tools, perf: Add asprintf replacement
PATCH 1/1 tools, perf: Add asprintf replacement
From: Andi Kleen <[email protected]>
asprintf corrupts memory on some older glibc versions. Provide a replacement. This fixes various segfaults with --branch-history on older Fedoras.
--- tools/perf/Makefile.perf | 1 + tools/perf/util/asprintf.c | 28 ++++++++++++++++++++++++++++ 2 files changed, 29 insertions(+) create mode 100644 tools/perf/util/asprintf.c
tools/perf/Makefile.perf +1 -0
@@ -373,6 +373,7 @@ LIB_OBJS += $(OUTPUT)util/vdso.o LIB_OBJS += $(OUTPUT)util/stat.o LIB_OBJS += $(OUTPUT)util/record.o LIB_OBJS += $(OUTPUT)util/srcline.o +LIB_OBJS += $(OUTPUT)util/asprintf.o LIB_OBJS += $(OUTPUT)util/data.o LIB_OBJS += $(OUTPUT)ui/setup.o
tools/perf/util/asprintf.c new file +28 -0
@@ -0,0 +1,28 @@ +/* Replacement for asprintf as it's buggy in older glibc versions */ +#include <stdio.h> +#include <stdarg.h> +#include <stdlib.h> +#include <string.h> + +int vasprintf (char **str,const char *fmt,va_list ap) +{ +char buf[1024 ]; +int len =vsnprintf (buf,sizeof buf, fmt, ap); + + *str =malloc (len +1 ); +if (!*str) +return -1 ; +strcpy (*str, buf); +return len; +} + +int asprintf (char **str,const char *fmt, ...) +{ +va_list ap; +int ret; + +va_start (ap, fmt); + ret =vasprintf (str, fmt, ap); +va_end (ap); +return ret; +}
Signature · 1 line
1.8.5.3
Em Mon, Mar 10, 2014 at 11:43:24PM -0700, Andi Kleen escreveu:
Quoted 5 lines
From: Andi Kleen <[email protected]>
asprintf corrupts memory on some older glibc versions. Provide a replacement. This fixes various segfaults with --branch-history on older Fedoras.
Humm, this unconditionally replaces it with an alternative that limits the buffer to a fixed size :-\
Do you recall at least one of those old glibc version/release number?
A reproducer? So that I can try to reproduce it here and try to polish this a bit more...
- Arnaldo
Quoted 55 lines
Signed-off-by Andi Kleen <[email protected]>--- tools/perf/Makefile.perf | 1 + tools/perf/util/asprintf.c | 28 ++++++++++++++++++++++++++++ 2 files changed, 29 insertions(+) create mode 100644 tools/perf/util/asprintf.ctools/perf/Makefile.perf +1
-0@@ -373,6 +373,7 @@ LIB_OBJS += $(OUTPUT)util/vdso.o LIB_OBJS += $(OUTPUT)util/stat.o LIB_OBJS += $(OUTPUT)util/record.o LIB_OBJS += $(OUTPUT)util/srcline.o +LIB_OBJS += $(OUTPUT)util/asprintf.o LIB_OBJS += $(OUTPUT)util/data.o LIB_OBJS += $(OUTPUT)ui/setup.otools/perf/util/asprintf.c new file +28
-0@@ -0,0 +1,28 @@ +/* Replacement for asprintf as it's buggy in older glibc versions */ +#include <stdio.h> +#include <stdarg.h> +#include <stdlib.h> +#include <string.h> + +int vasprintf (char **str,const char *fmt,va_list ap) +{ +char buf[1024 ]; +int len =vsnprintf (buf,sizeof buf, fmt, ap); + + *str =malloc (len +1 ); +if (!*str) +return -1 ; +strcpy (*str, buf); +return len; +} + +int asprintf (char **str,const char *fmt, ...) +{ +va_list ap; +int ret; + +va_start (ap, fmt); + ret =vasprintf (str, fmt, ap); +va_end (ap); +return ret; +}Signature · 1 line
1.8.5.3
Quoted 2 lines
Humm, this unconditionally replaces it with an alternative that limits the buffer to a fixed size :-\
Better than corrupting memory.
I guess you could use two passes to avoid the limit, but it would surprise me if anything in perf needs more than 1K of printf. One issue
with doing two passes is that I wasn't sure the snprintf return value would work properly on all libcs (e.g. the weirdo one Android uses)
Quoted 2 lines
Do you recall at least one of those old glibc version/release number?
glibc-2.13-2.x86_64 (FC14)
Quoted 3 lines
A reproducer? So that I can try to reproduce it here and try to polish this a bit more...
I saw it with perf report --branch-history in TUI mode and then pressing e. But even running valgrind in stdio mode showed some corruption. Without the patch also using some of the --call-graph options segfaulted.
-Andi
Em Tue, Mar 18, 2014 at 08:05:33AM -0700, Andi Kleen escreveu:
Quoted 4 lines
Quoted 2 lines
Humm, this unconditionally replaces it with an alternative that limits the buffer to a fixed size :-\
Better than corrupting memory.
Yes, it is better than corrupting memory, use the less ugly, good point.
Quoted 3 lines
I guess you could use two passes to avoid the limit, but it would surprise me if anything in perf needs more than 1K of printf. One issue
with doing two passes is that I wasn't sure the snprintf return
The return of snprintf is crazy, that is why we use scnprintf.
Quoted 5 lines
value would work properly on all libcs (e.g. the weirdo one Android uses)
Quoted 2 lines
Do you recall at least one of those old glibc version/release number?
Quoted 8 lines
glibc-2.13-2.x86_64 (FC14)
Quoted 2 lines
A reproducer? So that I can try to reproduce it here and try to polish this a bit more...
I saw it with perf report --branch-history in TUI mode and then pressing e. But even running valgrind in stdio mode showed some corruption. Without the patch also using some of the --call-graph options segfaulted.
Thanks for the data points,
- Arnaldo