Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (c) 2017 Nicira, Inc.
4 */
5
6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8#include <linux/if.h>
9#include <linux/skbuff.h>
10#include <linux/ip.h>
11#include <linux/kernel.h>
12#include <linux/openvswitch.h>
13#include <linux/netlink.h>
14#include <linux/rculist.h>
15
16#include <net/netlink.h>
17#include <net/genetlink.h>
18
19#include "datapath.h"
20#include "meter.h"
21
22#define METER_HASH_BUCKETS 1024
23
24static const struct nla_policy meter_policy[OVS_METER_ATTR_MAX + 1] = {
25 [OVS_METER_ATTR_ID] = { .type = NLA_U32, },
26 [OVS_METER_ATTR_KBPS] = { .type = NLA_FLAG },
27 [OVS_METER_ATTR_STATS] = { .len = sizeof(struct ovs_flow_stats) },
28 [OVS_METER_ATTR_BANDS] = { .type = NLA_NESTED },
29 [OVS_METER_ATTR_USED] = { .type = NLA_U64 },
30 [OVS_METER_ATTR_CLEAR] = { .type = NLA_FLAG },
31 [OVS_METER_ATTR_MAX_METERS] = { .type = NLA_U32 },
32 [OVS_METER_ATTR_MAX_BANDS] = { .type = NLA_U32 },
33};
34
35static const struct nla_policy band_policy[OVS_BAND_ATTR_MAX + 1] = {
36 [OVS_BAND_ATTR_TYPE] = { .type = NLA_U32, },
37 [OVS_BAND_ATTR_RATE] = { .type = NLA_U32, },
38 [OVS_BAND_ATTR_BURST] = { .type = NLA_U32, },
39 [OVS_BAND_ATTR_STATS] = { .len = sizeof(struct ovs_flow_stats) },
40};
41
42static void ovs_meter_free(struct dp_meter *meter)
43{
44 if (!meter)
45 return;
46
47 kfree_rcu(meter, rcu);
48}
49
50static struct hlist_head *meter_hash_bucket(const struct datapath *dp,
51 u32 meter_id)
52{
53 return &dp->meters[meter_id & (METER_HASH_BUCKETS - 1)];
54}
55
56/* Call with ovs_mutex or RCU read lock. */
57static struct dp_meter *lookup_meter(const struct datapath *dp,
58 u32 meter_id)
59{
60 struct dp_meter *meter;
61 struct hlist_head *head;
62
63 head = meter_hash_bucket(dp, meter_id);
64 hlist_for_each_entry_rcu(meter, head, dp_hash_node) {
65 if (meter->id == meter_id)
66 return meter;
67 }
68 return NULL;
69}
70
71static void attach_meter(struct datapath *dp, struct dp_meter *meter)
72{
73 struct hlist_head *head = meter_hash_bucket(dp, meter->id);
74
75 hlist_add_head_rcu(&meter->dp_hash_node, head);
76}
77
78static void detach_meter(struct dp_meter *meter)
79{
80 ASSERT_OVSL();
81 if (meter)
82 hlist_del_rcu(&meter->dp_hash_node);
83}
84
85static struct sk_buff *
86ovs_meter_cmd_reply_start(struct genl_info *info, u8 cmd,
87 struct ovs_header **ovs_reply_header)
88{
89 struct sk_buff *skb;
90 struct ovs_header *ovs_header = info->userhdr;
91
92 skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
93 if (!skb)
94 return ERR_PTR(-ENOMEM);
95
96 *ovs_reply_header = genlmsg_put(skb, info->snd_portid,
97 info->snd_seq,
98 &dp_meter_genl_family, 0, cmd);
99 if (!*ovs_reply_header) {
100 nlmsg_free(skb);
101 return ERR_PTR(-EMSGSIZE);
102 }
103 (*ovs_reply_header)->dp_ifindex = ovs_header->dp_ifindex;
104
105 return skb;
106}
107
108static int ovs_meter_cmd_reply_stats(struct sk_buff *reply, u32 meter_id,
109 struct dp_meter *meter)
110{
111 struct nlattr *nla;
112 struct dp_meter_band *band;
113 u16 i;
114
115 if (nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id))
116 goto error;
117
118 if (!meter)
119 return 0;
120
121 if (nla_put(reply, OVS_METER_ATTR_STATS,
122 sizeof(struct ovs_flow_stats), &meter->stats) ||
123 nla_put_u64_64bit(reply, OVS_METER_ATTR_USED, meter->used,
124 OVS_METER_ATTR_PAD))
125 goto error;
126
127 nla = nla_nest_start_noflag(reply, OVS_METER_ATTR_BANDS);
128 if (!nla)
129 goto error;
130
131 band = meter->bands;
132
133 for (i = 0; i < meter->n_bands; ++i, ++band) {
134 struct nlattr *band_nla;
135
136 band_nla = nla_nest_start_noflag(reply, OVS_BAND_ATTR_UNSPEC);
137 if (!band_nla || nla_put(reply, OVS_BAND_ATTR_STATS,
138 sizeof(struct ovs_flow_stats),
139 &band->stats))
140 goto error;
141 nla_nest_end(reply, band_nla);
142 }
143 nla_nest_end(reply, nla);
144
145 return 0;
146error:
147 return -EMSGSIZE;
148}
149
150static int ovs_meter_cmd_features(struct sk_buff *skb, struct genl_info *info)
151{
152 struct sk_buff *reply;
153 struct ovs_header *ovs_reply_header;
154 struct nlattr *nla, *band_nla;
155 int err;
156
157 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_FEATURES,
158 &ovs_reply_header);
159 if (IS_ERR(reply))
160 return PTR_ERR(reply);
161
