Convert the number of MSI IRQs on x86 from a constant to a tunable.

[jhb]

Summary

The number of MSI IRQs still defaults to 512, but it can now be changed at
boot time via the machdep.num_msi_irqs tunable.

Test Plan

• booted a bhyve VM similar to the setup described in D17976. With machdep.num_msi_irqs=11 it had the same correct behavior described in the test plan for D17976.

Note, xen_msi might need it's own validation as it doesn't seem prepared
to cope with num_msi_irqs == 0, for example. I have not tested XEN.

[markj]

sys/x86/x86/msi.c

157

We have SYSCTL_UINT.

sys/x86/xen/xen_msi.c

78

Get rid of redundant parens while here?

[royger]

I plan to get rid of the Xen specific MSI implementation soon.

[jhb]

sys/x86/x86/msi.c

342–344

I think at a minimum I should add an overflow check here actually.

[jhb]

• Fixes from markj.
• Add overflow checks.

[jhb]

sys/x86/xen/xen_intr.c

693

So sadly clang omits this overflow check though it keeps the one in msi.c. I was testing by setting machdep.num_msi_irqs to 0xfffffff00 which panicked in msi.c and then tried 0xfffffeff which should have panicked here but didn't and just overflowed num_io_irqs to 4095.

[markj]

sys/x86/xen/xen_intr.c

693

Do you need to cast NR_EVENT_CHANNELS to u_int?

[jhb]

• Rework the Xen overflow check due to size_t promotion.

[jhb]

sys/x86/xen/xen_intr.c

693

Thanks to Mark for noticing that NR_EVENT_CHANNELS ends up being a size_t which is why the original expression was always false. This one works. Note that trying to boot with a large number of interrupts that didn't overflow wasn't very productive as the interrupt_sources[] array can be nearly 4GB and that's an awfully large size to pass to malloc().

[markj]

sys/x86/xen/xen_msi.c

60

Hmm, it'd be consistent to convert this to if (num_msi_irqs > UINT_MAX - first_msi_irqs).

[jhb]

• Adjust other overflow checks.