trace-cmd: make sure we have a pending pid in trace-hist
PATCH 1/1 trace-cmd: make sure we have a pending pid in trace-hist
I have a trace.dat file where we lost some events and so we didn't have the corresponding event to go with the stack trace. This made trace-cmd hist segfault because it unconditionally pushes the pending_pid's function onto the stack trace. In this case pending_pid < 0, which means we didn't have a pending stack trace, so we got garbage in the call chain and hilarity ensued. Fix this by only pushing the pending call chain if there actually is one, and also don't reset the current call chain if there isn't actually one there either. This fixed my segfault. Thanks,
--- Makefile | 9 +++++---- trace-cmd.c | 3 +++ trace-hist.c | 14 +++++++++----- trace-local.h | 2 ++ trace-usage.c | 5 +++++ 5 files changed, 24 insertions(+), 9 deletions(-)
Makefile +5 -4
@@ -81,7 +81,7 @@ ifndef NO_PYTHON PYTHON := ctracecmd.so PYTHON_GUI := ctracecmd.so ctracecmdgui.so -PYTHON_VERS ?= python +PYTHON_VERS ?= python-2.7 # Can build python? ifeq ($(shell sh -c "pkg-config --cflags $(PYTHON_VERS) > /dev/null 2>&1 && echo y"), y)
@@ -164,6 +164,7 @@ LIBS = -L. -ltracecmd -ldl LIB_FILE = libtracecmd.a PACKAGES= gtk+-2.0 libxml-2.0 gthread-2.0 +NOGUI_PACKAGES= glib-2.0 ifndef BUILDGUI BUILDGUI = 0
@@ -194,8 +195,8 @@ N = @/bin/true || CONFIG_FLAGS += $(HELP_DIR_SQ) else -CONFIG_INCLUDES = -CONFIG_LIBS = +CONFIG_INCLUDES = $(shell pkg-config --cflags $(NOGUI_PACKAGES)) +CONFIG_LIBS = $(shell pkg-config --libs $(NOGUI_PACKAGES)) CONFIG_FLAGS = VERSION = $(TC_VERSION)
@@ -307,7 +308,7 @@ $(obj)/%.o: $(src)/%.c TRACE_GUI_OBJS = trace-filter.o trace-compat.o trace-hash.o trace-dialog.o \ trace-xml.o TRACE_CMD_OBJS = trace-cmd.o trace-record.o trace-read.o trace-split.o trace-listen.o \ - trace-stack.o trace-hist.o trace-mem.o trace-snapshot.o + trace-stack.o trace-hist.o trace-mem.o trace-snapshot.o trace-latency.o TRACE_VIEW_OBJS = trace-view.o trace-view-store.o TRACE_GRAPH_OBJS = trace-graph.o trace-plot.o trace-plot-cpu.o trace-plot-task.o TRACE_VIEW_MAIN_OBJS = trace-view-main.o $(TRACE_VIEW_OBJS) $(TRACE_GUI_OBJS)
trace-cmd.c +3 -0
@@ -432,6 +432,9 @@ int main (int argc, char **argv) }else if (strcmp (argv[1 ],"stack" ) ==0 ) {trace_stack (argc, argv);exit (0 ); + }else if (strcmp (argv[1 ],"latency" ) ==0 ) { +trace_latency (argc, argv); +exit (0 ); }else if (strcmp (argv[1 ],"check-events" ) ==0 ) {char *tracing;int ret;
trace-hist.c +9 -5
@@ -453,9 +453,11 @@ process_kernel_stack(struct pevent *pevent, struct pevent_record *record) } }else {/* function stack trace? */ -copy_stack_to_pending (current_pid); -free (ips); -reset_stack (); +if (current_pid >=0 ) { +copy_stack_to_pending (current_pid); +free (ips); +reset_stack (); + } } current_pid = pid;
