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(unsigned int, bucket_time_ms) = 50;
112 #define V_bucket_time_ms VNET(bucket_time_ms)
113 SYSCTL_UINT(_net_route_algo, OID_AUTO, bucket_time_ms, CTLFLAG_RW | CTLFLAG_VNET,
114 &VNET_NAME(bucket_time_ms), 0, "Time interval to calculate update rate");
116 /* Minimum update rate to delay sync */
117 VNET_DEFINE(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(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 */
156 struct fib_sync_status {
157 struct timeval diverge_time; /* ts when diverged */
158 uint32_t num_changes; /* number of changes since sync */
159 uint32_t bucket_changes; /* num changes within the current bucket */
160 uint64_t bucket_id; /* 50ms bucket # */
164 * Data structure for the fib lookup instance tied to the particular rib.
167 uint32_t number_nhops; /* current # of nhops */
168 uint8_t hit_nhops; /* true if out of nhop limit */
169 uint8_t init_done; /* true if init is competed */
170 uint32_t fd_dead:1; /* Scheduled for deletion */
171 uint32_t fd_linked:1; /* true if linked */
172 uint32_t fd_need_rebuild:1; /* true if rebuild scheduled */
173 uint8_t fd_family; /* family */
174 uint32_t fd_fibnum; /* fibnum */
175 uint32_t fd_failed_rebuilds; /* stat: failed rebuilds */
176 uint32_t fd_gen; /* instance gen# */
177 struct callout fd_callout; /* rebuild callout */
178 enum fib_callout_action fd_callout_action; /* Callout action to take */
179 void *fd_algo_data; /* algorithm data */
180 struct nhop_object **nh_idx; /* nhop idx->ptr array */
181 struct nhop_ref_table *nh_ref_table; /* array with # of nhop references */
182 struct rib_head *fd_rh; /* RIB table we're attached to */
183 struct rib_subscription *fd_rs; /* storing table subscription */
184 struct fib_dp fd_dp; /* fib datapath data */
185 struct vnet *fd_vnet; /* vnet fib belongs to */
186 struct epoch_context fd_epoch_ctx; /* epoch context for deletion */
187 struct fib_lookup_module *fd_flm;/* pointer to the lookup module */
188 struct fib_sync_status fd_ss; /* State relevant to the rib sync */
189 uint32_t fd_num_changes; /* number of changes since last callout */
190 TAILQ_ENTRY(fib_data) entries; /* list of all fds in vnet */
193 static bool rebuild_fd(struct fib_data *fd, const char *reason);
194 static bool rebuild_fd_flm(struct fib_data *fd, struct fib_lookup_module *flm_new);
195 static void handle_fd_callout(void *_data);
196 static void destroy_fd_instance_epoch(epoch_context_t ctx);
197 static enum flm_op_result attach_datapath(struct fib_data *fd);
198 static bool is_idx_free(struct fib_data *fd, uint32_t index);
199 static void set_algo_fixed(struct rib_head *rh);
200 static bool is_algo_fixed(struct rib_head *rh);
202 static uint32_t fib_ref_nhop(struct fib_data *fd, struct nhop_object *nh);
203 static void fib_unref_nhop(struct fib_data *fd, struct nhop_object *nh);
205 static struct fib_lookup_module *fib_check_best_algo(struct rib_head *rh,
206 struct fib_lookup_module *orig_flm);
207 static void fib_unref_algo(struct fib_lookup_module *flm);
208 static bool flm_error_check(const struct fib_lookup_module *flm, uint32_t fibnum);
211 #define FIB_MOD_LOCK() mtx_lock(&fib_mtx)
212 #define FIB_MOD_UNLOCK() mtx_unlock(&fib_mtx)
213 #define FIB_MOD_LOCK_ASSERT() mtx_assert(&fib_mtx, MA_OWNED)
215 MTX_SYSINIT(fib_mtx, &fib_mtx, "algo list mutex", MTX_DEF);
217 /* Algorithm has to be this percent better than the current to switch */
218 #define BEST_DIFF_PERCENT (5 * 256 / 100)
219 /* Schedule algo re-evaluation X seconds after a change */
220 #define ALGO_EVAL_DELAY_MS 30000
221 /* Force algo re-evaluation after X changes */
222 #define ALGO_EVAL_NUM_ROUTES 100
223 /* Try to setup algorithm X times */
224 #define FIB_MAX_TRIES 32
225 /* Max amount of supported nexthops */
226 #define FIB_MAX_NHOPS 262144
227 #define FIB_CALLOUT_DELAY_MS 50
231 static int flm_debug_level = LOG_NOTICE;
232 SYSCTL_INT(_net_route_algo, OID_AUTO, debug_level, CTLFLAG_RW | CTLFLAG_RWTUN,
233 &flm_debug_level, 0, "debuglevel");
234 #define FLM_MAX_DEBUG_LEVEL LOG_DEBUG
239 #define _PASS_MSG(_l) (flm_debug_level >= (_l))
240 #define ALGO_PRINTF(_fmt, ...) printf("[fib_algo] %s: " _fmt "\n", __func__, ##__VA_ARGS__)
241 #define _ALGO_PRINTF(_fib, _fam, _aname, _gen, _func, _fmt, ...) \
242 printf("[fib_algo] %s.%u (%s#%u) %s: " _fmt "\n",\
243 print_family(_fam), _fib, _aname, _gen, _func, ## __VA_ARGS__)
244 #define _RH_PRINTF(_fib, _fam, _func, _fmt, ...) \
245 printf("[fib_algo] %s.%u %s: " _fmt "\n", print_family(_fam), _fib, _func, ## __VA_ARGS__)
246 #define RH_PRINTF(_l, _rh, _fmt, ...) if (_PASS_MSG(_l)) { \
247 _RH_PRINTF(_rh->rib_fibnum, _rh->rib_family, __func__, _fmt, ## __VA_ARGS__);\
249 #define FD_PRINTF(_l, _fd, _fmt, ...) FD_PRINTF_##_l(_l, _fd, _fmt, ## __VA_ARGS__)
250 #define _FD_PRINTF(_l, _fd, _fmt, ...) if (_PASS_MSG(_l)) { \
251 _ALGO_PRINTF(_fd->fd_fibnum, _fd->fd_family, _fd->fd_flm->flm_name, \
252 _fd->fd_gen, __func__, _fmt, ## __VA_ARGS__); \
254 #if FLM_MAX_DEBUG_LEVEL>=LOG_DEBUG2
255 #define FD_PRINTF_LOG_DEBUG2 _FD_PRINTF
257 #define FD_PRINTF_LOG_DEBUG2(_l, _fd, _fmt, ...)
