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 <netinet/in.h>
9 : #include <sys/socket.h> /* MSG_DONTWAIT needed before importing the net seccomp filter */
10 : #include <linux/if_xdp.h>
11 :
12 : #include "../fd_net_common.h"
13 : #include "../../metrics/fd_metrics.h"
14 : #include "../../netlink/fd_netlink_tile.h" /* neigh4_solicit */
15 : #include "../../topo/fd_topo.h"
16 :
17 : #include "../../../waltz/ip/fd_fib4.h"
18 : #include "../../../waltz/neigh/fd_neigh4_map.h"
19 : #include "../../../waltz/mib/fd_netdev_tbl.h"
20 : #include "../../../waltz/mib/fd_dbl_buf.h"
21 : #include "../../../waltz/xdp/fd_xdp_redirect_user.h" /* fd_xsk_activate */
22 : #include "../../../waltz/xdp/fd_xsk.h"
23 : #include "../../../util/log/fd_dtrace.h"
24 : #include "../../../util/net/fd_eth.h"
25 : #include "../../../util/net/fd_ip4.h"
26 : #include "../../../util/net/fd_gre.h"
27 :
28 : #include <unistd.h>
29 : #include <linux/if.h> /* struct ifreq */
30 : #include <sys/ioctl.h>
31 : #include <linux/unistd.h>
32 : #include <linux/if_arp.h>
33 :
34 : #include "generated/xdp_seccomp.h"
35 :
36 : /* MAX_NET_INS controls the max number of TX links that a net tile can
37 : serve. */
38 :
39 : #define MAX_NET_INS (32UL)
40 :
41 : /* FD_XDP_STATS_INTERVAL_NS controls the XDP stats refresh interval.
42 : This should be lower than the interval at which the metrics tile
43 : collects metrics. */
44 :
45 0 : #define FD_XDP_STATS_INTERVAL_NS (11e6) /* 11ms */
46 :
47 : /* XSK_IDX_{MAIN,LO} are the hardcoded XSK indices in ctx->xsk[ ... ].
48 : Only net tile 0 has XSK_IDX_LO, all net tiles have XSK_IDX_MAIN. */
49 :
50 0 : #define XSK_IDX_MAIN 0
51 0 : #define XSK_IDX_LO 1
52 :
53 : /* fd_net_in_ctx_t contains consumer information for an incoming tango
54 : link. It is used as part of the TX path. */
55 :
56 : typedef struct {
57 : fd_wksp_t * mem;
58 : ulong chunk0;
59 : ulong wmark;
60 : } fd_net_in_ctx_t;
61 :
62 : /* fd_net_out_ctx_t contains publisher information for a link to a
63 : downstream app tile. It is used as part of the RX path. */
64 :
65 : typedef struct {
66 : fd_frag_meta_t * mcache;
67 : ulong * sync;
68 : ulong depth;
69 : ulong seq;
70 : } fd_net_out_ctx_t;
71 :
72 : /* fd_net_flusher_t controls the pacing of XDP sendto calls for flushing
73 : TX batches. In the 'wakeup' XDP mode, no TX occurs unless the net
74 : tile wakes up the kernel periodically using the sendto() syscall.
75 : If sendto() is called too frequently, time is wasted on context
76 : switches. If sendto() is called not often enough, packets are
77 : delayed or dropped. sendto() calls make almost no guarantees how
78 : much packets are sent out, nor do they indicate when the kernel
79 : finishes a wakeup call (asynchronously dispatched). The net tile
80 : thus uses a myraid of flush triggers that were tested for best
81 : performance. */
82 :
83 : struct fd_net_flusher {
84 :
85 : /* Packets that were enqueued after the last sendto() wakeup are
86 : considered "pending". If there are more than pending_wmark packets
87 : pending, a wakeup is dispatched. Thus, this dispatch trigger is
88 : proportional to packet rate, but does not trigger if I/O is seldom. */
89 : ulong pending_cnt;
90 : ulong pending_wmark;
91 :
92 : /* Sometimes, packets are not flushed out even after a sendto()
93 : wakeup. This can result in the tail of a burst getting delayed or
94 : overrun. If more than tail_flush_backoff ticks pass since the last
95 : sendto() wakeup and there are still unacknowledged packets in the
96 : TX ring, issues another wakeup. */
97 : long next_tail_flush_ticks;
98 : long tail_flush_backoff;
99 :
100 : };
101 :
102 : typedef struct fd_net_flusher fd_net_flusher_t;
103 :
104 : FD_PROTOTYPES_BEGIN
105 :
106 : /* fd_net_flusher_inc marks a new packet as enqueued. */
107 :
108 : static inline void
109 : fd_net_flusher_inc( fd_net_flusher_t * flusher,
110 0 : long now ) {
111 0 : flusher->pending_cnt++;
112 0 : long next_flush = now + flusher->tail_flush_backoff;
113 0 : flusher->next_tail_flush_ticks = fd_long_min( flusher->next_tail_flush_ticks, next_flush );
114 0 : }
115 :
116 : /* fd_net_flusher_check returns 1 if a sendto() wakeup should be issued
117 : immediately. now is a recent fd_tickcount() value.
