perf mem: add priv level filtering support

perf mem: add priv level filtering support

This patch adds the -u -and -k options to perf to allow filtering of load/store sampling based on priv levels. This may not be supported by all HW platforms.

By default, loads/stores are sampled at both user and kernel privilege levels.

To sample only at the user level:

$ perf mem -u -t load rec ......

To sample only at the kernel level:

$ perf mem -k -t load rec ......

Man page updated accordingly.

Signed-off-by Stephane Eranian <[email protected]>

---

tools/perf/Documentation/perf-mem.txt +0 -0

index 888d511..4c4e405 100644

tools/perf/Documentation/perf-mem.txt +1 -0

+
  SEE ALSO
  --------
  linkperf:perf-record[1], linkperf:perf-report[1]

tools/perf/builtin-mem.c +86 -20

index 706a1fa..8ac9d1e 100644
@@ -9,13 +9,18 @@
  #define MEM_OPERATION_LOAD	"load"
  #define MEM_OPERATION_STORE	"store"
 
-static const char	*mem_operation		= MEM_OPERATION_LOAD;
+#define OP_LOAD		0x1
+#define OP_STORE	0x2
+
 
  struct perf_mem {
  	struct perf_tool	tool;
  	char const		*input_name;
  	bool			hide_unresolved;
+	const char		*mem_op;
  	bool			dump_raw;
+	bool			user;
+	bool			kernel;
  	const char		*cpu_list;
  	DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  };

@@ -25,35 +30,88 @@ static const char * const mem_usage[] = {
  	NULL
  };
 
-static int __cmd_record(int argc, const char **argv)
+static inline const char *get_plm(struct perf_mem *mem)
+{
+	const char *plm = "";
+
+	if (mem->user && !mem->kernel) {
+		plm = "u";
+	} else if (!mem->user && mem->kernel) {
+		plm = "k";
+	}
+	return plm;
+}
+
+static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
  {
  	int rec_argc, i = 0, j;
  	const char **rec_argv;
-	char event[64];
-	int ret;
+	char *str;
+	int mode = 0;
+	int ki, ret;
+
+
+	if (!strcmp(mem->mem_op, MEM_OPERATION_STORE))
+		mode |= OP_STORE;
+	else if (!strcmp(mem->mem_op, MEM_OPERATION_LOAD))
+		mode |= OP_LOAD;
+	else  {
+		fprintf(stderr, "unknown sampling mode: %s\n", mem->mem_op);
+		return -1;
+	}
 
-	rec_argc = argc + 4;
+	rec_argc = argc + 6;
  	rec_argv = calloc(rec_argc + 1, sizeof(char *));
  	if (!rec_argv)
  		return -1;
 
  	rec_argv[i++] = strdup("record");
-	if (!strcmp(mem_operation, MEM_OPERATION_LOAD))
-		rec_argv[i++] = strdup("-W");
+
  	rec_argv[i++] = strdup("-d");
-	rec_argv[i++] = strdup("-e");
 
-	if (strcmp(mem_operation, MEM_OPERATION_LOAD))
-		sprintf(event, "cpu/mem-stores/pp");
-	else
-		sprintf(event, "cpu/mem-loads/pp");
+        if (mode & OP_LOAD) {
+                rec_argv[i++] = strdup("-W");
 
-	rec_argv[i++] = strdup(event);
-	for (j = 1; j < argc; j++, i++)
-		rec_argv[i] = argv[j];
+                rec_argv[i++] = strdup("-e");
+
+                str = malloc(strlen("cpu/mem-loads/pp") + 1 + 1);
+                if (!str) {
+                        ki = i;
+                        ret = -1;
+                        goto end;
+                }
+                sprintf(str, "cpu/mem-loads/%spp", get_plm(mem));
+                rec_argv[i++] = str;
+        }
+
+        if (mode & OP_STORE) {
+                rec_argv[i++] = strdup("-e");
+
+                str = malloc(strlen("cpu/mem-stores/pp") + 1 + 1);
+                if (!str) {
+                        ki = i;
+                        ret = -1;
+                        goto end;
+                }
+                sprintf(str, "cpu/mem-stores/%spp", get_plm(mem));
+                rec_argv[i++] = str;
+        }
+
+        /* arguments after i are not malloc'd */
+        ki = i;
 
