1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (c) 2017 Nicira, Inc.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/skbuff.h>
11 #include <linux/kernel.h>
12 #include <linux/openvswitch.h>
13 #include <linux/netlink.h>
14 #include <linux/rculist.h>
16 #include <net/netlink.h>
17 #include <net/genetlink.h>
22 static const struct nla_policy meter_policy[OVS_METER_ATTR_MAX + 1] = {
23 [OVS_METER_ATTR_ID] = { .type = NLA_U32, },
24 [OVS_METER_ATTR_KBPS] = { .type = NLA_FLAG },
25 [OVS_METER_ATTR_STATS] = { .len = sizeof(struct ovs_flow_stats) },
26 [OVS_METER_ATTR_BANDS] = { .type = NLA_NESTED },
27 [OVS_METER_ATTR_USED] = { .type = NLA_U64 },
28 [OVS_METER_ATTR_CLEAR] = { .type = NLA_FLAG },
29 [OVS_METER_ATTR_MAX_METERS] = { .type = NLA_U32 },
30 [OVS_METER_ATTR_MAX_BANDS] = { .type = NLA_U32 },
33 static const struct nla_policy band_policy[OVS_BAND_ATTR_MAX + 1] = {
34 [OVS_BAND_ATTR_TYPE] = { .type = NLA_U32, },
35 [OVS_BAND_ATTR_RATE] = { .type = NLA_U32, },
36 [OVS_BAND_ATTR_BURST] = { .type = NLA_U32, },
37 [OVS_BAND_ATTR_STATS] = { .len = sizeof(struct ovs_flow_stats) },
40 static u32 meter_hash(struct dp_meter_instance *ti, u32 id)
42 return id % ti->n_meters;
45 static void ovs_meter_free(struct dp_meter *meter)
50 kfree_rcu(meter, rcu);
53 /* Call with ovs_mutex or RCU read lock. */
54 static struct dp_meter *lookup_meter(const struct dp_meter_table *tbl,
57 struct dp_meter_instance *ti = rcu_dereference_ovsl(tbl->ti);
58 u32 hash = meter_hash(ti, meter_id);
59 struct dp_meter *meter;
61 meter = rcu_dereference_ovsl(ti->dp_meters[hash]);
62 if (meter && likely(meter->id == meter_id))
68 static struct dp_meter_instance *dp_meter_instance_alloc(const u32 size)
70 struct dp_meter_instance *ti;
72 ti = kvzalloc(sizeof(*ti) +
73 sizeof(struct dp_meter *) * size,
83 static void dp_meter_instance_free(struct dp_meter_instance *ti)
88 static void dp_meter_instance_free_rcu(struct rcu_head *rcu)
90 struct dp_meter_instance *ti;
92 ti = container_of(rcu, struct dp_meter_instance, rcu);
97 dp_meter_instance_realloc(struct dp_meter_table *tbl, u32 size)
99 struct dp_meter_instance *ti = rcu_dereference_ovsl(tbl->ti);
100 int n_meters = min(size, ti->n_meters);
101 struct dp_meter_instance *new_ti;
104 new_ti = dp_meter_instance_alloc(size);
108 for (i = 0; i < n_meters; i++)
109 if (rcu_dereference_ovsl(ti->dp_meters[i]))
110 new_ti->dp_meters[i] = ti->dp_meters[i];
112 rcu_assign_pointer(tbl->ti, new_ti);
113 call_rcu(&ti->rcu, dp_meter_instance_free_rcu);
118 static void dp_meter_instance_insert(struct dp_meter_instance *ti,
119 struct dp_meter *meter)
123 hash = meter_hash(ti, meter->id);
124 rcu_assign_pointer(ti->dp_meters[hash], meter);
127 static void dp_meter_instance_remove(struct dp_meter_instance *ti,
128 struct dp_meter *meter)
132 hash = meter_hash(ti, meter->id);
133 RCU_INIT_POINTER(ti->dp_meters[hash], NULL);
136 static int attach_meter(struct dp_meter_table *tbl, struct dp_meter *meter)
138 struct dp_meter_instance *ti = rcu_dereference_ovsl(tbl->ti);
139 u32 hash = meter_hash(ti, meter->id);
142 /* In generally, slots selected should be empty, because
143 * OvS uses id-pool to fetch a available id.
