bpf: Add tests for the LRU bpf_htab
authorMartin KaFai Lau <kafai@fb.com>
Fri, 11 Nov 2016 18:55:11 +0000 (10:55 -0800)
committerDavid S. Miller <davem@davemloft.net>
Tue, 15 Nov 2016 16:50:43 +0000 (11:50 -0500)
This patch has some unit tests and a test_lru_dist.

The test_lru_dist reads in the numeric keys from a file.
The files used here are generated by a modified fio-genzipf tool
originated from the fio test suit.  The sample data file can be
found here: https://github.com/iamkafai/bpf-lru

The zipf.* data files have 100k numeric keys and the key is also
ranged from 1 to 100k.

The test_lru_dist outputs the number of unique keys (nr_unique).
F.e. The following means, 61239 of them is unique out of 100k keys.
nr_misses means it cannot be found in the LRU map, so nr_misses
must be >= nr_unique. test_lru_dist also simulates a perfect LRU
map as a comparison:

[root@arch-fb-vm1 ~]# ~/devshare/fb-kernel/linux/samples/bpf/test_lru_dist \
/root/zipf.100k.a1_01.out 4000 1
...
test_parallel_lru_dist (map_type:9 map_flags:0x0):
    task:0 BPF LRU: nr_unique:23093(/100000) nr_misses:31603(/100000)
    task:0 Perfect LRU: nr_unique:23093(/100000 nr_misses:34328(/100000)
....
test_parallel_lru_dist (map_type:9 map_flags:0x2):
    task:0 BPF LRU: nr_unique:23093(/100000) nr_misses:31710(/100000)
    task:0 Perfect LRU: nr_unique:23093(/100000 nr_misses:34328(/100000)

[root@arch-fb-vm1 ~]# ~/devshare/fb-kernel/linux/samples/bpf/test_lru_dist \
/root/zipf.100k.a0_01.out 40000 1
...
test_parallel_lru_dist (map_type:9 map_flags:0x0):
    task:0 BPF LRU: nr_unique:61239(/100000) nr_misses:67054(/100000)
    task:0 Perfect LRU: nr_unique:61239(/100000 nr_misses:66993(/100000)
...
test_parallel_lru_dist (map_type:9 map_flags:0x2):
    task:0 BPF LRU: nr_unique:61239(/100000) nr_misses:67068(/100000)
    task:0 Perfect LRU: nr_unique:61239(/100000 nr_misses:66993(/100000)

LRU map has also been added to map_perf_test:
/* Global LRU */
[root@kerneltest003.31.prn1 ~]# for i in 1 4 8; do echo -n "$i cpus: "; \
./map_perf_test 16 $i | awk '{r += $3}END{print r " updates"}'; done
 1 cpus: 2934082 updates
 4 cpus: 7391434 updates
 8 cpus: 6500576 updates

/* Percpu LRU */
[root@kerneltest003.31.prn1 ~]# for i in 1 4 8; do echo -n "$i cpus: "; \
./map_perf_test 32 $i | awk '{r += $3}END{print r " updates"}'; done
  1 cpus: 2896553 updates
  4 cpus: 9766395 updates
  8 cpus: 17460553 updates

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
samples/bpf/Makefile
samples/bpf/map_perf_test_kern.c
samples/bpf/map_perf_test_user.c
samples/bpf/test_lru_dist.c [new file with mode: 0644]
tools/testing/selftests/bpf/Makefile
tools/testing/selftests/bpf/test_lru_map.c [new file with mode: 0644]

index ac87f9c068ae72a96ec63fedc7ddc51f665fa081..f394ac616ed8ea37b1929236ce862ac2260039d7 100644 (file)
@@ -2,6 +2,7 @@
 obj- := dummy.o
 
 # List of programs to build
+hostprogs-y := test_lru_dist
 hostprogs-y += sock_example
 hostprogs-y += fds_example
 hostprogs-y += sockex1
@@ -28,6 +29,7 @@ hostprogs-y += trace_event
 hostprogs-y += sampleip
 hostprogs-y += tc_l2_redirect
 
+test_lru_dist-objs := test_lru_dist.o libbpf.o
 sock_example-objs := sock_example.o libbpf.o
 fds_example-objs := bpf_load.o libbpf.o fds_example.o
 sockex1-objs := bpf_load.o libbpf.o sockex1_user.o
index 311538e5a7016e7bfdd250779d60491051ef8b4b..7ee1574c8ccff49ad7a2a301c4462f66e46d88da 100644 (file)
@@ -19,6 +19,21 @@ struct bpf_map_def SEC("maps") hash_map = {
        .max_entries = MAX_ENTRIES,
 };
 
+struct bpf_map_def SEC("maps") lru_hash_map = {
+       .type = BPF_MAP_TYPE_LRU_HASH,
+       .key_size = sizeof(u32),
+       .value_size = sizeof(long),
+       .max_entries = 10000,
+};
+
+struct bpf_map_def SEC("maps") percpu_lru_hash_map = {
+       .type = BPF_MAP_TYPE_LRU_HASH,
+       .key_size = sizeof(u32),
+       .value_size = sizeof(long),
+       .max_entries = 10000,
+       .map_flags = BPF_F_NO_COMMON_LRU,
+};
+
 struct bpf_map_def SEC("maps") percpu_hash_map = {
        .type = BPF_MAP_TYPE_PERCPU_HASH,
        .key_size = sizeof(u32),
@@ -53,6 +68,7 @@ int stress_hmap(struct pt_regs *ctx)
        value = bpf_map_lookup_elem(&hash_map, &key);
        if (value)
                bpf_map_delete_elem(&hash_map, &key);
+
        return 0;
 }
 
