Re: [PATCH] perf: script: add new output field 'dsoip'

From: Changbin Du
Date: Tue Apr 04 2023 - 02:54:43 EST

Next message: Sunil V L: "Re: [PATCH V3 00/20] Add basic ACPI support for RISC-V"
Previous message: yebin (H): "Re: [PATCH 2/2] lib/percpu_counter: fix dying cpu compare race"
In reply to: Adrian Hunter: "Re: [PATCH] perf: script: add new output field 'dsoip'"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Tue, Apr 04, 2023 at 09:38:40AM +0300, Adrian Hunter wrote:
> On 4/04/23 05:26, Changbin Du wrote:
> > On Mon, Apr 03, 2023 at 08:28:23AM +0300, Adrian Hunter wrote:
> >> On 31/03/23 11:44, Changbin Du wrote:
> >>> The goal of this change is to achieve consistent output for address. Before
> >>> this, the raw ip is printed for non-callchain and binary offset for
> >>> callchain. Mostly what we expect is the raw ip.
> >>>
> >>> This patch does two changes:
> >>> - Always print raw ip for symbols.
> >>> - Add a new 'dsoip' field if we really need the binary offset, and the
> >>
> >> IP stands for instruction pointer, which is always a memory address,
> >> so 'dsoip' does not seem like a good name. What about 'dso_offset'?
> >>
> > This name is inherited from the comments of map::map_ip. Actually now we have
> > three terms: dos ip, binary offset and dso offset. I also prefer to
> > 'dso_offset'.
> >
> >>> offset is appended to dso name.
> >>>
> >>> Default output now:
> >>> $ perf script
> >>> perf-exec 171459 5380927.621499: 1 cycles: ffffffff8c66e1b8 [unknown] ([unknown])
> >>> perf-exec 171459 5380927.621524: 1 cycles: ffffffff8c66e1b8 [unknown] ([unknown])
> >>> perf-exec 171459 5380927.621527: 2 cycles: ffffffff8c66e1b8 [unknown] ([unknown])
> >>> perf-exec 171459 5380927.621530: 17 cycles: ffffffff8c66e1b8 [unknown] ([unknown])
> >>> perf-exec 171459 5380927.621533: 194 cycles: ffffffff8c66e1bf [unknown] ([unknown])
> >>> perf-exec 171459 5380927.621536: 2153 cycles: ffffffff8c6375fe [unknown] ([unknown])
> >>> perf-exec 171459 5380927.621540: 23135 cycles: ffffffff8d001ac4 [unknown] ([unknown])
> >>> perf-exec 171459 5380927.621555: 224703 cycles: ffffffff8c7a5153 [unknown] ([unknown])
> >>> ls 171459 5380927.621686: 587397 cycles: ffffffff8c7c43de [unknown] ([unknown])
> >>> ls 171459 5380927.622016: 653693 cycles: 7fe1e47278c7 _dl_map_object_deps+0x307 (/lib/x86_64-linux-gnu/ld-2.27.so)
> >>> ls 171459 5380927.622382: 633896 cycles: 7fe1e41bd0d0 __wmemchr_ifunc+0x0 (/lib/x86_64-linux-gnu/libc-2.27.so)
> >>> ls 171459 5380927.622737: 608856 cycles: 7fe1e4194594 _int_malloc+0x264 (/lib/x86_64-linux-gnu/libc-2.27.so)
> >>> ls 171459 5380927.623078: 586376 cycles: 55ff5293d27c [unknown] (/bin/ls)
> >>> ls 171459 5380927.623405: 566812 cycles: 55ff5293b202 [unknown] (/bin/ls)
> >>> ls 171459 5380927.623723: 549960 cycles: ffffffff8c82b138 [unknown] ([unknown])
