2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2020 Alexander V. Chernikov
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
31 #include "opt_inet6.h"
32 #include "opt_route.h"
34 #include <sys/param.h>
35 #include <sys/eventhandler.h>
36 #include <sys/kernel.h>
39 #include <sys/rmlock.h>
40 #include <sys/malloc.h>
42 #include <sys/module.h>
43 #include <sys/kernel.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/syslog.h>
50 #include <sys/queue.h>
54 #include <net/if_var.h>
56 #include <netinet/in.h>
57 #include <netinet/in_var.h>
58 #include <netinet/ip.h>
59 #include <netinet/ip_var.h>
61 #include <netinet/ip6.h>
62 #include <netinet6/ip6_var.h>
65 #include <net/route.h>
66 #include <net/route/nhop.h>
67 #include <net/route/route_ctl.h>
68 #include <net/route/route_var.h>
69 #include <net/route/fib_algo.h>
71 #include <machine/stdarg.h>
74 * Fib lookup framework.
76 * This framework enables accelerated longest-prefix-match lookups for the
77 * routing tables by adding the ability to dynamically attach/detach lookup
78 * algorithms implementation to/from the datapath.
80 * flm - fib lookup modules - implementation of particular lookup algorithm
81 * fd - fib data - instance of an flm bound to specific routing table
83 * This file provides main framework functionality.
85 * The following are the features provided by the framework
87 * 1) nexhops abstraction -> provides transparent referencing, indexing
88 * and efficient idx->ptr mappings for nexthop and nexthop groups.
89 * 2) Routing table synchronisation
90 * 3) dataplane attachment points
91 * 4) automatic algorithm selection based on the provided preference.
95 * For each supported address family, there is a an allocated array of fib_dp
96 * structures, indexed by fib number. Each array entry contains callback function
97 * and its argument. This function will be called with a family-specific lookup key,
98 * scope and provided argument. This array gets re-created every time when new algo
99 * instance gets created. Please take a look at the replace_rtables_family() function
104 SYSCTL_DECL(_net_route);
105 SYSCTL_NODE(_net_route, OID_AUTO, algo, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
106 "Fib algorithm lookups");
108 /* Algorithm sync policy */
110 /* Time interval to bucket updates */
111 VNET_DEFINE_STATIC(unsigned int, update_bucket_time_ms) = 50;
112 #define V_update_bucket_time_ms VNET(update_bucket_time_ms)
113 SYSCTL_UINT(_net_route_algo, OID_AUTO, bucket_time_ms, CTLFLAG_RW | CTLFLAG_VNET,
114 &VNET_NAME(update_bucket_time_ms), 0, "Time interval to calculate update rate");
116 /* Minimum update rate to delay sync */
117 VNET_DEFINE_STATIC(unsigned int, bucket_change_threshold_rate) = 500;
118 #define V_bucket_change_threshold_rate VNET(bucket_change_threshold_rate)
119 SYSCTL_UINT(_net_route_algo, OID_AUTO, bucket_change_threshold_rate, CTLFLAG_RW | CTLFLAG_VNET,
120 &VNET_NAME(bucket_change_threshold_rate), 0, "Minimum update rate to delay sync");
122 /* Max allowed delay to sync */
123 VNET_DEFINE_STATIC(unsigned int, fib_max_sync_delay_ms) = 1000;
124 #define V_fib_max_sync_delay_ms VNET(fib_max_sync_delay_ms)
125 SYSCTL_UINT(_net_route_algo, OID_AUTO, fib_max_sync_delay_ms, CTLFLAG_RW | CTLFLAG_VNET,
126 &VNET_NAME(fib_max_sync_delay_ms), 0, "Maximum time to delay sync (ms)");
130 VNET_DEFINE_STATIC(bool, algo_fixed_inet6) = false;
131 #define V_algo_fixed_inet6 VNET(algo_fixed_inet6)
132 SYSCTL_NODE(_net_route_algo, OID_AUTO, inet6, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
133 "IPv6 longest prefix match lookups");
136 VNET_DEFINE_STATIC(bool, algo_fixed_inet) = false;
137 #define V_algo_fixed_inet VNET(algo_fixed_inet)
138 SYSCTL_NODE(_net_route_algo, OID_AUTO, inet, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
139 "IPv4 longest prefix match lookups");
142 /* Fib instance counter */
143 static uint32_t fib_gen = 0;
145 struct nhop_ref_table {
150 enum fib_callout_action {
151 FDA_NONE, /* No callout scheduled */
152 FDA_REBUILD, /* Asks to rebuild algo instance */
153 FDA_EVAL, /* Asks to evaluate if the current algo is still be best */
154 FDA_BATCH, /* Asks to submit batch of updates to the algo */
157 struct fib_sync_status {
158 struct timeval diverge_time; /* ts when diverged */
159 uint32_t num_changes; /* number of changes since sync */
160 uint32_t bucket_changes; /* num changes within the current bucket */
161 uint64_t bucket_id; /* 50ms bucket # */
162 struct fib_change_queue fd_change_queue;/* list of scheduled entries */
166 * Data structure for the fib lookup instance tied to the particular rib.
169 uint32_t number_nhops; /* current # of nhops */
170 uint8_t hit_nhops; /* true if out of nhop limit */
171 uint8_t init_done; /* true if init is competed */
172 uint32_t fd_dead:1; /* Scheduled for deletion */
173 uint32_t fd_linked:1; /* true if linked */
174 uint32_t fd_need_rebuild:1; /* true if rebuild scheduled */
175 uint32_t fd_batch:1; /* true if batched notification scheduled */
176 uint8_t fd_family; /* family */
177 uint32_t fd_fibnum; /* fibnum */
178 uint32_t fd_failed_rebuilds; /* stat: failed rebuilds */
179 uint32_t fd_gen; /* instance gen# */
180 struct callout fd_callout; /* rebuild callout */
181 enum fib_callout_action fd_callout_action; /* Callout action to take */
182 void *fd_algo_data; /* algorithm data */
183 struct nhop_object **nh_idx; /* nhop idx->ptr array */
184 struct nhop_ref_table *nh_ref_table; /* array with # of nhop references */
185 struct rib_head *fd_rh; /* RIB table we're attached to */
186 struct rib_subscription *fd_rs; /* storing table subscription */
187 struct fib_dp fd_dp; /* fib datapath data */
188 struct vnet *fd_vnet; /* vnet fib belongs to */
189 struct epoch_context fd_epoch_ctx; /* epoch context for deletion */
190 struct fib_lookup_module *fd_flm;/* pointer to the lookup module */
191 struct fib_sync_status fd_ss; /* State relevant to the rib sync */
192 uint32_t fd_num_changes; /* number of changes since last callout */
193 TAILQ_ENTRY(fib_data) entries; /* list of all fds in vnet */