@@ -96,5 +112,28 @@ int stress_percpu_hmap_alloc(struct pt_regs *ctx)
                bpf_map_delete_elem(&percpu_hash_map_alloc, &key);
        return 0;
 }
+
+SEC("kprobe/sys_getpid")
+int stress_lru_hmap_alloc(struct pt_regs *ctx)
+{
+       u32 key = bpf_get_prandom_u32();
+       long val = 1;
+
+       bpf_map_update_elem(&lru_hash_map, &key, &val, BPF_ANY);
+
+       return 0;
+}
+
+SEC("kprobe/sys_getppid")
+int stress_percpu_lru_hmap_alloc(struct pt_regs *ctx)
+{
+       u32 key = bpf_get_prandom_u32();
+       long val = 1;
+
+       bpf_map_update_elem(&percpu_lru_hash_map, &key, &val, BPF_ANY);
+
+       return 0;
+}
+
 char _license[] SEC("license") = "GPL";
 u32 _version SEC("version") = LINUX_VERSION_CODE;
index 3147377e8fd3c98fe455c6d96f5c774cb5392f11..9505b4d112f426790645ecd00c50263620ae68e6 100644 (file)
@@ -35,6 +35,8 @@ static __u64 time_get_ns(void)
 #define PERCPU_HASH_PREALLOC   (1 << 1)
 #define HASH_KMALLOC           (1 << 2)
 #define PERCPU_HASH_KMALLOC    (1 << 3)
+#define LRU_HASH_PREALLOC      (1 << 4)
+#define PERCPU_LRU_HASH_PREALLOC       (1 << 5)
 
 static int test_flags = ~0;
 
@@ -50,6 +52,30 @@ static void test_hash_prealloc(int cpu)
               cpu, MAX_CNT * 1000000000ll / (time_get_ns() - start_time));
 }
 
+static void test_lru_hash_prealloc(int cpu)
+{
+       __u64 start_time;
+       int i;
+
+       start_time = time_get_ns();
+       for (i = 0; i < MAX_CNT; i++)
+               syscall(__NR_getpid);
+       printf("%d:lru_hash_map_perf pre-alloc %lld events per sec\n",
+              cpu, MAX_CNT * 1000000000ll / (time_get_ns() - start_time));
+}
+
+static void test_percpu_lru_hash_prealloc(int cpu)
+{
+       __u64 start_time;
+       int i;
+
+       start_time = time_get_ns();
+       for (i = 0; i < MAX_CNT; i++)
+               syscall(__NR_getppid);
+       printf("%d:lru_hash_map_perf pre-alloc %lld events per sec\n",
+              cpu, MAX_CNT * 1000000000ll / (time_get_ns() - start_time));
+}
+
 static void test_percpu_hash_prealloc(int cpu)
 {
        __u64 start_time;
@@ -105,6 +131,12 @@ static void loop(int cpu)
 
        if (test_flags & PERCPU_HASH_KMALLOC)
                test_percpu_hash_kmalloc(cpu);
+
+       if (test_flags & LRU_HASH_PREALLOC)
+               test_lru_hash_prealloc(cpu);
+
+       if (test_flags & PERCPU_LRU_HASH_PREALLOC)
+               test_percpu_lru_hash_prealloc(cpu);
 }
 
