Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

Attendees

  • Mike
  • Arnd
  • Massami
  • Shashi
  • Alex
  • Bill
  • Mathieu
  • Matt  (Arm)
  • Stefano (Xilinx)
  • Mark
  • Joakim
  • Jean-Phillip
  • Illias
  • Bill
  • Randy
  • Sumit
  • Bogdan (NXP)
  • Srivatsa (QC)
  • Ruchika
  • Souvick
  • Diana (NXP)
  • Ed Doxat (Arm)
  • Pratik
  • Vincent

...

Excerpt


Update on the demo progress

...

Please ping Joakim if you have questions

Some resourses 

Kevin Brodsky's talk giving answers to many of Arnd's questions statements
- https://youtu.be/CFweo8YFevg?t=7214
- https://www.linuxplumbersconf.org/event/7/contributions/654/

The current state of  VirtIO-MMU / Limited Memory Sharing

Slides:

Widget Connector
urlhttps://docs.google.com/presentation/d/1uwP8iPFaPivS-5QU_0ZnuhUHgFCKjdTvJ-H9uygJOUQ/edit#slide=id.p

Arnd: discussed 3 different approaches - do they all suck? Which to start prototyping?

Option 1: host manages shared memory, use swiotlb to force use of shared memory

Option 2: guest uses IOMMU, signals pages meant for access, no copying, expensive transactions updating iommu?

Option 3: new virtquue design, shared memory part of queue (inline).

Steffano: can we increase the size of the queue? Arnd: it involves changing the binary layout of the queue

Steffano: had to increase ring size for 9pfs implementation - needed to go to 2Mb (up from 512K)

Arnd: having large queues risks cache thrashing.

Alex: Does swiotlb always imply copying Arnd: yes

Stefano: what about virtio managing page allocation?

Alex: sees the complexities with existing devices, talked about video - size it for "known" memory bandwidth

Stefano: what about expanding option 3 but allowing bigger sizes queus appropriate to the device?

Arnd: Option 3 has s simplicity, but offers a allocation per device.

Srivista: our implementation currently is closer to option 1

Arnd: so which option to implement, or do we see if QC can share their implementation

Alex: I'm leaning towards 3 which could then be expanded to larger queue sizes for given devices. We could then compare with QC swiotlb approach



What is the development platform ?


Action items

  •