Line data Source code
1 : #include "fd_quic_tile.h"
2 : #include "../metrics/fd_metrics.h"
3 : #include "../stem/fd_stem.h"
4 : #include "../topo/fd_topo.h"
5 : #include "fd_tpu.h"
6 : #include "../../waltz/quic/fd_quic_private.h"
7 : #include "generated/quic_seccomp.h"
8 : #include "../../util/net/fd_eth.h"
9 :
10 : #include <errno.h>
11 : #include <linux/unistd.h>
12 : #include <sys/random.h>
13 :
14 : /* fd_quic provides a TPU server tile.
15 :
16 : This tile handles incoming transactions that clients request to be
17 : included in blocks. Supported protocols currently include TPU/UDP
18 : and TPU/QUIC.
19 :
20 : The fd_quic tile acts as a plain old Tango producer writing to a cnc
21 : and an mcache. The tile will defragment multi-packet TPU streams
22 : coming in from QUIC, such that each mcache/dcache pair forms a
23 : complete txn. This requires the dcache mtu to be at least that of
24 : the largest allowed serialized txn size.
25 :
26 : QUIC tiles don't service network devices directly, but rely on
27 : packets being received by net tiles and forwarded on via. a mux
28 : (multiplexer). An arbitrary number of QUIC tiles can be run. Each
29 : UDP flow must stick to one QUIC tile. */
30 :
31 : static inline fd_quic_limits_t
32 0 : quic_limits( fd_topo_tile_t const * tile ) {
33 0 : fd_quic_limits_t limits = {
34 0 : .conn_cnt = tile->quic.max_concurrent_connections,
35 0 : .handshake_cnt = tile->quic.max_concurrent_handshakes,
36 :
37 : /* fd_quic will not issue nor use any new connection IDs after
38 : completing a handshake. Connection migration is not supported
39 : either. */
40 0 : .conn_id_cnt = FD_QUIC_MIN_CONN_ID_CNT,
41 0 : .inflight_frame_cnt = 64UL * tile->quic.max_concurrent_connections,
42 0 : .min_inflight_frame_cnt_conn = 32UL
43 0 : };
44 0 : if( FD_UNLIKELY( !fd_quic_footprint( &limits ) ) ) {
45 0 : FD_LOG_ERR(( "Invalid QUIC limits in config" ));
46 0 : }
47 0 : return limits;
48 0 : }
49 :
50 : FD_FN_CONST static inline ulong
51 0 : scratch_align( void ) {
52 0 : return fd_ulong_max( alignof(fd_quic_ctx_t), fd_quic_align() );
53 0 : }
54 :
55 : static inline ulong
56 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
57 0 : ulong out_depth = tile->quic.out_depth;
58 0 : ulong reasm_max = tile->quic.reasm_cnt;
59 :
60 0 : fd_quic_limits_t limits = quic_limits( tile ); /* May FD_LOG_ERR */
61 0 : ulong l = FD_LAYOUT_INIT;
62 0 : l = FD_LAYOUT_APPEND( l, alignof( fd_quic_ctx_t ), sizeof( fd_quic_ctx_t ) );
63 0 : l = FD_LAYOUT_APPEND( l, fd_quic_align(), fd_quic_footprint( &limits ) );
64 0 : l = FD_LAYOUT_APPEND( l, fd_tpu_reasm_align(), fd_tpu_reasm_footprint( out_depth, reasm_max ) );
65 0 : return FD_LAYOUT_FINI( l, scratch_align() );
66 0 : }
67 :
68 : /* legacy_stream_notify is called for transactions sent via TPU/UDP. For
69 : now both QUIC and non-QUIC transactions are accepted, with traffic
70 : type determined by port.