196 static bool rebuild_fd(struct fib_data *fd, const char *reason);
197 static bool rebuild_fd_flm(struct fib_data *fd, struct fib_lookup_module *flm_new);
198 static void handle_fd_callout(void *_data);
199 static void destroy_fd_instance_epoch(epoch_context_t ctx);
200 static bool is_idx_free(struct fib_data *fd, uint32_t index);
201 static void set_algo_fixed(struct rib_head *rh);
202 static bool is_algo_fixed(struct rib_head *rh);
204 static uint32_t fib_ref_nhop(struct fib_data *fd, struct nhop_object *nh);
205 static void fib_unref_nhop(struct fib_data *fd, struct nhop_object *nh);
207 static struct fib_lookup_module *fib_check_best_algo(struct rib_head *rh,
208 struct fib_lookup_module *orig_flm);
209 static void fib_unref_algo(struct fib_lookup_module *flm);
210 static bool flm_error_check(const struct fib_lookup_module *flm, uint32_t fibnum);
213 #define FIB_MOD_LOCK() mtx_lock(&fib_mtx)
214 #define FIB_MOD_UNLOCK() mtx_unlock(&fib_mtx)
215 #define FIB_MOD_LOCK_ASSERT() mtx_assert(&fib_mtx, MA_OWNED)
217 MTX_SYSINIT(fib_mtx, &fib_mtx, "algo list mutex", MTX_DEF);
219 /* Algorithm has to be this percent better than the current to switch */
220 #define BEST_DIFF_PERCENT (5 * 256 / 100)
221 /* Schedule algo re-evaluation X seconds after a change */
222 #define ALGO_EVAL_DELAY_MS 30000
223 /* Force algo re-evaluation after X changes */
224 #define ALGO_EVAL_NUM_ROUTES 100
225 /* Try to setup algorithm X times */
226 #define FIB_MAX_TRIES 32
227 /* Max amount of supported nexthops */
228 #define FIB_MAX_NHOPS 262144
229 #define FIB_CALLOUT_DELAY_MS 50
233 static int flm_debug_level = LOG_NOTICE;
234 SYSCTL_INT(_net_route_algo, OID_AUTO, debug_level, CTLFLAG_RW | CTLFLAG_RWTUN,
235 &flm_debug_level, 0, "debuglevel");
236 #define FLM_MAX_DEBUG_LEVEL LOG_DEBUG
241 #define _PASS_MSG(_l) (flm_debug_level >= (_l))
242 #define ALGO_PRINTF(_l, _fmt, ...) if (_PASS_MSG(_l)) { \
243 printf("[fib_algo] %s: " _fmt "\n", __func__, ##__VA_ARGS__); \
245 #define _ALGO_PRINTF(_fib, _fam, _aname, _gen, _func, _fmt, ...) \
246 printf("[fib_algo] %s.%u (%s#%u) %s: " _fmt "\n",\
247 print_family(_fam), _fib, _aname, _gen, _func, ## __VA_ARGS__)
248 #define _RH_PRINTF(_fib, _fam, _func, _fmt, ...) \
249 printf("[fib_algo] %s.%u %s: " _fmt "\n", print_family(_fam), _fib, _func, ## __VA_ARGS__)
250 #define RH_PRINTF(_l, _rh, _fmt, ...) if (_PASS_MSG(_l)) { \
251 _RH_PRINTF(_rh->rib_fibnum, _rh->rib_family, __func__, _fmt, ## __VA_ARGS__);\
253 #define FD_PRINTF(_l, _fd, _fmt, ...) FD_PRINTF_##_l(_l, _fd, _fmt, ## __VA_ARGS__)
254 #define _FD_PRINTF(_l, _fd, _fmt, ...) if (_PASS_MSG(_l)) { \
255 _ALGO_PRINTF(_fd->fd_fibnum, _fd->fd_family, _fd->fd_flm->flm_name, \
256 _fd->fd_gen, __func__, _fmt, ## __VA_ARGS__); \
258 #if FLM_MAX_DEBUG_LEVEL>=LOG_DEBUG2
259 #define FD_PRINTF_LOG_DEBUG2 _FD_PRINTF
261 #define FD_PRINTF_LOG_DEBUG2(_l, _fd, _fmt, ...)
263 #if FLM_MAX_DEBUG_LEVEL>=LOG_DEBUG
264 #define FD_PRINTF_LOG_DEBUG _FD_PRINTF
266 #define FD_PRINTF_LOG_DEBUG()
268 #if FLM_MAX_DEBUG_LEVEL>=LOG_INFO
269 #define FD_PRINTF_LOG_INFO _FD_PRINTF
271 #define FD_PRINTF_LOG_INFO()
273 #define FD_PRINTF_LOG_NOTICE _FD_PRINTF
274 #define FD_PRINTF_LOG_ERR _FD_PRINTF
275 #define FD_PRINTF_LOG_WARNING _FD_PRINTF
278 /* List of all registered lookup algorithms */
279 static TAILQ_HEAD(, fib_lookup_module) all_algo_list = TAILQ_HEAD_INITIALIZER(all_algo_list);
281 /* List of all fib lookup instances in the vnet */
282 VNET_DEFINE_STATIC(TAILQ_HEAD(fib_data_head, fib_data), fib_data_list);
283 #define V_fib_data_list VNET(fib_data_list)
285 /* Datastructure for storing non-transient fib lookup module failures */
288 uint32_t fe_fibnum; /* failed rtable */
289 struct fib_lookup_module *fe_flm; /* failed module */
290 TAILQ_ENTRY(fib_error) entries;/* list of all errored entries */
292 VNET_DEFINE_STATIC(TAILQ_HEAD(fib_error_head, fib_error), fib_error_list);
293 #define V_fib_error_list VNET(fib_error_list)
295 /* Per-family array of fibnum -> {func, arg} mappings used in datapath */
296 struct fib_dp_header {
297 struct epoch_context fdh_epoch_ctx;
298 uint32_t fdh_num_tables;
299 struct fib_dp fdh_idx[0];
303 * Tries to add new non-transient algorithm error to the list of
305 * Returns true on success.
308 flm_error_add(struct fib_lookup_module *flm, uint32_t fibnum)
310 struct fib_error *fe;
312 fe = malloc(sizeof(struct fib_error), M_TEMP, M_NOWAIT | M_ZERO);
316 fe->fe_family = flm->flm_family;
317 fe->fe_fibnum = fibnum;
320 /* Avoid duplicates by checking if error already exists first */
321 if (flm_error_check(flm, fibnum)) {
326 TAILQ_INSERT_HEAD(&V_fib_error_list, fe, entries);
333 * True if non-transient error has been registered for @flm in @fibnum.
336 flm_error_check(const struct fib_lookup_module *flm, uint32_t fibnum)
338 const struct fib_error *fe;
340 TAILQ_FOREACH(fe, &V_fib_error_list, entries) {
341 if ((fe->fe_flm == flm) && (fe->fe_fibnum == fibnum))
349 * Clear all errors of algo specified by @flm.
352 fib_error_clear_flm(struct fib_lookup_module *flm)
354 struct fib_error *fe, *fe_tmp;
356 FIB_MOD_LOCK_ASSERT();
358 TAILQ_FOREACH_SAFE(fe, &V_fib_error_list, entries, fe_tmp) {
359 if (fe->fe_flm == flm) {
360 TAILQ_REMOVE(&V_fib_error_list, fe, entries);
367 * Clears all errors in current VNET.
370 fib_error_clear(void)
372 struct fib_error *fe, *fe_tmp;
374 FIB_MOD_LOCK_ASSERT();
376 TAILQ_FOREACH_SAFE(fe, &V_fib_error_list, entries, fe_tmp) {
377 TAILQ_REMOVE(&V_fib_error_list, fe, entries);
383 print_op_result(enum flm_op_result result)
400 print_family(int family)
403 if (family == AF_INET)
405 else if (family == AF_INET6)
412 * Debug function used by lookup algorithms.
413 * Outputs message denoted by @fmt, prepended by "[fib_algo] inetX.Y (algo) "
416 fib_printf(int level, struct fib_data *fd, const char *func, char *fmt, ...)
421 if (level > flm_debug_level)
425 vsnprintf(buf, sizeof(buf), fmt, ap);
428 _ALGO_PRINTF(fd->fd_fibnum, fd->fd_family, fd->fd_flm->flm_name,
429 fd->fd_gen, func, "%s", buf);
433 * Outputs list of algorithms supported by the provided address family.