 static void run_perf_test(int tasks)
diff --git a/samples/bpf/test_lru_dist.c b/samples/bpf/test_lru_dist.c
new file mode 100644 (file)
index 0000000..2859977
--- /dev/null
@@ -0,0 +1,538 @@
+/*
+ * Copyright (c) 2016 Facebook
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ */
+#define _GNU_SOURCE
+#include <linux/types.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <linux/bpf.h>
+#include <errno.h>
+#include <string.h>
+#include <assert.h>
+#include <sched.h>
+#include <sys/wait.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <stdlib.h>
+#include <time.h>
+#include "libbpf.h"
+
+#define min(a, b) ((a) < (b) ? (a) : (b))
+#define offsetof(TYPE, MEMBER) ((size_t)&((TYPE *)0)->MEMBER)
+#define container_of(ptr, type, member) ({                     \
+       const typeof( ((type *)0)->member ) *__mptr = (ptr);    \
+       (type *)( (char *)__mptr - offsetof(type,member) );})
+
+static int nr_cpus;
+static unsigned long long *dist_keys;
+static unsigned int dist_key_counts;
+
+struct list_head {
+       struct list_head *next, *prev;
+};
+
+static inline void INIT_LIST_HEAD(struct list_head *list)
+{
+       list->next = list;
+       list->prev = list;
+}
+
+static inline int list_empty(const struct list_head *head)
+{
+       return head->next == head;
+}
+
+static inline void __list_add(struct list_head *new,
+                             struct list_head *prev,
+                             struct list_head *next)
+{
+       next->prev = new;
+       new->next = next;
+       new->prev = prev;
+       prev->next = new;
+}
+
+static inline void list_add(struct list_head *new, struct list_head *head)
+{
+       __list_add(new, head, head->next);
+}
+
+static inline void __list_del(struct list_head *prev, struct list_head *next)
+{
+       next->prev = prev;
+       prev->next = next;
+}
+
+static inline void __list_del_entry(struct list_head *entry)
+{
+       __list_del(entry->prev, entry->next);
+}
+
+static inline void list_move(struct list_head *list, struct list_head *head)
+{
+       __list_del_entry(list);
+       list_add(list, head);
+}
+
+#define list_entry(ptr, type, member) \
+       container_of(ptr, type, member)
+
+#define list_last_entry(ptr, type, member) \
+       list_entry((ptr)->prev, type, member)
+
+struct pfect_lru_node {
+       struct list_head list;
+       unsigned long long key;
+};
+
+struct pfect_lru {
+       struct list_head list;
+       struct pfect_lru_node *free_nodes;
+       unsigned int cur_size;
+       unsigned int lru_size;
+       unsigned int nr_unique;
+       unsigned int nr_misses;
+       unsigned int total;
+       int map_fd;
+};
+
+static void pfect_lru_init(struct pfect_lru *lru, unsigned int lru_size,
+                          unsigned int nr_possible_elems)
+{
+       lru->map_fd = bpf_create_map(BPF_MAP_TYPE_HASH,
+                                    sizeof(unsigned long long),
+                                    sizeof(struct pfect_lru_node *),
+                                    nr_possible_elems, 0);
+       assert(lru->map_fd != -1);
+
+       lru->free_nodes = malloc(lru_size * sizeof(struct pfect_lru_node));
+       assert(lru->free_nodes);
+
+       INIT_LIST_HEAD(&lru->list);
+       lru->cur_size = 0;
+       lru->lru_size = lru_size;
+       lru->nr_unique = lru->nr_misses = lru->total = 0;
+}
+
+static void pfect_lru_destroy(struct pfect_lru *lru)
+{
+       close(lru->map_fd);
+       free(lru->free_nodes);
+}
+
+static int pfect_lru_lookup_or_insert(struct pfect_lru *lru,
+                                     unsigned long long key)
+{
+       struct pfect_lru_node *node = NULL;
+       int seen = 0;
+
+       lru->total++;
+       if (!bpf_lookup_elem(lru->map_fd, &key, &node)) {
+               if (node) {
+                       list_move(&node->list, &lru->list);
+                       return 1;
+               }
+               seen = 1;
+       }
+
+       if (lru->cur_size < lru->lru_size) {
+               node =  &lru->free_nodes[lru->cur_size++];
+               INIT_LIST_HEAD(&node->list);
+       } else {
+               struct pfect_lru_node *null_node = NULL;
+
+               node = list_last_entry(&lru->list,
+                                      struct pfect_lru_node,
+                                      list);
+               bpf_update_elem(lru->map_fd, &node->key, &null_node, BPF_EXIST);
+       }
+
+       node->key = key;
+       list_move(&node->list, &lru->list);
+
+       lru->nr_misses++;
+       if (seen) {
+               assert(!bpf_update_elem(lru->map_fd, &key, &node, BPF_EXIST));
+       } else {
+               lru->nr_unique++;
+               assert(!bpf_update_elem(lru->map_fd, &key, &node, BPF_NOEXIST));
+       }
+
+       return seen;
+}
+
+static unsigned int read_keys(const char *dist_file,
+                             unsigned long long **keys)
+{
+       struct stat fst;
+       unsigned long long *retkeys;
+       unsigned int counts = 0;
+       int dist_fd;
+       char *b, *l;
+       int i;
+
+       dist_fd = open(dist_file, 0);
+       assert(dist_fd != -1);
+
+       assert(fstat(dist_fd, &fst) == 0);
+       b = malloc(fst.st_size);
+       assert(b);
+
+       assert(read(dist_fd, b, fst.st_size) == fst.st_size);
+       close(dist_fd);
+       for (i = 0; i < fst.st_size; i++) {
+               if (b[i] == '\n')
+                       counts++;
+       }
+       counts++; /* in case the last line has no \n */
+
