Skip to contentSkip to conversation

perf/core improvements and fixes

GIT PULL New Arnaldo Carvalho de Melo 6 participants 26 messages Jun 3, 2015 – Jun 10, 2015 perf events

lore
Share & tools
lore

Arnaldo Carvalho de Melo

author maintainer
#0 lore raw

Hi Ingo,

Please consider applying.

One of the next requests probably will have the eBPF work by Wang Nan,
but I am still going thru it and want to test it thoroughly.

BTW: Have you looked at it lately? It is at:

http://lkml.kernel.org/r/[email protected]

Super summary from the above cover letter:

---------------------
It enables 'perf record' to filter events using eBPF programs like:

 # perf record --event bpf-file.o sleep 1

Events are selected and filtered according to definitions in bpf-file.o.
---------------------

The first two patches from that series are in this pull req, as
they just move stuff into tools/include/linux/ from tools/perf/include.

Regards,

- Arnaldo

The following changes since commit 5c9b9bc67c684e40b3a5e7e9facde0fb7200cd8c:

Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core (2015-05-29 20:19:02 +0200)

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git tags/perf-core-for-mingo

for you to fetch changes up to 1f121b03d058dd07199d8924373d3c52a207f63b:

perf tools: Deal with kernel module names in '[]' correctly (2015-06-03 10:02:38 -0300)

----------------------------------------------------------------
perf/core improvements and fixes:

User visible:

- Fix 'perf probe' segfault when glob matching function without debuginfo (Wang Nan)

- Remove newline char when reading event scale and unit (Madhavan Srinivasan)

- Deal with kernel module names in '[]' correctly (Wang Nan)

Infrastructure:

- Fix the search for the kernel DSO on the unified list (Arnaldo Carvalho de Melo)

- Move tools/perf/util/include/linux/{kernel.h,list.h,poison.h} to tools/include, to be used in tools/lib/bpf/ (Wang Nan)

----------------------------------------------------------------
Arnaldo Carvalho de Melo (1):
perf machine: Fix the search for the kernel DSO on the unified list

Madhavan Srinivasan (1):
perf tools: Remove newline char when reading event scale and unit

Wang Nan (4):
perf probe: Fix segfault when glob matching function without debuginfo perf tools: Move linux/kernel.h to tools/include
tools: Move tools/perf/util/include/linux/{list.h,poison.h} to tools/include perf tools: Deal with kernel module names in '[]' correctly

 tools/{perf/util => }/include/linux/kernel.h |  4 +-
 tools/{perf/util => }/include/linux/list.h   |  6 +--
 tools/include/linux/poison.h                 |  1 +
 tools/perf/MANIFEST                          |  3 ++
 tools/perf/tests/kmod-path.c                 | 72 ++++++++++++++++++++++++++++
 tools/perf/util/dso.c                        | 47 ++++++++++++++++--
 tools/perf/util/dso.h                        |  2 +-
 tools/perf/util/header.c                     |  8 ++--
 tools/perf/util/include/linux/poison.h       |  1 -
 tools/perf/util/machine.c                    | 22 ++++++++-
 tools/perf/util/pmu.c                        | 11 ++++-
 tools/perf/util/probe-event.c                | 26 ++++++++--
 12 files changed, 179 insertions(+), 24 deletions(-)
 rename tools/{perf/util => }/include/linux/kernel.h (97%)
 rename tools/{perf/util => }/include/linux/list.h (90%)
 create mode 100644 tools/include/linux/poison.h
 delete mode 100644 tools/perf/util/include/linux/poison.h
author maintainer
#1 lore raw

PATCH 1/6 perf probe: Fix segfault when glob matching function without debuginfo

From: Wang Nan <[email protected]>

Commit 4c859351226c920b227fec040a3b447f0d482af3 ("perf probe: Support glob wildcards for function name") introduces segfault problems when debuginfo is not available:

 # perf probe 'sys_w*'
  Added new events:
  Segmentation fault

The first problem resides in find_probe_trace_events_from_map(). In that function, find_probe_functions() is called to match each symbol against glob to find the number of matching functions, but still use map__for_each_symbol_by_name() to find 'struct symbol' for matching functions. Unfortunately, map__for_each_symbol_by_name() does exact matching by searching in an rbtree.

It doesn't know glob matching, and not easy for it to support it because it use rbtree based binary search, but we are unable to ensure all names matched by the glob (any glob passed by user) reside in one subtree.

This patch drops map__for_each_symbol_by_name(). Since there is no rbtree again, re-matching all symbols costs a lot. This patch avoid it by saving all matching results into an array (syms).

The second problem is the lost of tp->realname. In __add_probe_trace_events(), if pev->point.function is glob, the event name should be set to tev->point.realname. This patch ensures its existence by strdup sym->name instead of leaving a NULL pointer there.

After this patch:

 # perf probe 'sys_w*'
 Added new events:
   probe:sys_waitid     (on sys_w*)
   probe:sys_wait4      (on sys_w*)
   probe:sys_waitpid    (on sys_w*)
   probe:sys_write      (on sys_w*)
   probe:sys_writev     (on sys_w*)

You can now use it in all perf tools, such as:

perf record -e probe:sys_writev -aR sleep 1

---
 tools/perf/util/probe-event.c | 26 +++++++++++++++++++++-----
 1 file changed, 21 insertions(+), 5 deletions(-)
tools/perf/util/probe-event.c +21 -5
index d27edef5eb5b..e6f215b7a052 100644
@@ -2494,7 +2494,8 @@ close_out:
 	return ret;
 }
 
-static int find_probe_functions(struct map *map, char *name)
+static int find_probe_functions(struct map *map, char *name,
+				struct symbol **syms)
 {
 	int found = 0;
 	struct symbol *sym;
@@ -2504,8 +2505,11 @@ static int find_probe_functions(struct map *map, char *name)
 		return 0;
 
 	map__for_each_symbol(map, sym, tmp) {
-		if (strglobmatch(sym->name, name))
+		if (strglobmatch(sym->name, name)) {
 			found++;
+			if (syms && found < probe_conf.max_probes)
+				syms[found - 1] = sym;
+		}
 	}
 
 	return found;
@@ -2528,11 +2532,12 @@ static int find_probe_trace_events_from_map(struct perf_probe_event *pev,
 	struct map *map = NULL;
 	struct ref_reloc_sym *reloc_sym = NULL;
 	struct symbol *sym;
+	struct symbol **syms = NULL;
 	struct probe_trace_event *tev;
 	struct perf_probe_point *pp = &pev->point;
 	struct probe_trace_point *tp;
 	int num_matched_functions;
-	int ret, i;
+	int ret, i, j;
 
 	map = get_target_map(pev->target, pev->uprobes);
 	if (!map) {
@@ -2540,11 +2545,17 @@ static int find_probe_trace_events_from_map(struct perf_probe_event *pev,
 		goto out;
 	}
 
+	syms = malloc(sizeof(struct symbol *) * probe_conf.max_probes);
+	if (!syms) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
 	/*
 	 * Load matched symbols: Since the different local symbols may have
 	 * same name but different addresses, this lists all the symbols.
 	 */
-	num_matched_functions = find_probe_functions(map, pp->function);
+	num_matched_functions = find_probe_functions(map, pp->function, syms);
 	if (num_matched_functions == 0) {
 		pr_err("Failed to find symbol %s in %s\n", pp->function,
 			pev->target ? : "kernel");
@@ -2575,7 +2586,9 @@ static int find_probe_trace_events_from_map(struct perf_probe_event *pev,
 
 	ret = 0;
 
-	map__for_each_symbol_by_name(map, pp->function, sym) {
+	for (j = 0; j < num_matched_functions; j++) {
+		sym = syms[j];
+
 		tev = (*tevs) + ret;
 		tp = &tev->point;
 		if (ret == num_matched_functions) {
@@ -2599,6 +2612,8 @@ static int find_probe_trace_events_from_map(struct perf_probe_event *pev,
 			tp->symbol = strdup_or_goto(sym->name, nomem_out);
 			tp->offset = pp->offset;
 		}
+		tp->realname = strdup_or_goto(sym->name, nomem_out);
+
 		tp->retprobe = pp->retprobe;
 		if (pev->target)
 			tev->point.module = strdup_or_goto(pev->target,
@@ -2629,6 +2644,7 @@ static int find_probe_trace_events_from_map(struct perf_probe_event *pev,
 
 out:
 	put_target_map(map, pev->uprobes);
+	free(syms);
 	return ret;
 
 nomem_out:
Signature · 1 line
2.1.0
author maintainer
#2 lore raw

PATCH 2/6 perf tools: Remove newline char when reading event scale and unit

In reply to Arnaldo Carvalho de Melo

From: Madhavan Srinivasan <[email protected]>

The <fd979c013207> commit intruduced the perf_event_sysfs_show function to display the event_str value of an attr in kernel/event/core.c. But the function returns the value with a newline char.

So, if a event also carries a event.unit file, when printing the counter data perf tool formatting goes for a spin.

That is, because of the event unit, event name is printed in the newline because of perf_event_sysfs_show returns with a newline char.

Now fixing perf core will break API, hencing proposing a fix in the perf tool.

Cc Jiri Olsa <[email protected]>
Cc Michael Ellerman <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Sukadev Bhattiprolu <[email protected]>
[ Add spaces around operators ]
---
 tools/perf/util/pmu.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)
tools/perf/util/pmu.c +9 -2
index 5d3ab7c8ceaf..0fcc624eb767 100644
@@ -112,7 +112,11 @@ static int perf_pmu__parse_scale(struct perf_pmu_alias *alias, char *dir, char *
 	if (sret < 0)
 		goto error;
 
-	scale[sret] = '\0';
+	if (scale[sret - 1] == '\n')
+		scale[sret - 1] = '\0';
+	else
+		scale[sret] = '\0';
+
 	/*
 	 * save current locale
 	 */
@@ -154,7 +158,10 @@ static int perf_pmu__parse_unit(struct perf_pmu_alias *alias, char *dir, char *n
 
 	close(fd);
 
-	alias->unit[sret] = '\0';
+	if (alias->unit[sret - 1] == '\n')
+		alias->unit[sret - 1] = '\0';
+	else
+		alias->unit[sret] = '\0';
 
 	return 0;
 error:
Signature · 1 line
2.1.0
author maintainer
#3 lore raw

PATCH 3/6 perf machine: Fix the search for the kernel DSO on the unified list

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

When unifying the user_dsos and kernel_dsos a bug was introduced by inverting the check for dso->kernel, fix it.