> >>>
> >>> Display 'dsoip' field:
> >>> $ perf script -F +dsoip
> >>> perf-exec 171459 5380927.621499: 1 cycles: ffffffff8c66e1b8 [unknown] ([unknown]+0x0)
> >>> perf-exec 171459 5380927.621524: 1 cycles: ffffffff8c66e1b8 [unknown] ([unknown]+0x0)
> >>> perf-exec 171459 5380927.621527: 2 cycles: ffffffff8c66e1b8 [unknown] ([unknown]+0x0)
> >>> perf-exec 171459 5380927.621530: 17 cycles: ffffffff8c66e1b8 [unknown] ([unknown]+0x0)
> >>> perf-exec 171459 5380927.621533: 194 cycles: ffffffff8c66e1bf [unknown] ([unknown]+0x0)
> >>> perf-exec 171459 5380927.621536: 2153 cycles: ffffffff8c6375fe [unknown] ([unknown]+0x0)
> >>> perf-exec 171459 5380927.621540: 23135 cycles: ffffffff8d001ac4 [unknown] ([unknown]+0x0)
> >>> perf-exec 171459 5380927.621555: 224703 cycles: ffffffff8c7a5153 [unknown] ([unknown]+0x0)
> >>> ls 171459 5380927.621686: 587397 cycles: ffffffff8c7c43de [unknown] ([unknown]+0x0)
> >>> ls 171459 5380927.622016: 653693 cycles: 7fe1e47278c7 _dl_map_object_deps+0x307 (/lib/x86_64-linux-gnu/ld-2.27.so+0xe8c7)
> >>> ls 171459 5380927.622382: 633896 cycles: 7fe1e41bd0d0 __wmemchr_ifunc+0x0 (/lib/x86_64-linux-gnu/libc-2.27.so+0xbd0d0)
> >>> ls 171459 5380927.622737: 608856 cycles: 7fe1e4194594 _int_malloc+0x264 (/lib/x86_64-linux-gnu/libc-2.27.so+0x94594)
> >>> ls 171459 5380927.623078: 586376 cycles: 55ff5293d27c [unknown] (/bin/ls+0x1227c)
> >>> ls 171459 5380927.623405: 566812 cycles: 55ff5293b202 [unknown] (/bin/ls+0x10202)
> >>> ls 171459 5380927.623723: 549960 cycles: ffffffff8c82b138 [unknown] ([unknown]+0x0)
> >>>
> >>> Signed-off-by: Changbin Du <changbin.du@xxxxxxxxxx>
> >>> ---
> >>> tools/perf/Documentation/perf-script.txt | 2 +-
> >>> tools/perf/builtin-script.c | 5 +++++
> >>> tools/perf/util/evsel_fprintf.c | 26 ++++++++++++------------
> >>> tools/perf/util/evsel_fprintf.h | 13 ++++++------
> >>> 4 files changed, 26 insertions(+), 20 deletions(-)
> >>>
> >>> diff --git a/tools/perf/Documentation/perf-script.txt b/tools/perf/Documentation/perf-script.txt
> >>> index 777a0d8ba7d1..ef392785f776 100644
> >>> --- a/tools/perf/Documentation/perf-script.txt
> >>> +++ b/tools/perf/Documentation/perf-script.txt
> >>> @@ -130,7 +130,7 @@ OPTIONS
> >>> -F::
> >>> --fields::
> >>> Comma separated list of fields to print. Options are:
> >>> - comm, tid, pid, time, cpu, event, trace, ip, sym, dso, addr, symoff,
> >>> + comm, tid, pid, time, cpu, event, trace, ip, sym, dso, dsoip, addr, symoff,
> >>> srcline, period, iregs, uregs, brstack, brstacksym, flags, bpf-output,
> >>> brstackinsn, brstackinsnlen, brstackoff, callindent, insn, insnlen, synth,
> >>> phys_addr, metric, misc, srccode, ipc, data_page_size, code_page_size, ins_lat,
> >>> diff --git a/tools/perf/builtin-script.c b/tools/perf/builtin-script.c
