Line data Source code
1 : /* The repair tile is responsible for repairing missing shreds that were
2 : not received via Turbine. The goal is to ensure that slots we "care"
3 : about have their FEC sets inserted into store.
4 :
5 : Generally there are two distinct traffic patterns:
6 :
7 : a. Firedancer boots up and fires off a large number of repairs to
8 : recover all the blocks between the snapshot on which it is booting
9 : and the head of the chain. In this mode, repair tile utilization
10 : is very high along with net and sign utilization.
11 :
12 : b. Firedancer catches up to the head of the chain and enters steady
13 : state where most shred traffic is delivered over turbine. In this
14 : state, repairs are only occasionally needed to recover shreds lost
15 : due to anomalies like packet loss, transmitter (leader) never sent
16 : them or even a malicious leader etc. On rare occasion, repair
17 : will also need to recover a different version of a block that
18 : equivocated.
19 :
20 : To accomplish the above, repair mainly processes 4 kinds of frags:
21 :
22 : 1. Shred data (from shred tile)
23 :
24 : Any shred (coding or data) that passes validation and filtering in
25 : the shred tile is forwarded to repair. Repair uses these to track
26 : which shreds have been received in `fd_forest`, a tree data
27 : structure that mirrors the block/slot ancestry chain. It also
28 : uses these shreds to discover slots or ancestries that were not
29 : known. fd_forest tracks metadata for each slot, including slot
30 : completion status, merkle roots, and metrics.
31 :
32 : Any shred that we can correlate with a repair request we made is
33 : used to update peer response latency metrics in fd_policy (See
34 : fd_policy.h for more details).
35 :
36 : 2. FEC status messages (from shred tile)
37 :
38 : These fall under two categories: FEC completion and FEC eviction.
39 : When all shreds in a FEC set have been recovered, the shred tile
40 : sends a completion message. This may trigger chained merkle
41 : verification if the slot has a confirmed block_id. The completed
42 : FEC message is always forwarded to replay via repair_out.
43 :
44 : When an incomplete FEC set is evicted from the shred tile's FEC
45 : resolver (e.g. due to capacity), it also notifies repair. Repair
46 : clears the corresponding FEC set entries from the forest so those
47 : shred indices can be re-requested if they are necessary. As
48 : mentioned in fd_forest.h, forest needs to maintain a strict subset
49 : of shreds that are known by fec_resolver, store, and reasm in
50 : order to guarantee forward progress always.
51 :
52 : 3. Pings (from net tile)
53 :
54 : Repair peers use a ping-pong protocol to verify liveness before
55 : serving repair requests. When a ping arrives over the network,
56 : repair validates the message and constructs a pong response. To
57 : prevent spam attacks, repair has stopgaps like tracking how many
58 : pongs per peer are currently in the sign queue and dropping pings
59 : from unknown peers. These are the only untrusted inputs to
60 : repair.
61 :
62 : 4. Sign task responses (from sign tile)
63 :
64 : Repair requests are signed asynchronously; the repair tile
65 : constructs a repair request, dispatches it to a sign tile via the
66 : repair_sign output link. The repair-sign communication is
67 : manually managed via credit tracking in the repair tile (see
68 : comment in out_ctx_t struct definition).
69 :
70 : After receiving the signature back from sign tile, repair injects
71 : the signature into the pending request and dispatches it to the
72 : net tile via the repair_net output link. The behavior depends on
73 : the request type:
74 : - Pong: the signed pong is sent to the peer that pinged us.
75 : - Warmup: a proactive request sent when a new peer's contact info
76 : first arrives, prepaying the ping-pong RTT cost. The signed
77 : request is sent to the peer if it is still active.
78 : - Regular shred request: the request is recorded in an inflight
79 : table (for tracking response latency and timeouts) and the
80 : signed packet is sent to the selected peer.
81 :
82 : Secondary "other" frags that are processed but not part of core
83 : repair logic:
84 :
85 : 5. Confirmation messages from tower
86 :
87 : Tower sends two kinds of messages relevant to repair:
88 : - slot_done: indicates a slot has finished replay and may advance
89 : the root. Repair publishes (roots) the forest up to that slot,
90 : pruning old ancestry.
91 : - slot_confirmed: indicates a slot has reached a confirmation
92 : level (e.g. duplicate-confirmed). If the slot is not yet in the
93 : forest, repair creates a sentinel block so it can be repaired.
94 : It also stores the confirmed block_id and may trigger chained
95 : merkle verification. See fd_forest.h on more details about
96 : chained merkle verification.
97 :
98 : 6. Eviction messages from replay (reasm)
99 :
100 : When the replay tile's reassembly buffer evicts a FEC set (e.g.
101 : due to pool capacity) from itself and from store, it notifies
102 : repair with the slot and fec_set_idx. Repair clears those FEC
103 : entries from the forest so the shreds can be re-requested.
104 :
105 : 7. Contact info messages from gossip
106 :
107 : Gossip forwards contact info updates and removals for other
108 : validators. Repair uses these to maintain a list of peers to
109 : make requests to.
110 :
111 : If fd_forest tracks what we know about each shred, fd_policy and
112 : fd_inflights is responsible for deciding what next repair request to
113 : make. fd_policy and fd_inflights split responsibility: fd_policy
114 : makes any new requests, orphan requests, and requests directly off
115 : the forest iterator, while fd_inflights re-requests anything that has
116 : been requested but not received yet within a timeout window. */
117 :
118 : #define _GNU_SOURCE
119 :
120 : #include "../genesis/fd_genesi_tile.h"
121 : #include "../../disco/topo/fd_topo.h"
122 : #include "../../disco/fd_clock_tile.h"
123 : #include "generated/fd_repair_tile_seccomp.h"
124 : #include "../../disco/keyguard/fd_keyload.h"
125 : #include "../../disco/keyguard/fd_keyguard.h"
126 : #include "../../disco/keyguard/fd_keyswitch.h"
127 : #include "../../disco/metrics/fd_metrics.h"
128 : #include "../../disco/net/fd_net_tile.h"
129 : #include "../../disco/shred/fd_rnonce_ss.h"
130 : #include "../../disco/shred/fd_shred_tile.h"
131 : #include "fd_repair_tile.h"
132 : #include "../../flamenco/gossip/fd_gossip_message.h"
133 : #include "../replay/fd_replay_tile.h"
134 : #include "../tower/fd_tower_tile.h"
135 : #include "../../discof/restore/utils/fd_ssmsg.h"
136 : #include "../../util/net/fd_net_headers.h"
137 : #include "../../util/pod/fd_pod_format.h"
138 : #include "../../tango/fd_tango_base.h"
139 :
140 : #include "../forest/fd_forest.h"
141 : #include "fd_repair_metrics.h"
142 : #include "fd_inflight.h"
143 : #include "fd_repair.h"
144 : #include "fd_policy.h"
145 :
146 : #define DEBUG_LOGGING 0
147 :
148 : #define IN_KIND_CONTACT (0)
149 0 : #define IN_KIND_NET (1)
150 0 : #define IN_KIND_TOWER (2)
151 0 : #define IN_KIND_SHRED (3)
152 0 : #define IN_KIND_SIGN (4)
153 0 : #define IN_KIND_SNAP (5)
154 0 : #define IN_KIND_GOSSIP (6)
155 0 : #define IN_KIND_GENESIS (7)
156 0 : #define IN_KIND_REPLAY (8)
157 :
158 : #define MAX_IN_LINKS (32)
159 : #define MAX_SHRED_TILE_CNT ( 16UL )
160 : #define MAX_SIGN_TILE_CNT ( 16UL )
161 :
162 : /* Max number of validators that can be actively queried */
163 0 : #define FD_REPAIR_PEER_MAX (FD_CONTACT_INFO_TABLE_SIZE)
164 :
165 : /* Max number of pending repair requests recently made to keep track of.
166 : Calculated generally as we estimate around 50k/s/core to sign
167 : requests. Assuming an over-provisioned 4 sign tiles just for repair,
168 : this means we can make up to ~200k requests per second. With a dedup
169 : timeout of 80ms, this means we can make up to ~16k requests within
170 : the dedup timeout window. We round up to the next power of two to
171 : get the dedup cache max. Since we are sizing the dedup cache for a
172 : generous margin, and this number not particularly fragile or
173 : sensitive, we can leave it static. */
174 0 : #define FD_REQLIM_CACHE_MAX (1<<20)
175 :
176 : /* static map from request type to metric array index */
177 : static uint metric_index[FD_REPAIR_KIND_ORPHAN + 1] = {
178 : [FD_REPAIR_KIND_PONG] = FD_METRICS_ENUM_REPAIR_SENT_REQUEST_TYPE_V_PONG_IDX,
179 : [FD_REPAIR_KIND_SHRED] = FD_METRICS_ENUM_REPAIR_SENT_REQUEST_TYPE_V_NEEDED_WINDOW_IDX,
180 : [FD_REPAIR_KIND_HIGHEST_SHRED] = FD_METRICS_ENUM_REPAIR_SENT_REQUEST_TYPE_V_NEEDED_HIGHEST_WINDOW_IDX,
181 : [FD_REPAIR_KIND_ORPHAN] = FD_METRICS_ENUM_REPAIR_SENT_REQUEST_TYPE_V_NEEDED_ORPHAN_IDX,
182 : };
183 :
184 : typedef union {
185 : struct {
186 : fd_wksp_t * mem;
187 : ulong chunk0;
188 : ulong wmark;
189 : ulong mtu;
190 : };
191 : fd_net_rx_bounds_t net_rx;
192 : } in_ctx_t;
193 :
194 : struct out_ctx {
195 : ulong idx;
196 : fd_wksp_t * mem;
197 : ulong chunk0;
198 : ulong wmark;
199 : ulong chunk;
200 :
201 : /* Repair tile directly tracks credit outside of stem for these
202 : asynchronous sign links. In particular, credits tracks the RETURN
203 : sign_repair link. This is because repair_sign and
204 : sign_repair are unreliable. If both links were reliable, and the
205 : links filled completely, stem would get into a deadlock. Neither
206 : repair or sign would have credits, which would prevent frags from
207 : getting polled in repair or sign, which would prevent any credits
208 : from getting returned back to the tiles. So the sign_repair return
209 : link must be unreliable. credits / max_credits are used by the
210 : repair_sign link, but credits tracks the RETURN
211 : sign_repair link.
212 :
213 : Consider the scenario:
214 :
215 : repair_sign (depth 128) sign_repair (depth 128)
216 : repair ----------------------> sign ------------------------> repair
217 : [rest free, r130, r129] [r128, r127, ... , r1] (full)
218 :
219 : If repair is publishing too many requests too fast(common in
220 : catchup), and not polling enough frags from sign, without manual
221 : management the sign_repair link would be overrun. Nothing is
222 : stopping repair from publishing more requests, because sign is
223 : functioning fast enough to handle the requests. However, nothing is
224 : stopping sign from polling the next request and signing it, and
225 : PUBLISHING it on the sign_repair link that is already full, because
226 : the sign_repair link is unreliable.
227 :
228 : This is why we need to manually track credits for the sign_repair
229 : link. We must ensure that there are never more than 128 items in
230 : the ENTIRE repair_sign -> sign tile -> sign_repair work queue, else
231 : there is always a possibility of an overrun in the sign_repair
232 : link.
233 :
234 : We can furthermore ensure some nice properties by having the
235 : repair_sign link have a greater depth than the sign_repair link.
236 : This way, we exclusively use manual credit management to control
237 : the rate at which we publish requests to sign. This allows for
238 : repair_sign to also be unreliable. Even when the repair sign link
239 : is "full", we can avoid backpressure and continue polling frags,
240 : without overruning the sign_repair link.
241 :
242 : To lose a frag to overrun isn't necessarily critical, but in
243 : general the repair tile relies on the fact that a signing task
244 : published to sign tile will always come back. If we lose a frag to
245 : overrun, then there will be an entry in the pending signs structure
246 : that is never removed, and theoretically the map could fill up.
