Optimizes tracing of events with the 'tcg' and 'vcpu' properties (e.g., memory
accesses), making it feasible to statically enable them by default on all QEMU
builds.
Some quick'n'dirty numbers with 400.perlbench (SPECcpu2006) on the train input
(medium size - suns.pl) and the guest_mem_before event:
* vanilla, statically disabled
real 0m2,259s
user 0m2,252s
sys 0m0,004s
* vanilla, statically enabled (overhead: 2.18x)
real 0m4,921s
user 0m4,912s
sys 0m0,008s
* multi-tb, statically disabled (overhead: 0.99x) [within noise range]
real 0m2,228s
user 0m2,216s
sys 0m0,008s
* multi-tb, statically enabled (overhead: 0.99x) [within noise range]
real 0m2,229s
user 0m2,224s
sys 0m0,004s
Right now, events with the 'tcg' property always generate TCG code to trace that
event at guest code execution time, where the event's dynamic state is checked.
This series adds a performance optimization where TCG code for events with the
'tcg' and 'vcpu' properties is not generated if the event is dynamically
disabled. This optimization raises two issues:
* An event can be dynamically disabled/enabled after the corresponding TCG code
has been generated (i.e., a new TB with the corresponding code should be
used).
* Each vCPU can have a different dynamic state for the same event (i.e., tracing
the memory accesses of only one process pinned to a vCPU).
To handle both issues, this series integrates the dynamic tracing event state
into the TB hashing function, so that vCPUs tracing different events will use
separate TBs. Note that only events with the 'vcpu' property are used for
hashing (as stored in the bitmap of CPUState->trace_dstate).
This makes dynamic event state changes on vCPUs very efficient, since they can
use TBs produced by other vCPUs while on the same event state combination (or
produced by the same vCPU, earlier).
Discarded alternatives:
* Emitting TCG code to check if an event needs tracing, where we should still
move the tracing call code to either a cold path (making tracing performance
worse), or leave it inlined (making non-tracing performance worse).
* Eliding TCG code only when *zero* vCPUs are tracing an event, since enabling
it on a single vCPU will impact the performance of all other vCPUs that are
not tracing that event.
Signed-off-by: Lluís Vilanova <address@hidden>
---
Changes in v9
=============
* Rebase on 931892e8a6.
* Undo renaming of tb->trace_vcpu_dstate to the shorter tb->trace_ds.
* Add measurements to commit enabling all guest events.
Changes in v8
=============
[Emilio G. Cota]
* Ported to current dev tree.
* Allocate cpu->trace_dstate statically. This
* allows us to drop the event_count inline patch.
* simplifies and improves the performance of accessing cpu->trace_dstate:
we just need to dereference, instead of going through bitmap_copy and
an intermediate unsigned long.
* If we try to register more CPU events than the max we support (there's a
constant for it), drop the event and tell the user with error_report. But
really this is a bug, since we control what CPU events are traceable. Should
we abort() as well?
* Added rth's R-b tag
* Addressed my own comments:
* rename tb->trace_vcpu_dstate to the shorter tb->trace_ds
* use uint32_t for tb->trace_ds instead of a typedef
* add BUILD_BUG_ON check to make sure tb->trace_ds is big enough
* fix xxhash
* Do not add trace_dstate to tb_htable_lookup, since we can grab it from
cpu->trace_dstate.
This patchset applies cleanly on top of rth's tcg-next (a01792e1e).
Changes in v7
=============
* Fix delayed dstate changes (now uses async_run_on_cpu() as suggested by Paolo
Bonzini).
* Note to Richard: patch 4 has been adapted to the new patch 3 async callback,
but is essentially the same.
Changes in v6
=============
* Check hashing size error with QEMU_BUILD_BUG_ON [Richard Henderson].
Changes in v5
=============
* Move define into "qemu-common.h" to allow compilation of tests.
Changes in v4
=============
* Incorporate trace_dstate into the TB hashing function instead of using
multiple physical TB caches [suggested by Richard Henderson].
Changes in v3
=============
* Rebase on 0737f32daf.
* Do not use reserved symbol prefixes ("__") [Stefan Hajnoczi].
* Refactor trace_get_vcpu_event_count() to be inlinable.
* Optimize cpu_tb_cache_set_requested() (hottest path).
Changes in v2
=============
* Fix bitmap copy in cpu_tb_cache_set_apply().
* Split generated code re-alignment into a separate patch [Daniel P. Berrange].
Lluís Vilanova (7):
exec: [tcg] Refactor flush of per-CPU virtual TB cache
trace: Allocate cpu->trace_dstate in place
trace: [tcg] Delay changes to dynamic state when translating
exec: [tcg] Use different TBs according to the vCPU's dynamic tracing
state
trace: [tcg] Do not generate TCG code to trace dinamically-disabled events
trace: [tcg,trivial] Re-align generated code
trace: [trivial] Statically enable all guest events
accel/tcg/cpu-exec.c | 8 ++++++--
accel/tcg/cputlb.c | 2 +-
accel/tcg/translate-all.c | 26 +++++++++++++++++++-------
include/exec/exec-all.h | 12 ++++++++++++
include/exec/tb-hash-xx.h | 7 +++++--
include/exec/tb-hash.h | 5 +++--
include/qom/cpu.h | 12 ++++++------
qom/cpu.c | 8 --------
scripts/tracetool/__init__.py | 3 ++-
scripts/tracetool/backend/dtrace.py | 4 ++--
scripts/tracetool/backend/ftrace.py | 20 ++++++++++----------
scripts/tracetool/backend/log.py | 19 ++++++++++---------
scripts/tracetool/backend/simple.py | 4 ++--
scripts/tracetool/backend/syslog.py | 6 +++---
scripts/tracetool/backend/ust.py | 4 ++--
scripts/tracetool/format/h.py | 26 +++++++++++++++++++-------
scripts/tracetool/format/tcg_h.py | 21 +++++++++++++++++----
scripts/tracetool/format/tcg_helper_c.py | 5 +++--
tcg/tcg-runtime.c | 3 ++-
tests/qht-bench.c | 2 +-
trace-events | 6 +++---
trace/control-target.c | 21 ++++++++++++++++++---
trace/control.c | 9 ++++++++-
trace/control.h | 3 +++
24 files changed, 157 insertions(+), 79 deletions(-)
To: qemu-devel@nongnu.org
Cc: Stefan Hajnoczi <stefa...@redhat.com>
Cc: Eduardo Habkost <ehabk...@redhat.com>
Cc: Eric Blake <ebl...@redhat.com>
Cc: Emilio G. Cota <c...@braap.org>
