1, SCTLR_EL1_UCT, 0);
}
So we may want heterogeneous values of SCTLR_EL1 on different cores,
unless for example for this routine we reimplement to trap accesses to
CTR_EL0 homogeneously.
Furthermore it is unclear that it would be valid to write to SCTLR_EL1
directly with the result of reading from a thread_struct field. My
belief that this behavior may be incorrect comes from the documented
semantics of reserved fields in SCTLR_EL1. These bits are defined as
RES0 which means that software must use an SBZP approach to write to
them, and an approach that reads a value from thread_struct and then
writes exactly that value to SCTLR_EL1 would not necessarily preserve
RES0 bits that are different between cores.
This led me to go back to considering approaches where we only update
EnIA on entry and exit and leave the updating of the other bits to
task switch at the same time as TCF0 in Catalin's MTE series.
Such approaches seem like they will perform better as well, if
extrapolating from existing hardware is anything to go by. It seems
that my previous measurements on Pixel 4 may have been inaccurate
because /sys/devices/system/cpu/*/cpufreq/scaling_governor was set to
its default value so scaling may have interfered with the results. I
ran a new set of measurements with scaling_governor set to powersave
which gave me much more stable results. With the newly stable results
I decided to take the median of 100 runs rather than 1000 to
compensate for the slower run time caused by the powersave governor.
These results showed that conditional MSR is faster than unconditional
MSR, and that LDR from thread_struct to obtain the state of the
incoming task is slightly faster than MRS. The full results with
references to patches are below:
Baseline: 316.7ns
Conditional MSR based on LDR [1]: 319.6ns
Unconditional MSR [2]: 353.4ns
Conditional MSR based on MRS [3]: 323.8ns
I also ran the same experiment on a DragonBoard 845c. Since this board
can run the mainline kernel it gives us results that we may expect to
be slightly more valid when making changes to the mainline kernel than
the 4.14 kernel used by Pixel 4. On this board I saw two "clusters" of
results so I took the median of the faster cluster which was more
stable (I don't think this was due to big.LITTLE because I saw very
different numbers if I pinned the process to a LITTLE core):
Baseline: 450.5ns
Conditional MSR based on LDR: 457.8ns
Unconditional MSR: 487.0ns
Conditional MSR based on MRS: 461.5ns
So we can see very similar perf impacts to Pixel 4. Although this
board is Cortex-A75/A55 which is the prior uarch generation to Pixel
4's Cortex-A76/A55, it also tells us that on big Cortex uarches
SCTLR_EL1 performance is somewhat stable over time, at least to the
extent that we can extrapolate that from two data points. Of course,
we may tune the instruction sequences again once hardware with PAC
support is publicly available.
Thus I think that to begin with we should use instruction sequences
similar to those in [1], protected via alternatives so we wouldn't be
impacting the performance on existing hardware. I will develop my
patch on top of the MTE series since they will interfere with each
other and it seems plausible that MTE will land first.
Peter
[1] https://android-review.googlesource.com/c/kernel/msm/+/1457681
[2] https://android-review.googlesource.com/c/kernel/msm/+/1456578
[3] https://android-review.googlesource.com/c/kernel/msm/+/1456579
Peter
More information about the linux-arm-kernel
mailing list