OSDN Git Service

samples: bpf: Add devmap_xmit tracepoint statistics support
authorKumar Kartikeya Dwivedi <memxor@gmail.com>
Sat, 21 Aug 2021 00:19:59 +0000 (05:49 +0530)
committerAlexei Starovoitov <ast@kernel.org>
Tue, 24 Aug 2021 21:48:41 +0000 (14:48 -0700)
This adds support for retrieval and printing for devmap_xmit total and
mutli mode tracepoint. For multi mode, we keep a hash map entry for each
redirection stream, such that we can dynamically add and remove entries
on output.

The from_match and to_match will be set by individual samples when
setting up the XDP program on these devices.

The multi mode tracepoint is also handy for xdp_redirect_map_multi,
where up to 32 devices can be specified.

Also add samples_init_pre_load macro to finally set up the resized maps
and mmap them in place for low overhead stats retrieval.

Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210821002010.845777-12-memxor@gmail.com
samples/bpf/xdp_sample_user.c
samples/bpf/xdp_sample_user.h

index e2692de..eb484c1 100644 (file)
@@ -77,6 +77,8 @@ enum map_type {
        MAP_CPUMAP_ENQUEUE,
        MAP_CPUMAP_KTHREAD,
        MAP_EXCEPTION,
+       MAP_DEVMAP_XMIT,
+       MAP_DEVMAP_XMIT_MULTI,
        NUM_MAP,
 };
 
@@ -103,6 +105,8 @@ struct stats_record {
        struct record redir_err[XDP_REDIRECT_ERR_MAX];
        struct record kthread;
        struct record exception[XDP_ACTION_MAX];
+       struct record devmap_xmit;
+       DECLARE_HASHTABLE(xmit_map, 5);
        struct record enq[];
 };
 
@@ -111,7 +115,9 @@ struct sample_output {
                __u64 rx;
                __u64 redir;
                __u64 drop;
+               __u64 drop_xmit;
                __u64 err;
+               __u64 xmit;
        } totals;
        struct {
                __u64 pps;
@@ -125,6 +131,12 @@ struct sample_output {
        struct {
                __u64 hits;
        } except_cnt;
+       struct {
+               __u64 pps;
+               __u64 drop;
+               __u64 err;
+               double bavg;
+       } xmit_cnt;
 };
 
 struct xdp_desc {
@@ -265,6 +277,16 @@ static void sample_print_help(int mask)
                       "  \t\t\t\thit/s     - Number of times the tracepoint was hit per second\n\n");
        }
 
+       if (mask & SAMPLE_DEVMAP_XMIT_CNT) {
+               printf("  devmap_xmit\t\tDisplays devmap_xmit tracepoint events\n"
+                      "  \t\t\tThis tracepoint is invoked for successful transmissions on output\n"
+                      "  \t\t\tdevice but these statistics are not available for generic XDP mode,\n"
+                      "  \t\t\thence they will be omitted from the output when using SKB mode\n"
+                      "  \t\t\t\txmit/s    - Number of packets that were transmitted per second\n"
+                      "  \t\t\t\tdrop/s    - Number of packets that failed transmissions per second\n"
+                      "  \t\t\t\tdrv_err/s - Number of internal driver errors per second\n"
+                      "  \t\t\t\tbulk-avg  - Average number of packets processed for each event\n\n");
+       }
 }
 
 void sample_usage(char *argv[], const struct option *long_options,
@@ -353,6 +375,74 @@ static void map_collect_percpu(struct datarec *values, struct record *rec)
        rec->total.xdp_redirect = sum_xdp_redirect;
 }
 