436 print_algos_sysctl(struct sysctl_req *req, int family)
438 struct fib_lookup_module *flm;
440 int error, count = 0;
442 error = sysctl_wire_old_buffer(req, 0);
444 sbuf_new_for_sysctl(&sbuf, NULL, 512, req);
445 TAILQ_FOREACH(flm, &all_algo_list, entries) {
446 if (flm->flm_family == family) {
448 sbuf_cat(&sbuf, ", ");
449 sbuf_cat(&sbuf, flm->flm_name);
452 error = sbuf_finish(&sbuf);
460 print_algos_sysctl_inet6(SYSCTL_HANDLER_ARGS)
463 return (print_algos_sysctl(req, AF_INET6));
465 SYSCTL_PROC(_net_route_algo_inet6, OID_AUTO, algo_list,
466 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
467 print_algos_sysctl_inet6, "A", "List of IPv6 lookup algorithms");
472 print_algos_sysctl_inet(SYSCTL_HANDLER_ARGS)
475 return (print_algos_sysctl(req, AF_INET));
477 SYSCTL_PROC(_net_route_algo_inet, OID_AUTO, algo_list,
478 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
479 print_algos_sysctl_inet, "A", "List of IPv4 lookup algorithms");
483 * Calculate delay between repeated failures.
484 * Returns current delay in milliseconds.
487 callout_calc_delay_ms(struct fib_data *fd)
491 if (fd->fd_failed_rebuilds > 10)
494 shift = fd->fd_failed_rebuilds;
496 return ((1 << shift) * FIB_CALLOUT_DELAY_MS);
500 schedule_callout(struct fib_data *fd, enum fib_callout_action action, int delay_ms)
503 FD_PRINTF(LOG_DEBUG, fd, "delay=%d action=%d", delay_ms, action);
504 fd->fd_callout_action = action;
505 callout_reset_sbt(&fd->fd_callout, SBT_1MS * delay_ms, 0,
506 handle_fd_callout, fd, 0);
510 schedule_fd_rebuild(struct fib_data *fd, const char *reason)
513 RIB_WLOCK_ASSERT(fd->fd_rh);
515 if (!fd->fd_need_rebuild) {
516 fd->fd_need_rebuild = true;
517 /* Stop batch updates */
518 fd->fd_batch = false;
521 * Potentially re-schedules pending callout
522 * initiated by schedule_algo_eval.
524 FD_PRINTF(LOG_INFO, fd, "Scheduling rebuild: %s (failures=%d)",
525 reason, fd->fd_failed_rebuilds);
526 schedule_callout(fd, FDA_REBUILD, callout_calc_delay_ms(fd));
531 sync_rib_gen(struct fib_data *fd)
533 FD_PRINTF(LOG_DEBUG, fd, "Sync gen %u -> %u", fd->fd_rh->rnh_gen, fd->fd_rh->rnh_gen_rib);
534 fd->fd_rh->rnh_gen = fd->fd_rh->rnh_gen_rib;
538 get_tv_diff_ms(const struct timeval *old_tv, const struct timeval *new_tv)
542 diff = ((int64_t)(new_tv->tv_sec - old_tv->tv_sec)) * 1000;
543 diff += (new_tv->tv_usec - old_tv->tv_usec) / 1000;
549 add_tv_diff_ms(struct timeval *tv, int ms)
551 tv->tv_sec += ms / 1000;
553 if (ms * 1000 + tv->tv_usec < 1000000)
554 tv->tv_usec += ms * 1000;
557 tv->tv_usec = ms * 1000 + tv->tv_usec - 1000000;
562 * Marks the time when algo state diverges from the rib state.
565 mark_diverge_time(struct fib_data *fd)
567 struct fib_sync_status *fd_ss = &fd->fd_ss;
569 getmicrouptime(&fd_ss->diverge_time);
570 fd_ss->bucket_id = 0;
571 fd_ss->bucket_changes = 0;
575 * Calculates and updates the next algorithm sync time, based on the current activity.
577 * The intent is to provide reasonable balance between the update
578 * latency and efficient batching when changing large amount of routes.
580 * High-level algorithm looks the following:
581 * 1) all changes are bucketed in 50ms intervals
582 * 2) If amount of changes within the bucket is greater than the threshold,
583 * the update gets delayed, up to maximum delay threshold.
586 update_rebuild_delay(struct fib_data *fd, enum fib_callout_action action)
588 uint32_t bucket_id, new_delay = 0;
591 /* Fetch all variables at once to ensure consistent reads */
592 uint32_t bucket_time_ms = V_update_bucket_time_ms;
593 uint32_t threshold_rate = V_bucket_change_threshold_rate;
594 uint32_t max_delay_ms = V_fib_max_sync_delay_ms;
596 if (bucket_time_ms == 0)
598 /* calculate per-bucket threshold rate */
599 threshold_rate = threshold_rate * bucket_time_ms / 1000;
603 struct fib_sync_status *fd_ss = &fd->fd_ss;
605 bucket_id = get_tv_diff_ms(&fd_ss->diverge_time, &tv) / bucket_time_ms;
607 if (fd_ss->bucket_id == bucket_id) {
608 fd_ss->bucket_changes++;
609 if (fd_ss->bucket_changes == threshold_rate) {
610 new_delay = (bucket_id + 2) * bucket_time_ms;
611 if (new_delay <= max_delay_ms) {
612 FD_PRINTF(LOG_DEBUG, fd,
613 "hit threshold of %u routes, delay update,"
614 "bucket: %u, total delay: %u",
615 threshold_rate, bucket_id + 1, new_delay);
618 FD_PRINTF(LOG_DEBUG, fd,
619 "maximum sync delay (%u ms) reached", max_delay_ms);
621 } else if ((bucket_id == 0) && (fd_ss->bucket_changes == 1))
622 new_delay = bucket_time_ms;
624 fd_ss->bucket_id = bucket_id;
625 fd_ss->bucket_changes = 1;
629 /* Calculated time has been updated */
630 struct timeval new_tv = fd_ss->diverge_time;
631 add_tv_diff_ms(&new_tv, new_delay);
633 int32_t delay_ms = get_tv_diff_ms(&tv, &new_tv);
634 schedule_callout(fd, action, delay_ms);
639 update_algo_state(struct fib_data *fd)
642 RIB_WLOCK_ASSERT(fd->fd_rh);
644 if (fd->fd_batch || fd->fd_need_rebuild) {
645 enum fib_callout_action action = fd->fd_need_rebuild ? FDA_REBUILD : FDA_BATCH;
646 update_rebuild_delay(fd, action);
650 if (fd->fd_num_changes++ == 0) {
651 /* Start callout to consider switch */
652 if (!callout_pending(&fd->fd_callout))
653 schedule_callout(fd, FDA_EVAL, ALGO_EVAL_DELAY_MS);
654 } else if (fd->fd_num_changes == ALGO_EVAL_NUM_ROUTES) {
655 /* Reset callout to exec immediately */
656 if (fd->fd_callout_action == FDA_EVAL)
657 schedule_callout(fd, FDA_EVAL, 1);
662 need_immediate_sync(struct fib_data *fd, struct rib_cmd_info *rc)
664 struct nhop_object *nh;
666 /* Sync addition/removal of interface routes */
667 switch (rc->rc_cmd) {
670 if (!NH_IS_NHGRP(nh)) {
671 if (!(nh->nh_flags & NHF_GATEWAY))
673 if (nhop_get_rtflags(nh) & RTF_STATIC)
679 if (!NH_IS_NHGRP(nh)) {
680 if (!(nh->nh_flags & NHF_GATEWAY))
682 if (nhop_get_rtflags(nh) & RTF_STATIC)
692 apply_rtable_changes(struct fib_data *fd)
694 enum flm_op_result result;
695 struct fib_change_queue *q = &fd->fd_ss.fd_change_queue;
697 result = fd->fd_flm->flm_change_rib_items_cb(fd->fd_rh, q, fd->fd_algo_data);
699 if (result == FLM_SUCCESS) {
701 for (int i = 0; i < q->count; i++)
702 if (q->entries[i].nh_old)
703 fib_unref_nhop(fd, q->entries[i].nh_old);
706 fd->fd_batch = false;
708 return (result == FLM_SUCCESS);
712 fill_change_entry(struct fib_data *fd, struct fib_change_entry *ce, struct rib_cmd_info *rc)
716 switch (fd->fd_family) {
719 rt_get_inet_prefix_plen(rc->rc_rt, &ce->addr4, &plen, &ce->scopeid);
724 rt_get_inet6_prefix_plen(rc->rc_rt, &ce->addr6, &plen, &ce->scopeid);
730 ce->nh_old = rc->rc_nh_old;
731 ce->nh_new = rc->rc_nh_new;
732 if (ce->nh_new != NULL) {
733 if (fib_ref_nhop(fd, ce->nh_new) == 0)
741 queue_rtable_change(struct fib_data *fd, struct rib_cmd_info *rc)
743 struct fib_change_queue *q = &fd->fd_ss.fd_change_queue;
745 if (q->count >= q->size) {
749 q_size = 256; /* ~18k memory */
751 q_size = q->size * 2;
753 size_t size = q_size * sizeof(struct fib_change_entry);
754 void *a = realloc(q->entries, size, M_TEMP, M_NOWAIT | M_ZERO);
756 FD_PRINTF(LOG_INFO, fd, "Unable to realloc queue for %u elements",
764 return (fill_change_entry(fd, &q->entries[q->count++], rc));
768 * Rib subscription handler. Checks if the algorithm is ready to
769 * receive updates, handles nexthop refcounting and passes change
770 * data to the algorithm callback.