+       retkeys = malloc(counts * sizeof(unsigned long long));
+       assert(retkeys);
+
+       counts = 0;
+       for (l = strtok(b, "\n"); l; l = strtok(NULL, "\n"))
+               retkeys[counts++] = strtoull(l, NULL, 10);
+       free(b);
+
+       *keys = retkeys;
+
+       return counts;
+}
+
+static int create_map(int map_type, int map_flags, unsigned int size)
+{
+       int map_fd;
+
+       map_fd = bpf_create_map(map_type, sizeof(unsigned long long),
+                               sizeof(unsigned long long), size, map_flags);
+
+       if (map_fd == -1)
+               perror("bpf_create_map");
+
+       return map_fd;
+}
+
+static int sched_next_online(int pid, int next_to_try)
+{
+       cpu_set_t cpuset;
+
+       if (next_to_try == nr_cpus)
+               return -1;
+
+       while (next_to_try < nr_cpus) {
+               CPU_ZERO(&cpuset);
+               CPU_SET(next_to_try++, &cpuset);
+               if (!sched_setaffinity(pid, sizeof(cpuset), &cpuset))
+                       break;
+       }
+
+       return next_to_try;
+}
+
+static void run_parallel(unsigned int tasks, void (*fn)(int i, void *data),
+                        void *data)
+{
+       int next_sched_cpu = 0;
+       pid_t pid[tasks];
+       int i;
+
+       for (i = 0; i < tasks; i++) {
+               pid[i] = fork();
+               if (pid[i] == 0) {
+                       next_sched_cpu = sched_next_online(0, next_sched_cpu);
+                       fn(i, data);
+                       exit(0);
+               } else if (pid[i] == -1) {
+                       printf("couldn't spawn #%d process\n", i);
+                       exit(1);
+               }
+               /* It is mostly redundant and just allow the parent
+                * process to update next_shced_cpu for the next child
+                * process
+                */
+               next_sched_cpu = sched_next_online(pid[i], next_sched_cpu);
+       }
+       for (i = 0; i < tasks; i++) {
+               int status;
+
+               assert(waitpid(pid[i], &status, 0) == pid[i]);
+               assert(status == 0);
+       }
+}
+
+static void do_test_lru_dist(int task, void *data)
+{
+       unsigned int nr_misses = 0;
+       struct pfect_lru pfect_lru;
+       unsigned long long key, value = 1234;
+       unsigned int i;
+
+       unsigned int lru_map_fd = ((unsigned int *)data)[0];
+       unsigned int lru_size = ((unsigned int *)data)[1];
+       unsigned long long key_offset = task * dist_key_counts;
+
+       pfect_lru_init(&pfect_lru, lru_size, dist_key_counts);
+
+       for (i = 0; i < dist_key_counts; i++) {
+               key = dist_keys[i] + key_offset;
+
+               pfect_lru_lookup_or_insert(&pfect_lru, key);
+
+               if (!bpf_lookup_elem(lru_map_fd, &key, &value))
+                       continue;
+
+               if (bpf_update_elem(lru_map_fd, &key, &value, BPF_NOEXIST)) {
+                       printf("bpf_update_elem(lru_map_fd, %llu): errno:%d\n",
+                              key, errno);
+                       assert(0);
+               }
+
+               nr_misses++;
+       }
+
+       printf("    task:%d BPF LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
+              task, pfect_lru.nr_unique, dist_key_counts, nr_misses,
+              dist_key_counts);
+       printf("    task:%d Perfect LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
+              task, pfect_lru.nr_unique, pfect_lru.total,
+              pfect_lru.nr_misses, pfect_lru.total);
+
+       pfect_lru_destroy(&pfect_lru);
+       close(lru_map_fd);
+}
+
+static void test_parallel_lru_dist(int map_type, int map_flags,
+                                  int nr_tasks, unsigned int lru_size)
+{
+       int child_data[2];
+       int lru_map_fd;
+
+       printf("%s (map_type:%d map_flags:0x%X):\n", __func__, map_type,
+              map_flags);
+
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               lru_map_fd = create_map(map_type, map_flags,
+                                       nr_cpus * lru_size);
+       else
+               lru_map_fd = create_map(map_type, map_flags,
+                                       nr_tasks * lru_size);
+       assert(lru_map_fd != -1);
+
+       child_data[0] = lru_map_fd;
+       child_data[1] = lru_size;
+
+       run_parallel(nr_tasks, do_test_lru_dist, child_data);
+
+       close(lru_map_fd);
+}
+
+static void test_lru_loss0(int map_type, int map_flags)
+{
+       unsigned long long key, value[nr_cpus];
+       unsigned int old_unused_losses = 0;
+       unsigned int new_unused_losses = 0;
+       unsigned int used_losses = 0;
+       int map_fd;
+
+       printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
+              map_flags);
+
+       assert(sched_next_online(0, 0) != -1);
+
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               map_fd = create_map(map_type, map_flags, 900 * nr_cpus);
+       else
+               map_fd = create_map(map_type, map_flags, 900);
+
+       assert(map_fd != -1);
+
+       value[0] = 1234;
+
+       for (key = 1; key <= 1000; key++) {
+               int start_key, end_key;
+
+               assert(bpf_update_elem(map_fd, &key, value, BPF_NOEXIST) == 0);
+
+               start_key = 101;
+               end_key = min(key, 900);
+
+               while (start_key <= end_key) {
+                       bpf_lookup_elem(map_fd, &start_key, value);
+                       start_key++;
+               }