Fixes 3d39ac538629 ("perf machine: No need to have two DSOs lists")
Cc Adrian Hunter <[email protected]>
Cc David Ahern <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
---
 tools/perf/util/machine.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/util/machine.c +1 -1
index 2ed61f59d415..4e29e80932e5 100644
@@ -1149,7 +1149,7 @@ static int machine__process_kernel_mmap_event(struct machine *machine,
 		struct dso *dso;
 
 		list_for_each_entry(dso, &machine->dsos.head, node) {
-			if (dso->kernel && is_kernel_module(dso->long_name))
+			if (!dso->kernel || is_kernel_module(dso->long_name))
 				continue;
 
 			kernel = dso;
Signature · 1 line
2.1.0
author maintainer
#4 lore raw

PATCH 4/6 perf tools: Move linux/kernel.h to tools/include

In reply to Arnaldo Carvalho de Melo

From: Wang Nan <[email protected]>

This patch moves kernel.h from tools/perf/util/include/linux/kernel.h to tools/include/linux/kernel.h to enable other libraries use macros in it, like libbpf which will be introduced by further patches.

MANIFEST is also updated for 'make perf-*-src-pkg'.

Cc Brendan Gregg <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc David Ahern <[email protected]>
Cc He Kuang <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Kaixu Xia <[email protected]>
Cc Masami Hiramatsu <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Zefan Li <[email protected]>
[ Fixed up the ifdef guard to match other entries in tools/include/linux ]
---
 tools/include/linux/kernel.h           | 107 +++++++++++++++++++++++++++++++++
 tools/perf/MANIFEST                    |   1 +
 tools/perf/util/include/linux/kernel.h | 107 ---------------------------------
 3 files changed, 108 insertions(+), 107 deletions(-)
 create mode 100644 tools/include/linux/kernel.h
 delete mode 100644 tools/perf/util/include/linux/kernel.h
tools/include/linux/kernel.h new file +107 -0
new file mode 100644
index 000000000000..76df53539c2a
@@ -0,0 +1,107 @@
+#ifndef __TOOLS_LINUX_KERNEL_H
+#define __TOOLS_LINUX_KERNEL_H
+
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+
+#define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
+
+#define PERF_ALIGN(x, a)	__PERF_ALIGN_MASK(x, (typeof(x))(a)-1)
+#define __PERF_ALIGN_MASK(x, mask)	(((x)+(mask))&~(mask))
+
+#ifndef offsetof
+#define offsetof(TYPE, MEMBER) ((size_t) &((TYPE *)0)->MEMBER)
+#endif
+
+#ifndef container_of
+/**
+ * container_of - cast a member of a structure out to the containing structure
+ * @ptr:	the pointer to the member.
+ * @type:	the type of the container struct this is embedded in.
+ * @member:	the name of the member within the struct.
+ *
+ */
+#define container_of(ptr, type, member) ({			\
+	const typeof(((type *)0)->member) * __mptr = (ptr);	\
+	(type *)((char *)__mptr - offsetof(type, member)); })
+#endif
+
+#define BUILD_BUG_ON_ZERO(e) (sizeof(struct { int:-!!(e); }))
+
+#ifndef max
+#define max(x, y) ({				\
+	typeof(x) _max1 = (x);			\
+	typeof(y) _max2 = (y);			\
+	(void) (&_max1 == &_max2);		\
+	_max1 > _max2 ? _max1 : _max2; })
+#endif
+
+#ifndef min
+#define min(x, y) ({				\
+	typeof(x) _min1 = (x);			\
+	typeof(y) _min2 = (y);			\
+	(void) (&_min1 == &_min2);		\
+	_min1 < _min2 ? _min1 : _min2; })
+#endif
+
+#ifndef roundup
+#define roundup(x, y) (                                \
+{                                                      \
+	const typeof(y) __y = y;		       \
+	(((x) + (__y - 1)) / __y) * __y;	       \
+}                                                      \
+)
+#endif
+
+#ifndef BUG_ON
+#ifdef NDEBUG
+#define BUG_ON(cond) do { if (cond) {} } while (0)
+#else
+#define BUG_ON(cond) assert(!(cond))
+#endif
+#endif
+
+/*
+ * Both need more care to handle endianness
+ * (Don't use bitmap_copy_le() for now)
+ */
+#define cpu_to_le64(x)	(x)
+#define cpu_to_le32(x)	(x)
+
+static inline int
+vscnprintf(char *buf, size_t size, const char *fmt, va_list args)
+{
+	int i;
+	ssize_t ssize = size;
+
+	i = vsnprintf(buf, size, fmt, args);
+
+	return (i >= ssize) ? (ssize - 1) : i;
+}
+
+static inline int scnprintf(char * buf, size_t size, const char * fmt, ...)
+{
+	va_list args;
+	ssize_t ssize = size;
+	int i;
+
+	va_start(args, fmt);
+	i = vsnprintf(buf, size, fmt, args);
+	va_end(args);
+
+	return (i >= ssize) ? (ssize - 1) : i;
+}
+
+/*
+ * This looks more complex than it should be. But we need to
+ * get the type for the ~ right in round_down (it needs to be
+ * as wide as the result!), and we want to evaluate the macro
+ * arguments just once each.
+ */
+#define __round_mask(x, y) ((__typeof__(x))((y)-1))
+#define round_up(x, y) ((((x)-1) | __round_mask(x, y))+1)
+#define round_down(x, y) ((x) & ~__round_mask(x, y))
+
+#endif
tools/perf/MANIFEST +1 -0
index a83cf75164e1..fce4a47347aa 100644
@@ -40,6 +40,7 @@ tools/include/linux/bitops.h
 tools/include/linux/compiler.h
 tools/include/linux/export.h
 tools/include/linux/hash.h
+tools/include/linux/kernel.h
 tools/include/linux/log2.h
 tools/include/linux/types.h
 include/asm-generic/bitops/arch_hweight.h
tools/perf/util/include/linux/kernel.h deleted +0 -107

107 lines not on this page · raw message

Signature · 1 line
2.1.0
author maintainer
#5 lore raw

PATCH 5/6 tools: Move tools/perf/util/include/linux/{list.h,poison.h} to tools/include

In reply to Arnaldo Carvalho de Melo

From: Wang Nan <[email protected]>

This patch moves list.h from tools/perf/util/include/linux/list.h to tools/include/linux/list.h to enable other libraries use macros in it, like libbpf which will be introduced by further patches. Since list.h depend on poison.h, poison.h is also moved.

Both file use relative path, so one '..' is removed for each header to make them suit for new directory.

MANIFEST is also updated for 'make perf-*-src-pkg'.

Cc Alexei Starovoitov <[email protected]>
Cc Brendan Gregg <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc David Ahern <[email protected]>
Cc He Kuang <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Kaixu Xia <[email protected]>
Cc Masami Hiramatsu <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Zefan Li <[email protected]>
---
 tools/include/linux/list.h             | 29 +++++++++++++++++++++++++++++
 tools/include/linux/poison.h           |  1 +
 tools/perf/MANIFEST                    |  2 ++
 tools/perf/util/include/linux/list.h   | 29 -----------------------------
 tools/perf/util/include/linux/poison.h |  1 -
 5 files changed, 32 insertions(+), 30 deletions(-)
 create mode 100644 tools/include/linux/list.h
 create mode 100644 tools/include/linux/poison.h
 delete mode 100644 tools/perf/util/include/linux/list.h
 delete mode 100644 tools/perf/util/include/linux/poison.h
tools/include/linux/list.h new file +29 -0
new file mode 100644
index 000000000000..76b014c96893
@@ -0,0 +1,29 @@
+#include <linux/kernel.h>
+#include <linux/types.h>
+
+#include "../../../include/linux/list.h"
+
+#ifndef TOOLS_LIST_H
+#define TOOLS_LIST_H
+/**
+ * list_del_range - deletes range of entries from list.
+ * @begin: first element in the range to delete from the list.
+ * @end: last element in the range to delete from the list.
+ * Note: list_empty on the range of entries does not return true after this,
+ * the entries is in an undefined state.
+ */
+static inline void list_del_range(struct list_head *begin,
+				  struct list_head *end)
+{
+	begin->prev->next = end->next;
+	end->next->prev = begin->prev;
+}
+
+/**
+ * list_for_each_from	-	iterate over a list from one of its nodes
+ * @pos:  the &struct list_head to use as a loop cursor, from where to start
+ * @head: the head for your list.
+ */
+#define list_for_each_from(pos, head) \
+	for (; pos != (head); pos = pos->next)
+#endif
tools/include/linux/poison.h new file +1 -0
new file mode 100644
index 000000000000..0c27bdf14233
@@ -0,0 +1 @@
+#include "../../../include/linux/poison.h"
tools/perf/MANIFEST +2 -0
index fce4a47347aa..a0bdd6124583 100644
@@ -41,7 +41,9 @@ tools/include/linux/compiler.h
 tools/include/linux/export.h
 tools/include/linux/hash.h
 tools/include/linux/kernel.h
+tools/include/linux/list.h
 tools/include/linux/log2.h
+tools/include/linux/poison.h
 tools/include/linux/types.h
 include/asm-generic/bitops/arch_hweight.h
 include/asm-generic/bitops/const_hweight.h
tools/perf/util/include/linux/list.h deleted +0 -29
deleted file mode 100644
index 76ddbc726343..000000000000
@@ -1,29 +0,0 @@
-#include <linux/kernel.h>
-#include <linux/types.h>
-
-#include "../../../../include/linux/list.h"
-
-#ifndef PERF_LIST_H
-#define PERF_LIST_H
-/**
- * list_del_range - deletes range of entries from list.
- * @begin: first element in the range to delete from the list.
- * @end: last element in the range to delete from the list.
- * Note: list_empty on the range of entries does not return true after this,
- * the entries is in an undefined state.
- */
-static inline void list_del_range(struct list_head *begin,
-				  struct list_head *end)
-{
-	begin->prev->next = end->next;
-	end->next->prev = begin->prev;
-}
-
-/**
- * list_for_each_from	-	iterate over a list from one of its nodes
- * @pos:  the &struct list_head to use as a loop cursor, from where to start
- * @head: the head for your list.
- */
-#define list_for_each_from(pos, head) \
-	for (; pos != (head); pos = pos->next)
-#endif
tools/perf/util/include/linux/poison.h deleted +0 -1
deleted file mode 100644
index fef6dbc9ce13..000000000000
@@ -1 +0,0 @@
-#include "../../../../include/linux/poison.h"
Signature · 1 line
2.1.0
author maintainer
#6 lore raw

