Line data Source code
1 : /* The xdp tile translates between AF_XDP and fd_tango
2 : traffic. It is responsible for setting up the XDP and
3 : XSK socket configuration. */
4 :
5 : #include <errno.h>
6 : #include <fcntl.h>
7 : #include <net/if.h>
8 : #include <sys/socket.h> /* MSG_DONTWAIT needed before importing the net seccomp filter */
9 : #include <linux/if_xdp.h>
10 :
11 : #include "../../metrics/fd_metrics.h"
12 : #include "../../netlink/fd_netlink_tile.h" /* neigh4_solicit */
13 : #include "../../topo/fd_topo.h"
14 :
15 : #include "../../../waltz/ip/fd_fib4.h"
16 : #include "../../../waltz/neigh/fd_neigh4_map.h"
17 : #include "../../../waltz/xdp/fd_xdp_redirect_user.h" /* fd_xsk_activate */
18 : #include "../../../waltz/xdp/fd_xsk.h"
19 : #include "../../../util/log/fd_dtrace.h"
20 : #include "../../../util/net/fd_eth.h"
21 : #include "../../../util/net/fd_ip4.h"
22 :
23 : #include <unistd.h>
24 : #include <linux/if.h> /* struct ifreq */
25 : #include <sys/ioctl.h>
26 : #include <linux/unistd.h>
27 :
28 : #include "generated/xdp_seccomp.h"
29 :
30 : /* MAX_NET_INS controls the max number of TX links that a net tile can
31 : serve. */
32 :
33 : #define MAX_NET_INS (32UL)
34 :
35 : /* FD_XDP_STATS_INTERVAL_NS controls the XDP stats refresh interval.
36 : This should be lower than the interval at which the metrics tile
37 : collects metrics. */
38 :
39 0 : #define FD_XDP_STATS_INTERVAL_NS (11e6) /* 11ms */
40 :
41 : /* fd_net_in_ctx_t contains consumer information for an incoming tango
42 : link. It is used as part of the TX path. */
43 :
44 : typedef struct {
45 : fd_wksp_t * mem;
46 : ulong chunk0;
47 : ulong wmark;
48 : } fd_net_in_ctx_t;
49 :
50 : /* fd_net_out_ctx_t contains publisher information for a link to a
51 : downstream app tile. It is used as part of the RX path. */
52 :
53 : typedef struct {
54 : fd_frag_meta_t * mcache;
55 : ulong * sync;
56 : ulong depth;
57 : ulong seq;
58 : } fd_net_out_ctx_t;
59 :
60 : /* fd_net_flusher_t controls the pacing of XDP sendto calls for flushing
61 : TX batches. In the 'wakeup' XDP mode, no TX occurs unless the net
62 : tile wakes up the kernel periodically using the sendto() syscall.
63 : If sendto() is called too frequently, time is wasted on context
64 : switches. If sendto() is called not often enough, packets are
65 : delayed or dropped. sendto() calls make almost no guarantees how
66 : much packets are sent out, nor do they indicate when the kernel
67 : finishes a wakeup call (asynchronously dispatched). The net tile
68 : thus uses a myraid of flush triggers that were tested for best
69 : performance. */
70 :
71 : struct fd_net_flusher {
72 :
73 : /* Packets that were enqueued after the last sendto() wakeup are
74 : considered "pending". If there are more than pending_wmark packets
75 : pending, a wakeup is dispatched. Thus, this dispatch trigger is
76 : proportional to packet rate, but does not trigger if I/O is seldom. */
77 : ulong pending_cnt;
78 : ulong pending_wmark;
79 :
80 : /* Sometimes, packets are not flushed out even after a sendto()
81 : wakeup. This can result in the tail of a burst getting delayed or
82 : overrun. If more than tail_flush_backoff ticks pass since the last
83 : sendto() wakeup and there are still unacknowledged packets in the
84 : TX ring, issues another wakeup. */
85 : long next_tail_flush_ticks;
86 : long tail_flush_backoff;
87 :
88 : };
89 :
90 : typedef struct fd_net_flusher fd_net_flusher_t;
91 :
92 : FD_PROTOTYPES_BEGIN
93 :
94 : /* fd_net_flusher_inc marks a new packet as enqueued. */
95 :
96 : static inline void
97 : fd_net_flusher_inc( fd_net_flusher_t * flusher,
98 0 : long now ) {
99 0 : flusher->pending_cnt++;
100 0 : long next_flush = now + flusher->tail_flush_backoff;
101 0 : flusher->next_tail_flush_ticks = fd_long_min( flusher->next_tail_flush_ticks, next_flush );
102 0 : }
103 :
104 : /* fd_net_flusher_check returns 1 if a sendto() wakeup should be issued
105 : immediately. now is a recent fd_tickcount() value.
106 : If tx_ring_empty==0 then the kernel is caught up with the net tile
107 : on the XDP TX ring. (Otherwise, the kernel is behind the net tile) */
108 :
109 : static inline int
110 : fd_net_flusher_check( fd_net_flusher_t * flusher,
111 : long now,
112 0 : int tx_ring_empty ) {
113 0 : int flush_level = flusher->pending_cnt >= flusher->pending_wmark;
114 0 : int flush_timeout = now >= flusher->next_tail_flush_ticks;
115 0 : int flush = flush_level || flush_timeout;
116 0 : if( !flush ) return 0;
117 0 : if( FD_UNLIKELY( tx_ring_empty ) ) {
118 : /* Flush requested but caught up */
119 0 : flusher->pending_cnt = 0UL;
120 0 : flusher->next_tail_flush_ticks = LONG_MAX;
121 0 : return 0;
122 0 : }
123 0 : return 1;
124 0 : }
125 :
126 : /* fd_net_flusher_wakeup signals a sendto() wakeup was done. now is a
127 : recent fd_tickcount() value. */
128 :
129 : static inline void
130 : fd_net_flusher_wakeup( fd_net_flusher_t * flusher,
131 0 : long now ) {
132 0 : flusher->pending_cnt = 0UL;
133 0 : flusher->next_tail_flush_ticks = now + flusher->tail_flush_backoff;
134 0 : }
135 :
136 : FD_PROTOTYPES_END
137 :
138 : /* fd_net_free_ring is a FIFO queue that stores pointers to free XDP TX
139 : frames. */
140 :
141 : struct fd_net_free_ring {
142 : ulong prod;
143 : ulong cons;
144 : ulong depth;
145 : ulong * queue;
146 : };
147 : typedef struct fd_net_free_ring fd_net_free_ring_t;
148 :
149 : typedef struct {
150 : /* An "XSK" is an AF_XDP socket */
151 : uint xsk_cnt;
152 : fd_xsk_t xsk[ 2 ];
153 : int prog_link_fds[ 2 ];
154 :
155 : /* UMEM frame region within dcache */
156 : void * umem_frame0; /* First UMEM frame */
157 : ulong umem_sz; /* Usable UMEM size starting at frame0 */
158 :
159 : /* UMEM chunk region within workspace */
160 : uint umem_chunk0; /* Lowest allowed chunk number */
161 : uint umem_wmark; /* Highest allowed chunk number */
162 :
163 : /* All net tiles are subscribed to the same TX links. (These are
164 : incoming links from app tiles asking the net tile to send out packets)
165 : The net tiles "take turns" doing TX jobs based on the L3+L4 dst hash.