118 : If tx_ring_empty==0 then the kernel is caught up with the net tile
119 : on the XDP TX ring. (Otherwise, the kernel is behind the net tile) */
120 :
121 : static inline int
122 : fd_net_flusher_check( fd_net_flusher_t * flusher,
123 : long now,
124 0 : int tx_ring_empty ) {
125 0 : int flush_level = flusher->pending_cnt >= flusher->pending_wmark;
126 0 : int flush_timeout = now >= flusher->next_tail_flush_ticks;
127 0 : int flush = flush_level || flush_timeout;
128 0 : if( !flush ) return 0;
129 0 : if( FD_UNLIKELY( tx_ring_empty ) ) {
130 : /* Flush requested but caught up */
131 0 : flusher->pending_cnt = 0UL;
132 0 : flusher->next_tail_flush_ticks = LONG_MAX;
133 0 : return 0;
134 0 : }
135 0 : return 1;
136 0 : }
137 :
138 : /* fd_net_flusher_wakeup signals a sendto() wakeup was done. now is a
139 : recent fd_tickcount() value. */
140 :
141 : static inline void
142 : fd_net_flusher_wakeup( fd_net_flusher_t * flusher,
143 0 : long now ) {
144 0 : flusher->pending_cnt = 0UL;
145 0 : flusher->next_tail_flush_ticks = now + flusher->tail_flush_backoff;
146 0 : }
147 :
148 : FD_PROTOTYPES_END
149 :
150 : /* fd_net_free_ring is a FIFO queue that stores pointers to free XDP TX
151 : frames. */
152 :
153 : struct fd_net_free_ring {
154 : ulong prod;
155 : ulong cons;
156 : ulong depth;
157 : ulong * queue;
158 : };
159 : typedef struct fd_net_free_ring fd_net_free_ring_t;
160 :
161 : typedef struct {
162 : /* An "XSK" is an AF_XDP socket */
163 : uint xsk_cnt;
164 : fd_xsk_t xsk[ 2 ];
165 : int prog_link_fds[ 2 ];
166 :
167 : /* UMEM frame region within dcache */
168 : void * umem_frame0; /* First UMEM frame */
169 : ulong umem_sz; /* Usable UMEM size starting at frame0 */
170 :
171 : /* UMEM chunk region within workspace */
172 : uint umem_chunk0; /* Lowest allowed chunk number */
173 : uint umem_wmark; /* Highest allowed chunk number */
174 :
175 : /* All net tiles are subscribed to the same TX links. (These are
176 : incoming links from app tiles asking the net tile to send out packets)
177 : The net tiles "take turns" doing TX jobs based on the L3+L4 dst hash.
