perf: add perf-inject builtin
[safe/jmp/linux-2.6] / tools / perf / builtin-kmem.c
index f315b05..31f60a2 100644 (file)
@@ -6,12 +6,12 @@
 #include "util/symbol.h"
 #include "util/thread.h"
 #include "util/header.h"
+#include "util/session.h"
 
 #include "util/parse-options.h"
 #include "util/trace-event.h"
 
 #include "util/debug.h"
-#include "util/data_map.h"
 
 #include <linux/rbtree.h>
 
@@ -20,32 +20,28 @@ typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
 
 static char const              *input_name = "perf.data";
 
-static struct perf_header      *header;
-static u64                     sample_type;
-
 static int                     alloc_flag;
 static int                     caller_flag;
 
-sort_fn_t                      alloc_sort_fn;
-sort_fn_t                      caller_sort_fn;
-
 static int                     alloc_lines = -1;
 static int                     caller_lines = -1;
 
-static char                    *cwd;
-static int                     cwdlen;
+static bool                    raw_ip;
+
+static char                    default_sort_order[] = "frag,hit,bytes";
+
+static int                     *cpunode_map;
+static int                     max_cpu_num;
 
 struct alloc_stat {
-       union {
-               struct {
-                       char    *name;
-                       u64     call_site;
-               };
-               u64     ptr;
-       };
+       u64     call_site;
+       u64     ptr;
        u64     bytes_req;
        u64     bytes_alloc;
        u32     hit;
+       u32     pingpong;
+
+       short   alloc_cpu;
 
        struct rb_node node;
 };
@@ -56,41 +52,71 @@ static struct rb_root root_caller_stat;
 static struct rb_root root_caller_sorted;
 
 static unsigned long total_requested, total_allocated;
+static unsigned long nr_allocs, nr_cross_allocs;
 
-struct raw_event_sample {
-       u32 size;
-       char data[0];
-};
+#define PATH_SYS_NODE  "/sys/devices/system/node"
 
-static int
-process_comm_event(event_t *event, unsigned long offset, unsigned long head)
+static void init_cpunode_map(void)
 {
-       struct thread *thread = threads__findnew(event->comm.pid);
-
-       dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
-               (void *)(offset + head),
-               (void *)(long)(event->header.size),
-               event->comm.comm, event->comm.pid);
+       FILE *fp;
+       int i;
 
-       if (thread == NULL ||
-           thread__set_comm(thread, event->comm.comm)) {
-               dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
-               return -1;
+       fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
+       if (!fp) {
+               max_cpu_num = 4096;
+               return;
        }
 
-       return 0;
+       if (fscanf(fp, "%d", &max_cpu_num) < 1)
+               die("Failed to read 'kernel_max' from sysfs");
+       max_cpu_num++;
+
+       cpunode_map = calloc(max_cpu_num, sizeof(int));
+       if (!cpunode_map)
+               die("calloc");
+       for (i = 0; i < max_cpu_num; i++)
+               cpunode_map[i] = -1;
+       fclose(fp);
 }
 
-static void insert_alloc_stat(unsigned long ptr,
-                             int bytes_req, int bytes_alloc)
+static void setup_cpunode_map(void)
+{
+       struct dirent *dent1, *dent2;
+       DIR *dir1, *dir2;
+       unsigned int cpu, mem;
+       char buf[PATH_MAX];
+
+       init_cpunode_map();
+
+       dir1 = opendir(PATH_SYS_NODE);
+       if (!dir1)
+               return;
+
+       while ((dent1 = readdir(dir1)) != NULL) {
+               if (dent1->d_type != DT_DIR ||
+                   sscanf(dent1->d_name, "node%u", &mem) < 1)
+                       continue;
+
+               snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
+               dir2 = opendir(buf);
+               if (!dir2)
+                       continue;
+               while ((dent2 = readdir(dir2)) != NULL) {
+                       if (dent2->d_type != DT_LNK ||
+                           sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
+                               continue;
+                       cpunode_map[cpu] = mem;
+               }
+       }
+}
+
+static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
+                             int bytes_req, int bytes_alloc, int cpu)
 {
        struct rb_node **node = &root_alloc_stat.rb_node;
        struct rb_node *parent = NULL;
        struct alloc_stat *data = NULL;
 