259 #if FLM_MAX_DEBUG_LEVEL>=LOG_DEBUG
260 #define FD_PRINTF_LOG_DEBUG _FD_PRINTF
262 #define FD_PRINTF_LOG_DEBUG()
264 #if FLM_MAX_DEBUG_LEVEL>=LOG_INFO
265 #define FD_PRINTF_LOG_INFO _FD_PRINTF
267 #define FD_PRINTF_LOG_INFO()
269 #define FD_PRINTF_LOG_NOTICE _FD_PRINTF
270 #define FD_PRINTF_LOG_ERR _FD_PRINTF
271 #define FD_PRINTF_LOG_WARNING _FD_PRINTF
274 /* List of all registered lookup algorithms */
275 static TAILQ_HEAD(, fib_lookup_module) all_algo_list = TAILQ_HEAD_INITIALIZER(all_algo_list);
277 /* List of all fib lookup instances in the vnet */
278 VNET_DEFINE_STATIC(TAILQ_HEAD(fib_data_head, fib_data), fib_data_list);
279 #define V_fib_data_list VNET(fib_data_list)
281 /* Datastructure for storing non-transient fib lookup module failures */
284 uint32_t fe_fibnum; /* failed rtable */
285 struct fib_lookup_module *fe_flm; /* failed module */
286 TAILQ_ENTRY(fib_error) entries;/* list of all errored entries */
288 VNET_DEFINE_STATIC(TAILQ_HEAD(fib_error_head, fib_error), fib_error_list);
289 #define V_fib_error_list VNET(fib_error_list)
291 /* Per-family array of fibnum -> {func, arg} mappings used in datapath */
292 struct fib_dp_header {
293 struct epoch_context fdh_epoch_ctx;
294 uint32_t fdh_num_tables;
295 struct fib_dp fdh_idx[0];
299 * Tries to add new non-transient algorithm error to the list of
301 * Returns true on success.
304 flm_error_add(struct fib_lookup_module *flm, uint32_t fibnum)
306 struct fib_error *fe;
308 fe = malloc(sizeof(struct fib_error), M_TEMP, M_NOWAIT | M_ZERO);
312 fe->fe_family = flm->flm_family;
313 fe->fe_fibnum = fibnum;
316 /* Avoid duplicates by checking if error already exists first */
317 if (flm_error_check(flm, fibnum)) {
322 TAILQ_INSERT_HEAD(&V_fib_error_list, fe, entries);
329 * True if non-transient error has been registered for @flm in @fibnum.
332 flm_error_check(const struct fib_lookup_module *flm, uint32_t fibnum)
334 const struct fib_error *fe;
336 TAILQ_FOREACH(fe, &V_fib_error_list, entries) {
337 if ((fe->fe_flm == flm) && (fe->fe_fibnum == fibnum))
345 * Clear all errors of algo specified by @flm.
348 fib_error_clear_flm(struct fib_lookup_module *flm)
350 struct fib_error *fe, *fe_tmp;
352 FIB_MOD_LOCK_ASSERT();
354 TAILQ_FOREACH_SAFE(fe, &V_fib_error_list, entries, fe_tmp) {
355 if (fe->fe_flm == flm) {
356 TAILQ_REMOVE(&V_fib_error_list, fe, entries);
363 * Clears all errors in current VNET.
368 struct fib_error *fe, *fe_tmp;
370 FIB_MOD_LOCK_ASSERT();
372 TAILQ_FOREACH_SAFE(fe, &V_fib_error_list, entries, fe_tmp) {
373 TAILQ_REMOVE(&V_fib_error_list, fe, entries);
379 print_op_result(enum flm_op_result result)
394 print_family(int family)
397 if (family == AF_INET)
399 else if (family == AF_INET6)
406 * Debug function used by lookup algorithms.
407 * Outputs message denoted by @fmt, prepended by "[fib_algo] inetX.Y (algo) "
410 fib_printf(int level, struct fib_data *fd, const char *func, char *fmt, ...)
415 if (level > flm_debug_level)
419 vsnprintf(buf, sizeof(buf), fmt, ap);
422 _ALGO_PRINTF(fd->fd_fibnum, fd->fd_family, fd->fd_flm->flm_name,
423 fd->fd_gen, func, "%s", buf);
427 * Outputs list of algorithms supported by the provided address family.
430 print_algos_sysctl(struct sysctl_req *req, int family)
432 struct fib_lookup_module *flm;
434 int error, count = 0;
436 error = sysctl_wire_old_buffer(req, 0);
438 sbuf_new_for_sysctl(&sbuf, NULL, 512, req);
439 TAILQ_FOREACH(flm, &all_algo_list, entries) {
440 if (flm->flm_family == family) {
442 sbuf_cat(&sbuf, ", ");
443 sbuf_cat(&sbuf, flm->flm_name);
446 error = sbuf_finish(&sbuf);
454 print_algos_sysctl_inet6(SYSCTL_HANDLER_ARGS)
457 return (print_algos_sysctl(req, AF_INET6));
459 SYSCTL_PROC(_net_route_algo_inet6, OID_AUTO, algo_list,
460 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
461 print_algos_sysctl_inet6, "A", "List of IPv6 lookup algorithms");
466 print_algos_sysctl_inet(SYSCTL_HANDLER_ARGS)
469 return (print_algos_sysctl(req, AF_INET));
471 SYSCTL_PROC(_net_route_algo_inet, OID_AUTO, algo_list,
472 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
473 print_algos_sysctl_inet, "A", "List of IPv4 lookup algorithms");
477 * Calculate delay between repeated failures.
478 * Returns current delay in milliseconds.
481 callout_calc_delay_ms(struct fib_data *fd)
485 if (fd->fd_failed_rebuilds > 10)
488 shift = fd->fd_failed_rebuilds;
490 return ((1 << shift) * FIB_CALLOUT_DELAY_MS);
494 schedule_callout(struct fib_data *fd, enum fib_callout_action action, int delay_ms)
497 FD_PRINTF(LOG_DEBUG, fd, "delay=%d action=%d", delay_ms, action);
498 fd->fd_callout_action = action;
499 callout_reset_sbt(&fd->fd_callout, SBT_1MS * delay_ms, 0,
500 handle_fd_callout, fd, 0);
504 schedule_fd_rebuild(struct fib_data *fd, const char *reason)
507 RIB_WLOCK_ASSERT(fd->fd_rh);
509 if (!fd->fd_need_rebuild) {
510 fd->fd_need_rebuild = true;
513 * Potentially re-schedules pending callout
514 * initiated by schedule_algo_eval.