178 : net_tile_id is the index of the current interface, net_tile_cnt is the
179 : total amount of interfaces. */
180 : uint net_tile_id;
181 : uint net_tile_cnt;
182 :
183 : /* Details pertaining to an inflight send op */
184 : struct {
185 : uint xsk_idx;
186 : void * frame;
187 : uchar mac_addrs[12]; /* First 12 bytes of Ethernet header */
188 : uint src_ip; /* src_ip in net order */
189 :
190 : uint use_gre; /* The tx packet will be GRE-encapsulated */
191 : uint gre_outer_src_ip; /* For GRE: Outer iphdr's src_ip in net order */
192 : uint gre_outer_dst_ip; /* For GRE: Outer iphdr's dst_ip in net order */
193 : } tx_op;
194 :
195 : /* Round-robin cycle serivce operations */
196 : uint rr_idx;
197 :
198 : /* Ring tracking free packet buffers */
199 : fd_net_free_ring_t free_tx;
200 :
201 : uchar src_mac_addr[6];
202 :
203 : uint default_address;
204 : uint bind_address;
205 : ushort shred_listen_port;
206 : ushort quic_transaction_listen_port;
207 : ushort legacy_transaction_listen_port;
208 : ushort gossip_listen_port;
209 : ushort repair_intake_listen_port;
210 : ushort repair_serve_listen_port;
211 : ushort send_src_port;
212 :
213 : ulong in_cnt;
214 : fd_net_in_ctx_t in[ MAX_NET_INS ];
215 :
216 : fd_net_out_ctx_t quic_out[1];
217 : fd_net_out_ctx_t shred_out[1];
218 : fd_net_out_ctx_t gossip_out[1];
219 : fd_net_out_ctx_t repair_out[1];
220 : fd_net_out_ctx_t send_out[1];
221 :
222 : /* XDP stats refresh timer */
223 : long xdp_stats_interval_ticks;
224 : long next_xdp_stats_refresh;
225 :
226 : /* TX flush timers */
227 : fd_net_flusher_t tx_flusher[2]; /* one per XSK */
228 :
229 : /* Route and neighbor tables */
230 : fd_fib4_t const * fib_local;
231 : fd_fib4_t const * fib_main;
232 : fd_neigh4_hmap_t neigh4[1];
233 : fd_netlink_neigh4_solicit_link_t neigh4_solicit[1];
234 :
235 : /* Netdev table */
236 : fd_netdev_tbl_join_t netdev_tbl_handle; // handle to the netdev table itself
237 : fd_dbl_buf_t * netdev_dbl_handle; // handle to the double buffer
238 : uchar * netdev_buf; // local buf to copy in the netdev table
239 : ulong netdev_buf_sz; // size of netdev_buf above
240 : int has_gre_interface; // netdev table has a GRE interface entry
241 :
242 : struct {
243 : ulong rx_pkt_cnt;
244 : ulong rx_bytes_total;
245 : ulong rx_undersz_cnt;
246 : ulong rx_fill_blocked_cnt;
247 : ulong rx_backp_cnt;
248 : long rx_busy_cnt;
249 : long rx_idle_cnt;
250 :
251 : ulong tx_submit_cnt;
252 : ulong tx_complete_cnt;
253 : ulong tx_bytes_total;
254 : ulong tx_route_fail_cnt;
255 : ulong tx_no_xdp_cnt;
256 : ulong tx_neigh_fail_cnt;
257 : ulong tx_full_fail_cnt;
258 : long tx_busy_cnt;
259 : long tx_idle_cnt;
260 :
261 : ulong xsk_tx_wakeup_cnt;
262 : ulong xsk_rx_wakeup_cnt;
263 : } metrics;
264 : } fd_net_ctx_t;
265 :
266 : FD_FN_CONST static inline ulong
267 0 : scratch_align( void ) {
268 0 : return 4096UL;
269 0 : }
270 :
271 : FD_FN_PURE static inline ulong