PATCH 6/6 perf tools: Deal with kernel module names in '[]' correctly

In reply to Arnaldo Carvalho de Melo

From: Wang Nan <[email protected]>

Before patch ba92732e9808 ('perf kmaps: Check kmaps to make code more robust'), 'perf report' and 'perf annotate' will segfault if trace data contains kernel module information like this:

 # perf report -D -i ./perf.data
 ...
 0 0 0x188 [0x50]: PERF_RECORD_MMAP -1/0: [0xffffffbff1018000(0xf068000) @ 0]: x [test_module]
 ...
 # perf report -i ./perf.data --objdump=/path/to/objdump --kallsyms=/path/to/kallsyms

perf: Segmentation fault
-------- backtrace -------- /path/to/perf[0x503478] /lib64/libc.so.6(+0x3545f)[0x7fb201f3745f] /path/to/perf[0x499b56] /path/to/perf(dso__load_kallsyms+0x13c)[0x49b56c] /path/to/perf(dso__load+0x72e)[0x49c21e] /path/to/perf(map__load+0x6e)[0x4ae9ee] /path/to/perf(thread__find_addr_map+0x24c)[0x47deec] /path/to/perf(perf_event__preprocess_sample+0x88)[0x47e238] /path/to/perf[0x43ad02]
/path/to/perf[0x4b55bc] /path/to/perf(ordered_events__flush+0xca)[0x4b57ea] /path/to/perf[0x4b1a01] /path/to/perf(perf_session__process_events+0x3be)[0x4b428e] /path/to/perf(cmd_report+0xf11)[0x43bfc1] /path/to/perf[0x474702] /path/to/perf(main+0x5f5)[0x42de95] /lib64/libc.so.6(__libc_start_main+0xf4)[0x7fb201f23bd4] /path/to/perf[0x42dfc4]

This is because __kmod_path__parse treats '[' leading names as kernel name instead of names of kernel module.

If perf.data contains build information and the buildid of such modules can be found, the dso->kernel of it will be set to DSO_TYPE_KERNEL by __event_process_build_id(), not kernel module.

It will then be passed to dso__load() -> dso__load_kernel_sym() -> dso__load_kcore() if --kallsyms is provided.

The refered patch adds NULL pointer checker to avoid segfault. However, such kernel modules are still processed incorrectly.

This patch fixes __kmod_path__parse, makes it treat names like '[test_module]' as kernel modules.

kmod-path.c is also update to reflect the above changes.

[ Fixed the merged with 0443f36b0de0 ("perf machine: Fix the search
for the kernel DSO on the unified list" ]

---
 tools/perf/tests/kmod-path.c | 72 ++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/dso.c        | 47 ++++++++++++++++++++++++++---
 tools/perf/util/dso.h        |  2 +-
 tools/perf/util/header.c     |  8 ++---
 tools/perf/util/machine.c    | 22 +++++++++++++-
 5 files changed, 140 insertions(+), 11 deletions(-)
tools/perf/tests/kmod-path.c +72 -0
index e8d7cbb9320c..08c433b4bf4f 100644
@@ -34,9 +34,21 @@ static int test(const char *path, bool alloc_name, bool alloc_ext,
 	return 0;
 }
 
+static int test_is_kernel_module(const char *path, int cpumode, bool expect)
+{
+	TEST_ASSERT_VAL("is_kernel_module",
+			(!!is_kernel_module(path, cpumode)) == (!!expect));
+	pr_debug("%s (cpumode: %d) - is_kernel_module: %s\n",
+			path, cpumode, expect ? "true" : "false");
+	return 0;
+}
+
 #define T(path, an, ae, k, c, n, e) \
 	TEST_ASSERT_VAL("failed", !test(path, an, ae, k, c, n, e))
 
+#define M(path, c, e) \
+	TEST_ASSERT_VAL("failed", !test_is_kernel_module(path, c, e))
+
 int test__kmod_path__parse(void)
 {
 	/* path                alloc_name  alloc_ext   kmod  comp   name     ext */
@@ -44,30 +56,90 @@ int test__kmod_path__parse(void)
 	T("/xxxx/xxxx/x-x.ko", false     , true      , true, false, NULL   , NULL);
 	T("/xxxx/xxxx/x-x.ko", true      , false     , true, false, "[x_x]", NULL);
 	T("/xxxx/xxxx/x-x.ko", false     , false     , true, false, NULL   , NULL);
+	M("/xxxx/xxxx/x-x.ko", PERF_RECORD_MISC_CPUMODE_UNKNOWN, true);
+	M("/xxxx/xxxx/x-x.ko", PERF_RECORD_MISC_KERNEL, true);
+	M("/xxxx/xxxx/x-x.ko", PERF_RECORD_MISC_USER, false);
 
 	/* path                alloc_name  alloc_ext   kmod  comp  name   ext */
 	T("/xxxx/xxxx/x.ko.gz", true     , true      , true, true, "[x]", "gz");
 	T("/xxxx/xxxx/x.ko.gz", false    , true      , true, true, NULL , "gz");
 	T("/xxxx/xxxx/x.ko.gz", true     , false     , true, true, "[x]", NULL);
 	T("/xxxx/xxxx/x.ko.gz", false    , false     , true, true, NULL , NULL);
+	M("/xxxx/xxxx/x.ko.gz", PERF_RECORD_MISC_CPUMODE_UNKNOWN, true);
+	M("/xxxx/xxxx/x.ko.gz", PERF_RECORD_MISC_KERNEL, true);
+	M("/xxxx/xxxx/x.ko.gz", PERF_RECORD_MISC_USER, false);
 
 	/* path              alloc_name  alloc_ext  kmod   comp  name    ext */
 	T("/xxxx/xxxx/x.gz", true      , true     , false, true, "x.gz" ,"gz");
 	T("/xxxx/xxxx/x.gz", false     , true     , false, true, NULL   ,"gz");
 	T("/xxxx/xxxx/x.gz", true      , false    , false, true, "x.gz" , NULL);
 	T("/xxxx/xxxx/x.gz", false     , false    , false, true, NULL   , NULL);
+	M("/xxxx/xxxx/x.gz", PERF_RECORD_MISC_CPUMODE_UNKNOWN, false);
+	M("/xxxx/xxxx/x.gz", PERF_RECORD_MISC_KERNEL, false);
+	M("/xxxx/xxxx/x.gz", PERF_RECORD_MISC_USER, false);
 
 	/* path   alloc_name  alloc_ext  kmod   comp  name     ext */
 	T("x.gz", true      , true     , false, true, "x.gz", "gz");
 	T("x.gz", false     , true     , false, true, NULL  , "gz");
 	T("x.gz", true      , false    , false, true, "x.gz", NULL);
 	T("x.gz", false     , false    , false, true, NULL  , NULL);
+	M("x.gz", PERF_RECORD_MISC_CPUMODE_UNKNOWN, false);
+	M("x.gz", PERF_RECORD_MISC_KERNEL, false);
+	M("x.gz", PERF_RECORD_MISC_USER, false);
 
 	/* path      alloc_name  alloc_ext  kmod  comp  name  ext */
 	T("x.ko.gz", true      , true     , true, true, "[x]", "gz");
 	T("x.ko.gz", false     , true     , true, true, NULL , "gz");
 	T("x.ko.gz", true      , false    , true, true, "[x]", NULL);
 	T("x.ko.gz", false     , false    , true, true, NULL , NULL);
+	M("x.ko.gz", PERF_RECORD_MISC_CPUMODE_UNKNOWN, true);
+	M("x.ko.gz", PERF_RECORD_MISC_KERNEL, true);
+	M("x.ko.gz", PERF_RECORD_MISC_USER, false);
+
+	/* path            alloc_name  alloc_ext  kmod  comp   name             ext */
+	T("[test_module]", true      , true     , true, false, "[test_module]", NULL);
+	T("[test_module]", false     , true     , true, false, NULL           , NULL);
+	T("[test_module]", true      , false    , true, false, "[test_module]", NULL);
+	T("[test_module]", false     , false    , true, false, NULL           , NULL);
+	M("[test_module]", PERF_RECORD_MISC_CPUMODE_UNKNOWN, true);
+	M("[test_module]", PERF_RECORD_MISC_KERNEL, true);
+	M("[test_module]", PERF_RECORD_MISC_USER, false);
+
+	/* path            alloc_name  alloc_ext  kmod  comp   name             ext */
+	T("[test.module]", true      , true     , true, false, "[test.module]", NULL);
+	T("[test.module]", false     , true     , true, false, NULL           , NULL);
+	T("[test.module]", true      , false    , true, false, "[test.module]", NULL);
+	T("[test.module]", false     , false    , true, false, NULL           , NULL);
+	M("[test.module]", PERF_RECORD_MISC_CPUMODE_UNKNOWN, true);
+	M("[test.module]", PERF_RECORD_MISC_KERNEL, true);
+	M("[test.module]", PERF_RECORD_MISC_USER, false);
+
+	/* path     alloc_name  alloc_ext  kmod   comp   name      ext */
+	T("[vdso]", true      , true     , false, false, "[vdso]", NULL);
+	T("[vdso]", false     , true     , false, false, NULL    , NULL);
+	T("[vdso]", true      , false    , false, false, "[vdso]", NULL);
+	T("[vdso]", false     , false    , false, false, NULL    , NULL);
+	M("[vdso]", PERF_RECORD_MISC_CPUMODE_UNKNOWN, false);
+	M("[vdso]", PERF_RECORD_MISC_KERNEL, false);
+	M("[vdso]", PERF_RECORD_MISC_USER, false);
+
+	/* path         alloc_name  alloc_ext  kmod   comp   name          ext */
+	T("[vsyscall]", true      , true     , false, false, "[vsyscall]", NULL);
+	T("[vsyscall]", false     , true     , false, false, NULL        , NULL);
+	T("[vsyscall]", true      , false    , false, false, "[vsyscall]", NULL);
+	T("[vsyscall]", false     , false    , false, false, NULL        , NULL);
+	M("[vsyscall]", PERF_RECORD_MISC_CPUMODE_UNKNOWN, false);
+	M("[vsyscall]", PERF_RECORD_MISC_KERNEL, false);
+	M("[vsyscall]", PERF_RECORD_MISC_USER, false);
+
+	/* path                alloc_name  alloc_ext  kmod   comp   name      ext */
+	T("[kernel.kallsyms]", true      , true     , false, false, "[kernel.kallsyms]", NULL);
+	T("[kernel.kallsyms]", false     , true     , false, false, NULL               , NULL);
+	T("[kernel.kallsyms]", true      , false    , false, false, "[kernel.kallsyms]", NULL);
+	T("[kernel.kallsyms]", false     , false    , false, false, NULL               , NULL);
+	M("[kernel.kallsyms]", PERF_RECORD_MISC_CPUMODE_UNKNOWN, false);
+	M("[kernel.kallsyms]", PERF_RECORD_MISC_KERNEL, false);
+	M("[kernel.kallsyms]", PERF_RECORD_MISC_USER, false);
 