773 handle_rtable_change_cb(struct rib_head *rnh, struct rib_cmd_info *rc,
776 struct fib_data *fd = (struct fib_data *)_data;
777 enum flm_op_result result;
779 RIB_WLOCK_ASSERT(rnh);
782 * There is a small gap between subscribing for route changes
783 * and initiating rtable dump. Avoid receiving route changes
784 * prior to finishing rtable dump by checking `init_done`.
789 bool immediate_sync = need_immediate_sync(fd, rc);
791 /* Consider scheduling algorithm re-evaluation */
792 update_algo_state(fd);
795 * If algo requested rebuild, stop sending updates by default.
796 * This simplifies nexthop refcount handling logic.
798 if (fd->fd_need_rebuild) {
800 rebuild_fd(fd, "rtable change type enforced sync");
805 * Algo requested updates to be delivered in batches.
806 * Add the current change to the queue and return.
809 if (immediate_sync) {
810 if (!queue_rtable_change(fd, rc) || !apply_rtable_changes(fd))
811 rebuild_fd(fd, "batch sync failed");
813 if (!queue_rtable_change(fd, rc))
814 schedule_fd_rebuild(fd, "batch queue failed");
820 * Maintain guarantee that every nexthop returned by the dataplane
821 * lookup has > 0 refcount, so can be safely referenced within current
824 if (rc->rc_nh_new != NULL) {
825 if (fib_ref_nhop(fd, rc->rc_nh_new) == 0) {
826 /* ran out of indexes */
827 schedule_fd_rebuild(fd, "ran out of nhop indexes");
832 result = fd->fd_flm->flm_change_rib_item_cb(rnh, rc, fd->fd_algo_data);
837 /* Unref old nexthop on success */
838 if (rc->rc_nh_old != NULL)
839 fib_unref_nhop(fd, rc->rc_nh_old);
844 * Algo asks to batch the changes.
846 if (queue_rtable_change(fd, rc)) {
847 if (!immediate_sync) {
849 mark_diverge_time(fd);
850 update_rebuild_delay(fd, FDA_BATCH);
853 if (apply_rtable_changes(fd))
856 FD_PRINTF(LOG_ERR, fd, "batched sync failed, force the rebuild");
861 * Algo is not able to apply the update.
862 * Schedule algo rebuild.
864 if (!immediate_sync) {
865 mark_diverge_time(fd);
866 schedule_fd_rebuild(fd, "algo requested rebuild");
870 FD_PRINTF(LOG_INFO, fd, "running sync rebuild");
871 rebuild_fd(fd, "rtable change type enforced sync");
876 * Algo reported a non-recoverable error.
877 * Record the error and schedule rebuild, which will
878 * trigger best algo selection.
880 FD_PRINTF(LOG_ERR, fd, "algo reported non-recoverable error");
881 if (!flm_error_add(fd->fd_flm, fd->fd_fibnum))
882 FD_PRINTF(LOG_ERR, fd, "failed to ban algo");
883 schedule_fd_rebuild(fd, "algo reported non-recoverable error");
888 estimate_nhop_scale(const struct fib_data *old_fd, struct fib_data *fd)
891 if (old_fd == NULL) {
892 // TODO: read from rtable
893 fd->number_nhops = 16;
897 if (old_fd->hit_nhops && old_fd->number_nhops < FIB_MAX_NHOPS)
898 fd->number_nhops = 2 * old_fd->number_nhops;
900 fd->number_nhops = old_fd->number_nhops;
906 enum flm_op_result result;
910 * Handler called after all rtenties have been dumped.
911 * Performs post-dump framework checks and calls
912 * algo:flm_dump_end_cb().
914 * Updates walk_cbdata result.
917 sync_algo_end_cb(struct rib_head *rnh, enum rib_walk_hook stage, void *_data)
919 struct walk_cbdata *w = (struct walk_cbdata *)_data;
920 struct fib_data *fd = w->fd;
922 RIB_WLOCK_ASSERT(w->fd->fd_rh);
924 if (rnh->rib_dying) {
925 w->result = FLM_ERROR;
930 FD_PRINTF(LOG_INFO, fd, "ran out of nexthops at %u nhops",
931 fd->nh_ref_table->count);
932 if (w->result == FLM_SUCCESS)
933 w->result = FLM_REBUILD;
937 if (stage != RIB_WALK_HOOK_POST || w->result != FLM_SUCCESS)
940 /* Post-dump hook, dump successful */
941 w->result = fd->fd_flm->flm_dump_end_cb(fd->fd_algo_data, &fd->fd_dp);
943 if (w->result == FLM_SUCCESS) {
944 /* Mark init as done to allow routing updates */
950 * Callback for each entry in rib.
951 * Calls algo:flm_dump_rib_item_cb func as a part of initial
952 * route table synchronisation.
955 sync_algo_cb(struct rtentry *rt, void *_data)
957 struct walk_cbdata *w = (struct walk_cbdata *)_data;
959 RIB_WLOCK_ASSERT(w->fd->fd_rh);
961 if (w->result == FLM_SUCCESS && w->func) {
964 * Reference nexthops to maintain guarantee that
965 * each nexthop returned by datapath has > 0 references
966 * and can be safely referenced within current epoch.
968 struct nhop_object *nh = rt_get_raw_nhop(rt);
969 if (fib_ref_nhop(w->fd, nh) != 0)
970 w->result = w->func(rt, w->fd->fd_algo_data);
972 w->result = FLM_REBUILD;
979 * Dump all routing table state to the algo instance.
981 static enum flm_op_result
982 sync_algo(struct fib_data *fd)
984 struct walk_cbdata w = {
986 .func = fd->fd_flm->flm_dump_rib_item_cb,
987 .result = FLM_SUCCESS,
990 rib_walk_ext_locked(fd->fd_rh, sync_algo_cb, sync_algo_end_cb, &w);
992 FD_PRINTF(LOG_INFO, fd,
993 "initial dump completed (rtable version: %d), result: %s",
994 fd->fd_rh->rnh_gen, print_op_result(w.result));
1000 * Schedules epoch-backed @fd instance deletion.