71 :
72 : UDP transactions must fit in one packet and cannot be fragmented, and
73 : notify here means the entire packet was received. */
74 :
75 : static void
76 : legacy_stream_notify( fd_quic_ctx_t * ctx,
77 : uchar * packet,
78 0 : ulong packet_sz ) {
79 :
80 0 : long tspub = fd_tickcount();
81 0 : fd_tpu_reasm_t * reasm = ctx->reasm;
82 0 : fd_stem_context_t * stem = ctx->stem;
83 0 : fd_frag_meta_t * mcache = stem->mcaches[0];
84 0 : void * base = ctx->verify_out_mem;
85 0 : ulong seq = stem->seqs[0];
86 :
87 0 : int err = fd_tpu_reasm_publish_fast( reasm, packet, packet_sz, mcache, base, seq, tspub );
88 0 : if( FD_LIKELY( err==FD_TPU_REASM_SUCCESS ) ) {
89 0 : fd_stem_advance( stem, 0UL );
90 0 : ctx->metrics.txns_received_udp++;
91 0 : }
92 0 : }
93 :
94 : /* Because of the separate mcache for publishing network fragments
95 : back to networking tiles, which is not managed by the mux, we
96 : need to periodically update the sync. */
97 : static void
98 0 : during_housekeeping( fd_quic_ctx_t * ctx ) {
99 0 : fd_mcache_seq_update( ctx->net_out_sync, ctx->net_out_seq );
100 0 : }
101 :
102 : /* This tile always publishes messages downstream, even if there are
103 : no credits available. It ignores the flow control of the downstream
104 : verify tile. This is OK as the verify tile is written to expect
105 : this behavior, and enables the QUIC tile to publish as fast as it
106 : can. It would currently be difficult trying to backpressure further
107 : up the stack to the network itself. */
108 : static inline void
109 : before_credit( fd_quic_ctx_t * ctx,
110 : fd_stem_context_t * stem,
111 0 : int * charge_busy ) {
112 0 : ctx->stem = stem;
113 :
114 : /* Publishes to mcache via callbacks */
115 0 : *charge_busy = fd_quic_service( ctx->quic );
116 0 : }
117 :
118 : static inline void
119 0 : metrics_write( fd_quic_ctx_t * ctx ) {
120 0 : FD_MCNT_SET ( QUIC, TXNS_RECEIVED_UDP, ctx->metrics.txns_received_udp );
121 0 : FD_MCNT_SET ( QUIC, TXNS_RECEIVED_QUIC_FAST, ctx->metrics.txns_received_quic_fast );
122 0 : FD_MCNT_SET ( QUIC, TXNS_RECEIVED_QUIC_FRAG, ctx->metrics.txns_received_quic_frag );
123 0 : FD_MCNT_SET ( QUIC, FRAGS_OK, ctx->metrics.frag_ok_cnt );
124 0 : FD_MCNT_SET ( QUIC, FRAGS_GAP, ctx->metrics.frag_gap_cnt );
125 0 : FD_MCNT_SET ( QUIC, FRAGS_DUP, ctx->metrics.frag_dup_cnt );
126 0 : FD_MCNT_SET ( QUIC, TXNS_OVERRUN, ctx->metrics.reasm_overrun );
127 0 : FD_MCNT_SET ( QUIC, TXNS_ABANDONED, ctx->metrics.reasm_abandoned );
128 0 : FD_MCNT_SET ( QUIC, TXN_REASMS_STARTED, ctx->metrics.reasm_started );
129 0 : FD_MGAUGE_SET( QUIC, TXN_REASMS_ACTIVE, (ulong)fd_long_max( ctx->metrics.reasm_active, 0L ) );
130 :
131 0 : FD_MCNT_SET( QUIC, LEGACY_TXN_UNDERSZ, ctx->metrics.udp_pkt_too_small );
132 0 : FD_MCNT_SET( QUIC, LEGACY_TXN_OVERSZ, ctx->metrics.udp_pkt_too_large );
133 0 : FD_MCNT_SET( QUIC, TXN_UNDERSZ, ctx->metrics.quic_txn_too_small );
134 0 : FD_MCNT_SET( QUIC, TXN_OVERSZ, ctx->metrics.quic_txn_too_large );
135 :
136 0 : FD_MCNT_SET( QUIC, RECEIVED_PACKETS, ctx->quic->metrics.net_rx_pkt_cnt );
137 0 : FD_MCNT_SET( QUIC, RECEIVED_BYTES, ctx->quic->metrics.net_rx_byte_cnt );
138 0 : FD_MCNT_SET( QUIC, SENT_PACKETS, ctx->quic->metrics.net_tx_pkt_cnt );
139 0 : FD_MCNT_SET( QUIC, SENT_BYTES, ctx->quic->metrics.net_tx_byte_cnt );
140 0 : FD_MCNT_SET( QUIC, RETRY_SENT, ctx->quic->metrics.retry_tx_cnt );
141 :
142 0 : FD_MGAUGE_SET( QUIC, CONNECTIONS_ACTIVE, ctx->quic->metrics.conn_active_cnt );
143 0 : FD_MCNT_SET( QUIC, CONNECTIONS_CREATED, ctx->quic->metrics.conn_created_cnt );
144 0 : FD_MCNT_SET( QUIC, CONNECTIONS_CLOSED, ctx->quic->metrics.conn_closed_cnt );
145 0 : FD_MCNT_SET( QUIC, CONNECTIONS_ABORTED, ctx->quic->metrics.conn_aborted_cnt );
146 0 : FD_MCNT_SET( QUIC, CONNECTIONS_TIMED_OUT, ctx->quic->metrics.conn_timeout_cnt );
