tools/lib/traceevent, tools/perf: Move struct tep_handler definition in a local header file
PATCH 1/1 tools/lib/traceevent, tools/perf: Move struct tep_handler definition in a local header file
From: Tzvetomir Stoyanov <[email protected]>
As traceevent is going to be transferred into a proper library, its local data should be protected from the library users. This patch encapsulates struct tep_handler into a local header, not visible outside of the library. It implements also a bunch of new APIs, which library users can use to access tep_handler members.
--- tools/lib/traceevent/Build | 1 + tools/lib/traceevent/event-parse-api.c | 275 +++++++++++++++++++++++++++++++ tools/lib/traceevent/event-parse-local.h | 92 +++++++++++ tools/lib/traceevent/event-parse.c | 2 + tools/lib/traceevent/event-parse.h | 228 ++++--------------------- tools/lib/traceevent/event-plugin.c | 1 + tools/lib/traceevent/parse-filter.c | 1 + tools/perf/util/trace-event-parse.c | 25 +-- tools/perf/util/trace-event-read.c | 2 +- 9 files changed, 416 insertions(+), 211 deletions(-) create mode 100644 tools/lib/traceevent/event-parse-api.c create mode 100644 tools/lib/traceevent/event-parse-local.h
tools/lib/traceevent/Build +1 -0
@@ -5,6 +5,7 @@ libtraceevent-y += parse-filter.o libtraceevent-y += parse-utils.o libtraceevent-y += kbuffer-parse.o libtraceevent-y += tep_strerror.o +libtraceevent-y += event-parse-api.o plugin_jbd2-y += plugin_jbd2.o plugin_hrtimer-y += plugin_hrtimer.o
tools/lib/traceevent/event-parse-api.c new file +275 -0
275 lines not on this page · raw message
tools/lib/traceevent/event-parse-local.h new file +92 -0
92 lines not on this page · raw message
tools/lib/traceevent/event-parse.c +2 -0
@@ -22,6 +22,8 @@#include <netinet/in.h> #include "event-parse.h" + +#include "event-parse-local.h" #include "event-utils.h" #include "trace-seq.h"
tools/lib/traceevent/event-parse.h +28 -200
253 lines not on this page · raw message
tools/lib/traceevent/event-plugin.c +1 -0
@@ -14,6 +14,7 @@#include <unistd.h> #include <dirent.h> #include "event-parse.h" +#include "event-parse-local.h" #include "event-utils.h" #include "trace-seq.h"
tools/lib/traceevent/parse-filter.c +1 -0
@@ -11,6 +11,7 @@#include <sys/types.h> #include "event-parse.h" +#include "event-parse-local.h" #include "event-utils.h" #define COMM "COMM"
tools/perf/util/trace-event-parse.c +15 -10
@@ -37,10 +37,11 @@ static int get_common_field(struct scripting_context *context,struct tep_format_field *field;if (!*size) { -if (!pevent->events) + + event =tep_get_first_event (pevent); +if (!event)return 0 ; - event = pevent->events[0 ]; field =tep_find_common_field (event, type);if (!field)return 0 ;
@@ -192,25 +193,29 @@ struct tep_event_format *trace_find_next_event(struct tep_handle *pevent, struct tep_event_format *event) {static int idx; +int events_count; +struct tep_event_format *all_events; -if (!pevent || !pevent->events) + all_events =tep_get_first_event (pevent); + events_count =tep_get_events_count (pevent); +if (!pevent || !all_events || events_count <1 )return NULL ;if (!event) { idx =0 ; -return pevent->events[0 ]; +return all_events; } -if (idx < pevent->nr_events && event == pevent->events[idx]) { +if (idx < events_count && event == (all_events + idx)) { idx++; -if (idx == pevent->nr_events) +if (idx == events_count)return NULL ; -return pevent->events[idx]; +return (all_events + idx); } -for (idx =1 ; idx < pevent->nr_events; idx++) { -if (event == pevent->events[idx -1 ]) -return pevent->events[idx]; +for (idx =1 ; idx < events_count; idx++) { +if (event == (all_events + (idx -1 ))) +return (all_events + idx); }return NULL ; }
tools/perf/util/trace-event-read.c +1 -1
@@ -241,7 +241,7 @@ static int read_header_files(struct tep_handle *pevent)* The commit field in the page is of type long, * use that instead, since it represents the kernel. */ -tep_set_long_size (pevent, pevent->header_page_size_size); +tep_set_long_size (pevent,tep_get_header_page_size (pevent)); }free (header_page);
Signature · 1 line
2.13.6
Quoted 5 lines · On Fri, 5 Oct 2018 12:22:25 -0400 Steven Rostedt <[email protected]> wrote:
tools/lib/traceevent/event-parse-api.c new file +5
-0+/** +* tep_get_first_event - returns the first event in the events array +* @tep: a handle to the tep_handle +* +* This returns pointer to the first element of the events array
tools/lib/traceevent/event-parse-api.c line 25
Quoted 10 lines
tools/lib/traceevent/event-parse-api.c new file +10
-0+* If @tep is NULL, NULL is returned. +*/ +struct tep_event_format *tep_get_first_event (struct tep_handle *tep) +{ +if (tep && tep->events) +return tep->events[0 ]; + +return NULL ; +} +
Hold off on pulling this in. This patch has some whitespace issues that need to be cleaned up.
