]> git.ipfire.org Git - thirdparty/systemd.git/blob - src/network/tc/qdisc.c
network: traffic control: drop meta from QDisc and TClass
[thirdparty/systemd.git] / src / network / tc / qdisc.c
1 /* SPDX-License-Identifier: LGPL-2.1-or-later
2 * Copyright © 2019 VMware, Inc. */
3
4 #include <linux/pkt_sched.h>
5
6 #include "alloc-util.h"
7 #include "conf-parser.h"
8 #include "in-addr-util.h"
9 #include "netlink-util.h"
10 #include "networkd-link.h"
11 #include "networkd-manager.h"
12 #include "networkd-network.h"
13 #include "networkd-queue.h"
14 #include "parse-util.h"
15 #include "qdisc.h"
16 #include "set.h"
17 #include "string-util.h"
18 #include "strv.h"
19 #include "tc-util.h"
20
21 const QDiscVTable * const qdisc_vtable[_QDISC_KIND_MAX] = {
22 [QDISC_KIND_BFIFO] = &bfifo_vtable,
23 [QDISC_KIND_CAKE] = &cake_vtable,
24 [QDISC_KIND_CODEL] = &codel_vtable,
25 [QDISC_KIND_DRR] = &drr_vtable,
26 [QDISC_KIND_ETS] = &ets_vtable,
27 [QDISC_KIND_FQ] = &fq_vtable,
28 [QDISC_KIND_FQ_CODEL] = &fq_codel_vtable,
29 [QDISC_KIND_FQ_PIE] = &fq_pie_vtable,
30 [QDISC_KIND_GRED] = &gred_vtable,
31 [QDISC_KIND_HHF] = &hhf_vtable,
32 [QDISC_KIND_HTB] = &htb_vtable,
33 [QDISC_KIND_NETEM] = &netem_vtable,
34 [QDISC_KIND_PIE] = &pie_vtable,
35 [QDISC_KIND_QFQ] = &qfq_vtable,
36 [QDISC_KIND_PFIFO] = &pfifo_vtable,
37 [QDISC_KIND_PFIFO_FAST] = &pfifo_fast_vtable,
38 [QDISC_KIND_PFIFO_HEAD_DROP] = &pfifo_head_drop_vtable,
39 [QDISC_KIND_SFB] = &sfb_vtable,
40 [QDISC_KIND_SFQ] = &sfq_vtable,
41 [QDISC_KIND_TBF] = &tbf_vtable,
42 [QDISC_KIND_TEQL] = &teql_vtable,
43 };
44
45 static int qdisc_new(QDiscKind kind, QDisc **ret) {
46 _cleanup_(qdisc_freep) QDisc *qdisc = NULL;
47 int r;
48
49 if (kind == _QDISC_KIND_INVALID) {
50 qdisc = new(QDisc, 1);
51 if (!qdisc)
52 return -ENOMEM;
53
54 *qdisc = (QDisc) {
55 .parent = TC_H_ROOT,
56 .kind = kind,
57 };
58 } else {
59 assert(kind >= 0 && kind < _QDISC_KIND_MAX);
60 qdisc = malloc0(qdisc_vtable[kind]->object_size);
61 if (!qdisc)
62 return -ENOMEM;
63
64 qdisc->parent = TC_H_ROOT;
65 qdisc->kind = kind;
66
67 if (QDISC_VTABLE(qdisc)->init) {
68 r = QDISC_VTABLE(qdisc)->init(qdisc);
69 if (r < 0)
70 return r;
71 }
72 }
73
74 *ret = TAKE_PTR(qdisc);
75
76 return 0;
77 }
78
79 int qdisc_new_static(QDiscKind kind, Network *network, const char *filename, unsigned section_line, QDisc **ret) {
80 _cleanup_(config_section_freep) ConfigSection *n = NULL;
81 _cleanup_(qdisc_freep) QDisc *qdisc = NULL;
82 QDisc *existing;
83 int r;
84
85 assert(network);
86 assert(ret);
87 assert(filename);
88 assert(section_line > 0);
89
90 r = config_section_new(filename, section_line, &n);
91 if (r < 0)
92 return r;
93
94 existing = hashmap_get(network->qdiscs_by_section, n);
95 if (existing) {
96 if (existing->kind != _QDISC_KIND_INVALID &&
97 kind != _QDISC_KIND_INVALID &&
98 existing->kind != kind)
99 return -EINVAL;
100
101 if (existing->kind == kind || kind == _QDISC_KIND_INVALID) {
102 *ret = existing;
103 return 0;