 	return 0;
 }
tools/perf/util/dso.c +42 -5

66 lines not on this page · raw message

tools/perf/util/dso.h +1 -1
index 24a507a54147..ba2d90ed881f 100644
@@ -220,7 +220,7 @@ char dso__symtab_origin(const struct dso *dso);
 int dso__read_binary_type_filename(const struct dso *dso, enum dso_binary_type type,
 				   char *root_dir, char *filename, size_t size);
 bool is_supported_compression(const char *ext);
-bool is_kernel_module(const char *pathname);
+bool is_kernel_module(const char *pathname, int cpumode);
 bool decompress_to_file(const char *ext, const char *filename, int output_fd);
 bool dso__needs_decompress(struct dso *dso);
 
tools/perf/util/header.c +4 -4
index 851143a7988d..ac5aaaeed7ff 100644
@@ -1239,7 +1239,7 @@ static int __event_process_build_id(struct build_id_event *bev,
 {
 	int err = -1;
 	struct machine *machine;
-	u16 misc;
+	u16 cpumode;
 	struct dso *dso;
 	enum dso_kernel_type dso_type;
 
@@ -1247,9 +1247,9 @@ static int __event_process_build_id(struct build_id_event *bev,
 	if (!machine)
 		goto out;
 
-	misc = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
+	cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 
-	switch (misc) {
+	switch (cpumode) {
 	case PERF_RECORD_MISC_KERNEL:
 		dso_type = DSO_TYPE_KERNEL;
 		break;
@@ -1270,7 +1270,7 @@ static int __event_process_build_id(struct build_id_event *bev,
 
 		dso__set_build_id(dso, &bev->build_id);
 
-		if (!is_kernel_module(filename))
+		if (!is_kernel_module(filename, cpumode))
 			dso->kernel = dso_type;
 
 		build_id__sprintf(dso->build_id, sizeof(dso->build_id),
tools/perf/util/machine.c +21 -1
index 4e29e80932e5..9e02c86f39f5 100644
@@ -1149,9 +1149,29 @@ static int machine__process_kernel_mmap_event(struct machine *machine,
 		struct dso *dso;
 
 		list_for_each_entry(dso, &machine->dsos.head, node) {
-			if (!dso->kernel || is_kernel_module(dso->long_name))
+
+			/*
+			 * The cpumode passed to is_kernel_module is not the
+			 * cpumode of *this* event. If we insist on passing
+			 * correct cpumode to is_kernel_module, we should
+			 * record the cpumode when we adding this dso to the
+			 * linked list.
+			 *
+			 * However we don't really need passing correct
+			 * cpumode.  We know the correct cpumode must be kernel
+			 * mode (if not, we should not link it onto kernel_dsos
+			 * list).
+			 *
+			 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
+			 * is_kernel_module() treats it as a kernel cpumode.
+			 */
+
+			if (!dso->kernel ||
+			    is_kernel_module(dso->long_name,
+					     PERF_RECORD_MISC_CPUMODE_UNKNOWN))
 				continue;
 
+
 			kernel = dso;
 			break;
 		}
Signature · 1 line
2.1.0
#7 lore raw
In reply to Arnaldo Carvalho de Melo
Quoted 11 lines · Arnaldo Carvalho de Melo

Hi Ingo,

Please consider applying.

One of the next requests probably will have the eBPF work by Wang Nan,
but I am still going thru it and want to test it thoroughly.

BTW: Have you looked at it lately? It is at:

http://lkml.kernel.org/r/[email protected]

Quoted 8 lines

Super summary from the above cover letter:

---------------------
It enables 'perf record' to filter events using eBPF programs like:

 # perf record --event bpf-file.o sleep 1

Events are selected and filtered according to definitions in bpf-file.o.

Looks useful, but I think the UI needs one more tweak: could you fix it to be able to filter based on the eBPF _source_ file, not just the object file?

People want to tweak such filters as they profile, so we should use the eBPF source code as the primary interface. We can compile it internally to the .o just fine. The .o file is a totally uninteresting intermediate product in itself.

I.e. we need to first think through such profiling workflows from beginning to end before allowing them upstream.

Quoted 63 lines

---------------------

The first two patches from that series are in this pull req, as
they just move stuff into tools/include/linux/ from tools/perf/include.

Regards,

- Arnaldo

The following changes since commit 5c9b9bc67c684e40b3a5e7e9facde0fb7200cd8c:

Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core (2015-05-29 20:19:02 +0200)

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git tags/perf-core-for-mingo

for you to fetch changes up to 1f121b03d058dd07199d8924373d3c52a207f63b:

perf tools: Deal with kernel module names in '[]' correctly (2015-06-03 10:02:38 -0300)

----------------------------------------------------------------
perf/core improvements and fixes:

User visible:

- Fix 'perf probe' segfault when glob matching function without debuginfo (Wang Nan)

- Remove newline char when reading event scale and unit (Madhavan Srinivasan)

- Deal with kernel module names in '[]' correctly (Wang Nan)

Infrastructure:

- Fix the search for the kernel DSO on the unified list (Arnaldo Carvalho de Melo)

- Move tools/perf/util/include/linux/{kernel.h,list.h,poison.h} to tools/include, to be used in tools/lib/bpf/ (Wang Nan)

----------------------------------------------------------------
Arnaldo Carvalho de Melo (1):
perf machine: Fix the search for the kernel DSO on the unified list

Madhavan Srinivasan (1):
perf tools: Remove newline char when reading event scale and unit

Wang Nan (4):
perf probe: Fix segfault when glob matching function without debuginfo perf tools: Move linux/kernel.h to tools/include
tools: Move tools/perf/util/include/linux/{list.h,poison.h} to tools/include perf tools: Deal with kernel module names in '[]' correctly

 tools/{perf/util => }/include/linux/kernel.h |  4 +-
 tools/{perf/util => }/include/linux/list.h   |  6 +--
 tools/include/linux/poison.h                 |  1 +
 tools/perf/MANIFEST                          |  3 ++
 tools/perf/tests/kmod-path.c                 | 72 ++++++++++++++++++++++++++++
 tools/perf/util/dso.c                        | 47 ++++++++++++++++--
 tools/perf/util/dso.h                        |  2 +-
 tools/perf/util/header.c                     |  8 ++--
 tools/perf/util/include/linux/poison.h       |  1 -
Quoted 8 lines
 tools/perf/util/machine.c                    | 22 ++++++++-
 tools/perf/util/pmu.c                        | 11 ++++-
 tools/perf/util/probe-event.c                | 26 ++++++++--
 12 files changed, 179 insertions(+), 24 deletions(-)
 rename tools/{perf/util => }/include/linux/kernel.h (97%)
 rename tools/{perf/util => }/include/linux/list.h (90%)
 create mode 100644 tools/include/linux/poison.h
 delete mode 100644 tools/perf/util/include/linux/poison.h

Pulled, thanks a lot Arnaldo!

Ingo

#8 lore raw
Quoted 24 lines · Ingo Molnar
Quoted 11 lines · Arnaldo Carvalho de Melo

Hi Ingo,

Please consider applying.

One of the next requests probably will have the eBPF work by Wang Nan,
but I am still going thru it and want to test it thoroughly.

BTW: Have you looked at it lately? It is at:

http://lkml.kernel.org/r/[email protected]

Quoted 8 lines

Super summary from the above cover letter:

---------------------
It enables 'perf record' to filter events using eBPF programs like:

  # perf record --event bpf-file.o sleep 1

Events are selected and filtered according to definitions in bpf-file.o.

Looks useful, but I think the UI needs one more tweak: could you fix it to be able to filter based on the eBPF _source_ file, not just the object file?

Quoted 6 lines

People want to tweak such filters as they profile, so we should use the eBPF source code as the primary interface. We can compile it internally to the .o just fine. The .o file is a totally uninteresting intermediate product in itself.

I.e. we need to first think through such profiling workflows from beginning to end before allowing them upstream.

In a private mail Alexei Starovoitov disscussed with me about this. He said that he is working on a shared object which can compile C program into BPF bytecode on the fly. After he done his work, I think perf can support dtrace-like profiling that, users will be able to feed source code to perf directly on cmdline. He said he can release it on June. I added him to the CC-list.

However I think the '.o' intermediate is still needed. I'd like to share a real profiling experience using eBPF today, please keep an eye on it. In my experience, since we are using C instead of dtrace, the code piece could be relative complex. Therefore, even if perf is able to compile the C source on the fly, I think user still need to transfer the profiling scripts to the target machine. Therefor, for him, precompiling and do some debugging on a high-end server then transfer it into target machine (like a smartphone) is tolerable, and useful for me.

Thank you.

Quoted 74 lines
Quoted 67 lines

---------------------

The first two patches from that series are in this pull req, as
they just move stuff into tools/include/linux/ from tools/perf/include.