516 FD_PRINTF(LOG_INFO, fd, "Scheduling rebuild: %s (failures=%d)",
517 reason, fd->fd_failed_rebuilds);
518 schedule_callout(fd, FDA_REBUILD, callout_calc_delay_ms(fd));
523 get_tv_diff_ms(const struct timeval *old_tv, const struct timeval *new_tv)
527 diff = ((int64_t)(new_tv->tv_sec - old_tv->tv_sec)) * 1000;
528 diff += (new_tv->tv_usec - old_tv->tv_usec) / 1000;
534 add_tv_diff_ms(struct timeval *tv, int ms)
536 tv->tv_sec += ms / 1000;
538 if (ms * 1000 + tv->tv_usec < 1000000)
539 tv->tv_usec += ms * 1000;
542 tv->tv_usec = ms * 1000 + tv->tv_usec - 1000000;
547 * Marks the time when algo state diverges from the rib state.
550 mark_diverge_time(struct fib_data *fd)
552 struct fib_sync_status *fd_ss = &fd->fd_ss;
554 getmicrouptime(&fd_ss->diverge_time);
555 fd_ss->bucket_id = 0;
556 fd_ss->bucket_changes = 0;
560 * Calculates and updates the next algorithm sync time, based on the current activity.
562 * The intent is to provide reasonable balance between the update
563 * latency and efficient batching when changing large amount of routes.
565 * High-level algorithm looks the following:
566 * 1) all changes are bucketed in 50ms intervals
567 * 2) If amount of changes within the bucket is greater than the threshold,
568 * the update gets delayed, up to maximum delay threshold.
571 update_rebuild_delay(struct fib_data *fd)
573 uint32_t bucket_id, new_delay = 0;
576 /* Fetch all variables at once to ensure consistent reads */
577 uint32_t bucket_time_ms = V_bucket_time_ms;
578 uint32_t threshold_rate = V_bucket_change_threshold_rate;
579 uint32_t max_delay_ms = V_fib_max_sync_delay_ms;
581 if (bucket_time_ms == 0)
583 /* calculate per-bucket threshold rate */
584 threshold_rate = threshold_rate * bucket_time_ms / 1000;
588 struct fib_sync_status *fd_ss = &fd->fd_ss;
590 bucket_id = get_tv_diff_ms(&fd_ss->diverge_time, &tv) / bucket_time_ms;
592 if (fd_ss->bucket_id == bucket_id) {
593 fd_ss->bucket_changes++;
594 if (fd_ss->bucket_changes == threshold_rate) {
595 new_delay = (bucket_id + 2) * bucket_time_ms;
596 if (new_delay <= max_delay_ms) {
597 FD_PRINTF(LOG_DEBUG, fd,
598 "hit threshold of %u routes, delay update,"
599 "bucket: %u, total delay: %u",
600 threshold_rate, bucket_id + 1, new_delay);
603 FD_PRINTF(LOG_DEBUG, fd,
604 "maximum sync delay (%u ms) reached", max_delay_ms);
606 } else if ((bucket_id == 0) && (fd_ss->bucket_changes == 1))
607 new_delay = bucket_time_ms;
609 fd_ss->bucket_id = bucket_id;
610 fd_ss->bucket_changes = 1;
614 /* Calculated time has been updated */
615 struct timeval new_tv = fd_ss->diverge_time;
616 add_tv_diff_ms(&new_tv, new_delay);
618 int32_t delay_ms = get_tv_diff_ms(&tv, &new_tv);
619 schedule_callout(fd, FDA_REBUILD, delay_ms);
624 update_algo_state(struct fib_data *fd)
627 RIB_WLOCK_ASSERT(fd->fd_rh);
629 if (fd->fd_need_rebuild) {
630 update_rebuild_delay(fd);
634 if (fd->fd_num_changes++ == 0) {
635 /* Start callout to consider switch */
636 if (!callout_pending(&fd->fd_callout))
637 schedule_callout(fd, FDA_EVAL, ALGO_EVAL_DELAY_MS);
638 } else if (fd->fd_num_changes == ALGO_EVAL_NUM_ROUTES) {
639 /* Reset callout to exec immediately */
640 if (fd->fd_callout_action == FDA_EVAL)
641 schedule_callout(fd, FDA_EVAL, 1);
646 need_immediate_sync(struct fib_data *fd, struct rib_cmd_info *rc)
648 struct nhop_object *nh;
650 /* Sync addition/removal of interface routes */
651 switch (rc->rc_cmd) {
654 if (!NH_IS_NHGRP(nh) && (!(nh->nh_flags & NHF_GATEWAY)))
659 if (!NH_IS_NHGRP(nh) && (!(nh->nh_flags & NHF_GATEWAY)))
668 * Rib subscription handler. Checks if the algorithm is ready to
669 * receive updates, handles nexthop refcounting and passes change
670 * data to the algorithm callback.
673 handle_rtable_change_cb(struct rib_head *rnh, struct rib_cmd_info *rc,
676 struct fib_data *fd = (struct fib_data *)_data;
677 enum flm_op_result result;
679 RIB_WLOCK_ASSERT(rnh);
682 * There is a small gap between subscribing for route changes
683 * and initiating rtable dump. Avoid receiving route changes
684 * prior to finishing rtable dump by checking `init_done`.
689 bool immediate_sync = need_immediate_sync(fd, rc);
691 /* Consider scheduling algorithm re-evaluation */
692 update_algo_state(fd);
695 * If algo requested rebuild, stop sending updates by default.
696 * This simplifies nexthop refcount handling logic.
698 if (fd->fd_need_rebuild)
702 * Maintain guarantee that every nexthop returned by the dataplane
703 * lookup has > 0 refcount, so can be safely referenced within current
706 if (rc->rc_nh_new != NULL) {
707 if (fib_ref_nhop(fd, rc->rc_nh_new) == 0) {
708 /* ran out of indexes */
709 schedule_fd_rebuild(fd, "ran out of nhop indexes");
714 result = fd->fd_flm->flm_change_rib_item_cb(rnh, rc, fd->fd_algo_data);
718 /* Unref old nexthop on success */
719 if (rc->rc_nh_old != NULL)
720 fib_unref_nhop(fd, rc->rc_nh_old);
725 * Algo is not able to apply the update.