> >>> index a792214d1af8..a181decf7922 100644
> >>> --- a/tools/perf/builtin-script.c
> >>> +++ b/tools/perf/builtin-script.c
> >>> @@ -133,6 +133,7 @@ enum perf_output_field {
> >>> PERF_OUTPUT_VCPU = 1ULL << 38,
> >>> PERF_OUTPUT_CGROUP = 1ULL << 39,
> >>> PERF_OUTPUT_RETIRE_LAT = 1ULL << 40,
> >>> + PERF_OUTPUT_DSOIP = 1ULL << 41,
> >>> };
> >>>
> >>> struct perf_script {
> >>> @@ -174,6 +175,7 @@ struct output_option {
> >>> {.str = "ip", .field = PERF_OUTPUT_IP},
> >>> {.str = "sym", .field = PERF_OUTPUT_SYM},
> >>> {.str = "dso", .field = PERF_OUTPUT_DSO},
> >>> + {.str = "dsoip", .field = PERF_OUTPUT_DSOIP},
> >>> {.str = "addr", .field = PERF_OUTPUT_ADDR},
> >>> {.str = "symoff", .field = PERF_OUTPUT_SYMOFFSET},
> >>> {.str = "srcline", .field = PERF_OUTPUT_SRCLINE},
> >>> @@ -574,6 +576,9 @@ static void set_print_ip_opts(struct perf_event_attr *attr)
> >>> if (PRINT_FIELD(DSO))
> >>> output[type].print_ip_opts |= EVSEL__PRINT_DSO;
> >>>
> >>> + if (PRINT_FIELD(DSOIP))
> >>> + output[type].print_ip_opts |= EVSEL__PRINT_DSOIP;
> >>> +
> >>> if (PRINT_FIELD(SYMOFFSET))
> >>> output[type].print_ip_opts |= EVSEL__PRINT_SYMOFFSET;
> >>>
> >>> diff --git a/tools/perf/util/evsel_fprintf.c b/tools/perf/util/evsel_fprintf.c
> >>> index bd22c4932d10..fd1791fab051 100644
> >>> --- a/tools/perf/util/evsel_fprintf.c
> >>> +++ b/tools/perf/util/evsel_fprintf.c
> >>> @@ -116,6 +116,7 @@ int sample__fprintf_callchain(struct perf_sample *sample, int left_alignment,
> >>> int print_ip = print_opts & EVSEL__PRINT_IP;
> >>> int print_sym = print_opts & EVSEL__PRINT_SYM;
> >>> int print_dso = print_opts & EVSEL__PRINT_DSO;
> >>> + int print_dsoip = print_opts & EVSEL__PRINT_DSOIP;
> >>> int print_symoffset = print_opts & EVSEL__PRINT_SYMOFFSET;
> >>> int print_oneline = print_opts & EVSEL__PRINT_ONELINE;
> >>> int print_srcline = print_opts & EVSEL__PRINT_SRCLINE;
> >>> @@ -133,7 +134,7 @@ int sample__fprintf_callchain(struct perf_sample *sample, int left_alignment,
> >>> while (1) {
> >>> struct symbol *sym;
> >>> struct map *map;
> >>> - u64 addr = 0;
> >>> + u64 dso_ip = 0;
> >>
> >> Renaming 'addr' just makes the patch more complicated.
> >> Just leave it as 'addr'
> >>
> > Just to make the variable name reflect its real content. It's not a memory
> > address, right?
>
> We already use addr e.g. in struct addr_location.
>
> If you want to rename it, make it a separate patch.
>
okay.

> >
> >>>
> >>> node = callchain_cursor_current(cursor);
> >>> if (!node)
> >>> @@ -141,6 +142,8 @@ int sample__fprintf_callchain(struct perf_sample *sample, int left_alignment,
> >>>
> >>> sym = node->ms.sym;
> >>> map = node->ms.map;
> >>> + if (map)
> >>> + dso_ip = map->map_ip(map, node->ip);
> >>
> >> There does not seem to be any reason to move these lines of code.