147 0 : FD_MCNT_SET( QUIC, CONNECTIONS_RETRIED, ctx->quic->metrics.conn_retry_cnt );
148 :
149 0 : FD_MCNT_SET( QUIC, CONNECTION_ERROR_NO_SLOTS, ctx->quic->metrics.conn_err_no_slots_cnt );
150 0 : FD_MCNT_SET( QUIC, CONNECTION_ERROR_RETRY_FAIL, ctx->quic->metrics.conn_err_retry_fail_cnt );
151 :
152 0 : FD_MCNT_ENUM_COPY( QUIC, PKT_CRYPTO_FAILED, ctx->quic->metrics.pkt_decrypt_fail_cnt );
153 0 : FD_MCNT_ENUM_COPY( QUIC, PKT_NO_KEY, ctx->quic->metrics.pkt_no_key_cnt );
154 0 : FD_MCNT_SET( QUIC, PKT_NO_CONN, ctx->quic->metrics.pkt_no_conn_cnt );
155 0 : FD_MCNT_ENUM_COPY( QUIC, FRAME_TX_ALLOC, ctx->quic->metrics.frame_tx_alloc_cnt );
156 0 : FD_MCNT_SET( QUIC, PKT_NET_HEADER_INVALID, ctx->quic->metrics.pkt_net_hdr_err_cnt );
157 0 : FD_MCNT_SET( QUIC, PKT_QUIC_HEADER_INVALID, ctx->quic->metrics.pkt_quic_hdr_err_cnt );
158 0 : FD_MCNT_SET( QUIC, PKT_UNDERSZ, ctx->quic->metrics.pkt_undersz_cnt );
159 0 : FD_MCNT_SET( QUIC, PKT_OVERSZ, ctx->quic->metrics.pkt_oversz_cnt );
160 0 : FD_MCNT_SET( QUIC, PKT_VERNEG, ctx->quic->metrics.pkt_verneg_cnt );
161 0 : FD_MCNT_SET( QUIC, PKT_RETRANSMISSIONS, ctx->quic->metrics.pkt_retransmissions_cnt );
162 :
163 0 : FD_MCNT_SET( QUIC, HANDSHAKES_CREATED, ctx->quic->metrics.hs_created_cnt );
164 0 : FD_MCNT_SET( QUIC, HANDSHAKE_ERROR_ALLOC_FAIL, ctx->quic->metrics.hs_err_alloc_fail_cnt );
165 0 : FD_MCNT_SET( QUIC, HANDSHAKE_EVICTED, ctx->quic->metrics.hs_evicted_cnt );
166 :
167 0 : FD_MCNT_SET( QUIC, STREAM_RECEIVED_EVENTS, ctx->quic->metrics.stream_rx_event_cnt );
168 0 : FD_MCNT_SET( QUIC, STREAM_RECEIVED_BYTES, ctx->quic->metrics.stream_rx_byte_cnt );
169 :
170 0 : FD_MCNT_ENUM_COPY( QUIC, RECEIVED_FRAMES, ctx->quic->metrics.frame_rx_cnt );
171 0 : FD_MCNT_SET ( QUIC, FRAME_FAIL_PARSE, ctx->quic->metrics.frame_rx_err_cnt );
172 :
173 0 : FD_MCNT_ENUM_COPY( QUIC, ACK_TX, ctx->quic->metrics.ack_tx );
174 :
175 0 : FD_MHIST_COPY( QUIC, SERVICE_DURATION_SECONDS, ctx->quic->metrics.service_duration );
176 0 : FD_MHIST_COPY( QUIC, RECEIVE_DURATION_SECONDS, ctx->quic->metrics.receive_duration );
177 0 : }
178 :
179 : static int
180 : before_frag( fd_quic_ctx_t * ctx,
181 : ulong in_idx,
182 : ulong seq,
183 0 : ulong sig ) {
184 0 : (void)in_idx;
185 0 : (void)seq;
186 :
187 0 : ulong proto = fd_disco_netmux_sig_proto( sig );
188 0 : if( FD_UNLIKELY( proto!=DST_PROTO_TPU_UDP && proto!=DST_PROTO_TPU_QUIC ) ) return 1;
189 :
190 0 : ulong hash = fd_disco_netmux_sig_hash( sig );
191 0 : if( FD_UNLIKELY( (hash % ctx->round_robin_cnt) != ctx->round_robin_id ) ) return 1;
192 :
193 0 : return 0;
194 0 : }
195 :
196 : static void
197 : during_frag( fd_quic_ctx_t * ctx,
198 : ulong in_idx,
199 : ulong seq FD_PARAM_UNUSED,
200 : ulong sig FD_PARAM_UNUSED,
201 : ulong chunk,
202 : ulong sz,
203 0 : ulong ctl ) {
204 0 : void const * src = fd_net_rx_translate_frag( &ctx->net_in_bounds[ in_idx ], chunk, ctl, sz );
205 :
206 : /* FIXME this copy could be eliminated by combining it with the decrypt operation */
207 0 : fd_memcpy( ctx->buffer, src, sz );
208 0 : }
209 :
210 : static void
211 : after_frag( fd_quic_ctx_t * ctx,
212 : ulong in_idx,
213 : ulong seq,
214 : ulong sig,
215 : ulong sz,
216 : ulong tsorig,
217 : ulong tspub,
218 0 : fd_stem_context_t * stem ) {
219 0 : (void)in_idx;
220 0 : (void)seq;
221 0 : (void)tsorig;
222 0 : (void)tspub;
223 0 : (void)stem;
224 :
225 0 : ulong proto = fd_disco_netmux_sig_proto( sig );
226 :
227 0 : if( FD_LIKELY( proto==DST_PROTO_TPU_QUIC ) ) {
228 0 : if( FD_UNLIKELY( sz<sizeof(fd_eth_hdr_t) ) ) FD_LOG_ERR(( "QUIC packet too small" ));
229 0 : uchar * ip_pkt = ctx->buffer + sizeof(fd_eth_hdr_t);
230 0 : ulong ip_sz = sz - sizeof(fd_eth_hdr_t);
231 :
232 0 : fd_quic_t * quic = ctx->quic;
233 0 : long dt = -fd_tickcount();
234 0 : fd_quic_process_packet( quic, ip_pkt, ip_sz );