272 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
273 0 : ulong l = FD_LAYOUT_INIT;
274 0 : l = FD_LAYOUT_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
275 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );
276 0 : l = FD_LAYOUT_APPEND( l, fd_netdev_tbl_align(), fd_netdev_tbl_footprint( NETDEV_MAX, BOND_MASTER_MAX ) );
277 0 : return FD_LAYOUT_FINI( l, scratch_align() );
278 0 : }
279 :
280 : static void
281 0 : metrics_write( fd_net_ctx_t * ctx ) {
282 0 : FD_MCNT_SET( NET, RX_PKT_CNT, ctx->metrics.rx_pkt_cnt );
283 0 : FD_MCNT_SET( NET, RX_BYTES_TOTAL, ctx->metrics.rx_bytes_total );
284 0 : FD_MCNT_SET( NET, RX_UNDERSZ_CNT, ctx->metrics.rx_undersz_cnt );
285 0 : FD_MCNT_SET( NET, RX_FILL_BLOCKED_CNT, ctx->metrics.rx_fill_blocked_cnt );
286 0 : FD_MCNT_SET( NET, RX_BACKPRESSURE_CNT, ctx->metrics.rx_backp_cnt );
287 0 : FD_MGAUGE_SET( NET, RX_BUSY_CNT, (ulong)fd_long_max( ctx->metrics.rx_busy_cnt, 0L ) );
288 0 : FD_MGAUGE_SET( NET, RX_IDLE_CNT, (ulong)fd_long_max( ctx->metrics.rx_idle_cnt, 0L ) );
289 0 : FD_MGAUGE_SET( NET, TX_BUSY_CNT, (ulong)fd_long_max( ctx->metrics.tx_busy_cnt, 0L ) );
290 0 : FD_MGAUGE_SET( NET, TX_IDLE_CNT, (ulong)fd_long_max( ctx->metrics.tx_idle_cnt, 0L ) );
291 :
292 0 : FD_MCNT_SET( NET, TX_SUBMIT_CNT, ctx->metrics.tx_submit_cnt );
293 0 : FD_MCNT_SET( NET, TX_COMPLETE_CNT, ctx->metrics.tx_complete_cnt );
294 0 : FD_MCNT_SET( NET, TX_BYTES_TOTAL, ctx->metrics.tx_bytes_total );
295 0 : FD_MCNT_SET( NET, TX_ROUTE_FAIL_CNT, ctx->metrics.tx_route_fail_cnt );
296 0 : FD_MCNT_SET( NET, TX_NEIGHBOR_FAIL_CNT, ctx->metrics.tx_neigh_fail_cnt );
297 0 : FD_MCNT_SET( NET, TX_FULL_FAIL_CNT, ctx->metrics.tx_full_fail_cnt );
298 :
299 0 : FD_MCNT_SET( NET, XSK_TX_WAKEUP_CNT, ctx->metrics.xsk_tx_wakeup_cnt );
300 0 : FD_MCNT_SET( NET, XSK_RX_WAKEUP_CNT, ctx->metrics.xsk_rx_wakeup_cnt );
301 0 : }
302 :
303 : struct xdp_statistics_v0 {
304 : __u64 rx_dropped; /* Dropped for other reasons */
305 : __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
306 : __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
307 : };
308 :
309 : struct xdp_statistics_v1 {
310 : __u64 rx_dropped; /* Dropped for other reasons */
311 : __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
312 : __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
313 : __u64 rx_ring_full; /* Dropped due to rx ring being full */
314 : __u64 rx_fill_ring_empty_descs; /* Failed to retrieve item from fill ring */
315 : __u64 tx_ring_empty_descs; /* Failed to retrieve item from tx ring */
316 : };
317 :
318 : static void
319 0 : poll_xdp_statistics( fd_net_ctx_t * ctx ) {
320 0 : struct xdp_statistics_v1 stats = {0};
321 0 : ulong xsk_cnt = ctx->xsk_cnt;
322 0 : for( ulong j=0UL; j<xsk_cnt; j++ ) {
323 0 : struct xdp_statistics_v1 sub_stats;
324 0 : uint optlen = (uint)sizeof(struct xdp_statistics_v1);
325 0 : if( FD_UNLIKELY( -1==getsockopt( ctx->xsk[ j ].xsk_fd, SOL_XDP, XDP_STATISTICS, &sub_stats, &optlen ) ) )