Regards,

- Arnaldo

The following changes since commit 5c9b9bc67c684e40b3a5e7e9facde0fb7200cd8c:

   Merge tag 'perf-core-for-mingo' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core (2015-05-29 20:19:02 +0200)

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git tags/perf-core-for-mingo

for you to fetch changes up to 1f121b03d058dd07199d8924373d3c52a207f63b:

perf tools: Deal with kernel module names in '[]' correctly (2015-06-03 10:02:38 -0300)

----------------------------------------------------------------
perf/core improvements and fixes:

User visible:

- Fix 'perf probe' segfault when glob matching function without debuginfo (Wang Nan)

- Remove newline char when reading event scale and unit (Madhavan Srinivasan)

- Deal with kernel module names in '[]' correctly (Wang Nan)

Infrastructure:

- Fix the search for the kernel DSO on the unified list (Arnaldo Carvalho de Melo)

- Move tools/perf/util/include/linux/{kernel.h,list.h,poison.h} to tools/include, to be used in tools/lib/bpf/ (Wang Nan)

----------------------------------------------------------------
Arnaldo Carvalho de Melo (1):
perf machine: Fix the search for the kernel DSO on the unified list

Madhavan Srinivasan (1):
perf tools: Remove newline char when reading event scale and unit

Wang Nan (4):
perf probe: Fix segfault when glob matching function without debuginfo perf tools: Move linux/kernel.h to tools/include
tools: Move tools/perf/util/include/linux/{list.h,poison.h} to tools/include perf tools: Deal with kernel module names in '[]' correctly

  tools/{perf/util => }/include/linux/kernel.h |  4 +-
  tools/{perf/util => }/include/linux/list.h   |  6 +--
  tools/include/linux/poison.h                 |  1 +
  tools/perf/MANIFEST                          |  3 ++
  tools/perf/tests/kmod-path.c                 | 72 ++++++++++++++++++++++++++++
  tools/perf/util/dso.c                        | 47 ++++++++++++++++--
  tools/perf/util/dso.h                        |  2 +-
  tools/perf/util/header.c                     |  8 ++--
  tools/perf/util/include/linux/poison.h       |  1 -
  tools/perf/util/machine.c                    | 22 ++++++++-
  tools/perf/util/pmu.c                        | 11 ++++-
  tools/perf/util/probe-event.c                | 26 ++++++++--
  12 files changed, 179 insertions(+), 24 deletions(-)
Quoted 4 lines

rename tools/{perf/util => }/include/linux/kernel.h (97%) rename tools/{perf/util => }/include/linux/list.h (90%) create mode 100644 tools/include/linux/poison.h delete mode 100644 tools/perf/util/include/linux/poison.h

Pulled, thanks a lot Arnaldo!

Ingo

#9 lore raw
Quoted 32 lines · Wangnan (F)
Quoted 24 lines · Ingo Molnar
Quoted 11 lines · Arnaldo Carvalho de Melo

Hi Ingo,

Please consider applying.

One of the next requests probably will have the eBPF work by Wang Nan,
but I am still going thru it and want to test it thoroughly.

BTW: Have you looked at it lately? It is at:

http://lkml.kernel.org/r/[email protected]

Quoted 8 lines

Super summary from the above cover letter:

---------------------
It enables 'perf record' to filter events using eBPF programs like:

 # perf record --event bpf-file.o sleep 1

Events are selected and filtered according to definitions in bpf-file.o.

Looks useful, but I think the UI needs one more tweak: could you fix it to be able to filter based on the eBPF _source_ file, not just the object file?

Quoted 6 lines

People want to tweak such filters as they profile, so we should use the eBPF source code as the primary interface. We can compile it internally to the .o just fine. The .o file is a totally uninteresting intermediate product in itself.

I.e. we need to first think through such profiling workflows from beginning to end before allowing them upstream.

Quoted 7 lines

In a private mail Alexei Starovoitov disscussed with me about this. He said that he is working on a shared object which can compile C program into BPF bytecode on the fly. After he done his work, I think perf can support dtrace-like profiling that, users will be able to feed source code to perf directly on cmdline. He said he can release it on June. I added him to the CC-list.

However I think the '.o' intermediate is still needed. [...]

So how do you generate the .o? Why cannot the tool, if it sees that the filter parameter is eBPF source code, do that automatically?

I.e. you are making the user jump through hoops for no good reason - that's a bad UI and a bad workflow. Please don't do that!

Thanks,

Ingo

#10 lore raw
Quoted 39 lines · Ingo Molnar
Quoted 37 lines · Wangnan (F)
Quoted 24 lines · Ingo Molnar
Quoted 11 lines · Arnaldo Carvalho de Melo

Hi Ingo,

Please consider applying.

One of the next requests probably will have the eBPF work by Wang Nan,
but I am still going thru it and want to test it thoroughly.

BTW: Have you looked at it lately? It is at:

http://lkml.kernel.org/r/[email protected]

Quoted 8 lines

Super summary from the above cover letter:

---------------------
It enables 'perf record' to filter events using eBPF programs like:

  # perf record --event bpf-file.o sleep 1

Events are selected and filtered according to definitions in bpf-file.o.

Looks useful, but I think the UI needs one more tweak: could you fix it to be able to filter based on the eBPF _source_ file, not just the object file?

Quoted 6 lines

People want to tweak such filters as they profile, so we should use the eBPF source code as the primary interface. We can compile it internally to the .o just fine. The .o file is a totally uninteresting intermediate product in itself.

I.e. we need to first think through such profiling workflows from beginning to end before allowing them upstream.

In a private mail Alexei Starovoitov disscussed with me about this. He said that he is working on a shared object which can compile C program into BPF bytecode on the fly. After he done his work, I think perf can support dtrace-like profiling that, users will be able to feed source code to perf directly on cmdline. He said he can release it on June. I added him to the CC-list.

Quoted 3 lines
Quoted 1 line

However I think the '.o' intermediate is still needed. [...]

So how do you generate the .o? Why cannot the tool, if it sees that the filter parameter is eBPF source code, do that automatically?

I think compiling on the fly is our goal, and Alexei is working on it. However, it looks like some limitations are still exist. For example, the BPF program may require the definition of kernel structure for it to derefrernce struct pointers. Therefore, the machine on which compiling occures should have kernel header and configuration installed. However, most of production environment doesn't have them for security and reasons.

I think BPF should be used to profile not only development environment but also production systems. If we only concern development systems, why not recompiling kernel with stub core or using kProbe modules?

What about this:

Give perf the ability to generate the '.o' file (to memory or dump to a real file) on development environment. If we are willing to profiling that system, we can make load the generated object into kernel and start trace by (perf record --event bpf_file.c ...). If we are willing to profile systems on which the compiling environment is missing, we can make perf to create a '.o' file (perf bpf compile bpf_file.c -o bpf_file.o) then copy it onto target environment, use perf record --event bpf_file.o ... to filter events.

Thank you.

Quoted 6 lines

I.e. you are making the user jump through hoops for no good reason - that's a bad UI and a bad workflow. Please don't do that!

Thanks,

Ingo

#11 lore raw
Quoted 4 lines · Wangnan (F)
Quoted 2 lines

So how do you generate the .o? Why cannot the tool, if it sees that the filter parameter is eBPF source code, do that automatically?

I think compiling on the fly is our goal, and Alexei is working on it.

So what exact command line are you using to create the .o?

What exactly should users type to create a simple eBPF filter profile?

Thanks,

Ingo

#12 lore raw

发自我的 iPhone

Quoted 5 lines · 在 2015年6月4日,下午8:40,Ingo Molnar <[email protected]> 写道:

* Wangnan (F) <[email protected]> wrote:

Quoted 9 lines
Quoted 4 lines
Quoted 2 lines

So how do you generate the .o? Why cannot the tool, if it sees that the filter parameter is eBPF source code, do that automatically?

I think compiling on the fly is our goal, and Alexei is working on it.

So what exact command line are you using to create the .o?

What exactly should users type to create a simple eBPF filter profile?

I have mentioned in previous mail:

Use

 # perf record -e bpf_source.c cmdline 

to create a eBPF filter from source,

Use

# perf record -e bpf_object.o cmdline 

to create a eBPF filter from object intermedia.

Use

# perf bpf compile bpf_source.c --kbuild=kernel-build-dir -o bpf_object.o

to create the .o

I think this should be enough. Currently only the second case has been implemented.

Thanks.

Quoted 3 lines

Thanks,

Ingo

#13 lore raw
Quoted 32 lines · * pi3orama <[email protected]> wrote:

发自我的 iPhone

Quoted 5 lines · 在 2015年6月4日,下午8:40,Ingo Molnar <[email protected]> 写道:

* Wangnan (F) <[email protected]> wrote:

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

So how do you generate the .o? Why cannot the tool, if it sees that the filter parameter is eBPF source code, do that automatically?

I think compiling on the fly is our goal, and Alexei is working on it.

So what exact command line are you using to create the .o?

What exactly should users type to create a simple eBPF filter profile?

I have mentioned in previous mail:

Use

 # perf record -e bpf_source.c cmdline 

to create a eBPF filter from source,

Use

# perf record -e bpf_object.o cmdline 

to create a eBPF filter from object intermedia.

Quoted 7 lines

Use

# perf bpf compile bpf_source.c --kbuild=kernel-build-dir -o bpf_object.o

to create the .o

I think this should be enough. Currently only the second case has been implemented.

So if users cannot actually generate .o files then it's premature to merge this in such an incomplete form!

It should be possible to use a feature that we are merging.

Thanks,

Ingo

#14 lore raw
Quoted 14 lines · Ingo Molnar
Quoted 14 lines
Quoted 14 lines
 # perf record -e bpf_source.c cmdline

to create a eBPF filter from source,

Use

# perf record -e bpf_object.o cmdline

to create a eBPF filter from object intermedia.

Use

# perf bpf compile bpf_source.c --kbuild=kernel-build-dir -o bpf_object.o
Quoted 7 lines
Quoted 3 lines
Quoted 3 lines

to create the .o

I think this should be enough. Currently only the second case has been implemented.

So if users cannot actually generate .o files then it's premature to merge this in such an incomplete form!

It should be possible to use a feature that we are merging.

of course it's usable :) There is some confusion here. To compile .c into .o one can easily use
clang -O2 -emit-llvm -c file.c -o - | llc -march=bpf -o file.o any version of clang is ok,
llc needs to be fresh with bpf backend.

For a lot of cases kernel headers are not needed, so above will work fine.
For our TC examples we recommend to use 'bcc' alias:
bcc() {
clang -O2 -emit-llvm -c $1 -o - | llc -march=bpf -filetype=obj -o
"basename $1 .c.o"
}
then compiling as easy as 'bcc file.c'

What Wang mentioned that we're working on is fully integrated 'bcc'. It will use clang/llvm as libraries, so no intermediate steps will be needed, but some folks will always have concerns about ultra-embedded environments where even 20Mb of libllvm.so is too much.

