| 1 |
On Wed, Jan 3, 2018 at 3:35 PM, Wols Lists <antlists@××××××××××××.uk> wrote: |
| 2 |
> |
| 3 |
> And as I understand it the code can be disabled with either a compile |
| 4 |
> time option or command line switch to the kernel. |
| 5 |
|
| 6 |
I suspect the compile-time option is PAGE_TABLE_ISOLATION (which was |
| 7 |
newly added in 4.14.11). The command line option nopti will disable |
| 8 |
it at runtime. |
| 9 |
|
| 10 |
Rumor has it that it will be disabled on AMD CPUs in 4.14.12, but I |
| 11 |
can't point to anywhere authoritative for that news so I'd consider it |
| 12 |
a rumor. I've also heard that Arch has deployed it early to 4.14.11, |
| 13 |
and I wouldn't be surprised if many distros do this if it is intended |
| 14 |
to go into the next stable, as there would be no point subjecting AMD |
| 15 |
users to performance issues. I haven't spoken to the Gentoo kernel |
| 16 |
team about what their plans are for it. In any case, nopti on the |
| 17 |
command line is probably the cleanest solution. I personally avoided |
| 18 |
disabling the feature in the compiled kernel because I don't want to |
| 19 |
be using the same config file on an Intel CPU in a year or two and |
| 20 |
forget I have it forced off. |
| 21 |
|
| 22 |
> The relevant code is |
| 23 |
> called KAISER, which forces kernel and user address space into different |
| 24 |
> contexts, and causes a nasty context-switching overhead on both Intel |
| 25 |
> and AMD cpus. |
| 26 |
> |
| 27 |
|
| 28 |
I believe the kernel went with "Page Table Isolation (PTI)" rather |
| 29 |
that KAISER, probably to avoid ethnic issues. Apparently this was |
| 30 |
deemed to have a more acceptable acronym than Forcefully Unmap |
| 31 |
Complete Kernel With Interrupt Trampolines. |
| 32 |
|
| 33 |
-- |
| 34 |
Rich |
Archives