247 : Conceptually, with a reliable (unreliable links, but strictly
248 : controlled count-per-link) sign->repair->sign structure, there
249 : should be no eviction needed in this pending signs structure. */
250 :
251 : ulong in_idx; /* index of the incoming link */
252 : ulong credits; /* available credits for link */
253 : ulong max_credits; /* maximum credits (depth) */
254 : };
255 : typedef struct out_ctx out_ctx_t;
256 :
257 : /* Data needed to sign and send a pong that is not contained in the
258 : pong msg itself. */
259 :
260 : struct pong_data {
261 : fd_ip4_port_t peer_addr;
262 : fd_hash_t hash;
263 : uint daddr;
264 : fd_pubkey_t key;
265 : };
266 : typedef struct pong_data pong_data_t;
267 :
268 : struct sign_req {
269 : ulong key; /* map key, ctx->pending_key_next */
270 : ulong buflen;
271 : union {
272 : uchar buf[sizeof(fd_repair_msg_t)];
273 : fd_repair_msg_t msg;
274 : };
275 : pong_data_t pong_data; /* populated only for pong msgs */
276 : };
277 : typedef struct sign_req sign_req_t;
278 :
279 : #define MAP_NAME fd_signs_map
280 0 : #define MAP_KEY key
281 0 : #define MAP_KEY_NULL ULONG_MAX
282 0 : #define MAP_KEY_INVAL(k) (k==ULONG_MAX)
283 0 : #define MAP_T sign_req_t
284 : #define MAP_MEMOIZE 0
285 : #include "../../util/tmpl/fd_map_dynamic.c"
286 :
287 : /* Because the sign tiles could be all busy when a contact info or a
288 : ping arrives, we need to save ping messages to be signed in a queue
289 : and dispatched in after_credit when there are sign tiles available.
290 : The size of the queue is sized to be the number of warm up
291 : requests we might burst to the queue all at once (at most
292 : FD_REPAIR_PEER_MAX), then doubled for good measure.
293 :
294 : There is a possibility that someone could spam pings to block other
295 : peers' pings (and prevent us from responding to those pings). To
296 : mitigate this, we track the number of pings currently living in the
297 : sign queue that belong to each peer. If a peer already has a pong
298 : living in the sign queue, we drop the pings from that peer.
299 :
300 : The peer could send us a new bogus ping every time we pop their ping
301 : from the sign queue, but there would be no way to prevent other
302 : peers' pings from getting processed, so the wasted work and impact
303 : would be minimal.
304 :
305 : Typical flow is that a pong will get added to the pong_queue during
306 : an after_frag call. Then on the following after_credit will get
307 : popped from the sign_queue and added to sign_map, and then dispatched
308 : to the sign tile.
309 :
310 : Note that after the first turbine shred arrives, the signs_queue also
311 : stores highest window index requests for slots between snapshot and
312 : turbine_slot0, which are dispatched first before any other requests
313 : as a catchup optimization. This doesn't break any of the inflight
314 : invariants as highest window index requests do not get added to the
315 : inflight table. */
316 :
317 : struct sign_pending {
318 : fd_repair_msg_t msg;
319 : pong_data_t pong_data; /* populated only for pong msgs */
320 : };
321 : typedef struct sign_pending sign_pending_t;
322 :
323 : #define QUEUE_NAME fd_signs_queue
324 0 : #define QUEUE_T sign_pending_t
325 0 : #define QUEUE_MAX (2*FD_REPAIR_PEER_MAX)
326 : #include "../../util/tmpl/fd_queue.c"
327 :
328 : struct ctx {
329 : long tsdebug; /* timestamp for debug printing */
330 :
331 : fd_clock_tile_t clock[1];
332 :
333 : ulong repair_seed;
334 :
335 : fd_keyswitch_t * keyswitch;
336 : int halt_signing;
337 :
338 : fd_ip4_port_t repair_intake_addr;
339 :
340 : fd_forest_t * forest;
341 : fd_policy_t * policy;
342 : fd_reqlim_t * dedup;
343 : fd_inflights_t * inflights;
344 : fd_repair_t * protocol;
345 :
346 : ulong enforce_fixed_fec_set; /* min slot where the feature is enforced */
347 :
348 : fd_pubkey_t identity_public_key;
349 :
350 : fd_wksp_t * wksp;
351 :
352 : fd_stem_context_t * stem;
353 :
354 : uchar in_kind[ MAX_IN_LINKS ];
355 : in_ctx_t in_links[ MAX_IN_LINKS ];
356 :
357 : int skip_frag;
358 :
359 : out_ctx_t net_out_ctx[1];
360 :
361 : out_ctx_t repair_out_ctx[1];
362 :
363 : /* repair_sign links (to sign tiles 1+) - for round-robin distribution */
364 :
365 : ulong repair_sign_cnt;
366 : out_ctx_t repair_sign_out_ctx[ MAX_SIGN_TILE_CNT ];
367 :
368 : ulong sign_rrobin_idx;
369 :
370 : /* Pending sign requests for async operations */
371 :
372 : uint pending_key_next;
373 : sign_req_t * signs_map; /* contains any request currently in the repair->sign or sign->repair dcache */
374 : sign_pending_t * pong_queue; /* contains any pong or initial warmup request waiting to be dispatched to repair->sign. Size is 2*FD_REPAIR_PEER_MAX */
375 :
376 : ushort net_id;
377 :
378 : /* Buffers for incoming unreliable frags */
379 : uchar net_buf[ FD_NET_MTU ];
380 : uchar sign_buf[ sizeof(fd_ed25519_sig_t) ];
381 :
382 : /* Store chunk for incoming reliable frags */
383 : ulong chunk;
384 : ulong snap_out_chunk; /* store second to last chunk for snap_out */
385 :
386 : fd_ip4_udp_hdrs_t intake_hdr[1];
387 :
388 : fd_rnonce_ss_t repair_nonce_ss[1];
389 :
390 : ulong manifest_slot;
391 : struct {
392 : ulong send_pkt_cnt;
393 : ulong sent_pkt_types[FD_METRICS_ENUM_REPAIR_SENT_REQUEST_TYPE_CNT];
394 : ulong current_slot;
395 : ulong old_shred;
396 : ulong last_requested_slot;
397 : ulong last_requested_orphan;
398 : ulong sign_tile_unavail;
399 : ulong rerequest;
400 : ulong malformed_ping;
401 : ulong unknown_peer_ping;
402 : ulong fail_sigverify_ping;
403 : fd_histf_t slot_compl_time[ 1 ];
404 : fd_histf_t response_latency[ 1 ];
405 : ulong blk_evicted;
406 : ulong blk_failed_insert;
407 :
408 : ulong slot_evicted;
409 : ulong slot_evicted_by;
410 : ulong slot_failed_insert;
411 :
412 : ulong failed_chain_verify_cnt;
413 : ulong failed_chain_verify_slot;
414 : } metrics[ 1 ];
415 :
416 : /* Slot-level metrics */
417 :
418 : fd_repair_metrics_t * slot_metrics;
419 : ulong turbine_slot0; // catchup considered complete after this slot
420 : };
421 : typedef struct ctx ctx_t;
422 :
423 : FD_FN_CONST static inline ulong
424 0 : scratch_align( void ) {
425 0 : return 128UL;
426 0 : }
427 :
428 : FD_FN_PURE static inline ulong
429 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
430 0 : ulong total_sign_depth = tile->repair.repair_sign_depth * tile->repair.repair_sign_cnt;
431 0 : int lg_sign_depth = fd_ulong_find_msb( fd_ulong_pow2_up(total_sign_depth) ) + 1;
432 :
433 0 : ulong l = FD_LAYOUT_INIT;
434 0 : l = FD_LAYOUT_APPEND( l, alignof(ctx_t), sizeof(ctx_t) );
435 0 : l = FD_LAYOUT_APPEND( l, fd_repair_align(), fd_repair_footprint () );
436 0 : l = FD_LAYOUT_APPEND( l, fd_forest_align(), fd_forest_footprint ( tile->repair.slot_max ) );
437 0 : l = FD_LAYOUT_APPEND( l, fd_policy_align(), fd_policy_footprint ( FD_REPAIR_PEER_MAX ) );
438 0 : l = FD_LAYOUT_APPEND( l, fd_reqlim_align(), fd_reqlim_footprint ( FD_REQLIM_CACHE_MAX ) );
439 0 : l = FD_LAYOUT_APPEND( l, fd_inflights_align(), fd_inflights_footprint () );
440 0 : l = FD_LAYOUT_APPEND( l, fd_signs_map_align(), fd_signs_map_footprint ( lg_sign_depth ) );
441 0 : l = FD_LAYOUT_APPEND( l, fd_signs_queue_align(), fd_signs_queue_footprint() );
442 0 : l = FD_LAYOUT_APPEND( l, fd_repair_metrics_align(), fd_repair_metrics_footprint() );
443 0 : return FD_LAYOUT_FINI( l, scratch_align() );
444 0 : }
445 :
446 : /* Below functions manage the current pending sign requests. */
447 :
448 : static sign_req_t *
449 : sign_map_insert( ctx_t * ctx,
450 : fd_repair_msg_t const * msg,
451 0 : pong_data_t const * opt_pong_data ) {
452 :
453 : /* Check if there is any space for a new pending sign request. Should never fail as long as credit management is working. */
454 0 : if( FD_UNLIKELY( fd_signs_map_key_cnt( ctx->signs_map )==fd_signs_map_key_max( ctx->signs_map ) ) ) return NULL;
455 :
456 0 : sign_req_t * pending = fd_signs_map_insert( ctx->signs_map, ctx->pending_key_next++ );
457 0 : if( FD_UNLIKELY( !pending ) ) return NULL; /* Not possible, unless the same key is used twice. */
458 0 : pending->msg = *msg;
459 0 : pending->buflen = fd_repair_sz( msg );
460 0 : if( FD_UNLIKELY( opt_pong_data ) ) pending->pong_data = *opt_pong_data;
461 0 : return pending;
462 0 : }
463 :
464 : static int
465 : sign_map_remove( ctx_t * ctx,
466 0 : ulong key ) {
467 0 : sign_req_t * pending = fd_signs_map_query( ctx->signs_map, key, NULL );
468 0 : if( FD_UNLIKELY( !pending ) ) return -1;
469 0 : fd_signs_map_remove( ctx->signs_map, pending );
470 0 : return 0;
471 0 : }
472 :
473 : static void
474 : send_packet( ctx_t * ctx,
475 : fd_stem_context_t * stem,
476 : uint dst_ip_addr,
477 : ushort dst_port,
478 : uint src_ip_addr,
479 : uchar const * payload,
480 : ulong payload_sz,
481 0 : ulong tsorig ) {
482 0 : ctx->metrics->send_pkt_cnt++;
483 0 : uchar * packet = fd_chunk_to_laddr( ctx->net_out_ctx->mem, ctx->net_out_ctx->chunk );
484 0 : fd_ip4_udp_hdrs_t * hdr = (fd_ip4_udp_hdrs_t *)packet;
485 0 : *hdr = *ctx->intake_hdr;
486 :
487 0 : fd_ip4_hdr_t * ip4 = hdr->ip4;
488 0 : ip4->saddr = src_ip_addr;
489 0 : ip4->daddr = dst_ip_addr;
490 0 : ip4->net_id = fd_ushort_bswap( ctx->net_id++ );
491 0 : ip4->check = 0U;
492 0 : ip4->net_tot_len = fd_ushort_bswap( (ushort)(payload_sz + sizeof(fd_ip4_hdr_t)+sizeof(fd_udp_hdr_t)) );
493 0 : ip4->check = fd_ip4_hdr_check_fast( ip4 );
494 :
495 0 : fd_udp_hdr_t * udp = hdr->udp;
496 0 : udp->net_dport = dst_port;
497 0 : udp->net_len = fd_ushort_bswap( (ushort)(payload_sz + sizeof(fd_udp_hdr_t)) );
498 0 : fd_memcpy( packet+sizeof(fd_ip4_udp_hdrs_t), payload, payload_sz );
499 0 : hdr->udp->check = 0U;
500 :
501 0 : ulong tspub = fd_frag_meta_ts_comp( fd_tickcount() );
502 0 : ulong sig = fd_disco_netmux_sig( dst_ip_addr, dst_port, dst_ip_addr, DST_PROTO_OUTGOING, sizeof(fd_ip4_udp_hdrs_t) );
503 0 : ulong packet_sz = payload_sz + sizeof(fd_ip4_udp_hdrs_t);
504 0 : ulong chunk = ctx->net_out_ctx->chunk;
505 0 : fd_stem_publish( stem, ctx->net_out_ctx->idx, sig, chunk, packet_sz, 0UL, tsorig, tspub );
506 0 : ctx->net_out_ctx->chunk = fd_dcache_compact_next( chunk, packet_sz, ctx->net_out_ctx->chunk0, ctx->net_out_ctx->wmark );
507 0 : }
508 :
509 : /* Returns a sign_out context with max available credits.