235 0 : dt += fd_tickcount();
236 0 : fd_histf_sample( quic->metrics.receive_duration, (ulong)dt );
237 0 : quic->metrics.net_rx_byte_cnt += sz;
238 0 : quic->metrics.net_rx_pkt_cnt++;
239 0 : } else if( FD_LIKELY( proto==DST_PROTO_TPU_UDP ) ) {
240 0 : ulong network_hdr_sz = fd_disco_netmux_sig_hdr_sz( sig );
241 0 : if( FD_UNLIKELY( sz<=network_hdr_sz ) ) {
242 : /* Transaction not valid if the packet isn't large enough for the network
243 : headers. */
244 0 : ctx->metrics.udp_pkt_too_small++;
245 0 : return;
246 0 : }
247 :
248 0 : ulong data_sz = sz - network_hdr_sz;
249 0 : if( FD_UNLIKELY( data_sz<FD_TXN_MIN_SERIALIZED_SZ ) ) {
250 : /* Smaller than the smallest possible transaction */
251 0 : ctx->metrics.udp_pkt_too_small++;
252 0 : return;
253 0 : }
254 :
255 0 : if( FD_UNLIKELY( data_sz>FD_TPU_MTU ) ) {
256 : /* Transaction couldn't possibly be valid if it's longer than transaction
257 : MTU so drop it. This is not required, as the txn will fail to parse,
258 : but it's a nice short circuit. */
259 0 : ctx->metrics.udp_pkt_too_large++;
260 0 : return;
261 0 : }
262 :
263 0 : legacy_stream_notify( ctx, ctx->buffer+network_hdr_sz, data_sz );
264 0 : }
265 0 : }
266 :
267 : static ulong
268 0 : quic_now( void * ctx FD_PARAM_UNUSED ) {
269 0 : return (ulong)fd_tickcount();
270 0 : }
271 :
272 : static void
273 : quic_conn_final( fd_quic_conn_t * conn,
274 0 : void * quic_ctx ) {
275 0 : fd_quic_ctx_t * ctx = quic_ctx;
276 0 : long abandon_cnt = fd_long_max( conn->srx->rx_streams_active, 0L );
277 0 : ctx->metrics.reasm_active -= abandon_cnt;
278 0 : ctx->metrics.reasm_abandoned += (ulong)abandon_cnt;
279 0 : }
280 :
281 : static int
282 : quic_stream_rx( fd_quic_conn_t * conn,
283 : ulong stream_id,
284 : ulong offset,
285 : uchar const * data,
286 : ulong data_sz,
287 0 : int fin ) {
288 :
289 0 : long tspub = fd_tickcount();
290 0 : fd_quic_t * quic = conn->quic;
291 0 : fd_quic_state_t * state = fd_quic_get_state( quic ); /* ugly */
292 0 : fd_quic_ctx_t * ctx = quic->cb.quic_ctx;
293 0 : fd_tpu_reasm_t * reasm = ctx->reasm;
294 0 : ulong conn_uid = fd_quic_conn_uid( conn );
295 0 : fd_stem_context_t * stem = ctx->stem;
296 0 : fd_frag_meta_t * mcache = stem->mcaches[0];
297 0 : void * base = ctx->verify_out_mem;
298 0 : ulong seq = stem->seqs[0];
299 :
300 0 : int oversz = offset+data_sz > FD_TPU_MTU;
301 :
302 0 : if( offset==0UL && fin ) {
303 : /* Fast path */
304 0 : if( FD_UNLIKELY( data_sz<FD_TXN_MIN_SERIALIZED_SZ ) ) {
305 0 : ctx->metrics.quic_txn_too_small++;
306 0 : return FD_QUIC_SUCCESS; /* drop */
307 0 : }
308 0 : if( FD_UNLIKELY( oversz ) ) {
309 0 : ctx->metrics.quic_txn_too_large++;
310 0 : return FD_QUIC_SUCCESS; /* drop */
311 0 : }
312 0 : int err = fd_tpu_reasm_publish_fast( reasm, data, data_sz, mcache, base, seq, tspub );
313 0 : if( FD_LIKELY( err==FD_TPU_REASM_SUCCESS ) ) {
314 0 : fd_stem_advance( stem, 0UL );
315 0 : ctx->metrics.txns_received_quic_fast++;
316 0 : }
317 0 : return FD_QUIC_SUCCESS;
318 0 : }
319 :
320 0 : if( data_sz==0UL && !fin ) return FD_QUIC_SUCCESS; /* noop */
321 :
322 0 : fd_tpu_reasm_slot_t * slot = fd_tpu_reasm_query( reasm, conn_uid, stream_id );
323 :
324 0 : if( !slot ) { /* start a new reassembly */
325 0 : if( offset>0 ) {
326 0 : ctx->metrics.frag_gap_cnt++;
327 0 : return FD_QUIC_SUCCESS;
328 0 : }
329 0 : if( data_sz==0 ) return FD_QUIC_SUCCESS; /* ignore empty */
330 0 : if( FD_UNLIKELY( oversz ) ) {
331 0 : ctx->metrics.quic_txn_too_large++;
332 0 : return FD_QUIC_SUCCESS; /* drop */
333 0 : }
334 :
335 : /* Was the reasm buffer we evicted busy? */
336 0 : fd_tpu_reasm_slot_t * victim = fd_tpu_reasm_peek_tail( reasm );
337 0 : int victim_busy = victim->k.state == FD_TPU_REASM_STATE_BUSY;
338 :
339 : /* If so, does the connection it refers to still exist?
