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(sizeof(*ti) +
73 sizeof(struct dp_meter *) * size,
74 GFP_KERNEL);
75 if (!ti)
76 return NULL;
77
78 ti->n_meters = size;
79
80 return ti;
81}
82
83static void dp_meter_instance_free(struct dp_meter_instance *ti)
84{
85 kvfree(ti);
86}
87
88static void dp_meter_instance_free_rcu(struct rcu_head *rcu)
89{
90 struct dp_meter_instance *ti;
91
92 ti = container_of(rcu, struct dp_meter_instance, rcu);
93 kvfree(ti);
94}
95
96static int
97dp_meter_instance_realloc(struct dp_meter_table *tbl, u32 size)
98{
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;
102 int i;
103
104 new_ti = dp_meter_instance_alloc(size);
105 if (!new_ti)
106 return -ENOMEM;
107
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];
111
112 rcu_assign_pointer(tbl->ti, new_ti);
113 call_rcu(&ti->rcu, dp_meter_instance_free_rcu);
114
115 return 0;
116}
117
118static void dp_meter_instance_insert(struct dp_meter_instance *ti,
119 struct dp_meter *meter)
120{
121 u32 hash;
122
123 hash = meter_hash(ti, meter->id);
124 rcu_assign_pointer(ti->dp_meters[hash], meter);
125}
126
127static void dp_meter_instance_remove(struct dp_meter_instance *ti,
128 struct dp_meter *meter)
129{
130 u32 hash;
131
132 hash = meter_hash(ti, meter->id);
133 RCU_INIT_POINTER(ti->dp_meters[hash], NULL);
134}
135
136static int attach_meter(struct dp_meter_table *tbl, struct dp_meter *meter)
137{
138 struct dp_meter_instance *ti = rcu_dereference_ovsl(tbl->ti);
139 u32 hash = meter_hash(ti, meter->id);
140 int err;
141
142 /* In generally, slots selected should be empty, because
143 * OvS uses id-pool to fetch a available id.
144 */
145 if (unlikely(rcu_dereference_ovsl(ti->dp_meters[hash])))
146 return -EBUSY;
147
148 dp_meter_instance_insert(ti, meter);
149
150 /* That function is thread-safe. */
151 tbl->count++;
152 if (tbl->count >= tbl->max_meters_allowed) {
153 err = -EFBIG;
154 goto attach_err;
155 }
156
157 if (tbl->count >= ti->n_meters &&
158 dp_meter_instance_realloc(tbl, ti->n_meters * 2)) {
159 err = -ENOMEM;
160 goto attach_err;
161 }
162
163 return 0;
164
165attach_err:
166 dp_meter_instance_remove(ti, meter);
167 tbl->count--;
168 return err;
169}
170
171static int detach_meter(struct dp_meter_table *tbl, struct dp_meter *meter)
172{
173 struct dp_meter_instance *ti;
174
175 ASSERT_OVSL();
176 if (!meter)
177 return 0;
178
179 ti = rcu_dereference_ovsl(tbl->ti);
180 dp_meter_instance_remove(ti, meter);
181
182 tbl->count--;
183
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;
188 int i;
189
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.