104 }
105 }
106
107 r = qdisc_new(kind, &qdisc);
108 if (r < 0)
109 return r;
110
111 if (existing) {
112 qdisc->handle = existing->handle;
113 qdisc->parent = existing->parent;
114 qdisc->tca_kind = TAKE_PTR(existing->tca_kind);
115
116 qdisc_free(existing);
117 }
118
119 qdisc->network = network;
120 qdisc->section = TAKE_PTR(n);
121 qdisc->source = NETWORK_CONFIG_SOURCE_STATIC;
122
123 r = hashmap_ensure_put(&network->qdiscs_by_section, &config_section_hash_ops, qdisc->section, qdisc);
124 if (r < 0)
125 return r;
126
127 *ret = TAKE_PTR(qdisc);
128 return 0;
129 }
130
131 QDisc* qdisc_free(QDisc *qdisc) {
132 if (!qdisc)
133 return NULL;
134
135 if (qdisc->network && qdisc->section)
136 hashmap_remove(qdisc->network->qdiscs_by_section, qdisc->section);
137
138 config_section_free(qdisc->section);
139
140 if (qdisc->link)
141 set_remove(qdisc->link->qdiscs, qdisc);
142
143 free(qdisc->tca_kind);
144 return mfree(qdisc);
145 }
146
147 static const char *qdisc_get_tca_kind(const QDisc *qdisc) {
148 assert(qdisc);
149
150 return (QDISC_VTABLE(qdisc) && QDISC_VTABLE(qdisc)->tca_kind) ?
151 QDISC_VTABLE(qdisc)->tca_kind : qdisc->tca_kind;
152 }
153
154 void qdisc_hash_func(const QDisc *qdisc, struct siphash *state) {
155 assert(qdisc);
156 assert(state);
157
158 siphash24_compress(&qdisc->handle, sizeof(qdisc->handle), state);
159 siphash24_compress(&qdisc->parent, sizeof(qdisc->parent), state);
160 siphash24_compress_string(qdisc_get_tca_kind(qdisc), state);
161 }
162
163 int qdisc_compare_func(const QDisc *a, const QDisc *b) {
164 int r;
165
166 assert(a);
167 assert(b);
168
169 r = CMP(a->handle, b->handle);
170 if (r != 0)
171 return r;
172
173 r = CMP(a->parent, b->parent);
174 if (r != 0)
175 return r;
176
177 return strcmp_ptr(qdisc_get_tca_kind(a), qdisc_get_tca_kind(b));
178 }
179
180 DEFINE_PRIVATE_HASH_OPS_WITH_KEY_DESTRUCTOR(
181 qdisc_hash_ops,
182 QDisc,
183 qdisc_hash_func,
184 qdisc_compare_func,
185 qdisc_free);
186
187 static int qdisc_get(Link *link, const QDisc *in, QDisc **ret) {
188 QDisc *existing;
189
190 assert(link);
191 assert(in);
192
193 existing = set_get(link->qdiscs, in);
194 if (!existing)
195 return -ENOENT;
196
197 if (ret)
198 *ret = existing;
199 return 0;
200 }
201
202 static int qdisc_add(Link *link, QDisc *qdisc) {
203 int r;
204
205 assert(link);
206 assert(qdisc);
207
208 r = set_ensure_put(&link->qdiscs, &qdisc_hash_ops, qdisc);
209 if (r < 0)
210 return r;
211 if (r == 0)
212 return -EEXIST;
213
214 qdisc->link = link;
215 return 0;
216 }
217
218 static int qdisc_dup(const QDisc *src, QDisc **ret) {
219 _cleanup_(qdisc_freep) QDisc *dst = NULL;
220
221 assert(src);
222 assert(ret);
223
224 if (QDISC_VTABLE(src))
225 dst = memdup(src, QDISC_VTABLE(src)->object_size);
226 else
227 dst = newdup(QDisc, src, 1);
228 if (!dst)
229 return -ENOMEM;
230
231 /* clear all pointers */
232 dst->network = NULL;
233 dst->section = NULL;
234 dst->link = NULL;
235 dst->tca_kind = NULL;
236
