Line data Source code
1 : #include "fd_ipecho_client.h"
2 : #include "fd_ipecho_server.h"
3 : #include "fd_ipecho_server_port_check.h"
4 :
5 : #include "../genesis/fd_genesi_tile.h"
6 : #include "../genesis/genesis_hash.h"
7 : #include "../../disco/topo/fd_topo.h"
8 : #include "../../disco/topo/fd_dns_resolve.h"
9 : #include "../../disco/metrics/fd_metrics.h"
10 : #include "../../disco/waker/fd_waker.h"
11 : #include "../../disco/fd_clock_tile.h"
12 :
13 : #include <netinet/in.h>
14 : #include <sys/socket.h>
15 : #include <time.h> /* CLOCK_REALTIME for seccomp filter */
16 : #include <poll.h>
17 :
18 : #include <linux/futex.h>
19 : #include "generated/fd_ipecho_tile_seccomp.h"
20 :
21 0 : #define FD_IPECHO_MAX_CONNECTION_CNT (1024UL)
22 :
23 : struct fd_ipecho_tile_ctx {
24 : int retrieving;
25 :
26 : fd_ipecho_server_t * server;
27 : fd_ipecho_client_t * client;
28 :
29 : ulong waker_client_idx;
30 : ulong * waker_fseq;
31 :
32 : fd_clock_tile_t clock[1];
33 :
34 : uint bind_address;
35 : ushort bind_port;
36 :
37 : ushort expected_shred_version;
38 :
39 : fd_ip4_port_t entrypoints[ FD_TOPO_GOSSIP_ENTRYPOINTS_MAX ];
40 : ulong entrypoints_cnt;
41 :
42 : fd_wksp_t * genesi_in_mem;
43 : ulong genesi_in_chunk0;
44 : ulong genesi_in_wmark;
45 : };
46 :
47 : typedef struct fd_ipecho_tile_ctx fd_ipecho_tile_ctx_t;
48 :
49 : FD_FN_CONST static inline ulong
50 0 : scratch_align( void ) {
51 0 : return alignof( fd_ipecho_tile_ctx_t );
52 0 : }
53 :
54 : FD_FN_PURE static inline ulong
55 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
56 0 : (void)tile;
57 :
58 0 : ulong l = FD_LAYOUT_INIT;
59 0 : l = FD_LAYOUT_APPEND( l, alignof(fd_ipecho_tile_ctx_t), sizeof(fd_ipecho_tile_ctx_t) );
60 0 : l = FD_LAYOUT_APPEND( l, fd_ipecho_client_align(), fd_ipecho_client_footprint() );
61 0 : l = FD_LAYOUT_APPEND( l, fd_ipecho_server_align(), fd_ipecho_server_footprint( FD_IPECHO_MAX_CONNECTION_CNT ) );
62 0 : return FD_LAYOUT_FINI( l, scratch_align() );
63 0 : }
64 :
65 : static inline void
66 0 : metrics_write( fd_ipecho_tile_ctx_t * ctx ) {
67 0 : fd_ipecho_server_metrics_t * metrics = fd_ipecho_server_metrics( ctx->server );
68 :
69 0 : FD_MGAUGE_SET( IPECHO, CONN_OPEN, metrics->connection_cnt );
70 0 : FD_MCNT_SET( IPECHO, CONN_RX_BYTES, metrics->bytes_read );
71 0 : FD_MCNT_SET( IPECHO, CONN_TX_BYTES, metrics->bytes_written );
72 0 : ulong conn_closed[ FD_METRICS_ENUM_CONN_CLOSE_RESULT_CNT ];
73 0 : conn_closed[ FD_METRICS_ENUM_CONN_CLOSE_RESULT_V_ACCEPTED_IDX ] = metrics->connections_closed_ok;
74 0 : conn_closed[ FD_METRICS_ENUM_CONN_CLOSE_RESULT_V_ERROR_IDX ] = metrics->connections_closed_error;
75 0 : FD_MCNT_ENUM_COPY( IPECHO, CONN_CLOSED, conn_closed );
76 :
77 0 : fd_ipecho_server_port_check_metrics_t * pc_metrics = fd_ipecho_server_port_check_metrics(
78 0 : fd_ipecho_server_port_check( ctx->server ) );
79 0 : FD_MCNT_SET( IPECHO, PORT_CHECK_UDP_TX, pc_metrics->udp_sent );
