linux-kernel - Re: [RFC PATCH v2] PCI: Waiting command completed in get_port_device

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 for Android: free password hash cracker in your pocket

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

Message-ID: <20220111215555.GA19605@wunner.de>
Date:   Tue, 11 Jan 2022 22:55:55 +0100
From:   Lukas Wunner <lukas@...ner.de>
To:     Bjorn Helgaas <helgaas@...nel.org>
Cc:     Yao Hongbo <yaohongbo@...ux.alibaba.com>, bhelgaas@...gle.com,
        zhangliguang@...ux.alibaba.com,
        alikernel-developer@...ux.alibaba.com, linux-pci@...r.kernel.org,
        linux-kernel@...r.kernel.org,
        "Rafael J. Wysocki" <rjw@...ysocki.net>
Subject: Re: [RFC PATCH v2] PCI: Waiting command completed in
 get_port_device_capability()

On Tue, Jan 11, 2022 at 12:55:38PM -0600, Bjorn Helgaas wrote:
> On Fri, Jan 07, 2022 at 11:22:49AM +0800, Yao Hongbo wrote:
> > According to the PCIe specification Revision 5.0, section
> > 7.5.3.11 (slot Status Register), if Command Complete notification

Hm, I only have the PCIe r4.0 Base Spec available and the section
number is 7.8.11 there.  Is 7.5.3.11 really correct for the r5.0 spec?

> > However, before probing the pcie hotplug service, there needs to set
> > HPIE bit in the slot ctrl register to disable hotplug interrupts,
> > and there is no wait currently.

There is a similar issue on resume.  The PCI core writes to the
Slot Control register in...

  pci_pm_resume_noirq()
    pci_pm_default_resume_early()
      pci_restore_state()
        pci_restore_pcie_state()

...and to ensure that the proper Command Completed dance is performed,
we do this in pciehp_resume_noirq() and pciehp_runtime_resume():

	/* pci_restore_state() just wrote to the Slot Control register */
	ctrl->cmd_started = jiffies;
	ctrl->cmd_busy = true;

That was introduced by 469e764c4a3c.

I'd prefer that we do the same in pciehp_probe() (or pcie_init())
if we come to the conlusion that we want to continue disabling HPIE
and CCIE in get_port_device_capability().  It's a much simpler and
smaller solution than the one proposed in the present patch.

> But now, on ACPI systems, we only clear HPIE and CCIE here if we *do*
> have the hotplug driver (because host->native_pcie_hotplug only
> remains set if we have been granted control via _OSC, and we only
> request control when CONFIG_HOTPLUG_PCI_PCIE is enabled).  On these
> systems, we should be able to remove this disable code because pciehp
> will do whatever it needs.

I guess an argument could be made that even though CONFIG_HOTPLUG_PCI_PCIE
is enabled, pciehp may fail to probe, e.g. because the call to kzalloc()
in pcie_init() returns NULL due to insufficient memory.  Then the interrupt
would remain enabled if the BIOS neglected to disable it.  That could be
accounted for by rearranging pciehp such that the interrupt is disabled
first thing on probe.  Of course that approach doesn't solve the problem
if CONFIG_HOTPLUG_PCI_PCIE is disabled and we'd still have to disable
the interrupt in get_port_device_capability() in that case.

I think it generally makes sense to assume the worst and compensate for
broken BIOSes.

Hotplug and PME share the same MSI or MSI-X vector per section 6.7.3.4
of the PCIe Base Spec, so enabled but unhandled hotplug interrupts will
cause the PME interrupt handler to run.  That's probably not fatal,
but still undesirable.

Thanks,

Lukas