-       if (!alloc_flag)
-               return;
-
        while (*node) {
                parent = *node;
                data = rb_entry(*node, struct alloc_stat, node);
@@ -106,10 +132,13 @@ static void insert_alloc_stat(unsigned long ptr,
        if (data && data->ptr == ptr) {
                data->hit++;
                data->bytes_req += bytes_req;
-               data->bytes_alloc += bytes_req;
+               data->bytes_alloc += bytes_alloc;
        } else {
                data = malloc(sizeof(*data));
+               if (!data)
+                       die("malloc");
                data->ptr = ptr;
+               data->pingpong = 0;
                data->hit = 1;
                data->bytes_req = bytes_req;
                data->bytes_alloc = bytes_alloc;
@@ -117,6 +146,8 @@ static void insert_alloc_stat(unsigned long ptr,
                rb_link_node(&data->node, parent, node);
                rb_insert_color(&data->node, &root_alloc_stat);
        }
+       data->call_site = call_site;
+       data->alloc_cpu = cpu;
 }
 
 static void insert_caller_stat(unsigned long call_site,
@@ -126,9 +157,6 @@ static void insert_caller_stat(unsigned long call_site,
        struct rb_node *parent = NULL;
        struct alloc_stat *data = NULL;
 
-       if (!caller_flag)
-               return;
-
        while (*node) {
                parent = *node;
                data = rb_entry(*node, struct alloc_stat, node);
@@ -144,10 +172,13 @@ static void insert_caller_stat(unsigned long call_site,
        if (data && data->call_site == call_site) {
                data->hit++;
                data->bytes_req += bytes_req;
-               data->bytes_alloc += bytes_req;
+               data->bytes_alloc += bytes_alloc;
        } else {
                data = malloc(sizeof(*data));
+               if (!data)
+                       die("malloc");
                data->call_site = call_site;
+               data->pingpong = 0;
                data->hit = 1;
                data->bytes_req = bytes_req;
                data->bytes_alloc = bytes_alloc;
@@ -157,102 +188,138 @@ static void insert_caller_stat(unsigned long call_site,
        }
 }
 
-static void process_alloc_event(struct raw_event_sample *raw,
+static void process_alloc_event(void *data,
                                struct event *event,
-                               int cpu __used,
+                               int cpu,
                                u64 timestamp __used,
                                struct thread *thread __used,
-                               int node __used)
+                               int node)
 {
        unsigned long call_site;
        unsigned long ptr;
        int bytes_req;
        int bytes_alloc;
+       int node1, node2;
 
-       ptr = raw_field_value(event, "ptr", raw->data);
-       call_site = raw_field_value(event, "call_site", raw->data);
-       bytes_req = raw_field_value(event, "bytes_req", raw->data);
-       bytes_alloc = raw_field_value(event, "bytes_alloc", raw->data);
+       ptr = raw_field_value(event, "ptr", data);
+       call_site = raw_field_value(event, "call_site", data);
+       bytes_req = raw_field_value(event, "bytes_req", data);
+       bytes_alloc = raw_field_value(event, "bytes_alloc", data);
 
-       insert_alloc_stat(ptr, bytes_req, bytes_alloc);
+       insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
        insert_caller_stat(call_site, bytes_req, bytes_alloc);
 
        total_requested += bytes_req;
        total_allocated += bytes_alloc;
+
+       if (node) {
+               node1 = cpunode_map[cpu];
+               node2 = raw_field_value(event, "node", data);
+               if (node1 != node2)
+                       nr_cross_allocs++;
+       }
+       nr_allocs++;
 }
 