193 */
194 for (i = half_size; i < ti->n_meters; i++)
195 if (rcu_dereference_ovsl(ti->dp_meters[i]))
196 goto out;
197
198 if (dp_meter_instance_realloc(tbl, half_size))
199 goto shrink_err;
200 }
201
202out:
203 return 0;
204
205shrink_err:
206 dp_meter_instance_insert(ti, meter);
207 tbl->count++;
208 return -ENOMEM;
209}
210
211static struct sk_buff *
212ovs_meter_cmd_reply_start(struct genl_info *info, u8 cmd,
213 struct ovs_header **ovs_reply_header)
214{
215 struct sk_buff *skb;
216 struct ovs_header *ovs_header = info->userhdr;
217
218 skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
219 if (!skb)
220 return ERR_PTR(-ENOMEM);
221
222 *ovs_reply_header = genlmsg_put(skb, info->snd_portid,
223 info->snd_seq,
224 &dp_meter_genl_family, 0, cmd);
225 if (!*ovs_reply_header) {
226 nlmsg_free(skb);
227 return ERR_PTR(-EMSGSIZE);
228 }
229 (*ovs_reply_header)->dp_ifindex = ovs_header->dp_ifindex;
230
231 return skb;
232}
233
234static int ovs_meter_cmd_reply_stats(struct sk_buff *reply, u32 meter_id,
235 struct dp_meter *meter)
236{
237 struct nlattr *nla;
238 struct dp_meter_band *band;
239 u16 i;
240
241 if (nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id))
242 goto error;
243
244 if (nla_put(reply, OVS_METER_ATTR_STATS,
245 sizeof(struct ovs_flow_stats), &meter->stats))
246 goto error;
247
248 if (nla_put_u64_64bit(reply, OVS_METER_ATTR_USED, meter->used,
249 OVS_METER_ATTR_PAD))
250 goto error;
251
252 nla = nla_nest_start_noflag(reply, OVS_METER_ATTR_BANDS);
253 if (!nla)
254 goto error;
255
256 band = meter->bands;
257
258 for (i = 0; i < meter->n_bands; ++i, ++band) {
259 struct nlattr *band_nla;
260
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),
264 &band->stats))
265 goto error;
266 nla_nest_end(reply, band_nla);
267 }
268 nla_nest_end(reply, nla);
269
270 return 0;
271error:
272 return -EMSGSIZE;
273}
274
275static int ovs_meter_cmd_features(struct sk_buff *skb, struct genl_info *info)
276{
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;
281 struct datapath *dp;
282 int err = -EMSGSIZE;
283
284 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_FEATURES,
285 &ovs_reply_header);
286 if (IS_ERR(reply))
287 return PTR_ERR(reply);
288
289 ovs_lock();
290 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
291 if (!dp) {
292 err = -ENODEV;
293 goto exit_unlock;
294 }
295
296 if (nla_put_u32(reply, OVS_METER_ATTR_MAX_METERS,
297 dp->meter_tbl.max_meters_allowed))
298 goto exit_unlock;
299
300 ovs_unlock();
301
302 if (nla_put_u32(reply, OVS_METER_ATTR_MAX_BANDS, DP_MAX_BANDS))
303 goto nla_put_failure;
304
305 nla = nla_nest_start_noflag(reply, OVS_METER_ATTR_BANDS);
306 if (!nla)
307 goto nla_put_failure;
308
309 band_nla = nla_nest_start_noflag(reply, OVS_BAND_ATTR_UNSPEC);
310 if (!band_nla)
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);
317
318 genlmsg_end(reply, ovs_reply_header);
319 return genlmsg_reply(reply, info);
320
321exit_unlock:
322 ovs_unlock();
323nla_put_failure:
324 nlmsg_free(reply);
325 return err;
326}
327
328static struct dp_meter *dp_meter_create(struct nlattr **a)
329{
330 struct nlattr *nla;
331 int rem;
332 u16 n_bands = 0;
333 struct dp_meter *meter;
334 struct dp_meter_band *band;
335 int err;
336
337 /* Validate attributes, count the bands. */
338 if (!a[OVS_METER_ATTR_BANDS])
339 return ERR_PTR(-EINVAL);
340
341 nla_for_each_nested(nla, a[OVS_METER_ATTR_BANDS], rem)
342 if (++n_bands > DP_MAX_BANDS)
343 return ERR_PTR(-EINVAL);
344
345 /* Allocate and set up the meter before locking anything. */
346 meter = kzalloc(struct_size(meter, bands, n_bands), GFP_KERNEL_ACCOUNT);
347 if (!meter)
348 return ERR_PTR(-ENOMEM);
349
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]);
358 }
359 meter->n_bands = n_bands;
360
361 /* Set up meter bands. */
362 band = 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;
366
367 err = nla_parse_deprecated((struct nlattr **)&attr,
368 OVS_BAND_ATTR_MAX, nla_data(nla),
369 nla_len(nla), band_policy, NULL);
370 if (err)
371 goto exit_free_meter;
372
373 if (!attr[OVS_BAND_ATTR_TYPE] ||
374 !attr[OVS_BAND_ATTR_RATE] ||
375 !attr[OVS_BAND_ATTR_BURST]) {
376 err = -EINVAL;
377 goto exit_free_meter;
378 }
379
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) {
383 err = -EINVAL;
384 goto exit_free_meter;
385 }
386
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.
