<div dir="ltr">One of the downsides of using the gdb protocol is that these packets are stateful meaning the thread id option is not there and the word btrace stands for branch trace which basically suggests execution tracing.</div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Apr 11, 2016 at 4:50 PM, Pavel Labath <span dir="ltr"><<a href="mailto:labath@google.com" target="_blank">labath@google.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">I think we should reuse packets from the gdb protocol whereever it<br>
makes sense. So, if they fit your needs (and a quick glance seems to<br>
confirm that), then I think you should use them.<br>
<div class="HOEnZb"><div class="h5"><br>
On 11 April 2016 at 15:28, Ravitheja Addepally <<a href="mailto:ravithejawork@gmail.com">ravithejawork@gmail.com</a>> wrote:<br>
> Hello,<br>
> Regarding the packet definitions for tracing, how about reusing the<br>
> existing btrace packets ?<br>
><br>
> <a href="https://sourceware.org/gdb/onlinedocs/gdb/General-Query-Packets.html#qXfer%20btrace%20read" rel="noreferrer" target="_blank">https://sourceware.org/gdb/onlinedocs/gdb/General-Query-Packets.html#qXfer%20btrace%20read</a><br>
><br>
> On Fri, Apr 1, 2016 at 7:13 PM, Greg Clayton <<a href="mailto:gclayton@apple.com">gclayton@apple.com</a>> wrote:<br>
>><br>
>> We also need to think about all other types of tracing. It might make more<br>
>> sense to keep these calls on SBProcess and have the calls simply be:<br>
>><br>
>><br>
>> lldb::SBTrace lldb::SBProcess::StartTrace(SBTraceOptions &trace_options,<br>
>> lldb::SBError &error);<br>
>><br>
>> And you would need to specify which threads in the SBTraceOptions object<br>
>> itself?:<br>
>><br>
>> SBTraceOptions trace_options;<br>
>><br>
>> And then do some like:<br>
>><br>
>> trace_options.SetTraceAllThreads();<br>
>><br>
>> And maybe tracing all threads is the default. Or one can limit this to one<br>
>> thread:<br>
>><br>
>> trace_options.SetThreadID (tid);<br>
>><br>
>> Then you start tracing using the "trace_options" which has the notion of<br>
>> which threads to trace.<br>
>><br>
>> lldb::SBError error;<br>
>> lldb::SBTrace trace = process.StartTrace(trace_options, error);<br>
>><br>
>> It really depends on how all different forms of trace are enabled for<br>
>> different kinds of tracing. It takes kernel support to trace only specific<br>
>> threads, but someone might be debugging a bare board that only has the<br>
>> option tracing all threads on each core. When making an API we can't assume<br>
>> we can limit this to any threads and even to any process.<br>
>><br>
>> Greg<br>
>><br>
>> > On Apr 1, 2016, at 2:00 AM, Pavel Labath <<a href="mailto:labath@google.com">labath@google.com</a>> wrote:<br>
>> ><br>
>> > I second Greg's suggestions, and I have some thoughts of my own:<br>
>> ><br>
>> > - with the proposed API, it is not possible to get a trace for newly<br>
>> > created threads - the process needs to be stopped first, so you can<br>
>> > enable trace collection. There certainly are cases where this could be<br>
>> > problematic, e.g. if you get a crash early during thread creation and<br>
>> > you want to figure out how you got there. For this to work, we might<br>
>> > need an API like<br>
>> > SBProcess::TraceNewThreads(...)<br>
>> > or<br>
>> > SBProcess::TraceAllThreads(...)<br>
>> > with the assumption that "all" also includes newly created threads in<br>
>> > the future.<br>
>> ><br>
>> > I'm not saying this needs to be done in the first implementation, but<br>
>> > I think that we should at least design the API in a way that will not<br>
>> > make adding this unnecessarily hard in the future (e.g. the idea of<br>
>> > returning an SBTrace object might be problematic, since you don't know<br>
>> > if/how many threads will be created).<br>
>> ><br>
>> ><br>
>> ><br>
>> > Also, thinking about new APIs, should we have a way to mark an API as<br>