510 : If no sign_out context has available credits, returns NULL. */
511 : static out_ctx_t *
512 0 : sign_avail_credits( ctx_t * ctx ) {
513 0 : out_ctx_t * sign_out = NULL;
514 0 : ulong max_credits = 0;
515 0 : for( uint i = 0; i < ctx->repair_sign_cnt; i++ ) {
516 0 : if( ctx->repair_sign_out_ctx[i].credits > max_credits ) {
517 0 : max_credits = ctx->repair_sign_out_ctx[i].credits;
518 0 : sign_out = &ctx->repair_sign_out_ctx[i];
519 0 : }
520 0 : }
521 0 : return sign_out;
522 0 : }
523 :
524 : /* Prepares the signing preimage and publishes a signing request that
525 : will be signed asynchronously by the sign tile. The signed data will
526 : be returned via dcache as a frag. */
527 : static void
528 : fd_repair_send_sign_request( ctx_t * ctx,
529 : out_ctx_t * sign_out,
530 : fd_repair_msg_t const * msg,
531 0 : pong_data_t const * opt_pong_data ) {
532 :
533 0 : if( FD_UNLIKELY( ctx->halt_signing ) ) FD_LOG_CRIT(( "can't dispatch sign requests while halting signing" ));
534 :
535 : /* New sign request */
536 0 : sign_req_t * pending = sign_map_insert( ctx, msg, opt_pong_data );
537 0 : if( FD_UNLIKELY( !pending ) ) return;
538 :
539 0 : ulong sig = 0;
540 0 : ulong preimage_sz = 0;
541 0 : uchar * dst = fd_chunk_to_laddr( sign_out->mem, sign_out->chunk );
542 :
543 0 : if( FD_UNLIKELY( msg->kind == FD_REPAIR_KIND_PONG ) ) {
544 0 : uchar pre_image[FD_REPAIR_PONG_PREIMAGE_SZ];
545 0 : preimage_pong( &opt_pong_data->hash, pre_image );
546 0 : preimage_sz = FD_REPAIR_PONG_PREIMAGE_SZ;
547 0 : fd_memcpy( dst, pre_image, preimage_sz );
548 0 : sig = ((ulong)pending->key << 32) | (uint)FD_KEYGUARD_SIGN_TYPE_SHA256_ED25519;
549 0 : } else {
550 : /* Sign and prepare the message directly into the pending buffer */
551 0 : uchar * preimage = preimage_req( &pending->msg, &preimage_sz );
552 0 : fd_memcpy( dst, preimage, preimage_sz );
553 0 : sig = ((ulong)pending->key << 32) | (uint)FD_KEYGUARD_SIGN_TYPE_ED25519;
554 0 : }
555 :
556 0 : fd_stem_publish( ctx->stem, sign_out->idx, sig, sign_out->chunk, preimage_sz, 0UL, 0UL, 0UL );
557 0 : sign_out->chunk = fd_dcache_compact_next( sign_out->chunk, preimage_sz, sign_out->chunk0, sign_out->wmark );
558 :
559 0 : ctx->metrics->sent_pkt_types[metric_index[msg->kind]]++;
560 0 : sign_out->credits--;
561 0 : }
562 :
563 : static inline int
564 : before_frag( ctx_t * ctx,
565 : ulong in_idx,
566 : ulong seq FD_PARAM_UNUSED,
567 0 : ulong sig ) {
568 0 : uint in_kind = ctx->in_kind[ in_idx ];
569 0 : if( FD_LIKELY ( in_kind==IN_KIND_NET ) ) return fd_disco_netmux_sig_proto( sig )!=DST_PROTO_REPAIR;
570 0 : if( FD_UNLIKELY( in_kind==IN_KIND_SHRED ) ) return fd_int_if( fd_forest_root_slot( ctx->forest )==ULONG_MAX, -1, 0 ); /* not ready to read frag */
571 0 : if( FD_UNLIKELY( in_kind==IN_KIND_GOSSIP ) ) {
572 0 : return sig!=FD_GOSSIP_UPDATE_TAG_CONTACT_INFO &&
573 0 : sig!=FD_GOSSIP_UPDATE_TAG_CONTACT_INFO_REMOVE;
574 0 : }
575 0 : if( FD_UNLIKELY( in_kind==IN_KIND_REPLAY ) ) return sig!=REPLAY_SIG_REASM_EVICTED;
576 0 : return 0;
577 0 : }
578 :
579 : static void
580 : during_frag( ctx_t * ctx,
581 : ulong in_idx,
582 : ulong seq FD_PARAM_UNUSED,
583 : ulong sig,
584 : ulong chunk,
585 : ulong sz,
586 0 : ulong ctl ) {
587 0 : ctx->skip_frag = 0;
588 :
589 0 : uint in_kind = ctx->in_kind[ in_idx ];
590 0 : in_ctx_t const * in_ctx = &ctx->in_links[ in_idx ];
591 0 : ctx->chunk = chunk;
592 :
593 0 : if( FD_UNLIKELY( in_kind==IN_KIND_NET ) ) {
594 0 : ulong hdr_sz = fd_disco_netmux_sig_hdr_sz( sig );
595 0 : FD_TEST( hdr_sz <= sz ); /* Should be ensured by the net tile */
596 0 : uchar const * dcache_entry = fd_net_rx_translate_frag( &in_ctx->net_rx, chunk, ctl, sz );
597 0 : fd_memcpy( ctx->net_buf, dcache_entry, sz );
598 0 : return;
599 0 : }
600 :
601 0 : if( FD_UNLIKELY( in_kind==IN_KIND_GENESIS ) ) {
602 0 : FD_TEST( sizeof(fd_genesis_meta_t)<=sig );
603 0 : return;
604 0 : }
605 :
606 0 : if( FD_UNLIKELY( sz!=0UL && ( chunk<in_ctx->chunk0 || chunk>in_ctx->wmark || sz>in_ctx->mtu ) ) )
607 0 : FD_LOG_ERR(( "chunk %lu %lu corrupt, not in range [%lu,%lu] in kind %u", chunk, sz, in_ctx->chunk0, in_ctx->wmark, in_kind ));
608 :
609 0 : if( FD_UNLIKELY( in_kind==IN_KIND_SNAP ) ) {
610 0 : if( FD_UNLIKELY( fd_ssmsg_sig_message( sig )!=FD_SSMSG_DONE ) ) ctx->snap_out_chunk = chunk;
611 0 : return;
612 0 : }
613 :
614 0 : if( FD_UNLIKELY( in_kind==IN_KIND_SIGN ) ) {
615 : /* sign_repair is unreliable, so we copy the frag for convention.
616 : Theoretically impossible to overrun. */
617 0 : uchar const * dcache_entry = fd_chunk_to_laddr_const( in_ctx->mem, chunk );
618 0 : fd_memcpy( ctx->sign_buf, dcache_entry, sz );
619 0 : return;
620 0 : }
621 0 : }
622 :
623 : static inline void
624 : after_snap( ctx_t * ctx,
625 : ulong sig,
626 0 : uchar const * chunk ) {
627 0 : if( FD_UNLIKELY( fd_ssmsg_sig_message( sig )!=FD_SSMSG_DONE ) ) return;
628 0 : fd_snapshot_manifest_t * manifest = (fd_snapshot_manifest_t *)chunk;
629 :
630 0 : fd_forest_init( ctx->forest, manifest->slot );
631 0 : }
632 :
633 : static inline void
634 0 : after_gossip( ctx_t * ctx, fd_gossip_update_message_t const * msg, ulong sig ) {
635 0 : switch( sig ) {
636 0 : case FD_GOSSIP_UPDATE_TAG_CONTACT_INFO_REMOVE: {
637 0 : fd_policy_peer_remove( ctx->policy, fd_type_pun_const( msg->origin ) );
638 0 : break;
639 0 : }
640 0 : case FD_GOSSIP_UPDATE_TAG_CONTACT_INFO: {
641 0 : fd_gossip_contact_info_t const * contact_info = msg->contact_info->value;
642 0 : fd_ip4_port_t repair_peer;
643 0 : repair_peer.addr = contact_info->sockets[ FD_GOSSIP_CONTACT_INFO_SOCKET_SERVE_REPAIR ].is_ipv6 ? 0U : contact_info->sockets[ FD_GOSSIP_CONTACT_INFO_SOCKET_SERVE_REPAIR ].ip4;
644 0 : repair_peer.port = contact_info->sockets[ FD_GOSSIP_CONTACT_INFO_SOCKET_SERVE_REPAIR ].port;
645 0 : if( FD_UNLIKELY( !repair_peer.addr || !repair_peer.port ) ) return;
646 0 : fd_policy_peer_t const * peer = fd_policy_peer_upsert( ctx->policy, fd_type_pun_const( msg->origin ), &repair_peer );
647 0 : if( FD_LIKELY( peer && !fd_signs_queue_full( ctx->pong_queue ) ) ) {
648 : /* The repair process uses a Ping-Pong protocol that incurs one
649 : round-trip time (RTT) for the initmial repair request. To
650 : optimize this, we proactively send a placeholder repair request
651 : as soon as we receive a peer's contact information for the first
652 : time, effectively prepaying the RTT cost. */
653 0 : fd_repair_msg_t * init = fd_repair_shred( ctx->protocol, fd_type_pun_const( msg->origin ), (ulong)fd_clock_tile_now( ctx->clock )/1000000UL, 0, 0, 0 );
654 0 : fd_signs_queue_push( ctx->pong_queue, (sign_pending_t){ .msg = *init } );
655 0 : }
656 0 : break;
657 0 : }
658 0 : default: FD_LOG_ERR(( "bad gossip sig %lu", sig ));
659 0 : }
660 0 : }
661 :
662 : static inline void
663 : after_sign( ctx_t * ctx,
664 : ulong in_idx,
665 : ulong sig,
666 0 : fd_stem_context_t * stem ) {
667 0 : ulong pending_key = sig >> 32;
668 : /* Look up the pending request. Since the repair_sign links are
669 : reliable, the incoming sign_repair fragments represent a complete
670 : set of the previously sent outgoing messages. However, with
671 : multiple sign tiles, the responses may arrive interleaved. */
672 :
673 : /* Find which sign tile sent this response and increment its credits */
674 0 : for( uint i = 0; i < ctx->repair_sign_cnt; i++ ) {
675 0 : if( ctx->repair_sign_out_ctx[i].in_idx == in_idx ) {
676 0 : if( FD_LIKELY( ctx->repair_sign_out_ctx[i].credits < ctx->repair_sign_out_ctx[i].max_credits ) ) ctx->repair_sign_out_ctx[i].credits++;
677 0 : break;
678 0 : }
679 0 : }
680 :
681 0 : sign_req_t * pending_ = fd_signs_map_query( ctx->signs_map, pending_key, NULL );
682 0 : if( FD_UNLIKELY( !pending_ ) ) FD_LOG_CRIT(( "No pending request found for key %lu", pending_key )); /* implies either bad programmer error or something happened with sign tile */
683 :
684 0 : sign_req_t pending[1] = { *pending_ }; /* Make a copy of the pending request so we can sign_map_remove immediately. */
685 0 : sign_map_remove( ctx, pending_key );
686 :
687 : /* This is a pong message */
688 0 : if( FD_UNLIKELY( pending->msg.kind == FD_REPAIR_KIND_PONG ) ) {
689 0 : fd_policy_peer_t * peer = fd_policy_peer_query( ctx->policy, &pending->pong_data.key );
690 0 : if( FD_LIKELY( peer && peer->ping ) ) peer->ping--; /* prevent underflow if the peer was removed/readded */
691 :
692 0 : fd_memcpy( pending->msg.pong.sig, ctx->sign_buf, 64UL );
693 0 : send_packet( ctx, stem, pending->pong_data.peer_addr.addr, pending->pong_data.peer_addr.port, pending->pong_data.daddr, pending->buf, fd_repair_sz( &pending->msg ), fd_frag_meta_ts_comp( fd_tickcount() ) );
694 0 : return;
695 0 : }
696 :
697 : /* Inject the signature into the pending request */
698 0 : fd_memcpy( pending->buf + 4, ctx->sign_buf, 64UL );
699 0 : uint src_ip4 = 0U;
700 :
701 : /* This is a warmup message */
702 0 : if( FD_UNLIKELY( pending->msg.kind == FD_REPAIR_KIND_SHRED && pending->msg.shred.slot == 0 ) ) {
703 0 : fd_policy_peer_t * peer = fd_policy_peer_query( ctx->policy, &pending->msg.shred.to );
704 0 : if( FD_UNLIKELY( peer ) ) send_packet( ctx, stem, peer->ip4, peer->port, src_ip4, pending->buf, pending->buflen, fd_frag_meta_ts_comp( fd_tickcount() ) );
705 0 : else { /* This is a warmup request for a peer that is no longer active. There's no reason to pick another peer for a warmup rq, so just drop it. */ }
706 0 : return;
707 0 : }
708 :
709 : /* This is a regular repair shred request
710 :
711 : We need to ensure we always send out any shred requests we have,
712 : because policy_next has no way to revisit a shred. But the fact
713 : that peers can drop out of the peer list makes this complicated.
