Re: [PATCH 06/23] perf doc: Add perf data file documentation

From: Andi Kleen
Date: Fri Jul 19 2013 - 11:46:54 EST

Next message: Naoya Horiguchi: "Re: [PATCH v3 0/8] extend hugepage migration"
Previous message: Greg Kroah-Hartman: "Re: [PATCH] lustre:libcfs: remove redundant code."
In reply to: Jiri Olsa: "Re: [PATCH 06/23] perf doc: Add perf data file documentation"
Next in thread: Jiri Olsa: "[PATCH 07/23] perf tools: Recognize version number for perf data file"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Wed, Jul 17, 2013 at 07:49:46PM +0200, Jiri Olsa wrote:
> Adding perf data file documentation.

FWIW I wrote the following python "construct" description of the data format
some time ago. It's more formal than yours. There were still a few things
missing though and one mysterious bug.

# A description of the perf.data file format in "construct"
#
from construct import *

def fork_exit(name):
return Struct(name,
SNInt32("pid"),
SNInt32("ppid"),
SNInt32("tid"),
SNInt32("ptid"),
UNInt64("time"))

def throttle(name):
return Struct(name,
UNInt64("time"),
UNInt64("id"),
UNInt64("stream_id"))

def event(sample_type, read_format):
return Embedded(
Struct("event",
If(lambda ctx: sample_type.ip,
UNInt64("ip")),
If(lambda ctx: sample_type.tid,
Struct("tid",
SNInt64("pid"),
SNInt64("tid"))),
If(lambda ctx: sample_type.time,
UNInt64("time")),
If(lambda ctx: sample_type.addr,
UNInt64("addr")),
If(lambda ctx: sample_type.stream_id,
UNInt64("stream_id")),
If(lambda ctx: sample_type.cpu,
Struct("cpu",
UNInt32("cpu"),
UNInt32("res"))),
If(lambda ctx: sample_type.period,
UNInt64("period")),
# XXX
#If(lambda ctx: sample_type.read,
# read_format(rformat)),
If(lambda ctx: sample_type.callchain,
Struct("callchain",
UNInt64("nr"),
Array(lambda ctx: ctx.nr,
UNInt64("addr")))),
If(lambda ctx: sample_type.raw,
Struct("raw",
UNInt32("size"),
Bytes("raw", lambda ctx: ctx.size))),
If(lambda ctx: sample_type.branch_stack,
Struct("branch_stack",
UNInt64("nr"),
Array(lambda ctx: ctx.nr,
Struct(None,
UNInt64("from"),
UNInt64("to"),
UNInt64("flags"))))), # XXX split
# both need parameters passed in
# sample reg
# stack user
If(lambda ctx: sample_type.weight,
UNInt64("weight")),
If(lambda ctx: sample_type.data_src,
UNInt64("data_src")),
Anchor("end_event"),
Padding(lambda ctx: max(0, ctx.size - ctx.end_event))))

# XXX need to make OnDemand for large files

def perf_event(sample_type, read_format):
return Struct("perf_event",
Anchor("start"),
Enum(UNInt32("type"),
MMAP = 1,
LOST = 2,
COMM = 3,
EXIT = 4,
THROTTLE = 5,
UNTHROTTLE = 6,
FORK = 7,
READ = 8,
SAMPLE = 9),
UNInt16("misc"),
UNInt16("size"),
Switch("data",
lambda ctx: ctx.type,
{
"MMAP": Struct("mmap",
SNInt32("pid"),
SNInt32("tid"),
UNInt64("addr"),
UNInt64("len"),
UNInt64("pgoff"),
CString("filename")),
"LOST": Struct("lost",
UNInt64("id"),
UNInt64("lost")),
"COMM": Struct("comm",
SNInt32("pid"),
SNInt32("tid"),
CString("comm")),
"EXIT": fork_exit("exit"),
"THROTTLE": throttle("thottle"),
"UNTHROTTLE": throttle("unthottle"),
"FORK": fork_exit("fork"),
#"READ": read_format(read_format),
"SAMPLE": event(sample_type, read_format),
}),
Anchor("end"),
Padding(lambda ctx: ctx.size - (ctx.end - ctx.start))
)

def perf_event_seq(sample_type, read_format):
return GreedyRange(perf_event(sample_type, read_format))

perf_event_attr_sizes = (64, 72, 80, 96)