So I think we need to support both 'perf record -e file.[co]'

#15 lore raw
Quoted 38 lines · Alexei Starovoitov
Quoted 14 lines · Ingo Molnar
Quoted 14 lines
Quoted 14 lines
 # perf record -e bpf_source.c cmdline

to create a eBPF filter from source,

Use

# perf record -e bpf_object.o cmdline

to create a eBPF filter from object intermedia.

Use

# perf bpf compile bpf_source.c --kbuild=kernel-build-dir -o bpf_object.o
Quoted 7 lines
Quoted 3 lines
Quoted 3 lines

to create the .o

I think this should be enough. Currently only the second case has been implemented.

So if users cannot actually generate .o files then it's premature to merge this in such an incomplete form!

It should be possible to use a feature that we are merging.

of course it's usable :) There is some confusion here. To compile .c into .o one can easily use
clang -O2 -emit-llvm -c file.c -o - | llc -march=bpf -o file.o any version of clang is ok,
llc needs to be fresh with bpf backend.

For a lot of cases kernel headers are not needed, so above will work fine.
For our TC examples we recommend to use 'bcc' alias:
bcc() {
clang -O2 -emit-llvm -c $1 -o - | llc -march=bpf -filetype=obj -o
"basename $1 .c.o"
}
then compiling as easy as 'bcc file.c'

Quoted 6 lines

What Wang mentioned that we're working on is fully integrated 'bcc'. It will use clang/llvm as libraries, so no intermediate steps will be needed, but some folks will always have concerns about ultra-embedded environments where even 20Mb of libllvm.so is too much.

So I think we need to support both 'perf record -e file.[co]'

I think we'd better make 'perf record -e file.c' default and '-e file.o' should be an option.

Thank you,

Quoted 2 lines
Signature · 5 lines
Masami HIRAMATSU
Linux Technology Research Center, System Productivity Research Dept.
Center for Technology Innovation - Systems Engineering
Hitachi, Ltd., Research & Development Group
E-mail: [email protected]
#16 lore raw
Quoted 44 lines · Masami Hiramatsu
Quoted 43 lines · Alexei Starovoitov
Quoted 14 lines · Ingo Molnar
Quoted 14 lines
Quoted 14 lines
  # perf record -e bpf_source.c cmdline

to create a eBPF filter from source,

Use

# perf record -e bpf_object.o cmdline

to create a eBPF filter from object intermedia.

Use

# perf bpf compile bpf_source.c --kbuild=kernel-build-dir -o bpf_object.o
Quoted 7 lines
Quoted 3 lines
Quoted 3 lines

to create the .o

I think this should be enough. Currently only the second case has been implemented.

So if users cannot actually generate .o files then it's premature to merge this in such an incomplete form!

It should be possible to use a feature that we are merging.

of course it's usable :) There is some confusion here. To compile .c into .o one can easily use
clang -O2 -emit-llvm -c file.c -o - | llc -march=bpf -o file.o any version of clang is ok,
llc needs to be fresh with bpf backend.

For a lot of cases kernel headers are not needed, so above will work fine.
For our TC examples we recommend to use 'bcc' alias:
bcc() {
clang -O2 -emit-llvm -c $1 -o - | llc -march=bpf -filetype=obj -o
"basename $1 .c.o"
}
then compiling as easy as 'bcc file.c'

What Wang mentioned that we're working on is fully integrated 'bcc'. It will use clang/llvm as libraries, so no intermediate steps will be needed, but some folks will always have concerns about ultra-embedded environments where even 20Mb of libllvm.so is too much.

Quoted 4 lines
Quoted 1 line

So I think we need to support both 'perf record -e file.[co]'

I think we'd better make 'perf record -e file.c' default and '-e file.o' should be an option.

what do you mean 'default' ? It's a command line :) .c is easier to use of course, no question.

#17 lore raw
In reply to Alexei Starovoitov
Quoted 19 lines · Alexei Starovoitov
Quoted 18 lines · Ingo Molnar
Quoted 17 lines
Quoted 17 lines
 # perf record -e bpf_source.c cmdline

to create a eBPF filter from source,

Use

# perf record -e bpf_object.o cmdline

to create a eBPF filter from object intermedia.

Use

# perf bpf compile bpf_source.c --kbuild=kernel-build-dir -o bpf_object.o

to create the .o

I think this should be enough. Currently only the second case has been implemented.

Quoted 8 lines
Quoted 4 lines

So if users cannot actually generate .o files then it's premature to merge this in such an incomplete form!

It should be possible to use a feature that we are merging.

of course it's usable :) There is some confusion here. To compile .c into .o one can easily use
clang -O2 -emit-llvm -c file.c -o - | llc -march=bpf -o file.o

There's no confusion here: you guys are trying to sell me what at this stage is incomplete and hard to use, and I'm resisting it as I should! :-)

We also have different definitions of 'easily'. It might be 'easy' to type:

clang -O2 -emit-llvm -c file.c -o - | llc -march=bpf -o file.o

... for some tooling developer intimate with eBPF, but to the first time user who found an interesting looking eBPF scriptlet on the net or in the documentation and wants to try his luck? It's absolutely non-obvious!

The current usage to get a _minimal_ eBPF script running is non-obvious and obscure to the level of being a show stopper.

I don't understand why you guys are even wasting time arguing about it: it's not that hard to auto-build from source code. It's one of the basic features of tooling. If you ever built perf you'll know that typing 'make install' will type in all those quirky build lines automatically for you, without requiring you to perform any other step, no matter how trivial.

Doubly annoying, you seem to have the UI principles wrong, you seem to think that a .o is a proper user interface. It absolutely is _not_ okay.

The Linux kernel project and as an extension the perf project deals with source code, and I'm 100% suspicious of approaches that somehow think that .o objects are the right UI for _anything_ except temporary files that sometimes show up in object directories...

Fix the 'newbie user' UI flow as a _first_ priority, not as a second thought!

Every single quirky line or nonsensical option you require a first time user to type halves the number of new users we'll get. You need to understand why dtrace is so popular:

- it's bloody easy to use

- it's a safe environment you can deploy in critical environments

- it's flexible

- instrumentation hacks are very easy to share

eBPF based scripting got 3 out of those 4 right, but please don't forget item 1 either, because without that we have nothing but a bunch of unusable functionality in the kernel and in tooling that benefits only very few people. Okay?

Quoted 1 line

So I think we need to support both 'perf record -e file.[co]'

Why do you even need to ask? Of course!

Think through how users will meet eBPF scripts and how they will interact with them:

- they'll see or download an eBPF scriptlet somewhere and will have a .c file.

- ideally there will be built-in eBPF scriptlets just like we have tracing plugins, and there's a good UI to query them and see their description and source code.

- then they will want to use it all with the minimum amount of fuss

- they don't care how the eBPF scriptlet gets to the kernel: whether the kernel can read and build the .c files, or whether there's some user tooling that turns it into bytecode. Most humans don't read bytecode!

- they will absolutely not download random .o's and we should not encourage that in any case - these things should be source code based.

These things compile in an eye blink, there's very little reason to ever deal with a .o, except some weird and rare usecases...

In fact I'm NAK-ing the whole .o based interface until the .c interface is made the _primary_ one and works well and until I see that you have thought through basic usability questions...

Thanks,

Ingo

#18 lore raw
Quoted 94 lines · Ingo Molnar
Quoted 17 lines · Alexei Starovoitov
Quoted 16 lines · Ingo Molnar
Quoted 16 lines
Quoted 16 lines
  # perf record -e bpf_source.c cmdline

to create a eBPF filter from source,

Use

# perf record -e bpf_object.o cmdline

to create a eBPF filter from object intermedia.

Use

# perf bpf compile bpf_source.c --kbuild=kernel-build-dir -o bpf_object.o

to create the .o

Quoted 8 lines
Quoted 5 lines
Quoted 1 line
Quoted 1 line

I think this should be enough. Currently only the second case has been implemented.

So if users cannot actually generate .o files then it's premature to merge this in such an incomplete form!

It should be possible to use a feature that we are merging.

of course it's usable :) There is some confusion here. To compile .c into .o one can easily use
clang -O2 -emit-llvm -c file.c -o - | llc -march=bpf -o file.o

There's no confusion here: you guys are trying to sell me what at this stage is incomplete and hard to use, and I'm resisting it as I should! :-)

We also have different definitions of 'easily'. It might be 'easy' to type:

clang -O2 -emit-llvm -c file.c -o - | llc -march=bpf -o file.o

... for some tooling developer intimate with eBPF, but to the first time user who found an interesting looking eBPF scriptlet on the net or in the documentation and wants to try his luck? It's absolutely non-obvious!

The current usage to get a _minimal_ eBPF script running is non-obvious and obscure to the level of being a show stopper.

I don't understand why you guys are even wasting time arguing about it: it's not that hard to auto-build from source code. It's one of the basic features of tooling. If you ever built perf you'll know that typing 'make install' will type in all those quirky build lines automatically for you, without requiring you to perform any other step, no matter how trivial.

Doubly annoying, you seem to have the UI principles wrong, you seem to think that a .o is a proper user interface. It absolutely is _not_ okay.

The Linux kernel project and as an extension the perf project deals with source code, and I'm 100% suspicious of approaches that somehow think that .o objects are the right UI for _anything_ except temporary files that sometimes show up in object directories...

Fix the 'newbie user' UI flow as a _first_ priority, not as a second thought!

Every single quirky line or nonsensical option you require a first time user to type halves the number of new users we'll get. You need to understand why dtrace is so popular:

- it's bloody easy to use

- it's a safe environment you can deploy in critical environments

- it's flexible

- instrumentation hacks are very easy to share

eBPF based scripting got 3 out of those 4 right, but please don't forget item 1 either, because without that we have nothing but a bunch of unusable functionality in the kernel and in tooling that benefits only very few people. Okay?

Quoted 1 line

So I think we need to support both 'perf record -e file.[co]'

Why do you even need to ask? Of course!

Think through how users will meet eBPF scripts and how they will interact with them:

- they'll see or download an eBPF scriptlet somewhere and will have a .c file.

- ideally there will be built-in eBPF scriptlets just like we have tracing plugins, and there's a good UI to query them and see their description and source code.

- then they will want to use it all with the minimum amount of fuss

- they don't care how the eBPF scriptlet gets to the kernel: whether the kernel can read and build the .c files, or whether there's some user tooling that turns it into bytecode. Most humans don't read bytecode!

- they will absolutely not download random .o's and we should not encourage that in any case - these things should be source code based.