714 : If the peer is still there (common), it's fine. If the peer is not
715 : there, we can add this request to the inflights table, pretend
716 : we've sent it and let the inflight timeout request it down the
717 : line. */
718 :
719 0 : fd_policy_peer_t * active = fd_policy_peer_query( ctx->policy, &pending->msg.shred.to );
720 0 : if( FD_UNLIKELY( !active ) ) {
721 : /* Already added to the inflights table, pretend we've sent it
722 : and let the inflight timeout request it down the line. */
723 0 : return;
724 0 : }
725 : /* Happy path - all is well, our peer didn't drop out from beneath us. */
726 0 : if( FD_UNLIKELY( pending->msg.kind == FD_REPAIR_KIND_ORPHAN ) ) ctx->metrics->last_requested_orphan = pending->msg.orphan.slot;
727 0 : else ctx->metrics->last_requested_slot = pending->msg.shred.slot;
728 :
729 0 : send_packet( ctx, stem, active->ip4, active->port, src_ip4, pending->buf, pending->buflen, fd_frag_meta_ts_comp( fd_tickcount() ) );
730 0 : }
731 :
732 : static int
733 0 : blk_insert_check( ctx_t * ctx, fd_forest_blk_t * new_blk, ulong new_slot, ulong evicted ) {
734 0 : if( FD_UNLIKELY( !new_blk ) ) {
735 0 : ctx->metrics->blk_failed_insert++;
736 0 : ctx->metrics->slot_failed_insert = new_slot;
737 0 : return 0;
738 0 : } else {
739 0 : if( FD_UNLIKELY( evicted != ULONG_MAX ) ) {
740 0 : ctx->metrics->blk_evicted++;
741 0 : ctx->metrics->slot_evicted = evicted;
742 0 : ctx->metrics->slot_evicted_by = new_slot;
743 0 : }
744 0 : return 1;
745 0 : }
746 0 : }
747 :
748 0 : static inline int shred_src( ulong sig ) {
749 0 : uint sig_src = fd_shred_sig_src( sig );
750 0 : switch( sig_src ) {
751 0 : case SHRED_SIG_SRC_TURBINE:
752 0 : return SHRED_SRC_TURBINE;
753 0 : case SHRED_SIG_SRC_RECONSTRUCTED:
754 0 : return SHRED_SRC_RECOVERED;
755 0 : case SHRED_SIG_SRC_REPAIR:
756 0 : return SHRED_SRC_REPAIR;
757 0 : case SHRED_SIG_SRC_BAD_REPAIR:
758 0 : return SHRED_SRC_TURBINE; /* like fec_resolver, treat bad repair shreds as turbine shreds */
759 0 : case SHRED_SIG_SRC_LEADER:
760 0 : return SHRED_SRC_LEADER;
761 0 : default: FD_LOG_CRIT(( "bad shred sig src %u", sig_src ));
762 0 : }
763 0 : }
764 :
765 : static inline void
766 : after_shred( ctx_t * ctx,
767 : ulong sig,
768 : fd_shred_t * shred,
769 : ulong nonce,
770 : fd_hash_t * mr,
771 0 : fd_hash_t * cmr ) {
772 :
773 : /* we don't want to add a slot to the forest that chains to a slot
774 : older than root, to avoid filling forest up with junk.
775 : Especially if we are close to full and we are having trouble
776 : rooting, we can't rely on publishing to prune these useless
777 : subtrees. TODO: do the same with reasm/store/shred? */
778 :
779 0 : if( FD_UNLIKELY( shred->slot <= fd_forest_root_slot( ctx->forest ) ||
780 0 : shred->slot - shred->data.parent_off < fd_forest_root_slot( ctx->forest ) ) ) {
781 0 : ctx->metrics->old_shred++;
782 0 : return;
783 0 : }
784 :
785 : /* Insert the shred sig (shared by all shred members in the FEC set)
786 : into the map. */
787 0 : int is_code = fd_shred_is_code( fd_shred_type( shred->variant ) );
788 0 : int src = shred_src( sig );
789 :
790 0 : if( FD_LIKELY( !is_code ) ) {
791 0 : long rtt = 0;
792 0 : fd_pubkey_t peer;
793 :
794 0 : int slot_complete = !!(shred->data.flags & FD_SHRED_DATA_FLAG_SLOT_COMPLETE);
795 0 : int ref_tick = shred->data.flags & FD_SHRED_DATA_REF_TICK_MASK;
796 0 : ulong evicted = ULONG_MAX;
797 0 : fd_forest_blk_t * blk = fd_forest_blk_insert( ctx->forest, shred->slot, shred->slot - shred->data.parent_off, &evicted );
798 0 : if( FD_UNLIKELY( !blk_insert_check( ctx, blk, shred->slot, evicted ) ) ) return;
799 :
800 0 : if( FD_LIKELY( fd_forest_data_shred_insert( ctx->forest, shred->slot, shred->slot - shred->data.parent_off, shred->idx, shred->fec_set_idx, slot_complete, ref_tick, src, mr, cmr ) ) ) {
801 0 : if( FD_UNLIKELY( src == SHRED_SRC_REPAIR && ( rtt = fd_inflights_request_match( ctx->inflights, nonce, shred->slot, shred->idx, &peer ) ) > 0 ) ) {
802 0 : fd_policy_peer_response_update( ctx->policy, &peer, rtt );
803 0 : fd_histf_sample( ctx->metrics->response_latency, (ulong)rtt );
804 0 : }
805 0 : }
806 0 : } else {
807 0 : fd_forest_code_shred_insert( ctx->forest, shred->slot, shred->idx );
808 0 : }
809 0 : }
810 :
811 : /* Kicks off the chained merkle verification starting at a slot with
812 : a confirmed, canonical block_id. Either finishes successfully and
813 : returns early, or detects an incorrect FEC set and clears it. In
814 : this case the verification is paused and state is saved at where
815 : it left off. Verification can be re-triggered in after_fec as well. */
816 : static inline void
817 : check_confirmed( ctx_t * ctx,
818 : fd_forest_blk_t * blk,
819 0 : fd_hash_t const * confirmed_bid ) {
820 :
821 0 : if( FD_LIKELY( !blk->chain_confirmed && blk->complete_idx != UINT_MAX ) ) {
822 : /* The above conditions say that all the shreds of the block have arrived. */
823 0 : fd_forest_blk_t * bad_blk = fd_forest_fec_chain_verify( ctx->forest, blk, confirmed_bid );
824 0 : if( FD_LIKELY( !bad_blk ) ) {
825 : /* chain verified successfully from blk to as far as we have fec data */
826 0 : return;
827 0 : }
828 :
829 0 : uint bad_fec_idx = fd_forest_merkle_last_incorrect_idx( bad_blk );
830 : /* bad_fec_idx UINT_MAX implies this slot is fully validated. Not
831 : possible here because that implies bad_blk is NULL and would
832 : early exit above. */
833 0 : FD_TEST( bad_fec_idx != UINT_MAX );
834 :
835 0 : fd_hash_t const * expected = (bad_fec_idx == bad_blk->complete_idx - (FD_FEC_SHRED_CNT - 1)) ? &bad_blk->confirmed_bid : &bad_blk->merkle_roots[(bad_fec_idx / 32) + 1].cmr;
836 :
837 0 : FD_BASE58_ENCODE_32_BYTES( confirmed_bid->uc, confirmed_bid_b58 );
838 0 : FD_BASE58_ENCODE_32_BYTES( expected->uc, expected_mr );
839 0 : FD_BASE58_ENCODE_32_BYTES( bad_blk->merkle_roots[bad_fec_idx / 32].mr.uc, recorded_mr );
840 :
841 0 : FD_LOG_WARNING(( "[%s] slot %lu block_id %s confirmation detected incorrect FECs. bad FEC is slot %lu fec set %u. expected mr (%s) != recorded mr (%s)",
842 0 : __func__,
843 0 : blk->slot,
844 0 : confirmed_bid_b58,
845 0 : bad_blk->slot,
846 0 : bad_fec_idx,
847 0 : expected_mr,
848 0 : recorded_mr ));
849 :
850 0 : ctx->metrics->failed_chain_verify_cnt++;
851 0 : ctx->metrics->failed_chain_verify_slot = bad_blk->slot;
852 :
853 : /* If we have a bad block, we need to dump and repair backwards from
854 : the point where the merkle root is incorrect.
855 : We start only by dumping the last incorrect FEC. It's possible that
856 : this is the only incorrect one. If it isn't though, when the slot
857 : recompletes, this function will trigger again and we will dump the
858 : second to last incorrect FEC. */
859 :
860 0 : fd_forest_fec_clear( ctx->forest, bad_blk->slot, bad_fec_idx, FD_FEC_SHRED_CNT - 1 );
861 0 : }
862 0 : }
863 :
864 : /* Returns 1 if the fec is guaranteed invalid, 0 otherwise. */
865 : static inline int
866 : after_fec( ctx_t * ctx,
867 : fd_shred_t * shred,
868 : fd_hash_t * mr,
869 0 : fd_hash_t * cmr ) {
870 :
871 : /* When this is a FEC completes msg, it is implied that all the
872 : other shreds in the FEC set can also be inserted. Shred inserts
873 : into the forest are idempotent so it is fine to insert the same
874 : shred multiple times. */
875 :
876 0 : int slot_complete = !!( shred->data.flags & FD_SHRED_DATA_FLAG_SLOT_COMPLETE );
877 0 : int ref_tick = shred->data.flags & FD_SHRED_DATA_REF_TICK_MASK;
878 :
879 : /* Similar to after_shred, do not insert a slot that chains to a slot older than root */
880 0 : if( FD_UNLIKELY( shred->slot <= fd_forest_root_slot( ctx->forest ) ||
881 0 : shred->slot - shred->data.parent_off < fd_forest_root_slot( ctx->forest ) ) ) return 0;
882 :
883 0 : ulong evicted = ULONG_MAX;
884 0 : fd_forest_blk_t * ele = fd_forest_blk_insert( ctx->forest, shred->slot, shred->slot - shred->data.parent_off, &evicted );
885 0 : if( FD_UNLIKELY( !blk_insert_check( ctx, ele, shred->slot, evicted ) ) ) return 0;
886 0 : if( FD_UNLIKELY( !fd_forest_fec_insert( ctx->forest, shred->slot, shred->slot - shred->data.parent_off, shred->idx, shred->fec_set_idx, slot_complete, ref_tick, mr, cmr ) ) ) return 1;
887 :
888 : /* metrics for completed slots */
889 0 : if( FD_UNLIKELY( ele->complete_idx != UINT_MAX && ele->buffered_idx==ele->complete_idx ) ) {
890 0 : long now = fd_tickcount();
891 0 : long start_ts = ele->first_req_ts == 0 || ele->slot >= ctx->turbine_slot0 ? ele->first_shred_ts : ele->first_req_ts;
892 0 : ulong duration_ticks = (ulong)(now - start_ts);
893 0 : fd_histf_sample( ctx->metrics->slot_compl_time, duration_ticks );
894 0 : fd_repair_metrics_add_slot( ctx->slot_metrics, ele->slot, start_ts, now, ele->repair_cnt, ele->turbine_cnt );
895 : /* Note: this log does not imply that the slot is fully executable.