perf_event_attr = Struct("perf_event_attr",
Anchor("start"),
Enum(UNInt32("type"),
HARDWARE = 0,
SOFTWARE = 1,
TRACEPOINT = 2,
HW_CACHE = 3,
RAW = 4,
BREAKPOINT = 5),
UNInt32("size"),
UNInt64("config"),
UNInt64("sample_period_freq"),
BitStruct("sample_type",
Flag("ip"),
Flag("tid"),
Flag("time"),
Flag("addr"),
Flag("read"),
Flag("callchain"),
Flag("id"),
Flag("cpu"),
Flag("period"),
Flag("stream_id"),
Flag("raw"),
Flag("branch_stack"),
Flag("regs_user"),
Flag("stack_user"),
Flag("weight"),
Flag("data_src"),
Padding(64 - 16)),
BitStruct("read_format",
Flag("total_time_enabled"),
Flag("total_time_running"),
Flag("id"),
Flag("group"),
Padding(64 - 4)),
Embedded(BitStruct(None,
Flag("disabled"),
Flag("inherit"),
Flag("pinned"),
Flag("exclusive"),
Flag("exclude_user"),
Flag("exclude_kernel"),
Flag("exclude_hv"),
Flag("exclude_idle"),
Flag("mmap"),
Flag("comm"),
Flag("freq"),
Flag("inherit_stat"),
Flag("enable_on_exec"),
Flag("task"),
Flag("watermark"),
BitField("precise_ip", 2),
Flag("mmap_data"),
Flag("sample_id_all"),
Flag("exclude_host"),
Flag("exclude_guest"),
Flag("exclude_callchain_kernel"),
Flag("exclude_callchain_user"),
Padding(41))),
UNInt32("wakeup_events"),
UNInt32("bp_type"),
UNInt64("config1"),
If(lambda ctx: ctx.size >= perf_event_attr_sizes[1],
UNInt64("config2")),
If(lambda ctx: ctx.size >= perf_event_attr_sizes[2],
UNInt64("branch_sample_type")),
If(lambda ctx: ctx.size >= perf_event_attr_sizes[3],
Embedded(Struct(None,
UNInt64("sample_regs_user"),
UNInt32("sample_stack_user")))),
Anchor("end"),
Value("perf_event_attr_size", lambda ctx: ctx.end - ctx.start),
Padding(lambda ctx: ctx.size - ctx.perf_event_attr_size))

# assumes all attributes are the same size

perf_file_attr = Struct("perf_file_attr",
Peek(Embedded(Struct(None, UNInt32("type"), UNInt32("size")))),
Array(lambda ctx: ctx._.size / ctx.size, perf_event_attr))

perf_event_types = Struct("perf_file_attr",
Anchor("here"),
Padding(lambda ctx: ctx._.size))

perf_data = OnDemand(Bytes("perf_data", lambda ctx: ctx.size))

def perf_file_section(name, target):
return Struct(name,
UNInt64("offset"),
UNInt64("size"),
Pointer(lambda ctx: ctx.offset, target))

perf_file = Struct("perf_file_header",
UNInt64("magic"), # XXX
UNInt64("size"),
UNInt64("attr_size"),
perf_file_section("attrs", perf_file_attr),
perf_file_section("data", perf_data),
perf_file_section("event_types", perf_event_types),
BitStruct("adds_features",
Flag("tracing_data"),
Flag("build_id"),
Flag("hostname"),
Flag("osrelease"),
Flag("version"),
Flag("arch"),
Flag("nrcpus"),
Flag("cpudesc"),
Flag("cpuid"),
Flag("total_mem"),
Flag("cmdline"),
Flag("event_desc"),
Flag("cpu_topology"),
Flag("numa_topology"),
Flag("branch_stack"),
Flag("pmu_mappings"),
Flag("group_desc"),
Padding(64 - 17)),
Padding(3 * 8))

def get_events(h):
data = h.data.perf_data.value
# assumes event 0 attributes applies to all samples?
# XXX
ev0 = h.attrs.perf_file_attr.perf_event_attr[0]
assert ev0.size in perf_event_attr_sizes
return perf_event_seq(ev0.sample_type, ev0.read_format).parse(data)

if __name__ == '__main__':
import sys

with open(sys.argv[1], "rb") as f:
h = perf_file.parse_stream(f)
print h
print get_events(h)

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Next message: Naoya Horiguchi: "Re: [PATCH v3 0/8] extend hugepage migration"
Previous message: Greg Kroah-Hartman: "Re: [PATCH] lustre:libcfs: remove redundant code."
In reply to: Jiri Olsa: "Re: [PATCH 06/23] perf doc: Add perf data file documentation"
Next in thread: Jiri Olsa: "[PATCH 07/23] perf tools: Recognize version number for perf data file"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]