@@ -482,8 +484,10 @@ process_kernel_stack(struct pevent *pevent, struct pevent_record *record)push_stack_func (func); } -push_stack_func (pending_ips[pending_ips_idx -1 ]); -reset_pending_stack (); +if (pending_pid >=0 ) { +push_stack_func (pending_ips[pending_ips_idx -1 ]); +reset_pending_stack (); + }save_call_chain (current_pid, ips, ips_idx,1 );if (do_restore)restore_stack (current_pid);
trace-local.h +2 -0
@@ -57,6 +57,8 @@ void trace_snapshot(int argc, char **argv);void trace_mem (int argc,char **argv); +void trace_latency (int argc,char **argv); +/* --- instance manipulation --- */ struct func_list {
trace-usage.c +5 -0
@@ -231,6 +231,11 @@ static struct usage_help usage_help[] = {" -N do not load any plugins\n" }, { +"latency" , +"parse latency events" , +"" + }, + { NULL, NULL, NULL } };
Signature · 1 line
2.0.0
Quoted 95 lines · On Wed, 30 Jul 2014 16:54:45 -0400 Josef Bacik <[email protected]> wrote:
I have a trace.dat file where we lost some events and so we didn't have the corresponding event to go with the stack trace. This made trace-cmd hist segfault because it unconditionally pushes the pending_pid's function onto the stack trace. In this case
pending_pid< 0, which means we didn't have a pending stack trace, so we got garbage in the call chain and hilarity ensued. Fix this by only pushing the pending call chain if there actually is one, and also don't reset the current call chain if there isn't actually one there either. This fixed my segfault. Thanks,Signed-off-by Josef Bacik <[email protected]>--- Makefile | 9 +++++---- trace-cmd.c | 3 +++ trace-hist.c | 14 +++++++++----- trace-local.h | 2 ++ trace-usage.c | 5 +++++ 5 files changed, 24 insertions(+), 9 deletions(-)Makefile +5
-4@@ -81,7 +81,7 @@ ifndef NO_PYTHON PYTHON := ctracecmd.so PYTHON_GUI := ctracecmd.so ctracecmdgui.so -PYTHON_VERS ?= python +PYTHON_VERS ?= python-2.7 # Can build python? ifeq ($(shell sh -c "pkg-config --cflags $(PYTHON_VERS) > /dev/null 2>&1 && echo y"), y)@@ -164,6 +164,7 @@ LIBS = -L. -ltracecmd -ldl LIB_FILE = libtracecmd.a PACKAGES= gtk+-2.0 libxml-2.0 gthread-2.0 +NOGUI_PACKAGES= glib-2.0 ifndef BUILDGUI BUILDGUI = 0@@ -194,8 +195,8 @@ N = @/bin/true || CONFIG_FLAGS += $(HELP_DIR_SQ) else -CONFIG_INCLUDES = -CONFIG_LIBS = +CONFIG_INCLUDES = $(shell pkg-config --cflags $(NOGUI_PACKAGES)) +CONFIG_LIBS = $(shell pkg-config --libs $(NOGUI_PACKAGES)) CONFIG_FLAGS = VERSION = $(TC_VERSION)@@ -307,7 +308,7 @@ $(obj)/%.o: $(src)/%.c TRACE_GUI_OBJS = trace-filter.o trace-compat.o trace-hash.o trace-dialog.o \ trace-xml.o TRACE_CMD_OBJS = trace-cmd.o trace-record.o trace-read.o trace-split.o trace-listen.o \ - trace-stack.o trace-hist.o trace-mem.o trace-snapshot.o + trace-stack.o trace-hist.o trace-mem.o trace-snapshot.o trace-latency.o TRACE_VIEW_OBJS = trace-view.o trace-view-store.o TRACE_GRAPH_OBJS = trace-graph.o trace-plot.o trace-plot-cpu.o trace-plot-task.o TRACE_VIEW_MAIN_OBJS = trace-view-main.o $(TRACE_VIEW_OBJS) $(TRACE_GUI_OBJS)trace-cmd.c +3
-0@@ -432,6 +432,9 @@ int main (int argc, char **argv) }else if (strcmp (argv[1 ],"stack" ) ==0 ) {trace_stack (argc, argv);exit (0 ); + }else if (strcmp (argv[1 ],"latency" ) ==0 ) { +trace_latency (argc, argv); +exit (0 ); }else if (strcmp (argv[1 ],"check-events" ) ==0 ) {char *tracing;int ret;trace-hist.c +5