80 0 : FD_MCNT_ENUM_COPY( IPECHO, PORT_CHECK_TCP_ATTEMPTED, pc_metrics->tcp );
81 0 : FD_MGAUGE_SET( IPECHO, PORT_CHECK_TCP_IN_FLIGHT, pc_metrics->active );
82 0 : }
83 :
84 : static inline void
85 : poll_client( fd_ipecho_tile_ctx_t * ctx,
86 : fd_stem_context_t * stem,
87 0 : int * charge_busy ) {
88 0 : if( FD_UNLIKELY( !ctx->client ) ) return;
89 :
90 0 : ushort shred_version;
91 0 : int result = fd_ipecho_client_poll( ctx->client, fd_clock_tile_now( ctx->clock ), &shred_version, charge_busy );
92 0 : if( FD_UNLIKELY( !result ) ) {
93 0 : if( FD_UNLIKELY( ctx->expected_shred_version && ctx->expected_shred_version!=shred_version ) ) {
94 0 : FD_LOG_ERR(( "Expected shred version %hu but entrypoint returned %hu",
95 0 : ctx->expected_shred_version, shred_version ));
96 0 : }
97 :
98 0 : FD_LOG_INFO(( "retrieved shred version %hu from entrypoint", shred_version ));
99 0 : FD_MGAUGE_SET( IPECHO, SHRED_VERSION, shred_version );
100 0 : fd_stem_publish( stem, 0UL, shred_version, 0UL, 0UL, 0UL, 0UL, 0UL );
101 0 : fd_ipecho_server_set_shred_version( ctx->server, shred_version );
102 0 : ctx->retrieving = 0;
103 0 : return;
104 0 : } else if( FD_UNLIKELY( -1==result ) ) {
105 0 : FD_LOG_ERR(( "Could not determine shred version from entrypoints. Please "
106 0 : "check you can connect to the entrypoints provided." ));
107 0 : }
108 0 : }
109 :
110 : static void
111 0 : during_housekeeping( fd_ipecho_tile_ctx_t * ctx ) {
112 0 : if( FD_UNLIKELY( fd_clock_tile_recal_due( ctx->clock ) ) ) fd_clock_tile_recal( ctx->clock );
113 0 : fd_ipecho_server_port_check_prune( fd_ipecho_server_port_check( ctx->server ), fd_tickcount() );
114 0 : }
115 :
116 : static inline long
117 0 : next_deadline( fd_ipecho_tile_ctx_t * ctx ) {
118 0 : if( FD_LIKELY( !ctx->retrieving || !ctx->client ) ) return LONG_MAX;
119 :
120 0 : long deadline = fd_ipecho_client_deadline_nanos( ctx->client );
121 0 : if( FD_UNLIKELY( deadline==LONG_MAX ) ) return 0L; /* first poll starts the clock */
122 0 : return fd_clock_tile_wallclock_to_tickcount( ctx->clock, deadline );
123 0 : }
124 :
125 : static inline void
126 : after_credit( fd_ipecho_tile_ctx_t * ctx,
127 : fd_stem_context_t * stem,
128 : int * opt_poll_in,
129 0 : int * charge_busy ) {
130 0 : (void)opt_poll_in;
131 :
132 0 : if( FD_UNLIKELY( ctx->retrieving ) ) {
133 0 : int fired = fd_fseq_query( ctx->waker_fseq )==1UL;
134 0 : if( FD_LIKELY( fired ) ) fd_fseq_update( ctx->waker_fseq, 0UL );
135 0 : poll_client( ctx, stem, charge_busy );
136 0 : if( FD_LIKELY( fired ) ) fd_waker_client_rearm( ctx->waker_client_idx );
137 0 : return;
138 0 : }
139 :
140 0 : if( FD_UNLIKELY( fd_fseq_query( ctx->waker_fseq )==1UL ) ) {
141 0 : fd_fseq_update( ctx->waker_fseq, 0UL );
142 0 : fd_ipecho_server_epoll_poll( ctx->server, fd_tickcount(), charge_busy ); /* one batch; the rearm re-fires leftovers */
143 0 : fd_waker_client_rearm( ctx->waker_client_idx );