326 0 : FD_LOG_ERR(( "getsockopt(SOL_XDP, XDP_STATISTICS) failed: %s", strerror( errno ) ));
327 0 : if( FD_UNLIKELY( optlen!=sizeof(struct xdp_statistics_v0) &&
328 0 : optlen!=sizeof(struct xdp_statistics_v1) ) ) {
329 0 : FD_LOG_ERR(( "getsockopt(SOL_XDP, XDP_STATISTICS) returned unexpected size %u", optlen ));
330 0 : }
331 0 : stats.rx_dropped += sub_stats.rx_dropped;
332 0 : stats.rx_invalid_descs += sub_stats.rx_invalid_descs;
333 0 : stats.tx_invalid_descs += sub_stats.tx_invalid_descs;
334 0 : stats.rx_ring_full += sub_stats.rx_ring_full;
335 0 : stats.rx_fill_ring_empty_descs += sub_stats.rx_fill_ring_empty_descs;
336 0 : stats.tx_ring_empty_descs += sub_stats.tx_ring_empty_descs;
337 0 : }
338 :
339 0 : FD_MCNT_SET( NET, XDP_RX_DROPPED_OTHER, stats.rx_dropped );
340 0 : FD_MCNT_SET( NET, XDP_RX_INVALID_DESCS, stats.rx_invalid_descs );
341 0 : FD_MCNT_SET( NET, XDP_TX_INVALID_DESCS, stats.tx_invalid_descs );
342 0 : FD_MCNT_SET( NET, XDP_RX_RING_FULL, stats.rx_ring_full );
343 0 : FD_MCNT_SET( NET, XDP_RX_FILL_RING_EMPTY_DESCS, stats.rx_fill_ring_empty_descs );
344 0 : FD_MCNT_SET( NET, XDP_TX_RING_EMPTY_DESCS, stats.tx_ring_empty_descs );
345 0 : }
346 :
347 : /* net_is_fatal_xdp_error returns 1 if the given errno returned by an
348 : XDP API indicates a non-recoverable error code. The net tile should
349 : crash if it sees such an error so the problem does not go undetected.
350 : Otherwise, returns 0. */
351 :
352 : static int
353 0 : net_is_fatal_xdp_error( int err ) {
354 0 : return err==ESOCKTNOSUPPORT || err==EOPNOTSUPP || err==EINVAL ||
355 0 : err==EPERM;
356 0 : }
357 :
358 : /* Load the netdev table to ctx->netdev_buf. Create a join in ctx->netdev_tbl_handle */
359 :
360 : static void
361 0 : net_load_netdev_tbl( fd_net_ctx_t * ctx ) {
362 : /* Copy from double buffer */
363 0 : if( FD_UNLIKELY( fd_dbl_buf_read( ctx->netdev_dbl_handle, ctx->netdev_buf_sz, ctx->netdev_buf, NULL ) ) == 0) FD_LOG_ERR(("netdev table load failed"));
364 :
365 : /* Recalculate the join handler to netdeb buf */
366 0 : if( FD_UNLIKELY( !fd_netdev_tbl_join( &ctx->netdev_tbl_handle, ctx->netdev_buf ) ) ) FD_LOG_ERR(("netdev table join failed"));
367 0 : }
368 :
369 : /* Query the netdev table. Return a fd_netdev_t pointer to the net device of the
370 : interface specified by if_idx. Null if the if_idx is invalid */
371 :
372 : static fd_netdev_t *
373 : net_query_netdev_tbl( fd_net_ctx_t *ctx,
374 0 : uint if_idx ) {
375 : /* dev_tbl is one-indexed */
376 0 : if( if_idx>ctx->netdev_tbl_handle.hdr->dev_cnt ) {
377 0 : FD_LOG_NOTICE(( "interface idx invalid (%u). dev_cnt (%u)", if_idx, ctx->netdev_tbl_handle.hdr->dev_cnt ));
378 0 : return NULL;
379 0 : }
380 :
381 0 : return &ctx->netdev_tbl_handle.dev_tbl[if_idx];
382 0 : }
383 :
384 : /* Iterates the netdev table and returns 1 if a GRE interface exists, 0 otherwise.
