linux-kernel - Re: [PATCH 1/3] x86/quirks: Scan all busses for early PCI quirks

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: <20181018221538.GN5906@bhelgaas-glaptop.roam.corp.google.com>
Date:   Thu, 18 Oct 2018 17:15:39 -0500
From:   Bjorn Helgaas <helgaas@...nel.org>
To:     "Guilherme G. Piccoli" <gpiccoli@...onical.com>
Cc:     linux-pci@...r.kernel.org, kexec@...ts.infradead.org,
        x86@...nel.org, linux-kernel@...r.kernel.org, bhelgaas@...gle.com,
        dyoung@...hat.com, bhe@...hat.com, vgoyal@...hat.com,
        tglx@...utronix.de, mingo@...hat.com, bp@...en8.de, hpa@...or.com,
        andi@...stfloor.org, lukas@...ner.de, billy.olsen@...onical.com,
        cascardo@...onical.com, ddstreet@...onical.com,
        fabiomirmar@...onical.com, gavin.guo@...onical.com,
        jay.vosburgh@...onical.com, kernel@...ccoli.net, mfo@...onical.com,
        shan.gavin@...ux.alibaba.com
Subject: Re: [PATCH 1/3] x86/quirks: Scan all busses for early PCI quirks

On Thu, Oct 18, 2018 at 03:37:19PM -0300, Guilherme G. Piccoli wrote:
> Recently was noticed in an HP GEN9 system that kdump couldn't succeed
> due to an irq storm coming from an Intel NIC, narrowed down to be lack
> of clearing the MSI/MSI-X enable bits during the kdump kernel boot.
> For that, we need an early quirk to manually turn off MSI/MSI-X for
> PCI devices - this was worked as an optional boot parameter in a
> subsequent patch.
> 
> Problem is that in our test system, the Intel NICs were not present in
> any secondary bus under the first PCIe root complex, so they couldn't
> be reached by the recursion in check_dev_quirk(). Modern systems,
> specially with multi-processors and multiple NUMA nodes expose multiple
> root complexes, describing more than one PCI hierarchy domain. Currently
> the simple recursion present in the early-quirks code from x86 starts a
> descending recursion from bus 0000:00, and reach many other busses by
> navigating this hierarchy walking through the bridges. This is not
> enough in systems with more than one root complex/host bridge, since
> the recursion won't "traverse" to other root complexes by starting
> statically in 0000:00 (for more details, see [0]).
> 
> This patch hence implements the full bus/device/function scan in
> early_quirks(), by checking all possible busses instead of using a
> recursion based on the first root bus or limiting the search scope to
> the first 32 busses (like it was done in the beginning [1]).

I don't want to expand the early quirk infrastructure unless there is
absolutely no other way to solve this.  The early quirk stuff is
x86-specific, and it's not obvious that this problem is x86-only.

This patch scans buses 0-255, but still only in domain 0, so it won't
help with even more complicated systems that use other domains.

I'm not an IRQ expert, but it seems wrong to me that we are enabling
this interrupt before we're ready for it.  The MSI should target an
IOAPIC.  Can't that IOAPIC entry be masked until later?  I guess the
kdump kernel doesn't know what MSI address the device might be using.

Could the IRQ core be more tolerant of this somehow, e.g., if it
notices incoming interrupts with no handler, could it disable the
IOAPIC entry and fall back to polling periodically until a handler is
added?

> [0] https://bugs.launchpad.net/bugs/1797990
> 
> [1] From historical perspective, early PCI scan dates back
> to BitKeeper, added by Andi Kleen's "[PATCH] APIC fixes for x86-64",
> on October/2003. It initially restricted the search to the first
> 32 busses and slots.
> 
> Due to a potential bug found in Nvidia chipsets, the scan
> was changed to run only in the first root bus: see
> commit 8659c406ade3 ("x86: only scan the root bus in early PCI quirks")
> 
> Finally, secondary busses reachable from the 1st bus were re-added back by:
> commit 850c321027c2 ("x86/quirks: Reintroduce scanning of secondary buses")
> 
> Reported-by: Dan Streetman <ddstreet@...onical.com>
> Signed-off-by: Guilherme G. Piccoli <gpiccoli@...onical.com>
> ---
>  arch/x86/kernel/early-quirks.c | 13 ++++++-------
>  1 file changed, 6 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/x86/kernel/early-quirks.c b/arch/x86/kernel/early-quirks.c
> index 50d5848bf22e..fd50f9e21623 100644
> --- a/arch/x86/kernel/early-quirks.c
> +++ b/arch/x86/kernel/early-quirks.c
> @@ -731,7 +731,6 @@ static int __init check_dev_quirk(int num, int slot, int func)
>  	u16 vendor;
>  	u16 device;
>  	u8 type;
> -	u8 sec;
>  	int i;
>  
>  	class = read_pci_config_16(num, slot, func, PCI_CLASS_DEVICE);
> @@ -760,11 +759,8 @@ static int __init check_dev_quirk(int num, int slot, int func)
>  	type = read_pci_config_byte(num, slot, func,
>  				    PCI_HEADER_TYPE);
>  
> -	if ((type & 0x7f) == PCI_HEADER_TYPE_BRIDGE) {
> -		sec = read_pci_config_byte(num, slot, func, PCI_SECONDARY_BUS);
> -		if (sec > num)
> -			early_pci_scan_bus(sec);
> -	}
> +	if ((type & 0x7f) == PCI_HEADER_TYPE_BRIDGE)
> +		return -1;
>  
>  	if (!(type & 0x80))
>  		return -1;
> @@ -787,8 +783,11 @@ static void __init early_pci_scan_bus(int bus)
>  
>  void __init early_quirks(void)
>  {
> +	int bus;
> +
>  	if (!early_pci_allowed())
>  		return;
>  
> -	early_pci_scan_bus(0);
> +	for (bus = 0; bus < 256; bus++)
> +		early_pci_scan_bus(bus);
>  }
> -- 
> 2.19.0
>