Hi,
This RFC does two principle things. It continues the push to reduce
the BQL lock contention to a minimum by dropping the BQL for most of
the run loop - eventually only holding it for timer processing and
sleeping. The second part is an attempt to fix the breakage caused to
record/replay by the previous work reducing the BQL in 2.9.
The patch breakdown is as follows:
- target/arm/arm-powertctl: drop BQL assertions
This just fixes a bogus assert. The async work doesn't need BQL
protection as it is all done in the context of the vCPU (meaning
nothing else should be messing with it). However going forward we will
need to audit all the async work calls to make sure they are fine to
run outside of the BQL.
- cpus: push BQL lock to qemu_*_wait_io_event
- cpus: only take BQL for sleeping threads
The BQL reduction is done in two stages, mainly to make bisection
easier.
- replay/replay-internal.c: track holding of replay_lock
- replay: make locking visible outside replay code
- replay: push replay_mutex_lock up the call tree
There is still more work to do here but essentially replay_lock now
replaces the BQL in keeping synchronisation between main-loop and the
TCG thread - it is no longer a fine lock for the replay log but a
gross lock that keeps batches of updates together. So far I have been
testing with the simple testcase:
./arm-softmmu/qemu-system-arm -machine type=vexpress-a9 -m 1024 \
-display none -smp 1 -kernel ../images/vexpress-kernel.img \
-dtb ../images/vexpress-v2p-ca9.dtb \
-append "console=ttyAMA0" -serial mon:stdio \
-icount shift=7,rr=record,rrfile=replay.bin
..and that works fine. However more complex testing with more devices
to exercise the async events code is needed here. One change that
needs some careful care is we are now not dropping the replay_mutex
lock between events.
And finally:
- scripts/qemu-gdb: add simple tcg lock status helper
- util/qemu-thread-*: add qemu_lock, locked and unlock trace events
- scripts/analyse-locks-simpletrace.py: script to analyse lock times
These are helpers I wrote while debugging the locking. The GDB helper
is fairly dump but does attempt to show where locks are held. The
qemu-lock tracing is a little hacky but could prove useful in doing
more detailed lock analysis.
Any thoughts/comments? Does this seem like a reasonable direction to
go?
Alex Bennée (9):
target/arm/arm-powertctl: drop BQL assertions
cpus: push BQL lock to qemu_*_wait_io_event
cpus: only take BQL for sleeping threads
replay/replay-internal.c: track holding of replay_lock
replay: make locking visible outside replay code
replay: push replay_mutex_lock up the call tree
scripts/qemu-gdb: add simple tcg lock status helper
util/qemu-thread-*: add qemu_lock, locked and unlock trace events
scripts/analyse-locks-simpletrace.py: script to analyse lock times
cpus-common.c | 13 ++---
cpus.c | 57 ++++++++++++++++++---
docs/replay.txt | 19 +++++++
include/qemu/thread.h | 7 ++-
include/sysemu/replay.h | 16 ++++++
kvm-all.c | 4 --
replay/replay-char.c | 21 +++-----
replay/replay-events.c | 18 ++-----
replay/replay-internal.c | 22 +++++++-
replay/replay-internal.h | 5 +-
replay/replay-time.c | 10 ++--
replay/replay.c | 40 +++++++--------
scripts/analyse-locks-simpletrace.py | 99 ++++++++++++++++++++++++++++++++++++
scripts/qemu-gdb.py | 3 +-
scripts/qemugdb/tcg.py | 46 +++++++++++++++++
stubs/replay.c | 15 ++++++
target/arm/arm-powerctl.c | 8 ---
target/i386/hax-all.c | 2 -
util/main-loop.c | 23 +++++++--
util/qemu-thread-posix.c | 11 +++-
util/trace-events | 5 ++
vl.c | 2 +
22 files changed, 353 insertions(+), 93 deletions(-)
create mode 100755 scripts/analyse-locks-simpletrace.py
create mode 100644 scripts/qemugdb/tcg.py
--
2.11.0