145 if (unlikely(rcu_dereference_ovsl(ti->dp_meters[hash])))
148 dp_meter_instance_insert(ti, meter);
150 /* That function is thread-safe. */
152 if (tbl->count >= tbl->max_meters_allowed) {
157 if (tbl->count >= ti->n_meters &&
158 dp_meter_instance_realloc(tbl, ti->n_meters * 2)) {
166 dp_meter_instance_remove(ti, meter);
171 static int detach_meter(struct dp_meter_table *tbl, struct dp_meter *meter)
173 struct dp_meter_instance *ti;
179 ti = rcu_dereference_ovsl(tbl->ti);
180 dp_meter_instance_remove(ti, meter);
184 /* Shrink the meter array if necessary. */
185 if (ti->n_meters > DP_METER_ARRAY_SIZE_MIN &&
186 tbl->count <= (ti->n_meters / 4)) {
187 int half_size = ti->n_meters / 2;
190 /* Avoid hash collision, don't move slots to other place.
191 * Make sure there are no references of meters in array
192 * which will be released.
194 for (i = half_size; i < ti->n_meters; i++)
195 if (rcu_dereference_ovsl(ti->dp_meters[i]))
198 if (dp_meter_instance_realloc(tbl, half_size))
206 dp_meter_instance_insert(ti, meter);
211 static struct sk_buff *
212 ovs_meter_cmd_reply_start(struct genl_info *info, u8 cmd,
213 struct ovs_header **ovs_reply_header)
216 struct ovs_header *ovs_header = info->userhdr;
218 skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
220 return ERR_PTR(-ENOMEM);
222 *ovs_reply_header = genlmsg_put(skb, info->snd_portid,
224 &dp_meter_genl_family, 0, cmd);
225 if (!*ovs_reply_header) {
227 return ERR_PTR(-EMSGSIZE);
229 (*ovs_reply_header)->dp_ifindex = ovs_header->dp_ifindex;
234 static int ovs_meter_cmd_reply_stats(struct sk_buff *reply, u32 meter_id,
235 struct dp_meter *meter)
238 struct dp_meter_band *band;
241 if (nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id))
244 if (nla_put(reply, OVS_METER_ATTR_STATS,
245 sizeof(struct ovs_flow_stats), &meter->stats))
248 if (nla_put_u64_64bit(reply, OVS_METER_ATTR_USED, meter->used,
252 nla = nla_nest_start_noflag(reply, OVS_METER_ATTR_BANDS);
258 for (i = 0; i < meter->n_bands; ++i, ++band) {
259 struct nlattr *band_nla;
261 band_nla = nla_nest_start_noflag(reply, OVS_BAND_ATTR_UNSPEC);
262 if (!band_nla || nla_put(reply, OVS_BAND_ATTR_STATS,
263 sizeof(struct ovs_flow_stats),
266 nla_nest_end(reply, band_nla);
268 nla_nest_end(reply, nla);
275 static int ovs_meter_cmd_features(struct sk_buff *skb, struct genl_info *info)
277 struct ovs_header *ovs_header = info->userhdr;
278 struct ovs_header *ovs_reply_header;
279 struct nlattr *nla, *band_nla;
280 struct sk_buff *reply;
284 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_FEATURES,
287 return PTR_ERR(reply);
290 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
296 if (nla_put_u32(reply, OVS_METER_ATTR_MAX_METERS,
297 dp->meter_tbl.max_meters_allowed))
302 if (nla_put_u32(reply, OVS_METER_ATTR_MAX_BANDS, DP_MAX_BANDS))
303 goto nla_put_failure;
305 nla = nla_nest_start_noflag(reply, OVS_METER_ATTR_BANDS);
307 goto nla_put_failure;
309 band_nla = nla_nest_start_noflag(reply, OVS_BAND_ATTR_UNSPEC);
311 goto nla_put_failure;
312 /* Currently only DROP band type is supported. */