237 if (src->tca_kind) {
238 dst->tca_kind = strdup(src->tca_kind);
239 if (!dst->tca_kind)
240 return -ENOMEM;
241 }
242
243 *ret = TAKE_PTR(dst);
244 return 0;
245 }
246
247 static void log_qdisc_debug(QDisc *qdisc, Link *link, const char *str) {
248 _cleanup_free_ char *state = NULL;
249
250 assert(qdisc);
251 assert(str);
252
253 if (!DEBUG_LOGGING)
254 return;
255
256 (void) network_config_state_to_string_alloc(qdisc->state, &state);
257
258 log_link_debug(link, "%s %s QDisc (%s): handle=%"PRIx32":%"PRIx32", parent=%"PRIx32":%"PRIx32", kind=%s",
259 str, strna(network_config_source_to_string(qdisc->source)), strna(state),
260 TC_H_MAJ(qdisc->handle) >> 16, TC_H_MIN(qdisc->handle),
261 TC_H_MAJ(qdisc->parent) >> 16, TC_H_MIN(qdisc->parent),
262 strna(qdisc_get_tca_kind(qdisc)));
263 }
264
265 int link_find_qdisc(Link *link, uint32_t handle, uint32_t parent, const char *kind, QDisc **ret) {
266 QDisc *qdisc;
267
268 assert(link);
269
270 handle = TC_H_MAJ(handle);
271
272 SET_FOREACH(qdisc, link->qdiscs) {
273 if (qdisc->handle != handle)
274 continue;
275
276 if (qdisc->parent != parent)
277 continue;
278
279 if (qdisc->source == NETWORK_CONFIG_SOURCE_FOREIGN)
280 continue;
281
282 if (!qdisc_exists(qdisc))
283 continue;
284
285 if (kind && !streq_ptr(kind, qdisc_get_tca_kind(qdisc)))
286 continue;
287
288 if (ret)
289 *ret = qdisc;
290 return 0;
291 }
292
293 return -ENOENT;
294 }
295
296 static int qdisc_handler(sd_netlink *rtnl, sd_netlink_message *m, Link *link) {
297 int r;
298
299 assert(link);
300 assert(link->tc_messages > 0);
301 link->tc_messages--;
302
303 if (IN_SET(link->state, LINK_STATE_FAILED, LINK_STATE_LINGER))
304 return 1;
305
306 r = sd_netlink_message_get_errno(m);
307 if (r < 0 && r != -EEXIST) {
308 log_link_message_error_errno(link, m, r, "Could not set QDisc");
309 link_enter_failed(link);
310 return 1;
311 }
312
313 if (link->tc_messages == 0) {
314 log_link_debug(link, "Traffic control configured");
315 link->tc_configured = true;
316 link_check_ready(link);
317 }
318
319 return 1;
320 }
321
322 static int qdisc_configure(QDisc *qdisc, Link *link) {
323 _cleanup_(sd_netlink_message_unrefp) sd_netlink_message *req = NULL;
324 int r;
325
326 assert(qdisc);
327 assert(link);
328 assert(link->manager);
329 assert(link->manager->rtnl);
330 assert(link->ifindex > 0);
331
332 log_qdisc_debug(qdisc, link, "Configuring");
333
334 r = sd_rtnl_message_new_traffic_control(link->manager->rtnl, &req, RTM_NEWQDISC,
335 link->ifindex, qdisc->handle, qdisc->parent);
336 if (r < 0)
337 return log_link_debug_errno(link, r, "Could not create RTM_NEWQDISC message: %m");
338
339 r = sd_netlink_message_append_string(req, TCA_KIND, qdisc_get_tca_kind(qdisc));
340 if (r < 0)
341 return r;
342
343 if (QDISC_VTABLE(qdisc) && QDISC_VTABLE(qdisc)->fill_message) {
344 r = QDISC_VTABLE(qdisc)->fill_message(link, qdisc, req);
345 if (r < 0)
346 return r;
347 }
348
349 r = netlink_call_async(link->manager->rtnl, NULL, req, qdisc_handler, link_netlink_destroy_callback, link);