166 : net_tile_id is the index of the current interface, net_tile_cnt is the
167 : total amount of interfaces. */
168 : uint net_tile_id;
169 : uint net_tile_cnt;
170 :
171 : /* Details pertaining to an inflight send op */
172 : struct {
173 : uint if_idx; /* 0: main interface, 1: loopback */
174 : void * frame;
175 : uchar mac_addrs[12]; /* First 12 bytes of Ethernet header */
176 : uint src_ip;
177 : } tx_op;
178 :
179 : /* Round-robin cycle serivce operations */
180 : uint rr_idx;
181 :
182 : /* Ring tracking free packet buffers */
183 : fd_net_free_ring_t free_tx;
184 :
185 : uchar src_mac_addr[6];
186 :
187 : ushort shred_listen_port;
188 : ushort quic_transaction_listen_port;
189 : ushort legacy_transaction_listen_port;
190 : ushort gossip_listen_port;
191 : ushort repair_intake_listen_port;
192 : ushort repair_serve_listen_port;
193 :
194 : ulong in_cnt;
195 : fd_net_in_ctx_t in[ MAX_NET_INS ];
196 :
197 : fd_net_out_ctx_t quic_out[1];
198 : fd_net_out_ctx_t shred_out[1];
199 : fd_net_out_ctx_t gossip_out[1];
200 : fd_net_out_ctx_t repair_out[1];
201 :
202 : /* XDP stats refresh timer */
203 : long xdp_stats_interval_ticks;
204 : long next_xdp_stats_refresh;
205 :
206 : /* TX flush timers */
207 : fd_net_flusher_t tx_flusher[2]; /* one per XSK */
208 :
209 : /* Route and neighbor tables */
210 : fd_fib4_t const * fib_local;
211 : fd_fib4_t const * fib_main;
212 : fd_neigh4_hmap_t neigh4[1];
213 : fd_netlink_neigh4_solicit_link_t neigh4_solicit[1];
214 :
215 : struct {
216 : ulong rx_pkt_cnt;
217 : ulong rx_bytes_total;
218 : ulong rx_undersz_cnt;
219 : ulong rx_fill_blocked_cnt;
220 : ulong rx_backp_cnt;
221 : long rx_busy_cnt;
222 : long rx_idle_cnt;
223 :
224 : ulong tx_submit_cnt;
225 : ulong tx_complete_cnt;
226 : ulong tx_bytes_total;
227 : ulong tx_route_fail_cnt;
228 : ulong tx_no_xdp_cnt;
229 : ulong tx_neigh_fail_cnt;
230 : ulong tx_full_fail_cnt;
231 : long tx_busy_cnt;
232 : long tx_idle_cnt;
233 :
234 : ulong xsk_tx_wakeup_cnt;
235 : ulong xsk_rx_wakeup_cnt;
236 : } metrics;
237 : } fd_net_ctx_t;
238 :
239 : FD_FN_CONST static inline ulong
240 3 : scratch_align( void ) {
241 3 : return 4096UL;
242 3 : }
243 :
244 : FD_FN_PURE static inline ulong
245 3 : scratch_footprint( fd_topo_tile_t const * tile ) {
246 3 : ulong l = FD_LAYOUT_INIT;
247 3 : l = FD_LAYOUT_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
248 3 : l = FD_LAYOUT_APPEND( l, alignof(ulong), tile->net.free_ring_depth * sizeof(ulong) );
249 3 : return FD_LAYOUT_FINI( l, scratch_align() );
250 3 : }
251 :
252 : static void
253 0 : metrics_write( fd_net_ctx_t * ctx ) {
254 0 : FD_MCNT_SET( NET, RX_PKT_CNT, ctx->metrics.rx_pkt_cnt );
255 0 : FD_MCNT_SET( NET, RX_BYTES_TOTAL, ctx->metrics.rx_bytes_total );
256 0 : FD_MCNT_SET( NET, RX_UNDERSZ_CNT, ctx->metrics.rx_undersz_cnt );
257 0 : FD_MCNT_SET( NET, RX_FILL_BLOCKED_CNT, ctx->metrics.rx_fill_blocked_cnt );
258 0 : FD_MCNT_SET( NET, RX_BACKPRESSURE_CNT, ctx->metrics.rx_backp_cnt );
259 0 : FD_MGAUGE_SET( NET, RX_BUSY_CNT, (ulong)fd_long_max( ctx->metrics.rx_busy_cnt, 0L ) );
260 0 : FD_MGAUGE_SET( NET, RX_IDLE_CNT, (ulong)fd_long_max( ctx->metrics.rx_idle_cnt, 0L ) );
261 0 : FD_MGAUGE_SET( NET, TX_BUSY_CNT, (ulong)fd_long_max( ctx->metrics.tx_busy_cnt, 0L ) );
262 0 : FD_MGAUGE_SET( NET, TX_IDLE_CNT, (ulong)fd_long_max( ctx->metrics.tx_idle_cnt, 0L ) );
263 :
264 0 : FD_MCNT_SET( NET, TX_SUBMIT_CNT, ctx->metrics.tx_submit_cnt );
265 0 : FD_MCNT_SET( NET, TX_COMPLETE_CNT, ctx->metrics.tx_complete_cnt );
266 0 : FD_MCNT_SET( NET, TX_BYTES_TOTAL, ctx->metrics.tx_bytes_total );
267 0 : FD_MCNT_SET( NET, TX_ROUTE_FAIL_CNT, ctx->metrics.tx_route_fail_cnt );
268 0 : FD_MCNT_SET( NET, TX_NEIGHBOR_FAIL_CNT, ctx->metrics.tx_neigh_fail_cnt );
269 0 : FD_MCNT_SET( NET, TX_FULL_FAIL_CNT, ctx->metrics.tx_full_fail_cnt );
270 :
271 0 : FD_MCNT_SET( NET, XSK_TX_WAKEUP_CNT, ctx->metrics.xsk_tx_wakeup_cnt );
272 0 : FD_MCNT_SET( NET, XSK_RX_WAKEUP_CNT, ctx->metrics.xsk_rx_wakeup_cnt );
273 0 : }
274 :
275 : struct xdp_statistics_v0 {
276 : __u64 rx_dropped; /* Dropped for other reasons */
277 : __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
278 : __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
279 : };
280 :
281 : struct xdp_statistics_v1 {
282 : __u64 rx_dropped; /* Dropped for other reasons */
283 : __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
284 : __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
285 : __u64 rx_ring_full; /* Dropped due to rx ring being full */
286 : __u64 rx_fill_ring_empty_descs; /* Failed to retrieve item from fill ring */
287 : __u64 tx_ring_empty_descs; /* Failed to retrieve item from tx ring */
288 : };
289 :
290 : static void
291 0 : poll_xdp_statistics( fd_net_ctx_t * ctx ) {
292 0 : struct xdp_statistics_v1 stats = {0};
293 0 : ulong xsk_cnt = ctx->xsk_cnt;
294 0 : for( ulong j=0UL; j<xsk_cnt; j++ ) {
295 0 : struct xdp_statistics_v1 sub_stats;
296 0 : uint optlen = (uint)sizeof(struct xdp_statistics_v1);
297 0 : if( FD_UNLIKELY( -1==getsockopt( ctx->xsk[ j ].xsk_fd, SOL_XDP, XDP_STATISTICS, &sub_stats, &optlen ) ) )
298 0 : FD_LOG_ERR(( "getsockopt(SOL_XDP, XDP_STATISTICS) failed: %s", strerror( errno ) ));
299 0 : if( FD_UNLIKELY( optlen!=sizeof(struct xdp_statistics_v0) &&
300 0 : optlen!=sizeof(struct xdp_statistics_v1) ) ) {
301 0 : FD_LOG_ERR(( "getsockopt(SOL_XDP, XDP_STATISTICS) returned unexpected size %u", optlen ));
302 0 : }
303 0 : stats.rx_dropped += sub_stats.rx_dropped;
304 0 : stats.rx_invalid_descs += sub_stats.rx_invalid_descs;
305 0 : stats.tx_invalid_descs += sub_stats.tx_invalid_descs;
306 0 : stats.rx_ring_full += sub_stats.rx_ring_full;
307 0 : stats.rx_fill_ring_empty_descs += sub_stats.rx_fill_ring_empty_descs;
308 0 : stats.tx_ring_empty_descs += sub_stats.tx_ring_empty_descs;
309 0 : }
310 :
311 0 : FD_MCNT_SET( NET, XDP_RX_DROPPED_OTHER, stats.rx_dropped );
312 0 : FD_MCNT_SET( NET, XDP_RX_INVALID_DESCS, stats.rx_invalid_descs );
313 0 : FD_MCNT_SET( NET, XDP_TX_INVALID_DESCS, stats.tx_invalid_descs );
314 0 : FD_MCNT_SET( NET, XDP_RX_RING_FULL, stats.rx_ring_full );
315 0 : FD_MCNT_SET( NET, XDP_RX_FILL_RING_EMPTY_DESCS, stats.rx_fill_ring_empty_descs );
316 0 : FD_MCNT_SET( NET, XDP_TX_RING_EMPTY_DESCS, stats.tx_ring_empty_descs );
317 0 : }
318 :
319 : /* net_is_fatal_xdp_error returns 1 if the given errno returned by an
320 : XDP API indicates a non-recoverable error code. The net tile should
321 : crash if it sees such an error so the problem does not go undetected.