896 : It's possible that we have a slot that doesn't chain verify,
897 : which could be un-executable. */
898 0 : FD_BASE58_ENCODE_32_BYTES( ele->merkle_roots[ele->complete_idx / 32].mr.uc, block_id );
899 0 : FD_BASE58_ENCODE_32_BYTES( mr->uc, fec_mr );
900 0 : FD_LOG_INFO(( "[%s] slot is complete %lu. num_data_shreds: %u, num_repaired: %u, num_turbine: %u, num_recovered: %u, duration: %.2f ms. last recvd fec: %u, mr %s. current block_id: %s",
901 0 : __func__,
902 0 : ele->slot,
903 0 : ele->complete_idx + 1,
904 0 : ele->repair_cnt,
905 0 : ele->turbine_cnt,
906 0 : ele->recovered_cnt,
907 0 : (double)fd_metrics_convert_ticks_to_nanoseconds(duration_ticks) / 1e6,
908 0 : shred->fec_set_idx,
909 0 : fec_mr,
910 0 : block_id ));
911 0 : }
912 :
913 : /* re-trigger continuation of chained merkle verification if slot is
914 : complete. TODO MOVE TO AFTER_SHRED? */
915 0 : fd_hash_t empty_mr = (fd_hash_t){ .ul = { 0, 0, 0, 0 } };
916 0 : if( FD_UNLIKELY( ele->buffered_idx == ele->complete_idx && !fd_hash_eq( &ele->confirmed_bid, &empty_mr ) ) ) {
917 0 : check_confirmed( ctx, ele, &ele->confirmed_bid /* if lowest_verified_fec is not UINT_MAX, confirmed_bid must be populated */ );
918 0 : }
919 0 : return 0;
920 0 : }
921 :
922 : static inline void
923 : after_net( ctx_t * ctx,
924 0 : ulong sz ) {
925 0 : fd_eth_hdr_t * eth; fd_ip4_hdr_t * ip4; fd_udp_hdr_t * udp;
926 0 : uchar * data; ulong data_sz;
927 0 : if( FD_UNLIKELY( !fd_ip4_udp_hdr_strip( ctx->net_buf, sz, &data, &data_sz, ð, &ip4, &udp ) ) ) {
928 0 : ctx->metrics->malformed_ping++;
929 0 : return;
930 0 : }
931 0 : fd_ip4_port_t peer_addr = { .addr=ip4->saddr, .port=udp->net_sport };
932 :
933 0 : fd_repair_ping_t ping[1];
934 0 : int err = fd_repair_ping_de( ping, data, data_sz );
935 0 : if( FD_UNLIKELY( err ) ) {
936 0 : ctx->metrics->malformed_ping++;
937 0 : return;
938 0 : }
939 :
940 0 : fd_policy_peer_t * peer = fd_policy_peer_query( ctx->policy, &ping->ping.from );
941 0 : if( FD_UNLIKELY( !peer ) ) {
942 0 : ctx->metrics->unknown_peer_ping++;
943 0 : return;
944 0 : }
945 0 : if( FD_UNLIKELY( peer->ping ) ) return;
946 0 : if( FD_UNLIKELY( fd_signs_queue_full( ctx->pong_queue ) ) ) return;
947 :
948 0 : fd_sha512_t sha[1];
949 0 : if( FD_UNLIKELY( FD_ED25519_SUCCESS != fd_ed25519_verify( ping->ping.hash.uc, 32UL, ping->ping.sig, ping->ping.from.uc, sha ) ) ) {
950 0 : ctx->metrics->fail_sigverify_ping++;
951 0 : return;
952 0 : }
953 :
954 : /* Any gossip peer can send a ping, but they are bounded to at most
955 : one ping in the queue so they can't evict others' pings without
956 : multiple gossip identities. */
957 :
958 0 : fd_repair_msg_t * pong = fd_repair_pong( ctx->protocol, &ping->ping.hash );
959 0 : fd_signs_queue_push( ctx->pong_queue, (sign_pending_t){ .msg = *pong, .pong_data = { .peer_addr = peer_addr, .hash = ping->ping.hash, .daddr = ip4->daddr, .key = ping->ping.from } } );
960 0 : peer->ping++;
961 0 : }
962 :
963 : static inline void
964 : after_evict( ctx_t * ctx,
965 0 : fd_fec_evicted_t * evicted ) {
966 0 : fd_forest_fec_clear( ctx->forest, evicted->slot, evicted->fec_set_idx, FD_FEC_SHRED_CNT - 1 );
967 0 : }
968 :
969 : static inline void
970 : after_tower( ctx_t * ctx,
971 : ulong sig,
972 0 : uchar const * chunk ) {
973 :
974 0 : switch( sig ) {
975 0 : case FD_TOWER_SIG_SLOT_DONE: {
976 0 : fd_tower_slot_done_t const * msg = (fd_tower_slot_done_t const *)fd_type_pun_const( chunk );
977 0 : if( FD_LIKELY( msg->root_slot!=ULONG_MAX && msg->root_slot > fd_forest_root_slot( ctx->forest ) ) ) fd_forest_publish( ctx->forest, msg->root_slot );
978 0 : break;
979 0 : }
980 0 : case FD_TOWER_SIG_SLOT_CONFIRMED: {
981 0 : fd_tower_slot_confirmed_t const * msg = (fd_tower_slot_confirmed_t const *)fd_type_pun_const( chunk );
982 0 : if( msg->slot > fd_forest_root_slot( ctx->forest ) && (msg->level >= FD_TOWER_SLOT_CONFIRMED_DUPLICATE ) ) {
983 0 : fd_forest_blk_t * blk = fd_forest_query( ctx->forest, msg->slot );
984 0 : if( FD_UNLIKELY( !blk ) ) {
985 : /* If we receive a confirmation for a slot we don't have,
986 : create a sentinel forest block that we can repair from. */
987 0 : ulong evicted = ULONG_MAX;
988 0 : blk = fd_forest_blk_insert( ctx->forest, msg->slot, ULONG_MAX, &evicted );
989 0 : FD_LOG_INFO(("[%s] creating sentinel for duplicate confirmed block %lu", __func__, msg->slot));
990 0 : if( FD_UNLIKELY( !blk_insert_check( ctx, blk, msg->slot, evicted ) ) ) return;
991 0 : }
992 :
993 : /* Confirm the block */
994 0 : blk->confirmed_bid = msg->block_id;
995 0 : check_confirmed( ctx, blk, &msg->block_id );
996 0 : }
997 0 : break;
998 0 : }
999 0 : default: return;
1000 0 : }
1001 0 : }
1002 :
1003 : static void
1004 : after_frag( ctx_t * ctx,
1005 : ulong in_idx,
1006 : ulong seq FD_PARAM_UNUSED,
1007 : ulong sig,
1008 : ulong sz,
1009 : ulong tsorig FD_PARAM_UNUSED,
1010 : ulong tspub,
1011 0 : fd_stem_context_t * stem ) {
1012 0 : if( FD_UNLIKELY( ctx->skip_frag ) ) return;
1013 :
1014 0 : ctx->stem = stem;
1015 0 : in_ctx_t const * in_ctx = &ctx->in_links[ in_idx ];
1016 0 : uint in_kind = ctx->in_kind[ in_idx ];
1017 :
1018 0 : switch( in_kind ) {
1019 : /* Unreliable frags */
1020 0 : case IN_KIND_NET: {
1021 0 : after_net( ctx, sz );
1022 0 : break;
1023 0 : }
1024 0 : case IN_KIND_SIGN: {
1025 0 : after_sign( ctx, in_idx, sig, stem );
1026 0 : break;
1027 0 : }
1028 : /* Reliable frags read directly from dcache */
1029 0 : case IN_KIND_SNAP: {
1030 0 : after_snap( ctx, sig, fd_chunk_to_laddr( ctx->in_links[ in_idx ].mem, ctx->snap_out_chunk ) );
1031 0 : break;
1032 0 : }
1033 0 : case IN_KIND_GENESIS: {
1034 0 : fd_genesis_meta_t const * meta = (fd_genesis_meta_t const *)fd_type_pun_const( fd_chunk_to_laddr( in_ctx->mem, ctx->chunk ) );
1035 0 : if( meta->bootstrap ) fd_forest_init( ctx->forest, 0 );
1036 0 : break;
1037 0 : }
1038 0 : case IN_KIND_GOSSIP: {
1039 0 : fd_gossip_update_message_t const * msg = (fd_gossip_update_message_t const *)fd_type_pun_const( fd_chunk_to_laddr( in_ctx->mem, ctx->chunk ) );
1040 0 : after_gossip( ctx, msg, sig );
1041 0 : break;
1042 0 : }
1043 0 : case IN_KIND_REPLAY: {
1044 0 : fd_replay_fec_evicted_t const * msg = (fd_replay_fec_evicted_t const *)fd_type_pun_const( fd_chunk_to_laddr( in_ctx->mem, ctx->chunk ) );
1045 0 : fd_forest_fec_clear( ctx->forest, msg->slot, msg->fec_set_idx, FD_FEC_SHRED_CNT - 1 );
1046 0 : break;
1047 0 : }
1048 0 : case IN_KIND_TOWER: {
1049 0 : after_tower( ctx, sig, fd_chunk_to_laddr( in_ctx->mem, ctx->chunk ) );
1050 0 : break;
1051 0 : }
1052 0 : case IN_KIND_SHRED: {
1053 :
1054 : /* There are 3 message types from shred:
1055 : 1. resolver evict - incomplete FEC set is evicted by resolver
1056 : 2. fec complete - FEC set is completed by resolver. Also contains a shred.