162 if (nla_put_u32(reply, OVS_METER_ATTR_MAX_METERS, U32_MAX) ||
163 nla_put_u32(reply, OVS_METER_ATTR_MAX_BANDS, DP_MAX_BANDS))
164 goto nla_put_failure;
165
166 nla = nla_nest_start_noflag(reply, OVS_METER_ATTR_BANDS);
167 if (!nla)
168 goto nla_put_failure;
169
170 band_nla = nla_nest_start_noflag(reply, OVS_BAND_ATTR_UNSPEC);
171 if (!band_nla)
172 goto nla_put_failure;
173 /* Currently only DROP band type is supported. */
174 if (nla_put_u32(reply, OVS_BAND_ATTR_TYPE, OVS_METER_BAND_TYPE_DROP))
175 goto nla_put_failure;
176 nla_nest_end(reply, band_nla);
177 nla_nest_end(reply, nla);
178
179 genlmsg_end(reply, ovs_reply_header);
180 return genlmsg_reply(reply, info);
181
182nla_put_failure:
183 nlmsg_free(reply);
184 err = -EMSGSIZE;
185 return err;
186}
187
188static struct dp_meter *dp_meter_create(struct nlattr **a)
189{
190 struct nlattr *nla;
191 int rem;
192 u16 n_bands = 0;
193 struct dp_meter *meter;
194 struct dp_meter_band *band;
195 int err;
196
197 /* Validate attributes, count the bands. */
198 if (!a[OVS_METER_ATTR_BANDS])
199 return ERR_PTR(-EINVAL);
200
201 nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem)
202 if (++n_bands > DP_MAX_BANDS)
203 return ERR_PTR(-EINVAL);
204
205 /* Allocate and set up the meter before locking anything. */
206 meter = kzalloc(struct_size(meter, bands, n_bands), GFP_KERNEL);
207 if (!meter)
208 return ERR_PTR(-ENOMEM);
209
210 meter->id = nla_get_u32(a[OVS_METER_ATTR_ID]);
211 meter->used = div_u64(ktime_get_ns(), 1000 * 1000);
212 meter->kbps = a[OVS_METER_ATTR_KBPS] ? 1 : 0;
213 meter->keep_stats = !a[OVS_METER_ATTR_CLEAR];
214 spin_lock_init(&meter->lock);
215 if (meter->keep_stats && a[OVS_METER_ATTR_STATS]) {
216 meter->stats = *(struct ovs_flow_stats *)
217 nla_data(a[OVS_METER_ATTR_STATS]);
218 }
219 meter->n_bands = n_bands;
220
221 /* Set up meter bands. */
222 band = meter->bands;
223 nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem) {
224 struct nlattr *attr[OVS_BAND_ATTR_MAX + 1];
225 u32 band_max_delta_t;
226
227 err = nla_parse_deprecated((struct nlattr **)&attr,
228 OVS_BAND_ATTR_MAX, nla_data(nla),
229 nla_len(nla), band_policy, NULL);
230 if (err)
231 goto exit_free_meter;
232
233 if (!attr[OVS_BAND_ATTR_TYPE] ||
234 !attr[OVS_BAND_ATTR_RATE] ||
235 !attr[OVS_BAND_ATTR_BURST]) {
236 err = -EINVAL;
237 goto exit_free_meter;
238 }
239
240 band->type = nla_get_u32(attr[OVS_BAND_ATTR_TYPE]);
241 band->rate = nla_get_u32(attr[OVS_BAND_ATTR_RATE]);
242 if (band->rate == 0) {
243 err = -EINVAL;
244 goto exit_free_meter;
245 }
246
247 band->burst_size = nla_get_u32(attr[OVS_BAND_ATTR_BURST]);
248 /* Figure out max delta_t that is enough to fill any bucket.
249 * Keep max_delta_t size to the bucket units:
250 * pkts => 1/1000 packets, kilobits => bits.
251 *
252 * Start with a full bucket.
253 */
254 band->bucket = (band->burst_size + band->rate) * 1000;
255 band_max_delta_t = band->bucket / band->rate;
256 if (band_max_delta_t > meter->max_delta_t)
257 meter->max_delta_t = band_max_delta_t;
258 band++;
259 }
260
261 return meter;
262
263exit_free_meter:
264 kfree(meter);
265 return ERR_PTR(err);
266}
267
268static int ovs_meter_cmd_set(struct sk_buff *skb, struct genl_info *info)
269{
270 struct nlattr **a = info->attrs;
271 struct dp_meter *meter, *old_meter;
272 struct sk_buff *reply;
273 struct ovs_header *ovs_reply_header;
274 struct ovs_header *ovs_header = info->userhdr;
275 struct datapath *dp;
276 int err;
277 u32 meter_id;
278 bool failed;
279
280 if (!a[OVS_METER_ATTR_ID]) {
281 return -ENODEV;
282 }
283
284 meter = dp_meter_create(a);
285 if (IS_ERR_OR_NULL(meter))
286 return PTR_ERR(meter);
287
288 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_SET,
289 &ovs_reply_header);
290 if (IS_ERR(reply)) {
291 err = PTR_ERR(reply);
292 goto exit_free_meter;
293 }
294
295 ovs_lock();
296 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
297 if (!dp) {
298 err = -ENODEV;
299 goto exit_unlock;
300 }
301
302 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
303
304 /* Cannot fail after this. */
305 old_meter = lookup_meter(dp, meter_id);
306 detach_meter(old_meter);
307 attach_meter(dp, meter);
308 ovs_unlock();
309
310 /* Build response with the meter_id and stats from
311 * the old meter, if any.