322 : Otherwise, returns 0. */
323 :
324 : static int
325 0 : net_is_fatal_xdp_error( int err ) {
326 0 : return err==ESOCKTNOSUPPORT || err==EOPNOTSUPP || err==EINVAL ||
327 0 : err==EPERM;
328 0 : }
329 :
330 : /* net_tx_ready returns 1 if the current XSK is ready to submit a TX send
331 : job. If the XSK is blocked for sends, returns 0. Reasons for block
332 : include:
333 : - No XSK TX buffer is available
334 : - XSK TX ring is full */
335 :
336 : static int
337 : net_tx_ready( fd_net_ctx_t * ctx,
338 0 : uint if_idx ) {
339 0 : fd_xsk_t * xsk = &ctx->xsk[ if_idx ];
340 0 : fd_xdp_ring_t * tx_ring = &xsk->ring_tx;
341 0 : fd_net_free_ring_t * free = &ctx->free_tx;
342 0 : if( free->prod == free->cons ) return 0; /* drop */
343 0 : if( tx_ring->prod - tx_ring->cons >= tx_ring->depth ) return 0; /* drop */
344 0 : return 1;
345 0 : }
346 :
347 : /* net_rx_wakeup triggers xsk_recvmsg to run in the kernel. Needs to be
348 : called periodically in order to receive packets. */
349 :
350 : static void
351 : net_rx_wakeup( fd_net_ctx_t * ctx,
352 : fd_xsk_t * xsk,
353 0 : int * charge_busy ) {
354 0 : if( !fd_xsk_rx_need_wakeup( xsk ) ) return;
355 0 : *charge_busy = 1;
356 0 : struct msghdr _ignored[ 1 ] = { 0 };
357 0 : if( FD_UNLIKELY( -1==recvmsg( xsk->xsk_fd, _ignored, MSG_DONTWAIT ) ) ) {
358 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
359 0 : FD_LOG_ERR(( "xsk recvmsg failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
360 0 : }
361 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
362 0 : long ts = fd_log_wallclock();
363 0 : if( ts > xsk->log_suppress_until_ns ) {
364 0 : FD_LOG_WARNING(( "xsk recvmsg failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
365 0 : xsk->log_suppress_until_ns = ts + (long)1e9;
366 0 : }
367 0 : }
368 0 : }
369 0 : ctx->metrics.xsk_rx_wakeup_cnt++;
370 0 : }
371 :
372 : /* net_tx_wakeup triggers xsk_sendmsg to run in the kernel. Needs to be
373 : called periodically in order to transmit packets. */
374 :
375 : static void
376 : net_tx_wakeup( fd_net_ctx_t * ctx,
377 : fd_xsk_t * xsk,
378 0 : int * charge_busy ) {
379 0 : if( !fd_xsk_tx_need_wakeup( xsk ) ) return;
380 0 : if( FD_VOLATILE_CONST( *xsk->ring_tx.prod )==FD_VOLATILE_CONST( *xsk->ring_tx.cons ) ) return;
381 0 : *charge_busy = 1;
382 0 : if( FD_UNLIKELY( -1==sendto( xsk->xsk_fd, NULL, 0, MSG_DONTWAIT, NULL, 0 ) ) ) {
383 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
384 0 : FD_LOG_ERR(( "xsk sendto failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
385 0 : }
386 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
387 0 : long ts = fd_log_wallclock();
388 0 : if( ts > xsk->log_suppress_until_ns ) {
389 0 : FD_LOG_WARNING(( "xsk sendto failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
390 0 : xsk->log_suppress_until_ns = ts + (long)1e9;
391 0 : }
392 0 : }
393 0 : }
394 0 : ctx->metrics.xsk_tx_wakeup_cnt++;
395 0 : }
396 :
397 : /* net_tx_periodic_wakeup does a timer based xsk_sendmsg wakeup. */
398 :
399 : static inline int
400 : net_tx_periodic_wakeup( fd_net_ctx_t * ctx,
401 : uint if_idx,
402 : long now,
403 0 : int * charge_busy ) {
404 0 : uint tx_prod = FD_VOLATILE_CONST( *ctx->xsk[ if_idx ].ring_tx.prod );
405 0 : uint tx_cons = FD_VOLATILE_CONST( *ctx->xsk[ if_idx ].ring_tx.cons );
406 0 : int tx_ring_empty = tx_prod==tx_cons;
407 0 : if( fd_net_flusher_check( ctx->tx_flusher+if_idx, now, tx_ring_empty ) ) {
408 0 : net_tx_wakeup( ctx, &ctx->xsk[ if_idx ], charge_busy );
409 0 : fd_net_flusher_wakeup( ctx->tx_flusher+if_idx, now );
410 0 : }
411 0 : return 0;
412 0 : }
413 :
414 : static void
415 0 : during_housekeeping( fd_net_ctx_t * ctx ) {
416 0 : long now = fd_tickcount();
417 :
418 0 : ctx->metrics.rx_busy_cnt = 0UL;
419 0 : ctx->metrics.rx_idle_cnt = 0UL;
420 0 : ctx->metrics.tx_busy_cnt = 0UL;
421 0 : ctx->metrics.tx_idle_cnt = fd_seq_diff( ctx->free_tx.prod, ctx->free_tx.cons );
422 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
423 0 : fd_xsk_t * xsk = &ctx->xsk[ j ];
424 : /* Refresh all sequence numbers (consumer first, then producer) */
425 0 : FD_COMPILER_MFENCE();
426 0 : xsk->ring_fr.cached_cons = FD_VOLATILE_CONST( *xsk->ring_fr.cons );
427 0 : xsk->ring_fr.cached_prod = FD_VOLATILE_CONST( *xsk->ring_fr.prod );
428 0 : xsk->ring_rx.cached_cons = FD_VOLATILE_CONST( *xsk->ring_rx.cons );
429 0 : xsk->ring_rx.cached_prod = FD_VOLATILE_CONST( *xsk->ring_rx.prod );
430 0 : xsk->ring_tx.cached_cons = FD_VOLATILE_CONST( *xsk->ring_tx.cons );
431 0 : xsk->ring_tx.cached_prod = FD_VOLATILE_CONST( *xsk->ring_tx.prod );
432 0 : xsk->ring_cr.cached_cons = FD_VOLATILE_CONST( *xsk->ring_cr.cons );
433 0 : xsk->ring_cr.cached_prod = FD_VOLATILE_CONST( *xsk->ring_cr.prod );
434 0 : FD_COMPILER_MFENCE();
435 0 : ctx->metrics.rx_busy_cnt += (long)(int)( xsk->ring_rx.cached_prod - xsk->ring_rx.cached_cons );
436 0 : ctx->metrics.rx_idle_cnt += (long)(int)( xsk->ring_fr.cached_prod - xsk->ring_fr.cached_cons );
437 0 : ctx->metrics.tx_busy_cnt += (long)(int)( xsk->ring_tx.cached_prod - xsk->ring_tx.cached_cons );
438 0 : ctx->metrics.tx_busy_cnt += (long)(int)( xsk->ring_cr.cached_prod - xsk->ring_cr.cached_cons );
439 0 : }
440 :
441 0 : if( now > ctx->next_xdp_stats_refresh ) {
442 0 : ctx->next_xdp_stats_refresh = now + ctx->xdp_stats_interval_ticks;
443 0 : poll_xdp_statistics( ctx );
444 0 : }
445 :
446 0 : int _charge_busy = 0;
447 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
448 0 : net_rx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
449 0 : }
450 0 : }
451 :
452 : /* net_tx_route resolves the destination interface index, src MAC address,
453 : and dst MAC address. Returns 1 on success, 0 on failure. On success,
454 : tx_op->{if_idx,mac_addrs} is set. */
455 :
456 : static int
457 : net_tx_route( fd_net_ctx_t * ctx,
458 0 : uint dst_ip ) {
459 :
460 : /* Route lookup */
461 :
462 0 : fd_fib4_hop_t hop[2] = {0};
463 0 : fd_fib4_lookup( ctx->fib_local, hop+0, dst_ip, 0UL );
464 0 : fd_fib4_lookup( ctx->fib_main, hop+1, dst_ip, 0UL );
465 0 : fd_fib4_hop_t const * next_hop = fd_fib4_hop_or( hop+0, hop+1 );
466 :
467 0 : uint rtype = next_hop->rtype;
468 0 : uint if_idx = next_hop->if_idx;
469 0 : uint ip4_src = next_hop->ip4_src;
470 :
471 0 : if( FD_UNLIKELY( rtype==FD_FIB4_RTYPE_LOCAL ) ) {
472 0 : rtype = FD_FIB4_RTYPE_UNICAST;
473 0 : if_idx = 1;
474 0 : }
475 :
476 0 : if( FD_UNLIKELY( rtype!=FD_FIB4_RTYPE_UNICAST ) ) {
477 0 : ctx->metrics.tx_route_fail_cnt++;
478 0 : return 0;
479 0 : }
480 :
481 0 : if( if_idx==1 ) {
482 : /* Set Ethernet src and dst address to 00:00:00:00:00:00 */
483 0 : memset( ctx->tx_op.mac_addrs, 0, 12UL );
484 0 : ctx->tx_op.if_idx = 1;
485 0 : return 1;
486 0 : }
487 :
488 0 : if( FD_UNLIKELY( if_idx!=ctx->xsk[ 0 ].if_idx ) ) {
489 0 : ctx->metrics.tx_no_xdp_cnt++;
490 0 : return 0;
491 0 : }