+static int map_collect_percpu_devmap(int map_fd, struct stats_record *rec)
+{
+       unsigned int nr_cpus = bpf_num_possible_cpus();
+       __u32 batch, count = 32;
+       struct datarec *values;
+       bool init = false;
+       __u64 *keys;
+       int i, ret;
+
+       keys = calloc(count, sizeof(__u64));
+       if (!keys)
+               return -ENOMEM;
+       values = calloc(count * nr_cpus, sizeof(struct datarec));
+       if (!values) {
+               free(keys);
+               return -ENOMEM;
+       }
+
+       for (;;) {
+               bool exit = false;
+
+               ret = bpf_map_lookup_batch(map_fd, init ? &batch : NULL, &batch,
+                                          keys, values, &count, NULL);
+               if (ret < 0 && errno != ENOENT)
+                       break;
+               if (errno == ENOENT)
+                       exit = true;
+
+               init = true;
+               for (i = 0; i < count; i++) {
+                       struct map_entry *e, *x = NULL;
+                       __u64 pair = keys[i];
+                       struct datarec *arr;
+
+                       arr = &values[i * nr_cpus];
+                       hash_for_each_possible(rec->xmit_map, e, node, pair) {
+                               if (e->pair == pair) {
+                                       x = e;
+                                       break;
+                               }
+                       }
+                       if (!x) {
+                               x = calloc(1, sizeof(*x));
+                               if (!x)
+                                       goto cleanup;
+                               if (map_entry_init(x, pair) < 0) {
+                                       free(x);
+                                       goto cleanup;
+                               }
+                               hash_add(rec->xmit_map, &x->node, pair);
+                       }
+                       map_collect_percpu(arr, &x->val);
+               }
+
+               if (exit)
+                       break;
+               count = 32;
+       }
+
+       free(values);
+       free(keys);
+       return 0;
+cleanup:
+       free(values);
+       free(keys);
+       return -ENOMEM;
+}
+
 static struct stats_record *alloc_stats_record(void)
 {
        struct stats_record *rec;
@@ -408,6 +498,16 @@ static struct stats_record *alloc_stats_record(void)
                        }
                }
        }
+       if (sample_mask & SAMPLE_DEVMAP_XMIT_CNT) {
+               rec->devmap_xmit.cpu = alloc_record_per_cpu();
+               if (!rec->devmap_xmit.cpu) {
+                       fprintf(stderr,
+                               "Failed to allocate devmap_xmit per-CPU array\n");
+                       goto end_exception;
+               }
+       }
+       if (sample_mask & SAMPLE_DEVMAP_XMIT_CNT_MULTI)
+               hash_init(rec->xmit_map);
        if (sample_mask & SAMPLE_CPUMAP_ENQUEUE_CNT) {
                for (i = 0; i < sample_n_cpus; i++) {
                        rec->enq[i].cpu = alloc_record_per_cpu();
@@ -418,13 +518,15 @@ static struct stats_record *alloc_stats_record(void)
                                        i);
                                while (i--)
                                        free(rec->enq[i].cpu);
-                               goto end_exception;
+                               goto end_devmap_xmit;
                        }
                }
        }
 
        return rec;
 
+end_devmap_xmit:
+       free(rec->devmap_xmit.cpu);
 end_exception:
        for (i = 0; i < XDP_ACTION_MAX; i++)
                free(rec->exception[i].cpu);
@@ -448,6 +550,12 @@ static void free_stats_record(struct stats_record *r)
 
        for (i = 0; i < sample_n_cpus; i++)
                free(r->enq[i].cpu);
+       hash_for_each_safe(r->xmit_map, i, tmp, e, node) {
+               hash_del(&e->node);
+               free(e->val.cpu);
+               free(e);
+       }
+       free(r->devmap_xmit.cpu);
        for (i = 0; i < XDP_ACTION_MAX; i++)
                free(r->exception[i].cpu);
        free(r->kthread.cpu);
@@ -835,6 +943,160 @@ static void stats_get_exception_cnt(struct stats_record *stats_rec,
        }
 }
 
