Some MSI are not routed correctly

[John Baldwin]

On Thursday, October 08, 2015 07:33:27 AM Maxim Sobolev wrote:
> Hi John & others,
> 
> We've came across a weird MSI routing issue on one of our newest dual
> E5-2690v3 (haswell) Supermicro X10DRL-i boxes running latest 10.2-p4. It is
> fitted with dual port Intel I350 card, in addition to the built-in I210
> chip that is not used. The hw.igb.num_queues is set to 4, and the driver
> reports binding to the CPUs 0-3 for the first port and CPUs 4-7 for the
> second, however when verified with top -P under the load, interrupts are
> only delivered to the CPUs 0-3, no interrupt time is recorded on the CPUs
> 4-7. systat -vm shows that all 8 queues are firing interrupts, so my guess
> that for whatever reason bus_bind_intr() is not doing what's expected to do
> for half of those interrupts.
> 
> What's interesting is that on a similar box (same chassis/mobo/cpu) but
> equipped with the quad-port X540-AT2 10Gig card, interrupts are routed
> properly. The latter is running with hw.ix.num_queues="3".
> 
> pcib2: <ACPI Host-PCI bridge> port 0xcf8-0xcff on acpi0
> pci0: <ACPI PCI bus> on pcib2
> pcib3: <ACPI PCI-PCI bridge> irq 26 at device 1.0 on pci0
> pci1: <ACPI PCI bus> on pcib3
> igb0: <Intel(R) PRO/1000 Network Connection version - 2.4.0> mem
> 0xc7200000-0xc72fffff,0xc7304000-0xc7307fff irq 26 at device 0.0 on pci1
> igb0: Using MSIX interrupts with 5 vectors
> igb0: Ethernet address: a0:36:9f:76:af:20
> igb0: Bound queue 0 to cpu0
> igb0: Bound queue 1 to cpu1
> igb0: Bound queue 2 to cpu2
> igb0: Bound queue 3 to cpu3
> igb0: netmap queues/slots: TX 4/4096, RX 4/4096
> igb1: <Intel(R) PRO/1000 Network Connection version - 2.4.0> mem
> 0xc7100000-0xc71fffff,0xc7300000-0xc7303fff irq 28 at device 0.1 on pci1
> igb1: Using MSIX interrupts with 5 vectors
> igb1: Ethernet address: a0:36:9f:76:af:21
> igb1: Bound queue 0 to cpu4
> igb1: Bound queue 1 to cpu5
> igb1: Bound queue 2 to cpu6
> igb1: Bound queue 3 to cpu7
> igb1: netmap queues/slots: TX 4/4096, RX 4/4096
> 
> pcib2: <ACPI Host-PCI bridge> port 0xcf8-0xcff on acpi0
> pci0: <ACPI PCI bus> on pcib2
> pcib3: <ACPI PCI-PCI bridge> irq 26 at device 1.0 on pci0
> pci1: <ACPI PCI bus> on pcib3
> pcib4: <ACPI PCI-PCI bridge> irq 32 at device 2.0 on pci0
> pci2: <ACPI PCI bus> on pcib4
> pcib5: <ACPI PCI-PCI bridge> irq 40 at device 3.0 on pci0
> pci3: <ACPI PCI bus> on pcib5
> ix0: <Intel(R) PRO/10GbE PCI-Express Network Driver, Version - 2.8.3> port
> 0x6020-0x603f mem 0xc7c00000-0xc7dfffff,0xc7e04000-0xc7e07fff irq 40 at
> device 0.0 on pci3
> ix0: Using MSIX interrupts with 4 vectors
> ix0: Bound queue 0 to cpu 0
> ix0: Bound queue 1 to cpu 1
> ix0: Bound queue 2 to cpu 2
> ix0: Ethernet address: 0c:c4:7a:5e:be:64
> ix0: PCI Express Bus: Speed 5.0GT/s Width x8
> ix0: netmap queues/slots: TX 3/4096, RX 3/4096
> ix1: <Intel(R) PRO/10GbE PCI-Express Network Driver, Version - 2.8.3> port
> 0x6000-0x601f mem 0xc7a00000-0xc7bfffff,0xc7e00000-0xc7e03fff irq 44 at
> device 0.1 on pci3
> ix1: Using MSIX interrupts with 4 vectors
> ix1: Bound queue 0 to cpu 3
> ix1: Bound queue 1 to cpu 4
> ix1: Bound queue 2 to cpu 5
> ix1: Ethernet address: 0c:c4:7a:5e:be:65
> ix1: PCI Express Bus: Speed 5.0GT/s Width x8
> ix1: netmap queues/slots: TX 3/4096, RX 3/4096
> 
> Some extra debug is here:
> 
> http://sobomax.sippysoft.com/haswell_bug/bad.dmesg
> http://sobomax.sippysoft.com/haswell_bug/lstopo_bad.png
> http://sobomax.sippysoft.com/haswell_bug/systat_vm_bad.png
> http://sobomax.sippysoft.com/haswell_bug/top_P_bad.png
> 
> http://sobomax.sippysoft.com/haswell_bug/good.dmesg
> http://sobomax.sippysoft.com/haswell_bug/lstopo_good.png
> http://sobomax.sippysoft.com/haswell_bug/systat_vm_good.png
> http://sobomax.sippysoft.com/haswell_bug/top_P_good.png
> 
> Any ideas on how to debug that further are welcome. The box in the
> production, but we can remove traffic during off-peak to run some
> test/debug code on.

Can you get procstat -S output for the interrupt threads?  (Usually interrupt
threads are in pid 12, so 'procstat -S 12' would suffice.)

-- 
John Baldwin