492 0 : ctx->tx_op.if_idx = 0;
493 :
494 : /* Neighbor resolve */
495 :
496 0 : uint neigh_ip = next_hop->ip4_gw;
497 0 : if( !neigh_ip ) neigh_ip = dst_ip;
498 :
499 0 : fd_neigh4_hmap_query_t neigh_query[1];
500 0 : int neigh_res = fd_neigh4_hmap_query_try( ctx->neigh4, &neigh_ip, NULL, neigh_query, 0 );
501 0 : if( FD_UNLIKELY( neigh_res!=FD_MAP_SUCCESS ) ) {
502 : /* Neighbor not found */
503 0 : fd_netlink_neigh4_solicit( ctx->neigh4_solicit, neigh_ip, if_idx, fd_frag_meta_ts_comp( fd_tickcount() ) );
504 0 : ctx->metrics.tx_neigh_fail_cnt++;
505 0 : return 0;
506 0 : }
507 0 : fd_neigh4_entry_t const * neigh = fd_neigh4_hmap_query_ele_const( neigh_query );
508 0 : if( FD_UNLIKELY( neigh->state != FD_NEIGH4_STATE_ACTIVE ) ) {
509 0 : ctx->metrics.tx_neigh_fail_cnt++;
510 0 : return 0;
511 0 : }
512 :
513 0 : ctx->tx_op.src_ip = ip4_src;
514 0 : memcpy( ctx->tx_op.mac_addrs+0, neigh->mac_addr, 6 );
515 0 : memcpy( ctx->tx_op.mac_addrs+6, ctx->src_mac_addr, 6 );
516 :
517 0 : if( FD_UNLIKELY( fd_neigh4_hmap_query_test( neigh_query ) ) ) {
518 0 : ctx->metrics.tx_neigh_fail_cnt++;
519 0 : return 0;
520 0 : }
521 :
522 0 : return 1;
523 0 : }
524 :
525 : /* before_frag is called when a new metadata descriptor for a TX job is
526 : found. This callback determines whether this net tile is responsible
527 : for the TX job. If so, it prepares the TX op for the during_frag and
528 : after_frag callbacks. */
529 :
530 : static inline int
531 : before_frag( fd_net_ctx_t * ctx,
532 : ulong in_idx,
533 : ulong seq,
534 0 : ulong sig ) {
535 0 : (void)in_idx; (void)seq;
536 :
537 : /* Find interface index of next packet */
538 :
539 0 : ulong proto = fd_disco_netmux_sig_proto( sig );
540 0 : if( FD_UNLIKELY( proto!=DST_PROTO_OUTGOING ) ) return 1;
541 :
542 0 : uint dst_ip = fd_disco_netmux_sig_dst_ip( sig );
543 0 : if( FD_UNLIKELY( !net_tx_route( ctx, dst_ip ) ) ) return 1;
544 :
545 0 : uint net_tile_id = ctx->net_tile_id;
546 0 : uint net_tile_cnt = ctx->net_tile_cnt;
547 0 : uint if_idx = ctx->tx_op.if_idx;
548 0 : if( FD_UNLIKELY( if_idx>=ctx->xsk_cnt ) ) return 1; /* ignore */
549 :
550 : /* Load balance TX */
551 :
552 0 : uint hash = (uint)fd_disco_netmux_sig_hash( sig );
553 0 : uint target_idx = hash % net_tile_cnt;
554 0 : if( if_idx==1 ) target_idx = 0; /* loopback always targets tile 0 */
555 :
556 : /* Skip if another net tile is responsible for this packet */
557 :
558 0 : if( net_tile_id!=target_idx ) return 1; /* ignore */
559 :
560 : /* Skip if TX is blocked */
561 :
562 0 : if( FD_UNLIKELY( !net_tx_ready( ctx, if_idx ) ) ) {
563 0 : ctx->metrics.tx_full_fail_cnt++;
564 0 : return 1;
565 0 : }
566 :
567 : /* Allocate buffer for receive */
568 :
569 0 : fd_net_free_ring_t * free = &ctx->free_tx;
570 0 : ulong alloc_seq = free->cons;
571 0 : void * frame = (void *)free->queue[ alloc_seq % free->depth ];
572 0 : free->cons = fd_seq_inc( alloc_seq, 1UL );
573 :
574 0 : ctx->tx_op.if_idx = if_idx;
575 0 : ctx->tx_op.frame = frame;
576 :
577 0 : return 0; /* continue */
578 0 : }
579 :
580 : /* during_frag is called when before_frag has committed to transmit an
581 : outgoing packet. */
582 :
583 : static inline void
584 : during_frag( fd_net_ctx_t * ctx,
585 : ulong in_idx,
586 : ulong seq FD_PARAM_UNUSED,
587 : ulong sig FD_PARAM_UNUSED,
588 : ulong chunk,
589 : ulong sz,
590 0 : ulong ctl FD_PARAM_UNUSED ) {
591 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz>FD_NET_MTU ) )
592 0 : FD_LOG_ERR(( "chunk %lu %lu corrupt, not in range [%lu,%lu]", chunk, sz, ctx->in[ in_idx ].chunk0, ctx->in[ in_idx ].wmark ));
593 :
594 0 : if( FD_UNLIKELY( sz<34UL ) )
595 0 : FD_LOG_ERR(( "packet too small %lu (in_idx=%lu)", sz, in_idx ));
596 :
597 0 : void * frame = ctx->tx_op.frame;
598 0 : if( FD_UNLIKELY( (ulong)frame < (ulong)ctx->umem_frame0 ) )
599 0 : FD_LOG_ERR(( "frame %p out of bounds (below %p)", frame, (void *)ctx->umem_frame0 ));
600 0 : ulong umem_off = (ulong)frame - (ulong)ctx->umem_frame0;
601 0 : if( FD_UNLIKELY( (ulong)umem_off > (ulong)ctx->umem_sz ) )
602 0 : FD_LOG_ERR(( "frame %p out of bounds (beyond %p)", frame, (void *)ctx->umem_sz ));
603 :
604 : /* Speculatively copy frame into XDP buffer */
605 0 : uchar const * src = fd_chunk_to_laddr_const( ctx->in[ in_idx ].mem, chunk );
606 0 : fd_memcpy( ctx->tx_op.frame, src, sz );
607 0 : }
608 :
609 : /* after_frag is called when the during_frag memcpy was _not_ overrun. */
610 :
611 : static void
612 : after_frag( fd_net_ctx_t * ctx,
613 : ulong in_idx,
614 : ulong seq,
615 : ulong sig,
616 : ulong sz,
617 : ulong tsorig,
618 : ulong tspub,
619 0 : fd_stem_context_t * stem ) {
620 0 : (void)in_idx; (void)seq; (void)sig; (void)tsorig; (void)tspub; (void)stem;
621 :
622 : /* Current send operation */
623 :
624 0 : uint if_idx = ctx->tx_op.if_idx;
625 0 : uchar * frame = ctx->tx_op.frame;
626 0 : fd_xsk_t * xsk = &ctx->xsk[ if_idx ];
627 :
628 : /* Select Ethernet addresses */
629 0 : memcpy( frame, ctx->tx_op.mac_addrs, 12 );
630 :
631 : /* Select IPv4 source address */
632 0 : uint ihl = frame[ 14 ] & 0x0f;
633 0 : ushort ethertype = FD_LOAD( ushort, frame+12 );
634 0 : uint ip4_saddr = FD_LOAD( uint, frame+26 );
635 0 : if( ethertype==fd_ushort_bswap( FD_ETH_HDR_TYPE_IP ) && ip4_saddr==0 ) {
636 0 : if( FD_UNLIKELY( ctx->tx_op.src_ip==0 ||
637 0 : ihl<5 || (14+(ihl<<2))>sz ) ) {
638 : /* Outgoing IPv4 packet with unknown src IP or invalid IHL */
639 : /* FIXME should select first IPv4 address of device table here */
640 0 : ctx->metrics.tx_route_fail_cnt++;
641 0 : return;
642 0 : }
643 :
644 : /* Recompute checksum after changing header */
645 0 : FD_STORE( uint, frame+26, ctx->tx_op.src_ip );
646 0 : FD_STORE( ushort, frame+24, 0 );
647 0 : FD_STORE( ushort, frame+24, fd_ip4_hdr_check( frame+14 ) );
648 0 : }
649 :
650 : /* Submit packet TX job
651 :
652 : Invariant for ring_tx: prod-cons<length
653 : (This invariant breaks if any other packet is sent over this ring
654 : between before_frag and this point, e.g. send_arp_probe.) */
655 :
656 0 : fd_xdp_ring_t * tx_ring = &xsk->ring_tx;
657 0 : uint tx_seq = FD_VOLATILE_CONST( *tx_ring->prod );
658 0 : uint tx_mask = tx_ring->depth - 1U;
659 0 : xsk->ring_tx.packet_ring[ tx_seq&tx_mask ] = (struct xdp_desc) {
660 0 : .addr = (ulong)frame - (ulong)ctx->umem_frame0,
661 0 : .len = (uint)sz,
662 0 : .options = 0
663 0 : };
664 :
665 : /* Frame is now owned by kernel. Clear tx_op. */
666 0 : ctx->tx_op.frame = NULL;
667 :
668 : /* Register newly enqueued packet */
669 0 : FD_VOLATILE( *xsk->ring_tx.prod ) = tx_ring->cached_prod = tx_seq+1U;
670 0 : ctx->metrics.tx_submit_cnt++;
671 0 : ctx->metrics.tx_bytes_total += sz;
672 0 : fd_net_flusher_inc( ctx->tx_flusher+if_idx, fd_tickcount() );
673 :
674 0 : }
675 :
676 : /* net_rx_packet is called when a new Ethernet frame is available.