>> > incubating/experimental? Maybe it would be good to mark these new APIs<br>
>> > as experimental for a while, so we have an option of changing them in<br>
>> > the future, if it turns out we have made the wrong decision. I was<br>
>> > thinking of either a naming convention<br>
>> > (SBThread::StartTraceExperimental) or some annotation/comment on the<br>
>> > methods. When we are confident this design is good, we can remove the<br>
>> > promote the api into the "stable" set.<br>
>> ><br>
>> > pl<br>
>> ><br>
>> ><br>
>> ><br>
>> ><br>
>> > On 31 March 2016 at 18:59, Greg Clayton via lldb-dev<br>
>> > <<a href="mailto:lldb-dev@lists.llvm.org">lldb-dev@lists.llvm.org</a>> wrote:<br>
>> >><br>
>> >>> On Mar 31, 2016, at 5:10 AM, Ravitheja Addepally via lldb-dev<br>
>> >>> <<a href="mailto:lldb-dev@lists.llvm.org">lldb-dev@lists.llvm.org</a>> wrote:<br>
>> >>><br>
>> >>> Hello all,<br>
>> >>> I am currently working on enabling Intel (R) Processor<br>
>> >>> Trace collection for lldb. I have done some previous discussions in this<br>
>> >>> mailing list on this topic but just to summarize , the path we chose was to<br>
>> >>> implement raw trace collection in lldb and the trace will be decoded outside<br>
>> >>> LLDB. I wanted to expose this feature through the SB API's and for trace<br>
>> >>> data transfer I wish to develop new communication packets. Now I want to get<br>
>> >>> the new API's and packet specifications reviewed by the dev list. Please<br>
>> >>> find the specification below -><br>
>> >>><br>
>> >>> lldb::SBError SBProcess::StartTrace(lldb::tid_t threadId, const<br>
>> >>> SBTraceConfig &config)<br>
>> >>> Start tracing for thread - threadId with trace configuration<br>
>> >>> config.<br>
>> >>> SBTraceConfig would contain the following fields-<br>
>> >>> -> TraceType - ProcessorTrace, SoftwareTrace , any trace<br>
>> >>> technology etc<br>
>> >>> -> size of trace buffer<br>
>> >>> -> size of meta data buffer<br>
>> >>> Returns error in case starting trace was unsuccessful, which could<br>
>> >>> occur by reasons such as<br>
>> >>> picking non existent thread, target does not support TraceType<br>
>> >>> selected etc.<br>
>> >><br>
>> >> If you are going to trace on a thread, we should be putting this API on<br>
>> >> SBThread. Also we have other config type classes in our public API and we<br>
>> >> have suffixed them with Options so SBTraceConfig should actually be<br>
>> >> SBTraceOptions. Also don't bother using "const" on any public APIs since the<br>
>> >> mean nothing and only cause issues. Why? All public classes usually contain<br>
>> >> a std::unique_ptr or a std::shared_ptr to a private class that exists only<br>
>> >> within LLDB itself. The "const" is just saying don't change my shared<br>
>> >> pointer, which doesn't actually do anything.<br>
>> >><br>
>> >> lldb::SBError SBThread::StartTrace(SBTraceOptions &trace_options);<br>
>> >><br>
>> >>><br>
>> >>> lldb::SBError SBProcess::StopTrace(lldb::tid_t threadId)<br>
>> >>> Stop tracing for thread - threadId. Tracing should be enabled<br>
>> >>> already for thread, else error is returned.<br>
>> >><br>
>> >> This should be:<br>
>> >><br>
>> >> lldb::SBError SBThread::StopTrace();<br>
>> >><br>
>> >> One question: can there only be one kind of trace going on at the same<br>
>> >> time? If we ever desire to support more than one at a time, we might need<br>
>> >> the above two calls to be:<br>
>> >><br>
>> >><br>
>> >> lldb::user_id_t SBThread::StartTrace(SBTraceOptions &trace_options,<br>
>> >> lldb::SBError &error);<br>
>> >> lldb::SBError SBThread::StopTrace(lldb::user_id_t trace_id);<br>
>> >><br>
>> >> The StartTrace could return a unique trace token that would need to be<br>
>> >> supplied back to any other trace calls like the ones below.<br>
>> >><br>
>> >>> size_t SBProcess::DumpTraceData(lldb::tid_t threadId, void *buf,<br>
>> >>> size_t size, SBError &sberror)<br>
>> >>> Dump the raw trace data for threadId in buffer described by pointer<br>
