mbox series

[v3,0/6] Implement icount=auto using TCG Plugins

Message ID 20240530220610.1245424-1-pierrick.bouvier@linaro.org (mailing list archive)
Headers show
Series Implement icount=auto using TCG Plugins | expand

Message

Pierrick Bouvier May 30, 2024, 10:06 p.m. UTC
The goal here is to be able to scale temporally execution of qemu-user/system,
using a given number of instructions per second.

We define a virtual clock, that can be late or in advance compared to real time.
When we are in advance, we slow execution (by sleeping) until catching real
time.

Finally, we should be able to cleanup icount=auto mode completely, and keep
icount usage for determistic purposes only.

It is built upon new TCG Plugins inline ops (store + conditional callbacks), now
merged on master.

Example in user-mode:

- Retrieve number of instructions to execute /bin/true
$ ./build/qemu-x86_64 -plugin ./build/tests/plugin/libinsn.so -d plugin /bin/true
cpu 0 insns: 120546
total insns: 120546
- Slow execution to match 5 seconds
$ time ./build/qemu-x86_64 -plugin ./build/contrib/plugins/libips,ips=$((120546/5)) /bin/true
real    0m4.985s

Tested in system mode by booting a full debian system, and using:
$ sysbench cpu run
Performance decrease linearly with the given number of ips.

v2
--

- Added missing personal Signed-off-by for commits from Alex
- Fix bad rebase in stubs/meson.build
- move qtest_{get,set}_virtual_clock to accel/qtest/qtest.c
- A race condition was identified for plugins init/idle/resume, but is not
  related to this series, and will be fixed in another one:
  https://lore.kernel.org/qemu-devel/20240529152219.825680-1-alex.bennee@linaro.org/

v3
--

- remove precise execution (per insn, per tb is enough)
- fixed algorithm used. Instead of comparing from start time of the system, we
  just check on a given quantum of time that we didn't run too fast. It is more
  simple, there is no need to track idle/resume events, and a vcpu can sleep
  correctly.
- use "sysbench cpu run" in system mode to check execution is slowed as expected.
- do not use idle/resume callback

Alex Bennée (4):
  sysemu: add set_virtual_time to accel ops
  qtest: use cpu interface in qtest_clock_warp
  sysemu: generalise qtest_warp_clock as qemu_clock_advance_virtual_time
  plugins: add time control API

Pierrick Bouvier (2):
  qtest: move qtest_{get,set}_virtual_clock to accel/qtest/qtest.c
  contrib/plugins: add ips plugin example for cost modeling

 include/qemu/qemu-plugin.h                    |  23 +++
 include/qemu/timer.h                          |  15 ++
 include/sysemu/accel-ops.h                    |  18 +-
 include/sysemu/cpu-timers.h                   |   3 +-
 include/sysemu/qtest.h                        |   2 -
 accel/qtest/qtest.c                           |  13 ++
 contrib/plugins/ips.c                         | 164 ++++++++++++++++++
 plugins/api.c                                 |  31 ++++
 ...t-virtual-clock.c => cpus-virtual-clock.c} |   5 +
 system/cpus.c                                 |  11 ++
 system/qtest.c                                |  37 +---
 util/qemu-timer.c                             |  26 +++
 contrib/plugins/Makefile                      |   1 +
 plugins/qemu-plugins.symbols                  |   2 +
 stubs/meson.build                             |   2 +-
 15 files changed, 314 insertions(+), 39 deletions(-)
 create mode 100644 contrib/plugins/ips.c
 rename stubs/{cpus-get-virtual-clock.c => cpus-virtual-clock.c} (68%)

Comments

Alex Bennée May 31, 2024, 4:27 p.m. UTC | #1
Pierrick Bouvier <pierrick.bouvier@linaro.org> writes:

> The goal here is to be able to scale temporally execution of qemu-user/system,
> using a given number of instructions per second.
>
> We define a virtual clock, that can be late or in advance compared to real time.
> When we are in advance, we slow execution (by sleeping) until catching real
> time.
>
> Finally, we should be able to cleanup icount=auto mode completely, and keep
> icount usage for determistic purposes only.
>
> It is built upon new TCG Plugins inline ops (store + conditional callbacks), now
> merged on master.
>
> Example in user-mode:
>
> - Retrieve number of instructions to execute /bin/true
> $ ./build/qemu-x86_64 -plugin ./build/tests/plugin/libinsn.so -d plugin /bin/true
> cpu 0 insns: 120546
> total insns: 120546
> - Slow execution to match 5 seconds
> $ time ./build/qemu-x86_64 -plugin ./build/contrib/plugins/libips,ips=$((120546/5)) /bin/true
> real    0m4.985s
>
> Tested in system mode by booting a full debian system, and using:
> $ sysbench cpu run
> Performance decrease linearly with the given number of ips.

Queued to plugins/next, thanks.