-3@@ -453,9 +453,11 @@ process_kernel_stack(struct pevent *pevent, struct pevent_record *record) } }else {/* function stack trace? */ -copy_stack_to_pending (current_pid); -free (ips); -reset_stack (); +if (current_pid >=0 ) { +copy_stack_to_pending (current_pid); +free (ips); +reset_stack (); + } } current_pid = pid;@@ -482,8 +484,10 @@ process_kernel_stack(struct pevent *pevent, struct pevent_record *record)push_stack_func (func); }
trace-hist.c process_kernel_stack(struct pevent *pevent, struct pevent_record *record) · line 493
Quoted 10 lines
trace-hist.c +4
-2-push_stack_func (pending_ips[pending_ips_idx -1 ]); -reset_pending_stack (); +if (pending_pid >=0 ) { +push_stack_func (pending_ips[pending_ips_idx -1 ]); +reset_pending_stack (); + }save_call_chain (current_pid, ips, ips_idx,1 );if (do_restore)restore_stack (current_pid);
This blob looks to be what the change log suggest. I think you added a bunch more than the fix.
Can you resend please with just the change from the change log. If you want to add more, please send separate patches. This shouldn't be like congress adding earmarks now, is it ;-)
Thanks!
-- Steve
Quoted 29 lines
trace-local.h +2
-0@@ -57,6 +57,8 @@ void trace_snapshot(int argc, char **argv);void trace_mem (int argc,char **argv); +void trace_latency (int argc,char **argv); +/* --- instance manipulation --- */ struct func_list {trace-usage.c +5
-0@@ -231,6 +231,11 @@ static struct usage_help usage_help[] = {" -N do not load any plugins\n" }, { +"latency" , +"parse latency events" , +"" + }, + { NULL, NULL, NULL } };
Oops crap that's the second time I've done that this week. When I get to my computer I'll send the actual fix by itself, sorry about that,
Josef
Steven Rostedt <[email protected]> wrote:
Quoted 95 lines · On Wed, 30 Jul 2014 16:54:45 -0400 Josef Bacik <[email protected]> wrote:
I have a trace.dat file where we lost some events and so we didn't have the corresponding event to go with the stack trace. This made trace-cmd hist segfault because it unconditionally pushes the pending_pid's function onto the stack trace. In this case
pending_pid< 0, which means we didn't have a pending stack trace, so we got garbage in the call chain and hilarity ensued. Fix this by only pushing the pending call chain if there actually is one, and also don't reset the current call chain if there isn't actually one there either. This fixed my segfault. Thanks,Signed-off-by Josef Bacik <[email protected]>--- Makefile | 9 +++++---- trace-cmd.c | 3 +++ trace-hist.c | 14 +++++++++----- trace-local.h | 2 ++ trace-usage.c | 5 +++++ 5 files changed, 24 insertions(+), 9 deletions(-)Makefile +5