-	ret = cmd_record(i, rec_argv, NULL);
+	for (j = 1; j < argc; j++, ki++)
+		rec_argv[ki] = argv[j];
+
+	ret = cmd_record(ki, rec_argv, NULL);
+
+end:
+	/*
+	 * XXX: free rec_argv[] entries, difficult because
+	 * cmd_record() drops some of them...
+	 */
  	free(rec_argv);
+
  	return ret;
  }
 

@@ -171,7 +229,7 @@ static int report_events(int argc, const char **argv, struct perf_mem *mem)
  	 * there is no weight (cost) associated with stores, so don't print
  	 * the column
  	 */
-	if (strcmp(mem_operation, MEM_OPERATION_LOAD))
+	if (strcmp(mem->mem_op, MEM_OPERATION_LOAD))
  		rep_argv[i++] = strdup("--sort=mem,sym,dso,symbol_daddr,"
  				       "dso_daddr,tlb,locked");
 

@@ -199,7 +257,7 @@ int cmd_mem(int argc, const char **argv, const char *prefix __maybe_unused)
  		.input_name		 = "perf.data",
  	};
  	const struct option mem_options[] = {
-	OPT_STRING('t', "type", &mem_operation,
+	OPT_STRING('t', "type", &mem.mem_op,
  		   "type", "memory operations(load/store)"),
  	OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw,
  		    "dump raw samples in ASCII"),

@@ -213,13 +271,18 @@ int cmd_mem(int argc, const char **argv, const char *prefix __maybe_unused)
  		   "separator",
  		   "separator for columns, no spaces will be added"
  		   " between columns '.' is reserved."),
+	OPT_BOOLEAN('u', "user-level", &mem.user,
+		    "include user-level accesses"),
+	OPT_BOOLEAN('k', "kernel-level", &mem.kernel,
+		    "include kernel-level accesses"),
  	OPT_END()
  	};
 
  	argc = parse_options(argc, argv, mem_options, mem_usage,
  			     PARSE_OPT_STOP_AT_NON_OPTION);
 
-	if (!argc || !(strncmp(argv[0], "rec", 3) || mem_operation))
+	if (!argc || !(strncmp(argv[0], "rec", 3)
+		       || strncmp(argv[0], "rep", 3)))
  		usage_with_options(mem_usage, mem_options);
 
  	if (!mem.input_name || !strlen(mem.input_name)) {

@@ -228,9 +291,12 @@ int cmd_mem(int argc, const char **argv, const char *prefix __maybe_unused)
  		else
  			mem.input_name = "perf.data";
  	}
+	/* default to load only, some processors only support loads */
+	if (!mem.mem_op)
+		mem.mem_op = MEM_OPERATION_LOAD;
 
  	if (!strncmp(argv[0], "rec", 3))
-		return __cmd_record(argc, argv);
+		return __cmd_record(argc, argv, &mem);
  	else if (!strncmp(argv[0], "rep", 3))
  		return report_events(argc, argv, &mem);
  	else

Quoted 29 lines · Stephane Eranian

perf mem: add priv level filtering support

This patch adds the -u -and -k options to perf to allow filtering of load/store sampling based on priv levels. This may not be supported by all HW platforms.

By default, loads/stores are sampled at both user and kernel privilege levels.

To sample only at the user level:

$ perf mem -u -t load rec ......

To sample only at the kernel level:

$ perf mem -k -t load rec ......

Man page updated accordingly.

Signed-off-by Stephane Eranian <[email protected]>

---

tools/perf/Documentation/perf-mem.txt +0 -0

index 888d511..4c4e405 100644
@@ -43,6 +43,12 @@ OPTIONS
        option can be passed in record mode. It will be interpreted the same way as perf
        record.
 

tools/perf/Documentation/perf-mem.txt OPTIONS · line 50

Quoted 5 lines

tools/perf/Documentation/perf-mem.txt +5 -0

+-k::
+       Only sample loads/stores at the user level (default: user + kernel)
+
+-u::
+       Only sample loads/stores at the kernel level (default: user + kernel)

Are the descriptions backwards? In the commit message yuo have -u means user level and -k means kernel level; the help message here seems backwards.