144 0 : } else if( FD_LIKELY( !stem->sleep ) ) {
145 0 : fd_log_sleep( (long)1e6 );
146 0 : }
147 0 : }
148 :
149 : static inline int
150 : returnable_frag( fd_ipecho_tile_ctx_t * ctx,
151 : ulong in_idx,
152 : ulong seq,
153 : ulong sig,
154 : ulong chunk,
155 : ulong sz,
156 : ulong ctl,
157 : ulong tsorig,
158 : ulong tspub,
159 0 : fd_stem_context_t * stem ) {
160 0 : (void)in_idx; (void)seq; (void)sig; (void)sz; (void)ctl; (void)tspub;
161 0 : fd_genesis_meta_t const * genesis_meta = fd_chunk_to_laddr( ctx->genesi_in_mem, chunk );
162 :
163 0 : if( FD_UNLIKELY( genesis_meta->bootstrap ) ) {
164 0 : ushort shred_version = compute_shred_version( genesis_meta->genesis_hash.uc, NULL, 0UL );
165 0 : FD_TEST( shred_version );
166 :
167 0 : FD_MGAUGE_SET( IPECHO, SHRED_VERSION, shred_version );
168 0 : fd_stem_publish( stem, 0UL, shred_version, 0UL, 0UL, 0UL, tsorig, fd_frag_meta_ts_comp( fd_tickcount() ) );
169 0 : fd_ipecho_server_set_shred_version( ctx->server, shred_version );
170 0 : ctx->retrieving = 0;
171 0 : }
172 :
173 0 : return 0;
174 0 : }
175 :
176 : static void
177 : privileged_init( fd_topo_t const * topo,
178 0 : fd_topo_tile_t const * tile ) {
179 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
180 :
181 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
182 0 : fd_ipecho_tile_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_ipecho_tile_ctx_t ), sizeof( fd_ipecho_tile_ctx_t ) );
183 0 : void * _client = FD_SCRATCH_ALLOC_APPEND( l, fd_ipecho_client_align(), fd_ipecho_client_footprint() );
184 0 : void * _server = FD_SCRATCH_ALLOC_APPEND( l, fd_ipecho_server_align(), fd_ipecho_server_footprint( FD_IPECHO_MAX_CONNECTION_CNT ) );
185 :
186 0 : ctx->bind_address = tile->ipecho.bind_address;
187 0 : ctx->bind_port = tile->ipecho.bind_port;
188 :
189 0 : ctx->expected_shred_version = tile->ipecho.expected_shred_version;
190 :
191 0 : ctx->entrypoints_cnt = tile->ipecho.entrypoints_cnt;
192 0 : fd_dns_resolve_peers( tile->ipecho.entrypoints[ 0 ], sizeof(tile->ipecho.entrypoints[ 0 ]), tile->ipecho.entrypoints_cnt, "gossip.entrypoints", ctx->entrypoints );
193 :
194 0 : ctx->retrieving = 1;
195 0 : if( FD_LIKELY( ctx->entrypoints_cnt ) ) {
196 0 : ctx->client = fd_ipecho_client_join( fd_ipecho_client_new( _client ) );
197 0 : FD_TEST( ctx->client );
198 0 : fd_ipecho_client_init( ctx->client, ctx->entrypoints, ctx->entrypoints_cnt, FD_WAKER_INNER_FD( tile->waker_client_idx ) );
199 0 : } else {
200 0 : ctx->client = NULL;
201 0 : }
202 :
203 0 : ctx->server = fd_ipecho_server_join( fd_ipecho_server_new( _server, FD_IPECHO_MAX_CONNECTION_CNT ) );
204 0 : FD_TEST( ctx->server );
205 0 : ctx->waker_client_idx = tile->waker_client_idx;
206 0 : FD_TEST( ctx->waker_client_idx!=ULONG_MAX );
207 0 : fd_ipecho_server_init( ctx->server, FD_WAKER_INNER_FD( ctx->waker_client_idx ), ctx->bind_address, ctx->bind_port, ctx->expected_shred_version );