> >>
> > yes. I cound undo this. (just a clean up to not place addr calculation betwwen
> > printings)
> >
> >>>
> >>> if (sym && sym->ignore && print_skip_ignored)
> >>> goto next;
> >>> @@ -150,20 +153,12 @@ int sample__fprintf_callchain(struct perf_sample *sample, int left_alignment,
> >>> if (print_arrow && !first)
> >>> printed += fprintf(fp, " <-");
> >>>
> >>> - if (map)
> >>> - addr = map->map_ip(map, node->ip);
> >>> -
> >>> - if (print_ip) {
> >>> - /* Show binary offset for userspace addr */
> >>> - if (map && !map->dso->kernel)
> >>
> >> I would have expected to keep existing functionality unchanged i.e.
> >>
> > But the goal is to make printed address consistent. See below difference.
>
> Then make it a separate patch.
>
okay. this should be done in muti-patches.

> >
> > $ perf script
> > ...
> > ls 1341034 2739463.008343: 2162417 cycles:
> > ffffffff992a917d [unknown] ([unknown])
> > ffffffff992a9b0a [unknown] ([unknown])
> > ffffffff992efa3d [unknown] ([unknown])
> > ffffffff992f10c9 [unknown] ([unknown])
> > ffffffff992f47cc [unknown] ([unknown])
> > ffffffff992f4ab8 [unknown] ([unknown])
> > ffffffff9909daa2 [unknown] ([unknown])
> > ffffffff99d657a7 [unknown] ([unknown])
> > ffffffff99e00b67 [unknown] ([unknown])
> > 235d3 memset+0x53 (/usr/lib/x86_64-linux-gnu/ld-2.31.so) # dso offset
> > a61b _dl_map_object+0x1bb (/usr/lib/x86_64-linux-gnu/ld-2.31.so)
> >
> > $ perf script -G
> > ...
> > ls 1341034 2739463.008876: 2053304 cycles: ffffffffc1596923 [unknown] ([unknown])
> > ls 1341034 2739463.009381: 1917049 cycles: 14def8e149e6 __strcoll_l+0xd96 (/usr/lib/x86_64-linux-gnu/libc-2.31.so) # raw ip
> >
> >
> >> - if (map && !map->dso->kernel)
> >> + if (!print_dsoip && map && !map->dso->kernel)
> >>
> >>> - printed += fprintf(fp, "%c%16" PRIx64, s, addr);
> >>> - else
> >>> - printed += fprintf(fp, "%c%16" PRIx64, s, node->ip);
> >>> - }
> >>> + if (print_ip)
> >>> + printed += fprintf(fp, "%c%16" PRIx64, s, node->ip);
> >>>
> >>> if (print_sym) {
> >>> printed += fprintf(fp, " ");
> >>> - node_al.addr = addr;
> >>> + node_al.addr = dso_ip;
> >>> node_al.map = map;
> >>>
> >>> if (print_symoffset) {
> >>> @@ -179,11 +174,13 @@ int sample__fprintf_callchain(struct perf_sample *sample, int left_alignment,
> >>> if (print_dso && (!sym || !sym->inlined)) {
> >>> printed += fprintf(fp, " (");
> >>> printed += map__fprintf_dsoname(map, fp);
> >>> + if (print_dsoip)
> >>
> >> Better to avoid '[unknown]+0x0' i.e.
> >>
> >> if (print_dsoip && map && map->dso)
> >>
> > This is also to keep output consistent which is frendly for parsers (FlameGraph
> > generator for example).
>
> And yet you want to change the existing output even when the
> new field is not requested. That is also unfriendly.
>
I'd think consistency is very important. It confuses people without reading the
code. So let's do them in seprated patches: one for consistency, another to add
new dso_offset field.

> The default perf script output is human readable. Programs
> should use the Python scripting interface (or the dlfilter
> interface) to get data, not scrape the human readable
> output which is, after all, not an API.
>
> Besides we don't do that (unknown+0x0) for unknown symbols, so
> why do it for unknown offsets.
Both (unknown) and (unknown+0x0) are good for me.

