Line data Source code
1 : #define _GNU_SOURCE /* dup3 */
2 : #include "fd_sock_tile_private.h"
3 : #include "../../fd_disco_base.h"
4 : #include "../../../discof/repair/fd_repair.h"
5 : #include "../../topo/fd_topo.h"
6 : #include "../../../util/net/fd_eth.h"
7 : #include "../../../util/net/fd_ip4.h"
8 : #include "../../../util/net/fd_udp.h"
9 :
10 : #include <stdalign.h> /* alignof */
11 : #include <errno.h>
12 : #include <fcntl.h> /* fcntl */
13 : #include <unistd.h> /* dup3, close */
14 : #include <netinet/in.h> /* sockaddr_in */
15 : #include <net/if.h> /* if_nametoindex */
16 : #include <linux/filter.h> /* SO_ATTACH_FILTER */
17 : #include <sys/socket.h> /* socket */
18 : #include "../../metrics/fd_metrics.h"
19 :
20 : #include "generated/fd_sock_tile_seccomp.h"
21 :
22 : /* recv/sendmmsg packet count in batch and tango burst depth
23 : FIXME make configurable in the future?
24 : FIXME keep in sync with fd_net_tile_topo.c */
25 0 : #define STEM_BURST (64UL)
26 :
27 : /* Place RX socket file descriptors in contiguous integer range. */
28 0 : #define RX_SOCK_FD_MIN (128)
29 :
30 : /* Controls max ancillary data size.
31 : Must be aligned by alignof(struct cmsghdr) */
32 0 : #define FD_SOCK_CMSG_MAX (64UL)
33 :
34 : /* Musl's CMSG_NXTHDR macro compares size_t against ptrdiff_t,
35 : triggering -Wsign-compare. */
36 : #pragma GCC diagnostic push
37 : #pragma GCC diagnostic ignored "-Wsign-compare"
38 : #pragma GCC diagnostic ignored "-Wsign-conversion"
39 : static inline struct cmsghdr *
40 : fd_cmsg_nxthdr( struct msghdr * mhdr,
41 0 : struct cmsghdr * cmsg ) {
42 0 : return CMSG_NXTHDR( mhdr, cmsg );
43 0 : }
44 : #pragma GCC diagnostic pop
45 :
46 : static ulong
47 : populate_allowed_seccomp( fd_topo_t const * topo,
48 : fd_topo_tile_t const * tile,
49 : ulong out_cnt,
50 0 : struct sock_filter * out ) {
51 0 : FD_SCRATCH_ALLOC_INIT( l, fd_topo_obj_laddr( topo, tile->tile_obj_id ) );
52 0 : fd_sock_tile_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sock_tile_t), sizeof(fd_sock_tile_t) );
53 :
54 0 : populate_sock_filter_policy_fd_sock_tile( out_cnt, out, (uint)fd_log_private_logfile_fd(), (uint)ctx->tx_sock, RX_SOCK_FD_MIN, RX_SOCK_FD_MIN+(uint)ctx->sock_cnt );
55 0 : return sock_filter_policy_fd_sock_tile_instr_cnt;
56 0 : }
57 :
58 : static ulong
59 : populate_allowed_fds( fd_topo_t const * topo,
60 : fd_topo_tile_t const * tile,
61 : ulong out_fds_cnt,
62 0 : int * out_fds ) {
63 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
64 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
65 0 : fd_sock_tile_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sock_tile_t), sizeof(fd_sock_tile_t) );
66 :
67 0 : ulong sock_cnt = ctx->sock_cnt;
68 0 : if( FD_UNLIKELY( out_fds_cnt<sock_cnt+3UL ) ) {
69 0 : FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
70 0 : }
71 :
72 0 : ulong out_cnt = 0UL;
73 :
74 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
75 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) ) {
76 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
77 0 : }
78 0 : if( -1!=ctx->tx_sock ) out_fds[ out_cnt++ ] = ctx->tx_sock;
79 0 : for( ulong j=0UL; j<sock_cnt; j++ ) {
80 0 : out_fds[ out_cnt++ ] = ctx->pollfd[ j ].fd;
81 0 : }
82 0 : return out_cnt;
83 0 : }
84 :
85 : FD_FN_CONST static inline ulong
86 0 : tx_scratch_footprint( void ) {
87 0 : return STEM_BURST * fd_ulong_align_up( FD_NET_MTU, FD_CHUNK_ALIGN );
88 0 : }
89 :
90 : FD_FN_CONST static inline ulong
91 0 : scratch_align( void ) {
92 0 : return 4096UL;
93 0 : }
94 :
95 : FD_FN_PURE static inline ulong
96 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
97 0 : ulong l = FD_LAYOUT_INIT;
98 0 : l = FD_LAYOUT_APPEND( l, alignof(fd_sock_tile_t), sizeof(fd_sock_tile_t) );
99 0 : l = FD_LAYOUT_APPEND( l, alignof(struct iovec), STEM_BURST*sizeof(struct iovec) );
100 0 : l = FD_LAYOUT_APPEND( l, alignof(struct cmsghdr), STEM_BURST*FD_SOCK_CMSG_MAX );
101 0 : l = FD_LAYOUT_APPEND( l, alignof(struct sockaddr_in), STEM_BURST*sizeof(struct sockaddr_in) );
102 0 : l = FD_LAYOUT_APPEND( l, alignof(struct mmsghdr), STEM_BURST*sizeof(struct mmsghdr) );
103 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong), STEM_BURST*sizeof(ulong) );
104 0 : if( !tile->sock.only_recv_lo ) l = FD_LAYOUT_APPEND( l, FD_CHUNK_ALIGN, tx_scratch_footprint() );
105 0 : return FD_LAYOUT_FINI( l, scratch_align() );
106 0 : }
107 :
108 : /* create_udp_socket creates and configures a new UDP socket for the
109 : sock tile at the given file descriptor ID. */
110 :
111 : static void
112 : create_udp_socket( int sock_fd,
113 : uint bind_addr,
114 : ushort udp_port,
115 : int so_rcvbuf,
116 0 : uint lo_ifindex ) {
117 :
118 0 : if( fcntl( sock_fd, F_GETFD, 0 )!=-1 ) {