208 :
209 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, scratch_align() );
210 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
211 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
212 0 : }
213 :
214 : static void
215 : unprivileged_init( fd_topo_t const * topo,
216 0 : fd_topo_tile_t const * tile ) {
217 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
218 :
219 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
220 0 : fd_ipecho_tile_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_ipecho_tile_ctx_t ), sizeof( fd_ipecho_tile_ctx_t ) );
221 :
222 0 : FD_MGAUGE_SET( IPECHO, SHRED_VERSION, tile->ipecho.expected_shred_version );
223 :
224 0 : FD_TEST( ctx->waker_client_idx!=ULONG_MAX );
225 0 : ctx->waker_fseq = fd_fseq_join( fd_topo_obj_laddr( topo, tile->waker_fseq_obj_id ) );
226 0 : FD_TEST( ctx->waker_fseq );
227 :
228 0 : fd_clock_tile_init( ctx->clock );
229 :
230 : /* In some topologies (e.g. firedancer-dev gossip), the ipecho tile
231 : has no input links. Guard against dereferencing a missing
232 : link/dcache. */
233 0 : if( FD_LIKELY( tile->in_cnt>0UL ) ) {
234 0 : ulong link_id = tile->in_link_id[ 0UL ];
235 0 : void * dcache = topo->links[ link_id ].dcache;
236 0 : ctx->genesi_in_mem = topo->workspaces[ topo->objs[ topo->links[ link_id ].dcache_obj_id ].wksp_id ].wksp;
237 0 : ctx->genesi_in_chunk0 = fd_dcache_compact_chunk0( ctx->genesi_in_mem, dcache );
238 0 : ctx->genesi_in_wmark = fd_dcache_compact_wmark ( ctx->genesi_in_mem, dcache, topo->links[ link_id ].mtu );
239 0 : } else {
240 0 : ctx->genesi_in_mem = NULL;
241 0 : ctx->genesi_in_chunk0 = 0UL;
242 0 : ctx->genesi_in_wmark = 0UL;
243 0 : }
244 0 : }
245 :
246 : static ulong
247 : rlimit_file_cnt( fd_topo_t const * topo FD_PARAM_UNUSED,
248 0 : fd_topo_tile_t const * tile ) {
249 : /* stderr, logfile, the listen socket, the waker's inner and outer
250 : epoll fds, and one spare for new accept() connections */
251 0 : ulong base = 6UL;
252 0 : return base +
253 0 : tile->ipecho.entrypoints_cnt + /* for the client */
254 0 : 1UL + /* for the port check's UDP socket */
255 0 : FD_IPECHO_MAX_CONNECTION_CNT + /* for the server's connections */
256 0 : FD_IPECHO_SERVER_PORT_CHECK_CONN_MAX; /* for the port check's outgoing TCP connections */
257 0 : }
258 :
259 : static ulong
260 : populate_allowed_seccomp( fd_topo_t const * topo,
261 : fd_topo_tile_t const * tile,
262 : ulong out_cnt,
263 0 : struct sock_filter * out ) {
264 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
265 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
266 0 : fd_ipecho_tile_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_ipecho_tile_ctx_t), sizeof(fd_ipecho_tile_ctx_t) );
267 :
268 0 : uint epoll_inner_fd = (uint)FD_WAKER_INNER_FD( tile->waker_client_idx );
269 0 : uint epoll_outer_fd = (uint)FD_WAKER_OUTER_FD;