312 */
313 failed = nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id);
314 WARN_ON(failed);
315 if (old_meter) {
316 spin_lock_bh(&old_meter->lock);
317 if (old_meter->keep_stats) {
318 err = ovs_meter_cmd_reply_stats(reply, meter_id,
319 old_meter);
320 WARN_ON(err);
321 }
322 spin_unlock_bh(&old_meter->lock);
323 ovs_meter_free(old_meter);
324 }
325
326 genlmsg_end(reply, ovs_reply_header);
327 return genlmsg_reply(reply, info);
328
329exit_unlock:
330 ovs_unlock();
331 nlmsg_free(reply);
332exit_free_meter:
333 kfree(meter);
334 return err;
335}
336
337static int ovs_meter_cmd_get(struct sk_buff *skb, struct genl_info *info)
338{
339 struct nlattr **a = info->attrs;
340 u32 meter_id;
341 struct ovs_header *ovs_header = info->userhdr;
342 struct ovs_header *ovs_reply_header;
343 struct datapath *dp;
344 int err;
345 struct sk_buff *reply;
346 struct dp_meter *meter;
347
348 if (!a[OVS_METER_ATTR_ID])
349 return -EINVAL;
350
351 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
352
353 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_GET,
354 &ovs_reply_header);
355 if (IS_ERR(reply))
356 return PTR_ERR(reply);
357
358 ovs_lock();
359
360 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
361 if (!dp) {
362 err = -ENODEV;
363 goto exit_unlock;
364 }
365
366 /* Locate meter, copy stats. */
367 meter = lookup_meter(dp, meter_id);
368 if (!meter) {
369 err = -ENOENT;
370 goto exit_unlock;
371 }
372
373 spin_lock_bh(&meter->lock);
374 err = ovs_meter_cmd_reply_stats(reply, meter_id, meter);
375 spin_unlock_bh(&meter->lock);
376 if (err)
377 goto exit_unlock;
378
379 ovs_unlock();
380
381 genlmsg_end(reply, ovs_reply_header);
382 return genlmsg_reply(reply, info);
383
384exit_unlock:
385 ovs_unlock();
386 nlmsg_free(reply);
387 return err;
388}
389
390static int ovs_meter_cmd_del(struct sk_buff *skb, struct genl_info *info)
391{
392 struct nlattr **a = info->attrs;
393 u32 meter_id;
394 struct ovs_header *ovs_header = info->userhdr;
395 struct ovs_header *ovs_reply_header;
396 struct datapath *dp;
397 int err;
398 struct sk_buff *reply;
399 struct dp_meter *old_meter;
400
401 if (!a[OVS_METER_ATTR_ID])
402 return -EINVAL;
403 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
404
405 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_DEL,
406 &ovs_reply_header);
407 if (IS_ERR(reply))
408 return PTR_ERR(reply);
409
410 ovs_lock();
411
412 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
413 if (!dp) {
414 err = -ENODEV;
415 goto exit_unlock;
416 }
417
418 old_meter = lookup_meter(dp, meter_id);
419 if (old_meter) {
420 spin_lock_bh(&old_meter->lock);
421 err = ovs_meter_cmd_reply_stats(reply, meter_id, old_meter);
422 WARN_ON(err);
423 spin_unlock_bh(&old_meter->lock);
424 detach_meter(old_meter);
425 }
426 ovs_unlock();
427 ovs_meter_free(old_meter);
428 genlmsg_end(reply, ovs_reply_header);
429 return genlmsg_reply(reply, info);
430
431exit_unlock:
432 ovs_unlock();
433 nlmsg_free(reply);
434 return err;
435}
436
437/* Meter action execution.
438 *
439 * Return true 'meter_id' drop band is triggered. The 'skb' should be
440 * dropped by the caller'.
441 */
442bool ovs_meter_execute(struct datapath *dp, struct sk_buff *skb,
443 struct sw_flow_key *key, u32 meter_id)
444{
445 struct dp_meter *meter;
446 struct dp_meter_band *band;
447 long long int now_ms = div_u64(ktime_get_ns(), 1000 * 1000);
448 long long int long_delta_ms;
449 u32 delta_ms;
450 u32 cost;
451 int i, band_exceeded_max = -1;
452 u32 band_exceeded_rate = 0;
453
454 meter = lookup_meter(dp, meter_id);
455 /* Do not drop the packet when there is no meter. */
456 if (!meter)
457 return false;
458
459 /* Lock the meter while using it. */
460 spin_lock(&meter->lock);
461
462 long_delta_ms = (now_ms - meter->used); /* ms */
463
464 /* Make sure delta_ms will not be too large, so that bucket will not
465 * wrap around below.
466 */
467 delta_ms = (long_delta_ms > (long long int)meter->max_delta_t)
468 ? meter->max_delta_t : (u32)long_delta_ms;
469
470 /* Update meter statistics.