Quoted 6 lines

These things compile in an eye blink, there's very little reason to ever deal with a .o, except some weird and rare usecases...

In fact I'm NAK-ing the whole .o based interface until the .c interface is made the _primary_ one and works well and until I see that you have thought through basic usability questions...

OK. Let's start making a nice UI.

At this stage, what about wrapping current clang and llc workflow into perf, let it call them to compile '.c' scripts? This is the way 'perf annotate' using objdump. I can do this job, but firstly I'd like to know people's opinion on it, and the value of the wrapper if Alexei Starovoitov's dynamic compiler shared object is coming.

Following functions will be added:

- perf searches clang and llc under current $PATH,

- Users are allowed to pass the position of those programs, if perf failed to find the automatically,

- Users are allowed to pass extra compiling options to clang and llc, like include directories,

- 'perf record' automatically calls them if a '.c' is passed using
'--event'.

- 'perf bpf compile' command will be added to compile a '.c' find into
'.o',

Further, basic header files should be shipped with kernel headers.

User interface update:

- --llc, --clang, --llc-opt and --clang-opt option will be added to
'perf record'
to indicate the position and extra options to them. Ideally they can
be leave
blank.

- 'perf bpf compile' will be added.

One problem I can find is that, the wrapper will make perf depend on llvm. I don't think the compiler will be deployed in production environments... And also, the embedded case...

Any suggestion? Do you think the above idea is on the right way?

Thanks.

Quoted 3 lines

Thanks,

Ingo

#19 lore raw
Quoted 1 line · Wangnan (F)

OK. Let's start making a nice UI.

Thanks!

Quoted 5 lines

[...]

One problem I can find is that, the wrapper will make perf depend on llvm. I don't think the compiler will be deployed in production environments... And also, the embedded case...

What dependencies are there?

On the usage side there should be very few outright dependencies: if the llvm binary is not available, or doesn't support what you need, or there's no runtime environment you can use to build the bytecode, you should display an informative error message so that the user knows what is missing and how to install it.

Thanks,

Ingo

author maintainer
#20 lore raw

Em Fri, Jun 05, 2015 at 02:05:50PM +0200, Ingo Molnar escreveu:

Quoted 1 line · Wangnan (F)

<SNIP>

Quoted 10 lines
Quoted 3 lines

One problem I can find is that, the wrapper will make perf depend on llvm. I don't think the compiler will be deployed in production environments... And also, the embedded case...

What dependencies are there?

On the usage side there should be very few outright dependencies: if the llvm binary is not available, or doesn't support what you need, or there's no runtime environment you can use to build the bytecode, you should display an informative error message so that the user knows what is missing and how to install it.

Right, something like:

[acme@zoo ~]$ perf trace -e nanosleep usleep 1
Error: No permissions to read /sys/kernel/debug/tracing/events/raw_syscalls/sys_(enter|exit)
Hint: Try 'sudo mount -o remount,mode=755 /sys/kernel/debug'
[acme@zoo ~]$ sudo mount -o remount,mode=755 /sys/kernel/debug [sudo] password for acme:
[acme@zoo ~]$ perf trace -e nanosleep usleep 1

     0.565 ( 0.060 ms): usleep/17648 nanosleep(rqtp: 0x7fff22baebf0) = 0

[acme@zoo ~]$ perf trace --all-cpus
Error: Operation not permitted.
Hint: Check /proc/sys/kernel/perf_event_paranoid setting.
Hint: For system wide tracing it needs to be set to -1.
Hint: Try: 'sudo sh -c "echo -1 > /proc/sys/kernel/perf_event_paranoid"'
Hint: The current value is 1.
[acme@zoo ~]$
[acme@zoo ~]$ trace -a -e poll usleep 1
[acme@zoo ~]$ trace -a -e poll usleep 1
0.041 ( 0.000 ms): firefox/1458 ... [continued]: poll()) = 1
0.267 ( 0.003 ms): firefox/1458 poll(ufds: 0x7f43d6ea1340, nfds: 5) = 0 Timeout 0.275 ( 0.001 ms): firefox/1458 poll(ufds: 0x7f43d6ea1340, nfds: 5) = 0 Timeout 0.283 ( 0.001 ms): firefox/1458 poll(ufds: 0x7f43d6ea1340, nfds: 5) = 0 Timeout 0.979 ( 0.000 ms): gnome-terminal/2572 ... [continued]: poll()) = 1
1.056 ( 0.768 ms): firefox/1458 poll(ufds: 0x7f43d6ea1340, nfds: 5, timeout_msecs: 4294967295) ... 1.065 ( 0.009 ms): gnome-terminal/2572 poll(ufds: 0x1934250, nfds: 23, timeout_msecs: 10) = 1 1.087 ( 0.007 ms): gnome-terminal/2572 poll(ufds: 0x1934250, nfds: 23, timeout_msecs: 10) = 2 1.132 ( 0.007 ms): gnome-terminal/2572 poll(ufds: 0x1934250, nfds: 23, timeout_msecs: 10) = 1 1.161 ( 0.013 ms): gnome-terminal/2572 poll(ufds: 0x1934250, nfds: 23, timeout_msecs: 10) = 1
[acme@zoo ~]$

I.e. Explain the mistake and provide a hint to solve it, as close to the actual commands needed to perform such corrective/enabling action as possible.

- Arnaldo

#21 lore raw
Quoted 45 lines · Arnaldo Carvalho de Melo

Em Fri, Jun 05, 2015 at 02:05:50PM +0200, Ingo Molnar escreveu:

Quoted 1 line · Wangnan (F)

<SNIP>

Quoted 10 lines
Quoted 3 lines

One problem I can find is that, the wrapper will make perf depend on llvm. I don't think the compiler will be deployed in production environments... And also, the embedded case...

What dependencies are there?

On the usage side there should be very few outright dependencies: if the llvm binary is not available, or doesn't support what you need, or there's no runtime environment you can use to build the bytecode, you should display an informative error message so that the user knows what is missing and how to install it.

Right, something like:

[acme@zoo ~]$ perf trace -e nanosleep usleep 1
Error: No permissions to read /sys/kernel/debug/tracing/events/raw_syscalls/sys_(enter|exit)
Hint: Try 'sudo mount -o remount,mode=755 /sys/kernel/debug'

[acme@zoo ~]$ sudo mount -o remount,mode=755 /sys/kernel/debug [sudo] password for acme:
[acme@zoo ~]$ perf trace -e nanosleep usleep 1

     0.565 ( 0.060 ms): usleep/17648 nanosleep(rqtp: 0x7fff22baebf0) = 0

[acme@zoo ~]$ perf trace --all-cpus
Error: Operation not permitted.
Hint: Check /proc/sys/kernel/perf_event_paranoid setting.
Hint: For system wide tracing it needs to be set to -1.
Hint: Try: 'sudo sh -c "echo -1 > /proc/sys/kernel/perf_event_paranoid"'
Hint: The current value is 1.
[acme@zoo ~]$
[acme@zoo ~]$ trace -a -e poll usleep 1
[acme@zoo ~]$ trace -a -e poll usleep 1
0.041 ( 0.000 ms): firefox/1458 ... [continued]: poll()) = 1
0.267 ( 0.003 ms): firefox/1458 poll(ufds: 0x7f43d6ea1340, nfds: 5) = 0 Timeout 0.275 ( 0.001 ms): firefox/1458 poll(ufds: 0x7f43d6ea1340, nfds: 5) = 0 Timeout 0.283 ( 0.001 ms): firefox/1458 poll(ufds: 0x7f43d6ea1340, nfds: 5) = 0 Timeout 0.979 ( 0.000 ms): gnome-terminal/2572 ... [continued]: poll()) = 1
1.056 ( 0.768 ms): firefox/1458 poll(ufds: 0x7f43d6ea1340, nfds: 5, timeout_msecs: 4294967295) ... 1.065 ( 0.009 ms): gnome-terminal/2572 poll(ufds: 0x1934250, nfds: 23, timeout_msecs: 10) = 1 1.087 ( 0.007 ms): gnome-terminal/2572 poll(ufds: 0x1934250, nfds: 23, timeout_msecs: 10) = 2 1.132 ( 0.007 ms): gnome-terminal/2572 poll(ufds: 0x1934250, nfds: 23, timeout_msecs: 10) = 1 1.161 ( 0.013 ms): gnome-terminal/2572 poll(ufds: 0x1934250, nfds: 23, timeout_msecs: 10) = 1
[acme@zoo ~]$

Quoted 2 lines

I.e. Explain the mistake and provide a hint to solve it, as close to the actual commands needed to perform such corrective/enabling action as possible.

Yeah, I absolutely love such tooling hints.

Thanks,

Ingo

author maintainer
#22 lore raw
In reply to Wangnan (F)

Em Fri, Jun 05, 2015 at 04:53:03PM +0800, Wangnan (F) escreveu:

Quoted 7 lines · Ingo Molnar
Quoted 5 lines · Alexei Starovoitov
Quoted 1 line · Ingo Molnar

In fact I'm NAK-ing the whole .o based interface until the .c interface is made the _primary_ one and works well and until I see that you have thought through basic usability questions...

Quoted 6 lines

OK. Let's start making a nice UI.

At this stage, what about wrapping current clang and llc workflow into perf, let it call them to compile '.c' scripts? This is the way 'perf annotate' using
objdump. I can do this job, but firstly I'd like to know people's opinion on

Right, no need for, at a first step, to save this into a cache, or use libraries, etc, just automate the bpf.c into bpf.o, load it and use it as an event.

Quoted 2 lines

it, and the value of the wrapper if Alexei Starovoitov's dynamic compiler shared object is coming.

No need to wait for that, when it comes we can use it, but Ingo established as the door for this to be accepted is that we could use:

perf record -e foo.c usleep

Right?

Quoted 3 lines

Following functions will be added:

- perf searches clang and llc under current $PATH,

Good for a first step

Quoted 2 lines

- Users are allowed to pass the position of those programs, if perf failed to find the automatically,

I would leave all this configurabilty for later, stating that it has to be in the PATH should be ok for a first step.

Quoted 2 lines

- Users are allowed to pass extra compiling options to clang and llc, like include directories,

For later too?

Quoted 2 lines

- 'perf record' automatically calls them if a '.c' is passed using
'--event'.

Right.