270 0 : uint port_check_udp_fd = (uint)fd_ipecho_server_port_check_udp_sockfd( fd_ipecho_server_port_check( ctx->server ) );
271 :
272 0 : populate_sock_filter_policy_fd_ipecho_tile( out_cnt, out, (uint)fd_log_private_logfile_fd(), epoll_inner_fd, epoll_outer_fd, port_check_udp_fd );
273 0 : return sock_filter_policy_fd_ipecho_tile_instr_cnt;
274 0 : }
275 :
276 : static ulong
277 : populate_allowed_fds( fd_topo_t const * topo,
278 : fd_topo_tile_t const * tile,
279 : ulong out_fds_cnt,
280 0 : int * out_fds ) {
281 :
282 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
283 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
284 0 : fd_ipecho_tile_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_ipecho_tile_ctx_t), sizeof(fd_ipecho_tile_ctx_t) );
285 :
286 0 : if( FD_UNLIKELY( out_fds_cnt<6UL+tile->ipecho.entrypoints_cnt ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
287 :
288 0 : ulong out_cnt = 0UL;
289 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
290 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
291 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
292 :
293 : /* All of the fds managed by the client. */
294 0 : for( ulong i=0UL; i<tile->ipecho.entrypoints_cnt; i++ ) {
295 0 : int fd = fd_ipecho_client_get_pollfds( ctx->client )[ i ].fd;
296 0 : if( FD_LIKELY( fd!=-1 ) ) out_fds[ out_cnt++ ] = fd;
297 0 : }
298 :
299 : /* The server's listen socket and the port check's UDP socket. */
300 0 : out_fds[ out_cnt++ ] = fd_ipecho_server_sockfd( ctx->server );
301 0 : out_fds[ out_cnt++ ] = fd_ipecho_server_port_check_udp_sockfd( fd_ipecho_server_port_check( ctx->server ) );
302 :
303 0 : out_fds[ out_cnt++ ] = FD_WAKER_OUTER_FD; /* waker outer epoll fd (rearm) */
304 0 : out_fds[ out_cnt++ ] = FD_WAKER_INNER_FD( tile->waker_client_idx ); /* waker inner epoll fd */
305 0 : return out_cnt;
306 0 : }
307 :
308 0 : #define STEM_BURST (2UL)
309 0 : #define STEM_LAZY ((long)10e6) /* 10ms */
310 :
311 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_ipecho_tile_ctx_t
312 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_ipecho_tile_ctx_t)
313 :
314 0 : #define STEM_CALLBACK_DURING_HOUSEKEEPING during_housekeeping
315 0 : #define STEM_CALLBACK_NEXT_DEADLINE next_deadline
316 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
317 0 : #define STEM_CALLBACK_AFTER_CREDIT after_credit
318 0 : #define STEM_CALLBACK_RETURNABLE_FRAG returnable_frag
319 :
320 : #include "../../disco/stem/fd_stem.c"
321 :
322 : fd_topo_run_tile_t fd_tile_ipecho = {
323 : .name = "ipecho",
324 : .rlimit_file_cnt_fn = rlimit_file_cnt,
325 : .populate_allowed_seccomp = populate_allowed_seccomp,
326 : .populate_allowed_fds = populate_allowed_fds,
327 : .scratch_align = scratch_align,
328 : .scratch_footprint = scratch_footprint,
329 : .privileged_init = privileged_init,
330 : .unprivileged_init = unprivileged_init,
331 : .run = stem_run,
332 : .allow_connect = 1,
333 : .keep_host_networking = 1
334 : };
|
