Messages in this thread | | | Date | Sun, 11 Oct 2009 22:23:54 +0200 | From | Frederic Weisbecker <> | Subject | Re: perf trace errors |
| |
On Sun, Oct 11, 2009 at 02:40:30PM +0200, Christoph Hellwig wrote: > On Sun, Oct 11, 2009 at 02:19:28PM +0200, Frederic Weisbecker wrote: > > Yeah, there are still some trace events that are not well supported > > in perf trace, for now... I'll fix that by the time. > > Any idea what is the missing part? I tried to understand the code but > I haven't even managed to grasp the data structures yet. > > The weird thing is that these message seem to come from parsing of > the filters, and a simple trace the whole system trace shouldn't > really have to set any filters anyway, nor should the filters depdend > on what trace point you use. > > Note that the in addition to all the block trace points also all my > new xfs tracepoints also give errors from perf trace. Things tend > to work quite nicely with the ASCII output from the kernel. >
It's about the trace event format file parsing, especially the format line:
print fmt: "%d,%d %s %llu + %u [%s]", ((unsigned int) ((REC->dev) >> 20)), ((unsigned int) ((REC->dev) & ((1U << 20) - 1))), REC->rwbs, (unsigned long long)REC->sector, REC->nr_sector, REC->comm
This is done from tools/perf/util/trace-event-parse.c
It is supposed to handle such cases already, not sure exactly where is the problem currently, that needs some investigation.
| |