Thanks!
-- Steve
Arnaldo Carvalho de Melo
maintainerEm Fri, Oct 05, 2018 at 12:28:46PM -0400, Steven Rostedt escreveu:
Quoted 13 lines
Quoted 10 lines · On Fri, 5 Oct 2018 12:22:25 -0400 Steven Rostedt <[email protected]> wrote:
tools/lib/traceevent/event-parse-api.c new file +10
-0+/** +* tep_get_first_event - returns the first event in the events array +* @tep: a handle to the tep_handle +* +* This returns pointer to the first element of the events array +* If @tep is NULL, NULL is returned. +*/ +struct tep_event_format *tep_get_first_event (struct tep_handle *tep) +{ +if (tep && tep->events)
Quoted 8 lines
tools/lib/traceevent/event-parse-api.c line 25
Quoted 5 lines
tools/lib/traceevent/event-parse-api.c new file +5
-0+return tep->events[0 ]; + +return NULL ; +} +Hold off on pulling this in. This patch has some whitespace issues that need to be cleaned up.
Ah, that if( thing? ?I can fix it here, holler if there is more than that.
Quoted 23 lines · On Mon, 8 Oct 2018 14:32:24 -0300 Arnaldo Carvalho de Melo <[email protected]>…
Em Fri, Oct 05, 2018 at 12:28:46PM -0400, Steven Rostedt escreveu:
Quoted 13 lines
Quoted 10 lines · On Fri, 5 Oct 2018 12:22:25 -0400 Steven Rostedt <[email protected]> wrote:
(quoted diff) +10
-0+/** + * tep_get_first_event - returns the first event in the events array + * @tep: a handle to the tep_handle + * + * This returns pointer to the first element of the events array + * If @tep is NULL, NULL is returned. + */ +struct tep_event_format *tep_get_first_event(struct tep_handle *tep) +{ + if(tep && tep->events)Quoted 8 lines
Review
Quoted 5 lines
(quoted diff) +5
-0+ return tep->events[0]; + + return NULL; +} +Hold off on pulling this in. This patch has some whitespace issues that need to be cleaned up.
Quoted 2 lines
Ah, that if( thing? ?I can fix it here, holler if there is more than that.
Well, here's a fixed up version anyway...
-- Steve
From 96dd7937a31d3a715c450c9a67f62a0022cd6a7c Mon Sep 17 00:00:00 2001
From: Tzvetomir Stoyanov <[email protected]>
Date: Mon, 8 Oct 2018 15:55:00 +0300
Subject: [PATCH] tools/lib/traceevent, tools/perf: Move struct tep_handler
definition in a local header file
As traceevent is going to be transferred into a proper library, its local data should be protected from the library users. This patch capsulates struct tep_handler into a local header, not visible outside of the library. It implements also a bunch of new APIs, which library users can use to access tep_handler members. Version 3 of the patch - fixed whitespace issues.
--- tools/lib/traceevent/Build | 1 + tools/lib/traceevent/event-parse-api.c | 275 +++++++++++++++++++++++++++++++ tools/lib/traceevent/event-parse-local.h | 92 +++++++++++ tools/lib/traceevent/event-parse.c | 2 + tools/lib/traceevent/event-parse.h | 230 ++++---------------------- tools/lib/traceevent/event-plugin.c | 1 + tools/lib/traceevent/parse-filter.c | 1 + tools/perf/util/trace-event-parse.c | 25 +-- tools/perf/util/trace-event-read.c | 2 +- 9 files changed, 417 insertions(+), 212 deletions(-) create mode 100644 tools/lib/traceevent/event-parse-api.c create mode 100644 tools/lib/traceevent/event-parse-local.h
tools/lib/traceevent/Build +1 -0
@@ -5,6 +5,7 @@ libtraceevent-y += parse-filter.o libtraceevent-y += parse-utils.o libtraceevent-y += kbuffer-parse.o libtraceevent-y += tep_strerror.o +libtraceevent-y += event-parse-api.o plugin_jbd2-y += plugin_jbd2.o plugin_hrtimer-y += plugin_hrtimer.o
tools/lib/traceevent/event-parse-api.c new file +275 -0
275 lines not on this page · raw message
tools/lib/traceevent/event-parse-local.h new file +92 -0
92 lines not on this page · raw message
tools/lib/traceevent/event-parse.c +2 -0
@@ -22,6 +22,8 @@#include <netinet/in.h> #include "event-parse.h" + +#include "event-parse-local.h" #include "event-utils.h" #include "trace-seq.h"