340 : (Or was the buffer previously abandoned by means of conn close) */
341 0 : uint victim_cidx = fd_quic_conn_uid_idx( victim->k.conn_uid );
342 0 : uint victim_gen = fd_quic_conn_uid_gen( victim->k.conn_uid );
343 0 : fd_quic_conn_t * victim_conn = fd_quic_conn_at_idx( state, victim_cidx ); /* possibly oob */
344 0 : if( victim_busy ) {
345 0 : uint victim_exists = (victim_conn->conn_gen == victim_gen) &
346 0 : (victim_conn->state == FD_QUIC_CONN_STATE_ACTIVE); /* in [0,1] */
347 0 : victim_conn->srx->rx_streams_active -= victim_exists;
348 0 : ctx->metrics.reasm_overrun += victim_exists;
349 0 : ctx->metrics.reasm_active -= victim_exists;
350 0 : }
351 :
352 0 : slot = fd_tpu_reasm_prepare( reasm, conn_uid, stream_id, tspub ); /* infallible */
353 0 : ctx->metrics.reasm_started++;
354 0 : ctx->metrics.reasm_active++;
355 0 : conn->srx->rx_streams_active++;
356 0 : } else if( slot->k.state != FD_TPU_REASM_STATE_BUSY ) {
357 0 : ctx->metrics.frag_dup_cnt++;
358 0 : return FD_QUIC_SUCCESS;
359 0 : }
360 :
361 0 : int reasm_res = fd_tpu_reasm_frag( reasm, slot, data, data_sz, offset );
362 0 : if( FD_UNLIKELY( reasm_res != FD_TPU_REASM_SUCCESS ) ) {
363 0 : int is_gap = reasm_res==FD_TPU_REASM_ERR_SKIP;
364 0 : int is_oversz = reasm_res==FD_TPU_REASM_ERR_SZ;
365 0 : ctx->metrics.frag_gap_cnt += (ulong)is_gap;
366 0 : ctx->metrics.quic_txn_too_large += (ulong)is_oversz;
367 0 : return is_gap ? FD_QUIC_FAILED : FD_QUIC_SUCCESS;
368 0 : }
369 0 : ctx->metrics.frag_ok_cnt++;
370 :
371 0 : if( fin ) {
372 0 : if( FD_UNLIKELY( slot->k.sz < FD_TXN_MIN_SERIALIZED_SZ ) ) {
373 0 : ctx->metrics.quic_txn_too_small++;
374 0 : return FD_QUIC_SUCCESS; /* ignore */
375 0 : }
376 0 : int pub_err = fd_tpu_reasm_publish( reasm, slot, mcache, base, seq, tspub );
377 0 : if( FD_UNLIKELY( pub_err!=FD_TPU_REASM_SUCCESS ) ) return FD_QUIC_SUCCESS; /* unreachable */
378 0 : ulong * rcv_cnt = (offset==0UL && fin) ? &ctx->metrics.txns_received_quic_fast : &ctx->metrics.txns_received_quic_frag;
379 0 : (*rcv_cnt)++;
380 0 : ctx->metrics.reasm_active--;
381 0 : conn->srx->rx_streams_active--;
382 :
383 0 : fd_stem_advance( stem, 0UL );
384 0 : }
385 :
386 0 : return FD_QUIC_SUCCESS;
387 0 : }
388 :
389 : static int
390 : quic_tx_aio_send( void * _ctx,
391 : fd_aio_pkt_info_t const * batch,
392 : ulong batch_cnt,
393 : ulong * opt_batch_idx,
394 0 : int flush ) {
395 0 : (void)flush;
396 :
397 0 : fd_quic_ctx_t * ctx = _ctx;
398 :
399 0 : for( ulong i=0; i<batch_cnt; i++ ) {
400 0 : if( FD_UNLIKELY( batch[ i ].buf_sz<FD_NETMUX_SIG_MIN_HDR_SZ ) ) continue;
401 :
402 0 : uint const ip_dst = FD_LOAD( uint, batch[ i ].buf+offsetof( fd_ip4_hdr_t, daddr_c ) );
403 0 : uchar * packet_l2 = fd_chunk_to_laddr( ctx->net_out_mem, ctx->net_out_chunk );
404 0 : uchar * packet_l3 = packet_l2 + sizeof(fd_eth_hdr_t);
405 0 : memset( packet_l2, 0, 12 );
406 0 : FD_STORE( ushort, packet_l2+offsetof( fd_eth_hdr_t, net_type ), fd_ushort_bswap( FD_ETH_HDR_TYPE_IP ) );
407 0 : fd_memcpy( packet_l3, batch[ i ].buf, batch[ i ].buf_sz );
408 0 : ulong sz_l2 = sizeof(fd_eth_hdr_t) + batch[ i ].buf_sz;
409 :
410 : /* send packets are just round-robined by sequence number, so for now
411 : just indicate where they came from so they don't bounce back */
412 0 : ulong sig = fd_disco_netmux_sig( ip_dst, 0U, ip_dst, DST_PROTO_OUTGOING, FD_NETMUX_SIG_MIN_HDR_SZ );