+       }
+
+       for (key = 1; key <= 1000; key++) {
+               if (bpf_lookup_elem(map_fd, &key, value)) {
+                       if (key <= 100)
+                               old_unused_losses++;
+                       else if (key <= 900)
+                               used_losses++;
+                       else
+                               new_unused_losses++;
+               }
+       }
+
+       close(map_fd);
+
+       printf("older-elem-losses:%d(/100) active-elem-losses:%d(/800) "
+              "newer-elem-losses:%d(/100)\n",
+              old_unused_losses, used_losses, new_unused_losses);
+}
+
+static void test_lru_loss1(int map_type, int map_flags)
+{
+       unsigned long long key, value[nr_cpus];
+       int map_fd;
+       unsigned int nr_losses = 0;
+
+       printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
+              map_flags);
+
+       assert(sched_next_online(0, 0) != -1);
+
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               map_fd = create_map(map_type, map_flags, 1000 * nr_cpus);
+       else
+               map_fd = create_map(map_type, map_flags, 1000);
+
+       assert(map_fd != -1);
+
+       value[0] = 1234;
+
+       for (key = 1; key <= 1000; key++)
+               assert(!bpf_update_elem(map_fd, &key, value, BPF_NOEXIST));
+
+       for (key = 1; key <= 1000; key++) {
+               if (bpf_lookup_elem(map_fd, &key, value))
+                       nr_losses++;
+       }
+
+       close(map_fd);
+
+       printf("nr_losses:%d(/1000)\n", nr_losses);
+}
+
+static void do_test_parallel_lru_loss(int task, void *data)
+{
+       const unsigned int nr_stable_elems = 1000;
+       const unsigned int nr_repeats = 100000;
+
+       int map_fd = *(int *)data;
+       unsigned long long stable_base;
+       unsigned long long key, value[nr_cpus];
+       unsigned long long next_ins_key;
+       unsigned int nr_losses = 0;
+       unsigned int i;
+
+       stable_base = task * nr_repeats * 2 + 1;
+       next_ins_key = stable_base;
+       value[0] = 1234;
+       for (i = 0; i < nr_stable_elems; i++) {
+               assert(bpf_update_elem(map_fd, &next_ins_key, value,
+                                      BPF_NOEXIST) == 0);
+               next_ins_key++;
+       }
+
+       for (i = 0; i < nr_repeats; i++) {
+               int rn;
+
+               rn = rand();
+
+               if (rn % 10) {
+                       key = rn % nr_stable_elems + stable_base;
+                       bpf_lookup_elem(map_fd, &key, value);
+               } else {
+                       bpf_update_elem(map_fd, &next_ins_key, value,
+                                       BPF_NOEXIST);
+                       next_ins_key++;
+               }
+       }
+
+       key = stable_base;
+       for (i = 0; i < nr_stable_elems; i++) {
+               if (bpf_lookup_elem(map_fd, &key, value))
+                       nr_losses++;
+               key++;
+       }
+
+       printf("    task:%d nr_losses:%u\n", task, nr_losses);
+}
+
+static void test_parallel_lru_loss(int map_type, int map_flags, int nr_tasks)
+{
+       int map_fd;
+
+       printf("%s (map_type:%d map_flags:0x%X):\n", __func__, map_type,
+              map_flags);
+
+       /* Give 20% more than the active working set */
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               map_fd = create_map(map_type, map_flags,
+                                   nr_cpus * (1000 + 200));
+       else
+               map_fd = create_map(map_type, map_flags,
+                                   nr_tasks * (1000 + 200));
+
+       assert(map_fd != -1);
+
+       run_parallel(nr_tasks, do_test_parallel_lru_loss, &map_fd);
+
+       close(map_fd);
+}
+
+int main(int argc, char **argv)
+{
+       struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
+       int map_flags[] = {0, BPF_F_NO_COMMON_LRU};
+       const char *dist_file;
+       int nr_tasks = 1;
+       int lru_size;
+       int f;
+
+       if (argc < 4) {
+               printf("Usage: %s <dist-file> <lru-size> <nr-tasks>\n",
+                      argv[0]);
+               return -1;
+       }
+
+       dist_file = argv[1];
+       lru_size = atoi(argv[2]);
+       nr_tasks = atoi(argv[3]);
+
+       setbuf(stdout, NULL);
+
+       assert(!setrlimit(RLIMIT_MEMLOCK, &r));
+
+       srand(time(NULL));
+
+       nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
+       assert(nr_cpus != -1);
+       printf("nr_cpus:%d\n\n", nr_cpus);
+
+       nr_tasks = min(nr_tasks, nr_cpus);
+
+       dist_key_counts = read_keys(dist_file, &dist_keys);
+       if (!dist_key_counts) {
+               printf("%s has no key\n", dist_file);
+               return -1;
+       }
+
+       for (f = 0; f < sizeof(map_flags) / sizeof(*map_flags); f++) {
+               test_lru_loss0(BPF_MAP_TYPE_LRU_HASH, map_flags[f]);
+               test_lru_loss1(BPF_MAP_TYPE_LRU_HASH, map_flags[f]);
+               test_parallel_lru_loss(BPF_MAP_TYPE_LRU_HASH, map_flags[f],
+                                      nr_tasks);
+               test_parallel_lru_dist(BPF_MAP_TYPE_LRU_HASH, map_flags[f],
+                                      nr_tasks, lru_size);
+               printf("\n");
+       }
+
+       free(dist_keys);
+
+       return 0;
+}
index 4761e2d65ab80a9e0c9f325edcda72c2d397f812..7a5f24543a5f06fc92e2c0c139f6af7a05f9ea37 100644 (file)
@@ -1,8 +1,8 @@
-CFLAGS += -Wall -O2
+CFLAGS += -Wall -O2 -I../../../../usr/include
 