tools/lib/traceevent/event-parse.h +29 -201
259 lines not on this page · raw message
tools/lib/traceevent/event-plugin.c +1 -0
@@ -14,6 +14,7 @@#include <unistd.h> #include <dirent.h> #include "event-parse.h" +#include "event-parse-local.h" #include "event-utils.h" #include "trace-seq.h"
tools/lib/traceevent/parse-filter.c +1 -0
@@ -11,6 +11,7 @@#include <sys/types.h> #include "event-parse.h" +#include "event-parse-local.h" #include "event-utils.h" #define COMM "COMM"
tools/perf/util/trace-event-parse.c +15 -10
@@ -37,10 +37,11 @@ static int get_common_field(struct scripting_context *context,struct tep_format_field *field;if (!*size) { -if (!pevent->events) + + event =tep_get_first_event (pevent); +if (!event)return 0 ; - event = pevent->events[0 ]; field =tep_find_common_field (event, type);if (!field)return 0 ;
@@ -192,25 +193,29 @@ struct tep_event_format *trace_find_next_event(struct tep_handle *pevent, struct tep_event_format *event) {static int idx; +int events_count; +struct tep_event_format *all_events; -if (!pevent || !pevent->events) + all_events =tep_get_first_event (pevent); + events_count =tep_get_events_count (pevent); +if (!pevent || !all_events || events_count <1 )return NULL ;if (!event) { idx =0 ; -return pevent->events[0 ]; +return all_events; } -if (idx < pevent->nr_events && event == pevent->events[idx]) { +if (idx < events_count && event == (all_events + idx)) { idx++; -if (idx == pevent->nr_events) +if (idx == events_count)return NULL ; -return pevent->events[idx]; +return (all_events + idx); } -for (idx =1 ; idx < pevent->nr_events; idx++) { -if (event == pevent->events[idx -1 ]) -return pevent->events[idx]; +for (idx =1 ; idx < events_count; idx++) { +if (event == (all_events + (idx -1 ))) +return (all_events + idx); }return NULL ; }
tools/perf/util/trace-event-read.c +1 -1
@@ -241,7 +241,7 @@ static int read_header_files(struct tep_handle *pevent)* The commit field in the page is of type long, * use that instead, since it represents the kernel. */ -tep_set_long_size (pevent, pevent->header_page_size_size); +tep_set_long_size (pevent,tep_get_header_page_size (pevent)); }free (header_page);
Signature · 1 line
2.13.6
Em Fri, Oct 05, 2018 at 12:22:25PM -0400, Steven Rostedt escreveu:
Quoted 11 lines
From: Tzvetomir Stoyanov <[email protected]>
As traceevent is going to be transferred into a proper library, its local data should be protected from the library users. This patch encapsulates struct
tep_handlerinto a local header, not visible outside of the library. It implements also a bunch of new APIs, which library users can use to accesstep_handlermembers.Signed-off-by Tzvetomir Stoyanov <[email protected]>Signed-off-by Steven Rostedt (VMware) <[email protected]>
Thanks, applied to perf/core.
- Arnaldo
tip-bot for Tzvetomir Stoyanov
merge bot[tip:perf/core] tools lib traceevent, perf tools: Move struct tep_handler definition in a local header file
In reply to Steven Rostedttools lib traceevent, perf tools: Move struct tep_handler definition in a local header file
Commit-ID: bb3dd7e7c4d5e024d607c0ec06c2a2fb9408cc99 Gitweb: https://git.kernel.org/tip/bb3dd7e7c4d5e024d607c0ec06c2a2fb9408cc99 Author: Tzvetomir Stoyanov <[email protected]> AuthorDate: Fri, 5 Oct 2018 12:22:25 -0400 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 8 Oct 2018 15:05:37 -0300
tools lib traceevent, perf tools: Move struct tep_handler definition in a local header file
As traceevent is going to be transferred into a proper library, its local data should be protected from the library users. This patch encapsulates struct tep_handler into a local header, not visible outside of the library. It implements also a bunch of new APIs, which library users can use to access tep_handler members.
