linux-kernel - Re: [PATCH] irqchip/gic: prevent buffer overflow in gic_ipi_send

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

Message-ID: <87wmjmv632.wl-maz@kernel.org>
Date: Sun, 08 Sep 2024 10:37:53 +0100
From: Marc Zyngier <maz@...nel.org>
To: Sergey Shtylyov <s.shtylyov@....ru>
Cc: Thomas Gleixner <tglx@...utronix.de>,
	<linux-kernel@...r.kernel.org>,
	<linux-arm-kernel@...ts.infradead.org>
Subject: Re: [PATCH] irqchip/gic: prevent buffer overflow in gic_ipi_send_mask()

On Fri, 06 Sep 2024 21:29:47 +0100,
Sergey Shtylyov <s.shtylyov@....ru> wrote:
> 
> On 9/5/24 10:47 AM, Marc Zyngier wrote:
> [...]
> 
> >>> ARM GIC arch v2 spec claims support for just 8 CPU interfaces.  However,
> >>> looking at the GIC driver's irq_set_affinity() method, it seems that the
> >>> passed CPU mask may contain the logical CPU #s beyond 8, and that method
> >>> filters them out before reading gic_cpu_map[], bailing out with
> >>> -EINVAL.
> >>
> >> The reasoning is correct in theory, but in reality it's a non problem.
> >>
> >> Simply because processors which use this GIC version cannot have more
> >> than 8 cores.
> >>
> >> That means num_possible_cpus() <= 8 so the cpumask handed in cannot have
> >> bits >= 8 set. Ergo for_each_cpu() can't return a bit which is >= 8.
> > 
> > That.
> 
>    That? :-)

What Thomas explained.

> 
> > The irq_set_affinity() check exists because the affinity can be
> > provided by userspace, and used to be be *anything*. Since
> 
>    In this case you mean gic_set_affinity(), right?

Yes.

> 
> > 33de0aa4bae98, the affinity that the driver gets is narrowed to what
> > is actually *online*.
> 
>    What I haven't quite understood from my (cursory) looking at the GICv2
> spec (and the GIC driver) is why only one CPU (with a lowest #) is selected
> from *mask_val before writing to GICD_GIC_DIST_TARGET, while the spec holds
> that an IRQ can be forwarded to any set of 8 CPU interfaces...

Because on all the existing implementations, having more than a single
target in GICD_ITARGETSRn results in all the targeted CPUs to be
interrupted, with the guarantee that only one will see the actual
interrupt (the read from GICC_IAR returns a value that is not 0x3ff),
and everyone else will only see a spurious interrupt (0x3ff). This is
because the distributor does not track which CPU is actually in a
position to handle the interrupt.

While this can be, under limited circumstances, beneficial from an
interrupt servicing latency, it is always bad from a global throughput
perspective. You end-up thrashing CPU caches, generating odd latencies
in unsuspecting code, and in general with disappointing performance.

Thankfully, GIC (v1/v2) is a dead horse, and v3 doesn't have this
particular problem (it replaced it with a bigger one in the form of
1:n distribution).

	M.

-- 
Without deviation from the norm, progress is not possible.