726 * Schedule algo rebuild.
728 if (!immediate_sync) {
729 mark_diverge_time(fd);
730 schedule_fd_rebuild(fd, "algo requested rebuild");
734 FD_PRINTF(LOG_INFO, fd, "running sync rebuild");
735 rebuild_fd(fd, "rtable change type enforced sync");
740 * Algo reported a non-recoverable error.
741 * Record the error and schedule rebuild, which will
742 * trigger best algo selection.
744 FD_PRINTF(LOG_ERR, fd, "algo reported non-recoverable error");
745 if (!flm_error_add(fd->fd_flm, fd->fd_fibnum))
746 FD_PRINTF(LOG_ERR, fd, "failed to ban algo");
747 schedule_fd_rebuild(fd, "algo reported non-recoverable error");
752 estimate_nhop_scale(const struct fib_data *old_fd, struct fib_data *fd)
755 if (old_fd == NULL) {
756 // TODO: read from rtable
757 fd->number_nhops = 16;
761 if (old_fd->hit_nhops && old_fd->number_nhops < FIB_MAX_NHOPS)
762 fd->number_nhops = 2 * old_fd->number_nhops;
764 fd->number_nhops = old_fd->number_nhops;
770 enum flm_op_result result;
774 * Handler called after all rtenties have been dumped.
775 * Performs post-dump framework checks and calls
776 * algo:flm_dump_end_cb().
778 * Updates walk_cbdata result.
781 sync_algo_end_cb(struct rib_head *rnh, enum rib_walk_hook stage, void *_data)
783 struct walk_cbdata *w = (struct walk_cbdata *)_data;
784 struct fib_data *fd = w->fd;
786 RIB_WLOCK_ASSERT(w->fd->fd_rh);
788 if (rnh->rib_dying) {
789 w->result = FLM_ERROR;
794 FD_PRINTF(LOG_INFO, fd, "ran out of nexthops at %u nhops",
795 fd->nh_ref_table->count);
796 if (w->result == FLM_SUCCESS)
797 w->result = FLM_REBUILD;
801 if (stage != RIB_WALK_HOOK_POST || w->result != FLM_SUCCESS)
804 /* Post-dump hook, dump successful */
805 w->result = fd->fd_flm->flm_dump_end_cb(fd->fd_algo_data, &fd->fd_dp);
807 if (w->result == FLM_SUCCESS) {
808 /* Mark init as done to allow routing updates */
814 * Callback for each entry in rib.
815 * Calls algo:flm_dump_rib_item_cb func as a part of initial
816 * route table synchronisation.
819 sync_algo_cb(struct rtentry *rt, void *_data)
821 struct walk_cbdata *w = (struct walk_cbdata *)_data;
823 RIB_WLOCK_ASSERT(w->fd->fd_rh);
825 if (w->result == FLM_SUCCESS && w->func) {
828 * Reference nexthops to maintain guarantee that
829 * each nexthop returned by datapath has > 0 references
830 * and can be safely referenced within current epoch.
832 struct nhop_object *nh = rt_get_raw_nhop(rt);
833 if (fib_ref_nhop(w->fd, nh) != 0)
834 w->result = w->func(rt, w->fd->fd_algo_data);
836 w->result = FLM_REBUILD;
843 * Dump all routing table state to the algo instance.
845 static enum flm_op_result
846 sync_algo(struct fib_data *fd)
848 struct walk_cbdata w = {
850 .func = fd->fd_flm->flm_dump_rib_item_cb,
851 .result = FLM_SUCCESS,
854 rib_walk_ext_locked(fd->fd_rh, sync_algo_cb, sync_algo_end_cb, &w);
856 FD_PRINTF(LOG_INFO, fd,
857 "initial dump completed (rtable version: %d), result: %s",
858 fd->fd_rh->rnh_gen, print_op_result(w.result));
864 * Schedules epoch-backed @fd instance deletion.
865 * * Unlinks @fd from the list of active algo instances.
866 * * Removes rib subscription.
868 * * Schedules actual deletion.
870 * Assume @fd is already unlinked from the datapath.
873 schedule_destroy_fd_instance(struct fib_data *fd, bool in_callout)
878 RIB_WLOCK_ASSERT(fd->fd_rh);
881 is_dead = fd->fd_dead;
885 TAILQ_REMOVE(&V_fib_data_list, fd, entries);
886 fd->fd_linked = false;
892 FD_PRINTF(LOG_INFO, fd, "DETACH");
894 if (fd->fd_rs != NULL)
895 rib_unsibscribe_locked(fd->fd_rs);
898 * After rib_unsubscribe() no _new_ handle_rtable_change_cb() calls
899 * will be executed, hence no _new_ callout schedules will happen.
901 callout_stop(&fd->fd_callout);
903 epoch_call(net_epoch_preempt, destroy_fd_instance_epoch,
910 * Wipe all fd instances from the list matching rib specified by @rh.
911 * If @keep_first is set, remove all but the first record.
914 fib_cleanup_algo(struct rib_head *rh, bool keep_first, bool in_callout)
916 struct fib_data_head tmp_head = TAILQ_HEAD_INITIALIZER(tmp_head);
917 struct fib_data *fd, *fd_tmp;
918 struct epoch_tracker et;
921 TAILQ_FOREACH_SAFE(fd, &V_fib_data_list, entries, fd_tmp) {
922 if (fd->fd_rh == rh) {
927 TAILQ_REMOVE(&V_fib_data_list, fd, entries);
928 fd->fd_linked = false;
929 TAILQ_INSERT_TAIL(&tmp_head, fd, entries);
934 /* Pass 2: remove each entry */
936 TAILQ_FOREACH_SAFE(fd, &tmp_head, entries, fd_tmp) {
938 RIB_WLOCK(fd->fd_rh);
939 schedule_destroy_fd_instance(fd, in_callout);
941 RIB_WUNLOCK(fd->fd_rh);
947 fib_destroy_rib(struct rib_head *rh)
951 * rnh has `is_dying` flag set, so setup of new fd's will fail at
952 * sync_algo() stage, preventing new entries to be added to the list
953 * of active algos. Remove all existing entries for the particular rib.