--- tools/lib/traceevent/Build | 1 + tools/lib/traceevent/event-parse-api.c | 275 +++++++++++++++++++++++++++++++ tools/lib/traceevent/event-parse-local.h | 92 +++++++++++ tools/lib/traceevent/event-parse.c | 2 + tools/lib/traceevent/event-parse.h | 228 ++++--------------------- tools/lib/traceevent/event-plugin.c | 1 + tools/lib/traceevent/parse-filter.c | 1 + tools/perf/util/trace-event-parse.c | 25 +-- tools/perf/util/trace-event-read.c | 2 +- 9 files changed, 416 insertions(+), 211 deletions(-)
tools/lib/traceevent/Build +1 -0
@@ -5,6 +5,7 @@ libtraceevent-y += parse-filter.o libtraceevent-y += parse-utils.o libtraceevent-y += kbuffer-parse.o libtraceevent-y += tep_strerror.o +libtraceevent-y += event-parse-api.o plugin_jbd2-y += plugin_jbd2.o plugin_hrtimer-y += plugin_hrtimer.o
tools/lib/traceevent/event-parse-api.c new file +275 -0
275 lines not on this page · raw message
tools/lib/traceevent/event-parse-local.h new file +92 -0
92 lines not on this page · raw message
tools/lib/traceevent/event-parse.c +2 -0
@@ -22,6 +22,8 @@#include <netinet/in.h> #include "event-parse.h" + +#include "event-parse-local.h" #include "event-utils.h" #include "trace-seq.h"
tools/lib/traceevent/event-parse.h +28 -200
253 lines not on this page · raw message
tools/lib/traceevent/event-plugin.c +1 -0
@@ -14,6 +14,7 @@#include <unistd.h> #include <dirent.h> #include "event-parse.h" +#include "event-parse-local.h" #include "event-utils.h" #include "trace-seq.h"
tools/lib/traceevent/parse-filter.c +1 -0
@@ -11,6 +11,7 @@#include <sys/types.h> #include "event-parse.h" +#include "event-parse-local.h" #include "event-utils.h" #define COMM "COMM"
tools/perf/util/trace-event-parse.c +15 -10
@@ -37,10 +37,11 @@ static int get_common_field(struct scripting_context *context,struct tep_format_field *field;if (!*size) { -if (!pevent->events) + + event =tep_get_first_event (pevent); +if (!event)return 0 ; - event = pevent->events[0 ]; field =tep_find_common_field (event, type);if (!field)return 0 ;
@@ -193,25 +194,29 @@ struct tep_event_format *trace_find_next_event(struct tep_handle *pevent, struct tep_event_format *event) {static int idx; +int events_count; +struct tep_event_format *all_events; -if (!pevent || !pevent->events) + all_events =tep_get_first_event (pevent); + events_count =tep_get_events_count (pevent); +if (!pevent || !all_events || events_count <1 )return NULL ;if (!event) { idx =0 ; -return pevent->events[0 ]; +return all_events; } -if (idx < pevent->nr_events && event == pevent->events[idx]) { +if (idx < events_count && event == (all_events + idx)) { idx++; -if (idx == pevent->nr_events) +if (idx == events_count)return NULL ; -return pevent->events[idx]; +return (all_events + idx); } -for (idx =1 ; idx < pevent->nr_events; idx++) { -if (event == pevent->events[idx -1 ]) -return pevent->events[idx]; +for (idx =1 ; idx < events_count; idx++) { +if (event == (all_events + (idx -1 ))) +return (all_events + idx); }return NULL ; }
tools/perf/util/trace-event-read.c +1 -1
@@ -241,7 +241,7 @@ static int read_header_files(struct tep_handle *pevent)* The commit field in the page is of type long, * use that instead, since it represents the kernel. */ -tep_set_long_size (pevent, pevent->header_page_size_size); +tep_set_long_size (pevent,tep_get_header_page_size (pevent)); }free (header_page);
Hi,
Quoted 7 lines · Steven Rostedt
From: Tzvetomir Stoyanov <[email protected]>
As traceevent is going to be transferred into a proper library, its local data should be protected from the library users. This patch encapsulates struct
tep_handlerinto a local header, not visible outside of the library. It implements also a bunch of new APIs, which library users can use to accesstep_handlermembers.
This commit appears to have broken perf script --gen-script / trace_find_next_event(). As far as I can tell:
@ -192,25 +193,29 @@ struct tep_event_format *trace_find_next_event(struct tep_handle *pevent,
Quoted 27 lines
(quoted diff) +9
-5struct tep_event_format *event) { static int idx; + int events_count; + struct tep_event_format *all_events; - if (!pevent || !pevent->events) + all_events = tep_get_first_event(pevent); + events_count = tep_get_events_count(pevent); + if (!pevent || !all_events || events_count < 1) return NULL; if (!event) { idx = 0; - return pevent->events[0]; + return all_events; } - if (idx < pevent->nr_events && event == pevent->events[idx]) { + if (idx < events_count && event == (all_events + idx)) { idx++; - if (idx == pevent->nr_events) + if (idx == events_count) return NULL; - return pevent->events[idx]; + return (all_events + idx); }
Review
Quoted 10 lines
(quoted diff) +3
-3- for (idx = 1; idx < pevent->nr_events; idx++) { - if (event == pevent->events[idx - 1]) - return pevent->events[idx]; + for (idx = 1; idx < events_count; idx++) { + if (event == (all_events + (idx - 1))) + return (all_events + idx); } return NULL; }
Is just plain wrong, as:
Review
Quoted 2 lines
(quoted diff) +1
-1- return pevent->events[idx]; + return (all_events + idx);
that's not a valid conversion. ->events isn't an array of tep_handle, it's an array of tep_handle* (and even if it were, the previous notation would have needed to dereference the value to make it comparable). What this does is look idx behind the individually allocated event. Which is incorrect.
