linux-kernel - Re: [PATCH -next] genirq: Add SPARSE_NR

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Thu, 10 Nov 2022 04:01:47 +0100
From:   Thomas Gleixner <tglx@...utronix.de>
To:     Chen Lifu <chenlifu@...wei.com>, linux-kernel@...r.kernel.org
Cc:     chenlifu@...wei.com
Subject: Re: [PATCH -next] genirq: Add SPARSE_NR_IRQS Kconfig option

On Fri, Sep 30 2022 at 16:58, Chen Lifu wrote:
> On a large-scale multi-core and NUMA platform, more than 1024 cores and
> 16 NUMA nodes for example, even if SPASE_IRQ is selected to increase the
> number of interrupt numbers by 8196 base on NR_IRQS, the interrupt numbers
> requirement cannot be met. Therefore, make the number of sparse interrupt
> numbers configurable.

Why?

Let me walk you through:

# git grep '#define\sNR_IRQS' arch/x86
arch/x86/include/asm/irq_vectors.h:#define NR_IRQS_LEGACY			16
arch/x86/include/asm/irq_vectors.h:#define NR_IRQS						\
arch/x86/include/asm/irq_vectors.h:#define	NR_IRQS				(NR_VECTORS + IO_APIC_VECTOR_LIMIT)
arch/x86/include/asm/irq_vectors.h:#define NR_IRQS				(NR_VECTORS + CPU_VECTOR_LIMIT)
arch/x86/include/asm/irq_vectors.h:#define NR_IRQS				NR_IRQS_LEGACY

Versus:

# git grep '#define\sNR_IRQS' arch/arm64

Empty. Oooops. Where does it get the define from on which the define you
try to influence with your config knob depends on, i.e.

    # define IRQ_BITMAP_BITS	(NR_IRQS + 8196)

Good question, right?

But not rocket science to answer. If there is no architecture specific
definition then there is a close to 100% probability that there is a
generic fallback define in include/asm-generic/. And indeed

# git grep '#define\sNR_IRQS' include/

include/asm-generic/irq.h:#define NR_IRQS 64

So let's do the math:

   (64 + 8196) / 1024 ~= 8

Unsurprisingly enough this does barely cope for the per CPU requirements
of an ARM64 system. Of course if you add a proper amount of periphery
then you surely run out of interrupt numbers....

Now let me go back to the grep I did on x86. That matches on some lines
w/o context, but let me show you the full context of the relevant
configuration of a enterprisy machine with all bells and whistels here
for illustration:

#define NR_VECTORS			256
#define MAX_IO_APICS			128
#define CPU_VECTOR_LIMIT		(64 * NR_CPUS)
#define IO_APIC_VECTOR_LIMIT		(32 * MAX_IO_APICS)

So in the case of a NR_CPUS=1024 configuration this evaluates to:

CPU_VECTOR_LIMIT	64*1024 = 65536
IO_APIC_VECTOR_LIMIT    32*128	=  4096

and then the relevant NR_IRQS define:

#define NR_IRQS						\
	(CPU_VECTOR_LIMIT > IO_APIC_VECTOR_LIMIT ?	\
		(NR_VECTORS + CPU_VECTOR_LIMIT)  :	\
		(NR_VECTORS + IO_APIC_VECTOR_LIMIT))

evaluates to:

 (65536 > 4096) ? (256 * 65536) : (256 * 4096) = 256 * 65536 = ....

While the resulting number is silly large, you should be able to figure
out the fundamental difference of the approach to limit the number of
interrupts between the x86 and the arm64 case, right?

There is no good reason to copy the insanely large numbers of x86 but
you surely can figure out how the key component in that calculation
which takes NR_CPUs into account avoids another random uncomprehensible
configuration knob:

    # define IRQ_BITMAP_BITS	(NR_IRQS + 8196)

No?

Thanks,

        tglx