677 : Attempts to copy out the frame to a downstream tile. */
678 :
679 : static void
680 : net_rx_packet( fd_net_ctx_t * ctx,
681 : fd_stem_context_t * stem,
682 : ulong umem_off,
683 : ulong sz,
684 0 : uint * freed_chunk ) {
685 :
686 0 : ulong umem_lowbits = umem_off & 0x3fUL;
687 :
688 0 : uchar const * packet = (uchar const *)ctx->umem_frame0 + umem_off;
689 0 : uchar const * packet_end = packet + sz;
690 0 : uchar const * iphdr = packet + 14U;
691 :
692 : /* Translate packet to UMEM frame index */
693 0 : ulong chunk = ctx->umem_chunk0 + (umem_off>>FD_CHUNK_LG_SZ);
694 :
695 : /* Filter for UDP/IPv4 packets. Test for ethtype and ipproto in 1
696 : branch */
697 0 : uint test_ethip = ( (uint)packet[12] << 16u ) | ( (uint)packet[13] << 8u ) | (uint)packet[23];
698 0 : if( FD_UNLIKELY( test_ethip!=0x080011 ) ) {
699 0 : FD_LOG_ERR(( "Firedancer received a packet from the XDP program that was either "
700 0 : "not an IPv4 packet, or not a UDP packet. It is likely your XDP program "
701 0 : "is not configured correctly." ));
702 0 : }
703 :
704 : /* IPv4 is variable-length, so lookup IHL to find start of UDP */
705 0 : uint iplen = ( ( (uint)iphdr[0] ) & 0x0FU ) * 4U;
706 0 : uchar const * udp = iphdr + iplen;
707 :
708 : /* Ignore if UDP header is too short */
709 0 : if( FD_UNLIKELY( udp+8U > packet_end ) ) {
710 0 : FD_DTRACE_PROBE( net_tile_err_rx_undersz );
711 0 : ctx->metrics.rx_undersz_cnt++;
712 0 : return;
713 0 : }
714 :
715 : /* Extract IP dest addr and UDP src/dest port */
716 0 : uint ip_srcaddr = *(uint *)( iphdr+12UL );
717 0 : ushort udp_srcport = fd_ushort_bswap( *(ushort *)( udp+0UL ) );
718 0 : ushort udp_dstport = fd_ushort_bswap( *(ushort *)( udp+2UL ) );
719 :
720 0 : FD_DTRACE_PROBE_4( net_tile_pkt_rx, ip_srcaddr, udp_srcport, udp_dstport, sz );
721 :
722 : /* Route packet to downstream tile */
723 0 : ushort proto;
724 0 : fd_net_out_ctx_t * out;
725 0 : if( FD_UNLIKELY( udp_dstport==ctx->shred_listen_port ) ) {
726 0 : proto = DST_PROTO_SHRED;
727 0 : out = ctx->shred_out;
728 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->quic_transaction_listen_port ) ) {
729 0 : proto = DST_PROTO_TPU_QUIC;
730 0 : out = ctx->quic_out;
731 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->legacy_transaction_listen_port ) ) {
732 0 : proto = DST_PROTO_TPU_UDP;
733 0 : out = ctx->quic_out;
734 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->gossip_listen_port ) ) {
735 0 : proto = DST_PROTO_GOSSIP;
736 0 : out = ctx->gossip_out;
737 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->repair_intake_listen_port ) ) {
738 0 : proto = DST_PROTO_REPAIR;
739 0 : out = ctx->repair_out;
740 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->repair_serve_listen_port ) ) {
741 0 : proto = DST_PROTO_REPAIR;
742 0 : out = ctx->repair_out;
743 0 : } else {
744 :
745 0 : FD_LOG_ERR(( "Firedancer received a UDP packet on port %hu which was not expected. "
746 0 : "Only the following ports should be configured to forward packets: "
747 0 : "%hu, %hu, %hu, %hu, %hu, %hu (excluding any 0 ports, which can be ignored)."
748 0 : "Please report this error to Firedancer maintainers.",
749 0 : udp_dstport,
750 0 : ctx->shred_listen_port,
751 0 : ctx->quic_transaction_listen_port,
752 0 : ctx->legacy_transaction_listen_port,
753 0 : ctx->gossip_listen_port,
754 0 : ctx->repair_intake_listen_port,
755 0 : ctx->repair_serve_listen_port ));
756 0 : }
757 :
758 : /* tile can decide how to partition based on src ip addr and src port */
759 0 : ulong sig = fd_disco_netmux_sig( ip_srcaddr, udp_srcport, 0U, proto, 14UL+8UL+iplen );
760 :
761 : /* Peek the mline for an old frame */
762 0 : fd_frag_meta_t * mline = out->mcache + fd_mcache_line_idx( out->seq, out->depth );
763 0 : *freed_chunk = mline->chunk;
764 :
765 : /* Overwrite the mline with the new frame */
766 0 : ulong tspub = (ulong)fd_frag_meta_ts_comp( fd_tickcount() );
767 0 : fd_mcache_publish( out->mcache, out->depth, out->seq, sig, chunk, sz, umem_lowbits, 0, tspub );
768 :
769 : /* Wind up for the next iteration */
770 0 : *stem->cr_avail -= stem->cr_decrement_amount;
771 0 : out->seq = fd_seq_inc( out->seq, 1UL );
772 :
773 0 : ctx->metrics.rx_pkt_cnt++;
774 0 : ctx->metrics.rx_bytes_total += sz;
775 :
776 0 : }
777 :
778 : /* net_comp_event is called when an XDP TX frame is free again. */
779 :
780 : static void
781 : net_comp_event( fd_net_ctx_t * ctx,
782 : fd_xsk_t * xsk,
783 0 : uint comp_seq ) {
784 :
785 : /* Locate the incoming frame */
786 :
787 0 : fd_xdp_ring_t * comp_ring = &xsk->ring_cr;
788 0 : uint comp_mask = comp_ring->depth - 1U;
789 0 : ulong frame = FD_VOLATILE_CONST( comp_ring->frame_ring[ comp_seq&comp_mask ] );
790 0 : ulong const frame_mask = FD_NET_MTU - 1UL;
791 0 : if( FD_UNLIKELY( frame+FD_NET_MTU > ctx->umem_sz ) ) {
792 0 : FD_LOG_ERR(( "Bounds check failed: frame=0x%lx umem_sz=0x%lx",
793 0 : frame, (ulong)ctx->umem_sz ));
794 0 : }
795 :
796 : /* Check if we have space to return the freed frame */
797 :
798 0 : fd_net_free_ring_t * free = &ctx->free_tx;
799 0 : ulong free_prod = free->prod;
800 0 : ulong free_mask = free->depth - 1UL;
801 0 : long free_cnt = fd_seq_diff( free_prod, free->cons );
802 0 : if( FD_UNLIKELY( free_cnt>=(long)free->depth ) ) return; /* blocked */
803 :
804 0 : free->queue[ free_prod&free_mask ] = (ulong)ctx->umem_frame0 + (frame & (~frame_mask));
805 0 : free->prod = fd_seq_inc( free_prod, 1UL );
806 :
807 : /* Wind up for next iteration */
808 :
809 0 : FD_VOLATILE( *comp_ring->cons ) = comp_ring->cached_cons = comp_seq+1U;
810 :
811 0 : ctx->metrics.tx_complete_cnt++;
812 :
813 0 : }
814 :
815 : /* net_rx_event is called when a new XDP RX frame is available. Calls
816 : net_rx_packet, then returns the packet back to the kernel via the fill
817 : ring. */
818 :
819 : static void
820 : net_rx_event( fd_net_ctx_t * ctx,
821 : fd_stem_context_t * stem,
822 : fd_xsk_t * xsk,
823 0 : uint rx_seq ) {
824 :
825 : // FIXME(topointon): Temporarily disabling backpressure feature because it triggers even with FD_TOPOB_UNRELIABLE
826 : //if( FD_UNLIKELY( *stem->cr_avail < stem->cr_decrement_amount ) ) {
827 : // ctx->metrics.rx_backp_cnt++;
828 : // return;
829 : //}
830 :
831 : /* Locate the incoming frame */
832 :
833 0 : fd_xdp_ring_t * rx_ring = &xsk->ring_rx;
834 0 : uint rx_mask = rx_ring->depth - 1U;
835 0 : struct xdp_desc frame = FD_VOLATILE_CONST( rx_ring->packet_ring[ rx_seq&rx_mask ] );
836 :
837 0 : if( FD_UNLIKELY( frame.len>FD_NET_MTU ) )
838 0 : FD_LOG_ERR(( "received a UDP packet with a too large payload (%u)", frame.len ));
839 :
840 : /* Check if we have space in the fill ring to free the frame */
841 :
842 0 : fd_xdp_ring_t * fill_ring = &xsk->ring_fr;
843 0 : uint fill_depth = fill_ring->depth;
844 0 : uint fill_mask = fill_depth-1U;
845 0 : ulong frame_mask = FD_NET_MTU - 1UL;
846 0 : uint fill_prod = FD_VOLATILE_CONST( *fill_ring->prod );
847 0 : uint fill_cons = FD_VOLATILE_CONST( *fill_ring->cons );
848 :
849 0 : if( FD_UNLIKELY( (int)(fill_prod-fill_cons) >= (int)fill_depth ) ) {
850 0 : ctx->metrics.rx_fill_blocked_cnt++;
851 0 : return; /* blocked */
852 0 : }
853 :