955 fib_cleanup_algo(rh, false, false);
959 * Finalises fd destruction by freeing all fd resources.
962 destroy_fd_instance(struct fib_data *fd)
965 FD_PRINTF(LOG_INFO, fd, "destroy fd %p", fd);
967 /* Call destroy callback first */
968 if (fd->fd_algo_data != NULL)
969 fd->fd_flm->flm_destroy_cb(fd->fd_algo_data);
972 if ((fd->nh_idx != NULL) && (fd->nh_ref_table != NULL)) {
973 for (int i = 0; i < fd->number_nhops; i++) {
974 if (!is_idx_free(fd, i)) {
975 FD_PRINTF(LOG_DEBUG2, fd, " FREE nhop %d %p",
977 nhop_free_any(fd->nh_idx[i]);
980 free(fd->nh_idx, M_RTABLE);
982 if (fd->nh_ref_table != NULL)
983 free(fd->nh_ref_table, M_RTABLE);
985 fib_unref_algo(fd->fd_flm);
991 * Epoch callback indicating fd is safe to destroy
994 destroy_fd_instance_epoch(epoch_context_t ctx)
998 fd = __containerof(ctx, struct fib_data, fd_epoch_ctx);
1000 destroy_fd_instance(fd);
1004 * Tries to setup fd instance.
1005 * - Allocates fd/nhop table
1006 * - Runs algo:flm_init_cb algo init
1007 * - Subscribes fd to the rib
1008 * - Runs rtable dump
1009 * - Adds instance to the list of active instances.
1011 * Returns: operation result. Fills in @pfd with resulting fd on success.
1014 static enum flm_op_result
1015 try_setup_fd_instance(struct fib_lookup_module *flm, struct rib_head *rh,
1016 struct fib_data *old_fd, struct fib_data **pfd)
1018 struct fib_data *fd;
1020 enum flm_op_result result;
1023 fd = malloc(sizeof(struct fib_data), M_RTABLE, M_NOWAIT | M_ZERO);
1026 RH_PRINTF(LOG_INFO, rh, "Unable to allocate fib_data structure");
1027 return (FLM_REBUILD);
1031 estimate_nhop_scale(old_fd, fd);
1034 fd->fd_gen = ++fib_gen;
1035 fd->fd_family = rh->rib_family;
1036 fd->fd_fibnum = rh->rib_fibnum;
1037 callout_init_rm(&fd->fd_callout, &rh->rib_lock, 0);
1038 fd->fd_vnet = curvnet;
1041 FD_PRINTF(LOG_DEBUG, fd, "allocated fd %p", fd);
1044 flm->flm_refcount++;
1047 /* Allocate nhidx -> nhop_ptr table */
1048 size = fd->number_nhops * sizeof(void *);
1049 fd->nh_idx = malloc(size, M_RTABLE, M_NOWAIT | M_ZERO);
1050 if (fd->nh_idx == NULL) {
1051 FD_PRINTF(LOG_INFO, fd, "Unable to allocate nhop table idx (sz:%zu)", size);
1052 return (FLM_REBUILD);
1055 /* Allocate nhop index refcount table */
1056 size = sizeof(struct nhop_ref_table);
1057 size += fd->number_nhops * sizeof(uint32_t);
1058 fd->nh_ref_table = malloc(size, M_RTABLE, M_NOWAIT | M_ZERO);
1059 if (fd->nh_ref_table == NULL) {
1060 FD_PRINTF(LOG_INFO, fd, "Unable to allocate nhop refcount table (sz:%zu)", size);
1061 return (FLM_REBUILD);
1063 FD_PRINTF(LOG_DEBUG, fd, "Allocated %u nhop indexes", fd->number_nhops);
1065 /* Okay, we're ready for algo init */
1066 void *old_algo_data = (old_fd != NULL) ? old_fd->fd_algo_data : NULL;
1067 result = flm->flm_init_cb(fd->fd_fibnum, fd, old_algo_data, &fd->fd_algo_data);
1068 if (result != FLM_SUCCESS) {
1069 FD_PRINTF(LOG_INFO, fd, "%s algo init failed", flm->flm_name);
1073 /* Try to subscribe */
1074 if (flm->flm_change_rib_item_cb != NULL) {
1075 fd->fd_rs = rib_subscribe_locked(fd->fd_rh,
1076 handle_rtable_change_cb, fd, RIB_NOTIFY_IMMEDIATE);
1077 if (fd->fd_rs == NULL) {
1078 FD_PRINTF(LOG_INFO, fd, "failed to subscribe to the rib changes");
1079 return (FLM_REBUILD);
1084 result = sync_algo(fd);
1085 if (result != FLM_SUCCESS) {
1086 FD_PRINTF(LOG_INFO, fd, "rib sync failed");
1089 FD_PRINTF(LOG_INFO, fd, "DUMP completed successfully.");
1093 * Insert fd in the beginning of a list, to maintain invariant
1094 * that first matching entry for the AF/fib is always the active
1097 TAILQ_INSERT_HEAD(&V_fib_data_list, fd, entries);
1098 fd->fd_linked = true;
1101 return (FLM_SUCCESS);
1105 * Sets up algo @flm for table @rh and links it to the datapath.
1108 static enum flm_op_result
1109 setup_fd_instance(struct fib_lookup_module *flm, struct rib_head *rh,
1110 struct fib_data *orig_fd, struct fib_data **pfd, bool attach)
1112 struct fib_data *prev_fd, *new_fd;
1113 enum flm_op_result result;
1116 RIB_WLOCK_ASSERT(rh);
1120 for (int i = 0; i < FIB_MAX_TRIES; i++) {
1121 result = try_setup_fd_instance(flm, rh, prev_fd, &new_fd);
1123 if ((result == FLM_SUCCESS) && attach)
1124 result = attach_datapath(new_fd);
1126 if ((prev_fd != NULL) && (prev_fd != orig_fd)) {
1127 schedule_destroy_fd_instance(prev_fd, false);
1131 RH_PRINTF(LOG_INFO, rh, "try %d: fib algo result: %s", i,
1132 print_op_result(result));
1134 if (result == FLM_REBUILD) {
1143 if (result != FLM_SUCCESS) {
1144 RH_PRINTF(LOG_WARNING, rh,
1145 "%s algo instance setup failed, failures=%d", flm->flm_name,
1146 orig_fd ? orig_fd->fd_failed_rebuilds + 1 : 0);
1147 /* update failure count */
1149 if (orig_fd != NULL)
1150 orig_fd->fd_failed_rebuilds++;
1153 /* Ban algo on non-recoverable error */
1154 if (result == FLM_ERROR)
1155 flm_error_add(flm, rh->rib_fibnum);
1157 if ((prev_fd != NULL) && (prev_fd != orig_fd))
1158 schedule_destroy_fd_instance(prev_fd, false);
1159 if (new_fd != NULL) {
1160 schedule_destroy_fd_instance(new_fd, false);
1170 * Tries to sync algo with the current rtable state, either
1171 * by executing batch update or rebuilding.