-static void process_free_event(struct raw_event_sample *raw __used,
-                              struct event *event __used,
-                              int cpu __used,
+static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
+static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
+
+static struct alloc_stat *search_alloc_stat(unsigned long ptr,
+                                           unsigned long call_site,
+                                           struct rb_root *root,
+                                           sort_fn_t sort_fn)
+{
+       struct rb_node *node = root->rb_node;
+       struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
+
+       while (node) {
+               struct alloc_stat *data;
+               int cmp;
+
+               data = rb_entry(node, struct alloc_stat, node);
+
+               cmp = sort_fn(&key, data);
+               if (cmp < 0)
+                       node = node->rb_left;
+               else if (cmp > 0)
+                       node = node->rb_right;
+               else
+                       return data;
+       }
+       return NULL;
+}
+
+static void process_free_event(void *data,
+                              struct event *event,
+                              int cpu,
                               u64 timestamp __used,
                               struct thread *thread __used)
 {
+       unsigned long ptr;
+       struct alloc_stat *s_alloc, *s_caller;
+
+       ptr = raw_field_value(event, "ptr", data);
+
+       s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
+       if (!s_alloc)
+               return;
+
+       if (cpu != s_alloc->alloc_cpu) {
+               s_alloc->pingpong++;
+
+               s_caller = search_alloc_stat(0, s_alloc->call_site,
+                                            &root_caller_stat, callsite_cmp);
+               assert(s_caller);
+               s_caller->pingpong++;
+       }
+       s_alloc->alloc_cpu = -1;
 }
 
 static void
-process_raw_event(event_t *raw_event __used, void *more_data,
+process_raw_event(event_t *raw_event __used, void *data,
                  int cpu, u64 timestamp, struct thread *thread)
 {
-       struct raw_event_sample *raw = more_data;
        struct event *event;
        int type;
 
-       type = trace_parse_common_type(raw->data);
+       type = trace_parse_common_type(data);
        event = trace_find_event(type);
 
        if (!strcmp(event->name, "kmalloc") ||
            !strcmp(event->name, "kmem_cache_alloc")) {
-               process_alloc_event(raw, event, cpu, timestamp, thread, 0);
+               process_alloc_event(data, event, cpu, timestamp, thread, 0);
                return;
        }
 
        if (!strcmp(event->name, "kmalloc_node") ||
            !strcmp(event->name, "kmem_cache_alloc_node")) {
-               process_alloc_event(raw, event, cpu, timestamp, thread, 1);
+               process_alloc_event(data, event, cpu, timestamp, thread, 1);
                return;
        }
 
        if (!strcmp(event->name, "kfree") ||
            !strcmp(event->name, "kmem_cache_free")) {
-               process_free_event(raw, event, cpu, timestamp, thread);
+               process_free_event(data, event, cpu, timestamp, thread);
                return;
        }
 }
 
-static int
-process_sample_event(event_t *event, unsigned long offset, unsigned long head)
+static int process_sample_event(event_t *event, struct perf_session *session)
 {
-       u64 ip = event->ip.ip;
-       u64 timestamp = -1;
-       u32 cpu = -1;
-       u64 period = 1;
-       void *more_data = event->ip.__more_data;
-       struct thread *thread = threads__findnew(event->ip.pid);
-
-       if (sample_type & PERF_SAMPLE_TIME) {
-               timestamp = *(u64 *)more_data;
-               more_data += sizeof(u64);
-       }
+       struct sample_data data;
+       struct thread *thread;
 
-       if (sample_type & PERF_SAMPLE_CPU) {
-               cpu = *(u32 *)more_data;
-               more_data += sizeof(u32);
-               more_data += sizeof(u32); /* reserved */
-       }
+       memset(&data, 0, sizeof(data));
+       data.time = -1;
+       data.cpu = -1;
+       data.period = 1;
 