413 :
414 0 : long tspub = fd_tickcount();
415 0 : fd_mcache_publish( ctx->net_out_mcache,
416 0 : ctx->net_out_depth,
417 0 : ctx->net_out_seq,
418 0 : sig,
419 0 : ctx->net_out_chunk,
420 0 : sz_l2,
421 0 : fd_frag_meta_ctl( 0UL, 1, 1, 0 ),
422 0 : 0,
423 0 : fd_frag_meta_ts_comp( tspub ) );
424 :
425 0 : ctx->net_out_seq = fd_seq_inc( ctx->net_out_seq, 1UL );
426 0 : ctx->net_out_chunk = fd_dcache_compact_next( ctx->net_out_chunk, FD_NET_MTU, ctx->net_out_chunk0, ctx->net_out_wmark );
427 0 : }
428 :
429 0 : if( FD_LIKELY( opt_batch_idx ) ) {
430 0 : *opt_batch_idx = batch_cnt;
431 0 : }
432 :
433 0 : return FD_AIO_SUCCESS;
434 0 : }
435 :
436 : static void
437 : privileged_init( fd_topo_t * topo,
438 0 : fd_topo_tile_t * tile ) {
439 0 : (void)topo; (void)tile;
440 :
441 : /* The fd_quic implementation calls fd_log_wallclock() internally
442 : which itself calls clock_gettime() which on most kernels is not a
443 : real syscall but a virtual one in the process via. the vDSO.
444 :
445 : The first time this virtual call is made to the vDSO it does an
446 : mmap(2) of some shared memory into userspace, which cannot
447 : happen while sandboxed so we need to ensure that initialization
448 : happens here. */
449 :
450 0 : fd_log_wallclock();
451 0 : }
452 :
453 : static void
454 : quic_tls_cv_sign( void * signer_ctx,
455 : uchar signature[ static 64 ],
456 0 : uchar const payload[ static 130 ] ) {
457 0 : fd_quic_ctx_t * ctx = signer_ctx;
458 0 : fd_sha512_t * sha512 = fd_sha512_join( ctx->sha512 );
459 0 : fd_ed25519_sign( signature, payload, 130UL, ctx->tls_pub_key, ctx->tls_priv_key, sha512 );
460 0 : fd_sha512_leave( sha512 );
461 0 : }
462 :
463 : static void
464 : unprivileged_init( fd_topo_t * topo,
465 0 : fd_topo_tile_t * tile ) {
466 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
467 0 : if( FD_UNLIKELY( topo->objs[ tile->tile_obj_id ].footprint < scratch_footprint( tile ) ) ) {
468 0 : FD_LOG_ERR(( "insufficient tile scratch space" ));
469 0 : }
470 :
471 0 : if( FD_UNLIKELY( tile->in_cnt==0 ) ) {
472 0 : FD_LOG_ERR(( "quic tile has no input links" ));
473 0 : }
474 0 : if( FD_UNLIKELY( tile->in_cnt > FD_QUIC_TILE_IN_MAX ) ) {
475 0 : FD_LOG_ERR(( "quic tile has too many input links (%lu), max %lu",
476 0 : tile->in_cnt, FD_QUIC_TILE_IN_MAX ));
477 0 : }
478 :
479 0 : if( FD_UNLIKELY( tile->out_cnt!=2UL ||
480 0 : strcmp( topo->links[ tile->out_link_id[ 0UL ] ].name, "quic_verify" ) ||
481 0 : strcmp( topo->links[ tile->out_link_id[ 1UL ] ].name, "quic_net" ) ) )
482 0 : FD_LOG_ERR(( "quic tile has none or unexpected output links %lu %s %s",
483 0 : tile->out_cnt, topo->links[ tile->out_link_id[ 0 ] ].name, topo->links[ tile->out_link_id[ 1 ] ].name ));
484 :
485 0 : ulong out_depth = topo->links[ tile->out_link_id[ 0 ] ].depth;
486 0 : if( FD_UNLIKELY( tile->quic.out_depth != out_depth ) )
487 0 : FD_LOG_ERR(( "tile->quic.out_depth (%u) does not match quic_verify link depth (%lu)",
488 0 : tile->quic.out_depth, out_depth ));
489 :
490 0 : void * txn_dcache = topo->links[ tile->out_link_id[ 0UL ] ].dcache;
491 0 : if( FD_UNLIKELY( !txn_dcache ) ) FD_LOG_ERR(( "Missing output dcache" ));
492 :
493 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
494 0 : fd_quic_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_quic_ctx_t ), sizeof( fd_quic_ctx_t ) );