471 */
472 meter->used = now_ms;
473 meter->stats.n_packets += 1;
474 meter->stats.n_bytes += skb->len;
475
476 /* Bucket rate is either in kilobits per second, or in packets per
477 * second. We maintain the bucket in the units of either bits or
478 * 1/1000th of a packet, correspondingly.
479 * Then, when rate is multiplied with milliseconds, we get the
480 * bucket units:
481 * msec * kbps = bits, and
482 * msec * packets/sec = 1/1000 packets.
483 *
484 * 'cost' is the number of bucket units in this packet.
485 */
486 cost = (meter->kbps) ? skb->len * 8 : 1000;
487
488 /* Update all bands and find the one hit with the highest rate. */
489 for (i = 0; i < meter->n_bands; ++i) {
490 long long int max_bucket_size;
491
492 band = &meter->bands[i];
493 max_bucket_size = (band->burst_size + band->rate) * 1000LL;
494
495 band->bucket += delta_ms * band->rate;
496 if (band->bucket > max_bucket_size)
497 band->bucket = max_bucket_size;
498
499 if (band->bucket >= cost) {
500 band->bucket -= cost;
501 } else if (band->rate > band_exceeded_rate) {
502 band_exceeded_rate = band->rate;
503 band_exceeded_max = i;
504 }
505 }
506
507 if (band_exceeded_max >= 0) {
508 /* Update band statistics. */
509 band = &meter->bands[band_exceeded_max];
510 band->stats.n_packets += 1;
511 band->stats.n_bytes += skb->len;
512
513 /* Drop band triggered, let the caller drop the 'skb'. */
514 if (band->type == OVS_METER_BAND_TYPE_DROP) {
515 spin_unlock(&meter->lock);
516 return true;
517 }
518 }
519
520 spin_unlock(&meter->lock);
521 return false;
522}
523
524static struct genl_ops dp_meter_genl_ops[] = {
525 { .cmd = OVS_METER_CMD_FEATURES,
526 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
527 .flags = 0, /* OK for unprivileged users. */
528 .doit = ovs_meter_cmd_features
529 },
530 { .cmd = OVS_METER_CMD_SET,
531 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
532 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN
533 * privilege.
534 */
535 .doit = ovs_meter_cmd_set,
536 },
537 { .cmd = OVS_METER_CMD_GET,
538 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
539 .flags = 0, /* OK for unprivileged users. */
540 .doit = ovs_meter_cmd_get,
541 },
542 { .cmd = OVS_METER_CMD_DEL,
543 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
544 .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN
545 * privilege.
546 */
547 .doit = ovs_meter_cmd_del
548 },
549};
550
551static const struct genl_multicast_group ovs_meter_multicast_group = {
552 .name = OVS_METER_MCGROUP,
553};
554
555struct genl_family dp_meter_genl_family __ro_after_init = {
556 .hdrsize = sizeof(struct ovs_header),
557 .name = OVS_METER_FAMILY,
558 .version = OVS_METER_VERSION,
559 .maxattr = OVS_METER_ATTR_MAX,
560 .policy = meter_policy,
561 .netnsok = true,
562 .parallel_ops = true,
563 .ops = dp_meter_genl_ops,
564 .n_ops = ARRAY_SIZE(dp_meter_genl_ops),
565 .mcgrps = &ovs_meter_multicast_group,
566 .n_mcgrps = 1,
567 .module = THIS_MODULE,
568};
569
570int ovs_meters_init(struct datapath *dp)
571{
572 int i;
573
574 dp->meters = kmalloc_array(METER_HASH_BUCKETS,
575 sizeof(struct hlist_head), GFP_KERNEL);
576
577 if (!dp->meters)
578 return -ENOMEM;
579
580 for (i = 0; i < METER_HASH_BUCKETS; i++)
581 INIT_HLIST_HEAD(&dp->meters[i]);
582
583 return 0;
584}
585
586void ovs_meters_exit(struct datapath *dp)
587{
588 int i;
589
590 for (i = 0; i < METER_HASH_BUCKETS; i++) {
591 struct hlist_head *head = &dp->meters[i];
592 struct dp_meter *meter;
593 struct hlist_node *n;
594
595 hlist_for_each_entry_safe(meter, n, head, dp_hash_node)
596 kfree(meter);
597 }
598
599 kfree(dp->meters);
600}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (c) 2017 Nicira, Inc.