-       if (sample_type & PERF_SAMPLE_PERIOD) {
-               period = *(u64 *)more_data;
-               more_data += sizeof(u64);
-       }
+       event__parse_sample(event, session->sample_type, &data);
 
-       dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
-               (void *)(offset + head),
-               (void *)(long)(event->header.size),
-               event->header.misc,
-               event->ip.pid, event->ip.tid,
-               (void *)(long)ip,
-               (long long)period);
+       dump_printf("(IP, %d): %d/%d: %#Lx period: %Ld\n", event->header.misc,
+                   data.pid, data.tid, data.ip, data.period);
 
+       thread = perf_session__findnew(session, event->ip.pid);
        if (thread == NULL) {
                pr_debug("problem processing %d event, skipping it.\n",
                         event->header.type);
@@ -261,40 +328,18 @@ process_sample_event(event_t *event, unsigned long offset, unsigned long head)
 
        dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 
-       process_raw_event(event, more_data, cpu, timestamp, thread);
+       process_raw_event(event, data.raw_data, data.cpu,
+                         data.time, thread);
 
        return 0;
 }
 
-static int sample_type_check(u64 type)
-{
-       sample_type = type;
-
-       if (!(sample_type & PERF_SAMPLE_RAW)) {
-               fprintf(stderr,
-                       "No trace sample to read. Did you call perf record "
-                       "without -R?");
-               return -1;
-       }
-
-       return 0;
-}
-
-static struct perf_file_handler file_handler = {
-       .process_sample_event   = process_sample_event,
-       .process_comm_event     = process_comm_event,
-       .sample_type_check      = sample_type_check,
+static struct perf_event_ops event_ops = {
+       .sample                 = process_sample_event,
+       .comm                   = event__process_comm,
+       .ordered_samples        = true,
 };
 
-static int read_events(void)
-{
-       register_idle_thread();
-       register_perf_file_handler(&file_handler);
-
-       return mmap_dispatch_perf_file(&header, input_name, 0, 0,
-                                      &cwdlen, &cwd);
-}
-
 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
 {
        if (n_alloc == 0)
@@ -303,42 +348,62 @@ static double fragmentation(unsigned long n_req, unsigned long n_alloc)
                return 100.0 - (100.0 * n_req / n_alloc);
 }
 
-static void __print_result(struct rb_root *root, int n_lines, int is_caller)
+static void __print_result(struct rb_root *root, struct perf_session *session,
+                          int n_lines, int is_caller)
 {
        struct rb_node *next;
+       struct machine *machine;
 
-       printf("\n ------------------------------------------------------------------------------\n");
-       if (is_caller)
-               printf(" Callsite          |");
-       else
-               printf(" Alloc Ptr         |");
-       printf(" Total_alloc/Per |  Total_req/Per  |  Hit   | Fragmentation\n");
-       printf(" ------------------------------------------------------------------------------\n");
+       printf("%.102s\n", graph_dotted_line);
+       printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
+       printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
+       printf("%.102s\n", graph_dotted_line);
 
        next = rb_first(root);
 
+       machine = perf_session__find_host_machine(session);
+       if (!machine) {
+               pr_err("__print_result: couldn't find kernel information\n");
+               return;
+       }
        while (next && n_lines--) {
-               struct alloc_stat *data;
-
-               data = rb_entry(next, struct alloc_stat, node);
+               struct alloc_stat *data = rb_entry(next, struct alloc_stat,
+                                                  node);
+               struct symbol *sym = NULL;
+               struct map *map;
+               char buf[BUFSIZ];
+               u64 addr;
+
+               if (is_caller) {
+                       addr = data->call_site;
+                       if (!raw_ip)
+                               sym = machine__find_kernel_function(machine, addr, &map, NULL);
+               } else
+                       addr = data->ptr;
+
+               if (sym != NULL)
+                       snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
+                                addr - map->unmap_ip(map, sym->start));
+               else
+                       snprintf(buf, sizeof(buf), "%#Lx", addr);
+               printf(" %-34s |", buf);
 