To reproduce the crash, just generating a trace file with at least two events suffices:
perf record -e raw_syscalls:sys_enter,raw_syscalls:sys_exit sleep 1
gdb --args perf script -g py
Program received signal SIGSEGV, Segmentation fault.__strlen_avx2 () at ../sysdeps/x86_64/multiarch/strlen-avx2.S:96
96 ../sysdeps/x86_64/multiarch/strlen-avx2.S: No such file or directory.
(gdb) bt
#0 __strlen_avx2 () at ../sysdeps/x86_64/multiarch/strlen-avx2.S:96
#1 0x00007ffff726f9ef in _IO_vfprintf_internal (s=s@entry=0x555555c23970, format=format@entry=0x5555558fed76 "def %s__%s(",
ap=ap@entry=0x7fffffffb900) at vfprintf.c:1638
#2 0x00007ffff7326536 in ___fprintf_chk (fp=fp@entry=0x555555c23970, flag=flag@entry=1,
format=format@entry=0x5555558fed76 "def %s__%s(") at fprintf_chk.c:35
#3 0x000055555587065c in fprintf (__fmt=0x5555558fed76 "def %s__%s(", __stream=0x555555c23970)
at /usr/include/x86_64-linux-gnu/bits/stdio2.h:100
#4 python_generate_script (pevent=0x555555c2e620, outfile=<optimized out>)
at util/scripting-engines/trace-event-python.c:1651
#5 0x0000555555745762 in cmd_script (argc=<optimized out>, argv=0x7fffffffe4d0) at builtin-script.c:3743
#6 0x000055555579821d in run_builtin (p=0x555555abeb38 <commands+408>, argc=3, argv=0x7fffffffe4d0) at perf.c:303
#7 0x0000555555712b7a in handle_internal_command (argv=0x7fffffffe4d0, argc=3) at perf.c:355
#8 run_argv (argcp=<synthetic pointer>, argv=<synthetic pointer>) at perf.c:399
#9 main (argc=3, argv=0x7fffffffe4d0) at perf.c:521The fix (in recent kernel versions) appears to just bee to use tep_get_event(), instead of the old pevent->events[...]. But it appears to me that
commit 80c5526c8544ae76cba31fb9702ab8accac1f0f3
Author: Tzvetomir Stoyanov <[email protected]> Date: 2019-04-01 12:43:12 -0400
tools lib traceevent: Implement new traceevent APIs for accessing struct tep_handler fields
ommitted adding it to event-parse.h. It appears to be intended as public API however, given that it got documented in
commit 747e942c3925bb85e2865371664499a98fca83b0
Author: Tzvetomir Stoyanov <[email protected]> Date: 2019-05-10 15:56:23 -0400
tools lib traceevent: Man pages for libtraceevent event get APIs
The following patch fixes this for me. I can polish it up, but I'm wondering if I'm missing something here?
tools/lib/traceevent/event-parse.h +1 -0
@@ -517,6 +517,7 @@ int tep_read_number_field(struct tep_format_field *field, const void *data, unsigned long long *value);struct tep_event *tep_get_first_event (struct tep_handle *tep); +struct tep_event *tep_get_event (struct tep_handle *tep,int index);int tep_get_events_count (struct tep_handle *tep);struct tep_event *tep_find_event (struct tep_handle *tep,int id);
tools/perf/util/trace-event-parse.c +6 -8
@@ -179,28 +179,26 @@ struct tep_event *trace_find_next_event(struct tep_handle *pevent, {static int idx;int events_count; -struct tep_event *all_events; - all_events =tep_get_first_event (pevent); events_count =tep_get_events_count (pevent); -if (!pevent || !all_events || events_count <1 ) +if (!pevent || events_count <1 )return NULL ;if (!event) { idx =0 ; -return all_events; +return tep_get_event (pevent,0 ); } -if (idx < events_count && event == (all_events + idx)) { +if (idx < events_count && event ==tep_get_event (pevent, idx)) { idx++;if (idx == events_count)return NULL ; -return (all_events + idx); +return tep_get_event (pevent, idx); }for (idx =1 ; idx < events_count; idx++) { -if (event == (all_events + (idx -1 ))) -return (all_events + idx); +if (event ==tep_get_event (pevent, idx -1 )) +return tep_get_event (pevent, idx); }return NULL ; }
Not related to this crash, but it also seems that the whole trace_find_next_event() API ought to be removed. Back when it was
-struct event *trace_find_next_event(struct event *event) -{
- if (!event)
- return event_list;
-
- return event->next;
-}
it made some sense, but the changes in
commit aaf045f72335653b24784d6042be8e4aee114403
Author: Steven Rostedt <[email protected]> Date: 2012-04-06 00:47:56 +0200
perf: Have perf use the new libtraceevent.a library
seem to make the current API somewhat absurd, as evidenced by the complexity in trace_find_next_event(). I mean even just removing that function and changing the two callsites to simple for loops with tep_get_events_count() & tep_get_event() ought to be a lot better.
Greetings,
Andres Freund
Hi Andres,
Quoted 3 lines · Andres Freund
Hi,
...