+static void stats_get_devmap_xmit(struct stats_record *stats_rec,
+                                 struct stats_record *stats_prev,
+                                 unsigned int nr_cpus,
+                                 struct sample_output *out)
+{
+       double pps, drop, info, err;
+       struct record *rec, *prev;
+       double t;
+       int i;
+
+       rec = &stats_rec->devmap_xmit;
+       prev = &stats_prev->devmap_xmit;
+       t = calc_period(rec, prev);
+       for (i = 0; i < nr_cpus; i++) {
+               struct datarec *r = &rec->cpu[i];
+               struct datarec *p = &prev->cpu[i];
+               char str[64];
+
+               pps = calc_pps(r, p, t);
+               drop = calc_drop_pps(r, p, t);
+               err = calc_errs_pps(r, p, t);
+
+               if (!pps && !drop && !err)
+                       continue;
+
+               snprintf(str, sizeof(str), "cpu:%d", i);
+               info = calc_info_pps(r, p, t);
+               if (info > 0)
+                       info = (pps + drop) / info; /* calc avg bulk */
+               print_default("     %-18s" FMT_COLUMNf FMT_COLUMNf FMT_COLUMNf
+                                     __COLUMN(".2f") "\n",
+                             str, XMIT(pps), DROP(drop), err, "drv_err/s",
+                             info, "bulk-avg");
+       }
+       if (out) {
+               pps = calc_pps(&rec->total, &prev->total, t);
+               drop = calc_drop_pps(&rec->total, &prev->total, t);
+               info = calc_info_pps(&rec->total, &prev->total, t);
+               if (info > 0)
+                       info = (pps + drop) / info; /* calc avg bulk */
+               err = calc_errs_pps(&rec->total, &prev->total, t);
+
+               out->xmit_cnt.pps = pps;
+               out->xmit_cnt.drop = drop;
+               out->xmit_cnt.bavg = info;
+               out->xmit_cnt.err = err;
+               out->totals.xmit += pps;
+               out->totals.drop_xmit += drop;
+               out->totals.err += err;
+       }
+}
+
+static void stats_get_devmap_xmit_multi(struct stats_record *stats_rec,
+                                       struct stats_record *stats_prev,
+                                       unsigned int nr_cpus,
+                                       struct sample_output *out,
+                                       bool xmit_total)
+{
+       double pps, drop, info, err;
+       struct map_entry *entry;
+       struct record *r, *p;
+       double t;
+       int bkt;
+
+       hash_for_each(stats_rec->xmit_map, bkt, entry, node) {
+               struct map_entry *e, *x = NULL;
+               char ifname_from[IFNAMSIZ];
+               char ifname_to[IFNAMSIZ];
+               const char *fstr, *tstr;
+               unsigned long prev_time;
+               struct record beg = {};
+               __u32 from_idx, to_idx;
+               char str[128];
+               __u64 pair;
+               int i;
+
+               prev_time = sample_interval * NANOSEC_PER_SEC;
+
+               pair = entry->pair;
+               from_idx = pair >> 32;
+               to_idx = pair & 0xFFFFFFFF;
+
+               r = &entry->val;
+               beg.timestamp = r->timestamp - prev_time;
+
+               /* Find matching entry from stats_prev map */
+               hash_for_each_possible(stats_prev->xmit_map, e, node, pair) {
+                       if (e->pair == pair) {
+                               x = e;
+                               break;
+                       }
+               }
+               if (x)
+                       p = &x->val;
+               else
+                       p = &beg;
+               t = calc_period(r, p);
+               pps = calc_pps(&r->total, &p->total, t);
+               drop = calc_drop_pps(&r->total, &p->total, t);
+               info = calc_info_pps(&r->total, &p->total, t);
+               if (info > 0)
+                       info = (pps + drop) / info; /* calc avg bulk */
+               err = calc_errs_pps(&r->total, &p->total, t);
+
+               if (out) {
+                       /* We are responsible for filling out totals */
+                       out->totals.xmit += pps;
+                       out->totals.drop_xmit += drop;
+                       out->totals.err += err;
+                       continue;
+               }
+
+               fstr = tstr = NULL;
+               if (if_indextoname(from_idx, ifname_from))
+                       fstr = ifname_from;
+               if (if_indextoname(to_idx, ifname_to))
+                       tstr = ifname_to;
+
+               snprintf(str, sizeof(str), "xmit %s->%s", fstr ?: "?",
+                        tstr ?: "?");
+               /* Skip idle streams of redirection */
+               if (pps || drop || err) {
+                       print_err(drop,
+                                 "  %-20s " FMT_COLUMNf FMT_COLUMNf FMT_COLUMNf
+                                 __COLUMN(".2f") "\n", str, XMIT(pps), DROP(drop),
+                                 err, "drv_err/s", info, "bulk-avg");
+               }
+
+               for (i = 0; i < nr_cpus; i++) {
+                       struct datarec *rc = &r->cpu[i];
+                       struct datarec *pc, p_beg = {};
+                       char str[64];
+
+                       pc = p == &beg ? &p_beg : &p->cpu[i];
+
+                       pps = calc_pps(rc, pc, t);
+                       drop = calc_drop_pps(rc, pc, t);
+                       err = calc_errs_pps(rc, pc, t);
+
+                       if (!pps && !drop && !err)
+                               continue;
+
+                       snprintf(str, sizeof(str), "cpu:%d", i);
+                       info = calc_info_pps(rc, pc, t);
+                       if (info > 0)
+                               info = (pps + drop) / info; /* calc avg bulk */
+
+                       print_default("     %-18s" FMT_COLUMNf FMT_COLUMNf FMT_COLUMNf
+                                     __COLUMN(".2f") "\n", str, XMIT(pps),
+                                     DROP(drop), err, "drv_err/s", info, "bulk-avg");
+               }
+       }
+}
+
 static void stats_print(const char *prefix, int mask, struct stats_record *r,
                        struct stats_record *p, struct sample_output *out)
 {
@@ -849,6 +1111,9 @@ static void stats_print(const char *prefix, int mask, struct stats_record *r,
        printf(FMT_COLUMNl,
               out->totals.err + out->totals.drop + out->totals.drop_xmit,
               "err,drop/s");
+       if (mask & SAMPLE_DEVMAP_XMIT_CNT ||
+           mask & SAMPLE_DEVMAP_XMIT_CNT_MULTI)
+               printf(FMT_COLUMNl, XMIT(out->totals.xmit));
        printf("\n");
 