-               printf(" %-16p  | %8llu/%-6lu | %8llu/%-6lu | %6lu | %8.3f%%\n",
-                      is_caller ? (void *)(unsigned long)data->call_site :
-                                  (void *)(unsigned long)data->ptr,
+               printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
                       (unsigned long long)data->bytes_alloc,
                       (unsigned long)data->bytes_alloc / data->hit,
                       (unsigned long long)data->bytes_req,
                       (unsigned long)data->bytes_req / data->hit,
                       (unsigned long)data->hit,
+                      (unsigned long)data->pingpong,
                       fragmentation(data->bytes_req, data->bytes_alloc));
 
                next = rb_next(next);
        }
 
        if (n_lines == -1)
-               printf(" ...               | ...             | ...             | ...    | ...   \n");
+               printf(" ...                                | ...             | ...             | ...    | ...      | ...   \n");
 
-       printf(" ------------------------------------------------------------------------------\n");
+       printf("%.102s\n", graph_dotted_line);
 }
 
 static void print_summary(void)
@@ -350,31 +415,46 @@ static void print_summary(void)
               total_allocated - total_requested);
        printf("Internal fragmentation: %f%%\n",
               fragmentation(total_requested, total_allocated));
+       printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
 }
 
-static void print_result(void)
+static void print_result(struct perf_session *session)
 {
        if (caller_flag)
-               __print_result(&root_caller_sorted, caller_lines, 1);
+               __print_result(&root_caller_sorted, session, caller_lines, 1);
        if (alloc_flag)
-               __print_result(&root_alloc_sorted, alloc_lines, 0);
+               __print_result(&root_alloc_sorted, session, alloc_lines, 0);
        print_summary();
 }
 
+struct sort_dimension {
+       const char              name[20];
+       sort_fn_t               cmp;
+       struct list_head        list;
+};
+
+static LIST_HEAD(caller_sort);
+static LIST_HEAD(alloc_sort);
+
 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
-                       sort_fn_t sort_fn)
+                       struct list_head *sort_list)
 {
        struct rb_node **new = &(root->rb_node);
        struct rb_node *parent = NULL;
+       struct sort_dimension *sort;
 
        while (*new) {
                struct alloc_stat *this;
-               int cmp;
+               int cmp = 0;
 
                this = rb_entry(*new, struct alloc_stat, node);
                parent = *new;
 
-               cmp = sort_fn(data, this);
+               list_for_each_entry(sort, sort_list, list) {
+                       cmp = sort->cmp(data, this);
+                       if (cmp)
+                               break;
+               }
 
                if (cmp > 0)
                        new = &((*new)->rb_left);
@@ -387,7 +467,7 @@ static void sort_insert(struct rb_root *root, struct alloc_stat *data,
 }
 
 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
-                         sort_fn_t sort_fn)
+                         struct list_head *sort_list)
 {
        struct rb_node *node;
        struct alloc_stat *data;
@@ -399,32 +479,45 @@ static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
 
                rb_erase(node, root);
                data = rb_entry(node, struct alloc_stat, node);
-               sort_insert(root_sorted, data, sort_fn);
+               sort_insert(root_sorted, data, sort_list);
        }
 }
 
 static void sort_result(void)
 {
-       __sort_result(&root_alloc_stat, &root_alloc_sorted, alloc_sort_fn);
-       __sort_result(&root_caller_stat, &root_caller_sorted, caller_sort_fn);
+       __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
+       __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
 }
 
 static int __cmd_kmem(void)
 {
+       int err = -EINVAL;
+       struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0, false);
+       if (session == NULL)
+               return -ENOMEM;
+
+       if (perf_session__create_kernel_maps(session) < 0)
+               goto out_delete;
+
+       if (!perf_session__has_traces(session, "kmem record"))
+               goto out_delete;
+
        setup_pager();
-       read_events();
+       err = perf_session__process_events(session, &event_ops);
+       if (err != 0)
+               goto out_delete;
        sort_result();
-       print_result();
-
-       return 0;
+       print_result(session);
+out_delete:
+       perf_session__delete(session);
+       return err;
 }
 
 static const char * const kmem_usage[] = {
-       "perf kmem [<options>] {record}",
+       "perf kmem [<options>] {record|stat}",
        NULL
 };
 