119 0 : FD_LOG_ERR(( "file descriptor %d already exists", sock_fd ));
120 0 : } else if( errno!=EBADF ) {
121 0 : FD_LOG_ERR(( "fcntl(F_GETFD) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
122 0 : }
123 :
124 0 : int orig_fd = socket( AF_INET, SOCK_DGRAM, IPPROTO_UDP );
125 0 : if( FD_UNLIKELY( orig_fd<0 ) ) {
126 0 : FD_LOG_ERR(( "socket(AF_INET,SOCK_DGRAM,IPPROTO_UDP) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
127 0 : }
128 :
129 0 : int reuseport = 1;
130 0 : if( FD_UNLIKELY( setsockopt( orig_fd, SOL_SOCKET, SO_REUSEPORT, &reuseport, sizeof(int) )<0 ) ) {
131 0 : FD_LOG_ERR(( "setsockopt(SOL_SOCKET,SO_REUSEPORT,1) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
132 0 : }
133 :
134 0 : int ip_pktinfo = 1;
135 0 : if( FD_UNLIKELY( setsockopt( orig_fd, IPPROTO_IP, IP_PKTINFO, &ip_pktinfo, sizeof(int) )<0 ) ) {
136 0 : FD_LOG_ERR(( "setsockopt(IPPROTO_IP,IP_PKTINFO,1) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
137 0 : }
138 :
139 0 : if( FD_UNLIKELY( 0!=setsockopt( orig_fd, SOL_SOCKET, SO_RCVBUF, &so_rcvbuf, sizeof(int) ) ) ) {
140 0 : FD_LOG_ERR(( "setsockopt(SOL_SOCKET,SO_RCVBUF,%i) failed (%i-%s)", so_rcvbuf, errno, fd_io_strerror( errno ) ));
141 0 : }
142 :
143 0 : struct sockaddr_in saddr = {
144 0 : .sin_family = AF_INET,
145 0 : .sin_addr.s_addr = bind_addr,
146 0 : .sin_port = fd_ushort_bswap( udp_port ),
147 0 : };
148 0 : if( lo_ifindex ) {
149 : /* SKF_AD_IFINDEX identifies the receive device, including local
150 : traffic addressed to a non loopback interface's IP address. */
151 0 : struct sock_filter filter[] = {
152 0 : BPF_STMT( BPF_LD|BPF_W|BPF_ABS, (uint)(SKF_AD_OFF+SKF_AD_IFINDEX) ),
153 0 : BPF_JUMP( BPF_JMP|BPF_JEQ|BPF_K, lo_ifindex, 0, 1 ),
154 0 : BPF_STMT( BPF_RET|BPF_K, UINT_MAX ),
155 0 : BPF_STMT( BPF_RET|BPF_K, 0 )
156 0 : };
157 0 : struct sock_fprog prog = { .len = 4, .filter = filter };
158 0 : if( FD_UNLIKELY( 0!=setsockopt( orig_fd, SOL_SOCKET, SO_ATTACH_FILTER, &prog, sizeof(prog) ) ) ) {
159 0 : FD_LOG_ERR(( "setsockopt(SO_ATTACH_FILTER) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
160 0 : }
161 0 : }
162 0 : if( FD_UNLIKELY( 0!=bind( orig_fd, fd_type_pun_const( &saddr ), sizeof(struct sockaddr_in) ) ) ) {
163 0 : FD_LOG_ERR(( "bind(0.0.0.0:%i) failed (%i-%s)", udp_port, errno, fd_io_strerror( errno ) ));
164 0 : }
165 :
166 0 : if( orig_fd==sock_fd ) {
167 0 : if( FD_UNLIKELY( fcntl( sock_fd, F_SETFD, FD_CLOEXEC ) ) ) {
168 0 : FD_LOG_ERR(( "fcntl(F_SETFD) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
169 0 : }
170 0 : return;
171 0 : }
172 0 : int dup_res = dup3( orig_fd, sock_fd, O_CLOEXEC );
173 0 : if( FD_UNLIKELY( dup_res!=sock_fd ) ) {
174 0 : FD_LOG_ERR(( "dup3 returned %i (%i-%s)", dup_res, errno, fd_io_strerror( errno ) ));
175 0 : }
176 :
177 0 : if( FD_UNLIKELY( 0!=close( orig_fd ) ) ) {
178 0 : FD_LOG_ERR(( "close(%d) failed (%i-%s)", orig_fd, errno, fd_io_strerror( errno ) ));
179 0 : }
180 0 : }
181 :
182 : static void
183 : privileged_init( fd_topo_t const * topo,
184 0 : fd_topo_tile_t const * tile ) {
185 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
186 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
187 0 : fd_sock_tile_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sock_tile_t), sizeof(fd_sock_tile_t) );
188 0 : struct iovec * batch_iov = FD_SCRATCH_ALLOC_APPEND( l, alignof(struct iovec), STEM_BURST*sizeof(struct iovec) );
189 0 : void * batch_cmsg = FD_SCRATCH_ALLOC_APPEND( l, alignof(struct cmsghdr), STEM_BURST*FD_SOCK_CMSG_MAX );
190 0 : struct sockaddr_in * batch_sa = FD_SCRATCH_ALLOC_APPEND( l, alignof(struct sockaddr_in), STEM_BURST*sizeof(struct sockaddr_in) );
191 0 : struct mmsghdr * batch_msg = FD_SCRATCH_ALLOC_APPEND( l, alignof(struct mmsghdr), STEM_BURST*sizeof(struct mmsghdr) );
192 0 : ulong * batch_fsz = FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), STEM_BURST*sizeof(ulong) );
193 0 : ulong tx_sz = tile->sock.only_recv_lo ? 0UL : tx_scratch_footprint();
194 0 : uchar * tx_scratch = FD_SCRATCH_ALLOC_APPEND( l, FD_CHUNK_ALIGN, tx_sz );
195 0 : FD_DCHECK_CRIT( scratch==ctx, "invalid layout" );
196 :
197 0 : fd_memset( ctx, 0, sizeof(fd_sock_tile_t) );
198 0 : fd_memset( batch_iov, 0, STEM_BURST*sizeof(struct iovec) );
199 0 : fd_memset( batch_sa, 0, STEM_BURST*sizeof(struct sockaddr_in) );
200 0 : fd_memset( batch_msg, 0, STEM_BURST*sizeof(struct mmsghdr) );
201 :
202 0 : ctx->batch_cnt = 0UL;
203 0 : ctx->batch_iov = batch_iov;
204 0 : ctx->batch_cmsg = batch_cmsg;
205 0 : ctx->batch_sa = batch_sa;
206 0 : ctx->batch_msg = batch_msg;
207 0 : ctx->batch_frame_sz = batch_fsz;
208 0 : ctx->tx_scratch0 = tx_scratch;
209 0 : ctx->tx_scratch1 = tx_scratch + tx_sz;
210 0 : ctx->tx_ptr = tx_scratch;