4 */
5
6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8#include <linux/if.h>
9#include <linux/skbuff.h>
10#include <linux/ip.h>
11#include <linux/kernel.h>
12#include <linux/openvswitch.h>
13#include <linux/netlink.h>
14#include <linux/rculist.h>
15
16#include <net/netlink.h>
17#include <net/genetlink.h>
18
19#include "datapath.h"
20#include "meter.h"
21
22static 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 },
31};
32
33static 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) },
38};
39
40static u32 meter_hash(struct dp_meter_instance *ti, u32 id)
41{
42 return id % ti->n_meters;
43}
44
45static void ovs_meter_free(struct dp_meter *meter)
46{
47 if (!meter)
48 return;
49
50 kfree_rcu(meter, rcu);
51}
52
53/* Call with ovs_mutex or RCU read lock. */
54static struct dp_meter *lookup_meter(const struct dp_meter_table *tbl,
55 u32 meter_id)
56{
57 struct dp_meter_instance *ti = rcu_dereference_ovsl(tbl->ti);
58 u32 hash = meter_hash(ti, meter_id);
59 struct dp_meter *meter;
60
61 meter = rcu_dereference_ovsl(ti->dp_meters[hash]);
62 if (meter && likely(meter->id == meter_id))
63 return meter;
64
65 return NULL;
66}
67
68static struct dp_meter_instance *dp_meter_instance_alloc(const u32 size)
69{
70 struct dp_meter_instance *ti;
71
72 ti = kvzalloc(struct_size(ti, dp_meters, size), GFP_KERNEL);
73 if (!ti)
74 return NULL;
75
76 ti->n_meters = size;
77
78 return ti;
79}
80
81static void dp_meter_instance_free(struct dp_meter_instance *ti)
82{
83 kvfree(ti);
84}
85
86static void dp_meter_instance_free_rcu(struct rcu_head *rcu)
87{
88 struct dp_meter_instance *ti;
89
90 ti = container_of(rcu, struct dp_meter_instance, rcu);
91 kvfree(ti);
92}
93
94static int
95dp_meter_instance_realloc(struct dp_meter_table *tbl, u32 size)
96{
97 struct dp_meter_instance *ti = rcu_dereference_ovsl(tbl->ti);
98 int n_meters = min(size, ti->n_meters);
99 struct dp_meter_instance *new_ti;
100 int i;
101
102 new_ti = dp_meter_instance_alloc(size);
103 if (!new_ti)
104 return -ENOMEM;
105
106 for (i = 0; i < n_meters; i++)
107 if (rcu_dereference_ovsl(ti->dp_meters[i]))
108 new_ti->dp_meters[i] = ti->dp_meters[i];
109
110 rcu_assign_pointer(tbl->ti, new_ti);
111 call_rcu(&ti->rcu, dp_meter_instance_free_rcu);
112
113 return 0;
114}
115
116static void dp_meter_instance_insert(struct dp_meter_instance *ti,
117 struct dp_meter *meter)
118{
119 u32 hash;
120
121 hash = meter_hash(ti, meter->id);
122 rcu_assign_pointer(ti->dp_meters[hash], meter);
123}
124
125static void dp_meter_instance_remove(struct dp_meter_instance *ti,
126 struct dp_meter *meter)
127{
128 u32 hash;
129
130 hash = meter_hash(ti, meter->id);
131 RCU_INIT_POINTER(ti->dp_meters[hash], NULL);
132}
133
134static int attach_meter(struct dp_meter_table *tbl, struct dp_meter *meter)
135{
136 struct dp_meter_instance *ti = rcu_dereference_ovsl(tbl->ti);
137 u32 hash = meter_hash(ti, meter->id);
138 int err;
139
140 /* In generally, slots selected should be empty, because
141 * OvS uses id-pool to fetch a available id.
142 */
143 if (unlikely(rcu_dereference_ovsl(ti->dp_meters[hash])))
144 return -EBUSY;
145
146 dp_meter_instance_insert(ti, meter);
147
148 /* That function is thread-safe. */
149 tbl->count++;
150 if (tbl->count >= tbl->max_meters_allowed) {
151 err = -EFBIG;
152 goto attach_err;
153 }
154
155 if (tbl->count >= ti->n_meters &&
156 dp_meter_instance_realloc(tbl, ti->n_meters * 2)) {
157 err = -ENOMEM;
158 goto attach_err;
159 }
160
161 return 0;
162
163attach_err:
164 dp_meter_instance_remove(ti, meter);
165 tbl->count--;
166 return err;
167}
168
169static int detach_meter(struct dp_meter_table *tbl, struct dp_meter *meter)
170{
171 struct dp_meter_instance *ti;
172
173 ASSERT_OVSL();
174 if (!meter)
175 return 0;
176
177 ti = rcu_dereference_ovsl(tbl->ti);
178 dp_meter_instance_remove(ti, meter);
179
180 tbl->count--;
181
182 /* Shrink the meter array if necessary. */
183 if (ti->n_meters > DP_METER_ARRAY_SIZE_MIN &&
184 tbl->count <= (ti->n_meters / 4)) {
185 int half_size = ti->n_meters / 2;
186 int i;
187
188 /* Avoid hash collision, don't move slots to other place.
189 * Make sure there are no references of meters in array
190 * which will be released.
