Re: swiotlb=force in Konrad's xen-pcifront-0.8.2 pvops domU kernel with PCI passthrough

[Dante Cinco <dantecinco@gmail.com>]

On Thu, Nov 18, 2010 at 9:19 AM, Konrad Rzeszutek Wilk
<konrad.wilk@oracle.com> wrote:
> Keir, Dan, Mathieu, Chris, Mukesh,
>
> This fellow is passing in a PCI device to his Xen PV guest and trying
> to get high IOPS. The kernel he is using is a 2.6.36 with tglx's
> sparse_irq rework.
>
>> I wanted to confirm that bounce buffering was indeed occurring so I
>> modified swiotlb.c in the kernel and added printks in the following
>> functions:
>> swiotlb_bounce
>> swiotlb_tbl_map_single
>> swiotlb_tbl_unmap_single
>> Sure enough we were calling all 3 five times per I/O. We took your
>> suggestion and replaced pci_map_single with pci_pool_alloc. The
>> swiotlb calls were gone but the I/O performance only improved 6% (29k
>> IOPS to 31k IOPS) which is still abysmal.
>
> Hey! 6% that is nothing to sneeze at.

When we were using an HVM kernel (2.6.32.15+drm33.5), our IOPS was at
least 20x (~700k IOPS).

>
>>
>> Any suggestions on where to look next? I have one question about the
>
> So since you are talking IOPS I figured you must be using fio to run those
> numbers. And since you mentioned HVM at some point, you are not running
> this PV domain as a back-end for another PV guest. You are probably going
> to run some form of iSCSI target and stuff those down the PCI device.

Our setup is pure Fibre Channel. We're using a physically separate
system (Linux-based also) to initiate the SCSI I/Os.

>
> Couple of things that pop in my head.. but lets first address your question.
>
>> P2M array: Does the P2M lookup occur every DMA or just during the
>> allocation? What I'm getting at is this: Is the Xen-SWIOTLB a central
>
> It only occurs during allocation. Also since you are bypassing the
> bounce buffer those calls are done without any spinlock. The lookup
> of P2M is bitshifting, division - and are constant - so O(1).
>
>> resource that could be a bottleneck?
>
> Doubt it. Your best bet to figure this out is to play with ftrace, or
> perf trace. But I don't know how well they work with Xen nowadays - Jeremy
> and Mathieu Desnoyers poked it a bit and I think I overheard that Mathieu got
> it working?
>
> So the next couple of possiblities are:
>  1). you are hitting the spinlock issues on 'struct request' or any of
>     the paths on the I/O. Oracle did a lot of work on those - and one
>     way to find this out is to look at tracing and see where the contention is.
>     I don't know where or if those patches have been posted upstream.. but as said,
>     if you are seeing the spinlock usage high  - that might be it.
>  1b). Spinlocks - make sure you have CONFIG_PVOPS_SPINLOCK enabled. Otherwise

I checked the config file and it is enabled: CONFIG_PARAVIRT_SPINLOCKS=y

>     you are going to hit dreadfull conditions.
>  2). You are hitting the 64-bit syscall wall. Basically your user-mode
>     application (fio) is doing a write(), which used to be int 0x80 but now
>     is a syscall. The syscall gets trapped in the hypervisor which has to
>     call in your PV kernel. You get hit with two context switches for each
>     'write()' call. The solution is to use a 32-bit DomU as the guest user
>     application and guest kernel run in different rings.

There is no user space application that is involved with the I/O. It's
all kernel driver code that handles the I/O.

>  3). Xen CPU pools. You didn't say where the application that sends the IOs
>     is located. But if it was in a seperate domain then you will want to use
>     Xen CPU pools. Basically this way you can get gang-scheduling where the
>     guest that submits the I/O and the guest that picks up the I/O are running
>     right after each other. I don't know much more details, but this is what
>     I understand it does.
>  4). CPU/MSI-X affinity. I think you already did this, but make sure you pin
>     your guest to specific CPUs and also pin the MSI-X (vectors) to the proper
>     destination. You can use the 'xm debug-keys i' to see the MSI-X affinity - it
>     is a mask and basically see if it overlays the CPUs you are running your guest
>     at. Not sure how to actually set the MSI-X affinity ... now that I think.
>     Keir or some of the Intel folks might know better.

There 16 devices (multi-function) that are PCI-passed through to domU.
There are 16 VCPUs in domU and all are pinned to individual PCPUs
(24-CPU platform). Each IRQ in domU is affinitized to a CPU. This
strategy has worked well for us with the HVM kernel. Here's the output
of 'xm debug-keys i'
(XEN)    IRQ:  67 affinity:ffffffff,ffffffff,ffffffff,ffffffff vec:7a
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:127(----),
(XEN)    IRQ:  68 affinity:00000000,00000000,00000000,00000200 vec:43
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:126(----),
(XEN)    IRQ:  69 affinity:00000000,00000000,00000000,00000400 vec:83
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:125(----),
(XEN)    IRQ:  70 affinity:00000000,00000000,00000000,00000800 vec:4b
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:124(----),
(XEN)    IRQ:  71 affinity:00000000,00000000,00000000,00001000 vec:8b
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:123(----),
(XEN)    IRQ:  72 affinity:00000000,00000000,00000000,00002000 vec:53
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:122(----),
(XEN)    IRQ:  73 affinity:00000000,00000000,00000000,00004000 vec:93
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:121(----),
(XEN)    IRQ:  74 affinity:00000000,00000000,00000000,00008000 vec:5b
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:120(----),
(XEN)    IRQ:  75 affinity:00000000,00000000,00000000,00010000 vec:9b
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:119(----),
(XEN)    IRQ:  76 affinity:00000000,00000000,00000000,00020000 vec:63
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:118(----),
(XEN)    IRQ:  77 affinity:00000000,00000000,00000000,00040000 vec:a3
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:117(----),
(XEN)    IRQ:  78 affinity:00000000,00000000,00000000,00080000 vec:6b
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:116(----),
(XEN)    IRQ:  79 affinity:00000000,00000000,00000000,00100000 vec:ab
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:115(----),
(XEN)    IRQ:  80 affinity:00000000,00000000,00000000,00200000 vec:73
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:114(----),
(XEN)    IRQ:  81 affinity:00000000,00000000,00000000,00400000 vec:b3
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:113(----),
(XEN)    IRQ:  82 affinity:00000000,00000000,00000000,00800000 vec:7b
type=PCI-MSI         status=00000010 in-flight=0
domain-list=1:112(----),

>  5). Andrew, Mukesh, Keir, Dan, any other ideas?
>

We're also trying Chris' 4 things to try and will consider Mathieu's
LTT suggestion.

- Dante