-test_objs = test_verifier test_maps
+test_objs = test_verifier test_maps test_lru_map
 
-TEST_PROGS := test_verifier test_maps test_kmod.sh
+TEST_PROGS := test_verifier test_maps test_lru_map test_kmod.sh
 TEST_FILES := $(test_objs)
 
 all: $(test_objs)
diff --git a/tools/testing/selftests/bpf/test_lru_map.c b/tools/testing/selftests/bpf/test_lru_map.c
new file mode 100644 (file)
index 0000000..627757e
--- /dev/null
@@ -0,0 +1,583 @@
+/*
+ * Copyright (c) 2016 Facebook
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of version 2 of the GNU General Public
+ * License as published by the Free Software Foundation.
+ */
+#define _GNU_SOURCE
+#include <stdio.h>
+#include <unistd.h>
+#include <errno.h>
+#include <string.h>
+#include <assert.h>
+#include <sched.h>
+#include <sys/wait.h>
+#include <stdlib.h>
+#include <time.h>
+#include "bpf_sys.h"
+
+#define LOCAL_FREE_TARGET      (128)
+#define PERCPU_FREE_TARGET     (16)
+
+static int nr_cpus;
+
+static int create_map(int map_type, int map_flags, unsigned int size)
+{
+       int map_fd;
+
+       map_fd = bpf_map_create(map_type, sizeof(unsigned long long),
+                               sizeof(unsigned long long), size, map_flags);
+
+       if (map_fd == -1)
+               perror("bpf_map_create");
+
+       return map_fd;
+}
+
+static int map_subset(int map0, int map1)
+{
+       unsigned long long next_key = 0;
+       unsigned long long value0[nr_cpus], value1[nr_cpus];
+       int ret;
+
+       while (!bpf_map_next_key(map1, &next_key, &next_key)) {
+               assert(!bpf_map_lookup(map1, &next_key, value1));
+               ret = bpf_map_lookup(map0, &next_key, value0);
+               if (ret) {
+                       printf("key:%llu not found from map. %s(%d)\n",
+                              next_key, strerror(errno), errno);
+                       return 0;
+               }
+               if (value0[0] != value1[0]) {
+                       printf("key:%llu value0:%llu != value1:%llu\n",
+                              next_key, value0[0], value1[0]);
+                       return 0;
+               }
+       }
+       return 1;
+}
+
+static int map_equal(int lru_map, int expected)
+{
+       return map_subset(lru_map, expected) && map_subset(expected, lru_map);
+}
+
+static int sched_next_online(int pid, int next_to_try)
+{
+       cpu_set_t cpuset;
+
+       if (next_to_try == nr_cpus)
+               return -1;
+
+       while (next_to_try < nr_cpus) {
+               CPU_ZERO(&cpuset);
+               CPU_SET(next_to_try++, &cpuset);
+               if (!sched_setaffinity(pid, sizeof(cpuset), &cpuset))
+                       break;
+       }
+
+       return next_to_try;
+}
+
+/* Size of the LRU amp is 2
+ * Add key=1 (+1 key)
+ * Add key=2 (+1 key)
+ * Lookup Key=1
+ * Add Key=3
+ *   => Key=2 will be removed by LRU
+ * Iterate map.  Only found key=1 and key=3
+ */
+static void test_lru_sanity0(int map_type, int map_flags)
+{
+       unsigned long long key, value[nr_cpus];
+       int lru_map_fd, expected_map_fd;
+
+       printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
+              map_flags);
+
+       assert(sched_next_online(0, 0) != -1);
+
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               lru_map_fd = create_map(map_type, map_flags, 2 * nr_cpus);
+       else
+               lru_map_fd = create_map(map_type, map_flags, 2);
+       assert(lru_map_fd != -1);
+
+       expected_map_fd = create_map(BPF_MAP_TYPE_HASH, 0, 2);
+       assert(expected_map_fd != -1);
+
+       value[0] = 1234;
+
+       /* insert key=1 element */
+
+       key = 1;
+       assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+       assert(!bpf_map_update(expected_map_fd, &key, value, BPF_NOEXIST));
+
+       /* BPF_NOEXIST means: add new element if it doesn't exist */
+       assert(bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST) == -1 &&
+              /* key=1 already exists */
+              errno == EEXIST);
+
+       assert(bpf_map_update(lru_map_fd, &key, value, -1) == -1 &&
+              errno == EINVAL);
+
+       /* insert key=2 element */
+
+       /* check that key=2 is not found */
+       key = 2;
+       assert(bpf_map_lookup(lru_map_fd, &key, value) == -1 &&
+              errno == ENOENT);
+
+       /* BPF_EXIST means: update existing element */
+       assert(bpf_map_update(lru_map_fd, &key, value, BPF_EXIST) == -1 &&
+              /* key=2 is not there */
+              errno == ENOENT);
+
+       assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+
+       /* insert key=3 element */
+
+       /* check that key=3 is not found */
+       key = 3;
+       assert(bpf_map_lookup(lru_map_fd, &key, value) == -1 &&
+              errno == ENOENT);
+
+       /* check that key=1 can be found and mark the ref bit to
+        * stop LRU from removing key=1
+        */
+       key = 1;
+       assert(!bpf_map_lookup(lru_map_fd, &key, value));
+       assert(value[0] == 1234);
+
+       key = 3;
+       assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+       assert(!bpf_map_update(expected_map_fd, &key, value, BPF_NOEXIST));
+
+       /* key=2 has been removed from the LRU */
+       key = 2;
+       assert(bpf_map_lookup(lru_map_fd, &key, value) == -1);
+
+       assert(map_equal(lru_map_fd, expected_map_fd));
+
+       close(expected_map_fd);
+       close(lru_map_fd);
+
+       printf("Pass\n");
+}
+
+/* Size of the LRU map is 1.5*tgt_free
+ * Insert 1 to tgt_free (+tgt_free keys)
+ * Lookup 1 to tgt_free/2
+ * Insert 1+tgt_free to 2*tgt_free (+tgt_free keys)
+ * => 1+tgt_free/2 to LOCALFREE_TARGET will be removed by LRU
+ */
+static void test_lru_sanity1(int map_type, int map_flags, unsigned int tgt_free)
+{
+       unsigned long long key, end_key, value[nr_cpus];
+       int lru_map_fd, expected_map_fd;
+       unsigned int batch_size;
+       unsigned int map_size;
+
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               /* Ther percpu lru list (i.e each cpu has its own LRU
+                * list) does not have a local free list.  Hence,
+                * it will only free old nodes till there is no free
+                * from the LRU list.  Hence, this test does not apply
+                * to BPF_F_NO_COMMON_LRU
+                */
+               return;
+
+       printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
+              map_flags);
+
+       assert(sched_next_online(0, 0) != -1);
+
+       batch_size = tgt_free / 2;
+       assert(batch_size * 2 == tgt_free);
+
+       map_size = tgt_free + batch_size;
