Re: [PATCH 00/31 v2] PTI support for x86_32

[Joerg Roedel]

Hi Ingo,

On Sun, Feb 11, 2018 at 08:13:12PM +0100, Ingo Molnar wrote:
> Could you please measure the PTI kernel vs. vanilla kernel?

Okay, did that, here is the data. The test machine is a Xeon E5-1620v2,
which is Ivy Bridge based (no PCIE) and has 4C/8T.

I ran the 2 tests you suggested:

	* Test-1: perf stat --null --sync --repeat 10 perf bench sched messaging -g 20

	* Test-2: perf stat --null --sync --repeat 10 perf bench sched messaging -g 20 -t

The tests ran on these kernels:

	* tip-32-pae: current top of tip/x86-tip-for-linus branch,
	              compiled as a 32 bit kernel with PAE
	              (commit b2ac58f90540e39324e7a29a7ad471407ae0bf48)

	* pti-32-pae: Same as above with my patches on-top, as on

		      git://git.kernel.org/pub/scm/linux/kernel/git/joro/linux.git pti-x32-v2

	              compiled as a 32 bit kernel with PAE
		      (commit dbb0074f778b396a11e0c897fef9d0c4583e7ccb)

	* pti-off-64: current top of tip/x86-tip-for-linus branch,
		      compiled as a 64 bit kernel, booted with pti=off
	              (commit b2ac58f90540e39324e7a29a7ad471407ae0bf48)

	* pti-on-64: current top of tip/x86-tip-for-linus branch,
		     compiled as a 64 bit kernel, booted with pti=on
	             (commit b2ac58f90540e39324e7a29a7ad471407ae0bf48)

Results are:
	            | Test-1             | Test-2          
	------------+--------------------+-----------------
	tip-32-pae  | 0.28s (+-0.44%)    | 0.27s (+-2.15%) 
	------------+--------------------+-----------------
	pti-32-pae  | 0.44s (+-0.40%)    | 0.42s (+-0.48%) 
	------------+--------------------+-----------------
	pti-off-64  | 0.24s (+-0.40%)    | 0.25s (+-1.31%) 
	------------+--------------------+-----------------
	pti-on-64   | 0.30s (+-0.47%)    | 0.31s (+-0.95%)

On 32 bit with PTI enabled the test needs 157% (non-threaded) and
156% (threaded) of time compared to the non-PTI baseline.

On 64 bit these numbers are 125% (non-threaded) and 124% (threaded).

The pti-32-pae kernel still used 'rep movsb' in the entry code. I
replaced that with 'rep movsl' and measured again, but overhead is still
around 152%.

I also measured cycles with 'perf record' to see where the additional
time is spent. The report showed around 25% in entry_SYSENTER_32 for
the pti-32-pae kernel. The same report on the tip-32-pae kernel shows
around 2.5% for the same symbol.

The entry_SYSENTER_32 path does no stack-copy on entry (it only
push/pops 8 bytes for the cr3 switch), but one full pt_regs copy on
exit. The exit-path was easy to optimize, I got it to the point where it
only copied 8 bytes to the entry stack (flags and eax).  This way I got
the 'perf report' numbers for entry_SYSENTER_32 down to around 20%, but
the overall numbers for Test-1 and Test-2 are still at around 150% of
the baseline.

So it seems that most of the additional time is actually spent switching
the cr3s.

Regards,

	Joerg