391 *
392 * Start with a full bucket.
393 */
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;
398 band++;
399 }
400
401 return meter;
402
403exit_free_meter:
404 kfree(meter);
405 return ERR_PTR(err);
406}
407
408static int ovs_meter_cmd_set(struct sk_buff *skb, struct genl_info *info)
409{
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;
416 struct datapath *dp;
417 int err;
418 u32 meter_id;
419 bool failed;
420
421 if (!a[OVS_METER_ATTR_ID])
422 return -EINVAL;
423
424 meter = dp_meter_create(a);
425 if (IS_ERR(meter))
426 return PTR_ERR(meter);
427
428 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_SET,
429 &ovs_reply_header);
430 if (IS_ERR(reply)) {
431 err = PTR_ERR(reply);
432 goto exit_free_meter;
433 }
434
435 ovs_lock();
436 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
437 if (!dp) {
438 err = -ENODEV;
439 goto exit_unlock;
440 }
441
442 meter_tbl = &dp->meter_tbl;
443 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
444
445 old_meter = lookup_meter(meter_tbl, meter_id);
446 err = detach_meter(meter_tbl, old_meter);
447 if (err)
448 goto exit_unlock;
449
450 err = attach_meter(meter_tbl, meter);
451 if (err)
452 goto exit_free_old_meter;
453
454 ovs_unlock();
455
456 /* Build response with the meter_id and stats from
457 * the old meter, if any.
458 */
459 failed = nla_put_u32(reply, OVS_METER_ATTR_ID, meter_id);
460 WARN_ON(failed);
461 if (old_meter) {
462 spin_lock_bh(&old_meter->lock);
463 if (old_meter->keep_stats) {
464 err = ovs_meter_cmd_reply_stats(reply, meter_id,
465 old_meter);
466 WARN_ON(err);
467 }
468 spin_unlock_bh(&old_meter->lock);
469 ovs_meter_free(old_meter);
470 }
471
472 genlmsg_end(reply, ovs_reply_header);
473 return genlmsg_reply(reply, info);
474
475exit_free_old_meter:
476 ovs_meter_free(old_meter);
477exit_unlock:
478 ovs_unlock();
479 nlmsg_free(reply);
480exit_free_meter:
481 kfree(meter);
482 return err;
483}
484
485static int ovs_meter_cmd_get(struct sk_buff *skb, struct genl_info *info)
486{
487 struct ovs_header *ovs_header = info->userhdr;
488 struct ovs_header *ovs_reply_header;
489 struct nlattr **a = info->attrs;
490 struct dp_meter *meter;
491 struct sk_buff *reply;
492 struct datapath *dp;
493 u32 meter_id;
494 int err;
495
496 if (!a[OVS_METER_ATTR_ID])
497 return -EINVAL;
498
499 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
500
501 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_GET,
502 &ovs_reply_header);
503 if (IS_ERR(reply))
504 return PTR_ERR(reply);
505
506 ovs_lock();
507
508 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
509 if (!dp) {
510 err = -ENODEV;
511 goto exit_unlock;
512 }
513
514 /* Locate meter, copy stats. */
515 meter = lookup_meter(&dp->meter_tbl, meter_id);
516 if (!meter) {
517 err = -ENOENT;
518 goto exit_unlock;
519 }
520
521 spin_lock_bh(&meter->lock);
522 err = ovs_meter_cmd_reply_stats(reply, meter_id, meter);
523 spin_unlock_bh(&meter->lock);
524 if (err)
525 goto exit_unlock;
526
527 ovs_unlock();
528
529 genlmsg_end(reply, ovs_reply_header);
530 return genlmsg_reply(reply, info);
531
532exit_unlock:
533 ovs_unlock();
534 nlmsg_free(reply);
535 return err;
536}
