On Wed, Aug 26, 2020 at 7:27 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> On Wed, Aug 26 2020 at 13:53, Alexander Graf wrote:
> > Commit 633260fa143 ("x86/irq: Convey vector as argument and not in ptregs")
> > changed the handover logic of the vector identifier from ~vector in orig_ax
> > to purely register based. Unfortunately, this field has another consumer
> > in the APIC code which the commit did not touch. The net result was that
> > IRQ balancing did not work and instead resulted in interrupt storms, slowing
> > down the system.
>
> The net result is an observationof the symptom but that does not explain
> what the underlying technical issue is.
>
> > This patch restores the original semantics that orig_ax contains the vector.
> > When we receive an interrupt now, the actual vector number stays stored in
> > the orig_ax field which then gets consumed by the APIC code.
> >
> > To ensure that nobody else trips over this in the future, the patch also adds
> > comments at strategic places to warn anyone who would refactor the code that
> > there is another consumer of the field.
> >
> > With this patch in place, IRQ balancing works as expected and performance
> > levels are restored to previous levels.
>
> There's a lot of 'This patch and we' in that changelog. Care to grep
> for 'This patch' in Documentation/process/ ?
>
> > diff --git a/arch/x86/entry/entry_32.S b/arch/x86/entry/entry_32.S
> > index df8c017..22e829c 100644
> > --- a/arch/x86/entry/entry_32.S
> > +++ b/arch/x86/entry/entry_32.S
> > @@ -727,7 +727,7 @@ SYM_CODE_START_LOCAL(asm_\cfunc)
> > ENCODE_FRAME_POINTER
> > movl %esp, %eax
> > movl PT_ORIG_EAX(%esp), %edx /* get the vector from stack */
> > - movl $-1, PT_ORIG_EAX(%esp) /* no syscall to restart */
> > + /* keep vector on stack for APIC's irq_complete_move() */
>
> Yes that's fixing your observed wreckage, but it introduces a worse one.
>
> user space
> -> interrupt
> push vector into orig_ax (values are in the ranges of 0-127 and -128 - 255
> except for the system vectors which do
> not go through this code)
> handle()
> ...
> exit_to_user_mode_loop()
> arch_do_signal()
> /* Did we come from a system call? */
> if (syscall_get_nr(current, regs) >= 0) {
>
> ---> BOOM for any vector 0-127 because syscall_get_nr()
> resolves to regs->orig_ax
>
> Going to be fun to debug.
>
> The below nasty hack cures it, but I hate it with a passion. I'll look
> deeper for a sane variant.
>
Fundamentally, the way we overload orig_ax is problematic. I have a
half-written series to improve it, but my series is broken. I think
it's fixable, though.
First is this patch to use some __csh bits to indicate the entry type.
As far as I know, this patch is correct:
https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/entry&id=dfff54208072a27909ae97ebce644c251a233ff2
Then I wrote this incorrect patch:
https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/entry&id=3a5087acb8a2cc1e88b1a55fa36c2f8bef370572
That one is wrong because the orig_ax wreckage seems to have leaked
into user ABI -- user programs think that orig_ax has certain
semantics on user-visible entries.
But I think that the problem in this thread could be fixed quite
nicely by the first patch, plus a new CS_ENTRY_IRQ and allocating
eight bits of __csh to store the vector. Then we could read out the
vector.
--Andy