+       lru_map_fd = create_map(map_type, map_flags, map_size);
+       assert(lru_map_fd != -1);
+
+       expected_map_fd = create_map(BPF_MAP_TYPE_HASH, 0, map_size);
+       assert(expected_map_fd != -1);
+
+       value[0] = 1234;
+
+       /* Insert 1 to tgt_free (+tgt_free keys) */
+       end_key = 1 + tgt_free;
+       for (key = 1; key < end_key; key++)
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+
+       /* Lookup 1 to tgt_free/2 */
+       end_key = 1 + batch_size;
+       for (key = 1; key < end_key; key++) {
+               assert(!bpf_map_lookup(lru_map_fd, &key, value));
+               assert(!bpf_map_update(expected_map_fd, &key, value,
+                                      BPF_NOEXIST));
+       }
+
+       /* Insert 1+tgt_free to 2*tgt_free
+        * => 1+tgt_free/2 to LOCALFREE_TARGET will be
+        * removed by LRU
+        */
+       key = 1 + tgt_free;
+       end_key = key + tgt_free;
+       for (; key < end_key; key++) {
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+               assert(!bpf_map_update(expected_map_fd, &key, value,
+                                      BPF_NOEXIST));
+       }
+
+       assert(map_equal(lru_map_fd, expected_map_fd));
+
+       close(expected_map_fd);
+       close(lru_map_fd);
+
+       printf("Pass\n");
+}
+
+/* Size of the LRU map 1.5 * tgt_free
+ * Insert 1 to tgt_free (+tgt_free keys)
+ * Update 1 to tgt_free/2
+ *   => The original 1 to tgt_free/2 will be removed due to
+ *      the LRU shrink process
+ * Re-insert 1 to tgt_free/2 again and do a lookup immeidately
+ * Insert 1+tgt_free to tgt_free*3/2
+ * Insert 1+tgt_free*3/2 to tgt_free*5/2
+ *   => Key 1+tgt_free to tgt_free*3/2
+ *      will be removed from LRU because it has never
+ *      been lookup and ref bit is not set
+ */
+static void test_lru_sanity2(int map_type, int map_flags, unsigned int tgt_free)
+{
+       unsigned long long key, value[nr_cpus];
+       unsigned long long end_key;
+       int lru_map_fd, expected_map_fd;
+       unsigned int batch_size;
+       unsigned int map_size;
+
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               /* Ther percpu lru list (i.e each cpu has its own LRU
+                * list) does not have a local free list.  Hence,
+                * it will only free old nodes till there is no free
+                * from the LRU list.  Hence, this test does not apply
+                * to BPF_F_NO_COMMON_LRU
+                */
+               return;
+
+       printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
+              map_flags);
+
+       assert(sched_next_online(0, 0) != -1);
+
+       batch_size = tgt_free / 2;
+       assert(batch_size * 2 == tgt_free);
+
+       map_size = tgt_free + batch_size;
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               lru_map_fd = create_map(map_type, map_flags,
+                                       map_size * nr_cpus);
+       else
+               lru_map_fd = create_map(map_type, map_flags, map_size);
+       assert(lru_map_fd != -1);
+
+       expected_map_fd = create_map(BPF_MAP_TYPE_HASH, 0, map_size);
+       assert(expected_map_fd != -1);
+
+       value[0] = 1234;
+
+       /* Insert 1 to tgt_free (+tgt_free keys) */
+       end_key = 1 + tgt_free;
+       for (key = 1; key < end_key; key++)
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+
+       /* Any bpf_map_update will require to acquire a new node
+        * from LRU first.
+        *
+        * The local list is running out of free nodes.
+        * It gets from the global LRU list which tries to
+        * shrink the inactive list to get tgt_free
+        * number of free nodes.
+        *
+        * Hence, the oldest key 1 to tgt_free/2
+        * are removed from the LRU list.
+        */
+       key = 1;
+       if (map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+               assert(!bpf_map_delete(lru_map_fd, &key));
+       } else {
+               assert(bpf_map_update(lru_map_fd, &key, value, BPF_EXIST));
+       }
+
+       /* Re-insert 1 to tgt_free/2 again and do a lookup
+        * immeidately.
+        */
+       end_key = 1 + batch_size;
+       value[0] = 4321;
+       for (key = 1; key < end_key; key++) {
+               assert(bpf_map_lookup(lru_map_fd, &key, value));
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+               assert(!bpf_map_lookup(lru_map_fd, &key, value));
+               assert(value[0] == 4321);
+               assert(!bpf_map_update(expected_map_fd, &key, value,
+                                      BPF_NOEXIST));
+       }
+
+       value[0] = 1234;
+
+       /* Insert 1+tgt_free to tgt_free*3/2 */
+       end_key = 1 + tgt_free + batch_size;
+       for (key = 1 + tgt_free; key < end_key; key++)
+               /* These newly added but not referenced keys will be
+                * gone during the next LRU shrink.
+                */
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+
+       /* Insert 1+tgt_free*3/2 to  tgt_free*5/2 */
+       end_key = key + tgt_free;
+       for (; key < end_key; key++) {
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+               assert(!bpf_map_update(expected_map_fd, &key, value,
+                                      BPF_NOEXIST));
+       }
+
+       assert(map_equal(lru_map_fd, expected_map_fd));
+
+       close(expected_map_fd);
+       close(lru_map_fd);
+
+       printf("Pass\n");
+}
+
+/* Size of the LRU map is 2*tgt_free
+ * It is to test the active/inactive list rotation
+ * Insert 1 to 2*tgt_free (+2*tgt_free keys)
+ * Lookup key 1 to tgt_free*3/2
+ * Add 1+2*tgt_free to tgt_free*5/2 (+tgt_free/2 keys)
+ *  => key 1+tgt_free*3/2 to 2*tgt_free are removed from LRU
+ */
+static void test_lru_sanity3(int map_type, int map_flags, unsigned int tgt_free)
+{
+       unsigned long long key, end_key, value[nr_cpus];
+       int lru_map_fd, expected_map_fd;
+       unsigned int batch_size;
+       unsigned int map_size;
+
+       printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
+              map_flags);
+
+       assert(sched_next_online(0, 0) != -1);
+
+       batch_size = tgt_free / 2;
+       assert(batch_size * 2 == tgt_free);
+
+       map_size = tgt_free * 2;
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               lru_map_fd = create_map(map_type, map_flags,
+                                       map_size * nr_cpus);
+       else
+               lru_map_fd = create_map(map_type, map_flags, map_size);
+       assert(lru_map_fd != -1);
+
+       expected_map_fd = create_map(BPF_MAP_TYPE_HASH, 0, map_size);
+       assert(expected_map_fd != -1);