385 : Only called in privileged_init and during_housekeeping */
386 :
387 : static int
388 0 : net_check_gre_interface_exists( fd_net_ctx_t * ctx ) {
389 0 : fd_netdev_t * dev_tbl = ctx->netdev_tbl_handle.dev_tbl;
390 0 : ushort dev_cnt = ctx->netdev_tbl_handle.hdr->dev_cnt;
391 :
392 0 : for( ushort if_idx = 0; if_idx<dev_cnt; if_idx++ ) {
393 0 : if( dev_tbl[if_idx].dev_type==ARPHRD_IPGRE ) return 1;
394 0 : }
395 0 : return 0;
396 0 : }
397 :
398 :
399 : /* net_tx_ready returns 1 if the current XSK is ready to submit a TX send
400 : job. If the XSK is blocked for sends, returns 0. Reasons for block
401 : include:
402 : - No XSK TX buffer is available
403 : - XSK TX ring is full */
404 :
405 : static int
406 : net_tx_ready( fd_net_ctx_t * ctx,
407 0 : uint xsk_idx ) {
408 0 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
409 0 : fd_xdp_ring_t * tx_ring = &xsk->ring_tx;
410 0 : fd_net_free_ring_t * free = &ctx->free_tx;
411 0 : if( free->prod == free->cons ) return 0; /* drop */
412 0 : if( tx_ring->prod - tx_ring->cons >= tx_ring->depth ) return 0; /* drop */
413 0 : return 1;
414 0 : }
415 :
416 : /* net_rx_wakeup triggers xsk_recvmsg to run in the kernel. Needs to be
417 : called periodically in order to receive packets. */
418 :
419 : static void
420 : net_rx_wakeup( fd_net_ctx_t * ctx,
421 : fd_xsk_t * xsk,
422 0 : int * charge_busy ) {
423 0 : if( !fd_xsk_rx_need_wakeup( xsk ) ) return;
424 0 : *charge_busy = 1;
425 0 : struct msghdr _ignored[ 1 ] = { 0 };
426 0 : if( FD_UNLIKELY( -1==recvmsg( xsk->xsk_fd, _ignored, MSG_DONTWAIT ) ) ) {
427 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
428 0 : FD_LOG_ERR(( "xsk recvmsg failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
429 0 : }
430 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
431 0 : long ts = fd_log_wallclock();
432 0 : if( ts > xsk->log_suppress_until_ns ) {
433 0 : FD_LOG_WARNING(( "xsk recvmsg failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
434 0 : xsk->log_suppress_until_ns = ts + (long)1e9;
435 0 : }
436 0 : }
437 0 : }
438 0 : ctx->metrics.xsk_rx_wakeup_cnt++;
439 0 : }
440 :
441 : /* net_tx_wakeup triggers xsk_sendmsg to run in the kernel. Needs to be
442 : called periodically in order to transmit packets. */
443 :
444 : static void
445 : net_tx_wakeup( fd_net_ctx_t * ctx,
446 : fd_xsk_t * xsk,
447 0 : int * charge_busy ) {
448 0 : if( !fd_xsk_tx_need_wakeup( xsk ) ) return;
449 0 : if( FD_VOLATILE_CONST( *xsk->ring_tx.prod )==FD_VOLATILE_CONST( *xsk->ring_tx.cons ) ) return;
450 0 : *charge_busy = 1;
451 0 : if( FD_UNLIKELY( -1==sendto( xsk->xsk_fd, NULL, 0, MSG_DONTWAIT, NULL, 0 ) ) ) {
452 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
453 0 : FD_LOG_ERR(( "xsk sendto failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
454 0 : }
455 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
456 0 : long ts = fd_log_wallclock();
457 0 : if( ts > xsk->log_suppress_until_ns ) {
458 0 : FD_LOG_WARNING(( "xsk sendto failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