495 0 : fd_memset( ctx, 0, sizeof(fd_quic_ctx_t) );
496 :
497 0 : for( ulong i=0; i<tile->in_cnt; i++ ) {
498 0 : fd_topo_link_t * link = &topo->links[ tile->in_link_id[ i ] ];
499 0 : if( FD_UNLIKELY( 0!=strcmp( link->name, "net_quic" ) ) ) {
500 0 : FD_LOG_ERR(( "unexpected input link %s", link->name ));
501 0 : }
502 0 : fd_net_rx_bounds_init( &ctx->net_in_bounds[ i ], link->dcache );
503 0 : }
504 :
505 0 : if( FD_UNLIKELY( getrandom( ctx->tls_priv_key, ED25519_PRIV_KEY_SZ, 0 )!=ED25519_PRIV_KEY_SZ ) ) {
506 0 : FD_LOG_ERR(( "getrandom failed (%i-%s)", errno, fd_io_strerror( errno ) ));
507 0 : }
508 0 : fd_sha512_t * sha512 = fd_sha512_join( fd_sha512_new( ctx->sha512 ) );
509 0 : fd_ed25519_public_from_private( ctx->tls_pub_key, ctx->tls_priv_key, sha512 );
510 0 : fd_sha512_leave( sha512 );
511 :
512 0 : fd_aio_t * quic_tx_aio = fd_aio_join( fd_aio_new( ctx->quic_tx_aio, ctx, quic_tx_aio_send ) );
513 0 : if( FD_UNLIKELY( !quic_tx_aio ) ) FD_LOG_ERR(( "fd_aio_join failed" ));
514 :
515 0 : fd_quic_limits_t limits = quic_limits( tile );
516 0 : fd_quic_t * quic = fd_quic_join( fd_quic_new( FD_SCRATCH_ALLOC_APPEND( l, fd_quic_align(), fd_quic_footprint( &limits ) ), &limits ) );
517 0 : if( FD_UNLIKELY( !quic ) ) FD_LOG_ERR(( "fd_quic_new failed" ));
518 :
519 0 : ulong orig = 0UL; /* fd_tango origin ID */
520 0 : ulong reasm_max = tile->quic.reasm_cnt;
521 0 : void * reasm_mem = FD_SCRATCH_ALLOC_APPEND( l, fd_tpu_reasm_align(), fd_tpu_reasm_footprint( out_depth, reasm_max ) );
522 0 : ctx->reasm = fd_tpu_reasm_join( fd_tpu_reasm_new( reasm_mem, out_depth, reasm_max, orig, txn_dcache ) );
523 0 : if( FD_UNLIKELY( !ctx->reasm ) ) FD_LOG_ERR(( "fd_tpu_reasm_new failed" ));
524 :
525 0 : if( FD_UNLIKELY( tile->quic.ack_delay_millis == 0 ) ) {
526 0 : FD_LOG_ERR(( "Invalid `ack_delay_millis`: must be greater than zero" ));
527 0 : }
528 0 : if( FD_UNLIKELY( tile->quic.ack_delay_millis >= tile->quic.idle_timeout_millis ) ) {
529 0 : FD_LOG_ERR(( "Invalid `ack_delay_millis`: must be lower than `idle_timeout_millis`" ));
530 0 : }
531 :
532 0 : quic->config.role = FD_QUIC_ROLE_SERVER;
533 0 : quic->config.idle_timeout = tile->quic.idle_timeout_millis * (ulong)1e6;
534 0 : quic->config.ack_delay = tile->quic.ack_delay_millis * (ulong)1e6;
535 0 : quic->config.initial_rx_max_stream_data = FD_TXN_MTU;
536 0 : quic->config.retry = tile->quic.retry;
537 0 : fd_memcpy( quic->config.identity_public_key, ctx->tls_pub_key, ED25519_PUB_KEY_SZ );
538 :
539 0 : quic->config.sign = quic_tls_cv_sign;
540 0 : quic->config.sign_ctx = ctx;
541 :
542 0 : quic->cb.conn_final = quic_conn_final;
543 0 : quic->cb.stream_rx = quic_stream_rx;
544 0 : quic->cb.now = quic_now;
545 0 : quic->cb.now_ctx = ctx;
546 0 : quic->cb.quic_ctx = ctx;
547 :
548 0 : fd_quic_set_aio_net_tx( quic, quic_tx_aio );
549 0 : fd_quic_set_clock_tickcount( quic );
550 0 : if( FD_UNLIKELY( !fd_quic_init( quic ) ) ) FD_LOG_ERR(( "fd_quic_init failed" ));
551 :
552 0 : fd_topo_link_t * net_out = &topo->links[ tile->out_link_id[ 1 ] ];
553 :
554 0 : ctx->net_out_mcache = net_out->mcache;
555 0 : ctx->net_out_sync = fd_mcache_seq_laddr( ctx->net_out_mcache );
556 0 : ctx->net_out_depth = fd_mcache_depth( ctx->net_out_mcache );
557 0 : ctx->net_out_seq = fd_mcache_seq_query( ctx->net_out_sync );