854 : /* Pass it to the receive handler */
855 :
856 0 : uint freed_chunk = UINT_MAX;
857 0 : net_rx_packet( ctx, stem, frame.addr, frame.len, &freed_chunk );
858 :
859 0 : FD_COMPILER_MFENCE();
860 0 : FD_VOLATILE( *rx_ring->cons ) = rx_ring->cached_cons = rx_seq+1U;
861 :
862 : /* If this mcache publish shadowed a previous publish, mark the old
863 : frame as free. */
864 :
865 0 : if( FD_LIKELY( freed_chunk!=UINT_MAX ) ) {
866 0 : if( FD_UNLIKELY( ( freed_chunk < ctx->umem_chunk0 ) |
867 0 : ( freed_chunk > ctx->umem_wmark ) ) ) {
868 0 : FD_LOG_ERR(( "mcache corruption detected: chunk=%u chunk0=%u wmark=%u",
869 0 : freed_chunk, ctx->umem_chunk0, ctx->umem_wmark ));
870 0 : }
871 0 : ulong freed_off = (freed_chunk - ctx->umem_chunk0)<<FD_CHUNK_LG_SZ;
872 0 : fill_ring->frame_ring[ fill_prod&fill_mask ] = freed_off & (~frame_mask);
873 0 : FD_VOLATILE( *fill_ring->prod ) = fill_ring->cached_prod = fill_prod+1U;
874 0 : }
875 :
876 0 : }
877 :
878 : /* before_credit is called every loop iteration. */
879 :
880 : static void
881 : before_credit( fd_net_ctx_t * ctx,
882 : fd_stem_context_t * stem,
883 0 : int * charge_busy ) {
884 : /* A previous send attempt was overrun. A corrupt copy of the packet was
885 : placed into an XDP frame, but the frame was not yet submitted to the
886 : TX ring. Return the tx buffer to the free list. */
887 :
888 0 : if( ctx->tx_op.frame ) {
889 0 : *charge_busy = 1;
890 0 : fd_net_free_ring_t * free = &ctx->free_tx;
891 0 : ulong alloc_seq = free->prod;
892 0 : free->queue[ alloc_seq % free->depth ] = (ulong)ctx->tx_op.frame;
893 0 : free->prod = fd_seq_inc( alloc_seq, 1UL );
894 0 : ctx->tx_op.frame = NULL;
895 0 : }
896 :
897 : /* Check if new packets are available or if TX frames are free again
898 : (Round-robin through sockets) */
899 :
900 0 : uint rr_idx = ctx->rr_idx;
901 0 : fd_xsk_t * rr_xsk = &ctx->xsk[ rr_idx ];
902 0 : ctx->rr_idx++;
903 0 : ctx->rr_idx = fd_uint_if( ctx->rr_idx>=ctx->xsk_cnt, 0, ctx->rr_idx );
904 :
905 0 : net_tx_periodic_wakeup( ctx, rr_idx, fd_tickcount(), charge_busy );
906 :
907 0 : uint rx_cons = rr_xsk->ring_rx.cached_cons;
908 0 : uint rx_prod = FD_VOLATILE_CONST( *rr_xsk->ring_rx.prod );
909 0 : if( rx_cons!=rx_prod ) {
910 0 : *charge_busy = 1;
911 0 : rr_xsk->ring_rx.cached_prod = rx_prod;
912 0 : net_rx_event( ctx, stem, rr_xsk, rx_cons );
913 0 : } else {
914 0 : net_rx_wakeup( ctx, rr_xsk, charge_busy );
915 0 : }
916 :
917 0 : uint comp_cons = FD_VOLATILE_CONST( *rr_xsk->ring_cr.cons );
918 0 : uint comp_prod = FD_VOLATILE_CONST( *rr_xsk->ring_cr.prod );
919 0 : if( comp_cons!=comp_prod ) {
920 0 : *charge_busy = 1;
921 0 : rr_xsk->ring_cr.cached_prod = comp_prod;
922 0 : net_comp_event( ctx, rr_xsk, comp_cons );
923 0 : }
924 :
925 0 : }
926 :
927 : /* net_xsk_bootstrap assigns UMEM frames to the FILL ring. */
928 :
929 : static ulong
930 : net_xsk_bootstrap( fd_net_ctx_t * ctx,
931 : uint xsk_idx,
932 0 : ulong frame_off ) {
933 0 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
934 :
935 0 : ulong const frame_sz = FD_NET_MTU;
936 0 : ulong const fr_depth = ctx->xsk[ xsk_idx ].ring_fr.depth/2UL;
937 :
938 0 : fd_xdp_ring_t * fill = &xsk->ring_fr;
939 0 : uint fill_prod = fill->cached_prod;
940 0 : for( ulong j=0UL; j<fr_depth; j++ ) {
941 0 : fill->frame_ring[ j ] = frame_off;
942 0 : frame_off += frame_sz;
943 0 : }
944 0 : FD_VOLATILE( *fill->prod ) = fill->cached_prod = fill_prod + (uint)fr_depth;
945 :
946 0 : return frame_off;
947 0 : }
948 :
949 : /* FIXME source MAC address from netlnk tile instead */
950 :
951 : static void
952 : mac_address( const char * interface,
953 0 : uchar * mac ) {
954 0 : int fd = socket( AF_INET, SOCK_DGRAM, 0 );
955 0 : struct ifreq ifr;
956 0 : ifr.ifr_addr.sa_family = AF_INET;
957 0 : strncpy( ifr.ifr_name, interface, IFNAMSIZ );
958 0 : if( FD_UNLIKELY( ioctl( fd, SIOCGIFHWADDR, &ifr ) ) )
959 0 : FD_LOG_ERR(( "could not get MAC address of interface `%s`: (%i-%s)", interface, errno, fd_io_strerror( errno ) ));
960 0 : if( FD_UNLIKELY( close(fd) ) )
961 0 : FD_LOG_ERR(( "could not close socket (%i-%s)", errno, fd_io_strerror( errno ) ));
962 0 : fd_memcpy( mac, ifr.ifr_hwaddr.sa_data, 6 );
963 0 : }
964 :
965 : /* privileged_init does the following initialization steps:
966 :
967 : - Create an AF_XDP socket
968 : - Map XDP metadata rings
969 : - Register UMEM data region with socket
970 : - Insert AF_XDP socket into xsk_map
971 :
972 : Net tile 0 also runs fd_xdp_install and repeats the above step for
973 : the loopback device. (Unless the main interface is already loopback)
974 :
975 : Kernel object references:
976 :
977 : BPF_LINK file descriptor
978 : |
979 : +-> XDP program installation on NIC
980 : | |
981 : | +-> XDP program <-- BPF_PROG file descriptor (prog_fd)
982 : |
983 : +-> XSKMAP object <-- BPF_MAP file descriptor (xsk_map) */
984 :
985 : static void
986 : privileged_init( fd_topo_t * topo,
987 0 : fd_topo_tile_t * tile ) {
988 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
989 :
990 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
991 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
992 0 : ulong * free_tx = FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), tile->net.free_ring_depth * sizeof(ulong) );;
993 :
994 0 : fd_memset( ctx, 0, sizeof(fd_net_ctx_t) );
995 :
996 0 : uint if_idx = if_nametoindex( tile->net.interface );
997 0 : if( FD_UNLIKELY( !if_idx ) ) FD_LOG_ERR(( "if_nametoindex(%s) failed", tile->net.interface ));
998 :
999 0 : mac_address( tile->net.interface, ctx->src_mac_addr );
1000 :
1001 : /* Load up dcache containing UMEM */
1002 :
1003 0 : void * const dcache_mem = fd_topo_obj_laddr( topo, tile->net.umem_dcache_obj_id );
1004 0 : void * const umem_dcache = fd_dcache_join( dcache_mem );
1005 0 : ulong const umem_dcache_data_sz = fd_dcache_data_sz( umem_dcache );
1006 0 : ulong const umem_frame_sz = 2048UL;
1007 :
1008 : /* Left shrink UMEM region to be 4096 byte aligned */
1009 :
1010 0 : void * const umem_frame0 = (void *)fd_ulong_align_up( (ulong)umem_dcache, 4096UL );
1011 0 : ulong umem_sz = umem_dcache_data_sz - ((ulong)umem_frame0 - (ulong)umem_dcache);
1012 0 : umem_sz = fd_ulong_align_dn( umem_sz, umem_frame_sz );
1013 :
1014 : /* Derive chunk bounds */
1015 :
1016 0 : void * const umem_base = fd_wksp_containing( dcache_mem );
1017 0 : ulong const umem_chunk0 = ( (ulong)umem_frame0 - (ulong)umem_base )>>FD_CHUNK_LG_SZ;
1018 0 : ulong const umem_wmark = umem_chunk0 + ( ( umem_sz-umem_frame_sz )>>FD_CHUNK_LG_SZ );
1019 0 : if( FD_UNLIKELY( umem_chunk0>UINT_MAX || umem_wmark>UINT_MAX || umem_chunk0>umem_wmark ) ) {
1020 0 : FD_LOG_ERR(( "Calculated invalid UMEM bounds [%lu,%lu]", umem_chunk0, umem_wmark ));
1021 0 : }
1022 :
1023 0 : if( FD_UNLIKELY( !umem_base ) ) FD_LOG_ERR(( "UMEM dcache is not in a workspace" ));
1024 0 : if( FD_UNLIKELY( !umem_dcache ) ) FD_LOG_ERR(( "Failed to join UMEM dcache" ));
1025 :
1026 0 : ctx->umem_frame0 = umem_frame0;
1027 0 : ctx->umem_sz = umem_sz;
1028 0 : ctx->umem_chunk0 = (uint)umem_chunk0;
1029 0 : ctx->umem_wmark = (uint)umem_wmark;
1030 :
1031 0 : ctx->free_tx.queue = free_tx;
1032 0 : ctx->free_tx.depth = tile->net.xdp_tx_queue_size;