191 */
192 for (i = half_size; i < ti->n_meters; i++)
193 if (rcu_dereference_ovsl(ti->dp_meters[i]))
194 goto out;
195
196 if (dp_meter_instance_realloc(tbl, half_size))
197 goto shrink_err;
198 }
199
200out:
201 return 0;
202
203shrink_err:
204 dp_meter_instance_insert(ti, meter);
205 tbl->count++;
206 return -ENOMEM;
207}
208
209static struct sk_buff *
210ovs_meter_cmd_reply_start(struct genl_info *info, u8 cmd,
211 struct ovs_header **ovs_reply_header)
212{
213 struct sk_buff *skb;
214 struct ovs_header *ovs_header = genl_info_userhdr(info);
215
216 skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
217 if (!skb)
218 return ERR_PTR(-ENOMEM);
219
220 *ovs_reply_header = genlmsg_put(skb, info->snd_portid,
221 info->snd_seq,
222 &dp_meter_genl_family, 0, cmd);
223 if (!*ovs_reply_header) {
224 nlmsg_free(skb);
225 return ERR_PTR(-EMSGSIZE);
226 }
227 (*ovs_reply_header)->dp_ifindex = ovs_header->dp_ifindex;
228
229 return skb;
230}
231
232static int ovs_meter_cmd_reply_stats(struct sk_buff *reply, u32 meter_id,
233 struct dp_meter *meter)
234{
235 struct nlattr *nla;
236 struct dp_meter_band *band;
237 u16 i;
238
239 if (nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id))
240 goto error;
241
242 if (nla_put(reply, OVS_METER_ATTR_STATS,
243 sizeof(struct ovs_flow_stats), &meter->stats))
244 goto error;
245
246 if (nla_put_u64_64bit(reply, OVS_METER_ATTR_USED, meter->used,
247 OVS_METER_ATTR_PAD))
248 goto error;
249
250 nla = nla_nest_start_noflag(reply, OVS_METER_ATTR_BANDS);
251 if (!nla)
252 goto error;
253
254 band = meter->bands;
255
256 for (i = 0; i < meter->n_bands; ++i, ++band) {
257 struct nlattr *band_nla;
258
259 band_nla = nla_nest_start_noflag(reply, OVS_BAND_ATTR_UNSPEC);
260 if (!band_nla || nla_put(reply, OVS_BAND_ATTR_STATS,
261 sizeof(struct ovs_flow_stats),
262 &band->stats))
263 goto error;
264 nla_nest_end(reply, band_nla);
265 }
266 nla_nest_end(reply, nla);
267
268 return 0;
269error:
270 return -EMSGSIZE;
271}
272
273static int ovs_meter_cmd_features(struct sk_buff *skb, struct genl_info *info)
274{
275 struct ovs_header *ovs_header = genl_info_userhdr(info);
276 struct ovs_header *ovs_reply_header;
277 struct nlattr *nla, *band_nla;
278 struct sk_buff *reply;
279 struct datapath *dp;
280 int err = -EMSGSIZE;
281
282 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_FEATURES,
283 &ovs_reply_header);
284 if (IS_ERR(reply))
285 return PTR_ERR(reply);
286
287 ovs_lock();
288 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
289 if (!dp) {
290 err = -ENODEV;
291 goto exit_unlock;
292 }
293
294 if (nla_put_u32(reply, OVS_METER_ATTR_MAX_METERS,
295 dp->meter_tbl.max_meters_allowed))
296 goto exit_unlock;
297
298 ovs_unlock();
299
300 if (nla_put_u32(reply, OVS_METER_ATTR_MAX_BANDS, DP_MAX_BANDS))
301 goto nla_put_failure;
302
303 nla = nla_nest_start_noflag(reply, OVS_METER_ATTR_BANDS);
304 if (!nla)
305 goto nla_put_failure;
306
307 band_nla = nla_nest_start_noflag(reply, OVS_BAND_ATTR_UNSPEC);
308 if (!band_nla)
309 goto nla_put_failure;
310 /* Currently only DROP band type is supported. */
311 if (nla_put_u32(reply, OVS_BAND_ATTR_TYPE, OVS_METER_BAND_TYPE_DROP))
312 goto nla_put_failure;
313 nla_nest_end(reply, band_nla);
314 nla_nest_end(reply, nla);
315
316 genlmsg_end(reply, ovs_reply_header);
317 return genlmsg_reply(reply, info);
318
319exit_unlock:
320 ovs_unlock();
321nla_put_failure:
322 nlmsg_free(reply);
323 return err;
324}
325
326static struct dp_meter *dp_meter_create(struct nlattr **a)
327{
328 struct nlattr *nla;
329 int rem;
330 u16 n_bands = 0;
331 struct dp_meter *meter;
332 struct dp_meter_band *band;
333 int err;
334
335 /* Validate attributes, count the bands. */
336 if (!a[OVS_METER_ATTR_BANDS])
337 return ERR_PTR(-EINVAL);
338
339 nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem)
340 if (++n_bands > DP_MAX_BANDS)
341 return ERR_PTR(-EINVAL);
342
343 /* Allocate and set up the meter before locking anything. */
344 meter = kzalloc(struct_size(meter, bands, n_bands), GFP_KERNEL_ACCOUNT);
345 if (!meter)
346 return ERR_PTR(-ENOMEM);
347
348 meter->id = nla_get_u32(a[OVS_METER_ATTR_ID]);
349 meter->used = div_u64(ktime_get_ns(), 1000 * 1000);
350 meter->kbps = a[OVS_METER_ATTR_KBPS] ? 1 : 0;
351 meter->keep_stats = !a[OVS_METER_ATTR_CLEAR];
352 spin_lock_init(&meter->lock);
353 if (meter->keep_stats && a[OVS_METER_ATTR_STATS]) {
354 meter->stats = *(struct ovs_flow_stats *)
355 nla_data(a[OVS_METER_ATTR_STATS]);
356 }
357 meter->n_bands = n_bands;
358
359 /* Set up meter bands. */
360 band = meter->bands;
361 nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem) {
362 struct nlattr *attr[OVS_BAND_ATTR_MAX + 1];
363 u32 band_max_delta_t;
364
365 err = nla_parse_deprecated((struct nlattr **)&attr,
366 OVS_BAND_ATTR_MAX, nla_data(nla),
367 nla_len(nla), band_policy, NULL);
368 if (err)
369 goto exit_free_meter;
370
371 if (!attr[OVS_BAND_ATTR_TYPE] ||
372 !attr[OVS_BAND_ATTR_RATE] ||
373 !attr[OVS_BAND_ATTR_BURST]) {
374 err = -EINVAL;
375 goto exit_free_meter;
376 }
377
378 band->type = nla_get_u32(attr[OVS_BAND_ATTR_TYPE]);
379 band->rate = nla_get_u32(attr[OVS_BAND_ATTR_RATE]);
380 if (band->rate == 0) {
381 err = -EINVAL;
382 goto exit_free_meter;
383 }
384
385 band->burst_size = nla_get_u32(attr[OVS_BAND_ATTR_BURST]);
386 /* Figure out max delta_t that is enough to fill any bucket.