459 0 : xsk->log_suppress_until_ns = ts + (long)1e9;
460 0 : }
461 0 : }
462 0 : }
463 0 : ctx->metrics.xsk_tx_wakeup_cnt++;
464 0 : }
465 :
466 : /* net_tx_periodic_wakeup does a timer based xsk_sendmsg wakeup. */
467 :
468 : static inline int
469 : net_tx_periodic_wakeup( fd_net_ctx_t * ctx,
470 : uint xsk_idx,
471 : long now,
472 0 : int * charge_busy ) {
473 0 : uint tx_prod = FD_VOLATILE_CONST( *ctx->xsk[ xsk_idx ].ring_tx.prod );
474 0 : uint tx_cons = FD_VOLATILE_CONST( *ctx->xsk[ xsk_idx ].ring_tx.cons );
475 0 : int tx_ring_empty = tx_prod==tx_cons;
476 0 : if( fd_net_flusher_check( ctx->tx_flusher+xsk_idx, now, tx_ring_empty ) ) {
477 0 : net_tx_wakeup( ctx, &ctx->xsk[ xsk_idx ], charge_busy );
478 0 : fd_net_flusher_wakeup( ctx->tx_flusher+xsk_idx, now );
479 0 : }
480 0 : return 0;
481 0 : }
482 :
483 : static void
484 0 : during_housekeeping( fd_net_ctx_t * ctx ) {
485 0 : long now = fd_tickcount();
486 0 : net_load_netdev_tbl( ctx );
487 0 : ctx->has_gre_interface = net_check_gre_interface_exists( ctx );
488 :
489 0 : ctx->metrics.rx_busy_cnt = 0UL;
490 0 : ctx->metrics.rx_idle_cnt = 0UL;
491 0 : ctx->metrics.tx_busy_cnt = 0UL;
492 0 : ctx->metrics.tx_idle_cnt = fd_seq_diff( ctx->free_tx.prod, ctx->free_tx.cons );
493 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
494 0 : fd_xsk_t * xsk = &ctx->xsk[ j ];
495 : /* Refresh all sequence numbers (consumer first, then producer) */
496 0 : FD_COMPILER_MFENCE();
497 0 : xsk->ring_fr.cached_cons = FD_VOLATILE_CONST( *xsk->ring_fr.cons );
498 0 : xsk->ring_fr.cached_prod = FD_VOLATILE_CONST( *xsk->ring_fr.prod );
499 0 : xsk->ring_rx.cached_cons = FD_VOLATILE_CONST( *xsk->ring_rx.cons );
500 0 : xsk->ring_rx.cached_prod = FD_VOLATILE_CONST( *xsk->ring_rx.prod );
501 0 : xsk->ring_tx.cached_cons = FD_VOLATILE_CONST( *xsk->ring_tx.cons );
502 0 : xsk->ring_tx.cached_prod = FD_VOLATILE_CONST( *xsk->ring_tx.prod );
503 0 : xsk->ring_cr.cached_cons = FD_VOLATILE_CONST( *xsk->ring_cr.cons );
504 0 : xsk->ring_cr.cached_prod = FD_VOLATILE_CONST( *xsk->ring_cr.prod );
505 0 : FD_COMPILER_MFENCE();
506 0 : ctx->metrics.rx_busy_cnt += (long)(int)( xsk->ring_rx.cached_prod - xsk->ring_rx.cached_cons );
507 0 : ctx->metrics.rx_idle_cnt += (long)(int)( xsk->ring_fr.cached_prod - xsk->ring_fr.cached_cons );
508 0 : ctx->metrics.tx_busy_cnt += (long)(int)( xsk->ring_tx.cached_prod - xsk->ring_tx.cached_cons );
509 0 : ctx->metrics.tx_busy_cnt += (long)(int)( xsk->ring_cr.cached_prod - xsk->ring_cr.cached_cons );
510 0 : }
511 :
512 0 : if( now > ctx->next_xdp_stats_refresh ) {
513 0 : ctx->next_xdp_stats_refresh = now + ctx->xdp_stats_interval_ticks;
514 0 : poll_xdp_statistics( ctx );
515 0 : }
516 0 : }
517 :
518 :
519 : /* net_tx_route resolves the destination interface index, src MAC address,
520 : and dst MAC address. Returns 1 on success, 0 on failure. On success,
521 : tx_op->{if_idx,mac_addrs} is set. */
522 :
523 : static int
524 : net_tx_route( fd_net_ctx_t * ctx,
525 0 : uint dst_ip ) {