David

Quoted 185 lines

tools/perf/Documentation/perf-mem.txt +1 -0

+
  SEE ALSO
  --------
  linkperf:perf-record[1], linkperf:perf-report[1]

tools/perf/builtin-mem.c +86 -20

index 706a1fa..8ac9d1e 100644
@@ -9,13 +9,18 @@
  #define MEM_OPERATION_LOAD    "load"
  #define MEM_OPERATION_STORE   "store"
 
-static const char      *mem_operation          = MEM_OPERATION_LOAD;
+#define OP_LOAD                0x1
+#define OP_STORE       0x2
+
 
  struct perf_mem {
        struct perf_tool        tool;
        char const              *input_name;
        bool                    hide_unresolved;
+       const char              *mem_op;
        bool                    dump_raw;
+       bool                    user;
+       bool                    kernel;
        const char              *cpu_list;
        DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  };

@@ -25,35 +30,88 @@ static const char * const mem_usage[] = {
        NULL
  };
 
-static int __cmd_record(int argc, const char **argv)
+static inline const char *get_plm(struct perf_mem *mem)
+{
+       const char *plm = "";
+
+       if (mem->user && !mem->kernel) {
+               plm = "u";
+       } else if (!mem->user && mem->kernel) {
+               plm = "k";
+       }
+       return plm;
+}
+
+static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
  {
        int rec_argc, i = 0, j;
        const char **rec_argv;
-       char event[64];
-       int ret;
+       char *str;
+       int mode = 0;
+       int ki, ret;
+
+
+       if (!strcmp(mem->mem_op, MEM_OPERATION_STORE))
+               mode |= OP_STORE;
+       else if (!strcmp(mem->mem_op, MEM_OPERATION_LOAD))
+               mode |= OP_LOAD;
+       else  {
+               fprintf(stderr, "unknown sampling mode: %s\n", mem->mem_op);
+               return -1;
+       }
 
-       rec_argc = argc + 4;
+       rec_argc = argc + 6;
        rec_argv = calloc(rec_argc + 1, sizeof(char *));
        if (!rec_argv)
                return -1;
 
        rec_argv[i++] = strdup("record");
-       if (!strcmp(mem_operation, MEM_OPERATION_LOAD))
-               rec_argv[i++] = strdup("-W");
+
        rec_argv[i++] = strdup("-d");
-       rec_argv[i++] = strdup("-e");
 
-       if (strcmp(mem_operation, MEM_OPERATION_LOAD))
-               sprintf(event, "cpu/mem-stores/pp");
-       else
-               sprintf(event, "cpu/mem-loads/pp");
+        if (mode & OP_LOAD) {
+                rec_argv[i++] = strdup("-W");
 
-       rec_argv[i++] = strdup(event);
-       for (j = 1; j < argc; j++, i++)
-               rec_argv[i] = argv[j];
+                rec_argv[i++] = strdup("-e");
+
+                str = malloc(strlen("cpu/mem-loads/pp") + 1 + 1);
+                if (!str) {
+                        ki = i;
+                        ret = -1;
+                        goto end;
+                }
+                sprintf(str, "cpu/mem-loads/%spp", get_plm(mem));
+                rec_argv[i++] = str;
+        }
+
+        if (mode & OP_STORE) {
+                rec_argv[i++] = strdup("-e");
+
+                str = malloc(strlen("cpu/mem-stores/pp") + 1 + 1);
+                if (!str) {
+                        ki = i;
+                        ret = -1;
+                        goto end;
+                }
+                sprintf(str, "cpu/mem-stores/%spp", get_plm(mem));
+                rec_argv[i++] = str;
+        }
+
+        /* arguments after i are not malloc'd */
+        ki = i;
 
