Re: [RFC PATCH 0/7] x86/entry: Atomic statck switching for IST

[Lai Jiangshan]

On Tue, Apr 4, 2023 at 12:53 AM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
>
> On 4/3/23 07:05, Lai Jiangshan wrote:
> >  Documentation/x86/kernel-stacks.rst   |   2 +
> >  arch/x86/entry/Makefile               |   3 +
> >  arch/x86/entry/entry_64.S             | 615 +++++++-------------------
> >  arch/x86/entry/ist_entry.c            | 299 +++++++++++++
> >  arch/x86/include/asm/cpu_entry_area.h |   8 +-
> >  arch/x86/include/asm/idtentry.h       |  15 +-
> >  arch/x86/include/asm/sev.h            |  14 -
> >  arch/x86/include/asm/traps.h          |   1 -
> >  arch/x86/kernel/asm-offsets_64.c      |   7 +
> >  arch/x86/kernel/callthunks.c          |   2 +
> >  arch/x86/kernel/dumpstack_64.c        |   6 +-
> >  arch/x86/kernel/nmi.c                 |   8 -
> >  arch/x86/kernel/sev.c                 | 108 -----
> >  arch/x86/kernel/traps.c               |  43 --
> >  arch/x86/kvm/vmx/vmx.c                | 441 +++++++++++++++++-
> >  arch/x86/kvm/x86.c                    |  10 +-
> >  arch/x86/mm/cpu_entry_area.c          |   2 +-
> >  tools/objtool/check.c                 |   7 +-
> >  18 files changed, 937 insertions(+), 654 deletions(-)
> >  create mode 100644 arch/x86/entry/ist_entry.c
>
> Some high-level comments...
>
> The diffstat looks a lot nastier because of the testing patch.  If you
> that patch and trim the diffstat a bit, it starts to look a _lot_ more
> appealing:
>
> >  arch/x86/entry/entry_64.S             |  615 ++++++++----------------------------
> >  arch/x86/entry/ist_entry.c            |  299 +++++++++++++++++
> >  arch/x86/kernel/sev.c                 |  108 ------
> >  arch/x86/kernel/traps.c               |   43 --
> ...
> >  16 files changed, 490 insertions(+), 650 deletions(-)
>
> It removes more code than it adds and also removes a bunch of assembly.
> If it were me posting this, I'd have shouted that from the rooftops
> instead of obscuring it with a testing patch and leaving it as an
> exercise to the reader to figure out.

The cover letter has 800+ lines of comments.  About 100-300 lines
of comments will be moved into the code which would make the diffstat
not so appealing.


P.S.

One of the reasons I didn't move them is that the comments are much more
complicated than the code which is a sign of improvement-required.

I'm going to change the narration from save-touch-replicate-copy-commit
to save-copy-build-commit and avoid using "replicate".

copy=copy_outmost(), build=build_interrupted(), the new narration
will change the comments mainly, and it will not change the actual
work. If the new narration is not simpler, I will not send it out to
add any confusion.

>  * Flesh out the testing story.  What kind of hardware was this tested
>    on?  How much was bare metal vs. VMs, etc...?

It is tested on bare metal and VM.  It is hard to stress the bare
metal on atomic-IST-entry.  The testing code tests it in VM and the
super exceptions can be injected at will.

>  * Reinforce what the benefits to end users are here.  Are folks
>    _actually_ running into the #VC fragility, for instance?
>
> Also, let's say we queue this, it starts getting linux-next testing, and
> we start getting bug reports of hangs.  It'll have to get reverted if we
> can't find the bug fast.
>
> How much of a pain would it be to make atomic-IST-entry _temporarily_
> selectable at boot time?  It would obviously need to keep the old code
> around and would not have the nice diffstat.  But that way, folks would
> at least have a workaround while we're bug hunting.

It is easy to make atomic-IST-entry selectable at boot time since IDT
is an indirect table.  I will do it and temporarily keep the old code.