1033 :
1034 : /* Create and install XSKs */
1035 :
1036 0 : fd_xsk_params_t params0 = {
1037 0 : .if_idx = if_idx,
1038 0 : .if_queue_id = (uint)tile->kind_id,
1039 :
1040 : /* Some kernels produce EOPNOTSUP errors on sendto calls when
1041 : starting up without either XDP_ZEROCOPY or XDP_COPY
1042 : (e.g. 5.14.0-503.23.1.el9_5 with i40e) */
1043 0 : .bind_flags = tile->net.zero_copy ? XDP_ZEROCOPY : XDP_COPY,
1044 :
1045 0 : .fr_depth = tile->net.xdp_rx_queue_size*2,
1046 0 : .rx_depth = tile->net.xdp_rx_queue_size,
1047 0 : .cr_depth = tile->net.xdp_tx_queue_size,
1048 0 : .tx_depth = tile->net.xdp_tx_queue_size,
1049 :
1050 0 : .umem_addr = umem_frame0,
1051 0 : .frame_sz = umem_frame_sz,
1052 0 : .umem_sz = umem_sz
1053 0 : };
1054 :
1055 0 : int xsk_map_fd = 123462;
1056 0 : ctx->prog_link_fds[ 0 ] = 123463;
1057 : /* Init XSK */
1058 0 : if( FD_UNLIKELY( !fd_xsk_init( &ctx->xsk[ 0 ], ¶ms0 ) ) ) FD_LOG_ERR(( "failed to bind xsk for net tile %lu", tile->kind_id ));
1059 0 : if( FD_UNLIKELY( !fd_xsk_activate( &ctx->xsk[ 0 ], xsk_map_fd ) ) ) FD_LOG_ERR(( "failed to activate xsk for net tile %lu", tile->kind_id ));
1060 0 : ctx->xsk_cnt = 1;
1061 :
1062 0 : if( FD_UNLIKELY( fd_sandbox_gettid()==fd_sandbox_getpid() ) ) {
1063 : /* Kind of gross.. in single threaded mode we don't want to close the xsk_map_fd
1064 : since it's shared with other net tiles. Just check for that by seeing if we
1065 : are the only thread in the process. */
1066 0 : if( FD_UNLIKELY( -1==close( xsk_map_fd ) ) ) FD_LOG_ERR(( "close(%d) failed (%d-%s)", xsk_map_fd, errno, fd_io_strerror( errno ) ));
1067 0 : }
1068 :
1069 : /* Networking tile at index 0 also binds to loopback (only queue 0 available on lo) */
1070 :
1071 0 : if( FD_UNLIKELY( strcmp( tile->net.interface, "lo" ) && !tile->kind_id ) ) {
1072 0 : ctx->xsk_cnt = 2;
1073 :
1074 0 : ushort udp_port_candidates[] = {
1075 0 : (ushort)tile->net.legacy_transaction_listen_port,
1076 0 : (ushort)tile->net.quic_transaction_listen_port,
1077 0 : (ushort)tile->net.shred_listen_port,
1078 0 : (ushort)tile->net.gossip_listen_port,
1079 0 : (ushort)tile->net.repair_intake_listen_port,
1080 0 : (ushort)tile->net.repair_serve_listen_port,
1081 0 : };
1082 :
1083 0 : uint lo_idx = if_nametoindex( "lo" );
1084 0 : if( FD_UNLIKELY( !lo_idx ) ) FD_LOG_ERR(( "if_nametoindex(lo) failed" ));
1085 :
1086 : /* FIXME move this to fd_topo_run */
1087 0 : fd_xdp_fds_t lo_fds = fd_xdp_install( lo_idx,
1088 0 : sizeof(udp_port_candidates)/sizeof(udp_port_candidates[0]),
1089 0 : udp_port_candidates,
1090 0 : "skb" );
1091 :
1092 0 : ctx->prog_link_fds[ 1 ] = lo_fds.prog_link_fd;
1093 : /* init xsk 1 */
1094 0 : fd_xsk_params_t params1 = params0;
1095 0 : params1.if_idx = lo_idx; /* probably always 1 */
1096 0 : params1.if_queue_id = 0;
1097 0 : params1.bind_flags = 0;
1098 0 : if( FD_UNLIKELY( !fd_xsk_init( &ctx->xsk[ 1 ], ¶ms1 ) ) ) FD_LOG_ERR(( "failed to bind lo_xsk" ));
1099 0 : if( FD_UNLIKELY( !fd_xsk_activate( &ctx->xsk[ 1 ], lo_fds.xsk_map_fd ) ) ) FD_LOG_ERR(( "failed to activate lo_xsk" ));
1100 0 : if( FD_UNLIKELY( -1==close( lo_fds.xsk_map_fd ) ) ) FD_LOG_ERR(( "close(%d) failed (%d-%s)", xsk_map_fd, errno, fd_io_strerror( errno ) ));
1101 0 : }
1102 :
1103 0 : double tick_per_ns = fd_tempo_tick_per_ns( NULL );
1104 0 : ctx->xdp_stats_interval_ticks = (long)( FD_XDP_STATS_INTERVAL_NS * tick_per_ns );
1105 :
1106 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, 1UL );
1107 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
1108 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
1109 0 : }
1110 :
1111 : static void
1112 : unprivileged_init( fd_topo_t * topo,
1113 0 : fd_topo_tile_t * tile ) {
1114 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1115 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1116 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
1117 0 : FD_TEST( ctx->xsk_cnt!=0 );
1118 :
1119 0 : ctx->net_tile_id = (uint)tile->kind_id;
1120 0 : ctx->net_tile_cnt = (uint)fd_topo_tile_name_cnt( topo, tile->name );
1121 :
1122 0 : ctx->shred_listen_port = tile->net.shred_listen_port;
1123 0 : ctx->quic_transaction_listen_port = tile->net.quic_transaction_listen_port;
1124 0 : ctx->legacy_transaction_listen_port = tile->net.legacy_transaction_listen_port;
1125 0 : ctx->gossip_listen_port = tile->net.gossip_listen_port;
1126 0 : ctx->repair_intake_listen_port = tile->net.repair_intake_listen_port;
1127 0 : ctx->repair_serve_listen_port = tile->net.repair_serve_listen_port;
1128 :
1129 : /* Put a bound on chunks we read from the input, to make sure they
1130 : are within in the data region of the workspace. */
1131 :
1132 0 : if( FD_UNLIKELY( !tile->in_cnt ) ) FD_LOG_ERR(( "net tile in link cnt is zero" ));
1133 0 : if( FD_UNLIKELY( tile->in_cnt>MAX_NET_INS ) ) FD_LOG_ERR(( "net tile in link cnt %lu exceeds MAX_NET_INS %lu", tile->in_cnt, MAX_NET_INS ));
1134 0 : FD_TEST( tile->in_cnt<=32 );
1135 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
1136 0 : fd_topo_link_t * link = &topo->links[ tile->in_link_id[ i ] ];
1137 0 : if( FD_UNLIKELY( link->mtu!=FD_NET_MTU ) ) FD_LOG_ERR(( "net tile in link does not have a normal MTU" ));
1138 :
1139 0 : ctx->in[ i ].mem = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
1140 0 : ctx->in[ i ].chunk0 = fd_dcache_compact_chunk0( ctx->in[ i ].mem, link->dcache );
1141 0 : ctx->in[ i ].wmark = fd_dcache_compact_wmark( ctx->in[ i ].mem, link->dcache, link->mtu );
1142 0 : }
1143 :
1144 0 : for( ulong i = 0; i < tile->out_cnt; i++ ) {
1145 0 : fd_topo_link_t * out_link = &topo->links[ tile->out_link_id[ i ] ];
1146 0 : if( strcmp( out_link->name, "net_quic" ) == 0 ) {
1147 0 : fd_topo_link_t * quic_out = out_link;
1148 0 : ctx->quic_out->mcache = quic_out->mcache;
1149 0 : ctx->quic_out->sync = fd_mcache_seq_laddr( ctx->quic_out->mcache );
1150 0 : ctx->quic_out->depth = fd_mcache_depth( ctx->quic_out->mcache );
1151 0 : ctx->quic_out->seq = fd_mcache_seq_query( ctx->quic_out->sync );
1152 0 : } else if( strcmp( out_link->name, "net_shred" ) == 0 ) {
1153 0 : fd_topo_link_t * shred_out = out_link;
1154 0 : ctx->shred_out->mcache = shred_out->mcache;
1155 0 : ctx->shred_out->sync = fd_mcache_seq_laddr( ctx->shred_out->mcache );
1156 0 : ctx->shred_out->depth = fd_mcache_depth( ctx->shred_out->mcache );
1157 0 : ctx->shred_out->seq = fd_mcache_seq_query( ctx->shred_out->sync );
1158 0 : } else if( strcmp( out_link->name, "net_gossip" ) == 0 ) {
1159 0 : fd_topo_link_t * gossip_out = out_link;
1160 0 : ctx->gossip_out->mcache = gossip_out->mcache;
1161 0 : ctx->gossip_out->sync = fd_mcache_seq_laddr( ctx->gossip_out->mcache );
1162 0 : ctx->gossip_out->depth = fd_mcache_depth( ctx->gossip_out->mcache );
1163 0 : ctx->gossip_out->seq = fd_mcache_seq_query( ctx->gossip_out->sync );
1164 0 : } else if( strcmp( out_link->name, "net_repair" ) == 0 ) {
1165 0 : fd_topo_link_t * repair_out = out_link;
1166 0 : ctx->repair_out->mcache = repair_out->mcache;
1167 0 : ctx->repair_out->sync = fd_mcache_seq_laddr( ctx->repair_out->mcache );
1168 0 : ctx->repair_out->depth = fd_mcache_depth( ctx->repair_out->mcache );
