Re: [v2] genirq/msi: Add the address and data that show MSI/MSIX
From: Hans Zhang
Date: Sun Mar 02 2025 - 09:47:00 EST
Hi Thomas(tglx),
On 2025/3/2 17:01, Thomas Gleixner wrote:
Hans!
On Sat, Mar 01 2025 at 20:39, Hans Zhang wrote:
The debug_show() callback function is implemented in the MSI core code.
And assign it to the domain ops::debug_show() creation.
cat /sys/kernel/debug/irq/irqs/msi_irq_num, the address and data stored
in the MSI capability or the address and data stored in the MSIX vector
table will be displayed.
So this explains what the patch is doing and what the output is. But it
fails to explain the _why_. Documentation gives proper guidance:
https://www.kernel.org/doc/html/latest/process/maintainer-tip.html#changelog
https://www.kernel.org/doc/html/latest/process/submitting-patches.html#describe-changes
Is the following explanation OK?
Displaying the address and data of the MSI/MSIX interrupt in the debugfs
helps with debugging.
e.g.
root@root:/sys/kernel/debug/irq/irqs# cat /proc/interrupts | grep ITS
85: 0 0 0 0 0 0 0 0 0 0 0 0 ITS-MSI 75497472 Edge PCIe PME, aerdrv
86: 0 30 0 0 0 0 0 0 0 0 0 0 ITS-MSI 76021760 Edge nvme0q0
87: 287 0 0 0 0 0 0 0 0 0 0 0 ITS-MSI 76021761 Edge nvme0q1
88: 0 265 0 0 0 0 0 0 0 0 0 0 ITS-MSI 76021762 Edge nvme0q2
89: 0 0 177 0 0 0 0 0 0 0 0 0 ITS-MSI 76021763 Edge nvme0q3
90: 0 0 0 76 0 0 0 0 0 0 0 0 ITS-MSI 76021764 Edge nvme0q4
91: 0 0 0 0 161 0 0 0 0 0 0 0 ITS-MSI 76021765 Edge nvme0q5
92: 0 0 0 0 0 991 0 0 0 0 0 0 ITS-MSI 76021766 Edge nvme0q6
93: 0 0 0 0 0 0 194 0 0 0 0 0 ITS-MSI 76021767 Edge nvme0q7
94: 0 0 0 0 0 0 0 94 0 0 0 0 ITS-MSI 76021768 Edge nvme0q8
95: 0 0 0 0 0 0 0 0 148 0 0 0 ITS-MSI 76021769 Edge nvme0q9
96: 0 0 0 0 0 0 0 0 0 261 0 0 ITS-MSI 76021770 Edge nvme0q10
97: 0 0 0 0 0 0 0 0 0 0 127 0 ITS-MSI 76021771 Edge nvme0q11
98: 0 0 0 0 0 0 0 0 0 0 0 317 ITS-MSI 76021772 Edge nvme0q12
How is this relevant to describe the patch?
The preceding information will be deleted in v4 patch.
root@root:/sys/kernel/debug/irq/irqs#
root@root:/sys/kernel/debug/irq/irqs# cat 87
handler: handle_fasteoi_irq
device: 0000:91:00.0
status: 0x00000000
istate: 0x00004000
ddepth: 0
wdepth: 0
dstate: 0x31600200
IRQD_ACTIVATED
IRQD_IRQ_STARTED
IRQD_SINGLE_TARGET
IRQD_AFFINITY_MANAGED
IRQD_AFFINITY_ON_ACTIVATE
IRQD_HANDLE_ENFORCE_IRQCTX
node: 0
affinity: 0
effectiv: 0
domain: :soc@0:interrupt-controller@0e001000:its@0e050000-3
hwirq: 0x4880001
chip: ITS-MSI
This output is from a pre 6.11 kernel...
I will delete all other information except for what my patch will display.
flags: 0x20
IRQCHIP_ONESHOT_SAFE
msix:
address_hi: 0x00000000
address_lo: 0x0e060040
msg_data: 0x00000001
For demonstration it's enough to stop here, no?
Yes, I will change it to the following:
msix:
address_hi: 0x00000000
address_lo: 0x0e060040
msg_data: 0x00000001
+static void msi_domain_debug_show(struct seq_file *m, struct irq_domain *d,
+ struct irq_data *irqd, int ind)
+{
+ struct msi_desc *desc;
+ bool is_msix;
+
+ desc = irq_get_msi_desc(irqd->irq);
Move this up to the declaration.
+ if (!desc)
+ return;
+
+ is_msix = desc->pci.msi_attrib.is_msix;
That's not valid for non PCI MSI interrupts.
Do you mean to remove the following two lines of code?
is_msix = desc->pci.msi_attrib.is_msix;
seq_printf(m, "%*s%s:", ind, "", is_msix ? "msix" : "msi");
This function is used for all types of MSI interrupts. So for non PCI
MSI interrupts this will output random garbage. Just print the address
and be done with it. The MSI variant is visible from the chip name on
current kernels. It's either ITS-PCI-MSI or ITS-PCI-MSIX and not
ITS-MSI.
+ seq_printf(m, "%*s%s:", ind, "", is_msix ? "msix" : "msi");
+ seq_printf(m, "\n%*saddress_hi: 0x%08x", ind + 1, "", desc->msg.address_hi);
+ seq_printf(m, "\n%*saddress_lo: 0x%08x", ind + 1, "", desc->msg.address_lo);
+ seq_printf(m, "\n%*smsg_data: 0x%08x\n", ind + 1, "", desc->msg.data);
+}
+
static const struct irq_domain_ops msi_domain_ops = {
.alloc = msi_domain_alloc,
.free = msi_domain_free,
.activate = msi_domain_activate,
.deactivate = msi_domain_deactivate,
.translate = msi_domain_translate,
+ .debug_show = msi_domain_debug_show,
This does not build when CONFIG_GENERIC_IRQ_DEBUGFS=n.
Kernel test robot has reported a compilation error, and I have submitted
v3 patch to solve this problem. I will fix all your questions in v4 patch.
Finally, thank you very much for all your comments.
Best regards
Hans