350 if (r < 0)
351 return log_link_debug_errno(link, r, "Could not send netlink message: %m");
352
353 link_ref(link);
354 return 0;
355 }
356
357 static bool qdisc_is_ready_to_configure(QDisc *qdisc, Link *link) {
358 assert(qdisc);
359 assert(link);
360
361 if (!IN_SET(link->state, LINK_STATE_CONFIGURING, LINK_STATE_CONFIGURED))
362 return false;
363
364 if (IN_SET(qdisc->parent, TC_H_ROOT, TC_H_CLSACT)) /* TC_H_CLSACT == TC_H_INGRESS */
365 return true;
366
367 return link_find_tclass(link, qdisc->parent, NULL) >= 0;
368 }
369
370 int request_process_qdisc(Request *req) {
371 QDisc *qdisc;
372 Link *link;
373 int r;
374
375 assert(req);
376 assert_se(link = req->link);
377 assert_se(qdisc = req->qdisc);
378
379 if (!qdisc_is_ready_to_configure(qdisc, link))
380 return 0;
381
382 r = qdisc_configure(qdisc, link);
383 if (r < 0)
384 return log_link_warning_errno(link, r, "Failed to configure QDisc: %m");
385
386 qdisc_enter_configuring(qdisc);
387 return 1;
388 }
389
390 int link_request_qdisc(Link *link, QDisc *qdisc) {
391 QDisc *existing;
392 int r;
393
394 assert(link);
395 assert(qdisc);
396
397 if (qdisc_get(link, qdisc, &existing) < 0) {
398 _cleanup_(qdisc_freep) QDisc *tmp = NULL;
399
400 r = qdisc_dup(qdisc, &tmp);
401 if (r < 0)
402 return log_oom();
403
404 r = qdisc_add(link, tmp);
405 if (r < 0)
406 return log_link_warning_errno(link, r, "Failed to store QDisc: %m");
407
408 existing = TAKE_PTR(tmp);
409 } else
410 existing->source = qdisc->source;
411
412 log_qdisc_debug(existing, link, "Requesting");
413 r = link_queue_request(link, REQUEST_TYPE_TC_QDISC, existing, false,
414 &link->tc_messages, NULL, NULL);
415 if (r < 0)
416 return log_link_warning_errno(link, r, "Failed to request QDisc: %m");
417 if (r == 0)
418 return 0;
419
420 qdisc_enter_requesting(existing);
421 return 1;
422 }
423
424 int manager_rtnl_process_qdisc(sd_netlink *rtnl, sd_netlink_message *message, Manager *m) {
425 _cleanup_(qdisc_freep) QDisc *tmp = NULL;
426 QDisc *qdisc = NULL;
427 Link *link;
428 uint16_t type;
429 int ifindex, r;
430
431 assert(rtnl);
432 assert(message);
433 assert(m);
434
435 if (sd_netlink_message_is_error(message)) {
436 r = sd_netlink_message_get_errno(message);
437 if (r < 0)
438 log_message_warning_errno(message, r, "rtnl: failed to receive QDisc message, ignoring");
439
440 return 0;
441 }
442
443 r = sd_netlink_message_get_type(message, &type);
444 if (r < 0) {
445 log_warning_errno(r, "rtnl: could not get message type, ignoring: %m");
446 return 0;
447 } else if (!IN_SET(type, RTM_NEWQDISC, RTM_DELQDISC)) {
448 log_warning("rtnl: received unexpected message type %u when processing QDisc, ignoring.", type);
449 return 0;
450 }
451
452 r = sd_rtnl_message_traffic_control_get_ifindex(message, &ifindex);
453 if (r < 0) {
454 log_warning_errno(r, "rtnl: could not get ifindex from message, ignoring: %m");
455 return 0;
456 } else if (ifindex <= 0) {