-       ret = cmd_record(i, rec_argv, NULL);
+       for (j = 1; j < argc; j++, ki++)
+               rec_argv[ki] = argv[j];
+
+       ret = cmd_record(ki, rec_argv, NULL);
+
+end:
+       /*
+        * XXX: free rec_argv[] entries, difficult because
+        * cmd_record() drops some of them...
+        */
        free(rec_argv);
+
        return ret;
  }
 

@@ -171,7 +229,7 @@ static int report_events(int argc, const char **argv, struct perf_mem *mem)
         * there is no weight (cost) associated with stores, so don't print
         * the column
         */
-       if (strcmp(mem_operation, MEM_OPERATION_LOAD))
+       if (strcmp(mem->mem_op, MEM_OPERATION_LOAD))
                rep_argv[i++] = strdup("--sort=mem,sym,dso,symbol_daddr,"
                                       "dso_daddr,tlb,locked");
 

@@ -199,7 +257,7 @@ int cmd_mem(int argc, const char **argv, const char *prefix __maybe_unused)
                .input_name              = "perf.data",
        };
        const struct option mem_options[] = {
-       OPT_STRING('t', "type", &mem_operation,
+       OPT_STRING('t', "type", &mem.mem_op,
                   "type", "memory operations(load/store)"),
        OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw,
                    "dump raw samples in ASCII"),

@@ -213,13 +271,18 @@ int cmd_mem(int argc, const char **argv, const char *prefix __maybe_unused)
                   "separator",
                   "separator for columns, no spaces will be added"
                   " between columns '.' is reserved."),
+       OPT_BOOLEAN('u', "user-level", &mem.user,
+                   "include user-level accesses"),
+       OPT_BOOLEAN('k', "kernel-level", &mem.kernel,
+                   "include kernel-level accesses"),
        OPT_END()
        };
 
        argc = parse_options(argc, argv, mem_options, mem_usage,
                             PARSE_OPT_STOP_AT_NON_OPTION);
 
-       if (!argc || !(strncmp(argv[0], "rec", 3) || mem_operation))
+       if (!argc || !(strncmp(argv[0], "rec", 3)
+                      || strncmp(argv[0], "rep", 3)))
                usage_with_options(mem_usage, mem_options);
 
        if (!mem.input_name || !strlen(mem.input_name)) {

@@ -228,9 +291,12 @@ int cmd_mem(int argc, const char **argv, const char *prefix __maybe_unused)
                else
                        mem.input_name = "perf.data";
        }
+       /* default to load only, some processors only support loads */
+       if (!mem.mem_op)
+               mem.mem_op = MEM_OPERATION_LOAD;
 
        if (!strncmp(argv[0], "rec", 3))
-               return __cmd_record(argc, argv);
+               return __cmd_record(argc, argv, &mem);
        else if (!strncmp(argv[0], "rep", 3))
                return report_events(argc, argv, &mem);
        else

Quoted 4 lines

This patch adds the -u -and -k options to perf to allow filtering of load/store sampling based on priv levels. This may not be supported by all HW platforms.

tools/perf/Documentation/perf-mem.txt OPTIONS · line 50

Quoted 5 lines

+++ b/tools/perf/Documentation/perf-mem.txt
+-k::

tools/perf/Documentation/perf-mem.txt +3 -0

+	Only sample loads/stores at the user level (default: user + kernel)
+-u::
+	Only sample loads/stores at the kernel level (default: user + kernel)

Are the descriptions backwards? In the commit message yuo have -u means user level and -k means kernel level; the help message here seems backwards.

Review

Quoted 2 lines

(quoted diff) +2 -0

+	OPT_BOOLEAN('u', "user-level", &mem.user, "include user-level accesses"),
+	OPT_BOOLEAN('k', "kernel-level", &mem.kernel, "include kernel-level accesses"),

Em Mon, Aug 26, 2013 at 04:21:44PM -0600, David Ahern escreveu:

Quoted 7 lines · Stephane Eranian

Quoted 4 lines

This patch adds the -u -and -k options to perf to allow filtering of load/store sampling based on priv levels. This may not be supported by all HW platforms.