-
 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
 {
        if (l->ptr < r->ptr)
@@ -434,6 +527,11 @@ static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
        return 0;
 }
 
+static struct sort_dimension ptr_sort_dimension = {
+       .name   = "ptr",
+       .cmp    = ptr_cmp,
+};
+
 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
 {
        if (l->call_site < r->call_site)
@@ -443,6 +541,25 @@ static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
        return 0;
 }
 
+static struct sort_dimension callsite_sort_dimension = {
+       .name   = "callsite",
+       .cmp    = callsite_cmp,
+};
+
+static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
+{
+       if (l->hit < r->hit)
+               return -1;
+       else if (l->hit > r->hit)
+               return 1;
+       return 0;
+}
+
+static struct sort_dimension hit_sort_dimension = {
+       .name   = "hit",
+       .cmp    = hit_cmp,
+};
+
 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
 {
        if (l->bytes_alloc < r->bytes_alloc)
@@ -452,43 +569,122 @@ static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
        return 0;
 }
 
-static int parse_sort_opt(const struct option *opt __used,
-                         const char *arg, int unset __used)
+static struct sort_dimension bytes_sort_dimension = {
+       .name   = "bytes",
+       .cmp    = bytes_cmp,
+};
+
+static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
 {
-       sort_fn_t sort_fn;
+       double x, y;
 
-       if (!arg)
+       x = fragmentation(l->bytes_req, l->bytes_alloc);
+       y = fragmentation(r->bytes_req, r->bytes_alloc);
+
+       if (x < y)
                return -1;
+       else if (x > y)
+               return 1;
+       return 0;
+}
 
-       if (strcmp(arg, "ptr") == 0)
-               sort_fn = ptr_cmp;
-       else if (strcmp(arg, "call_site") == 0)
-               sort_fn = callsite_cmp;
-       else if (strcmp(arg, "bytes") == 0)
-               sort_fn = bytes_cmp;
-       else
+static struct sort_dimension frag_sort_dimension = {
+       .name   = "frag",
+       .cmp    = frag_cmp,
+};
+
+static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
+{
+       if (l->pingpong < r->pingpong)
                return -1;
+       else if (l->pingpong > r->pingpong)
+               return 1;
+       return 0;
+}
 
-       if (caller_flag > alloc_flag)
-               caller_sort_fn = sort_fn;
-       else
-               alloc_sort_fn = sort_fn;
+static struct sort_dimension pingpong_sort_dimension = {
+       .name   = "pingpong",
+       .cmp    = pingpong_cmp,
+};
+
+static struct sort_dimension *avail_sorts[] = {
+       &ptr_sort_dimension,
+       &callsite_sort_dimension,
+       &hit_sort_dimension,
+       &bytes_sort_dimension,
+       &frag_sort_dimension,
+       &pingpong_sort_dimension,
+};
+
+#define NUM_AVAIL_SORTS        \
+       (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
+
+static int sort_dimension__add(const char *tok, struct list_head *list)
+{
+       struct sort_dimension *sort;
+       int i;
+
+       for (i = 0; i < NUM_AVAIL_SORTS; i++) {
+               if (!strcmp(avail_sorts[i]->name, tok)) {
+                       sort = malloc(sizeof(*sort));
+                       if (!sort)
+                               die("malloc");
+                       memcpy(sort, avail_sorts[i], sizeof(*sort));
+                       list_add_tail(&sort->list, list);
+                       return 0;
+               }
+       }
+
+       return -1;
+}
 