1001 * * Unlinks @fd from the list of active algo instances.
1002 * * Removes rib subscription.
1004 * * Schedules actual deletion.
1006 * Assume @fd is already unlinked from the datapath.
1009 schedule_destroy_fd_instance(struct fib_data *fd, bool in_callout)
1014 RIB_WLOCK_ASSERT(fd->fd_rh);
1017 is_dead = fd->fd_dead;
1020 if (fd->fd_linked) {
1021 TAILQ_REMOVE(&V_fib_data_list, fd, entries);
1022 fd->fd_linked = false;
1028 FD_PRINTF(LOG_INFO, fd, "DETACH");
1030 if (fd->fd_rs != NULL)
1031 rib_unsubscribe_locked(fd->fd_rs);
1034 * After rib_unsubscribe() no _new_ handle_rtable_change_cb() calls
1035 * will be executed, hence no _new_ callout schedules will happen.
1037 callout_stop(&fd->fd_callout);
1039 fib_epoch_call(destroy_fd_instance_epoch, &fd->fd_epoch_ctx);
1045 * Wipe all fd instances from the list matching rib specified by @rh.
1046 * If @keep_first is set, remove all but the first record.
1049 fib_cleanup_algo(struct rib_head *rh, bool keep_first, bool in_callout)
1051 struct fib_data_head tmp_head = TAILQ_HEAD_INITIALIZER(tmp_head);
1052 struct fib_data *fd, *fd_tmp;
1053 struct epoch_tracker et;
1056 TAILQ_FOREACH_SAFE(fd, &V_fib_data_list, entries, fd_tmp) {
1057 if (fd->fd_rh == rh) {
1062 TAILQ_REMOVE(&V_fib_data_list, fd, entries);
1063 fd->fd_linked = false;
1064 TAILQ_INSERT_TAIL(&tmp_head, fd, entries);
1069 /* Pass 2: remove each entry */
1070 NET_EPOCH_ENTER(et);
1071 TAILQ_FOREACH_SAFE(fd, &tmp_head, entries, fd_tmp) {
1073 RIB_WLOCK(fd->fd_rh);
1074 schedule_destroy_fd_instance(fd, in_callout);
1076 RIB_WUNLOCK(fd->fd_rh);
1082 fib_destroy_rib(struct rib_head *rh)
1086 * rnh has `is_dying` flag set, so setup of new fd's will fail at
1087 * sync_algo() stage, preventing new entries to be added to the list
1088 * of active algos. Remove all existing entries for the particular rib.
1090 fib_cleanup_algo(rh, false, false);
1094 * Finalises fd destruction by freeing all fd resources.
1097 destroy_fd_instance(struct fib_data *fd)
1100 FD_PRINTF(LOG_INFO, fd, "destroy fd %p", fd);
1102 /* Call destroy callback first */
1103 if (fd->fd_algo_data != NULL)
1104 fd->fd_flm->flm_destroy_cb(fd->fd_algo_data);
1107 if ((fd->nh_idx != NULL) && (fd->nh_ref_table != NULL)) {
1108 for (int i = 0; i < fd->number_nhops; i++) {
1109 if (!is_idx_free(fd, i)) {
1110 FD_PRINTF(LOG_DEBUG2, fd, " FREE nhop %d %p",
1112 nhop_free_any(fd->nh_idx[i]);
1115 free(fd->nh_idx, M_RTABLE);
1117 if (fd->nh_ref_table != NULL)
1118 free(fd->nh_ref_table, M_RTABLE);
1120 if (fd->fd_ss.fd_change_queue.entries != NULL)
1121 free(fd->fd_ss.fd_change_queue.entries, M_TEMP);
1123 fib_unref_algo(fd->fd_flm);
1129 * Epoch callback indicating fd is safe to destroy
1132 destroy_fd_instance_epoch(epoch_context_t ctx)
1134 struct fib_data *fd;
1136 fd = __containerof(ctx, struct fib_data, fd_epoch_ctx);
1138 destroy_fd_instance(fd);
1142 * Tries to setup fd instance.
1143 * - Allocates fd/nhop table
1144 * - Runs algo:flm_init_cb algo init
1145 * - Subscribes fd to the rib
1146 * - Runs rtable dump
1147 * - Adds instance to the list of active instances.
1149 * Returns: operation result. Fills in @pfd with resulting fd on success.
1152 static enum flm_op_result
1153 try_setup_fd_instance(struct fib_lookup_module *flm, struct rib_head *rh,
1154 struct fib_data *old_fd, struct fib_data **pfd)
1156 struct fib_data *fd;
1158 enum flm_op_result result;
1161 fd = malloc(sizeof(struct fib_data), M_RTABLE, M_NOWAIT | M_ZERO);
1164 RH_PRINTF(LOG_INFO, rh, "Unable to allocate fib_data structure");
1165 return (FLM_REBUILD);
1169 estimate_nhop_scale(old_fd, fd);
1172 fd->fd_family = rh->rib_family;
1173 fd->fd_fibnum = rh->rib_fibnum;
1174 callout_init_rm(&fd->fd_callout, &rh->rib_lock, 0);
1175 fd->fd_vnet = curvnet;
1179 flm->flm_refcount++;
1180 fd->fd_gen = ++fib_gen;
1183 FD_PRINTF(LOG_DEBUG, fd, "allocated fd %p", fd);
1185 /* Allocate nhidx -> nhop_ptr table */
1186 size = fd->number_nhops * sizeof(void *);
1187 fd->nh_idx = malloc(size, M_RTABLE, M_NOWAIT | M_ZERO);
1188 if (fd->nh_idx == NULL) {
1189 FD_PRINTF(LOG_INFO, fd, "Unable to allocate nhop table idx (sz:%zu)", size);
1190 return (FLM_REBUILD);
1193 /* Allocate nhop index refcount table */
1194 size = sizeof(struct nhop_ref_table);
1195 size += fd->number_nhops * sizeof(uint32_t);
1196 fd->nh_ref_table = malloc(size, M_RTABLE, M_NOWAIT | M_ZERO);
1197 if (fd->nh_ref_table == NULL) {
1198 FD_PRINTF(LOG_INFO, fd, "Unable to allocate nhop refcount table (sz:%zu)", size);
1199 return (FLM_REBUILD);
1201 FD_PRINTF(LOG_DEBUG, fd, "Allocated %u nhop indexes", fd->number_nhops);
1203 /* Okay, we're ready for algo init */
1204 void *old_algo_data = (old_fd != NULL) ? old_fd->fd_algo_data : NULL;
1205 result = flm->flm_init_cb(fd->fd_fibnum, fd, old_algo_data, &fd->fd_algo_data);
1206 if (result != FLM_SUCCESS) {
1207 FD_PRINTF(LOG_INFO, fd, "%s algo init failed", flm->flm_name);
1211 /* Try to subscribe */
1212 if (flm->flm_change_rib_item_cb != NULL) {
1213 fd->fd_rs = rib_subscribe_locked(fd->fd_rh,
1214 handle_rtable_change_cb, fd, RIB_NOTIFY_IMMEDIATE);
1215 if (fd->fd_rs == NULL) {
1216 FD_PRINTF(LOG_INFO, fd, "failed to subscribe to the rib changes");
1217 return (FLM_REBUILD);
1222 result = sync_algo(fd);
1223 if (result != FLM_SUCCESS) {
1224 FD_PRINTF(LOG_INFO, fd, "rib sync failed");
1227 FD_PRINTF(LOG_INFO, fd, "DUMP completed successfully.");
1231 * Insert fd in the beginning of a list, to maintain invariant
1232 * that first matching entry for the AF/fib is always the active
1235 TAILQ_INSERT_HEAD(&V_fib_data_list, fd, entries);
1236 fd->fd_linked = true;
1239 return (FLM_SUCCESS);
1243 * Sets up algo @flm for table @rh and links it to the datapath.