1057 : 3. shred - new shred
1058 :
1059 : Msgs 2 and 3 have a shred header in the dcache. Msg 1 is empty. */
1060 :
1061 0 : uint sig_src = fd_shred_sig_src( sig );
1062 0 : int sig_res = fd_shred_sig_res( sig );
1063 :
1064 0 : if( FD_UNLIKELY( sig_src==SHRED_SIG_FEC_EVICTED ) ) {
1065 0 : fd_fec_evicted_t * evicted = (fd_fec_evicted_t *)fd_type_pun( fd_chunk_to_laddr( in_ctx->mem, ctx->chunk ) );
1066 0 : after_evict( ctx, evicted );
1067 0 : return;
1068 0 : }
1069 :
1070 0 : uchar * src = fd_chunk_to_laddr( in_ctx->mem, ctx->chunk );
1071 0 : fd_shred_base_t * shred_msg = (fd_shred_base_t *)fd_type_pun( src );
1072 0 : fd_shred_t * shred = &shred_msg->shred; /* completes & shred messages all have a shred header at the same offset (after merkle root) */
1073 :
1074 0 : if( FD_UNLIKELY( shred->slot > ctx->metrics->current_slot && sig_src == SHRED_SIG_SRC_TURBINE ) ) {
1075 0 : FD_LOG_INFO(( "[Turbine] slot: %lu, root: %lu", shred->slot, fd_forest_root_slot( ctx->forest ) ));
1076 0 : ctx->metrics->current_slot = shred->slot;
1077 0 : }
1078 :
1079 0 : if( FD_UNLIKELY( ctx->turbine_slot0 == ULONG_MAX && sig_src == SHRED_SIG_SRC_TURBINE ) ) {
1080 0 : ctx->turbine_slot0 = shred->slot;
1081 0 : fd_repair_metrics_set_turbine_slot0( ctx->slot_metrics, shred->slot );
1082 0 : fd_policy_set_turbine_slot0( ctx->policy, shred->slot );
1083 :
1084 : /* On first turbine shred, seed repair by queuing highest_shred
1085 : requests for slots between snapshot and turbine_slot0. This
1086 : bypasses forest entirely and dispatches directly via the sign
1087 : queue. Cap at half queue capacity to leave room for pongs. */
1088 0 : ulong root = fd_forest_root_slot( ctx->forest );
1089 0 : if( FD_LIKELY( root != ULONG_MAX && shred->slot > root ) ) {
1090 0 : ulong capacity = fd_signs_queue_max( ctx->pong_queue ) - fd_signs_queue_cnt( ctx->pong_queue );
1091 0 : ulong seed_cnt = fd_ulong_min( shred->slot-root, capacity/2 );
1092 0 : long now_ms = fd_clock_tile_now( ctx->clock )/(long)1e6;
1093 0 : for( ulong i=1; i<=seed_cnt; i++ ) {
1094 0 : ulong slot = root + i;
1095 0 : fd_pubkey_t const * peer = fd_policy_peer_select( ctx->policy );
1096 0 : if( FD_UNLIKELY( !peer ) ) break;
1097 0 : fd_repair_msg_t * msg = fd_repair_highest_shred( ctx->protocol, peer, (ulong)now_ms, 0, slot, 0 );
1098 0 : if( FD_LIKELY( msg ) ) fd_signs_queue_push( ctx->pong_queue, (sign_pending_t){ .msg = *msg } );
1099 0 : }
1100 0 : }
1101 0 : }
1102 :
1103 0 : if( FD_UNLIKELY( sig==SHRED_SIG_FEC_COMPLETE || sig==SHRED_SIG_FEC_COMPLETE_LEADER ) ) {
1104 0 : fd_fec_complete_t * complete_msg = (fd_fec_complete_t *)fd_type_pun( src );
1105 0 : int invalid = after_fec( ctx, &complete_msg->last_shred_hdr, &complete_msg->merkle_root, &complete_msg->chained_merkle_root );
1106 0 : ulong fwd_sig = invalid ? REPAIR_SIG_FEC_INVALID : (sig==SHRED_SIG_FEC_COMPLETE_LEADER ? REPAIR_SIG_FEC_LEADER : REPAIR_SIG_FEC);
1107 :
1108 : /* indiscriminately forward FEC complete messages along to replay */
1109 0 : memcpy( fd_chunk_to_laddr( ctx->repair_out_ctx->mem, ctx->repair_out_ctx->chunk ), src, sz );
1110 0 : fd_stem_publish( ctx->stem, ctx->repair_out_ctx->idx, fwd_sig, ctx->repair_out_ctx->chunk, sz, 0UL, 0UL, tspub );
1111 0 : ctx->repair_out_ctx->chunk = fd_dcache_compact_next( ctx->repair_out_ctx->chunk, sz, ctx->repair_out_ctx->chunk0, ctx->repair_out_ctx->wmark );
1112 0 : } else if( FD_LIKELY( sig_res!=SHRED_SIG_RESULT_EQVOC ) ) {
1113 0 : fd_hash_t * cmr = (fd_hash_t *)fd_type_pun(shred_msg->shred_ + fd_shred_chain_off( shred->variant ));
1114 0 : after_shred( ctx, sig, shred, shred_msg->rnonce, &shred_msg->merkle_root, cmr );
1115 0 : }
1116 0 : return;
1117 0 : }
1118 0 : default: FD_LOG_ERR(( "bad in_kind %u", in_kind )); /* Should never reach here since before_frag should have filtered out any unexpected frags. */
1119 0 : }
1120 0 : }
1121 :
1122 : /* Defer a request by adding it to the outstanding inflights table so it
1123 : can be re-requested after a timeout window. Nonce is 0 because these
1124 : are not real requests made to the network, and cannot be matched
1125 : by a shred response. */
1126 : static void
1127 0 : defer_inflight_request( ctx_t * ctx, ulong slot, ulong shred_idx ) {
1128 0 : fd_hash_t hash = { .ul[0] = 0 };
1129 0 : fd_inflight_key_t inflight_req = { .slot = slot, .shred_idx = shred_idx, .nonce = 0 };
1130 0 : if( FD_LIKELY( !fd_inflight_map_ele_query( ctx->inflights->map, &inflight_req, NULL, ctx->inflights->pool ) ) ) {
1131 0 : fd_inflights_request_insert( ctx->inflights, 0, &hash, slot, shred_idx );
1132 0 : }
1133 0 : }
1134 :
1135 : /* Should be called for any regular FD_REPAIR_KIND_SHRED request made. */
1136 : static void
1137 0 : record_inflight_request( ctx_t * ctx, ulong nonce, fd_pubkey_t const * peer, ulong slot, ulong shred_idx ) {
1138 0 : fd_inflights_request_insert( ctx->inflights, nonce, peer, slot, shred_idx );
1139 0 : fd_policy_peer_request_update( ctx->policy, peer );
1140 0 : }
1141 :
1142 :
1143 : /* Repair request invariants:
1144 :
1145 : Highest window index requests and orphan requests are much less
1146 : fragile than regular shred requests. This is because we request
1147 : orphans and highest window index requests whenever we "need" them,
1148 : i.e. we fire and forget them.
1149 :
1150 : But regular shred requests are only iterated once per shred per slot.
1151 : This is for performance reasons - we don't want to iterate over the
1152 : forest more than necessary. If we let knowledge of a regular shred
1153 : request get dropped, it's possible the request could get lost
1154 : forever. This is why any regular shred request that is conceived
1155 : needs to be handled at the repair tile level. All regular shred
1156 : requests must be added to the inflights outstanding table so it can
1157 : be re-requested after a timeout window.
1158 :
1159 : An inflight entry should only ever be permanently removed upon
1160 : receipt of a shred response.
1161 :
1162 : There are two methods through which we make regular shred requests:
1163 : 1. policy_next
1164 : 2. fd_inflights_request_pop
1165 :
1166 : In general, policy_next makes the first request for shred X, and
1167 : fd_inflights_request_pop makes all subsequent requests for shred X at
1168 : DEDUP_TIMEOUT intervals. This is to give each request a fair chance
1169 : to be received, but also to avoid colliding nonces. This is because
1170 : we want to be as accurate as possible when tracking per-peer response
1171 : latency.
1172 :
1173 : With eviction, it's possible for policy_next to make the same request
1174 : for shred X multiple times, in short timeout intervals. Therefore we
1175 : need to have both policy_next and fd_inflights_request_pop requests
1176 : pass through the same dedup cache. At the point of eviction, we
1177 : leave requests for that slot in the dedup cache and in the inflights
1178 : table. If an old request from before eviction happens to
1179 : dedup a request on the new insertion of the slot, these requests must
1180 : be queued up in inflights table so they can be re-requested after a
1181 : timeout window. */
1182 :
1183 : static inline void
1184 : after_credit( ctx_t * ctx,
1185 : fd_stem_context_t * stem FD_PARAM_UNUSED,
1186 : int * opt_poll_in FD_PARAM_UNUSED,
1187 0 : int * charge_busy ) {
1188 0 : long now = fd_clock_tile_now( ctx->clock );
1189 :
1190 0 : if( FD_UNLIKELY( ctx->halt_signing ) ) {
1191 0 : *charge_busy = 1;
1192 0 : return;
1193 0 : }
1194 :
1195 : /* Verify that there is at least one sign tile with available credits.
1196 : If not, we can't send any requests and leave early. */
1197 0 : out_ctx_t * sign_out = sign_avail_credits( ctx );
1198 0 : if( FD_UNLIKELY( !sign_out ) ) {
1199 0 : ctx->metrics->sign_tile_unavail++;
1200 0 : return;
1201 0 : }
1202 :
1203 : /* If inflights is at capacity, then the only thing we can send is:
1204 : pongs, initial highest window index requests, or resend things that
1205 : are already inflight. Any new requests that would cause an
1206 : inflight to be added to the queue must be deferred. */
1207 :
1208 0 : if( FD_UNLIKELY( !fd_signs_queue_empty( ctx->pong_queue ) ) ) {
1209 0 : sign_pending_t signable = fd_signs_queue_pop( ctx->pong_queue );
1210 0 : fd_repair_send_sign_request( ctx, sign_out, &signable.msg, signable.msg.kind == FD_REPAIR_KIND_PONG ? &signable.pong_data : NULL );
1211 0 : *charge_busy = 1;
1212 0 : return;
1213 0 : }
1214 :
1215 0 : if( FD_UNLIKELY( fd_inflights_should_drain( ctx->inflights, now ) ) ) {
1216 0 : ulong nonce; ulong slot; ulong shred_idx;
1217 0 : *charge_busy = 1;
1218 0 : fd_inflights_request_pop( ctx->inflights, &nonce, &slot, &shred_idx );
1219 :
1220 0 : fd_forest_blk_t * blk = fd_forest_query( ctx->forest, slot );
1221 0 : if( FD_UNLIKELY( blk && !fd_forest_blk_idxs_test( blk->idxs, shred_idx ) ) ) {
1222 0 : fd_pubkey_t const * peer = fd_policy_peer_select( ctx->policy );
1223 :
1224 0 : if( FD_UNLIKELY( !peer || fd_reqlim_next( ctx->dedup, fd_reqlim_key( FD_REPAIR_KIND_SHRED, slot, (uint)shred_idx ), now ) ) ) {
1225 : /* No peers available, park the request in inflights. */
1226 0 : defer_inflight_request( ctx, slot, shred_idx );
1227 0 : } else {
1228 0 : ctx->metrics->rerequest++;
1229 0 : nonce = fd_rnonce_ss_compute( ctx->repair_nonce_ss, 1, slot, (uint)shred_idx, now );
1230 0 : fd_repair_msg_t * msg = fd_repair_shred( ctx->protocol, peer, (ulong)now/(ulong)1e6, (uint)nonce, slot, shred_idx );