387 * Keep max_delta_t size to the bucket units:
388 * pkts => 1/1000 packets, kilobits => bits.
389 *
390 * Start with a full bucket.
391 */
392 band->bucket = band->burst_size * 1000ULL;
393 band_max_delta_t = div_u64(band->bucket, band->rate);
394 if (band_max_delta_t > meter->max_delta_t)
395 meter->max_delta_t = band_max_delta_t;
396 band++;
397 }
398
399 return meter;
400
401exit_free_meter:
402 kfree(meter);
403 return ERR_PTR(err);
404}
405
406static int ovs_meter_cmd_set(struct sk_buff *skb, struct genl_info *info)
407{
408 struct nlattr **a = info->attrs;
409 struct dp_meter *meter, *old_meter;
410 struct sk_buff *reply;
411 struct ovs_header *ovs_reply_header;
412 struct ovs_header *ovs_header = genl_info_userhdr(info);
413 struct dp_meter_table *meter_tbl;
414 struct datapath *dp;
415 int err;
416 u32 meter_id;
417 bool failed;
418
419 if (!a[OVS_METER_ATTR_ID])
420 return -EINVAL;
421
422 meter = dp_meter_create(a);
423 if (IS_ERR(meter))
424 return PTR_ERR(meter);
425
426 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_SET,
427 &ovs_reply_header);
428 if (IS_ERR(reply)) {
429 err = PTR_ERR(reply);
430 goto exit_free_meter;
431 }
432
433 ovs_lock();
434 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
435 if (!dp) {
436 err = -ENODEV;
437 goto exit_unlock;
438 }
439
440 meter_tbl = &dp->meter_tbl;
441 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
442
443 old_meter = lookup_meter(meter_tbl, meter_id);
444 err = detach_meter(meter_tbl, old_meter);
445 if (err)
446 goto exit_unlock;
447
448 err = attach_meter(meter_tbl, meter);
449 if (err)
450 goto exit_free_old_meter;
451
452 ovs_unlock();
453
454 /* Build response with the meter_id and stats from
455 * the old meter, if any.
456 */
457 failed = nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id);
458 WARN_ON(failed);
459 if (old_meter) {
460 spin_lock_bh(&old_meter->lock);
461 if (old_meter->keep_stats) {
462 err = ovs_meter_cmd_reply_stats(reply, meter_id,
463 old_meter);
464 WARN_ON(err);
465 }
466 spin_unlock_bh(&old_meter->lock);
467 ovs_meter_free(old_meter);
468 }
469
470 genlmsg_end(reply, ovs_reply_header);
471 return genlmsg_reply(reply, info);
472
473exit_free_old_meter:
474 ovs_meter_free(old_meter);
475exit_unlock:
476 ovs_unlock();
477 nlmsg_free(reply);
478exit_free_meter:
479 kfree(meter);
480 return err;
481}
482
483static int ovs_meter_cmd_get(struct sk_buff *skb, struct genl_info *info)
484{
485 struct ovs_header *ovs_header = genl_info_userhdr(info);
486 struct ovs_header *ovs_reply_header;
487 struct nlattr **a = info->attrs;
488 struct dp_meter *meter;
489 struct sk_buff *reply;
490 struct datapath *dp;
491 u32 meter_id;
492 int err;
493
494 if (!a[OVS_METER_ATTR_ID])
495 return -EINVAL;
496
497 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
498
499 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_GET,
500 &ovs_reply_header);
501 if (IS_ERR(reply))
502 return PTR_ERR(reply);
503
504 ovs_lock();
505
506 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
507 if (!dp) {
508 err = -ENODEV;
509 goto exit_unlock;
510 }
511
512 /* Locate meter, copy stats. */
513 meter = lookup_meter(&dp->meter_tbl, meter_id);
514 if (!meter) {
515 err = -ENOENT;
516 goto exit_unlock;
517 }
518
519 spin_lock_bh(&meter->lock);
520 err = ovs_meter_cmd_reply_stats(reply, meter_id, meter);
521 spin_unlock_bh(&meter->lock);
522 if (err)
523 goto exit_unlock;
524
525 ovs_unlock();
526
527 genlmsg_end(reply, ovs_reply_header);
528 return genlmsg_reply(reply, info);
529
530exit_unlock:
531 ovs_unlock();
532 nlmsg_free(reply);
533 return err;
534}
535
536static int ovs_meter_cmd_del(struct sk_buff *skb, struct genl_info *info)
537{
538 struct ovs_header *ovs_header = genl_info_userhdr(info);
539 struct ovs_header *ovs_reply_header;
540 struct nlattr **a = info->attrs;
541 struct dp_meter *old_meter;
542 struct sk_buff *reply;
543 struct datapath *dp;
544 u32 meter_id;
545 int err;
546
547 if (!a[OVS_METER_ATTR_ID])
548 return -EINVAL;
549
550 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_DEL,
551 &ovs_reply_header);
552 if (IS_ERR(reply))
553 return PTR_ERR(reply);
554
555 ovs_lock();
556
557 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
558 if (!dp) {
559 err = -ENODEV;
560 goto exit_unlock;
561 }
562
563 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
564 old_meter = lookup_meter(&dp->meter_tbl, meter_id);
565 if (old_meter) {
566 spin_lock_bh(&old_meter->lock);
567 err = ovs_meter_cmd_reply_stats(reply, meter_id, old_meter);
568 WARN_ON(err);
569 spin_unlock_bh(&old_meter->lock);
570
571 err = detach_meter(&dp->meter_tbl, old_meter);
572 if (err)
573 goto exit_unlock;
574 }
575
576 ovs_unlock();
577 ovs_meter_free(old_meter);
578 genlmsg_end(reply, ovs_reply_header);
579 return genlmsg_reply(reply, info);
580
581exit_unlock:
582 ovs_unlock();
583 nlmsg_free(reply);
584 return err;
585}
586
587/* Meter action execution.