>> >>> buf and size. Tracing should be enabled already for thread else error<br>
>> >>> is sent in sberror. The actual size of filled buffer is returned by<br>
>> >>> API.<br>
>> >>><br>
>> >>> size_t SBProcess::DumpTraceMetaData(lldb::tid_t threadId, void *buf,<br>
>> >>> size_t size, SBError &sberror)<br>
>> >>> Dump the raw trace meta data for threadId in buffer described by<br>
>> >>> pointer buf and size. Tracing should be enabled already for thread<br>
>> >>> else error is sent in sberror. The actual size of filled buffer is<br>
>> >>> returned by API.<br>
>> >><br>
>> >> These would be on lldb::SBThread and remove the lldb::tid_t parameter,<br>
>> >> possibly adding "lldb::user_id_t trace_id" as the first parameter.<br>
>> >><br>
>> >> The other way to do this is to create a lldb::SBTrace object. Then the<br>
>> >> APIs become:<br>
>> >><br>
>> >> lldb::SBTrace SBThread::StartTrace(SBTraceOptions &trace_options,<br>
>> >> lldb::SBError &error);<br>
>> >><br>
>> >> lldb::SBError SBTrace::StopTrace();<br>
>> >> size_t SBTrace::GetData(void *buf, size_t size, SBError &sberror);<br>
>> >> size_t SBTrace::GetMetaData(void *buf, size_t size, SBError &sberror);<br>
>> >> lldb::SBThread SBTrace::GetThread();<br>
>> >><br>
>> >>><br>
>> >>> LLDB Trace Packet Specification<br>
>> >>><br>
>> >>> QTrace:1:<threadid>,<type>,<buffersize>,<metabuffersize><br>
>> >>> Packet for starting tracing, where -<br>
>> >>> -> threadid - stands for thread to trace<br>
>> >>> -> type - Type of tracing to use, it will be like type of<br>
>> >>> trace mechanism to use.<br>
>> >>> For e.g ProcessorTrace, SoftwareTrace , any trace<br>
>> >>> technology etc and if<br>
>> >>> that trace is not supported by target error will be<br>
>> >>> returned. In Future<br>
>> >>> we can also add more parameters in the packet<br>
>> >>> specification, which can be type specific<br>
>> >>> and the server can parse them based on what type it<br>
>> >>> read in the beginning.<br>
>> >>> -> buffersize - Size for trace buffer<br>
>> >>> -> metabuffersize - Size of Meta Data<br>
>> >><br>
>> >> If we design this, we should have the arguments be in key/value format:<br>
>> >><br>
>> >>> QTrace:1:<key>:<value>;<key>:<value>;<key>:<value>;<br>
>> >><br>
>> >> Then this packet currently could be sent as:<br>
>> >><br>
>> >><br>
>> >> QTrace:1:threadid:<threadid>;type:<type>;buffersize=<buffersize>;metabuffersize=<metabuffersize>;<br>
>> >><br>
>> >> This way if we ever need to add new key value pairs, we don't need to<br>
>> >> make a new QTrace2 packet if the args ever change.<br>
>> >><br>
>> >><br>
>> >>> QTrace:0:<threadid><br>
>> >>> Stop tracing thread with threadid,{Trace needs to be started<br>
>> >>> of-course else error}<br>
>> >><br>
>> >> again, this should be key/value pair encoded<br>
>> >><br>
>> >> QTrace:0:threadid:<threadid>;<br>
>> >>><br>
>> >>> qXfer:trace:buffer:read:annex:<threadid>,<byte_count><br>
>> >>> Packet for reading the trace buffer<br>
>> >>> -> threadid - thread ID, of-course if tracing is not<br>
>> >>> started for this thread error will be returned.<br>
>> >>> -> byte_count - number of bytes to read, in case trace captured<br>
>> >>> is<br>
>> >>> less than byte_count, then only that much trace<br>
>> >>> will<br>
>> >>> be returned in response packet.<br>
>> >>><br>
>> >>> qXfer:trace:meta:read:annex:<threadid>,<byte_count><br>
>> >>> Similar Packet as above except it reads meta data<br>
>> >><br>
>> >> Hopefully we can key/value pair encode the args text that is<br>
>> >> "<threadid>,<byte_count>".<br>
>> >><br>
>> >><br>
>> >><br>
>> >> _______________________________________________<br>
>> >> lldb-dev mailing list<br>
>> >> <a href="mailto:lldb-dev@lists.llvm.org">lldb-dev@lists.llvm.org</a><br>
>> >> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/lldb-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/lldb-dev</a><br>
>><br>
><br>
</div></div></blockquote></div><br></div>