Quoted 9 lines

tools/perf/Documentation/perf-mem.txt OPTIONS · line 50

Quoted 5 lines

+++ b/tools/perf/Documentation/perf-mem.txt
+-k::

tools/perf/Documentation/perf-mem.txt +3 -0

+    Only sample loads/stores at the user level (default: user + kernel)
+-u::
+    Only sample loads/stores at the kernel level (default: user + kernel)

Are the descriptions backwards? In the commit message yuo have -u means user level and -k means kernel level; the help message here seems backwards.

Looks like it is reversed, yes.

Quoted 2 lines

Review

Quoted 2 lines

(quoted diff) +2 -0

+    OPT_BOOLEAN('u', "user-level", &mem.user, "include user-level accesses"),
+    OPT_BOOLEAN('k', "kernel-level", &mem.kernel, "include kernel-level accesses"),

And its not clear to say that using -u will _exclude_ kernel samples, just that it includes user samples, perf top has:

    -K, --hide_kernel_symbols   hide kernel symbols
    -U, --hide_user_symbols     hide user symbols

So perhaps we should change both to (and add this to 'report' as well):

    -U, --hide_kernel_symbols   hide kernel symbols
    -K, --hide_user_symbols     hide user symbols

To state that:

perf top -K
and

        perf top --hide_user_symbols

Are equivalent and asks for 'kernel only' samples, like it seems its the intent (filtering) of Stephane here, and seems to clarify things?

- Arnaldo

Quoted 26 lines · Arnaldo Carvalho de Melo

Em Mon, Aug 26, 2013 at 04:21:44PM -0600, David Ahern escreveu:

Quoted 7 lines · Stephane Eranian

Quoted 4 lines

This patch adds the -u -and -k options to perf to allow filtering of load/store sampling based on priv levels. This may not be supported by all HW platforms.

Quoted 9 lines

tools/perf/Documentation/perf-mem.txt OPTIONS · line 50

Quoted 5 lines

+++ b/tools/perf/Documentation/perf-mem.txt
+-k::

tools/perf/Documentation/perf-mem.txt +3 -0

+    Only sample loads/stores at the user level (default: user + kernel)
+-u::
+    Only sample loads/stores at the kernel level (default: user + kernel)

Are the descriptions backwards? In the commit message yuo have -u means user level and -k means kernel level; the help message here seems backwards.

Looks like it is reversed, yes.

Quoted 2 lines

Review

Quoted 2 lines

(quoted diff) +2 -0

+    OPT_BOOLEAN('u', "user-level", &mem.user, "include user-level accesses"),
+    OPT_BOOLEAN('k', "kernel-level", &mem.kernel, "include kernel-level accesses"),

And its not clear to say that using -u will _exclude_ kernel samples, just that it includes user samples, perf top has:

Quoted 8 lines

    -K, --hide_kernel_symbols   hide kernel symbols
    -U, --hide_user_symbols     hide user symbols

So perhaps we should change both to (and add this to 'report' as well):

    -U, --hide_kernel_symbols   hide kernel symbols
    -K, --hide_user_symbols     hide user symbols

Well, I don't know what perf top does here but I don't want to hide the samples. I simply don't want to collect them (do not appear in the perf.data file). If that's what is happening in perf top, then I'll be glad to use the same options.

Quoted 10 lines

To state that:

perf top -K
and

        perf top --hide_user_symbols

Are equivalent and asks for 'kernel only' samples, like it seems its the intent (filtering) of Stephane here, and seems to clarify things?

- Arnaldo

Quoted 5 lines

So perhaps we should change both to (and add this to 'report' as well):

    -U, --hide_kernel_symbols   hide kernel symbols
    -K, --hide_user_symbols     hide user symbols

Well, I don't know what perf top does here but I don't want to hide the samples. I simply don't want to collect them (do not appear in the perf.data file). If that's what is happening in perf top, then I'll be glad to use the same options.

But my worry here is about consistency accross tools for the single letter options, so perhaps if you could use:

-U collect only user level samples
-K collect only kernel level samples