>
> >
> >>> + printed += fprintf(fp, "+0x%lx", (map && map->dso) ? dso_ip : 0);
> >>> printed += fprintf(fp, ")");
> >>> }
> >>>
> >>> if (print_srcline)
> >>> - printed += map__fprintf_srcline(map, addr, "\n ", fp);
> >>> + printed += map__fprintf_srcline(map, dso_ip, "\n ", fp);
> >>>
> >>> if (sym && sym->inlined)
> >>> printed += fprintf(fp, " (inlined)");
> >>> @@ -214,6 +211,7 @@ int sample__fprintf_sym(struct perf_sample *sample, struct addr_location *al,
> >>> int print_ip = print_opts & EVSEL__PRINT_IP;
> >>> int print_sym = print_opts & EVSEL__PRINT_SYM;
> >>> int print_dso = print_opts & EVSEL__PRINT_DSO;
> >>> + int print_dsoip = print_opts & EVSEL__PRINT_DSOIP;
> >>> int print_symoffset = print_opts & EVSEL__PRINT_SYMOFFSET;
> >>> int print_srcline = print_opts & EVSEL__PRINT_SRCLINE;
> >>> int print_unknown_as_addr = print_opts & EVSEL__PRINT_UNKNOWN_AS_ADDR;
> >>> @@ -242,6 +240,8 @@ int sample__fprintf_sym(struct perf_sample *sample, struct addr_location *al,
> >>> if (print_dso) {
> >>> printed += fprintf(fp, " (");
> >>> printed += map__fprintf_dsoname(al->map, fp);
> >>> + if (print_dsoip)
> >>
> >> Better to avoid '[unknown]+0x0' i.e.
> >>
> >> if (print_dsoip && al->map && al->map->dso)
> >>
> >>
> > Ditto.
> >
> >>> + printed += fprintf(fp, "+0x%lx", (al->map && al->map->dso) ? al->addr : 0);
> >>> printed += fprintf(fp, ")");
> >>> }
> >>>
> >>> diff --git a/tools/perf/util/evsel_fprintf.h b/tools/perf/util/evsel_fprintf.h
> >>> index 3093d096c29f..3f282c5840da 100644
> >>> --- a/tools/perf/util/evsel_fprintf.h
> >>> +++ b/tools/perf/util/evsel_fprintf.h
> >>> @@ -20,12 +20,13 @@ int evsel__fprintf(struct evsel *evsel, struct perf_attr_details *details, FILE
> >>> #define EVSEL__PRINT_IP (1<<0)
> >>> #define EVSEL__PRINT_SYM (1<<1)
> >>> #define EVSEL__PRINT_DSO (1<<2)
> >>> -#define EVSEL__PRINT_SYMOFFSET (1<<3)
> >>> -#define EVSEL__PRINT_ONELINE (1<<4)
> >>> -#define EVSEL__PRINT_SRCLINE (1<<5)
> >>> -#define EVSEL__PRINT_UNKNOWN_AS_ADDR (1<<6)
> >>> -#define EVSEL__PRINT_CALLCHAIN_ARROW (1<<7)
> >>> -#define EVSEL__PRINT_SKIP_IGNORED (1<<8)
> >>> +#define EVSEL__PRINT_DSOIP (1<<3)
> >>> +#define EVSEL__PRINT_SYMOFFSET (1<<4)
> >>> +#define EVSEL__PRINT_ONELINE (1<<5)
> >>> +#define EVSEL__PRINT_SRCLINE (1<<6)
> >>> +#define EVSEL__PRINT_UNKNOWN_AS_ADDR (1<<7)
> >>> +#define EVSEL__PRINT_CALLCHAIN_ARROW (1<<8)
> >>> +#define EVSEL__PRINT_SKIP_IGNORED (1<<9)
> >>>
> >>> struct addr_location;
> >>> struct perf_event_attr;
> >>
> >>
> >
>

--
Cheers,
Changbin Du

Next message: Sunil V L: "Re: [PATCH V3 00/20] Add basic ACPI support for RISC-V"
Previous message: yebin (H): "Re: [PATCH 2/2] lib/percpu_counter: fix dying cpu compare race"
In reply to: Adrian Hunter: "Re: [PATCH] perf: script: add new output field 'dsoip'"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]