457 log_warning("rtnl: received QDisc message with invalid ifindex %d, ignoring.", ifindex);
458 return 0;
459 }
460
461 if (link_get_by_index(m, ifindex, &link) < 0) {
462 if (!m->enumerating)
463 log_warning("rtnl: received QDisc for link '%d' we don't know about, ignoring.", ifindex);
464 return 0;
465 }
466
467 r = qdisc_new(_QDISC_KIND_INVALID, &tmp);
468 if (r < 0)
469 return log_oom();
470
471 r = sd_rtnl_message_traffic_control_get_handle(message, &tmp->handle);
472 if (r < 0) {
473 log_link_warning_errno(link, r, "rtnl: received QDisc message without handle, ignoring: %m");
474 return 0;
475 }
476
477 r = sd_rtnl_message_traffic_control_get_parent(message, &tmp->parent);
478 if (r < 0) {
479 log_link_warning_errno(link, r, "rtnl: received QDisc message without parent, ignoring: %m");
480 return 0;
481 }
482
483 r = sd_netlink_message_read_string_strdup(message, TCA_KIND, &tmp->tca_kind);
484 if (r < 0) {
485 log_link_warning_errno(link, r, "rtnl: received QDisc message without kind, ignoring: %m");
486 return 0;
487 }
488
489 (void) qdisc_get(link, tmp, &qdisc);
490
491 switch (type) {
492 case RTM_NEWQDISC:
493 if (qdisc) {
494 qdisc_enter_configured(qdisc);
495 log_qdisc_debug(qdisc, link, "Received remembered");
496 } else {
497 qdisc_enter_configured(tmp);
498 log_qdisc_debug(tmp, link, "Received new");
499
500 r = qdisc_add(link, tmp);
501 if (r < 0) {
502 log_link_warning_errno(link, r, "Failed to remember QDisc, ignoring: %m");
503 return 0;
504 }
505
506 qdisc = TAKE_PTR(tmp);
507 }
508
509 break;
510
511 case RTM_DELQDISC:
512 if (qdisc) {
513 qdisc_enter_removed(qdisc);
514 if (qdisc->state == 0) {
515 log_qdisc_debug(qdisc, link, "Forgetting");
516 qdisc_free(qdisc);
517 } else
518 log_qdisc_debug(qdisc, link, "Removed");
519 } else
520 log_qdisc_debug(tmp, link, "Kernel removed unknown");
521
522 break;
523
524 default:
525 assert_not_reached();
526 }
527
528 return 1;
529 }
530
531 static int qdisc_section_verify(QDisc *qdisc, bool *has_root, bool *has_clsact) {
532 int r;
533
534 assert(qdisc);
535 assert(has_root);
536 assert(has_clsact);
537
538 if (section_is_invalid(qdisc->section))
539 return -EINVAL;
540
541 if (QDISC_VTABLE(qdisc) && QDISC_VTABLE(qdisc)->verify) {
542 r = QDISC_VTABLE(qdisc)->verify(qdisc);
543 if (r < 0)
544 return r;
545 }
546
547 if (qdisc->parent == TC_H_ROOT) {
548 if (*has_root)
549 return log_warning_errno(SYNTHETIC_ERRNO(EINVAL),
550 "%s: More than one root qdisc section is defined. "
551 "Ignoring the qdisc section from line %u.",
552 qdisc->section->filename, qdisc->section->line);
553 *has_root = true;
554 } else if (qdisc->parent == TC_H_CLSACT) { /* TC_H_CLSACT == TC_H_INGRESS */
555 if (*has_clsact)
556 return log_warning_errno(SYNTHETIC_ERRNO(EINVAL),
557 "%s: More than one clsact or ingress qdisc section is defined. "
558 "Ignoring the qdisc section from line %u.",
559 qdisc->section->filename, qdisc->section->line);
560 *has_clsact = true;
561 }
562
563 return 0;
564 }
565