Quoted 12 lines
Is just plain wrong, as:
Review
Quoted 2 lines
(quoted diff) +1
-1- return pevent->events[idx]; + return (all_events + idx);that's not a valid conversion. ->events isn't an array of
tep_handle, it's an array of tep_handle* (and even if it were, the previous notation would have needed to dereference the value to make it comparable). What this does is look idx behind the individually allocated event. Which is incorrect.
You are right, it is a bug.
Quoted 1 line
...
Quoted 47 lines
The following patch fixes this for me. I can polish it up, but I'm wondering if I'm missing something here?
tools/lib/traceevent/event-parse.h +1
-0@@ -517,6 +517,7 @@ int tep_read_number_field(struct tep_format_field *field, const void *data, unsigned long long *value);struct tep_event *tep_get_first_event (struct tep_handle *tep); +struct tep_event *tep_get_event (struct tep_handle *tep,int index);int tep_get_events_count (struct tep_handle *tep);struct tep_event *tep_find_event (struct tep_handle *tep,int id);tools/perf/util/trace-event-parse.c +4
-6@@ -179,28 +179,26 @@ struct tep_event *trace_find_next_event(struct tep_handle *pevent, {static int idx;int events_count; -struct tep_event *all_events; - all_events =tep_get_first_event (pevent); events_count =tep_get_events_count (pevent); -if (!pevent || !all_events || events_count <1 ) +if (!pevent || events_count <1 )return NULL ;if (!event) { idx =0 ; -return all_events; +return tep_get_event (pevent,0 ); } -if (idx < events_count && event == (all_events + idx)) { +if (idx < events_count && event ==tep_get_event (pevent, idx)) { idx++;if (idx == events_count)return NULL ; -return (all_events + idx); +return tep_get_event (pevent, idx);
tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 204
Quoted 10 lines
tools/perf/util/trace-event-parse.c +2
-2}for (idx =1 ; idx < events_count; idx++) { -if (event == (all_events + (idx -1 ))) -return (all_events + idx); +if (event ==tep_get_event (pevent, idx -1 )) +return tep_get_event (pevent, idx); }return NULL ; }
I'm OK with the first part of the patch, the changes in tools/lib/traceevent/event-parse.h.tep_get_event() is an API and is omitted in the header by mistake, it should be there.
As for the fix for the crash itself, I think it would be better to implement your suggestion -
removing trace_find_next_event() at all and replacing it with tep_get_events_count() andtep_get_event() loop.
Quoted 21 lines
Not related to this crash, but it also seems that the whole
trace_find_next_event()API ought to be removed. Back when it wastools/perf/util/trace-event-parse.c +0
-7-struct event *trace_find_next_event (struct event *event) -{ -if (!event) -return event_list; - -return event->next; -}it made some sense, but the changes in
commit
aaf045f72335653b24784d6042be8e4aee114403Author: Steven Rostedt <[email protected]> Date: 2012-04-06 00:47:56 +0200perf: Have perf use the new libtraceevent.a library
Quoted 8 lines
seem to make the current API somewhat absurd, as evidenced by the complexity in
trace_find_next_event(). I mean even just removing that function and changing the two callsites to simple for loops withtep_get_events_count()&tep_get_event()ought to be a lot better.Greetings,
Andres Freund
Thanks!
Signature · 2 lines
Tzvetomir (Ceco) Stoyanov VMware Open Source Technology Center
Quoted 2 lines · On Thu, 25 Jul 2019 20:58:29 -0700 Andres Freund <[email protected]> wrote:
Hi,
Hi Andres,
Quoted 26 lines · Steven Rostedt
Quoted 7 lines
From: Tzvetomir Stoyanov <[email protected]>
As traceevent is going to be transferred into a proper library, its local data should be protected from the library users. This patch encapsulates struct
tep_handlerinto a local header, not visible outside of the library. It implements also a bunch of new APIs, which library users can use to accesstep_handlermembers.This commit appears to have broken perf script --gen-script /
trace_find_next_event(). As far as I can tell:@ -192,25 +193,29 @@ struct
tep_event_format*trace_find_next_event(structtep_handle*pevent,Review
Quoted 9 lines
(quoted diff) +4
-1struct tep_event_format *event) { static int idx; + int events_count; + struct tep_event_format *all_events; - if (!pevent || !pevent->events) + all_events = tep_get_first_event(pevent); + events_count = tep_get_events_count(pevent);Is just plain wrong, as:
Quoted 5 lines
Review
Quoted 2 lines
(quoted diff) +1
-1- return pevent->events[idx]; + return (all_events + idx);that's not a valid conversion. ->events isn't an array of
tep_handle, it's an array of tep_handle* (and even if it were, the previous notation
You're right, it is wrong, but it's not tep_handle* but tep_event_format*.
The tep_get_first_event() returns a pointer to the first event, but that's not an array. We can't use that to index to other events.
Quoted 13 lines
would have needed to dereference the value to make it comparable). What this does is look idx behind the individually allocated event. Which is incorrect.
