introduce TRACE_EVENT_ABI (was Re: TRACE_EVENT_ABI ?)
I'm not sure we can support any as an ABI yet. The text format seems to volatile in general - not just the output of the individual trace events but also the common file format, and for the binary format we need to
figure a good way to tag the output yet. Also when we define one as
one ABI we should make very clear what that means, e.g. does it have to stay exactly as is? Or can we add new fields but not remove old one?
Quoted 6 lines · Christoph Hellwig
I'm not sure we can support any as an ABI yet. The text format seems to volatile in general - not just the output of the individual trace events but also the common file format, and for the binary format we need to
figure a good way to tag the output yet. Also when we define one as
one ABI we should make very clear what that means, e.g. does it have to stay exactly as is? Or can we add new fields but not remove old one?
I had this discussion with people in Portland. We seem to agree that this should just lock the old fields in, but you can add new ones at the end.
-- Steve
Quoted 3 lines · Steven Rostedt
I had this discussion with people in Portland. We seem to agree that this should just lock the old fields in, but you can add new ones at the end.
Please document this very clearly.
Quoted 10 lines · Steven Rostedt
Quoted 7 lines · Christoph Hellwig
I'm not sure we can support any as an ABI yet. The text format seems to volatile in general - not just the output of the individual trace events but also the common file format, and for the binary format we need to figure a good way to tag the output yet. Also when we define one as one ABI we should make very clear what that means, e.g. does it have to stay exactly as is? Or can we add new fields but not remove old one?
Quoted 3 lines
I had this discussion with people in Portland. We seem to agree that this should just lock the old fields in, but you can add new ones at the end.
Yeah, that's the sanest approach i was thinking about when i suggested TRACE_EVENT_ABI() to Arjan and Peter.
The raw record is opaque, comes with a length field and goes into the ring-buffer so it's nicely extensible. Existing bits shouldnt change.
User-space that relies on a record can define a structure of that and copy that over from the ring-buffer - and ignore any new bits.
What i'd also like to see is the use of typical ABI-safe type fields in the trace definitions themselves: u8, u32, u64, etc. 'long' is obviously not good. Could we do some automation for that perhaps? I.e. emit a warning (boot time or so) if TRACE_EVENT_ABI() is used with unsafe type fields.
( Endianness is another detail, if perf.data is shipped to a
different-endian system. Best is probably to define a new perf.data attribute extension with the endianness of the generator system included. That way the perf.data parser can convert endianness if it wants/needs to. )
Ingo
Quoted 32 lines · * Ingo Molnar ([email protected]) wrote:
Quoted 10 lines · Steven Rostedt
Quoted 7 lines · Christoph Hellwig
I'm not sure we can support any as an ABI yet. The text format seems to volatile in general - not just the output of the individual trace events but also the common file format, and for the binary format we need to figure a good way to tag the output yet. Also when we define one as one ABI we should make very clear what that means, e.g. does it have to stay exactly as is? Or can we add new fields but not remove old one?
Quoted 3 lines
I had this discussion with people in Portland. We seem to agree that this should just lock the old fields in, but you can add new ones at the end.
Yeah, that's the sanest approach i was thinking about when i suggested
TRACE_EVENT_ABI()to Arjan and Peter.The raw record is opaque, comes with a length field and goes into the ring-buffer so it's nicely extensible. Existing bits shouldnt change.
User-space that relies on a record can define a structure of that and copy that over from the ring-buffer - and ignore any new bits.
What i'd also like to see is the use of typical ABI-safe type fields in the trace definitions themselves: u8, u32, u64, etc. 'long' is obviously not good. Could we do some automation for that perhaps? I.e. emit a warning (boot time or so) if
TRACE_EVENT_ABI()is used with unsafe type fields.
Quoted 6 lines
( Endianness is another detail, if perf.data is shipped to a
different-endian system. Best is probably to define a new perf.data attribute extension with the endianness of the generator system included. That way the perf.data parser can convert endianness if it wants/needs to. )
You can ship sizeof(long) along with the endianness information to access "long" typed variables from the parser too. If we export this information, then I don't see how exporting a 'long' variable would be obviously not good. It would become as portable as u32 or u64, yet more efficient when only the space for a 'long' record is required. Same applies to pointers, obviously.
All we need is a small "accessor" library to deal with the type size and endianness issues.
Mathieu
Quoted 1 line
Ingo
Signature · 2 lines
Mathieu Desnoyers OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68