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