211 0 : ctx->repair_shred_sock_idx = UINT_MAX;
212 0 : ctx->tx_sock = -1;
213 0 : ctx->bind_address = tile->sock.net.bind_address;
214 0 : ctx->tick_per_ns = fd_tempo_tick_per_ns( NULL );
215 0 : if( tile->sock.only_recv_lo ) {
216 0 : ctx->lo_ifindex = if_nametoindex( "lo" );
217 0 : if( FD_UNLIKELY( !ctx->lo_ifindex ) ) {
218 0 : FD_LOG_ERR(( "if_nametoindex(lo) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
219 0 : }
220 0 : }
221 :
222 : /* Create receive sockets. Incrementally assign them to file
223 : descriptors starting at sock_fd_min. */
224 :
225 0 : int sock_fd_min = RX_SOCK_FD_MIN;
226 0 : ushort udp_port_candidates[] = {
227 0 : (ushort)tile->sock.net.legacy_transaction_listen_port,
228 0 : (ushort)tile->sock.net.quic_transaction_listen_port,
229 0 : (ushort)tile->sock.net.shred_listen_port,
230 0 : (ushort)tile->sock.net.gossip_listen_port,
231 0 : (ushort)tile->sock.net.repair_client_listen_port,
232 0 : (ushort)tile->sock.net.repair_serve_listen_port,
233 0 : (ushort)tile->sock.net.votor_quic_client_listen_port,
234 0 : (ushort)tile->sock.net.votor_quic_server_listen_port
235 0 : };
236 0 : static char const * udp_port_links[] = {
237 0 : "net_quic", /* legacy_transaction_listen_port */
238 0 : "net_quic", /* quic_transaction_listen_port */
239 0 : "net_shred", /* shred_listen_port (turbine) */
240 0 : "net_gossvf", /* gossip_listen_port */
241 0 : "net_shred", /* shred_listen_port (repair) */
242 0 : "net_rserve", /* repair_serve_listen_port */
243 0 : "net_votor", /* votor_quic_client_listen_port */
244 0 : "net_votor" /* votor_quic_server_listen_port */
245 0 : };
246 0 : static uchar const udp_port_protos[] = {
247 0 : DST_PROTO_TPU_UDP, /* legacy_transaction_listen_port */
248 0 : DST_PROTO_TPU_QUIC, /* quic_transaction_listen_port */
249 0 : DST_PROTO_SHRED, /* shred_listen_port (turbine) */
250 0 : DST_PROTO_GOSSIP, /* gossip_listen_port */
251 0 : DST_PROTO_REPAIR, /* shred_listen_port (repair) */
252 0 : DST_PROTO_RSERVE, /* repair_serve_listen_port */
253 0 : DST_PROTO_VOTOR, /* votor_quic_client_listen_port */
254 : DST_PROTO_VOTOR /* votor_quic_server_listen_port */
255 0 : };
256 0 : for( uint candidate_idx=0U; candidate_idx<sizeof(udp_port_candidates)/sizeof(udp_port_candidates[0]); candidate_idx++ ) {
257 0 : if( !udp_port_candidates[ candidate_idx ] ) continue;
258 0 : uint sock_idx = ctx->sock_cnt;
259 0 : if( sock_idx>=FD_SOCK_TILE_MAX_SOCKETS ) FD_LOG_ERR(( "too many sockets" ));
260 0 : ushort port = (ushort)udp_port_candidates[ candidate_idx ];
261 :
262 0 : char const * target_link = udp_port_links[ candidate_idx ];
263 0 : ctx->link_rx_map[ sock_idx ] = 0xFF;
264 0 : for( ulong j=0UL; j<(tile->out_cnt); j++ ) {
265 0 : if( 0==strcmp( topo->links[ tile->out_link_id[ j ] ].name, target_link ) ) {
266 0 : ctx->proto_id [ sock_idx ] = (uchar)udp_port_protos[ candidate_idx ];
267 0 : ctx->link_rx_map [ sock_idx ] = (uchar)j;
268 0 : ctx->rx_sock_port[ sock_idx ] = (ushort)port;
269 0 : break;
270 0 : }
271 0 : }
272 0 : if( ctx->link_rx_map[ sock_idx ]==0xFF ) {
273 : /* listen port number has no associated links,
274 : i.e. the repair server is disabled, then no net_rserve link. */
275 0 : continue;
276 0 : }
277 :
278 : /* Record the socket index of the repair intake socket, so repair
279 : ping packets can be routed to the repair tile at runtime. */
280 0 : if( tile->sock.net.repair_client_listen_port &&
281 0 : udp_port_candidates[ candidate_idx ]==tile->sock.net.repair_client_listen_port )
282 0 : ctx->repair_shred_sock_idx = sock_idx;
283 :
284 0 : int sock_fd = sock_fd_min + (int)sock_idx;
285 0 : create_udp_socket( sock_fd, tile->sock.net.bind_address, port, tile->sock.so_rcvbuf, ctx->lo_ifindex );
286 0 : ctx->pollfd[ sock_idx ].fd = sock_fd;
287 0 : ctx->pollfd[ sock_idx ].events = POLLIN;
288 0 : ctx->sock_cnt++;
289 0 : }
290 :
291 0 : if( tile->sock.only_recv_lo ) return;
292 :
293 : /* Create transmit socket */
294 :
295 0 : int tx_sock = socket( AF_INET, SOCK_RAW|SOCK_CLOEXEC, FD_IP4_HDR_PROTOCOL_UDP );
296 0 : if( FD_UNLIKELY( tx_sock<0 ) ) {
297 0 : FD_LOG_ERR(( "socket(AF_INET,SOCK_RAW|SOCK_CLOEXEC,17) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
298 0 : }
299 :
300 0 : if( FD_UNLIKELY( 0!=setsockopt( tx_sock, SOL_SOCKET, SO_SNDBUF, &tile->sock.so_sndbuf, sizeof(int) ) ) ) {
301 0 : FD_LOG_ERR(( "setsockopt(SOL_SOCKET,SO_SNDBUF,%i) failed (%i-%s)", tile->sock.so_sndbuf, errno, fd_io_strerror( errno ) ));
302 0 : }
303 :
304 0 : uchar mcast_ttl = 64;
305 0 : if( FD_UNLIKELY( 0!=setsockopt( tx_sock, IPPROTO_IP, IP_MULTICAST_TTL, &mcast_ttl, sizeof(mcast_ttl) ) ) ) {
306 0 : FD_LOG_ERR(( "setsockopt(IPPROTO_IP,IP_MULTICAST_TTL,%u) failed (%i-%s)", (uint)mcast_ttl, errno, fd_io_strerror( errno ) ));