313 if (nla_put_u32(reply, OVS_BAND_ATTR_TYPE, OVS_METER_BAND_TYPE_DROP))
314 goto nla_put_failure;
315 nla_nest_end(reply, band_nla);
316 nla_nest_end(reply, nla);
318 genlmsg_end(reply, ovs_reply_header);
319 return genlmsg_reply(reply, info);
328 static struct dp_meter *dp_meter_create(struct nlattr **a)
333 struct dp_meter *meter;
334 struct dp_meter_band *band;
337 /* Validate attributes, count the bands. */
338 if (!a[OVS_METER_ATTR_BANDS])
339 return ERR_PTR(-EINVAL);
341 nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem)
342 if (++n_bands > DP_MAX_BANDS)
343 return ERR_PTR(-EINVAL);
345 /* Allocate and set up the meter before locking anything. */
346 meter = kzalloc(struct_size(meter, bands, n_bands), GFP_KERNEL);
348 return ERR_PTR(-ENOMEM);
350 meter->id = nla_get_u32(a[OVS_METER_ATTR_ID]);
351 meter->used = div_u64(ktime_get_ns(), 1000 * 1000);
352 meter->kbps = a[OVS_METER_ATTR_KBPS] ? 1 : 0;
353 meter->keep_stats = !a[OVS_METER_ATTR_CLEAR];
354 spin_lock_init(&meter->lock);
355 if (meter->keep_stats && a[OVS_METER_ATTR_STATS]) {
356 meter->stats = *(struct ovs_flow_stats *)
357 nla_data(a[OVS_METER_ATTR_STATS]);
359 meter->n_bands = n_bands;
361 /* Set up meter bands. */
363 nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem) {
364 struct nlattr *attr[OVS_BAND_ATTR_MAX + 1];
365 u32 band_max_delta_t;
367 err = nla_parse_deprecated((struct nlattr **)&attr,
368 OVS_BAND_ATTR_MAX, nla_data(nla),
369 nla_len(nla), band_policy, NULL);
371 goto exit_free_meter;
373 if (!attr[OVS_BAND_ATTR_TYPE] ||
374 !attr[OVS_BAND_ATTR_RATE] ||
375 !attr[OVS_BAND_ATTR_BURST]) {
377 goto exit_free_meter;
380 band->type = nla_get_u32(attr[OVS_BAND_ATTR_TYPE]);
381 band->rate = nla_get_u32(attr[OVS_BAND_ATTR_RATE]);
382 if (band->rate == 0) {
384 goto exit_free_meter;
387 band->burst_size = nla_get_u32(attr[OVS_BAND_ATTR_BURST]);
388 /* Figure out max delta_t that is enough to fill any bucket.
389 * Keep max_delta_t size to the bucket units:
390 * pkts => 1/1000 packets, kilobits => bits.
392 * Start with a full bucket.
394 band->bucket = band->burst_size * 1000ULL;
395 band_max_delta_t = div_u64(band->bucket, band->rate);
396 if (band_max_delta_t > meter->max_delta_t)
397 meter->max_delta_t = band_max_delta_t;
408 static int ovs_meter_cmd_set(struct sk_buff *skb, struct genl_info *info)
410 struct nlattr **a = info->attrs;
411 struct dp_meter *meter, *old_meter;
412 struct sk_buff *reply;
413 struct ovs_header *ovs_reply_header;
414 struct ovs_header *ovs_header = info->userhdr;
415 struct dp_meter_table *meter_tbl;
421 if (!a[OVS_METER_ATTR_ID])
424 meter = dp_meter_create(a);
426 return PTR_ERR(meter);
428 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_SET,
431 err = PTR_ERR(reply);
432 goto exit_free_meter;
436 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
442 meter_tbl = &dp->meter_tbl;
443 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
445 old_meter = lookup_meter(meter_tbl, meter_id);
446 err = detach_meter(meter_tbl, old_meter);
450 err = attach_meter(meter_tbl, meter);
456 /* Build response with the meter_id and stats from
457 * the old meter, if any.