588 *
589 * Return true 'meter_id' drop band is triggered. The 'skb' should be
590 * dropped by the caller'.
591 */
592bool ovs_meter_execute(struct datapath *dp, struct sk_buff *skb,
593 struct sw_flow_key *key, u32 meter_id)
594{
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;
601 u32 delta_ms;
602 u32 cost;
603
604 meter = lookup_meter(&dp->meter_tbl, meter_id);
605 /* Do not drop the packet when there is no meter. */
606 if (!meter)
607 return false;
608
609 /* Lock the meter while using it. */
610 spin_lock(&meter->lock);
611
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.
618 */
619 long_delta_ms = 0;
620 }
621
622 /* Make sure delta_ms will not be too large, so that bucket will not
623 * wrap around below.
624 */
625 delta_ms = (long_delta_ms > (long long int)meter->max_delta_t)
626 ? meter->max_delta_t : (u32)long_delta_ms;
627
628 /* Update meter statistics.
629 */
630 meter->used = now_ms;
631 meter->stats.n_packets += 1;
632 meter->stats.n_bytes += skb->len;
633
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
638 * bucket units:
639 * msec * kbps = bits, and
640 * msec * packets/sec = 1/1000 packets.
641 *
642 * 'cost' is the number of bucket units in this packet.
643 */
644 cost = (meter->kbps) ? skb->len * 8 : 1000;
645
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;
649
650 band = &meter->bands[i];
651 max_bucket_size = band->burst_size * 1000LL;
652
653 band->bucket += delta_ms * band->rate;
654 if (band->bucket > max_bucket_size)
655 band->bucket = max_bucket_size;
656
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;
662 }
663 }
664
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;
670
671 /* Drop band triggered, let the caller drop the 'skb'. */
672 if (band->type == OVS_METER_BAND_TYPE_DROP) {
673 spin_unlock(&meter->lock);
674 return true;
675 }
676 }
677
678 spin_unlock(&meter->lock);
679 return false;
680}
681
682static 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
687 },
688 { .cmd = OVS_METER_CMD_SET,
689 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
690 .flags = GENL_UNS_ADMIN_PERM, /* Requires CAP_NET_ADMIN
691 * privilege.
692 */
693 .doit = ovs_meter_cmd_set,
694 },
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,
699 },
700 { .cmd = OVS_METER_CMD_DEL,
701 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
702 .flags = GENL_UNS_ADMIN_PERM, /* Requires CAP_NET_ADMIN
703 * privilege.
704 */
705 .doit = ovs_meter_cmd_del
706 },
707};
708
709static const struct genl_multicast_group ovs_meter_multicast_group = {
710 .name = OVS_METER_MCGROUP,
711};
712
713struct 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,
719 .netnsok = true,
720 .parallel_ops = true,
721 .small_ops = dp_meter_genl_ops,
722 .n_small_ops = ARRAY_SIZE(dp_meter_genl_ops),
723 .resv_start_op = OVS_METER_CMD_GET + 1,
724 .mcgrps = &ovs_meter_multicast_group,
725 .n_mcgrps = 1,
726 .module = THIS_MODULE,
727};
728
729int ovs_meters_init(struct datapath *dp)
730{
731 struct dp_meter_table *tbl = &dp->meter_tbl;
732 struct dp_meter_instance *ti;
733 unsigned long free_mem_bytes;
734
735 ti = dp_meter_instance_alloc(DP_METER_ARRAY_SIZE_MIN);
736 if (!ti)
737 return -ENOMEM;
738
739 /* Allow meters in a datapath to use ~3.12% of physical memory. */
740 free_mem_bytes = nr_free_buffer_pages() * (PAGE_SIZE >> 5);
741 tbl->max_meters_allowed = min(free_mem_bytes / sizeof(struct dp_meter),
742 DP_METER_NUM_MAX);
743 if (!tbl->max_meters_allowed)
744 goto out_err;
745
746 rcu_assign_pointer(tbl->ti, ti);
747 tbl->count = 0;
748
749 return 0;
750
751out_err:
752 dp_meter_instance_free(ti);
753 return -ENOMEM;
754}
755
756void ovs_meters_exit(struct datapath *dp)
757{
758 struct dp_meter_table *tbl = &dp->meter_tbl;
759 struct dp_meter_instance *ti = rcu_dereference_raw(tbl->ti);
760 int i;
761
762 for (i = 0; i < ti->n_meters; i++)
763 ovs_meter_free(rcu_dereference_raw(ti->dp_meters[i]));
764
765 dp_meter_instance_free(ti);
766}