307 0 : }
308 :
309 0 : ctx->tx_sock = tx_sock;
310 0 : ctx->bind_address = tile->sock.net.bind_address;
311 0 : }
312 :
313 : static void
314 : unprivileged_init( fd_topo_t const * topo,
315 0 : fd_topo_tile_t const * tile ) {
316 0 : fd_sock_tile_t * ctx = fd_topo_obj_laddr( topo, tile->tile_obj_id );
317 :
318 0 : if( FD_UNLIKELY( tile->out_cnt > MAX_NET_OUTS ) ) {
319 0 : FD_LOG_ERR(( "sock tile has %lu out links which exceeds the max (%lu)", tile->out_cnt, MAX_NET_OUTS ));
320 0 : }
321 :
322 0 : ctx->link_rx_cnt = tile->out_cnt;
323 0 : ctx->repair_rx = 0xFF;
324 0 : for( ulong i=0UL; i<(tile->out_cnt); i++ ) {
325 0 : if( 0!=strncmp( topo->links[ tile->out_link_id[ i ] ].name, "net_", 4 ) ) {
326 0 : FD_LOG_ERR(( "out link %lu is not a net RX link", i ));
327 0 : }
328 0 : if( 0==strcmp( topo->links[ tile->out_link_id[ i ] ].name, "net_repair" ) ) {
329 0 : if( FD_UNLIKELY( ctx->repair_rx!=0xFF ) ) FD_LOG_ERR(( "multiple net_repair out links" ));
330 0 : ctx->repair_rx = (uchar)i;
331 0 : }
332 0 : fd_topo_link_t const * link = &topo->links[ tile->out_link_id[ i ] ];
333 0 : ctx->link_rx[ i ].base = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
334 0 : ctx->link_rx[ i ].chunk0 = fd_dcache_compact_chunk0( ctx->link_rx[ i ].base, link->dcache );
335 0 : ctx->link_rx[ i ].wmark = fd_dcache_compact_wmark( ctx->link_rx[ i ].base, link->dcache, link->mtu );
336 0 : ctx->link_rx[ i ].chunk = ctx->link_rx[ i ].chunk0;
337 0 : if( FD_UNLIKELY( link->burst < STEM_BURST ) ) {
338 0 : FD_LOG_ERR(( "link %lu dcache burst is too low (%lu<%lu)",
339 0 : tile->out_link_id[ i ], link->burst, STEM_BURST ));
340 0 : }
341 0 : }
342 :
343 0 : if( FD_UNLIKELY( ctx->repair_shred_sock_idx!=UINT_MAX && ctx->repair_rx==0xFF ) ) {
344 0 : FD_LOG_ERR(( "repair intake socket configured but no net_repair out link was found" ));
345 0 : }
346 :
347 0 : FD_CHECK_ERR( tile->in_cnt<=MAX_NET_INS, "too many input links" );
348 :
349 : /* The loopback-only sock tile has no TX links and is not counted */
350 0 : ctx->tx_tile_idx = 0UL;
351 0 : ctx->tx_tile_cnt = 0UL;
352 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
353 0 : fd_topo_tile_t const * t = &topo->tiles[ i ];
354 0 : if( strcmp( t->name, "sock" ) || t->sock.only_recv_lo ) continue;
355 0 : if( t->id==tile->id ) ctx->tx_tile_idx = ctx->tx_tile_cnt;
356 0 : ctx->tx_tile_cnt++;
357 0 : }
358 0 : ctx->tx_tile_cnt = fd_ulong_max( ctx->tx_tile_cnt, 1UL );
359 :
360 0 : for( ulong i=0UL; i<(tile->in_cnt); i++ ) {
361 0 : if( !strstr( topo->links[ tile->in_link_id[ i ] ].name, "_net" ) ) {
362 0 : FD_LOG_ERR(( "in link %lu is not a net TX link", i ));
363 0 : }
364 0 : fd_topo_link_t const * link = &topo->links[ tile->in_link_id[ i ] ];
365 0 : ctx->link_tx[ i ].base = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
366 0 : ctx->link_tx[ i ].chunk0 = fd_dcache_compact_chunk0( ctx->link_tx[ i ].base, link->dcache );
367 0 : ctx->link_tx[ i ].wmark = fd_dcache_compact_wmark( ctx->link_tx[ i ].base, link->dcache, link->mtu );
368 0 : }
369 :
370 0 : }
371 :
372 : /* RX PATH (socket->tango) ********************************************/
373 :
374 : /* FIXME Pace RX polling and interleave it with TX jobs to reduce TX
375 : tail latency */
376 :
377 : /* poll_rx_socket does one recvmmsg batch receive on the given socket
378 : index. Returns the number of packets returned by recvmmsg. */
379 :
380 : static ulong
381 : poll_rx_socket( fd_sock_tile_t * ctx,
382 : fd_stem_context_t * stem,
383 : uint sock_idx,
384 : int sock_fd,
385 0 : ushort proto ) {
386 0 : ulong hdr_sz = sizeof(fd_eth_hdr_t) + sizeof(fd_ip4_hdr_t) + sizeof(fd_udp_hdr_t);
387 0 : ulong payload_max = FD_NET_MTU-hdr_sz;
388 0 : uchar rx_link = ctx->link_rx_map[ sock_idx ];
389 0 : ushort dport = ctx->rx_sock_port[ sock_idx ];
390 :
391 0 : fd_sock_link_rx_t * link = ctx->link_rx + rx_link;
392 0 : void * const base = link->base;
393 0 : ulong const chunk0 = link->chunk0;
394 0 : ulong const wmark = link->wmark;
395 0 : ulong chunk_next = link->chunk;
396 0 : uchar * cmsg_next = ctx->batch_cmsg;
397 :
398 0 : for( ulong j=0UL; j<STEM_BURST; j++ ) {
399 0 : ctx->batch_iov[ j ].iov_base = (uchar *)fd_chunk_to_laddr( base, chunk_next ) + hdr_sz;
400 0 : ctx->batch_iov[ j ].iov_len = payload_max;
401 0 : ctx->batch_msg[ j ].msg_hdr = (struct msghdr) {
402 0 : .msg_iov = ctx->batch_iov+j,
403 0 : .msg_iovlen = 1,
404 0 : .msg_name = ctx->batch_sa+j,
405 0 : .msg_namelen = sizeof(struct sockaddr_in),
406 0 : .msg_control = cmsg_next,
407 0 : .msg_controllen = FD_SOCK_CMSG_MAX,
408 0 : };
409 0 : cmsg_next += FD_SOCK_CMSG_MAX;
410 : /* Speculatively prepare all chunk indexes for a receive.