+
+       value[0] = 1234;
+
+       /* Insert 1 to 2*tgt_free (+2*tgt_free keys) */
+       end_key = 1 + (2 * tgt_free);
+       for (key = 1; key < end_key; key++)
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+
+       /* Lookup key 1 to tgt_free*3/2 */
+       end_key = tgt_free + batch_size;
+       for (key = 1; key < end_key; key++) {
+               assert(!bpf_map_lookup(lru_map_fd, &key, value));
+               assert(!bpf_map_update(expected_map_fd, &key, value,
+                                      BPF_NOEXIST));
+       }
+
+       /* Add 1+2*tgt_free to tgt_free*5/2
+        * (+tgt_free/2 keys)
+        */
+       key = 2 * tgt_free + 1;
+       end_key = key + batch_size;
+       for (; key < end_key; key++) {
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+               assert(!bpf_map_update(expected_map_fd, &key, value,
+                                      BPF_NOEXIST));
+       }
+
+       assert(map_equal(lru_map_fd, expected_map_fd));
+
+       close(expected_map_fd);
+       close(lru_map_fd);
+
+       printf("Pass\n");
+}
+
+/* Test deletion */
+static void test_lru_sanity4(int map_type, int map_flags, unsigned int tgt_free)
+{
+       int lru_map_fd, expected_map_fd;
+       unsigned long long key, value[nr_cpus];
+       unsigned long long end_key;
+
+       printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
+              map_flags);
+
+       assert(sched_next_online(0, 0) != -1);
+
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               lru_map_fd = create_map(map_type, map_flags,
+                                       3 * tgt_free * nr_cpus);
+       else
+               lru_map_fd = create_map(map_type, map_flags, 3 * tgt_free);
+       assert(lru_map_fd != -1);
+
+       expected_map_fd = create_map(BPF_MAP_TYPE_HASH, 0,
+                                    3 * tgt_free);
+       assert(expected_map_fd != -1);
+
+       value[0] = 1234;
+
+       for (key = 1; key <= 2 * tgt_free; key++)
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+
+       key = 1;
+       assert(bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+
+       for (key = 1; key <= tgt_free; key++) {
+               assert(!bpf_map_lookup(lru_map_fd, &key, value));
+               assert(!bpf_map_update(expected_map_fd, &key, value,
+                                      BPF_NOEXIST));
+       }
+
+       for (; key <= 2 * tgt_free; key++) {
+               assert(!bpf_map_delete(lru_map_fd, &key));
+               assert(bpf_map_delete(lru_map_fd, &key));
+       }
+
+       end_key = key + 2 * tgt_free;
+       for (; key < end_key; key++) {
+               assert(!bpf_map_update(lru_map_fd, &key, value, BPF_NOEXIST));
+               assert(!bpf_map_update(expected_map_fd, &key, value,
+                                      BPF_NOEXIST));
+       }
+
+       assert(map_equal(lru_map_fd, expected_map_fd));
+
+       close(expected_map_fd);
+       close(lru_map_fd);
+
+       printf("Pass\n");
+}
+
+static void do_test_lru_sanity5(unsigned long long last_key, int map_fd)
+{
+       unsigned long long key, value[nr_cpus];
+
+       /* Ensure the last key inserted by previous CPU can be found */
+       assert(!bpf_map_lookup(map_fd, &last_key, value));
+
+       value[0] = 1234;
+
+       key = last_key + 1;
+       assert(!bpf_map_update(map_fd, &key, value, BPF_NOEXIST));
+       assert(!bpf_map_lookup(map_fd, &key, value));
+
+       /* Cannot find the last key because it was removed by LRU */
+       assert(bpf_map_lookup(map_fd, &last_key, value));
+}
+
+/* Test map with only one element */
+static void test_lru_sanity5(int map_type, int map_flags)
+{
+       unsigned long long key, value[nr_cpus];
+       int next_sched_cpu = 0;
+       int map_fd;
+       int i;
+
+       if (map_flags & BPF_F_NO_COMMON_LRU)
+               return;
+
+       printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
+              map_flags);
+
+       map_fd = create_map(map_type, map_flags, 1);
+       assert(map_fd != -1);
+
+       value[0] = 1234;
+       key = 0;
+       assert(!bpf_map_update(map_fd, &key, value, BPF_NOEXIST));
+
+       for (i = 0; i < nr_cpus; i++) {
+               pid_t pid;
+
+               pid = fork();
+               if (pid == 0) {
+                       next_sched_cpu = sched_next_online(0, next_sched_cpu);
+                       if (next_sched_cpu != -1)
+                               do_test_lru_sanity5(key, map_fd);
+                       exit(0);
+               } else if (pid == -1) {
+                       printf("couldn't spawn #%d process\n", i);
+                       exit(1);
+               } else {
+                       int status;
+
+                       /* It is mostly redundant and just allow the parent
+                        * process to update next_shced_cpu for the next child
+                        * process
+                        */
+                       next_sched_cpu = sched_next_online(pid, next_sched_cpu);
+
+                       assert(waitpid(pid, &status, 0) == pid);
+                       assert(status == 0);
+                       key++;
+               }
+       }
+
+       close(map_fd);
+
+       printf("Pass\n");
+}
+
+int main(int argc, char **argv)
+{
+       struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
+       int map_types[] = {BPF_MAP_TYPE_LRU_HASH,
+                            BPF_MAP_TYPE_LRU_PERCPU_HASH};
+       int map_flags[] = {0, BPF_F_NO_COMMON_LRU};
+       int t, f;
+
+       setbuf(stdout, NULL);
+
+       assert(!setrlimit(RLIMIT_MEMLOCK, &r));
+
+       nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
+       assert(nr_cpus != -1);
+       printf("nr_cpus:%d\n\n", nr_cpus);
+
+       for (f = 0; f < sizeof(map_flags) / sizeof(*map_flags); f++) {
+               unsigned int tgt_free = (map_flags[f] & BPF_F_NO_COMMON_LRU) ?
+                       PERCPU_FREE_TARGET : LOCAL_FREE_TARGET;
+
+               for (t = 0; t < sizeof(map_types) / sizeof(*map_types); t++) {
+                       test_lru_sanity0(map_types[t], map_flags[f]);
+                       test_lru_sanity1(map_types[t], map_flags[f], tgt_free);
+                       test_lru_sanity2(map_types[t], map_flags[f], tgt_free);
+                       test_lru_sanity3(map_types[t], map_flags[f], tgt_free);
+                       test_lru_sanity4(map_types[t], map_flags[f], tgt_free);
+                       test_lru_sanity5(map_types[t], map_flags[f]);
+
+                       printf("\n");
+               }
+       }
+
+       return 0;
+}