1246 static enum flm_op_result
1247 setup_fd_instance(struct fib_lookup_module *flm, struct rib_head *rh,
1248 struct fib_data *orig_fd, struct fib_data **pfd, bool attach)
1250 struct fib_data *prev_fd, *new_fd;
1251 enum flm_op_result result;
1254 RIB_WLOCK_ASSERT(rh);
1258 for (int i = 0; i < FIB_MAX_TRIES; i++) {
1259 result = try_setup_fd_instance(flm, rh, prev_fd, &new_fd);
1261 if ((result == FLM_SUCCESS) && attach) {
1262 if (fib_set_datapath_ptr(new_fd, &new_fd->fd_dp))
1263 sync_rib_gen(new_fd);
1265 result = FLM_REBUILD;
1268 if ((prev_fd != NULL) && (prev_fd != orig_fd)) {
1269 schedule_destroy_fd_instance(prev_fd, false);
1273 RH_PRINTF(LOG_INFO, rh, "try %d: fib algo result: %s", i,
1274 print_op_result(result));
1276 if (result == FLM_REBUILD) {
1285 if (result != FLM_SUCCESS) {
1286 RH_PRINTF(LOG_WARNING, rh,
1287 "%s algo instance setup failed, failures=%d", flm->flm_name,
1288 orig_fd ? orig_fd->fd_failed_rebuilds + 1 : 0);
1289 /* update failure count */
1291 if (orig_fd != NULL)
1292 orig_fd->fd_failed_rebuilds++;
1295 /* Ban algo on non-recoverable error */
1296 if (result == FLM_ERROR)
1297 flm_error_add(flm, rh->rib_fibnum);
1299 if ((prev_fd != NULL) && (prev_fd != orig_fd))
1300 schedule_destroy_fd_instance(prev_fd, false);
1301 if (new_fd != NULL) {
1302 schedule_destroy_fd_instance(new_fd, false);
1312 * Tries to sync algo with the current rtable state, either
1313 * by executing batch update or rebuilding.
1314 * Returns true on success.
1317 execute_callout_action(struct fib_data *fd)
1319 enum fib_callout_action action = fd->fd_callout_action;
1320 struct fib_lookup_module *flm_new = NULL;
1324 RIB_WLOCK_ASSERT(fd->fd_rh);
1326 fd->fd_need_rebuild = false;
1327 fd->fd_batch = false;
1328 fd->fd_num_changes = 0;
1330 /* First, check if we're still OK to use this algo */
1331 if (!is_algo_fixed(fd->fd_rh))
1332 flm_new = fib_check_best_algo(fd->fd_rh, fd->fd_flm);
1333 if (flm_new != NULL)
1334 action = FDA_REBUILD;
1336 if (action == FDA_BATCH) {
1338 if (!apply_rtable_changes(fd))
1339 action = FDA_REBUILD;
1342 if (action == FDA_REBUILD)
1343 result = rebuild_fd_flm(fd, flm_new != NULL ? flm_new : fd->fd_flm);
1344 if (flm_new != NULL)
1345 fib_unref_algo(flm_new);
1351 * Callout for all scheduled fd-related work.
1352 * - Checks if the current algo is still the best algo
1353 * - Synchronises algo instance to the rtable (batch usecase)
1354 * - Creates a new instance of an algo for af/fib if desired.
1357 handle_fd_callout(void *_data)
1359 struct fib_data *fd = (struct fib_data *)_data;
1360 struct epoch_tracker et;
1362 FD_PRINTF(LOG_INFO, fd, "running callout type=%d", fd->fd_callout_action);
1364 NET_EPOCH_ENTER(et);
1365 CURVNET_SET(fd->fd_vnet);
1366 execute_callout_action(fd);
1372 * Tries to create new algo instance based on @fd data.
1373 * Returns true on success.
1376 rebuild_fd_flm(struct fib_data *fd, struct fib_lookup_module *flm_new)
1378 struct fib_data *fd_new, *fd_tmp = NULL;
1381 if (flm_new == fd->fd_flm)
1384 FD_PRINTF(LOG_NOTICE, fd, "switching algo to %s", flm_new->flm_name);
1386 result = setup_fd_instance(flm_new, fd->fd_rh, fd_tmp, &fd_new, true);
1387 if (result != FLM_SUCCESS) {
1388 FD_PRINTF(LOG_NOTICE, fd, "table rebuild failed");
1391 FD_PRINTF(LOG_INFO, fd_new, "switched to new instance");
1393 /* Remove old instance */
1394 schedule_destroy_fd_instance(fd, true);
1400 rebuild_fd(struct fib_data *fd, const char *reason)
1402 struct fib_lookup_module *flm_new = NULL;
1405 if (!is_algo_fixed(fd->fd_rh))
1406 flm_new = fib_check_best_algo(fd->fd_rh, fd->fd_flm);
1408 FD_PRINTF(LOG_INFO, fd, "running sync rebuild: %s", reason);
1409 result = rebuild_fd_flm(fd, flm_new != NULL ? flm_new : fd->fd_flm);
1410 if (flm_new != NULL)
1411 fib_unref_algo(flm_new);
1414 FD_PRINTF(LOG_ERR, fd, "sync rebuild failed");
1415 schedule_fd_rebuild(fd, "sync rebuild failed");
1422 * Finds algo by name/family.
1423 * Returns referenced algo or NULL.
1425 static struct fib_lookup_module *
1426 fib_find_algo(const char *algo_name, int family)
1428 struct fib_lookup_module *flm;
1431 TAILQ_FOREACH(flm, &all_algo_list, entries) {
1432 if ((strcmp(flm->flm_name, algo_name) == 0) &&
1433 (family == flm->flm_family)) {
1434 flm->flm_refcount++;
1445 fib_unref_algo(struct fib_lookup_module *flm)
1449 flm->flm_refcount--;
1454 set_fib_algo(uint32_t fibnum, int family, struct sysctl_oid *oidp, struct sysctl_req *req)
1456 struct fib_lookup_module *flm = NULL;
1457 struct fib_data *fd = NULL;
1458 char old_algo_name[32], algo_name[32];
1459 struct rib_head *rh = NULL;
1460 enum flm_op_result result;
1461 struct epoch_tracker et;
1464 /* Fetch current algo/rib for af/family */
1466 TAILQ_FOREACH(fd, &V_fib_data_list, entries) {
1467 if ((fd->fd_family == family) && (fd->fd_fibnum == fibnum))
1475 strlcpy(old_algo_name, fd->fd_flm->flm_name,
1476 sizeof(old_algo_name));
1479 strlcpy(algo_name, old_algo_name, sizeof(algo_name));
1480 error = sysctl_handle_string(oidp, algo_name, sizeof(algo_name), req);
1481 if (error != 0 || req->newptr == NULL)
1484 if (strcmp(algo_name, old_algo_name) == 0)
1487 /* New algorithm name is different */
1488 flm = fib_find_algo(algo_name, family);
1490 RH_PRINTF(LOG_INFO, rh, "unable to find algo %s", algo_name);
1495 NET_EPOCH_ENTER(et);
1497 result = setup_fd_instance(flm, rh, NULL, &fd, true);
1500 fib_unref_algo(flm);
1501 if (result != FLM_SUCCESS)
1504 /* Disable automated jumping between algos */
1508 /* Remove old instance(s) */
1509 fib_cleanup_algo(rh, true, false);
1511 /* Drain cb so user can unload the module after userret if so desired */
1512 epoch_drain_callbacks(net_epoch_preempt);
1519 set_algo_inet_sysctl_handler(SYSCTL_HANDLER_ARGS)
1522 return (set_fib_algo(curthread->td_proc->p_fibnum, AF_INET, oidp, req));
1524 SYSCTL_PROC(_net_route_algo_inet, OID_AUTO, algo,
1525 CTLFLAG_VNET | CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
1526 set_algo_inet_sysctl_handler, "A", "Set IPv4 lookup algo");
1531 set_algo_inet6_sysctl_handler(SYSCTL_HANDLER_ARGS)
1534 return (set_fib_algo(curthread->td_proc->p_fibnum, AF_INET6, oidp, req));
1536 SYSCTL_PROC(_net_route_algo_inet6, OID_AUTO, algo,
1537 CTLFLAG_VNET | CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
1538 set_algo_inet6_sysctl_handler, "A", "Set IPv6 lookup algo");
1541 static struct nhop_object *
1542 dummy_lookup(void *algo_data, const struct flm_lookup_key key, uint32_t scopeid)
1548 destroy_fdh_epoch(epoch_context_t ctx)
1550 struct fib_dp_header *fdh;
1552 fdh = __containerof(ctx, struct fib_dp_header, fdh_epoch_ctx);
1553 free(fdh, M_RTABLE);
1556 static struct fib_dp_header *
1557 alloc_fib_dp_array(uint32_t num_tables, bool waitok)
1560 struct fib_dp_header *fdh;
1562 sz = sizeof(struct fib_dp_header);
1563 sz += sizeof(struct fib_dp) * num_tables;
1564 fdh = malloc(sz, M_RTABLE, (waitok ? M_WAITOK : M_NOWAIT) | M_ZERO);
1566 fdh->fdh_num_tables = num_tables;
1568 * Set dummy lookup function ptr always returning NULL, so
1569 * we can delay algo init.