411 : At function exit, chunks into which a packet was received are
412 : committed, all others are freed. */
413 0 : chunk_next = fd_dcache_compact_next( chunk_next, FD_NET_MTU, chunk0, wmark );
414 0 : }
415 :
416 0 : int msg_cnt = recvmmsg( sock_fd, ctx->batch_msg, STEM_BURST, MSG_DONTWAIT, NULL );
417 0 : if( FD_UNLIKELY( msg_cnt<0 ) ) {
418 0 : if( FD_UNLIKELY( errno==EINTR ) ) return 0UL;
419 0 : if( FD_LIKELY( errno==EAGAIN ) ) return 0UL;
420 : /* unreachable if socket is in a valid state */
421 0 : FD_LOG_ERR(( "recvmmsg failed (%i-%s)", errno, fd_io_strerror( errno ) ));
422 0 : }
423 0 : long ts = fd_tickcount();
424 0 : ctx->metrics.sys_recvmmsg_cnt++;
425 :
426 0 : if( FD_UNLIKELY( msg_cnt==0 ) ) return 0UL;
427 :
428 : /* Track the chunk index of the last frag populated, so we can derive
429 : the chunk indexes for the next poll_rx_socket call.
430 : Guaranteed to be set since msg_cnt>0. */
431 0 : ulong last_chunk;
432 :
433 0 : for( ulong j=0; j<(ulong)msg_cnt; j++ ) {
434 0 : uchar * payload = ctx->batch_iov[ j ].iov_base;
435 0 : ulong payload_sz = ctx->batch_msg[ j ].msg_len;
436 0 : struct sockaddr_in * sa = ctx->batch_msg[ j ].msg_hdr.msg_name;
437 0 : ulong frame_sz = payload_sz + hdr_sz;
438 0 : last_chunk = fd_laddr_to_chunk( base, payload-hdr_sz );
439 0 : if( FD_UNLIKELY( ctx->lo_ifindex &&
440 0 : (ctx->batch_msg[ j ].msg_hdr.msg_flags & (MSG_TRUNC|MSG_CTRUNC)) ) ) continue;
441 0 : ctx->metrics.rx_bytes_total += frame_sz;
442 0 : if( FD_UNLIKELY( sa->sin_family!=AF_INET ) ) {
443 : /* unreachable */
444 0 : FD_LOG_ERR(( "Received packet with unexpected sin_family %i", sa->sin_family ));
445 0 : }
446 :
447 0 : long daddr = -1;
448 0 : struct cmsghdr * cmsg = CMSG_FIRSTHDR( &ctx->batch_msg[ j ].msg_hdr );
449 0 : if( FD_LIKELY( cmsg ) ) {
450 0 : do {
451 0 : if( FD_LIKELY( (cmsg->cmsg_level==IPPROTO_IP) &
452 0 : (cmsg->cmsg_type ==IP_PKTINFO) ) ) {
453 0 : struct in_pktinfo const * pi = (struct in_pktinfo const *)CMSG_DATA( cmsg );
454 0 : daddr = pi->ipi_addr.s_addr;
455 0 : }
456 0 : cmsg = fd_cmsg_nxthdr( &ctx->batch_msg[ j ].msg_hdr, cmsg );
457 0 : } while( FD_UNLIKELY( cmsg ) ); /* optimize for 1 cmsg */
458 0 : }
459 0 : if( FD_UNLIKELY( daddr<0L ) ) {
460 : /* unreachable because IP_PKTINFO was set */
461 0 : FD_LOG_ERR(( "Missing IP_PKTINFO on incoming packet" ));
462 0 : }
463 :
464 0 : fd_eth_hdr_t * eth_hdr = (fd_eth_hdr_t *)( payload-42UL );
465 0 : fd_ip4_hdr_t * ip_hdr = (fd_ip4_hdr_t *)( payload-28UL );
466 0 : fd_udp_hdr_t * udp_hdr = (fd_udp_hdr_t *)( payload- 8UL );
467 0 : memset( eth_hdr->dst, 0, 6 );
468 0 : memset( eth_hdr->src, 0, 6 );
469 0 : eth_hdr->net_type = fd_ushort_bswap( FD_ETH_HDR_TYPE_IP );
470 0 : *ip_hdr = (fd_ip4_hdr_t) {
471 0 : .verihl = FD_IP4_VERIHL( 4, 5 ),
472 0 : .net_tot_len = fd_ushort_bswap( (ushort)( payload_sz+28UL ) ),
473 0 : .ttl = 1,
474 0 : .protocol = FD_IP4_HDR_PROTOCOL_UDP,
475 0 : };
476 0 : uint daddr_ = (uint)(ulong)daddr;
477 0 : memcpy( ip_hdr->saddr_c, &sa->sin_addr.s_addr, 4 );
478 0 : memcpy( ip_hdr->daddr_c, &daddr_, 4 );
479 0 : *udp_hdr = (fd_udp_hdr_t) {
480 0 : .net_sport = sa->sin_port,
481 0 : .net_dport = (ushort)fd_ushort_bswap( (ushort)dport ),
482 0 : .net_len = (ushort)fd_ushort_bswap( (ushort)( payload_sz+8UL ) ),
483 0 : .check = 0
484 0 : };
485 :
486 0 : ctx->metrics.rx_pkt_cnt++;
487 0 : ulong chunk = fd_laddr_to_chunk( base, eth_hdr );
488 0 : ulong sig = fd_disco_netmux_sig( sa->sin_addr.s_addr, fd_ushort_bswap( sa->sin_port ), sa->sin_addr.s_addr, proto, hdr_sz );
489 0 : ulong tspub = fd_frag_meta_ts_comp( ts );
490 :
491 : /* When a message arrives on the repair intake port, it is sent to
492 : the shred tile, unless it is a ping message or an alpenglow
493 : repair response (identified by the frame size), then it is sent
494 : to the repair tile. The repair tile does not own any sockets, so
495 : we look up the net_repair link directly. */
496 0 : if( FD_UNLIKELY( sock_idx==ctx->repair_shred_sock_idx && payload_sz<=AG_REPAIR_RESPONSE_MAX_SZ ) ) {
497 0 : fd_sock_link_rx_t * repair_link = ctx->link_rx + ctx->repair_rx;
498 0 : uchar * repair_buf = fd_chunk_to_laddr( repair_link->base, repair_link->chunk );
499 0 : memcpy( repair_buf, eth_hdr, frame_sz );
500 0 : fd_stem_publish( stem, ctx->repair_rx, sig, repair_link->chunk, frame_sz, 0UL, 0UL, tspub );
501 0 : repair_link->chunk = fd_dcache_compact_next( repair_link->chunk, FD_NET_MTU, repair_link->chunk0, repair_link->wmark );