        if (mask & SAMPLE_RX_CNT) {
@@ -899,6 +1164,25 @@ static void stats_print(const char *prefix, int mask, struct stats_record *r,
                stats_get_exception_cnt(r, p, nr_cpus, NULL);
        }
 
+       if (mask & SAMPLE_DEVMAP_XMIT_CNT) {
+               str = (sample_log_level & LL_DEFAULT) && out->xmit_cnt.pps ?
+                                   "devmap_xmit total" :
+                                   "devmap_xmit";
+
+               print_err(out->xmit_cnt.err || out->xmit_cnt.drop,
+                         "  %-20s " FMT_COLUMNl FMT_COLUMNl FMT_COLUMNl
+                                 __COLUMN(".2f") "\n",
+                         str, XMIT(out->xmit_cnt.pps),
+                         DROP(out->xmit_cnt.drop), out->xmit_cnt.err,
+                         "drv_err/s", out->xmit_cnt.bavg, "bulk-avg");
+
+               stats_get_devmap_xmit(r, p, nr_cpus, NULL);
+       }
+
+       if (mask & SAMPLE_DEVMAP_XMIT_CNT_MULTI)
+               stats_get_devmap_xmit_multi(r, p, nr_cpus, NULL,
+                                           mask & SAMPLE_DEVMAP_XMIT_CNT);
+
        if (sample_log_level & LL_DEFAULT ||
            ((sample_log_level & LL_SIMPLE) && sample_err_exp)) {
                sample_err_exp = false;
@@ -910,12 +1194,13 @@ int sample_setup_maps(struct bpf_map **maps)
 {
        sample_n_cpus = libbpf_num_possible_cpus();
 
-       for (int i = 0; i < NUM_MAP; i++) {
+       for (int i = 0; i < MAP_DEVMAP_XMIT_MULTI; i++) {
                sample_map[i] = maps[i];
 
                switch (i) {
                case MAP_RX:
                case MAP_CPUMAP_KTHREAD:
+               case MAP_DEVMAP_XMIT:
                        sample_map_count[i] = sample_n_cpus;
                        break;
                case MAP_REDIRECT_ERR:
@@ -933,12 +1218,13 @@ int sample_setup_maps(struct bpf_map **maps)
                if (bpf_map__resize(sample_map[i], sample_map_count[i]) < 0)
                        return -errno;
        }
+       sample_map[MAP_DEVMAP_XMIT_MULTI] = maps[MAP_DEVMAP_XMIT_MULTI];
        return 0;
 }
 
 static int sample_setup_maps_mappings(void)
 {
-       for (int i = 0; i < NUM_MAP; i++) {
+       for (int i = 0; i < MAP_DEVMAP_XMIT_MULTI; i++) {
                size_t size = sample_map_count[i] * sizeof(struct datarec);
 
                sample_mmap[i] = mmap(NULL, size, PROT_READ | PROT_WRITE,
@@ -1057,9 +1343,20 @@ static void sample_summary_print(void)
        if (sample_out.totals.drop)
                print_always("  Rx dropped          : %'-10llu\n",
                             sample_out.totals.drop);
+       if (sample_out.totals.drop_xmit)
+               print_always("  Tx dropped          : %'-10llu\n",
+                            sample_out.totals.drop_xmit);
        if (sample_out.totals.err)
                print_always("  Errors recorded     : %'-10llu\n",
                             sample_out.totals.err);
+       if (sample_out.totals.xmit) {
+               double pkts = sample_out.totals.xmit;
+
+               print_always("  Packets transmitted : %'-10llu\n",
+                            sample_out.totals.xmit);
+               print_always("  Average transmit/s  : %'-10.0f\n",
+                            sample_round(pkts / period));
+       }
 }
 
 void sample_exit(int status)
@@ -1115,6 +1412,13 @@ static int sample_stats_collect(struct stats_record *rec)
                        map_collect_percpu(&sample_mmap[MAP_EXCEPTION][i * sample_n_cpus],
                                           &rec->exception[i]);
 