1571 for (uint32_t i = 0; i < num_tables; i++)
1572 fdh->fdh_idx[i].f = dummy_lookup;
1577 static struct fib_dp_header *
1578 get_fib_dp_header(struct fib_dp *dp)
1581 return (__containerof((void *)dp, struct fib_dp_header, fdh_idx));
1585 * Replace per-family index pool @pdp with a new one which
1586 * contains updated callback/algo data from @fd.
1587 * Returns true on success.
1590 replace_rtables_family(struct fib_dp **pdp, struct fib_data *fd, struct fib_dp *dp)
1592 struct fib_dp_header *new_fdh, *old_fdh;
1596 FD_PRINTF(LOG_DEBUG, fd, "[vnet %p] replace with f:%p arg:%p",
1597 curvnet, dp->f, dp->arg);
1600 old_fdh = get_fib_dp_header(*pdp);
1602 if (old_fdh->fdh_idx[fd->fd_fibnum].f == dp->f) {
1604 * Function is the same, data pointer needs update.
1605 * Perform in-line replace without reallocation.
1607 old_fdh->fdh_idx[fd->fd_fibnum].arg = dp->arg;
1608 FD_PRINTF(LOG_DEBUG, fd, "FDH %p inline update", old_fdh);
1613 new_fdh = alloc_fib_dp_array(old_fdh->fdh_num_tables, false);
1614 FD_PRINTF(LOG_DEBUG, fd, "OLD FDH: %p NEW FDH: %p", old_fdh, new_fdh);
1615 if (new_fdh == NULL) {
1617 FD_PRINTF(LOG_WARNING, fd, "error attaching datapath");
1621 memcpy(&new_fdh->fdh_idx[0], &old_fdh->fdh_idx[0],
1622 old_fdh->fdh_num_tables * sizeof(struct fib_dp));
1623 /* Update relevant data structure for @fd */
1624 new_fdh->fdh_idx[fd->fd_fibnum] = *dp;
1626 /* Ensure memcpy() writes have completed */
1627 atomic_thread_fence_rel();
1628 /* Set new datapath pointer */
1629 *pdp = &new_fdh->fdh_idx[0];
1631 FD_PRINTF(LOG_DEBUG, fd, "update %p -> %p", old_fdh, new_fdh);
1633 fib_epoch_call(destroy_fdh_epoch, &old_fdh->fdh_epoch_ctx);
1638 static struct fib_dp **
1639 get_family_dp_ptr(int family)
1644 return (&V_inet_dp);
1648 return (&V_inet6_dp);
1655 * Make datapath use fib instance @fd
1658 fib_set_datapath_ptr(struct fib_data *fd, struct fib_dp *dp)
1660 struct fib_dp **pdp;
1662 pdp = get_family_dp_ptr(fd->fd_family);
1663 return (replace_rtables_family(pdp, fd, dp));
1667 * Grow datapath pointers array.
1668 * Called from sysctl handler on growing number of routing tables.
1671 grow_rtables_family(struct fib_dp **pdp, uint32_t new_num_tables)
1673 struct fib_dp_header *new_fdh, *old_fdh = NULL;
1675 new_fdh = alloc_fib_dp_array(new_num_tables, true);
1679 old_fdh = get_fib_dp_header(*pdp);
1680 memcpy(&new_fdh->fdh_idx[0], &old_fdh->fdh_idx[0],
1681 old_fdh->fdh_num_tables * sizeof(struct fib_dp));
1684 /* Wait till all writes completed */
1685 atomic_thread_fence_rel();
1687 *pdp = &new_fdh->fdh_idx[0];
1690 if (old_fdh != NULL)
1691 fib_epoch_call(destroy_fdh_epoch, &old_fdh->fdh_epoch_ctx);
1695 * Grows per-AF arrays of datapath pointers for each supported family.
1696 * Called from fibs resize sysctl handler.
1699 fib_grow_rtables(uint32_t new_num_tables)
1703 grow_rtables_family(get_family_dp_ptr(AF_INET), new_num_tables);
1706 grow_rtables_family(get_family_dp_ptr(AF_INET6), new_num_tables);
1711 fib_get_rtable_info(struct rib_head *rh, struct rib_rtable_info *rinfo)
1714 bzero(rinfo, sizeof(struct rib_rtable_info));
1715 rinfo->num_prefixes = rh->rnh_prefixes;
1716 rinfo->num_nhops = nhops_get_count(rh);
1718 rinfo->num_nhgrp = nhgrp_get_count(rh);
1723 * Updates pointer to the algo data for the @fd.
1726 fib_set_algo_ptr(struct fib_data *fd, void *algo_data)
1728 RIB_WLOCK_ASSERT(fd->fd_rh);
1730 fd->fd_algo_data = algo_data;
1734 * Calls @callback with @ctx after the end of a current epoch.
1737 fib_epoch_call(epoch_callback_t callback, epoch_context_t ctx)
1739 epoch_call(net_epoch_preempt, callback, ctx);
1743 * Accessor to get rib instance @fd is attached to.
1746 fib_get_rh(struct fib_data *fd)
1753 * Accessor to export idx->nhop array
1755 struct nhop_object **
1756 fib_get_nhop_array(struct fib_data *fd)
1759 return (fd->nh_idx);
1763 get_nhop_idx(struct nhop_object *nh)
1766 if (NH_IS_NHGRP(nh))
1767 return (nhgrp_get_idx((struct nhgrp_object *)nh));
1770 return (nhop_get_idx(nh));
1774 fib_get_nhop_idx(struct fib_data *fd, struct nhop_object *nh)
1777 return (get_nhop_idx(nh));
1781 is_idx_free(struct fib_data *fd, uint32_t index)
1784 return (fd->nh_ref_table->refcnt[index] == 0);
1788 fib_ref_nhop(struct fib_data *fd, struct nhop_object *nh)
1790 uint32_t idx = get_nhop_idx(nh);
1792 if (idx >= fd->number_nhops) {
1797 if (is_idx_free(fd, idx)) {
1799 fd->nh_idx[idx] = nh;
1800 fd->nh_ref_table->count++;
1801 FD_PRINTF(LOG_DEBUG2, fd, " REF nhop %u %p", idx, fd->nh_idx[idx]);
1803 fd->nh_ref_table->refcnt[idx]++;
1808 struct nhop_release_data {
1809 struct nhop_object *nh;
1810 struct epoch_context ctx;
1814 release_nhop_epoch(epoch_context_t ctx)
1816 struct nhop_release_data *nrd;
1818 nrd = __containerof(ctx, struct nhop_release_data, ctx);
1819 nhop_free_any(nrd->nh);
1824 * Delays nexthop refcount release.
