Hi!
At long last a new version of the call depth tracking patches.
v1: https://lkml.kernel.org/r/[email protected]
This version is significantly different from the last in that it no longer
makes use of external call thunks allocated from the module space. Instead
every function gets aligned to 16 bytes and gets 16 bytes of (pre-symbol)
padding. (This padding will also come in handy for other things, like the
kCFI/FineIBT work.)
Prior to these patches function alignment is basically non-existent, as such
any instruction fetch for the first instructions of a function will have (on
average) half the fetch window filled with whatever comes before. By pushing
the alignment up to 16 bytes this improves matters for chips that happen to
have a 16 byte i-fetch window size (Intel) while not making matters worse for
chips that have a larger 32 byte i-fetch window (AMD Zen). In fact, it improves
the worst case for Zen from 31 bytes of garbage to 16 bytes of garbage.
As such the first many patches of the series fix up lots of alignment quirks.
The second big difference is the introduction of struct pcpu_hot. Because the
compiler managed to place two adjacent (in code) DEFINE_PER_CPU() variables in
random cachelines (it is absolutely free to do so) the introduction of the
per-cpu x86_call_depth variable sometimes introduced significant additional
cache pressure, while other times it would sit nicely in the same line with
preempt_count and not show up at all.
In order to alleviate this problem; introduce struct pcpu_hot and collect a
number of hot per-cpu variables in a way the compiler can't mess up.
Since these changes are 'unconditional', Mel was gracious enough to help test
this on his test setup across all the relevant uarchs (very much including both
Intel and AMD machines) and found that while these changes cause some very
small wins and losses across the board it is mostly noise.
Aside from these changes; the core of the depth tracking is still the same.
- objtool creates a list of (function) call sites.
- for every call; overwrite the padding of the target function with the
accounting thunk (if not already done) and adjust the call site to
target this thunk.
- the retbleed return thunk mechanism is used for a custom return thunk
that includes return accounting and does RSB stuffing when required.
This ensures no new compiler is required and avoids almost all overhead for
non affected machines. This new option can still be selected using:
"retbleed=stuff"
on the kernel command line.
As a refresher; the theory behind call depth tracking is:
The Return-Stack-Buffer (RSB) is a 16 deep stack that is filled on every call.
On the return path speculation will "pop" an entry and takes that as the return
target. Once the RSB is empty, the CPU falls back to other predictors, e.g. the
Branch History Buffer, which can be mistrained by user space and misguides the
(return) speculation path to a disclosure gadget of your choice -- as described
in the retbleed paper.
Call depth tracking is designed to break this speculation path by stuffing
speculation trap calls into the RSB whenver the RSB is running low. This way
the speculation stalls and never falls back to other predictors.
The assumption is that stuffing at the 12th return is sufficient to break the
speculation before it hits the underflow and the fallback to the other
predictors. Testing confirms that it works. Johannes, one of the retbleed
researchers, tried to attack this approach and confirmed that it brings the
signal to noise ratio down to the crystal ball level.
Excerpts from IBRS vs stuff from Mel's testing:
perfsyscall
6.0.0-rc1 6.0.0-rc1
tglx-mit-spectre-ibrs tglx-mit-spectre-retpoline-retstuff
Duration User 136.16 69.10
Duration System 100.50 33.04
Duration Elapsed 237.20 102.65
That's a massive improvement with a major reduction in system CPU usage.
Kernel compilation is variable. Skylake-X was modest with 2-18% gain depending
on degree of parallelisation.
Git checkouts are roughly 14% faster on Skylake-X
Network test were localhost only so are limited but even so, the gain is
large. Skylake-X again;
Netperf-TCP
6.0.0-rc1 6.0.0-rc1
tglx-mit-spectre-ibrs tglx-mit-spectre-retpoline-retstuff
Hmean send-64 241.39 ( 0.00%) 298.00 * 23.45%*
Hmean send-128 489.55 ( 0.00%) 610.46 * 24.70%*
Hmean send-256 990.85 ( 0.00%) 1201.73 * 21.28%*
Hmean send-1024 4051.84 ( 0.00%) 5006.19 * 23.55%*
Hmean send-2048 7924.75 ( 0.00%) 9777.14 * 23.37%*
Hmean send-3312 12319.98 ( 0.00%) 15210.07 * 23.46%*
Hmean send-4096 14770.62 ( 0.00%) 17941.32 * 21.47%*
Hmean send-8192 26302.00 ( 0.00%) 30170.04 * 14.71%*
Hmean send-16384 42449.51 ( 0.00%) 48036.45 * 13.16%*
While this is UDP_STREAM, TCP_STREAM is similarly impressive.
FIO measurements done by Tim Chen:
read (kIOPs) Mean stdev mitigations=off retbleed=off CPU util
================================================================================
mitigations=off 357.33 3.79 0.00% 6.14% 98.93%
retbleed=off 336.67 6.43 -5.78% 0.00% 99.01%
retbleed=ibrs 242.00 0.00 -32.28% -28.12% 99.41%
retbleed=stuff (pad) 314.33 1.53 -12.03% -6.63% 99.31%
read/write Baseline Baseline
70/30 (kIOPs) Mean stdev mitigations=off retbleed=off CPU util
================================================================================
mitigations=off 349.00 5.29 0.00% 9.06% 96.66%
retbleed=off 320.00 5.05 -8.31% 0.00% 95.54%
retbleed=ibrs 238.60 0.17 -31.63% -25.44% 98.18%
retbleed=stuff (pad) 293.37 0.81 -15.94% -8.32% 97.71%
Baseline Baseline
write (kIOPs) Mean stdev mitigations=off retbleed=off CPU util
================================================================================
mitigations=off 296.33 8.08 0.00% 6.21% 93.96%
retbleed=off 279.00 2.65 -5.85% 0.00% 93.63%
retbleed=ibrs 230.33 0.58 -22.27% -17.44% 95.92%
retbleed=stuff (pad) 266.67 1.53 -10.01% -4.42% 94.75%
The patches can also be found in git here:
git://git.kernel.org/pub/scm/linux/kernel/git/tglx/devel.git depthtracking
On Fri, Sep 02, 2022 at 03:06:25PM +0200, Peter Zijlstra wrote:
> Excerpts from IBRS vs stuff from Mel's testing:
>
FWIW, I retested this version as there were slight changes and retbleed=stuff
on Skylake is still far faster than the default. The default performance
differences are mostly within the noise versus a vanilla 6.0-rc3 kernel.
Other generations of Intel machine showed mostly noise. Zen[1-3] showed
no significant difference (immune to retbleed but functional alignment or
percpu structure changes might have mattered). For the series;
Tested-by: Mel Gorman <[email protected]>
--
Mel Gorman
SUSE Labs