Messages in this thread |  | | Date | Thu, 12 Dec 2013 16:05:25 -0300 | From | Arnaldo Carvalho de Melo <> | Subject | Re: [PATCH v0 07/71] perf tools: Record whether a dso is 64-bit |
| |
Em Thu, Dec 12, 2013 at 09:45:26AM -0700, David Ahern escreveu: > On 12/12/13, 5:05 AM, Adrian Hunter wrote: > > >>>diff --git a/tools/perf/util/dso.h b/tools/perf/util/dso.h > >>>index 384f2d9..62680e1 100644 > >>>--- a/tools/perf/util/dso.h > >>>+++ b/tools/perf/util/dso.h > >>>@@ -91,6 +91,7 @@ struct dso { > >>> u8 annotate_warned:1; > >>> u8 sname_alloc:1; > >>> u8 lname_alloc:1; > >>>+ u8 is_64_bit:1; > >> > >>The is_64_bit name seems a bit hardcoded. We need something similar for > >>perf-trace to set the audit machine type for resolving syscalls. How about > >>having this field set a machine type rather than a "64-bit" flag? > > > >I am not sure what you mean by "machine type". For itrace the > >implementation only deals with its own architecture (e.g. the intel_pt > >pmu is only on Intel architecture) so it is not necessary to record > >the architecture. > > > >is_64_bit corresponds to ELFCLASS64 (vs ELFCLASS32) which is needed > >to determine whether the instruction set is 64-bit. That should > >work for other architectures too. > > > > perf-trace needs something similar -- an audit machine type to know > how to convert syscall numbers to functions. One of the following > per task: > > typedef enum { > MACH_X86=0, > MACH_86_64, > MACH_IA64, > MACH_PPC64, > MACH_PPC, > MACH_S390X, > MACH_S390, > MACH_ALPHA, > MACH_ARMEB > } machine_t; > > I was pondering how the 2 can be combined into a common flag.
Well, if we can pass somehow the magic number of an executable mmap in the PERF_RECORD_MMAP2 record, we would be able, together with the data we already have in the perf.data header (uname in a live session), to figure that out, no?
I.e. we wouldn't be limiting ourselves to the ELF executable format.
Time to read the kernel loader for the various formats we support...
- Arnaldo
|  |