1172 * Returns true on success.
1175 execute_callout_action(struct fib_data *fd)
1177 enum fib_callout_action action = fd->fd_callout_action;
1178 struct fib_lookup_module *flm_new = NULL;
1182 RIB_WLOCK_ASSERT(fd->fd_rh);
1184 fd->fd_need_rebuild = false;
1185 fd->fd_num_changes = 0;
1187 /* First, check if we're still OK to use this algo */
1188 if (!is_algo_fixed(fd->fd_rh))
1189 flm_new = fib_check_best_algo(fd->fd_rh, fd->fd_flm);
1190 if (flm_new != NULL)
1191 action = FDA_REBUILD;
1193 if (action == FDA_REBUILD)
1194 result = rebuild_fd_flm(fd, flm_new != NULL ? flm_new : fd->fd_flm);
1195 if (flm_new != NULL)
1196 fib_unref_algo(flm_new);
1202 * Callout for all scheduled fd-related work.
1203 * - Checks if the current algo is still the best algo
1204 * - Creates a new instance of an algo for af/fib if desired.
1207 handle_fd_callout(void *_data)
1209 struct fib_data *fd = (struct fib_data *)_data;
1210 struct epoch_tracker et;
1212 FD_PRINTF(LOG_INFO, fd, "running callout type=%d", fd->fd_callout_action);
1214 NET_EPOCH_ENTER(et);
1215 CURVNET_SET(fd->fd_vnet);
1216 execute_callout_action(fd);
1222 * Tries to create new algo instance based on @fd data.
1223 * Returns true on success.
1226 rebuild_fd_flm(struct fib_data *fd, struct fib_lookup_module *flm_new)
1228 struct fib_data *fd_new, *fd_tmp = NULL;
1231 if (flm_new == fd->fd_flm)
1234 FD_PRINTF(LOG_NOTICE, fd, "switching algo to %s", flm_new->flm_name);
1236 result = setup_fd_instance(flm_new, fd->fd_rh, fd_tmp, &fd_new, true);
1237 if (result != FLM_SUCCESS) {
1238 FD_PRINTF(LOG_NOTICE, fd, "table rebuild failed");
1241 FD_PRINTF(LOG_INFO, fd_new, "switched to new instance");
1243 /* Remove old instance */
1244 schedule_destroy_fd_instance(fd, true);
1250 rebuild_fd(struct fib_data *fd, const char *reason)
1252 struct fib_lookup_module *flm_new = NULL;
1255 if (!is_algo_fixed(fd->fd_rh))
1256 flm_new = fib_check_best_algo(fd->fd_rh, fd->fd_flm);
1258 FD_PRINTF(LOG_INFO, fd, "running sync rebuild: %s", reason);
1259 result = rebuild_fd_flm(fd, flm_new != NULL ? flm_new : fd->fd_flm);
1260 if (flm_new != NULL)
1261 fib_unref_algo(flm_new);
1264 FD_PRINTF(LOG_ERR, fd, "sync rebuild failed");
1265 schedule_fd_rebuild(fd, "sync rebuild failed");
1272 * Finds algo by name/family.
1273 * Returns referenced algo or NULL.
1275 static struct fib_lookup_module *
1276 fib_find_algo(const char *algo_name, int family)
1278 struct fib_lookup_module *flm;
1281 TAILQ_FOREACH(flm, &all_algo_list, entries) {
1282 if ((strcmp(flm->flm_name, algo_name) == 0) &&
1283 (family == flm->flm_family)) {
1284 flm->flm_refcount++;
1295 fib_unref_algo(struct fib_lookup_module *flm)
1299 flm->flm_refcount--;
1304 set_fib_algo(uint32_t fibnum, int family, struct sysctl_oid *oidp, struct sysctl_req *req)
1306 struct fib_lookup_module *flm = NULL;
1307 struct fib_data *fd = NULL;
1308 char old_algo_name[32], algo_name[32];
1309 struct rib_head *rh = NULL;
1310 enum flm_op_result result;
1311 struct epoch_tracker et;
1314 /* Fetch current algo/rib for af/family */
1316 TAILQ_FOREACH(fd, &V_fib_data_list, entries) {
1317 if ((fd->fd_family == family) && (fd->fd_fibnum == fibnum))
1325 strlcpy(old_algo_name, fd->fd_flm->flm_name,
1326 sizeof(old_algo_name));
1329 strlcpy(algo_name, old_algo_name, sizeof(algo_name));
1330 error = sysctl_handle_string(oidp, algo_name, sizeof(algo_name), req);
1331 if (error != 0 || req->newptr == NULL)
1334 if (strcmp(algo_name, old_algo_name) == 0)
1337 /* New algorithm name is different */
1338 flm = fib_find_algo(algo_name, family);
1340 RH_PRINTF(LOG_INFO, rh, "unable to find algo %s", algo_name);
1345 NET_EPOCH_ENTER(et);
1347 result = setup_fd_instance(flm, rh, NULL, &fd, true);
1350 fib_unref_algo(flm);
1351 if (result != FLM_SUCCESS)
1354 /* Disable automated jumping between algos */
1358 /* Remove old instance(s) */
1359 fib_cleanup_algo(rh, true, false);
1361 /* Drain cb so user can unload the module after userret if so desired */
1362 epoch_drain_callbacks(net_epoch_preempt);
1369 set_algo_inet_sysctl_handler(SYSCTL_HANDLER_ARGS)
1372 return (set_fib_algo(curthread->td_proc->p_fibnum, AF_INET, oidp, req));
1374 SYSCTL_PROC(_net_route_algo_inet, OID_AUTO, algo,
1375 CTLFLAG_VNET | CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
1376 set_algo_inet_sysctl_handler, "A", "Set IPv4 lookup algo");
1381 set_algo_inet6_sysctl_handler(SYSCTL_HANDLER_ARGS)
1384 return (set_fib_algo(curthread->td_proc->p_fibnum, AF_INET6, oidp, req));
1386 SYSCTL_PROC(_net_route_algo_inet6, OID_AUTO, algo,
1387 CTLFLAG_VNET | CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
1388 set_algo_inet6_sysctl_handler, "A", "Set IPv6 lookup algo");
1392 destroy_fdh_epoch(epoch_context_t ctx)
1394 struct fib_dp_header *fdh;
1396 fdh = __containerof(ctx, struct fib_dp_header, fdh_epoch_ctx);
1397 free(fdh, M_RTABLE);
1400 static struct fib_dp_header *
1401 alloc_fib_dp_array(uint32_t num_tables, bool waitok)
1404 struct fib_dp_header *fdh;
1406 sz = sizeof(struct fib_dp_header);
1407 sz += sizeof(struct fib_dp) * num_tables;
1408 fdh = malloc(sz, M_RTABLE, (waitok ? M_WAITOK : M_NOWAIT) | M_ZERO);
1410 fdh->fdh_num_tables = num_tables;
1414 static struct fib_dp_header *
1415 get_fib_dp_header(struct fib_dp *dp)
1418 return (__containerof((void *)dp, struct fib_dp_header, fdh_idx));
1422 * Replace per-family index pool @pdp with a new one which
1423 * contains updated callback/algo data from @fd.