459 failed = nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id);
462 spin_lock_bh(&old_meter->lock);
463 if (old_meter->keep_stats) {
464 err = ovs_meter_cmd_reply_stats(reply, meter_id,
468 spin_unlock_bh(&old_meter->lock);
469 ovs_meter_free(old_meter);
472 genlmsg_end(reply, ovs_reply_header);
473 return genlmsg_reply(reply, info);
483 static int ovs_meter_cmd_get(struct sk_buff *skb, struct genl_info *info)
485 struct ovs_header *ovs_header = info->userhdr;
486 struct ovs_header *ovs_reply_header;
487 struct nlattr **a = info->attrs;
488 struct dp_meter *meter;
489 struct sk_buff *reply;
494 if (!a[OVS_METER_ATTR_ID])
497 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
499 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_GET,
502 return PTR_ERR(reply);
506 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
512 /* Locate meter, copy stats. */
513 meter = lookup_meter(&dp->meter_tbl, meter_id);
519 spin_lock_bh(&meter->lock);
520 err = ovs_meter_cmd_reply_stats(reply, meter_id, meter);
521 spin_unlock_bh(&meter->lock);
527 genlmsg_end(reply, ovs_reply_header);
528 return genlmsg_reply(reply, info);
536 static int ovs_meter_cmd_del(struct sk_buff *skb, struct genl_info *info)
538 struct ovs_header *ovs_header = info->userhdr;
539 struct ovs_header *ovs_reply_header;
540 struct nlattr **a = info->attrs;
541 struct dp_meter *old_meter;
542 struct sk_buff *reply;
547 if (!a[OVS_METER_ATTR_ID])
550 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_DEL,
553 return PTR_ERR(reply);
557 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
563 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
564 old_meter = lookup_meter(&dp->meter_tbl, meter_id);
566 spin_lock_bh(&old_meter->lock);
567 err = ovs_meter_cmd_reply_stats(reply, meter_id, old_meter);
569 spin_unlock_bh(&old_meter->lock);
571 err = detach_meter(&dp->meter_tbl, old_meter);
577 ovs_meter_free(old_meter);
578 genlmsg_end(reply, ovs_reply_header);
579 return genlmsg_reply(reply, info);
587 /* Meter action execution.
589 * Return true 'meter_id' drop band is triggered. The 'skb' should be
590 * dropped by the caller'.
592 bool ovs_meter_execute(struct datapath *dp, struct sk_buff *skb,
593 struct sw_flow_key *key, u32 meter_id)
595 long long int now_ms = div_u64(ktime_get_ns(), 1000 * 1000);
596 long long int long_delta_ms;
597 struct dp_meter_band *band;
598 struct dp_meter *meter;
599 int i, band_exceeded_max = -1;
600 u32 band_exceeded_rate = 0;
604 meter = lookup_meter(&dp->meter_tbl, meter_id);
605 /* Do not drop the packet when there is no meter. */
609 /* Lock the meter while using it. */
610 spin_lock(&meter->lock);
612 long_delta_ms = (now_ms - meter->used); /* ms */
613 if (long_delta_ms < 0) {
614 /* This condition means that we have several threads fighting
615 * for a meter lock, and the one who received the packets a
616 * bit later wins. Assuming that all racing threads received
617 * packets at the same time to avoid overflow.
622 /* Make sure delta_ms will not be too large, so that bucket will not
625 delta_ms = (long_delta_ms > (long long int)meter->max_delta_t)
626 ? meter->max_delta_t : (u32)long_delta_ms;
628 /* Update meter statistics.
630 meter->used = now_ms;
631 meter->stats.n_packets += 1;
632 meter->stats.n_bytes += skb->len;
634 /* Bucket rate is either in kilobits per second, or in packets per
635 * second. We maintain the bucket in the units of either bits or
636 * 1/1000th of a packet, correspondingly.