+static int setup_sorting(struct list_head *sort_list, const char *arg)
+{
+       char *tok;
+       char *str = strdup(arg);
+
+       if (!str)
+               die("strdup");
+
+       while (true) {
+               tok = strsep(&str, ",");
+               if (!tok)
+                       break;
+               if (sort_dimension__add(tok, sort_list) < 0) {
+                       error("Unknown --sort key: '%s'", tok);
+                       return -1;
+               }
+       }
+
+       free(str);
        return 0;
 }
 
-static int parse_stat_opt(const struct option *opt __used,
+static int parse_sort_opt(const struct option *opt __used,
                          const char *arg, int unset __used)
 {
        if (!arg)
                return -1;
 
-       if (strcmp(arg, "alloc") == 0)
-               alloc_flag = (caller_flag + 1);
-       else if (strcmp(arg, "caller") == 0)
-               caller_flag = (alloc_flag + 1);
+       if (caller_flag > alloc_flag)
+               return setup_sorting(&caller_sort, arg);
        else
-               return -1;
+               return setup_sorting(&alloc_sort, arg);
+
+       return 0;
+}
+
+static int parse_caller_opt(const struct option *opt __used,
+                         const char *arg __used, int unset __used)
+{
+       caller_flag = (alloc_flag + 1);
+       return 0;
+}
+
+static int parse_alloc_opt(const struct option *opt __used,
+                         const char *arg __used, int unset __used)
+{
+       alloc_flag = (caller_flag + 1);
        return 0;
 }
 
@@ -513,15 +709,19 @@ static int parse_line_opt(const struct option *opt __used,
 static const struct option kmem_options[] = {
        OPT_STRING('i', "input", &input_name, "file",
                   "input file name"),
-       OPT_CALLBACK(0, "stat", NULL, "<alloc>|<caller>",
-                    "stat selector, Pass 'alloc' or 'caller'.",
-                    parse_stat_opt),
-       OPT_CALLBACK('s', "sort", NULL, "key",
-                    "sort by key: ptr, call_site, hit, bytes",
+       OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
+                          "show per-callsite statistics",
+                          parse_caller_opt),
+       OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
+                          "show per-allocation statistics",
+                          parse_alloc_opt),
+       OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
+                    "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
                     parse_sort_opt),
        OPT_CALLBACK('l', "line", NULL, "num",
-                    "show n lins",
+                    "show n lines",
                     parse_line_opt),
+       OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
        OPT_END()
 };
 
@@ -529,7 +729,6 @@ static const char *record_args[] = {
        "record",
        "-a",
        "-R",
-       "-M",
        "-f",
        "-c", "1",
        "-e", "kmem:kmalloc",
@@ -559,20 +758,27 @@ static int __cmd_record(int argc, const char **argv)
 
 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
 {
-       symbol__init(0);
-
        argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
 
-       if (argc && !strncmp(argv[0], "rec", 3))
-               return __cmd_record(argc, argv);
-       else if (argc)
+       if (!argc)
                usage_with_options(kmem_usage, kmem_options);
 
-       if (!alloc_sort_fn)
-               alloc_sort_fn = bytes_cmp;
-       if (!caller_sort_fn)
-               caller_sort_fn = bytes_cmp;
+       symbol__init();
+
+       if (!strncmp(argv[0], "rec", 3)) {
+               return __cmd_record(argc, argv);
+       } else if (!strcmp(argv[0], "stat")) {
+               setup_cpunode_map();
+
+               if (list_empty(&caller_sort))
+                       setup_sorting(&caller_sort, default_sort_order);
+               if (list_empty(&alloc_sort))
+                       setup_sorting(&alloc_sort, default_sort_order);
+
+               return __cmd_kmem();
+       } else
+               usage_with_options(kmem_usage, kmem_options);
 
-       return __cmd_kmem();
+       return 0;
 }