1424 * Returns 0 on success.
1426 static enum flm_op_result
1427 replace_rtables_family(struct fib_dp **pdp, struct fib_data *fd)
1429 struct fib_dp_header *new_fdh, *old_fdh;
1433 FD_PRINTF(LOG_DEBUG, fd, "[vnet %p] replace with f:%p arg:%p",
1434 curvnet, fd->fd_dp.f, fd->fd_dp.arg);
1437 old_fdh = get_fib_dp_header(*pdp);
1438 new_fdh = alloc_fib_dp_array(old_fdh->fdh_num_tables, false);
1439 FD_PRINTF(LOG_DEBUG, fd, "OLD FDH: %p NEW FDH: %p", old_fdh, new_fdh);
1440 if (new_fdh == NULL) {
1442 FD_PRINTF(LOG_WARNING, fd, "error attaching datapath");
1443 return (FLM_REBUILD);
1446 memcpy(&new_fdh->fdh_idx[0], &old_fdh->fdh_idx[0],
1447 old_fdh->fdh_num_tables * sizeof(struct fib_dp));
1448 /* Update relevant data structure for @fd */
1449 new_fdh->fdh_idx[fd->fd_fibnum] = fd->fd_dp;
1451 /* Ensure memcpy() writes have completed */
1452 atomic_thread_fence_rel();
1453 /* Set new datapath pointer */
1454 *pdp = &new_fdh->fdh_idx[0];
1456 FD_PRINTF(LOG_DEBUG, fd, "update %p -> %p", old_fdh, new_fdh);
1458 epoch_call(net_epoch_preempt, destroy_fdh_epoch,
1459 &old_fdh->fdh_epoch_ctx);
1461 return (FLM_SUCCESS);
1464 static struct fib_dp **
1465 get_family_dp_ptr(int family)
1469 return (&V_inet_dp);
1471 return (&V_inet6_dp);
1477 * Make datapath use fib instance @fd
1479 static enum flm_op_result
1480 attach_datapath(struct fib_data *fd)
1482 struct fib_dp **pdp;
1484 pdp = get_family_dp_ptr(fd->fd_family);
1485 return (replace_rtables_family(pdp, fd));
1489 * Grow datapath pointers array.
1490 * Called from sysctl handler on growing number of routing tables.
1493 grow_rtables_family(struct fib_dp **pdp, uint32_t new_num_tables)
1495 struct fib_dp_header *new_fdh, *old_fdh = NULL;
1497 new_fdh = alloc_fib_dp_array(new_num_tables, true);
1501 old_fdh = get_fib_dp_header(*pdp);
1502 memcpy(&new_fdh->fdh_idx[0], &old_fdh->fdh_idx[0],
1503 old_fdh->fdh_num_tables * sizeof(struct fib_dp));
1506 /* Wait till all writes completed */
1507 atomic_thread_fence_rel();
1509 *pdp = &new_fdh->fdh_idx[0];
1512 if (old_fdh != NULL)
1513 epoch_call(net_epoch_preempt, destroy_fdh_epoch,
1514 &old_fdh->fdh_epoch_ctx);
1518 * Grows per-AF arrays of datapath pointers for each supported family.
1519 * Called from fibs resize sysctl handler.
1522 fib_grow_rtables(uint32_t new_num_tables)
1526 grow_rtables_family(get_family_dp_ptr(AF_INET), new_num_tables);
1529 grow_rtables_family(get_family_dp_ptr(AF_INET6), new_num_tables);
1534 fib_get_rtable_info(struct rib_head *rh, struct rib_rtable_info *rinfo)
1537 bzero(rinfo, sizeof(struct rib_rtable_info));
1538 rinfo->num_prefixes = rh->rnh_prefixes;
1539 rinfo->num_nhops = nhops_get_count(rh);
1541 rinfo->num_nhgrp = nhgrp_get_count(rh);
1546 * Accessor to get rib instance @fd is attached to.
1549 fib_get_rh(struct fib_data *fd)
1556 * Accessor to export idx->nhop array
1558 struct nhop_object **
1559 fib_get_nhop_array(struct fib_data *fd)
1562 return (fd->nh_idx);
1566 get_nhop_idx(struct nhop_object *nh)
1569 if (NH_IS_NHGRP(nh))
1570 return (nhgrp_get_idx((struct nhgrp_object *)nh) * 2 - 1);
1572 return (nhop_get_idx(nh) * 2);
1574 return (nhop_get_idx(nh));
1579 fib_get_nhop_idx(struct fib_data *fd, struct nhop_object *nh)
1582 return (get_nhop_idx(nh));
1586 is_idx_free(struct fib_data *fd, uint32_t index)
1589 return (fd->nh_ref_table->refcnt[index] == 0);
1593 fib_ref_nhop(struct fib_data *fd, struct nhop_object *nh)
1595 uint32_t idx = get_nhop_idx(nh);
1597 if (idx >= fd->number_nhops) {
1602 if (is_idx_free(fd, idx)) {
1604 fd->nh_idx[idx] = nh;
1605 fd->nh_ref_table->count++;
1606 FD_PRINTF(LOG_DEBUG2, fd, " REF nhop %u %p", idx, fd->nh_idx[idx]);
1608 fd->nh_ref_table->refcnt[idx]++;
1613 struct nhop_release_data {
1614 struct nhop_object *nh;
1615 struct epoch_context ctx;
1619 release_nhop_epoch(epoch_context_t ctx)
1621 struct nhop_release_data *nrd;
1623 nrd = __containerof(ctx, struct nhop_release_data, ctx);
1624 nhop_free_any(nrd->nh);
1629 * Delays nexthop refcount release.