-4@@ -81,7 +81,7 @@ ifndef NO_PYTHON PYTHON := ctracecmd.so PYTHON_GUI := ctracecmd.so ctracecmdgui.so -PYTHON_VERS ?= python +PYTHON_VERS ?= python-2.7 # Can build python? ifeq ($(shell sh -c "pkg-config --cflags $(PYTHON_VERS) > /dev/null 2>&1 && echo y"), y)@@ -164,6 +164,7 @@ LIBS = -L. -ltracecmd -ldl LIB_FILE = libtracecmd.a PACKAGES= gtk+-2.0 libxml-2.0 gthread-2.0 +NOGUI_PACKAGES= glib-2.0 ifndef BUILDGUI BUILDGUI = 0@@ -194,8 +195,8 @@ N = @/bin/true || CONFIG_FLAGS += $(HELP_DIR_SQ) else -CONFIG_INCLUDES = -CONFIG_LIBS = +CONFIG_INCLUDES = $(shell pkg-config --cflags $(NOGUI_PACKAGES)) +CONFIG_LIBS = $(shell pkg-config --libs $(NOGUI_PACKAGES)) CONFIG_FLAGS = VERSION = $(TC_VERSION)@@ -307,7 +308,7 @@ $(obj)/%.o: $(src)/%.c TRACE_GUI_OBJS = trace-filter.o trace-compat.o trace-hash.o trace-dialog.o \ trace-xml.o TRACE_CMD_OBJS = trace-cmd.o trace-record.o trace-read.o trace-split.o trace-listen.o \ - trace-stack.o trace-hist.o trace-mem.o trace-snapshot.o + trace-stack.o trace-hist.o trace-mem.o trace-snapshot.o trace-latency.o TRACE_VIEW_OBJS = trace-view.o trace-view-store.o TRACE_GRAPH_OBJS = trace-graph.o trace-plot.o trace-plot-cpu.o trace-plot-task.o TRACE_VIEW_MAIN_OBJS = trace-view-main.o $(TRACE_VIEW_OBJS) $(TRACE_GUI_OBJS)trace-cmd.c +3
-0@@ -432,6 +432,9 @@ int main (int argc, char **argv) }else if (strcmp (argv[1 ],"stack" ) ==0 ) {trace_stack (argc, argv);exit (0 ); + }else if (strcmp (argv[1 ],"latency" ) ==0 ) { +trace_latency (argc, argv); +exit (0 ); }else if (strcmp (argv[1 ],"check-events" ) ==0 ) {char *tracing;int ret;trace-hist.c +5
-3@@ -453,9 +453,11 @@ process_kernel_stack(struct pevent *pevent, struct pevent_record *record) } }else {/* function stack trace? */ -copy_stack_to_pending (current_pid); -free (ips); -reset_stack (); +if (current_pid >=0 ) { +copy_stack_to_pending (current_pid); +free (ips); +reset_stack (); + } } current_pid = pid;@@ -482,8 +484,10 @@ process_kernel_stack(struct pevent *pevent, struct pevent_record *record)push_stack_func (func); }
trace-hist.c process_kernel_stack(struct pevent *pevent, struct pevent_record *record) · line 493
Quoted 10 lines
trace-hist.c +4
-2-push_stack_func (pending_ips[pending_ips_idx -1 ]); -reset_pending_stack (); +if (pending_pid >=0 ) { +push_stack_func (pending_ips[pending_ips_idx -1 ]); +reset_pending_stack (); + }save_call_chain (current_pid, ips, ips_idx,1 );if (do_restore)restore_stack (current_pid);
This blob looks to be what the change log suggest. I think you added a bunch more than the fix.
Can you resend please with just the change from the change log. If you want to add more, please send separate patches. This shouldn't be like congress adding earmarks now, is it ;-)
Thanks!
-- Steve
Quoted 29 lines
trace-local.h +2
-0@@ -57,6 +57,8 @@ void trace_snapshot(int argc, char **argv);void trace_mem (int argc,char **argv); +void trace_latency (int argc,char **argv); +/* --- instance manipulation --- */ struct func_list {trace-usage.c +5
-0@@ -231,6 +231,11 @@ static struct usage_help usage_help[] = {" -N do not load any plugins\n" }, { +"latency" , +"parse latency events" , +"" + }, + { NULL, NULL, NULL } };