537
538static int ovs_meter_cmd_del(struct sk_buff *skb, struct genl_info *info)
539{
540 struct ovs_header *ovs_header = info->userhdr;
541 struct ovs_header *ovs_reply_header;
542 struct nlattr **a = info->attrs;
543 struct dp_meter *old_meter;
544 struct sk_buff *reply;
545 struct datapath *dp;
546 u32 meter_id;
547 int err;
548
549 if (!a[OVS_METER_ATTR_ID])
550 return -EINVAL;
551
552 reply = ovs_meter_cmd_reply_start(info, OVS_METER_CMD_DEL,
553 &ovs_reply_header);
554 if (IS_ERR(reply))
555 return PTR_ERR(reply);
556
557 ovs_lock();
558
559 dp = get_dp(sock_net(skb->sk), ovs_header->dp_ifindex);
560 if (!dp) {
561 err = -ENODEV;
562 goto exit_unlock;
563 }
564
565 meter_id = nla_get_u32(a[OVS_METER_ATTR_ID]);
566 old_meter = lookup_meter(&dp->meter_tbl, meter_id);
567 if (old_meter) {
568 spin_lock_bh(&old_meter->lock);
569 err = ovs_meter_cmd_reply_stats(reply, meter_id, old_meter);
570 WARN_ON(err);
571 spin_unlock_bh(&old_meter->lock);
572
573 err = detach_meter(&dp->meter_tbl, old_meter);
574 if (err)
575 goto exit_unlock;
576 }
577
578 ovs_unlock();
579 ovs_meter_free(old_meter);
580 genlmsg_end(reply, ovs_reply_header);
581 return genlmsg_reply(reply, info);
582
583exit_unlock:
584 ovs_unlock();
585 nlmsg_free(reply);
586 return err;
587}
588
589/* Meter action execution.
590 *
591 * Return true 'meter_id' drop band is triggered. The 'skb' should be
592 * dropped by the caller'.
593 */
594bool ovs_meter_execute(struct datapath *dp, struct sk_buff *skb,
595 struct sw_flow_key *key, u32 meter_id)
596{
597 long long int now_ms = div_u64(ktime_get_ns(), 1000 * 1000);
598 long long int long_delta_ms;
599 struct dp_meter_band *band;
600 struct dp_meter *meter;
601 int i, band_exceeded_max = -1;
602 u32 band_exceeded_rate = 0;
603 u32 delta_ms;
604 u32 cost;
605
606 meter = lookup_meter(&dp->meter_tbl, meter_id);
607 /* Do not drop the packet when there is no meter. */
608 if (!meter)
609 return false;
610
611 /* Lock the meter while using it. */
612 spin_lock(&meter->lock);
613
614 long_delta_ms = (now_ms - meter->used); /* ms */
615 if (long_delta_ms < 0) {
616 /* This condition means that we have several threads fighting
617 * for a meter lock, and the one who received the packets a
618 * bit later wins. Assuming that all racing threads received
619 * packets at the same time to avoid overflow.
620 */
621 long_delta_ms = 0;
622 }
623
624 /* Make sure delta_ms will not be too large, so that bucket will not
625 * wrap around below.
626 */
627 delta_ms = (long_delta_ms > (long long int)meter->max_delta_t)
628 ? meter->max_delta_t : (u32)long_delta_ms;
629
630 /* Update meter statistics.
631 */
632 meter->used = now_ms;
633 meter->stats.n_packets += 1;
634 meter->stats.n_bytes += skb->len;
635
636 /* Bucket rate is either in kilobits per second, or in packets per
637 * second. We maintain the bucket in the units of either bits or
638 * 1/1000th of a packet, correspondingly.
639 * Then, when rate is multiplied with milliseconds, we get the
640 * bucket units:
641 * msec * kbps = bits, and
642 * msec * packets/sec = 1/1000 packets.
643 *
644 * 'cost' is the number of bucket units in this packet.