1169 0 : ctx->repair_out->seq = fd_mcache_seq_query( ctx->repair_out->sync );
1170 0 : } else if( strcmp( out_link->name, "net_netlnk" ) == 0 ) {
1171 0 : fd_topo_link_t * netlink_out = out_link;
1172 0 : ctx->neigh4_solicit->mcache = netlink_out->mcache;
1173 0 : ctx->neigh4_solicit->depth = fd_mcache_depth( ctx->neigh4_solicit->mcache );
1174 0 : ctx->neigh4_solicit->seq = fd_mcache_seq_query( fd_mcache_seq_laddr( ctx->neigh4_solicit->mcache ) );
1175 0 : } else {
1176 0 : FD_LOG_ERR(( "unrecognized out link `%s`", out_link->name ));
1177 0 : }
1178 0 : }
1179 :
1180 : /* Check if any of the tiles we set a listen port for do not have an outlink. */
1181 0 : if( FD_UNLIKELY( ctx->shred_listen_port!=0 && ctx->shred_out->mcache==NULL ) ) {
1182 0 : FD_LOG_ERR(( "shred listen port set but no out link was found" ));
1183 0 : } else if( FD_UNLIKELY( ctx->quic_transaction_listen_port!=0 && ctx->quic_out->mcache==NULL ) ) {
1184 0 : FD_LOG_ERR(( "quic transaction listen port set but no out link was found" ));
1185 0 : } else if( FD_UNLIKELY( ctx->legacy_transaction_listen_port!=0 && ctx->quic_out->mcache==NULL ) ) {
1186 0 : FD_LOG_ERR(( "legacy transaction listen port set but no out link was found" ));
1187 0 : } else if( FD_UNLIKELY( ctx->gossip_listen_port!=0 && ctx->gossip_out->mcache==NULL ) ) {
1188 0 : FD_LOG_ERR(( "gossip listen port set but no out link was found" ));
1189 0 : } else if( FD_UNLIKELY( ctx->repair_intake_listen_port!=0 && ctx->repair_out->mcache==NULL ) ) {
1190 0 : FD_LOG_ERR(( "repair intake port set but no out link was found" ));
1191 0 : } else if( FD_UNLIKELY( ctx->repair_serve_listen_port!=0 && ctx->repair_out->mcache==NULL ) ) {
1192 0 : FD_LOG_ERR(( "repair serve listen port set but no out link was found" ));
1193 0 : } else if( FD_UNLIKELY( ctx->neigh4_solicit->mcache==NULL ) ) {
1194 0 : FD_LOG_ERR(( "netlink request link not found" ));
1195 0 : }
1196 :
1197 0 : for( uint j=0U; j<2U; j++ ) {
1198 0 : ctx->tx_flusher[ j ].pending_wmark = (ulong)( (double)tile->net.xdp_tx_queue_size * 0.7 );
1199 0 : ctx->tx_flusher[ j ].tail_flush_backoff = (long)( (double)tile->net.tx_flush_timeout_ns * fd_tempo_tick_per_ns( NULL ) );
1200 0 : ctx->tx_flusher[ j ].next_tail_flush_ticks = LONG_MAX;
1201 0 : }
1202 :
1203 : /* Join netbase objects */
1204 0 : ctx->fib_local = fd_fib4_join( fd_topo_obj_laddr( topo, tile->net.fib4_local_obj_id ) );
1205 0 : ctx->fib_main = fd_fib4_join( fd_topo_obj_laddr( topo, tile->net.fib4_main_obj_id ) );
1206 0 : if( FD_UNLIKELY( !ctx->fib_local || !ctx->fib_main ) ) FD_LOG_ERR(( "fd_fib4_join failed" ));
1207 0 : if( FD_UNLIKELY( !fd_neigh4_hmap_join(
1208 0 : ctx->neigh4,
1209 0 : fd_topo_obj_laddr( topo, tile->net.neigh4_obj_id ),
1210 0 : fd_topo_obj_laddr( topo, tile->net.neigh4_ele_obj_id ) ) ) ) {
1211 0 : FD_LOG_ERR(( "fd_neigh4_hmap_join failed" ));
1212 0 : }
1213 :
1214 : /* Initialize TX free ring */
1215 :
1216 0 : ulong const frame_sz = 2048UL;
1217 0 : ulong frame_off = 0UL;
1218 0 : ulong const tx_depth = ctx->free_tx.depth;
1219 0 : for( ulong j=0; j<tx_depth; j++ ) {
1220 0 : ctx->free_tx.queue[ j ] = (ulong)ctx->umem_frame0 + frame_off;
1221 0 : frame_off += frame_sz;
1222 0 : }
1223 0 : ctx->free_tx.prod = tx_depth;
1224 :
1225 : /* Initialize RX mcache chunks */
1226 :
1227 0 : for( ulong i=0UL; i<(tile->out_cnt); i++ ) {
1228 0 : fd_topo_link_t * out_link = &topo->links[ tile->out_link_id[ i ] ];
1229 0 : fd_frag_meta_t * mcache = out_link->mcache;
1230 0 : for( ulong j=0UL; j<fd_mcache_depth( mcache ); j++ ) {
1231 0 : mcache[ j ].chunk = (uint)( ctx->umem_chunk0 + (frame_off>>FD_CHUNK_LG_SZ) );
1232 0 : frame_off += frame_sz;
1233 0 : }
1234 0 : }
1235 :
1236 : /* Initialize FILL ring */
1237 :
1238 0 : int _charge_busy = 0;
1239 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
1240 0 : frame_off = net_xsk_bootstrap( ctx, j, frame_off );
1241 0 : net_rx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
1242 0 : net_tx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
1243 0 : }
1244 :
1245 0 : if( FD_UNLIKELY( frame_off > ctx->umem_sz ) ) {
1246 0 : FD_LOG_ERR(( "UMEM is too small" ));
1247 0 : }
1248 0 : }
1249 :
1250 : static ulong
1251 : populate_allowed_seccomp( fd_topo_t const * topo,
1252 : fd_topo_tile_t const * tile,
1253 : ulong out_cnt,
1254 0 : struct sock_filter * out ) {
1255 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1256 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1257 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_net_ctx_t ), sizeof( fd_net_ctx_t ) );
1258 :
1259 : /* A bit of a hack, if there is no loopback XSK for this tile, we still need to pass
1260 : two "allow" FD arguments to the net policy, so we just make them both the same. */
1261 0 : int allow_fd2 = ctx->xsk_cnt>1UL ? ctx->xsk[ 1 ].xsk_fd : ctx->xsk[ 0 ].xsk_fd;
1262 0 : FD_TEST( ctx->xsk[ 0 ].xsk_fd >= 0 && allow_fd2 >= 0 );
1263 0 : populate_sock_filter_policy_xdp( out_cnt, out, (uint)fd_log_private_logfile_fd(), (uint)ctx->xsk[ 0 ].xsk_fd, (uint)allow_fd2 );
1264 0 : return sock_filter_policy_xdp_instr_cnt;
1265 0 : }
1266 :
1267 : static ulong
1268 : populate_allowed_fds( fd_topo_t const * topo,
1269 : fd_topo_tile_t const * tile,
1270 : ulong out_fds_cnt,
1271 0 : int * out_fds ) {
1272 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1273 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1274 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_net_ctx_t ), sizeof( fd_net_ctx_t ) );
1275 :
1276 0 : if( FD_UNLIKELY( out_fds_cnt<6UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
1277 :
1278 0 : ulong out_cnt = 0UL;
1279 :
1280 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
1281 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
1282 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
1283 :
1284 0 : out_fds[ out_cnt++ ] = ctx->xsk[ 0 ].xsk_fd;
1285 0 : out_fds[ out_cnt++ ] = ctx->prog_link_fds[ 0 ];
1286 0 : if( FD_LIKELY( ctx->xsk_cnt>1UL ) ) out_fds[ out_cnt++ ] = ctx->xsk[ 1 ].xsk_fd;
1287 0 : if( FD_LIKELY( ctx->xsk_cnt>1UL ) ) out_fds[ out_cnt++ ] = ctx->prog_link_fds[ 1 ];
1288 0 : return out_cnt;
1289 0 : }
1290 :
1291 0 : #define STEM_BURST (1UL)
1292 0 : #define STEM_LAZY ((ulong)30e3) /* 30 us */
1293 :
1294 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_net_ctx_t
1295 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_net_ctx_t)
1296 :
1297 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
1298 0 : #define STEM_CALLBACK_DURING_HOUSEKEEPING during_housekeeping
1299 0 : #define STEM_CALLBACK_BEFORE_CREDIT before_credit
1300 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
1301 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
1302 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
1303 :
1304 : #include "../../stem/fd_stem.c"
1305 :
1306 : fd_topo_run_tile_t fd_tile_net = {
1307 : .name = "net",
1308 : .populate_allowed_seccomp = populate_allowed_seccomp,
1309 : .populate_allowed_fds = populate_allowed_fds,
1310 : .scratch_align = scratch_align,
1311 : .scratch_footprint = scratch_footprint,
1312 : .privileged_init = privileged_init,
1313 : .unprivileged_init = unprivileged_init,
1314 : .run = stem_run,
1315 : };
|