QEMU/KVM Call notes for Jan 25, 2011

Hello,

Please find attached my notes from the weekly QEMU/KVM call January 25, 2011. My apologies if I got something wrong.

Jes

- QEMU 0.14/0.15 releases
- Feb 1st 2011 branch to stable tree for 0.14. Anthony would like to do a formal release quickly, so hopefully within 1-2 weeks after snapshot (Feb 15, 2011).
- We should plan a formal release schedule for 0.15
http://wiki.qemu.org/Planning/0.15-example

- coroutines
- Proposal from Stefan Hajnoczi
http://www.mail-archive.com/[email protected]/msg52522.html
- First example is to try and calls to synchronous system calls in QCOW2 codebase.
- Suggest every image format in block layer have it's own coroutine layer.
- Long term, can/should we apply this to entire block layer?
Q: Can we go to real threads in the block layer?
A: coroutines are light weight, and we have more control, but it is
an option. Going to threads would require a serious amount of work making the full QEMU stack thread safe.
Doesn't solve the parallelism problem, so not the final long term solution to the scalability problem.
Q: Should we do coroutines in the block layer rather than the image
formats? This would provide aio support for formats that do not currently implement aio calls.
A: Avi suggests mixed approach, to benefit from not having explicit
synchronization points in high performance formats.
Follow-on longer discussion about implementation details, which is to be followed up in email.

- glib integration
- Proposal from Anthony
http://www.mail-archive.com/[email protected]/msg52798.html
- Provides portable interfaces to threading, data structures, and utility functions.
- Idea would be to get rid of a lot of duplicated code which we would get from glib instead.
- Also consider gio and gobject.
- gobject could be a replacement for qdev
- gobject could help make vnc server become standalone
- Not suggesting to convert everything to glib types (gint, guint, etc). Clarification of which types are recommended will be needed in CODING_STYLE
Q: Which would be the first subsystem to convert
A: Convert QEMU threads to gthreads
- Comment, doing a conversion of QMP would be good, but Luiz will not have time to look at this for maybe 2 months.
- Suggestion from Paolo Bonzini to make a standalone library for JSON, rather than relying on the glib implementation

- Google Summer of Code 2011 - please post suggestions to the list, and we should revisit this item next week.