645 */
646 cost = (meter->kbps) ? skb->len * 8 : 1000;
647
648 /* Update all bands and find the one hit with the highest rate. */
649 for (i = 0; i < meter->n_bands; ++i) {
650 long long int max_bucket_size;
651
652 band = &meter->bands[i];
653 max_bucket_size = band->burst_size * 1000LL;
654
655 band->bucket += delta_ms * band->rate;
656 if (band->bucket > max_bucket_size)
657 band->bucket = max_bucket_size;
658
659 if (band->bucket >= cost) {
660 band->bucket -= cost;
661 } else if (band->rate > band_exceeded_rate) {
662 band_exceeded_rate = band->rate;
663 band_exceeded_max = i;
664 }
665 }
666
667 if (band_exceeded_max >= 0) {
668 /* Update band statistics. */
669 band = &meter->bands[band_exceeded_max];
670 band->stats.n_packets += 1;
671 band->stats.n_bytes += skb->len;
672
673 /* Drop band triggered, let the caller drop the 'skb'. */
674 if (band->type == OVS_METER_BAND_TYPE_DROP) {
675 spin_unlock(&meter->lock);
676 return true;
677 }
678 }
679
680 spin_unlock(&meter->lock);
681 return false;
682}
683
684static const struct genl_small_ops dp_meter_genl_ops[] = {
685 { .cmd = OVS_METER_CMD_FEATURES,
686 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
687 .flags = 0, /* OK for unprivileged users. */
688 .doit = ovs_meter_cmd_features
689 },
690 { .cmd = OVS_METER_CMD_SET,
691 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
692 .flags = GENL_UNS_ADMIN_PERM, /* Requires CAP_NET_ADMIN
693 * privilege.
694 */
695 .doit = ovs_meter_cmd_set,
696 },
697 { .cmd = OVS_METER_CMD_GET,
698 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
699 .flags = 0, /* OK for unprivileged users. */
700 .doit = ovs_meter_cmd_get,
701 },
702 { .cmd = OVS_METER_CMD_DEL,
703 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
704 .flags = GENL_UNS_ADMIN_PERM, /* Requires CAP_NET_ADMIN
705 * privilege.
706 */
707 .doit = ovs_meter_cmd_del
708 },
709};
710
711static const struct genl_multicast_group ovs_meter_multicast_group = {
712 .name = OVS_METER_MCGROUP,
713};
714
715struct genl_family dp_meter_genl_family __ro_after_init = {
716 .hdrsize = sizeof(struct ovs_header),
717 .name = OVS_METER_FAMILY,
718 .version = OVS_METER_VERSION,
719 .maxattr = OVS_METER_ATTR_MAX,
720 .policy = meter_policy,
721 .netnsok = true,
722 .parallel_ops = true,
723 .small_ops = dp_meter_genl_ops,
724 .n_small_ops = ARRAY_SIZE(dp_meter_genl_ops),
725 .resv_start_op = OVS_METER_CMD_GET + 1,
726 .mcgrps = &ovs_meter_multicast_group,
727 .n_mcgrps = 1,
728 .module = THIS_MODULE,
729};
730
731int ovs_meters_init(struct datapath *dp)
732{
733 struct dp_meter_table *tbl = &dp->meter_tbl;
734 struct dp_meter_instance *ti;
735 unsigned long free_mem_bytes;
736
737 ti = dp_meter_instance_alloc(DP_METER_ARRAY_SIZE_MIN);
738 if (!ti)
739 return -ENOMEM;
740
741 /* Allow meters in a datapath to use ~3.12% of physical memory. */
742 free_mem_bytes = nr_free_buffer_pages() * (PAGE_SIZE >> 5);
743 tbl->max_meters_allowed = min(free_mem_bytes / sizeof(struct dp_meter),
744 DP_METER_NUM_MAX);
745 if (!tbl->max_meters_allowed)
746 goto out_err;
747
748 rcu_assign_pointer(tbl->ti, ti);
749 tbl->count = 0;
750
751 return 0;
752
753out_err:
754 dp_meter_instance_free(ti);
755 return -ENOMEM;
756}
757
758void ovs_meters_exit(struct datapath *dp)
759{
760 struct dp_meter_table *tbl = &dp->meter_tbl;
761 struct dp_meter_instance *ti = rcu_dereference_raw(tbl->ti);
762 int i;
763
764 for (i = 0; i < ti->n_meters; i++)
765 ovs_meter_free(rcu_dereference_raw(ti->dp_meters[i]));
766
767 dp_meter_instance_free(ti);
768}