Re: [PATCH v2 05/13] x86/irq: Reserve a per CPU IDT vector for posted MSIs

From: Arnaldo Carvalho de Melo
Date: Fri Apr 19 2024 - 16:07:31 EST


On Fri, Apr 19, 2024 at 06:00:24AM +0200, Thomas Gleixner wrote:
> On Mon, Apr 15 2024 at 13:43, Jacob Pan wrote:
> > On Mon, 15 Apr 2024 11:53:58 -0700, Jacob Pan <jacob.jun.pan@xxxxxxxxxxxxxxx> wrote:
> >> On Thu, 11 Apr 2024 18:51:14 +0200, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >> > If we really care then we do it proper for _all_ of them. Something like
> >> > the uncompiled below. There is certainly a smarter way to do the build
> >> > thing, but my kbuild foo is rusty.
> >> I too had the concern of the wasting system vectors, but did not know how
> >> to fix it. But now your code below works well. Tested without KVM in
> >> .config to show the gaps:
> >>
> >> In VECTOR IRQ domain.
> >>
> >> BEFORE:
> >> System: 46: 0-31,50,235-236,244,246-255
> >>
> >> AFTER:
> >> System: 46: 0-31,50,241-242,245-255
> >>
> >> The only gap is MANAGED_IRQ_SHUTDOWN_VECTOR(243), which is expected on a
> >> running system.
> >>
> >> Verified in irqvectors.s: .ascii "->MANAGED_IRQ_SHUTDOWN_VECTOR $243
> >>
> >> POSTED MSI/first system vector moved up from 235 to 241 for this case.
> >>
> >> Will try to let tools/arch/x86/include/asm/irq_vectors.h also use it
> >> instead of manually copy over each time. Any suggestions greatly
> >> appreciated.
> >>
> > On a second thought, if we make system IRQ vector determined at compile
> > time based on different CONFIG options, will it break userspace tools such
> > as perf? More importantly the rule of not breaking userspace.

The rule for tools/perf is "don't impose _any requirement_ on the kernel
developers, they don't have to test if any change they do outside of
tools/ will break something inside tools/."

> tools/arch/x86/include/asm/irq_vectors.h is only used to generate the
> list of system vectors for pretty output. And your change already broke
> that.

Yeah, I even moved that from tools/arch/x86/include/asm/irq_vectors.h
to tools/perf/trace/beauty/arch/x86/include/asm/irq_vectors.h (for next
merge window).

Having it in tools/arch/x86/include/asm/irq_vectors.h was a bad decision
as it, as you mentinoned, is only used to generate string tables:

⬢[acme@toolbox perf-tools-next]$ tools/perf/trace/beauty/tracepoints/x86_irq_vectors.sh
static const char *x86_irq_vectors[] = {
[0x02] = "NMI",
[0x80] = "IA32_SYSCALL",
[0xec] = "LOCAL_TIMER",
[0xed] = "HYPERV_STIMER0",
[0xee] = "HYPERV_REENLIGHTENMENT",
[0xef] = "MANAGED_IRQ_SHUTDOWN",
[0xf0] = "POSTED_INTR_NESTED",
[0xf1] = "POSTED_INTR_WAKEUP",
[0xf2] = "POSTED_INTR",
[0xf3] = "HYPERVISOR_CALLBACK",
[0xf4] = "DEFERRED_ERROR",
[0xf6] = "IRQ_WORK",
[0xf7] = "X86_PLATFORM_IPI",
[0xf8] = "REBOOT",
[0xf9] = "THRESHOLD_APIC",
[0xfa] = "THERMAL_APIC",
[0xfb] = "CALL_FUNCTION_SINGLE",
[0xfc] = "CALL_FUNCTION",
[0xfd] = "RESCHEDULE",
[0xfe] = "ERROR_APIC",
[0xff] = "SPURIOUS_APIC",
};

⬢[acme@toolbox perf-tools-next]$

Used in:

root@number:~# perf trace -a -e irq_vectors:irq_work_entry/max-stack=32/ --max-events=1
0.000 kworker/u57:0-/9912 irq_vectors:irq_work_entry(vector: IRQ_WORK)
__sysvec_irq_work ([kernel.kallsyms])
__sysvec_irq_work ([kernel.kallsyms])
sysvec_irq_work ([kernel.kallsyms])
asm_sysvec_irq_work ([kernel.kallsyms])
_raw_spin_unlock_irqrestore ([kernel.kallsyms])
dma_fence_wait_timeout ([kernel.kallsyms])
intel_atomic_commit_tail ([kernel.kallsyms])
process_one_work ([kernel.kallsyms])
worker_thread ([kernel.kallsyms])
kthread ([kernel.kallsyms])
ret_from_fork ([kernel.kallsyms])
ret_from_fork_asm ([kernel.kallsyms])
root@number:~#

But as the original cset introducing this explains, these irq_vectors:
tracepoins operate on just one of the vectors, so irq_work_entry(vector:
IRQ_WORK), irq_vectors:reschedule_exit(vector: RESCHEDULE), etc.

> The obvious solution to that is to expose that list in sysfs for
> consumption by perf.

nah, the best thing these days is stop using 'int' for vector and use
'enum irq_vector', then since we have BTF we can use that to do the enum
-> string translation, like with (using /sys/kernel/btf/vmlinux, that is
pretty much available everywhere these days):

root@number:~# pahole clocksource_ids
enum clocksource_ids {
CSID_GENERIC = 0,
CSID_ARM_ARCH_COUNTER = 1,
CSID_MAX = 2,
};

root@number:~# pahole skb_drop_reason | head
enum skb_drop_reason {
SKB_NOT_DROPPED_YET = 0,
SKB_CONSUMED = 1,
SKB_DROP_REASON_NOT_SPECIFIED = 2,
SKB_DROP_REASON_NO_SOCKET = 3,
SKB_DROP_REASON_PKT_TOO_SMALL = 4,
SKB_DROP_REASON_TCP_CSUM = 5,
SKB_DROP_REASON_SOCKET_FILTER = 6,
SKB_DROP_REASON_UDP_CSUM = 7,
SKB_DROP_REASON_NETFILTER_DROP = 8,
root@number:~#

Then its easy to go from 0 to CSID_GENERIC, etc.

⬢[acme@toolbox pahole]$ perf stat -e cycles pahole skb_drop_reason > /dev/null

Performance counter stats for 'pahole skb_drop_reason':

6,095,427 cpu_atom/cycles:u/ (2.82%)
103,694,633 cpu_core/cycles:u/ (97.18%)

0.039031759 seconds time elapsed

0.016028000 seconds user
0.023007000 seconds sys


⬢[acme@toolbox pahole]$

- Arnaldo

> But we don't have to do any of that right away. It's an orthogonal
> issue. Just waste the extra system vector to start with and then we can
> add the compile time dependend change on top if we really care about
> gaining back the vectors.
>
> Thanks,
>
> tglx