+       if (sample_mask & SAMPLE_DEVMAP_XMIT_CNT)
+               map_collect_percpu(sample_mmap[MAP_DEVMAP_XMIT], &rec->devmap_xmit);
+
+       if (sample_mask & SAMPLE_DEVMAP_XMIT_CNT_MULTI) {
+               if (map_collect_percpu_devmap(bpf_map__fd(sample_map[MAP_DEVMAP_XMIT_MULTI]), rec) < 0)
+                       return -EINVAL;
+       }
        return 0;
 }
 
@@ -1123,7 +1427,9 @@ static void sample_summary_update(struct sample_output *out, int interval)
        sample_out.totals.rx += out->totals.rx;
        sample_out.totals.redir += out->totals.redir;
        sample_out.totals.drop += out->totals.drop;
+       sample_out.totals.drop_xmit += out->totals.drop_xmit;
        sample_out.totals.err += out->totals.err;
+       sample_out.totals.xmit += out->totals.xmit;
        sample_out.rx_cnt.pps += interval;
 }
 
@@ -1141,6 +1447,11 @@ static void sample_stats_print(int mask, struct stats_record *cur,
                stats_get_redirect_err_cnt(cur, prev, 0, &out);
        if (mask & SAMPLE_EXCEPTION_CNT)
                stats_get_exception_cnt(cur, prev, 0, &out);
+       if (mask & SAMPLE_DEVMAP_XMIT_CNT)
+               stats_get_devmap_xmit(cur, prev, 0, &out);
+       else if (mask & SAMPLE_DEVMAP_XMIT_CNT_MULTI)
+               stats_get_devmap_xmit_multi(cur, prev, 0, &out,
+                                           mask & SAMPLE_DEVMAP_XMIT_CNT);
        sample_summary_update(&out, interval);
 
        stats_print(prog_name, mask, cur, prev, &out);
index 2037326..3a67898 100644 (file)
@@ -14,9 +14,11 @@ enum stats_mask {
        SAMPLE_CPUMAP_ENQUEUE_CNT   = 1U << 3,
        SAMPLE_CPUMAP_KTHREAD_CNT   = 1U << 4,
        SAMPLE_EXCEPTION_CNT        = 1U << 5,
+       SAMPLE_DEVMAP_XMIT_CNT      = 1U << 6,
        SAMPLE_REDIRECT_CNT         = 1U << 7,
        SAMPLE_REDIRECT_MAP_CNT     = SAMPLE_REDIRECT_CNT | _SAMPLE_REDIRECT_MAP,
        SAMPLE_REDIRECT_ERR_MAP_CNT = SAMPLE_REDIRECT_ERR_CNT | _SAMPLE_REDIRECT_MAP,
+       SAMPLE_DEVMAP_XMIT_CNT_MULTI = 1U << 8,
 };
 
 /* Exit return codes */
@@ -63,6 +65,17 @@ static inline char *safe_strncpy(char *dst, const char *src, size_t size)
                        return -errno;                                         \
        })
 
+#define sample_init_pre_load(skel)                                             \
+       ({                                                                     \
+               skel->rodata->nr_cpus = libbpf_num_possible_cpus();            \
+               sample_setup_maps((struct bpf_map *[]){                        \
+                       skel->maps.rx_cnt, skel->maps.redir_err_cnt,           \
+                       skel->maps.cpumap_enqueue_cnt,                         \
+                       skel->maps.cpumap_kthread_cnt,                         \
+                       skel->maps.exception_cnt, skel->maps.devmap_xmit_cnt,  \
+                       skel->maps.devmap_xmit_cnt_multi });                   \
+       })
+
 #define DEFINE_SAMPLE_INIT(name)                                               \
        static int sample_init(struct name *skel, int mask)                    \
        {                                                                      \
@@ -84,6 +97,10 @@ static inline char *safe_strncpy(char *dst, const char *src, size_t size)
                        __attach_tp(tp_xdp_cpumap_kthread);                    \
                if (mask & SAMPLE_EXCEPTION_CNT)                               \
                        __attach_tp(tp_xdp_exception);                         \
+               if (mask & SAMPLE_DEVMAP_XMIT_CNT)                             \
+                       __attach_tp(tp_xdp_devmap_xmit);                       \
+               if (mask & SAMPLE_DEVMAP_XMIT_CNT_MULTI)                       \
+                       __attach_tp(tp_xdp_devmap_xmit_multi);                 \
                return 0;                                                      \
        }