558 0 : ctx->net_out_mem = topo->workspaces[ topo->objs[ net_out->dcache_obj_id ].wksp_id ].wksp;
559 0 : ctx->net_out_chunk0 = fd_dcache_compact_chunk0( ctx->net_out_mem, net_out->dcache );
560 0 : ctx->net_out_wmark = fd_dcache_compact_wmark ( ctx->net_out_mem, net_out->dcache, net_out->mtu );
561 0 : ctx->net_out_chunk = ctx->net_out_chunk0;
562 :
563 0 : fd_topo_link_t * verify_out = &topo->links[ tile->out_link_id[ 0 ] ];
564 :
565 0 : ctx->verify_out_mem = topo->workspaces[ topo->objs[ verify_out->dcache_obj_id ].wksp_id ].wksp;
566 :
567 0 : ctx->quic = quic;
568 :
569 0 : ctx->round_robin_cnt = fd_topo_tile_name_cnt( topo, tile->name );
570 0 : ctx->round_robin_id = tile->kind_id;
571 0 : if( FD_UNLIKELY( ctx->round_robin_id >= ctx->round_robin_cnt ) ) {
572 0 : FD_LOG_ERR(( "invalid round robin configuration" ));
573 0 : }
574 :
575 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, 1UL );
576 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
577 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
578 :
579 0 : fd_histf_join( fd_histf_new( ctx->quic->metrics.service_duration, FD_MHIST_SECONDS_MIN( QUIC, SERVICE_DURATION_SECONDS ),
580 0 : FD_MHIST_SECONDS_MAX( QUIC, SERVICE_DURATION_SECONDS ) ) );
581 0 : fd_histf_join( fd_histf_new( ctx->quic->metrics.receive_duration, FD_MHIST_SECONDS_MIN( QUIC, RECEIVE_DURATION_SECONDS ),
582 0 : FD_MHIST_SECONDS_MAX( QUIC, RECEIVE_DURATION_SECONDS ) ) );
583 0 : }
584 :
585 : static ulong
586 : populate_allowed_seccomp( fd_topo_t const * topo,
587 : fd_topo_tile_t const * tile,
588 : ulong out_cnt,
589 0 : struct sock_filter * out ) {
590 0 : (void)topo;
591 0 : (void)tile;
592 :
593 0 : populate_sock_filter_policy_quic( out_cnt, out, (uint)fd_log_private_logfile_fd() );
594 0 : return sock_filter_policy_quic_instr_cnt;
595 0 : }
596 :
597 : static ulong
598 : populate_allowed_fds( fd_topo_t const * topo,
599 : fd_topo_tile_t const * tile,
600 : ulong out_fds_cnt,
601 0 : int * out_fds ) {
602 0 : (void)topo;
603 0 : (void)tile;
604 :
605 0 : if( FD_UNLIKELY( out_fds_cnt<2UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
606 :
607 0 : ulong out_cnt = 0UL;
608 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
609 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
610 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
611 0 : return out_cnt;
612 0 : }
613 :
614 0 : #define STEM_BURST (1UL)
615 :
616 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_quic_ctx_t
617 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_quic_ctx_t)
618 :
619 0 : #define STEM_CALLBACK_DURING_HOUSEKEEPING during_housekeeping
620 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
621 0 : #define STEM_CALLBACK_BEFORE_CREDIT before_credit
622 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
623 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
624 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
625 :
626 : #include "../stem/fd_stem.c"
627 :
628 : fd_topo_run_tile_t fd_tile_quic = {
629 : .name = "quic",
630 : .populate_allowed_seccomp = populate_allowed_seccomp,
631 : .populate_allowed_fds = populate_allowed_fds,
632 : .scratch_align = scratch_align,
633 : .scratch_footprint = scratch_footprint,
634 : .privileged_init = privileged_init,
635 : .unprivileged_init = unprivileged_init,
636 : .run = stem_run,
637 : };
|
