Hi,
here are 3 patches which update the PTI-x86-32 patches recently merged
into the tip-tree. The patches are ordered by importance:
Patch 1: Very important, it fixes a vmalloc-fault in NMI context
when PTI is enabled. This is pretty unlikely to hit
when starting perf on an idle machine, which is why I
didn't find it earlier in my testing. I always started
'perf top' first :/ But when I start 'perf top' last
when the kernel-compile already runs, it hits almost
immediatly.
Patch 2: Fix the 'from-kernel-check' in SWITCH_TO_KERNEL_STACK
to also take VM86 into account. This is not strictly
necessary because the slow-path also works for VM86
mode but it is not how the code was intended to work.
And it breaks when Patch 3 is applied on-top.
Patch 3: Implement the reduced copying in the paranoid
entry/exit path as suggested by Andy Lutomirski while
reviewing version 7 of the original patches.
I have the x86/tip branch with these patches on-top running my test for
6h now, with no issues so far. So for now it looks like there are no
scheduling points or irq-enabled sections reached from the paranoid
entry/exit paths and we always return to the entry-stack we came from.
I keep the test running over the weekend at least.
Please review.
[ If Patch 1 looks good to the maintainers I suggest applying it soon,
before too many linux-next testers run into this issue. It is actually
the reason why I send out the patches _now_ and didn't wait until next
week when the other two patches got more testing from my side. ]
Thanks,
Joerg
Joerg Roedel (3):
perf/core: Make sure the ring-buffer is mapped in all page-tables
x86/entry/32: Check for VM86 mode in slow-path check
x86/entry/32: Copy only ptregs on paranoid entry/exit path
arch/x86/entry/entry_32.S | 82 ++++++++++++++++++++++++++-------------------
kernel/events/ring_buffer.c | 10 ++++++
2 files changed, 58 insertions(+), 34 deletions(-)
--
2.7.4
