Re: Qemu Guest kernel 4.20-rc1 PCIe hotplug issue

[mika.westerberg@xxxxxxxxxxxxxxx]

+Lukas

On Tue, Nov 13, 2018 at 11:45:42AM +0000, Shameerali Kolothum Thodi wrote:
> Hi Mika,

Hi,

> Since the commit commit 720d6a671a6e("PCI: pciehp: Do not handle events
> if interrupts are masked"), the hotplug support on Qemu Guest(4.120-rc1)
> with a vfio passthrough device seems to be broken. This is on an ARM64 platform.
> 
> I am booting a Guest with below command line options with the intention of
> hot add a ixgbevf dev later,
> 
> ./qemu-system-aarch64 -machine virt,kernel_irqchip=on,gic-version=3 -cpu host \
>  -kernel Image_4.20-rc1 \
>  -initrd rootfs-iperf.cpio \
>  -device ioh3420,id=rp1 \
>  -net none \
>  -m 4096 \
>  -nographic -D -d -enable-kvm \
>  -append "console=ttyAMA0 root=/dev/vda -m 4096 rw pciehp.pciehp_debug=1
>   pcie_ports=native searlycon=pl011,0x9000000"
> 
> But receives this on boot,
> 
> [    1.327852] pciehp 0000:00:01.0:pcie004: Timeout 
> on hotplug command 0x03f1 (issued 1016 msec ago)
> [    1.335842] pciehp 0000:00:01.0:pcie004: Timeout on hotplug command
> 0x03f1 (issued 1024 msec ago)
> [    3.847843] pciehp 0000:00:01.0:pcie004: Failed to check link status
> [    3.855843] pciehp 0000:00:01.0:pcie004: Timeout on hotplug command
> 0x02f1 (issued 2520 msec ago)
> [    4.879846] pciehp 0000:00:01.0:pcie004: Timeout on hotplug command
> 0x06f1 (issued 1024 msec ago)
> [    5.911840] pciehp 0000:00:01.0:pcie004: Timeout on hotplug command
> 0x06f1 (issued 2056 msec ago)
> [    6.927844] pciehp 0000:00:01.0:pcie004: Timeout on hotplug command
> 0x07f1 (issued 1016 msec ago)
> [    7.951843] pciehp 0000:00:01.0:pcie004: Timeout on hotplug command
> 0x0771 (issued 1024 msec ago)
> 
> Trying to hot add using "device_addvfio-pci,host=0000:01:10.1,id=net0,bus=rp1"
> doesn't work either. And if I boot the guest with an assigned device
> (-device vfio-pci,host=0000:01:10.1,id=net0,bus=rp1), I can see the dev listed in
> the Guest but then hot remove doesn't work.
> 
> This all works on 4.19 and bisect points to the above mentioned commit, where an
> additional check is added in pciehp_isr(),
> 
> -	 * Interrupts only occur in D3hot or shallower (PCIe r4.0, sec 6.7.3.4).
> +	 * Interrupts only occur in D3hot or shallower and only if enabled
> +	 * in the Slot Control register (PCIe r4.0, sec 6.7.3.4).
>  	 */
> -	if (pdev->current_state == PCI_D3cold)
> +	if (pdev->current_state == PCI_D3cold ||
> +	    (!(ctrl->slot_ctrl & PCI_EXP_SLTCTL_HPIE) && !pciehp_poll_mode))
>  		return IRQ_NONE;
> 
> I think this doesn't work for the first time, where the cmd with PCI_EXP_SLTCTL_HPIE bit set
> is written,
> pciehp_probe()
>   pcie_init_notification()
>     pcie_enable_notification()
>        pcie_do_write_cmd()
> 
> to begin with, ctrl->slot_ctrl = 0 in pciehp_isr() as this is only set once the write
> is returned.
> 
> Or else I am missing something here. Please take a look and let me know.

Thanks for the detailed report and analysis. I think you are right and
the ctrl->slot_ctrl is only set after the actual value has been
programmed to the hardware, so if there was interrupt "pending" it will
trigger immediately (just to find ctrl->slot_ctrl == 0).

I wonder if the following change helps here?

diff --git a/drivers/pci/hotplug/pciehp_hpc.c b/drivers/pci/hotplug/pciehp_hpc.c
index 7dd443aea5a5..cd9eae650aa5 100644
--- a/drivers/pci/hotplug/pciehp_hpc.c
+++ b/drivers/pci/hotplug/pciehp_hpc.c
@@ -156,9 +156,9 @@ static void pcie_do_write_cmd(struct controller *ctrl, u16 cmd,
 	slot_ctrl |= (cmd & mask);
 	ctrl->cmd_busy = 1;
 	smp_mb();
+	ctrl->slot_ctrl = slot_ctrl;
 	pcie_capability_write_word(pdev, PCI_EXP_SLTCTL, slot_ctrl);
 	ctrl->cmd_started = jiffies;
-	ctrl->slot_ctrl = slot_ctrl;
 
 	/*
 	 * Controllers with the Intel CF118 and similar errata advertise