502 0 : } else {
503 0 : fd_stem_publish( stem, rx_link, sig, chunk, frame_sz, 0UL, 0UL, tspub );
504 0 : }
505 :
506 0 : last_chunk = chunk;
507 0 : }
508 :
509 : /* Rewind the chunk index to the first free index. */
510 0 : link->chunk = fd_dcache_compact_next( last_chunk, FD_NET_MTU, chunk0, wmark );
511 0 : return (ulong)msg_cnt;
512 0 : }
513 :
514 : static ulong
515 : poll_rx( fd_sock_tile_t * ctx,
516 0 : fd_stem_context_t * stem ) {
517 0 : ulong pkt_cnt = 0UL;
518 0 : if( FD_UNLIKELY( ctx->batch_cnt ) ) {
519 0 : FD_LOG_ERR(( "Batch is not clean" ));
520 0 : }
521 0 : ctx->tx_idle_cnt = 0; /* restart TX polling */
522 0 : long wait_ns = ctx->lo_ifindex ? (long)((double)fd_long_max( stem->then-fd_tickcount(), 0L )/ctx->tick_per_ns) : 0L;
523 0 : struct timespec ts = { .tv_sec = wait_ns/1000000000L, .tv_nsec = wait_ns%1000000000L };
524 0 : if( FD_UNLIKELY( ppoll( ctx->pollfd, ctx->sock_cnt, &ts, NULL )<0 ) ) {
525 0 : if( FD_UNLIKELY( errno==EINTR ) ) return 0UL;
526 0 : FD_LOG_ERR(( "ppoll failed (%i-%s)", errno, fd_io_strerror( errno ) ));
527 0 : }
528 0 : for( uint j=0U; j<ctx->sock_cnt; j++ ) {
529 0 : if( ctx->pollfd[ j ].revents & (POLLIN|POLLERR) ) {
530 0 : pkt_cnt += poll_rx_socket(
531 0 : ctx,
532 0 : stem,
533 0 : j,
534 0 : ctx->pollfd[ j ].fd,
535 0 : ctx->proto_id[ j ]
536 0 : );
537 0 : }
538 0 : ctx->pollfd[ j ].revents = 0;
539 0 : }
540 0 : return pkt_cnt;
541 0 : }
542 :
543 : /* TX PATH (tango->socket) ********************************************/
544 :
545 : static void
546 0 : flush_tx_batch( fd_sock_tile_t * ctx ) {
547 0 : ulong batch_cnt = ctx->batch_cnt;
548 0 : for( int j = 0; j < (int)batch_cnt; /* incremented in loop */ ) {
549 0 : int remain = (int)batch_cnt - j;
550 0 : int send_cnt = sendmmsg( ctx->tx_sock, ctx->batch_msg + j, (uint)remain, MSG_DONTWAIT );
551 0 : if( send_cnt>=0 ) {
552 0 : ctx->metrics.sys_sendmmsg_cnt[ FD_METRICS_ENUM_SENDMMSG_RESULT_V_SUCCESS_IDX ]++;
553 0 : }
554 0 : if( FD_UNLIKELY( send_cnt < remain ) ) {
555 0 : ctx->metrics.tx_drop_cnt++;
556 0 : if( FD_UNLIKELY( send_cnt < 0 ) ) {
557 0 : switch( errno ) {
558 0 : case EAGAIN:
559 0 : case ENOBUFS:
560 0 : ctx->metrics.sys_sendmmsg_cnt[ FD_METRICS_ENUM_SENDMMSG_RESULT_V_BUFFER_FULL_IDX ]++;
561 0 : break;
562 0 : case EPERM:
563 0 : ctx->metrics.sys_sendmmsg_cnt[ FD_METRICS_ENUM_SENDMMSG_RESULT_V_PERMISSION_IDX ]++;
564 0 : break;
565 0 : case ENETUNREACH:
566 0 : case EHOSTUNREACH:
567 0 : ctx->metrics.sys_sendmmsg_cnt[ FD_METRICS_ENUM_SENDMMSG_RESULT_V_UNREACHABLE_IDX ]++;
568 0 : break;
569 0 : case ENONET:
570 0 : case ENETDOWN:
571 0 : case EHOSTDOWN:
572 0 : ctx->metrics.sys_sendmmsg_cnt[ FD_METRICS_ENUM_SENDMMSG_RESULT_V_DOWN_IDX ]++;
573 0 : break;
574 0 : default:
575 0 : ctx->metrics.sys_sendmmsg_cnt[ FD_METRICS_ENUM_SENDMMSG_RESULT_V_OTHER_IDX ]++;
576 : /* log with NOTICE, since flushing has a significant negative performance impact */
577 0 : FD_LOG_NOTICE(( "sendmmsg failed (%i-%s)", errno, fd_io_strerror( errno ) ));
578 0 : }
579 :
580 : /* first message failed, so skip failing message and continue */
581 0 : j++;
582 0 : } else {
583 : /* send_cnt succeeded, so skip those and also the failing message */
584 0 : j += send_cnt + 1;
585 :
586 : /* add the successful count */
587 0 : ctx->metrics.tx_pkt_cnt += (ulong)send_cnt;
588 0 : for( int k=j-send_cnt-1; k<j-1; k++ ) ctx->metrics.tx_bytes_total += ctx->batch_frame_sz[ k ];
589 0 : }
590 :
591 0 : continue;
592 0 : }
593 :
594 : /* send_cnt == batch_cnt, so we sent everything */
595 0 : ctx->metrics.tx_pkt_cnt += (ulong)send_cnt;
596 0 : for( int k=j; k<j+send_cnt; k++ ) ctx->metrics.tx_bytes_total += ctx->batch_frame_sz[ k ];
597 0 : break;
598 0 : }
599 :
600 0 : ctx->tx_ptr = ctx->tx_scratch0;
601 0 : ctx->batch_cnt = 0;
602 0 : }
603 :
604 : /* before_frag is called when a new frag has been detected. Every sock
605 : tile reads every TX link, so each sends only its share of packets. */
606 :
607 : static inline int
608 : before_frag( fd_sock_tile_t * ctx,
609 : ulong in_idx FD_PARAM_UNUSED,
610 : ulong seq FD_PARAM_UNUSED,
611 0 : ulong sig ) {
612 0 : ulong proto = fd_disco_netmux_sig_proto( sig );
613 0 : if( FD_UNLIKELY( proto!=DST_PROTO_OUTGOING ) ) return 1;
614 0 : if( fd_disco_netmux_sig_hash( sig )%ctx->tx_tile_cnt!=ctx->tx_tile_idx ) return 1;
615 0 : return 0; /* continue */
616 0 : }
617 :
618 : /* during_frag is called when a new frag passed early filtering.