1231 0 : fd_repair_send_sign_request( ctx, sign_out, msg, NULL );
1232 0 : record_inflight_request( ctx, nonce, peer, slot, shred_idx ); /* Request is definitely a regular shred request. */
1233 0 : return;
1234 0 : }
1235 0 : }
1236 0 : }
1237 :
1238 0 : if( FD_UNLIKELY( fd_inflights_outstanding_free( ctx->inflights ) <= fd_signs_map_key_cnt( ctx->signs_map ) ) ) return; /* no new requests allowed */
1239 :
1240 0 : fd_repair_msg_t const * cout = fd_policy_next( ctx->policy, ctx->dedup, ctx->forest, ctx->protocol, now, ctx->metrics->current_slot, charge_busy );
1241 0 : if( FD_UNLIKELY( !cout ) ) return;
1242 :
1243 0 : if( ( cout->kind == FD_REPAIR_KIND_SHRED && fd_reqlim_next( ctx->dedup, fd_reqlim_key( FD_REPAIR_KIND_SHRED, cout->shred.slot, (uint)cout->shred.shred_idx ), now ) ) ) {
1244 : /* Here if policy_next is re-requesting a shred that's already been
1245 : requested. This could be happen during a race - imagine we make a
1246 : request for shred 0 in slot Y. Then eviction causes slot Y to be
1247 : removed and then readded. policy_next will re-request shred 0,
1248 : but if we let it get dropped here, it's possible the request
1249 : could get lost forever. */
1250 0 : defer_inflight_request( ctx, cout->shred.slot, cout->shred.shred_idx );
1251 0 : return;
1252 0 : }
1253 :
1254 : /* finally, send the request made by policy */
1255 0 : fd_repair_send_sign_request( ctx, sign_out, cout, NULL );
1256 0 : if( FD_LIKELY( cout->kind == FD_REPAIR_KIND_SHRED ) ) record_inflight_request( ctx, cout->shred.nonce, &cout->shred.to, cout->shred.slot, cout->shred.shred_idx );
1257 0 : }
1258 :
1259 : static void
1260 0 : signs_queue_update_identity( ctx_t * ctx ) {
1261 0 : ulong queue_cnt = fd_signs_queue_cnt( ctx->pong_queue );
1262 0 : for( ulong i=0UL; i<queue_cnt; i++ ) {
1263 0 : sign_pending_t signable = fd_signs_queue_pop( ctx->pong_queue );
1264 0 : switch( signable.msg.kind ) {
1265 0 : case FD_REPAIR_KIND_PONG:
1266 0 : memcpy( signable.msg.pong.from.uc, ctx->identity_public_key.uc, sizeof(fd_pubkey_t) );
1267 0 : break;
1268 0 : case FD_REPAIR_KIND_SHRED:
1269 0 : memcpy( signable.msg.shred.from.uc, ctx->identity_public_key.uc, sizeof(fd_pubkey_t) );
1270 0 : break;
1271 0 : case FD_REPAIR_KIND_HIGHEST_SHRED:
1272 0 : memcpy( signable.msg.highest_shred.from.uc, ctx->identity_public_key.uc, sizeof(fd_pubkey_t) );
1273 0 : break;
1274 0 : case FD_REPAIR_KIND_ORPHAN:
1275 0 : memcpy( signable.msg.orphan.from.uc, ctx->identity_public_key.uc, sizeof(fd_pubkey_t) );
1276 0 : break;
1277 0 : default:
1278 0 : FD_LOG_CRIT(( "Unhandled repair kind %u", signable.msg.kind ));
1279 0 : break;
1280 0 : }
1281 0 : fd_signs_queue_push( ctx->pong_queue, signable );
1282 0 : }
1283 0 : }
1284 :
1285 : static inline void
1286 0 : during_housekeeping( ctx_t * ctx ) {
1287 0 : if( FD_UNLIKELY( fd_clock_tile_recal_due( ctx->clock ) ) ) fd_clock_tile_recal( ctx->clock );
1288 :
1289 : # if DEBUG_LOGGING
1290 : long now = fd_log_wallclock();
1291 : if( FD_UNLIKELY( now - ctx->tsdebug > (long)10e9 ) ) {
1292 : fd_forest_print( ctx->forest );
1293 : ctx->tsdebug = fd_log_wallclock();
1294 : }
1295 : # endif
1296 :
1297 0 : if( FD_UNLIKELY( fd_keyswitch_state_query( ctx->keyswitch )==FD_KEYSWITCH_STATE_UNHALT_PENDING ) ) {
1298 0 : FD_LOG_DEBUG(( "keyswitch: unhalting" ));
1299 0 : FD_CHECK_CRIT( ctx->halt_signing, "state machine corruption" );
1300 0 : ctx->halt_signing = 0;
1301 0 : fd_keyswitch_state( ctx->keyswitch, FD_KEYSWITCH_STATE_COMPLETED );
1302 0 : }
1303 :
1304 0 : if( FD_UNLIKELY( fd_keyswitch_state_query( ctx->keyswitch )==FD_KEYSWITCH_STATE_SWITCH_PENDING ) ) {
1305 :
1306 0 : if( !ctx->halt_signing ) {
1307 : /* At this point, stop sending new sign requests to the sign tile
1308 : and wait for all outstanding sign requests to be received back
1309 : from the sign tile. We also need to update any pending
1310 : outgoing sign requests with the new identity key. */
1311 0 : FD_LOG_DEBUG(( "keyswitch: halting signing" ));
1312 0 : ctx->halt_signing = 1;
1313 0 : memcpy( ctx->identity_public_key.uc, ctx->keyswitch->bytes, 32UL );
1314 0 : ctx->protocol->identity_key = ctx->identity_public_key;
1315 0 : signs_queue_update_identity( ctx );
1316 0 : }
1317 :
1318 0 : if( fd_signs_map_key_cnt( ctx->signs_map )==0UL ) {
1319 : /* Once there are no more in flight sign requests, we are ready to
1320 : say that the keyswitch is completed. */
1321 0 : FD_LOG_DEBUG(( "keyswitch: completed, no more outstanding stale sign requests" ));
1322 0 : fd_keyswitch_state( ctx->keyswitch, FD_KEYSWITCH_STATE_COMPLETED );
1323 0 : }
1324 0 : }
1325 0 : }
1326 :
1327 : static void
1328 : privileged_init( fd_topo_t const * topo,
1329 0 : fd_topo_tile_t const * tile ) {
1330 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1331 :
1332 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1333 0 : ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(ctx_t), sizeof(ctx_t) );
1334 0 : fd_memset( ctx, 0, sizeof(ctx_t) );
1335 :
1336 0 : uchar const * identity_key = fd_keyload_load( tile->repair.identity_key_path, /* pubkey only: */ 1 );
1337 0 : fd_memcpy( ctx->identity_public_key.uc, identity_key, sizeof(fd_pubkey_t) );
1338 :
1339 0 : FD_TEST( fd_rng_secure( &ctx->repair_seed, sizeof(ulong) ) );
1340 :
1341 0 : ulong rnonce_ss_id = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "rnonce_ss" );
1342 0 : FD_TEST( rnonce_ss_id!=ULONG_MAX );
1343 0 : memcpy( ctx->repair_nonce_ss, fd_topo_obj_laddr( topo, rnonce_ss_id ), sizeof(fd_rnonce_ss_t) );
1344 0 : }
1345 :
1346 : static void
1347 : unprivileged_init( fd_topo_t const * topo,
1348 0 : fd_topo_tile_t const * tile ) {
1349 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1350 :
1351 0 : ulong total_sign_depth = tile->repair.repair_sign_depth * tile->repair.repair_sign_cnt;
1352 0 : int lg_sign_depth = fd_ulong_find_msb( fd_ulong_pow2_up(total_sign_depth) ) + 1;
1353 :
1354 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1355 0 : ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(ctx_t), sizeof(ctx_t) );
1356 0 : ctx->protocol = FD_SCRATCH_ALLOC_APPEND( l, fd_repair_align(), fd_repair_footprint() );
1357 0 : ctx->forest = FD_SCRATCH_ALLOC_APPEND( l, fd_forest_align(), fd_forest_footprint( tile->repair.slot_max ) );
1358 0 : ctx->policy = FD_SCRATCH_ALLOC_APPEND( l, fd_policy_align(), fd_policy_footprint( FD_REPAIR_PEER_MAX ) );
1359 0 : ctx->dedup = FD_SCRATCH_ALLOC_APPEND( l, fd_reqlim_align(), fd_reqlim_footprint( FD_REQLIM_CACHE_MAX ) );
1360 0 : ctx->inflights = FD_SCRATCH_ALLOC_APPEND( l, fd_inflights_align(), fd_inflights_footprint() );
1361 0 : ctx->signs_map = FD_SCRATCH_ALLOC_APPEND( l, fd_signs_map_align(), fd_signs_map_footprint( lg_sign_depth ) );
1362 0 : ctx->pong_queue = FD_SCRATCH_ALLOC_APPEND( l, fd_signs_queue_align(), fd_signs_queue_footprint() );
1363 0 : ctx->slot_metrics = FD_SCRATCH_ALLOC_APPEND( l, fd_repair_metrics_align(), fd_repair_metrics_footprint() );
1364 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, scratch_align() );
1365 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
1366 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
1367 :
1368 0 : ctx->protocol = fd_repair_join ( fd_repair_new ( ctx->protocol, &ctx->identity_public_key ) );
1369 0 : ctx->forest = fd_forest_join ( fd_forest_new ( ctx->forest, tile->repair.slot_max, ctx->repair_seed ) );
1370 0 : ctx->policy = fd_policy_join ( fd_policy_new ( ctx->policy, FD_REPAIR_PEER_MAX, ctx->repair_seed, ctx->repair_nonce_ss ) );
1371 0 : ctx->dedup = fd_reqlim_join ( fd_reqlim_new ( ctx->dedup, FD_REQLIM_CACHE_MAX, ctx->repair_seed ) );
1372 0 : ctx->inflights = fd_inflights_join ( fd_inflights_new ( ctx->inflights, ctx->repair_seed+1234UL ) );
1373 0 : ctx->signs_map = fd_signs_map_join ( fd_signs_map_new ( ctx->signs_map, lg_sign_depth, 0UL ) );
1374 0 : ctx->pong_queue = fd_signs_queue_join ( fd_signs_queue_new ( ctx->pong_queue ) );
1375 0 : ctx->slot_metrics = fd_repair_metrics_join( fd_repair_metrics_new( ctx->slot_metrics ) );
1376 :
1377 0 : ctx->keyswitch = fd_keyswitch_join( fd_topo_obj_laddr( topo, tile->id_keyswitch_obj_id ) );
1378 0 : FD_TEST( ctx->keyswitch );
1379 :
1380 0 : ctx->halt_signing = 0;
1381 :
1382 : /* Process in links */
1383 :
1384 0 : if( FD_UNLIKELY( tile->in_cnt > MAX_IN_LINKS ) ) FD_LOG_ERR(( "repair tile has too many input links" ));
1385 :
1386 0 : uint sign_repair_in_idx[ MAX_SIGN_TILE_CNT ] = {0};
1387 0 : uint sign_repair_idx = 0;
1388 0 : ulong sign_link_depth = 0;
1389 :
1390 0 : for( uint in_idx=0U; in_idx<(tile->in_cnt); in_idx++ ) {
1391 0 : fd_topo_link_t const * link = &topo->links[ tile->in_link_id[ in_idx ] ];
1392 0 : if( 0==strcmp( link->name, "net_repair" ) ) {
1393 0 : ctx->in_kind[ in_idx ] = IN_KIND_NET;
1394 0 : fd_net_rx_bounds_init( &ctx->in_links[ in_idx ].net_rx, link->dcache );
1395 0 : continue;
1396 0 : } else if( 0==strcmp( link->name, "sign_repair" ) ) {
1397 0 : ctx->in_kind[ in_idx ] = IN_KIND_SIGN;
1398 0 : sign_repair_in_idx[ sign_repair_idx++ ] = in_idx;
1399 0 : sign_link_depth = link->depth;
1400 0 : }
1401 0 : else if( 0==strcmp( link->name, "gossip_out" ) ) ctx->in_kind[ in_idx ] = IN_KIND_GOSSIP;
1402 0 : else if( 0==strcmp( link->name, "tower_out" ) ) ctx->in_kind[ in_idx ] = IN_KIND_TOWER;
1403 0 : else if( 0==strcmp( link->name, "shred_out" ) ) ctx->in_kind[ in_idx ] = IN_KIND_SHRED;
1404 0 : else if( 0==strcmp( link->name, "snapin_manif" ) ) ctx->in_kind[ in_idx ] = IN_KIND_SNAP;