1825 * Datapath may have the datastructures not updated yet, so the old
1826 * nexthop may still be returned till the end of current epoch. Delay
1827 * refcount removal, as we may be removing the last instance, which will
1828 * trigger nexthop deletion, rendering returned nexthop invalid.
1831 fib_schedule_release_nhop(struct fib_data *fd, struct nhop_object *nh)
1833 struct nhop_release_data *nrd;
1835 nrd = malloc(sizeof(struct nhop_release_data), M_TEMP, M_NOWAIT | M_ZERO);
1838 fib_epoch_call(release_nhop_epoch, &nrd->ctx);
1841 * Unable to allocate memory. Leak nexthop to maintain guarantee
1842 * that each nhop can be referenced.
1844 FD_PRINTF(LOG_ERR, fd, "unable to schedule nhop %p deletion", nh);
1849 fib_unref_nhop(struct fib_data *fd, struct nhop_object *nh)
1851 uint32_t idx = get_nhop_idx(nh);
1853 KASSERT((idx < fd->number_nhops), ("invalid nhop index"));
1854 KASSERT((nh == fd->nh_idx[idx]), ("index table contains whong nh"));
1856 fd->nh_ref_table->refcnt[idx]--;
1857 if (fd->nh_ref_table->refcnt[idx] == 0) {
1858 FD_PRINTF(LOG_DEBUG, fd, " FREE nhop %d %p", idx, fd->nh_idx[idx]);
1859 fib_schedule_release_nhop(fd, fd->nh_idx[idx]);
1864 set_algo_fixed(struct rib_head *rh)
1866 switch (rh->rib_family) {
1869 V_algo_fixed_inet = true;
1874 V_algo_fixed_inet6 = true;
1881 is_algo_fixed(struct rib_head *rh)
1884 switch (rh->rib_family) {
1887 return (V_algo_fixed_inet);
1891 return (V_algo_fixed_inet6);
1898 * Runs the check on what would be the best algo for rib @rh, assuming
1899 * that the current algo is the one specified by @orig_flm. Note that
1900 * it can be NULL for initial selection.
1902 * Returns referenced new algo or NULL if the current one is the best.
1904 static struct fib_lookup_module *
1905 fib_check_best_algo(struct rib_head *rh, struct fib_lookup_module *orig_flm)
1907 uint8_t preference, curr_preference = 0, best_preference = 0;
1908 struct fib_lookup_module *flm, *best_flm = NULL;
1909 struct rib_rtable_info rinfo;
1910 int candidate_algos = 0;
1912 fib_get_rtable_info(rh, &rinfo);
1915 TAILQ_FOREACH(flm, &all_algo_list, entries) {
1916 if (flm->flm_family != rh->rib_family)
1919 preference = flm->flm_get_pref(&rinfo);
1920 if (preference > best_preference) {
1921 if (!flm_error_check(flm, rh->rib_fibnum)) {
1922 best_preference = preference;
1926 if (flm == orig_flm)
1927 curr_preference = preference;
1929 if ((best_flm != NULL) && (curr_preference + BEST_DIFF_PERCENT < best_preference))
1930 best_flm->flm_refcount++;
1935 RH_PRINTF(LOG_DEBUG, rh, "candidate_algos: %d, curr: %s(%d) result: %s(%d)",
1936 candidate_algos, orig_flm ? orig_flm->flm_name : "NULL", curr_preference,
1937 best_flm ? best_flm->flm_name : (orig_flm ? orig_flm->flm_name : "NULL"),
1944 * Called when new route table is created.
1945 * Selects, allocates and attaches fib algo for the table.
1948 fib_select_algo_initial(struct rib_head *rh, struct fib_dp *dp)
1950 struct fib_lookup_module *flm;
1951 struct fib_data *fd = NULL;
1952 enum flm_op_result result;
1953 struct epoch_tracker et;
1955 flm = fib_check_best_algo(rh, NULL);
1957 RH_PRINTF(LOG_CRIT, rh, "no algo selected");
1960 RH_PRINTF(LOG_INFO, rh, "selected algo %s", flm->flm_name);
1962 NET_EPOCH_ENTER(et);
1964 result = setup_fd_instance(flm, rh, NULL, &fd, false);
1968 RH_PRINTF(LOG_DEBUG, rh, "result=%d fd=%p", result, fd);
1969 if (result == FLM_SUCCESS)
1972 RH_PRINTF(LOG_CRIT, rh, "unable to setup algo %s", flm->flm_name);
1974 fib_unref_algo(flm);
1976 return (result == FLM_SUCCESS);
1980 * Sets up fib algo instances for the non-initialized RIBs in the @family.
1981 * Allocates temporary datapath index to amortize datapaint index updates
1982 * with large @num_tables.
1985 fib_setup_family(int family, uint32_t num_tables)
1987 struct fib_dp_header *new_fdh = alloc_fib_dp_array(num_tables, false);
1988 if (new_fdh == NULL) {
1989 ALGO_PRINTF(LOG_CRIT, "Unable to setup framework for %s", print_family(family));
1993 for (int i = 0; i < num_tables; i++) {
1994 struct rib_head *rh = rt_tables_get_rnh(i, family);
1995 if (rh->rib_algo_init)
1997 if (!fib_select_algo_initial(rh, &new_fdh->fdh_idx[i]))
2000 rh->rib_algo_init = true;
2004 struct fib_dp **pdp = get_family_dp_ptr(family);
2005 struct fib_dp_header *old_fdh = get_fib_dp_header(*pdp);
2007 /* Update the items not touched by the new init, from the old data pointer */
2008 for (int i = 0; i < num_tables; i++) {
2009 if (new_fdh->fdh_idx[i].f == dummy_lookup)
2010 new_fdh->fdh_idx[i] = old_fdh->fdh_idx[i];
2013 /* Ensure all index writes have completed */
2014 atomic_thread_fence_rel();
2015 /* Set new datapath pointer */
2016 *pdp = &new_fdh->fdh_idx[0];
2020 fib_epoch_call(destroy_fdh_epoch, &old_fdh->fdh_epoch_ctx);
2024 * Registers fib lookup module within the subsystem.
2027 fib_module_register(struct fib_lookup_module *flm)
2031 ALGO_PRINTF(LOG_INFO, "attaching %s to %s", flm->flm_name,
2032 print_family(flm->flm_family));
2033 TAILQ_INSERT_TAIL(&all_algo_list, flm, entries);
2040 * Tries to unregister fib lookup module.
2042 * Returns 0 on success, EBUSY if module is still used
2043 * by some of the tables.
2046 fib_module_unregister(struct fib_lookup_module *flm)
2050 if (flm->flm_refcount > 0) {
2054 fib_error_clear_flm(flm);
2055 ALGO_PRINTF(LOG_INFO, "detaching %s from %s", flm->flm_name,
2056 print_family(flm->flm_family));
2057 TAILQ_REMOVE(&all_algo_list, flm, entries);
2067 TAILQ_INIT(&V_fib_data_list);
2071 vnet_fib_destroy(void)