Quoted 2 lines

- 'perf bpf compile' command will be added to compile a '.c' find into
'.o',

for later? I.e. 'perf record -e foo.c usleep' could start by generating the foo.o and not deleting it, so:

perf record -e foo.c usleep

Followed by:

perf record -e foo.o usleep

Would work, the later would be like a quick hack so that we could have access to a pre-compiled foo.o quickly, at this introductory stage.

Quoted 9 lines

Further, basic header files should be shipped with kernel headers.

User interface update:

- --llc, --clang, --llc-opt and --clang-opt option will be added to 'perf
record'
to indicate the position and extra options to them. Ideally they can be
leave
blank.

Couldn't this be left to a section in a .perfconfig file to avoid having so many command line options? We could have one of those files per "project", after we add a --config option to 'perf record', to override whatever it finds in the config file search it already does, i.e.:

[clang]

path = /a/b/clang
opt = -a -b -c -d

[llc]

path = /d/e/llc
opt = -r -t -y -u

But even this can be left for a second step.

Quoted 5 lines

- 'perf bpf compile' will be added.

One problem I can find is that, the wrapper will make perf depend on llvm. I don't think the compiler will be deployed in production environments... And also, the embedded case...

Well, we can always build a subset of perf, using the command line options to disable certain features.

- Arnaldo

#23 lore raw

[EXPERIENCE] My experience on using perf record BPF filter on a real usecase

In reply to Wangnan (F)

Hi all,

I'd like to share my exprience on using 'perf record' BPF filter in a real usecase to show the power and shortcome in my patch series:

https://lkml.kernel.org/r/[email protected]

and other works on eBPF.

My usecase shows that such filter is useful. Also, I hope it can help us to find way to further improve it.

My task is to find the reason why iozone test result is bad on some specific cases. The development environment is a x86_64 server, the target machine is a smartphone with Android. By previous analysis I have already got some useful information:

1. iozone computes bandwidth by averaging time of each sys_write.

2. In our case, 1% sys_write takes 75% of total time, so what I need to do now should be finding the reason why those sys_write take so long.

3. By sampling call stack on sched:sched_switch, I find that those sys_write calls lock_page() and blocks on it.

I decide to use BPF filter to find the other side of this locking contention. The idea is simple:

1. For all calls of lock_page(), probe at entry and exit points of it. Measure the execution time of the lock_page() call. If it takes too long (longer than 0.1 second) then there should have a lock contention. Take the sample at exit point.

2. For all calls of unlock_page(), if the page is acquiring by other on at least 0.1 second before, take a sample at this point.

Currently making the above idea work is possible but not very straightforward. One problem I can identify is:

Doesn't like ftrace, there is no way for eBPF program to access call stack information. Without extra information, eBPF programs are unable to match lock_page events and corresponding lock_page%return events. Currently the only way for passing information between programs are maps. To simulate call stack matching, I create a BPF_FUNC_git_tid() which returns current->pid, and a proc_locking_page_map map which records the acquired page and time of calling lock_page.

Another problem is: at the entry of lock_page() and unlock_page(), for fetching the page pointer I have to directly use 'ctx->regs[0]' (I am on aarch64). Which is not protable.

The final program I used is attached at the bottom of this email. It costs more than 100 lines of code. I have to do some debugging to ensure it works correctly on a virtual machine.

It is compiled using:

 # $CLANG ${INCLUDE} -D__KERNEL__ -Wno-unused-value -Wno-pointer-sign -O2 \
       -emit-llvm -c lock_page.c -o - | $LLC -march=bpf -filetype=obj -o  \
       lock_page.o

Then the lock_page.o is transfered onto target system.

I loaded it using following command:

 # perf record -e syscalls:sys_enter_write -e syscalls:sys_exit_write \
         -e lock_page.o -a iozone ...

Here is another inconvenience. Currently I only concern on write syscall issued by iozone. However, without '-a' I'm unable to collect information of the locker. If I want to filter sys_{enter,exit}_write belong to iozone out using eBPF, I need to implement another function like BPF_FUNC_git_comm. Another method is to use perf '--filter' after the two events. However it looks strange to use two filter mechanisms together. This time I choose to do filtering offline using perf script.

The result is resonable. Finaly I found the two side of lock contention. It shows the way to improve. I'm sorry I can't share the call stack in this list.

One inconvenience in this stage is: the information is printed into ring buffer while the samples are stored into perf.data. By analysing perf.data without ftrace ring buffer I don't know how long the lock_page() cost becasue I don't sample at the entry of lock_page().

The final part is the BPF program I used. I think there should have better way to do it. If any know how to make it shorter please let me know.

Thank you.

/* ------------- START OF BPF PROGRAM ------------- */

    /* __lock_page pass to unlock_page, key is pid */
    struct proc_locking_page {
        unsigned long page;
        unsigned long time;
    };

    struct bpf_map_def SEC("maps") proc_locking_page_map = {
        .type = BPF_MAP_TYPE_HASH,
        .key_size = sizeof(unsigned long),
        .value_size = sizeof(struct proc_locking_page),
        .max_entries = 1000000,
    };

    /* from page get pid */
    struct page_being_locked_by_proc {
        unsigned long tid;
        unsigned long time;
    };

    struct bpf_map_def SEC("maps") page_being_locked_by_proc_map = {
        .type = BPF_MAP_TYPE_HASH,
        .key_size = sizeof(unsigned long),
        .value_size = sizeof(struct page_being_locked_by_proc),
        .max_entries = 1000000,
    };

    SEC("lock_page=__lock_page")
    int lock_page_recorder(struct pt_regs *ctx)
    {
        unsigned long tid = bpf_get_tid();
        unsigned long page = ctx->regs[0];
        unsigned long curr_ns = bpf_ktime_get_ns();

        struct proc_locking_page locking_page;
        struct page_being_locked_by_proc being_locked;

        locking_page.page = page;
        locking_page.time = curr_ns;
        being_locked.tid = tid;
        being_locked.time = curr_ns;

        bpf_map_update_elem(&proc_locking_page_map, &tid, 
&locking_page, BPF_ANY);
        bpf_map_update_elem(&page_being_locked_by_proc_map, &page, 
&being_locked, BPF_ANY);
        return 0;
    }

    SEC("lock_page_ret=__lock_page%return")
    int lock_page_return_recorder(struct pt_regs *ctx)
    {
        unsigned long tid = bpf_get_tid();
        unsigned long curr_ns = bpf_ktime_get_ns();
        unsigned long page;
        unsigned long diff_time;
        struct proc_locking_page *p_locking_page;


        p_locking_page = bpf_map_lookup_elem(&proc_locking_page_map, &tid);


        /* BAD!! */
        if (!p_locking_page)
                return 0;

        page = p_locking_page->page;
        diff_time = curr_ns - p_locking_page->time;
        bpf_map_delete_elem(&proc_locking_page_map, &tid);
        bpf_map_delete_elem(&page_being_locked_by_proc_map, &page);

        if (diff_time > 10000000) {
                char fmt[] = "tid %d get page %lx using %d ns\n";
                bpf_trace_printk(fmt, sizeof(fmt), tid, page, diff_time);
                return 1;
        }

        return 0;
    }

    SEC("unlock_page=unlock_page")
    int unlock_page_recorder(struct pt_regs *ctx)
    {
        unsigned long tid = bpf_get_tid();
        unsigned long page = ctx->regs[0];
        unsigned long time = bpf_ktime_get_ns();
        unsigned long diff_time;
        struct page_being_locked_by_proc *p_being_locked;
        char fmt[] = "%d vs %d, %d ns\n";

        p_being_locked = 
bpf_map_lookup_elem(&page_being_locked_by_proc_map, &page);
        if (!p_being_locked)
                return 0;
        diff_time = time - p_being_locked->time;
        if (diff_time > 10000000) {
                bpf_trace_printk(fmt, sizeof(fmt), tid, 
p_being_locked->tid, diff_time);
                return 1;
        }
        return 0;
    }

/* ------------- END OF BPF PROGRAM ------------- */

#24 lore raw

Re: [EXPERIENCE] My experience on using perf record BPF filter on a real usecase

Quoted 4 lines · Wangnan (F)

Hi all,

I'd like to share my exprience on using 'perf record' BPF filter in a real usecase to show the power and shortcome in my patch series:

thanks for sharing!

Quoted 7 lines

Here is another inconvenience. Currently I only concern on write syscall issued by iozone. However, without '-a' I'm unable to collect information of the locker. If I want to filter sys_{enter,exit}_write belong to iozone out using eBPF, I need to implement another function like BPF_FUNC_git_comm. Another method is to use perf '--filter' after the two events. However it looks strange to use two filter mechanisms together. This time I choose to do filtering offline using perf script.

that doesn't sound clean.
btw, I've been playing for a while with bpf_get_current_task_info() helper:
https://git.kernel.org/cgit/linux/kernel/git/ast/bpf.git/commit/?id=c5453ffa107ddf95a91920cc947bb8bf9eab16d6
I think it's a better mechanism.
The user can get pid only via:
u32 pid = 0;
bpf_get_current_task_info(&pid, sizeof(pid));
or full pid + comm + future fields via full 'struct bpf_task_info' Thoughts?

#25 lore raw

Re: [EXPERIENCE] My experience on using perf record BPF filter on a real usecase

Quoted 21 lines · Alexei Starovoitov
Quoted 4 lines · Wangnan (F)

Hi all,

I'd like to share my exprience on using 'perf record' BPF filter in a real usecase to show the power and shortcome in my patch series:

thanks for sharing!

Quoted 7 lines

Here is another inconvenience. Currently I only concern on write syscall issued by iozone. However, without '-a' I'm unable to collect information of the locker. If I want to filter sys_{enter,exit}_write belong to iozone out using eBPF, I need to implement another function like BPF_FUNC_git_comm. Another method is to use perf '--filter' after the two events. However it looks strange to use two filter mechanisms together. This time I choose to do filtering offline using perf script.

that doesn't sound clean.
btw, I've been playing for a while with
bpf_get_current_task_info() helper:
https://git.kernel.org/cgit/linux/kernel/git/ast/bpf.git/commit/?id=c5453ffa107ddf95a91920cc947bb8bf9eab16d6

Quoted 7 lines

I think it's a better mechanism.
The user can get pid only via:
u32 pid = 0;
bpf_get_current_task_info(&pid, sizeof(pid));
or full pid + comm + future fields via full 'struct bpf_task_info' Thoughts?

Looks good. Thank you for your information!