[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <8df3b328-0e6e-4117-9707-de0e39f74d2e@linux.dev>
Date: Mon, 4 Aug 2025 16:20:26 -0400
From: Sean Anderson <sean.anderson@...ux.dev>
To: Lorenzo Pieralisi <lpieralisi@...nel.org>,
Krzysztof WilczyĆski <kw@...ux.com>,
Manivannan Sadhasivam <manivannan.sadhasivam@...aro.org>,
linux-pci@...r.kernel.org, Thomas Gleixner <tglx@...utronix.de>
Cc: Rob Herring <robh@...nel.org>, Mahesh J Salgaonkar
<mahesh@...ux.ibm.com>, Oliver O'Halloran <oohall@...il.com>,
Bjorn Helgaas <bhelgaas@...gle.com>, Michal Simek <michal.simek@....com>,
linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org
Subject: Re: [BUG] pci: nwl: Unhandled AER correctable error
On 8/1/25 13:43, Sean Anderson wrote:
> Hi,
>
> AER correctable errors are pretty rare. I only saw one once before and
> came up with commit 78457cae24cb ("PCI: xilinx-nwl: Rate-limit misc
> interrupt messages") in response. I saw another today and,
> unfortunately, clearing the correctable AER bit in MSGF_MISC_STATUS is
> not sufficient to handle the IRQ. It gets immediately re-raised,
> preventing the system from making any other progress. I suspect that it
> needs to be cleared in PCI_ERR_ROOT_STATUS. But since the AER IRQ never
> gets delivered to aer_irq, those registers never get tickled.
>
> The underlying problem is that pcieport thinks that the IRQ is going to
> be one of the MSIs or a legacy interrupt, but it's actually a native
> interrupt:
>
> CPU0 CPU1 CPU2 CPU3
> 42: 0 0 0 0 GICv2 150 Level nwl_pcie:misc
> 45: 0 0 0 0 nwl_pcie:legacy 0 Level PCIe PME, aerdrv
> 46: 25 0 0 0 nwl_pcie:msi 524288 Edge nvme0q0
> 47: 0 0 0 0 nwl_pcie:msi 524289 Edge nvme0q1
> 48: 0 0 0 0 nwl_pcie:msi 524290 Edge nvme0q2
> 49: 46 0 0 0 nwl_pcie:msi 524291 Edge nvme0q3
> 50: 0 0 0 0 nwl_pcie:msi 524292 Edge nvme0q4
>
> In the above example, AER errors will trigger interrupt 42, not 45.
> Actually, there are a bunch of different interrupts in MSGF_MISC_STATUS,
> so maybe nwl_pcie_misc_handler should be an interrupt controller
> instead? But even then pcie_port_enable_irq_vec() won't figure out the
> correct IRQ. Any ideas on how to fix this?
OK, so as a first pass, maybe something like
if (misc_stat & (MSGF_MISC_SR_FATAL_AER | MSGF_MISC_SR_NON_FATAL_AER
MSGF_MISC_SR_CORR_AER))
generic_handle_domain_irq(pcie->legacy_irq_domain, 0);
to simulate the correct IRQ. I have no idea whether it's safe to call
generic_handle_domain_irq in this context. It wasn't OK for AER (see
commit 9ae052253785 ("PCI/AER: Fix the broken interrupt injection")),
but maybe it's OK for us since the legacy irqchip doesn't support
affinity? I CC'd Thomas and maybe he can comment.
Otherwise, maybe the best thing is to just add an API to manually trigger AER.
> Additionally, any tips on actually triggering AER/PME stuff in a
> consistent way? Are there any off-the-shelf cards for sending weird PCIe
> stuff over a link for testing? Right now all I have
But I still don't know how to test this. I can inject a misc interrupt
since the GIC supports irq_set_irqchip_state, but that won't really
simulate an AER interrupt since MSGF_MISC_STATUS won't have the right
bit set. Maybe I can wiggle a card around in its slot? Maybe PME or link
bandwidth notification could trigger this as well?
--Sean
Powered by blists - more mailing lists