On Wed, Sep 28, 2016 at 02:08:03PM +0100, Daniel P. Berrange wrote: > These patches were previously posted as part of my giant > trace events modular build series > > v1: https://lists.gnu.org/archive/html/qemu-devel/2016-08/msg01714.html > v2: https://lists.gnu.org/archive/html/qemu-devel/2016-09/msg03335.html > v3: https://lists.gnu.org/archive/html/qemu-devel/2016-09/msg04282.html > v4: https://lists.gnu.org/archive/html/qemu-devel/2016-09/msg05467.html > > This series does all the refactoring required to support a fully > modular build of the trace probe points, but does not actually > convert anything to use it. The 40+ patches to convert each subdir > to use modular build will only be posted again, once the refactoring > is reviewed & queued, in order to avoid patch-bombing the list > more than is needed. The full series is visible at > > https://github.com/berrange/qemu/tree/trace-events-3 > > The key problem being tackled by this series is the assumption > that there is a single statically declared enum which provides > globally unique event IDs. Inside QEMU the event IDs were used > as list indexes into the 'trace_events' array, while the event > VCPU IDs were used as bitmap indexes in trace_dstate against > the CPUState struct. Externally to QEMU, the event IDs were > also written in the simpletrace binary data format and used > to lookup the entry in the trace-events file afterwards. > > Inside QEMU the refactoring work managed to remove all need > for event IDs for purposes of 'trace_events' array lookups. > Instead we now have global variables per-event which can be > referenced directly. > > When QEMU starts up and the various event groups are registered, > we now dynamically assign event IDs and VCPU IDs to each event. > This removes the limitation in the v1 posting that all vCPU > events had to be in one file. We also removed the limitation > on the total number of vCPU events. So there is no regression > in functionality of VCPU event support compared to current > GIT master. > Since the event IDs are allocated dynamically at runtime, > the simpletrace.py script cannot assume they map directly > to the 'trace-events' file entries. Thus, the simpletrace > binary format is extended to include a record type that > maps trace event IDs to trace event names. While it would > be possible to take this even further and make the > simpletrace binary format 100% self-describing this is left > as an exercise for future developers, as it is not a > pre-requisite for this modular build. > > While some of the intermediate patches may seem pointless > on their own, they exist in order to facilitate the review > of later patches by ensuring each patch does the minimum > possible refactoring work. > > Changed in v5: > > - Use single '_' instead of double/triple '_' in > constants (Lluís) > - Use more pythonic loop iterator (Lluís) > - Misc typos (Lluís/Stefan) > - Fix filtering of QMP trace events (Lluís) > - Fix some new mistakes in trace-events in master
Great, this looks very close. There are only a few comments from Lluís and me. Stefan
signature.asc
Description: PGP signature