The fix (in recent kernel versions) appears to just bee to use
tep_get_event(), instead of the old pevent->events[...]. But it appears to me that
commit80c5526c8544ae76cba31fb9702ab8accac1f0f3
Quoted 7 lines
Author: Tzvetomir Stoyanov <[email protected]> Date: 2019-04-01 12:43:12 -0400tools lib traceevent: Implement new traceevent APIs for accessing struct
tep_handlerfieldsommitted adding it to
event-parse.h. It appears to be intended as public API however, given that it got documented in
And it appears that we missed to add it ;-)
Quoted 16 lines
commit
747e942c3925bb85e2865371664499a98fca83b0Author: Tzvetomir Stoyanov <[email protected]> Date: 2019-05-10 15:56:23 -0400tools lib traceevent: Man pages for libtraceevent event get APIs
The following patch fixes this for me. I can polish it up, but I'm wondering if I'm missing something here?
tools/lib/traceevent/event-parse.h +0
-0
tools/lib/traceevent/event-parse.h int tep_read_number_field(struct tep_format_field *field, const void *data, · line 520
Quoted 4 lines
tools/lib/traceevent/event-parse.h +1
-0@@ -517,6 +517,7 @@ int tep_read_number_field(struct tep_format_field *field, const void *data, unsigned long long *value);struct tep_event *tep_get_first_event (struct tep_handle *tep); +struct tep_event *tep_get_event (struct tep_handle *tep,int index);
I was looking at the tep_get_event() code, and we should switch that to "unsigned int index" otherwise passing in a negative number will return an address outside the array.
Quoted 8 lines
tools/lib/traceevent/event-parse.h +0
-0int tep_get_events_count (struct tep_handle *tep);struct tep_event *tep_find_event (struct tep_handle *tep,int id);tools/perf/util/trace-event-parse.c +0
-0
tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 183
Quoted 7 lines
tools/perf/util/trace-event-parse.c +0
-2@@ -179,28 +179,26 @@ struct tep_event *trace_find_next_event(struct tep_handle *pevent, {static int idx;int events_count; -struct tep_event *all_events; - all_events =tep_get_first_event (pevent); events_count =tep_get_events_count (pevent);
I think we can get rid of the events_count and all its checks, as the same check is done within tep_get_event().
tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 184
Quoted 2 lines
tools/perf/util/trace-event-parse.c +1
-1-if (!pevent || !all_events || events_count <1 ) +if (!pevent || events_count <1 )
if (!pevent)
tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 192
Quoted 10 lines
tools/perf/util/trace-event-parse.c +2
-2return NULL ;if (!event) { idx =0 ; -return all_events; +return tep_get_event (pevent,0 ); } -if (idx < events_count && event == (all_events + idx)) { +if (idx < events_count && event ==tep_get_event (pevent, idx)) {
if (event == tep_get_event(pevent, idx)) return tep_get_event(pevent, ++idx);
tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 198
Quoted 7 lines
tools/perf/util/trace-event-parse.c +1
-1idx++;if (idx == events_count)return NULL ; -return (all_events + idx); +return tep_get_event (pevent, idx); }
struct tep_event_format *next_event; for (idx = 0; next_event = tep_get_event(pevent, idx); idx++) if (event == next_event) return tep_get_event(pevent, ++idx);
Also, I think setting the idx to 1 in the loop is wrong. Why? think of this:
first_event = trace_find_next_event(pevent, NULL); next_event = trace_find_next_event(pevent, first_event); for (i = 0; i < 5; i++) next_event = trace_find_next_event(pevent, next_event); second_event = trace_find_next_event(pevent, first_event);
second_event would become NULL.
Care to send a formal patch?
Thanks!
-- Steve
Quoted 40 lines
tools/perf/util/trace-event-parse.c +2
-2for (idx =1 ; idx < events_count; idx++) { -if (event == (all_events + (idx -1 ))) -return (all_events + idx); +if (event ==tep_get_event (pevent, idx -1 )) +return tep_get_event (pevent, idx); }return NULL ;}
Not related to this crash, but it also seems that the whole
trace_find_next_event()API ought to be removed. Back when it wastools/perf/util/trace-event-parse.c +0
-7-struct event *trace_find_next_event (struct event *event) -{ -if (!event) -return event_list; - -return event->next; -}it made some sense, but the changes in
commit
aaf045f72335653b24784d6042be8e4aee114403Author: Steven Rostedt <[email protected]> Date: 2012-04-06 00:47:56 +0200perf: Have perf use the new libtraceevent.a library
seem to make the current API somewhat absurd, as evidenced by the complexity in
trace_find_next_event(). I mean even just removing that function and changing the two callsites to simple for loops withtep_get_events_count()&tep_get_event()ought to be a lot better.Greetings,
Andres Freund
Hi,
Quoted 5 lines · Steven Rostedt
Quoted 3 lines · On Thu, 25 Jul 2019 20:58:29 -0700 Andres Freund <[email protected]> wrote:
Is just plain wrong, as:
Quoted 8 lines
Quoted 5 lines
Review
Quoted 2 lines
(quoted diff) +1
-1- return pevent->events[idx]; + return (all_events + idx);that's not a valid conversion. ->events isn't an array of
tep_handle, it's an array of tep_handle* (and even if it were, the previous notationYou're right, it is wrong, but it's not tep_handle* but tep_event_format*.