619 : Speculatively copies data into a sendmmsg buffer. (If all tiles
620 : respect backpressure could eliminate this copy) */
621 :
622 : static inline void
623 : during_frag( fd_sock_tile_t * ctx,
624 : ulong in_idx,
625 : ulong seq FD_PARAM_UNUSED,
626 : ulong sig,
627 : ulong chunk,
628 : ulong sz,
629 0 : ulong ctl FD_PARAM_UNUSED ) {
630 0 : if( FD_UNLIKELY( chunk<ctx->link_tx[ in_idx ].chunk0 || chunk>ctx->link_tx[ in_idx ].wmark || sz>FD_NET_MTU ) ) {
631 0 : FD_LOG_ERR(( "chunk %lu %lu corrupt, not in range [%lu,%lu]", chunk, sz, ctx->link_tx[ in_idx ].chunk0, ctx->link_tx[ in_idx ].wmark ));
632 0 : }
633 :
634 0 : ctx->parsed.invalid = 0;
635 :
636 0 : ulong const hdr_min = sizeof(fd_eth_hdr_t)+sizeof(fd_ip4_hdr_t)+sizeof(fd_udp_hdr_t);
637 0 : if( FD_UNLIKELY( sz<hdr_min ) ) {
638 : /* FIXME support ICMP messages in the future?
639 : Defer the error to after_frag, where we know we weren't
640 : overrun. */
641 0 : ctx->parsed.invalid = 1;
642 0 : return;
643 0 : }
644 :
645 0 : uchar const * frame = fd_chunk_to_laddr_const( ctx->link_tx[ in_idx ].base, chunk );
646 0 : ulong hdr_sz = fd_disco_netmux_sig_hdr_sz( sig );
647 0 : uchar const * payload = frame+hdr_sz;
648 0 : if( FD_UNLIKELY( hdr_sz>sz || hdr_sz<hdr_min ) ) {
649 0 : FD_LOG_ERR(( "packet from in_idx=%lu corrupt: hdr_sz=%lu total_sz=%lu",
650 0 : in_idx, hdr_sz, sz ));
651 0 : }
652 0 : ulong payload_sz = sz-hdr_sz;
653 :
654 0 : fd_ip4_hdr_t const * ip_hdr = (fd_ip4_hdr_t const *)( frame +sizeof(fd_eth_hdr_t) );
655 0 : fd_udp_hdr_t const * udp_hdr = (fd_udp_hdr_t const *)( payload-sizeof(fd_udp_hdr_t) );
656 0 : ctx->parsed.ip_version = FD_IP4_GET_VERSION( *ip_hdr );
657 0 : ctx->parsed.ip_protocol = ip_hdr->protocol;
658 0 : if( FD_UNLIKELY( ( ctx->parsed.ip_version !=4 ) |
659 0 : ( ctx->parsed.ip_protocol!=FD_IP4_HDR_PROTOCOL_UDP ) ) ) {
660 : /* sock tile only supports IPv4 UDP for now. Defer the error to
661 : after_frag, where we know we weren't overrun. */
662 0 : ctx->parsed.invalid = 1;
663 0 : return;
664 0 : }
665 :
666 0 : ulong msg_sz = sizeof(fd_udp_hdr_t) + payload_sz;
667 :
668 0 : ulong batch_idx = ctx->batch_cnt;
669 0 : FD_DCHECK_CRIT( batch_idx<STEM_BURST, "flow control error" );
670 0 : struct mmsghdr * msg = ctx->batch_msg + batch_idx;
671 0 : struct sockaddr_in * sa = ctx->batch_sa + batch_idx;
672 0 : struct iovec * iov = ctx->batch_iov + batch_idx;
673 0 : struct cmsghdr * cmsg = (void *)( (ulong)ctx->batch_cmsg + batch_idx*FD_SOCK_CMSG_MAX );
674 0 : uchar * buf = ctx->tx_ptr;
675 :
676 0 : *iov = (struct iovec) {
677 0 : .iov_base = buf,
678 0 : .iov_len = msg_sz,
679 0 : };
680 0 : sa->sin_family = AF_INET;
681 0 : sa->sin_addr.s_addr = FD_LOAD( uint, ip_hdr->daddr_c );
682 0 : sa->sin_port = 0; /* ignored */
683 :
684 0 : cmsg->cmsg_level = IPPROTO_IP;
685 0 : cmsg->cmsg_type = IP_PKTINFO;
686 0 : cmsg->cmsg_len = CMSG_LEN( sizeof(struct in_pktinfo) );
687 0 : struct in_pktinfo * pi = (struct in_pktinfo *)CMSG_DATA( cmsg );
688 0 : pi->ipi_ifindex = 0;
689 0 : pi->ipi_addr.s_addr = 0;
690 0 : pi->ipi_spec_dst.s_addr = fd_uint_if( !!ip_hdr->saddr, ip_hdr->saddr, ctx->bind_address );
691 :
692 0 : *msg = (struct mmsghdr) {
693 0 : .msg_hdr = {
694 0 : .msg_name = sa,
695 0 : .msg_namelen = sizeof(struct sockaddr_in),
696 0 : .msg_iov = iov,
697 0 : .msg_iovlen = 1,
698 0 : .msg_control = cmsg,
699 0 : .msg_controllen = CMSG_LEN( sizeof(struct in_pktinfo) )
700 0 : }
701 0 : };
702 :
703 0 : memcpy( buf, udp_hdr, sizeof(fd_udp_hdr_t) );
704 0 : fd_memcpy( buf+sizeof(fd_udp_hdr_t), payload, payload_sz );
705 0 : ctx->batch_frame_sz[ batch_idx ] = sz;
706 0 : }
707 :
708 : /* after_frag is called when a frag was copied into a sendmmsg buffer. */
709 :
710 : static void
711 : after_frag( fd_sock_tile_t * ctx,
712 : ulong in_idx,
713 : ulong seq FD_PARAM_UNUSED,
714 : ulong sig FD_PARAM_UNUSED,
715 : ulong sz,
716 : ulong tsorig FD_PARAM_UNUSED,
717 : ulong tspub FD_PARAM_UNUSED,
718 0 : fd_stem_context_t * stem FD_PARAM_UNUSED ) {
719 : /* Commit the packet added in during_frag. during_frag defers fatal