1630 * Datapath may have the datastructures not updated yet, so the old
1631 * nexthop may still be returned till the end of current epoch. Delay
1632 * refcount removal, as we may be removing the last instance, which will
1633 * trigger nexthop deletion, rendering returned nexthop invalid.
1636 fib_schedule_release_nhop(struct fib_data *fd, struct nhop_object *nh)
1638 struct nhop_release_data *nrd;
1640 nrd = malloc(sizeof(struct nhop_release_data), M_TEMP, M_NOWAIT | M_ZERO);
1643 epoch_call(net_epoch_preempt, release_nhop_epoch, &nrd->ctx);
1646 * Unable to allocate memory. Leak nexthop to maintain guarantee
1647 * that each nhop can be referenced.
1649 FD_PRINTF(LOG_ERR, fd, "unable to schedule nhop %p deletion", nh);
1654 fib_unref_nhop(struct fib_data *fd, struct nhop_object *nh)
1656 uint32_t idx = get_nhop_idx(nh);
1658 KASSERT((idx < fd->number_nhops), ("invalid nhop index"));
1659 KASSERT((nh == fd->nh_idx[idx]), ("index table contains whong nh"));
1661 fd->nh_ref_table->refcnt[idx]--;
1662 if (fd->nh_ref_table->refcnt[idx] == 0) {
1663 FD_PRINTF(LOG_DEBUG, fd, " FREE nhop %d %p", idx, fd->nh_idx[idx]);
1664 fib_schedule_release_nhop(fd, fd->nh_idx[idx]);
1669 set_algo_fixed(struct rib_head *rh)
1671 switch (rh->rib_family) {
1674 V_algo_fixed_inet = true;
1679 V_algo_fixed_inet6 = true;
1686 is_algo_fixed(struct rib_head *rh)
1689 switch (rh->rib_family) {
1692 return (V_algo_fixed_inet);
1696 return (V_algo_fixed_inet6);
1703 * Runs the check on what would be the best algo for rib @rh, assuming
1704 * that the current algo is the one specified by @orig_flm. Note that
1705 * it can be NULL for initial selection.
1707 * Returns referenced new algo or NULL if the current one is the best.
1709 static struct fib_lookup_module *
1710 fib_check_best_algo(struct rib_head *rh, struct fib_lookup_module *orig_flm)
1712 uint8_t preference, curr_preference = 0, best_preference = 0;
1713 struct fib_lookup_module *flm, *best_flm = NULL;
1714 struct rib_rtable_info rinfo;
1715 int candidate_algos = 0;
1717 fib_get_rtable_info(rh, &rinfo);
1720 TAILQ_FOREACH(flm, &all_algo_list, entries) {
1721 if (flm->flm_family != rh->rib_family)
1724 preference = flm->flm_get_pref(&rinfo);
1725 if (preference > best_preference) {
1726 if (!flm_error_check(flm, rh->rib_fibnum)) {
1727 best_preference = preference;
1731 if (flm == orig_flm)
1732 curr_preference = preference;
1734 if ((best_flm != NULL) && (curr_preference + BEST_DIFF_PERCENT < best_preference))
1735 best_flm->flm_refcount++;
1740 RH_PRINTF(LOG_DEBUG, rh, "candidate_algos: %d, curr: %s(%d) result: %s(%d)",
1741 candidate_algos, orig_flm ? orig_flm->flm_name : "NULL", curr_preference,
1742 best_flm ? best_flm->flm_name : (orig_flm ? orig_flm->flm_name : "NULL"),
1749 * Called when new route table is created.
1750 * Selects, allocates and attaches fib algo for the table.
1753 fib_select_algo_initial(struct rib_head *rh)
1755 struct fib_lookup_module *flm;
1756 struct fib_data *fd = NULL;
1757 enum flm_op_result result;
1758 struct epoch_tracker et;
1761 flm = fib_check_best_algo(rh, NULL);
1763 RH_PRINTF(LOG_CRIT, rh, "no algo selected");
1766 RH_PRINTF(LOG_INFO, rh, "selected algo %s", flm->flm_name);
1768 NET_EPOCH_ENTER(et);
1770 result = setup_fd_instance(flm, rh, NULL, &fd, false);
1774 RH_PRINTF(LOG_DEBUG, rh, "result=%d fd=%p", result, fd);
1775 if (result == FLM_SUCCESS) {
1778 * Attach datapath directly to avoid multiple reallocations
1781 struct fib_dp_header *fdp;
1782 struct fib_dp **pdp;
1784 pdp = get_family_dp_ptr(rh->rib_family);
1786 fdp = get_fib_dp_header(*pdp);
1787 fdp->fdh_idx[fd->fd_fibnum] = fd->fd_dp;
1788 FD_PRINTF(LOG_INFO, fd, "datapath attached");
1792 RH_PRINTF(LOG_CRIT, rh, "unable to setup algo %s", flm->flm_name);
1795 fib_unref_algo(flm);
1801 * Registers fib lookup module within the subsystem.
1804 fib_module_register(struct fib_lookup_module *flm)
1808 ALGO_PRINTF("attaching %s to %s", flm->flm_name,
1809 print_family(flm->flm_family));
1810 TAILQ_INSERT_TAIL(&all_algo_list, flm, entries);
1817 * Tries to unregister fib lookup module.
1819 * Returns 0 on success, EBUSY if module is still used
1820 * by some of the tables.
1823 fib_module_unregister(struct fib_lookup_module *flm)
1827 if (flm->flm_refcount > 0) {
1831 fib_error_clear_flm(flm);
1832 ALGO_PRINTF("detaching %s from %s", flm->flm_name,
1833 print_family(flm->flm_family));
1834 TAILQ_REMOVE(&all_algo_list, flm, entries);
1844 TAILQ_INIT(&V_fib_data_list);
1848 vnet_fib_destroy(void)