637 * Then, when rate is multiplied with milliseconds, we get the
639 * msec * kbps = bits, and
640 * msec * packets/sec = 1/1000 packets.
642 * 'cost' is the number of bucket units in this packet.
644 cost = (meter->kbps) ? skb->len * 8 : 1000;
646 /* Update all bands and find the one hit with the highest rate. */
647 for (i = 0; i < meter->n_bands; ++i) {
648 long long int max_bucket_size;
650 band = &meter->bands[i];
651 max_bucket_size = band->burst_size * 1000LL;
653 band->bucket += delta_ms * band->rate;
654 if (band->bucket > max_bucket_size)
655 band->bucket = max_bucket_size;
657 if (band->bucket >= cost) {
658 band->bucket -= cost;
659 } else if (band->rate > band_exceeded_rate) {
660 band_exceeded_rate = band->rate;
661 band_exceeded_max = i;
665 if (band_exceeded_max >= 0) {
666 /* Update band statistics. */
667 band = &meter->bands[band_exceeded_max];
668 band->stats.n_packets += 1;
669 band->stats.n_bytes += skb->len;
671 /* Drop band triggered, let the caller drop the 'skb'. */
672 if (band->type == OVS_METER_BAND_TYPE_DROP) {
673 spin_unlock(&meter->lock);
678 spin_unlock(&meter->lock);
682 static const struct genl_small_ops dp_meter_genl_ops[] = {
683 { .cmd = OVS_METER_CMD_FEATURES,
684 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
685 .flags = 0, /* OK for unprivileged users. */
686 .doit = ovs_meter_cmd_features
688 { .cmd = OVS_METER_CMD_SET,
689 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
690 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN
693 .doit = ovs_meter_cmd_set,
695 { .cmd = OVS_METER_CMD_GET,
696 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
697 .flags = 0, /* OK for unprivileged users. */
698 .doit = ovs_meter_cmd_get,
700 { .cmd = OVS_METER_CMD_DEL,
701 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
702 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN
705 .doit = ovs_meter_cmd_del
709 static const struct genl_multicast_group ovs_meter_multicast_group = {
710 .name = OVS_METER_MCGROUP,
713 struct genl_family dp_meter_genl_family __ro_after_init = {
714 .hdrsize = sizeof(struct ovs_header),
715 .name = OVS_METER_FAMILY,
716 .version = OVS_METER_VERSION,
717 .maxattr = OVS_METER_ATTR_MAX,
718 .policy = meter_policy,
720 .parallel_ops = true,
721 .small_ops = dp_meter_genl_ops,
722 .n_small_ops = ARRAY_SIZE(dp_meter_genl_ops),
723 .mcgrps = &ovs_meter_multicast_group,
725 .module = THIS_MODULE,
728 int ovs_meters_init(struct datapath *dp)
730 struct dp_meter_table *tbl = &dp->meter_tbl;
731 struct dp_meter_instance *ti;
732 unsigned long free_mem_bytes;
734 ti = dp_meter_instance_alloc(DP_METER_ARRAY_SIZE_MIN);
738 /* Allow meters in a datapath to use ~3.12% of physical memory. */
739 free_mem_bytes = nr_free_buffer_pages() * (PAGE_SIZE >> 5);
740 tbl->max_meters_allowed = min(free_mem_bytes / sizeof(struct dp_meter),
742 if (!tbl->max_meters_allowed)
745 rcu_assign_pointer(tbl->ti, ti);
751 dp_meter_instance_free(ti);
755 void ovs_meters_exit(struct datapath *dp)
757 struct dp_meter_table *tbl = &dp->meter_tbl;
758 struct dp_meter_instance *ti = rcu_dereference_raw(tbl->ti);
761 for (i = 0; i < ti->n_meters; i++)
762 ovs_meter_free(rcu_dereference_raw(ti->dp_meters[i]));
764 dp_meter_instance_free(ti);