720 : errors to here (where we know we weren't overrun) by setting the
721 : invalid flag. */
722 :
723 0 : if( FD_UNLIKELY( ctx->parsed.invalid ) ) {
724 0 : ulong const hdr_min = sizeof(fd_eth_hdr_t)+sizeof(fd_ip4_hdr_t)+sizeof(fd_udp_hdr_t);
725 0 : if( FD_UNLIKELY( sz<hdr_min ) ) {
726 : /* FIXME support ICMP messages in the future? */
727 0 : FD_LOG_ERR(( "packet too small %lu (in_idx=%lu)", sz, in_idx ));
728 0 : }
729 0 : if( FD_UNLIKELY( ( ctx->parsed.ip_version !=4 ) |
730 0 : ( ctx->parsed.ip_protocol!=FD_IP4_HDR_PROTOCOL_UDP ) ) ) {
731 0 : FD_LOG_ERR(( "packet from in_idx=%lu: sock tile only supports IPv4 UDP for now", in_idx ));
732 0 : }
733 0 : }
734 :
735 0 : ctx->tx_idle_cnt = 0;
736 0 : ctx->batch_cnt++;
737 : /* Technically leaves a gap. sz is always larger than the payload
738 : written to tx_ptr because Ethernet & IPv4 headers were stripped. */
739 0 : ctx->tx_ptr += fd_ulong_align_up( sz, FD_CHUNK_ALIGN );
740 :
741 0 : if( ctx->batch_cnt >= STEM_BURST ) {
742 0 : flush_tx_batch( ctx );
743 0 : }
744 0 : }
745 :
746 : /* End TX path ********************************************************/
747 :
748 : static void
749 0 : metrics_write( fd_sock_tile_t * ctx ) {
750 0 : FD_MCNT_SET( SOCK, SYSCALL_RX, ctx->metrics.sys_recvmmsg_cnt );
751 0 : FD_MCNT_ENUM_COPY( SOCK, SYSCALL_TX, ctx->metrics.sys_sendmmsg_cnt );
752 0 : FD_MCNT_SET( SOCK, PKT_RX, ctx->metrics.rx_pkt_cnt );
753 0 : FD_MCNT_SET( SOCK, PKT_TX, ctx->metrics.tx_pkt_cnt );
754 0 : FD_MCNT_SET( SOCK, PKT_TX_FAILED, ctx->metrics.tx_drop_cnt );
755 0 : FD_MCNT_SET( SOCK, PKT_TX_BYTES, ctx->metrics.tx_bytes_total );
756 0 : FD_MCNT_SET( SOCK, PKT_RX_BYTES, ctx->metrics.rx_bytes_total );
757 0 : }
758 :
759 : /* after_credit is called every stem iteration when there are enough
760 : flow control credits to publish a burst of fragments. */
761 :
762 : static inline void
763 : after_credit( fd_sock_tile_t * ctx,
764 : fd_stem_context_t * stem,
765 : int * poll_in FD_PARAM_UNUSED,
766 0 : int * charge_busy ) {
767 0 : if( ctx->lo_ifindex ) {
768 : /* Housekeeping cannot publish progress while the tile sleeps. */
769 0 : FD_COMPILER_MFENCE();
770 0 : metrics_write( ctx );
771 0 : FD_MGAUGE_SET( TILE, HEARTBEAT_TIME, (ulong)fd_log_wallclock() );
772 0 : for( ulong i=0UL; i<ctx->link_rx_cnt; i++ ) {
773 0 : fd_mcache_seq_update( fd_mcache_seq_laddr( stem->mcaches[ i ] ), stem->seqs[ i ] );
774 0 : }
775 0 : FD_COMPILER_MFENCE();
776 0 : *charge_busy |= poll_rx( ctx, stem )!=0UL;
777 0 : return;
778 0 : }
779 0 : if( ctx->tx_idle_cnt > 512 ) {
780 0 : if( ctx->batch_cnt ) {
781 0 : flush_tx_batch( ctx );
782 0 : }
783 0 : ulong pkt_cnt = poll_rx( ctx, stem );
784 0 : *charge_busy = pkt_cnt!=0;
785 0 : }
786 0 : ctx->tx_idle_cnt++;
787 0 : }
788 :
789 : static ulong
790 : rlimit_file_cnt( fd_topo_t const * topo,
791 0 : fd_topo_tile_t const * tile ) {
792 0 : fd_sock_tile_t * ctx = fd_topo_obj_laddr( topo, tile->tile_obj_id );
793 0 : return RX_SOCK_FD_MIN + ctx->sock_cnt;
794 0 : }
795 :
796 : #define STEM_NEVER_PARK 1
797 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_sock_tile_t
798 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_sock_tile_t)
799 :
800 0 : #define STEM_LAZY ((long)10e6) /* 10ms */
801 :
802 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
803 0 : #define STEM_CALLBACK_AFTER_CREDIT after_credit
804 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
805 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
806 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
807 :
808 : #include "../../stem/fd_stem.c"
809 :
810 : fd_topo_run_tile_t fd_tile_sock = {
811 : .name = "sock",
812 : .rlimit_file_cnt_fn = rlimit_file_cnt,
813 : .populate_allowed_seccomp = populate_allowed_seccomp,
814 : .populate_allowed_fds = populate_allowed_fds,
815 : .scratch_align = scratch_align,
816 : .scratch_footprint = scratch_footprint,
817 : .privileged_init = privileged_init,
818 : .unprivileged_init = unprivileged_init,
819 : .run = stem_run,
820 : };
|