1405 0 : else if( 0==strcmp( link->name, "genesi_out" ) ) ctx->in_kind[ in_idx ] = IN_KIND_GENESIS;
1406 0 : else if( 0==strcmp( link->name, "replay_out" ) ) ctx->in_kind[ in_idx ] = IN_KIND_REPLAY;
1407 0 : else FD_LOG_ERR(( "repair tile has unexpected input link %s", link->name ));
1408 :
1409 0 : ctx->in_links[ in_idx ].mem = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
1410 0 : ctx->in_links[ in_idx ].chunk0 = fd_dcache_compact_chunk0( ctx->in_links[ in_idx ].mem, link->dcache );
1411 0 : ctx->in_links[ in_idx ].wmark = fd_dcache_compact_wmark ( ctx->in_links[ in_idx ].mem, link->dcache, link->mtu );
1412 0 : ctx->in_links[ in_idx ].mtu = link->mtu;
1413 :
1414 0 : FD_TEST( fd_dcache_compact_is_safe( ctx->in_links[in_idx].mem, link->dcache, link->mtu, link->depth ) );
1415 0 : }
1416 :
1417 0 : ctx->net_out_ctx->idx = UINT_MAX;
1418 0 : ctx->repair_out_ctx->idx = UINT_MAX;
1419 0 : ctx->repair_sign_cnt = 0;
1420 0 : ctx->sign_rrobin_idx = 0;
1421 :
1422 0 : for( uint out_idx=0U; out_idx<(tile->out_cnt); out_idx++ ) {
1423 0 : fd_topo_link_t const * link = &topo->links[ tile->out_link_id[ out_idx ] ];
1424 :
1425 0 : if( 0==strcmp( link->name, "repair_net" ) ) {
1426 :
1427 0 : if( ctx->net_out_ctx->idx!=UINT_MAX ) continue; /* only use first net link */
1428 0 : ctx->net_out_ctx->idx = out_idx;
1429 0 : ctx->net_out_ctx->mem = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
1430 0 : ctx->net_out_ctx->chunk0 = fd_dcache_compact_chunk0( ctx->net_out_ctx->mem, link->dcache );
1431 0 : ctx->net_out_ctx->wmark = fd_dcache_compact_wmark( ctx->net_out_ctx->mem, link->dcache, link->mtu );
1432 0 : ctx->net_out_ctx->chunk = ctx->net_out_ctx->chunk0;
1433 :
1434 0 : } else if( 0==strcmp( link->name, "repair_out" ) ) {
1435 :
1436 0 : out_ctx_t * replay_out = ctx->repair_out_ctx;
1437 0 : replay_out->idx = out_idx;
1438 0 : replay_out->mem = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
1439 0 : replay_out->chunk0 = fd_dcache_compact_chunk0( replay_out->mem, link->dcache );
1440 0 : replay_out->wmark = fd_dcache_compact_wmark( replay_out->mem, link->dcache, link->mtu );
1441 0 : replay_out->chunk = replay_out->chunk0;
1442 :
1443 0 : } else if( 0==strcmp( link->name, "repair_sign" ) ) {
1444 :
1445 0 : out_ctx_t * repair_sign_out = &ctx->repair_sign_out_ctx[ ctx->repair_sign_cnt ];
1446 0 : repair_sign_out->idx = out_idx;
1447 0 : repair_sign_out->mem = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
1448 0 : repair_sign_out->chunk0 = fd_dcache_compact_chunk0( repair_sign_out->mem, link->dcache );
1449 0 : repair_sign_out->wmark = fd_dcache_compact_wmark( repair_sign_out->mem, link->dcache, link->mtu );
1450 0 : repair_sign_out->chunk = repair_sign_out->chunk0;
1451 0 : repair_sign_out->in_idx = sign_repair_in_idx[ ctx->repair_sign_cnt++ ]; /* match to the sign_repair input link */
1452 0 : repair_sign_out->max_credits = sign_link_depth;
1453 0 : repair_sign_out->credits = sign_link_depth;
1454 :
1455 0 : } else {
1456 0 : FD_LOG_ERR(( "repair tile has unexpected output link %s", link->name ));
1457 0 : }
1458 0 : }
1459 0 : FD_TEST( ctx->net_out_ctx->idx!=UINT_MAX );
1460 0 : FD_TEST( ctx->repair_out_ctx->idx!=UINT_MAX );
1461 0 : if( FD_UNLIKELY( ctx->repair_sign_cnt!=sign_repair_idx ) ) {
1462 0 : FD_LOG_ERR(( "Mismatch between repair_sign output links (%lu) and sign_repair input links (%u)", ctx->repair_sign_cnt, sign_repair_idx ));
1463 0 : }
1464 0 : if( FD_UNLIKELY( fd_signs_map_key_max( ctx->signs_map ) < tile->repair.repair_sign_depth * tile->repair.repair_sign_cnt ) ) {
1465 0 : FD_LOG_ERR(( "Repair pending signs tracking map is too small: %lu < %lu.", fd_signs_map_key_max( ctx->signs_map ), tile->repair.repair_sign_depth * tile->repair.repair_sign_cnt ));
1466 0 : }
1467 :
1468 0 : ctx->wksp = topo->workspaces[ topo->objs[ tile->tile_obj_id ].wksp_id ].wksp;
1469 0 : ctx->repair_intake_addr.port = fd_ushort_bswap( tile->repair.repair_client_listen_port );
1470 :
1471 : /* TODO clean these up */
1472 0 : ctx->net_id = (ushort)0;
1473 0 : fd_ip4_udp_hdr_init( ctx->intake_hdr, 0, 0, tile->repair.repair_client_listen_port );
1474 :
1475 : /* Repair set up */
1476 :
1477 0 : ctx->turbine_slot0 = ULONG_MAX;
1478 0 : FD_LOG_INFO(( "repair my addr - intake addr: " FD_IP4_ADDR_FMT ":%u",
1479 0 : FD_IP4_ADDR_FMT_ARGS( ctx->repair_intake_addr.addr ), fd_ushort_bswap( ctx->repair_intake_addr.port )
1480 0 : ));
1481 :
1482 0 : memset( ctx->metrics, 0, sizeof(ctx->metrics) );
1483 :
1484 0 : fd_histf_join( fd_histf_new( ctx->metrics->slot_compl_time, FD_MHIST_SECONDS_MIN( REPAIR, SLOT_COMPLETE_DURATION_SECONDS ),
1485 0 : FD_MHIST_SECONDS_MAX( REPAIR, SLOT_COMPLETE_DURATION_SECONDS ) ) );
1486 0 : fd_histf_join( fd_histf_new( ctx->metrics->response_latency, FD_MHIST_MIN( REPAIR, RESPONSE_LATENCY_NANOS ),
1487 0 : FD_MHIST_MAX( REPAIR, RESPONSE_LATENCY_NANOS ) ) );
1488 :
1489 0 : fd_clock_tile_init( ctx->clock );
1490 :
1491 0 : ctx->tsdebug = fd_log_wallclock();
1492 0 : ctx->pending_key_next = 0;
1493 0 : }
1494 :
1495 : static ulong
1496 : populate_allowed_seccomp( fd_topo_t const * topo FD_PARAM_UNUSED,
1497 : fd_topo_tile_t const * tile FD_PARAM_UNUSED,
1498 : ulong out_cnt,
1499 0 : struct sock_filter * out ) {
1500 0 : populate_sock_filter_policy_fd_repair_tile( out_cnt, out, (uint)fd_log_private_logfile_fd() );
1501 0 : return sock_filter_policy_fd_repair_tile_instr_cnt;
1502 0 : }
1503 :
1504 : static ulong
1505 : populate_allowed_fds( fd_topo_t const * topo FD_PARAM_UNUSED,
1506 : fd_topo_tile_t const * tile FD_PARAM_UNUSED,
1507 : ulong out_fds_cnt,
1508 0 : int * out_fds ) {
1509 0 : if( FD_UNLIKELY( out_fds_cnt<2UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
1510 :
1511 0 : ulong out_cnt = 0UL;
1512 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
1513 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
1514 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
1515 0 : return out_cnt;
1516 0 : }
1517 :
1518 : static inline void
1519 0 : metrics_write( ctx_t * ctx ) {
1520 0 : FD_MGAUGE_SET( REPAIR, SLOT_CURRENT, ctx->metrics->current_slot );
1521 0 : FD_MGAUGE_SET( REPAIR, SLOT_HIGHEST_REPAIRED, fd_forest_highest_repaired_slot( ctx->forest ) );
1522 0 : FD_MCNT_SET( REPAIR, SHRED_OLD, ctx->metrics->old_shred );
1523 0 : FD_MCNT_SET( REPAIR, PEER_REQUESTED, fd_policy_peer_pool_used( ctx->policy->peers.pool ) );
1524 0 : FD_MCNT_SET( REPAIR, SIGN_TILE_UNAVAILABLE, ctx->metrics->sign_tile_unavail );
1525 0 : FD_MCNT_SET( REPAIR, SHRED_REREQUESTED, ctx->metrics->rerequest );
1526 :
1527 0 : FD_MGAUGE_SET( REPAIR, SLOT_LAST_REQUESTED, ctx->metrics->last_requested_slot );
1528 0 : FD_MGAUGE_SET( REPAIR, ORPHAN_LAST_REQUESTED, ctx->metrics->last_requested_orphan );
1529 0 : FD_MGAUGE_SET( REPAIR, REQUEST_INFLIGHT, fd_inflight_pool_used( ctx->inflights->pool ) - ctx->inflights->popped_cnt );
1530 :
1531 0 : FD_MCNT_SET ( REPAIR, PKT_TX, ctx->metrics->send_pkt_cnt );
1532 0 : FD_MCNT_ENUM_COPY( REPAIR, REQUEST_TX, ctx->metrics->sent_pkt_types );
1533 :
1534 0 : FD_MHIST_COPY( REPAIR, SLOT_COMPLETE_DURATION_SECONDS, ctx->metrics->slot_compl_time );
1535 0 : FD_MHIST_COPY( REPAIR, RESPONSE_LATENCY_NANOS, ctx->metrics->response_latency );
1536 :
1537 0 : FD_MCNT_SET ( REPAIR, BLOCK_EVICTED, ctx->metrics->blk_evicted );
1538 0 : FD_MCNT_SET ( REPAIR, BLOCK_INSERT_FAILED, ctx->metrics->blk_failed_insert );
1539 0 : FD_MGAUGE_SET( REPAIR, SLOT_LAST_EVICTED, ctx->metrics->slot_evicted );
1540 0 : FD_MGAUGE_SET( REPAIR, SLOT_LAST_EVICTION_CAUSE, ctx->metrics->slot_evicted_by );
1541 0 : FD_MGAUGE_SET( REPAIR, SLOT_LAST_INSERT_FAILED, ctx->metrics->slot_failed_insert );
1542 :
1543 0 : FD_MCNT_SET ( REPAIR, CHAIN_VERIFY_FAILED, ctx->metrics->failed_chain_verify_cnt );
1544 0 : FD_MGAUGE_SET( REPAIR, SLOT_LAST_CHAIN_VERIFY_FAILED, ctx->metrics->failed_chain_verify_slot );
1545 :
1546 0 : FD_MCNT_SET( REPAIR, PING_UNKNOWN_PEER, ctx->metrics->unknown_peer_ping );
1547 0 : FD_MCNT_SET( REPAIR, PING_MALFORMED, ctx->metrics->malformed_ping );
1548 0 : FD_MCNT_SET( REPAIR, PING_SIGNATURE_FAILED, ctx->metrics->fail_sigverify_ping );
1549 0 : }
1550 :
1551 : #undef DEBUG_LOGGING
1552 :
1553 : /* At most one sign request is made in after_credit. Then at most one
1554 : message is published in after_frag. */
1555 0 : #define STEM_BURST (2UL)
1556 :
1557 : /* Set LAZY to a reasonable value that keeps housekeeping time low.
1558 : Repair tile's only reliable consumer is replay. */
1559 0 : #define STEM_LAZY (64000)
1560 :
1561 0 : #define STEM_CALLBACK_CONTEXT_TYPE ctx_t
1562 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(ctx_t)
1563 :
1564 0 : #define STEM_CALLBACK_AFTER_CREDIT after_credit
1565 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
1566 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
1567 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
1568 0 : #define STEM_CALLBACK_DURING_HOUSEKEEPING during_housekeeping
1569 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
1570 :
1571 : #include "../../disco/stem/fd_stem.c"
1572 :
1573 : fd_topo_run_tile_t fd_tile_repair = {
1574 : .name = "repair",
1575 : .populate_allowed_seccomp = populate_allowed_seccomp,
1576 : .populate_allowed_fds = populate_allowed_fds,
1577 : .scratch_align = scratch_align,
1578 : .scratch_footprint = scratch_footprint,
1579 : .unprivileged_init = unprivileged_init,
1580 : .privileged_init = privileged_init,
1581 : .run = stem_run,
1582 : };
|