566 void network_drop_invalid_qdisc(Network *network) {
567 bool has_root = false, has_clsact = false;
568 QDisc *qdisc;
569
570 assert(network);
571
572 HASHMAP_FOREACH(qdisc, network->qdiscs_by_section)
573 if (qdisc_section_verify(qdisc, &has_root, &has_clsact) < 0)
574 qdisc_free(qdisc);
575 }
576
577 int config_parse_qdisc_parent(
578 const char *unit,
579 const char *filename,
580 unsigned line,
581 const char *section,
582 unsigned section_line,
583 const char *lvalue,
584 int ltype,
585 const char *rvalue,
586 void *data,
587 void *userdata) {
588
589 _cleanup_(qdisc_free_or_set_invalidp) QDisc *qdisc = NULL;
590 Network *network = data;
591 int r;
592
593 assert(filename);
594 assert(lvalue);
595 assert(rvalue);
596 assert(data);
597
598 r = qdisc_new_static(ltype, network, filename, section_line, &qdisc);
599 if (r == -ENOMEM)
600 return log_oom();
601 if (r < 0) {
602 log_syntax(unit, LOG_WARNING, filename, line, r,
603 "More than one kind of queueing discipline, ignoring assignment: %m");
604 return 0;
605 }
606
607 if (streq(rvalue, "root"))
608 qdisc->parent = TC_H_ROOT;
609 else if (streq(rvalue, "clsact")) {
610 qdisc->parent = TC_H_CLSACT;
611 qdisc->handle = TC_H_MAKE(TC_H_CLSACT, 0);
612 } else if (streq(rvalue, "ingress")) {
613 qdisc->parent = TC_H_INGRESS;
614 qdisc->handle = TC_H_MAKE(TC_H_INGRESS, 0);
615 } else {
616 r = parse_handle(rvalue, &qdisc->parent);
617 if (r < 0) {
618 log_syntax(unit, LOG_WARNING, filename, line, r,
619 "Failed to parse 'Parent=', ignoring assignment: %s",
620 rvalue);
621 return 0;
622 }
623 }
624
625 if (STR_IN_SET(rvalue, "clsact", "ingress")) {
626 r = free_and_strdup(&qdisc->tca_kind, rvalue);
627 if (r < 0)
628 return log_oom();
629 } else
630 qdisc->tca_kind = mfree(qdisc->tca_kind);
631
632 TAKE_PTR(qdisc);
633
634 return 0;
635 }
636
637 int config_parse_qdisc_handle(
638 const char *unit,
639 const char *filename,
640 unsigned line,
641 const char *section,
642 unsigned section_line,
643 const char *lvalue,
644 int ltype,
645 const char *rvalue,
646 void *data,
647 void *userdata) {
648
649 _cleanup_(qdisc_free_or_set_invalidp) QDisc *qdisc = NULL;
650 Network *network = data;
651 uint16_t n;
652 int r;
653
654 assert(filename);
655 assert(lvalue);
656 assert(rvalue);
657 assert(data);
658
659 r = qdisc_new_static(ltype, network, filename, section_line, &qdisc);
660 if (r == -ENOMEM)
661 return log_oom();
662 if (r < 0) {
663 log_syntax(unit, LOG_WARNING, filename, line, r,
664 "More than one kind of queueing discipline, ignoring assignment: %m");
665 return 0;
666 }
667
668 if (isempty(rvalue)) {
669 qdisc->handle = TC_H_UNSPEC;
670 TAKE_PTR(qdisc);
671 return 0;
672 }
673
674 r = safe_atou16_full(rvalue, 16, &n);
675 if (r < 0) {
676 log_syntax(unit, LOG_WARNING, filename, line, r,
677 "Failed to parse 'Handle=', ignoring assignment: %s",
678 rvalue);
679 return 0;
680 }
681
682 qdisc->handle = (uint32_t) n << 16;
683 TAKE_PTR(qdisc);
684
685 return 0;
686 }