Err, yea. Typo.
Quoted 7 lines
Quoted 7 lines
tools/lib/traceevent/event-parse.h +0
-0@@ -517,6 +517,7 @@ int tep_read_number_field(struct tep_format_field *field, const void *data, unsigned long long *value);
Quoted 6 lines
tools/lib/traceevent/event-parse.h int tep_read_number_field(struct tep_format_field *field, const void *data, · line 520
Quoted 2 lines
tools/lib/traceevent/event-parse.h +1
-0struct tep_event *tep_get_first_event (struct tep_handle *tep); +struct tep_event *tep_get_event (struct tep_handle *tep,int index);I was looking at the
tep_get_event()code, and we should switch that to "unsigned int index" otherwise passing in a negative number will return an address outside the array.
Makes sense.
Quoted 63 lines
Quoted 8 lines
tools/lib/traceevent/event-parse.h +0
-0int tep_get_events_count (struct tep_handle *tep);struct tep_event *tep_find_event (struct tep_handle *tep,int id);tools/perf/util/trace-event-parse.c +0
-0tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 183
Quoted 7 lines
tools/perf/util/trace-event-parse.c +0
-2@@ -179,28 +179,26 @@ struct tep_event *trace_find_next_event(struct tep_handle *pevent, {static int idx;int events_count; -struct tep_event *all_events; - all_events =tep_get_first_event (pevent); events_count =tep_get_events_count (pevent);I think we can get rid of the
events_countand all its checks, as the same check is done withintep_get_event().tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 184
Quoted 2 lines
tools/perf/util/trace-event-parse.c +1
-1-if (!pevent || !all_events || events_count <1 ) +if (!pevent || events_count <1 )if (!pevent)
tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 192
Quoted 10 lines
tools/perf/util/trace-event-parse.c +2
-2return NULL ;if (!event) { idx =0 ; -return all_events; +return tep_get_event (pevent,0 ); } -if (idx < events_count && event == (all_events + idx)) { +if (idx < events_count && event ==tep_get_event (pevent, idx)) {if (event == tep_get_event(pevent, idx)) return tep_get_event(pevent, ++idx);tools/perf/util/trace-event-parse.c struct tep_event *trace_find_next_event(struct tep_handle *pevent, · line 198
Quoted 7 lines
tools/perf/util/trace-event-parse.c +1
-1idx++;if (idx == events_count)return NULL ; -return (all_events + idx); +return tep_get_event (pevent, idx); }struct tep_event_format *next_event; for (idx = 0; next_event = tep_get_event(pevent, idx); idx++) if (event == next_event) return tep_get_event(pevent, ++idx);Also, I think setting the idx to 1 in the loop is wrong. Why? think of this:
first_event = trace_find_next_event(pevent, NULL); next_event = trace_find_next_event(pevent, first_event); for (i = 0; i < 5; i++) next_event = trace_find_next_event(pevent, next_event); second_event = trace_find_next_event(pevent, first_event);
Quoted 1 line
second_eventwould become NULL.
How about my proposal to instead change the loops in trace-event-{python,perl}.c, the only callers of trace_find_next_event, to be something akin to
[idx_type_for_tep_get_event] event_count = tep_get_events_count(pevent); for ([idx_type_for_tep_get_event] idx = 0; idx < event_count; idx++)
{
struct tep_event *event = tep_get_events(...);}
and just removing trace_find_next_event()? It's not a nice API imo, and seems unnecessary given that the events aren't a linked list anymore.
Quoted 1 line
Care to send a formal patch?
Will do.
Greetings,
Andres Freund
Quoted 6 lines · On Fri, 26 Jul 2019 13:55:44 -0700 Andres Freund <[email protected]> wrote:
How about my proposal to instead change the loops in trace-event-{python,perl}.c, the only callers of
trace_find_next_event, to be something akin to[
idx_type_for_tep_get_event]event_count=tep_get_events_count(pevent); for ([idx_type_for_tep_get_event] idx = 0; idx <event_count; idx++)
Quoted 7 lines
{ struct tep_event *event = tep_get_events(...);}
and just removing
trace_find_next_event()? It's not a nice API imo, and seems unnecessary given that the events aren't a linked list anymore.
Yep, go for it :-)
Quoted 5 lines
Quoted 3 lines
Care to send a formal patch?
Will do.
Thanks!
-- Steve
Quoted 3 lines · On Fri, 26 Jul 2019 13:55:44 -0700 Andres Freund <[email protected]> wrote:
Quoted 1 line
Care to send a formal patch?
Will do.
Hi Andres,
Have you had a chance to send a patch?
Thanks!
-- Steve