526 :
527 : /* Route lookup */
528 :
529 0 : fd_fib4_hop_t hop[2] = {0};
530 0 : fd_fib4_lookup( ctx->fib_local, hop+0, dst_ip, 0UL );
531 0 : fd_fib4_lookup( ctx->fib_main, hop+1, dst_ip, 0UL );
532 0 : fd_fib4_hop_t const * next_hop = fd_fib4_hop_or( hop+0, hop+1 );
533 :
534 0 : uint rtype = next_hop->rtype;
535 0 : uint if_idx = next_hop->if_idx;
536 0 : uint ip4_src = next_hop->ip4_src;
537 :
538 0 : if( FD_UNLIKELY( rtype==FD_FIB4_RTYPE_LOCAL ) ) {
539 0 : rtype = FD_FIB4_RTYPE_UNICAST;
540 0 : if_idx = 1;
541 0 : }
542 :
543 0 : if( FD_UNLIKELY( rtype!=FD_FIB4_RTYPE_UNICAST ) ) {
544 0 : ctx->metrics.tx_route_fail_cnt++;
545 0 : return 0;
546 0 : }
547 :
548 0 : fd_netdev_t * netdev = net_query_netdev_tbl( ctx, if_idx );
549 0 : if( !netdev ) {
550 0 : ctx->metrics.tx_route_fail_cnt++;
551 0 : return 0;
552 0 : }
553 :
554 0 : ip4_src = fd_uint_if( !!ctx->bind_address, ctx->bind_address, ip4_src );
555 0 : ctx->tx_op.src_ip = ip4_src;
556 0 : ctx->tx_op.xsk_idx = UINT_MAX;
557 :
558 0 : if( netdev->dev_type==ARPHRD_LOOPBACK ) {
559 : /* Set Ethernet src and dst address to 00:00:00:00:00:00 */
560 0 : memset( ctx->tx_op.mac_addrs, 0, 12UL );
561 0 : ctx->tx_op.xsk_idx = XSK_IDX_LO;
562 : /* Set preferred src address to 127.0.0.1 if no bind address is set */
563 0 : if( !ctx->tx_op.src_ip ) ctx->tx_op.src_ip = FD_IP4_ADDR( 127,0,0,1 );
564 0 : return 1;
565 0 : } else if( netdev->dev_type==ARPHRD_IPGRE ) {
566 : /* skip MAC addrs lookup for GRE inner dst ip */
567 0 : if( netdev->gre_src_ip ) ctx->tx_op.gre_outer_src_ip = netdev->gre_src_ip;
568 0 : ctx->tx_op.gre_outer_dst_ip = netdev->gre_dst_ip;
569 0 : ctx->tx_op.use_gre = 1;
570 0 : return 1;
571 0 : }
572 :
573 0 : if( FD_UNLIKELY( netdev->dev_type!=ARPHRD_ETHER ) ) return 0; // drop
574 :
575 0 : if( FD_UNLIKELY( if_idx!=ctx->xsk[ XSK_IDX_MAIN ].if_idx ) ) {
576 0 : ctx->metrics.tx_no_xdp_cnt++;
577 0 : return 0;
578 0 : }
579 0 : ctx->tx_op.xsk_idx = XSK_IDX_MAIN;
580 :
581 : /* Neighbor resolve */
582 0 : uint neigh_ip = next_hop->ip4_gw;
583 0 : if( !neigh_ip ) neigh_ip = dst_ip;
584 :
585 0 : fd_neigh4_hmap_query_t neigh_query[1];
586 0 : int neigh_res = fd_neigh4_hmap_query_try( ctx->neigh4, &neigh_ip, NULL, neigh_query, 0 );
587 0 : if( FD_UNLIKELY( neigh_res!=FD_MAP_SUCCESS ) ) {
588 : /* Neighbor not found */
589 0 : fd_netlink_neigh4_solicit( ctx->neigh4_solicit, neigh_ip, if_idx, fd_frag_meta_ts_comp( fd_tickcount() ) );
590 0 : ctx->metrics.tx_neigh_fail_cnt++;
591 0 : return 0;
592 0 : }
593 0 : fd_neigh4_entry_t const * neigh = fd_neigh4_hmap_query_ele_const( neigh_query );
594 0 : if( FD_UNLIKELY( neigh->state != FD_NEIGH4_STATE_ACTIVE ) ) {
595 0 : ctx->metrics.tx_neigh_fail_cnt++;
596 0 : return 0;
597 0 : }
598 0 : ip4_src = fd_uint_if( !ip4_src, ctx->default_address, ip4_src );
599 0 : ctx->tx_op.src_ip = ip4_src;
600 0 : memcpy( ctx->tx_op.mac_addrs+0, neigh->mac_addr, 6 );
601 0 : memcpy( ctx->tx_op.mac_addrs+6, netdev->mac_addr, 6 );
602 :
603 0 : if( FD_UNLIKELY( fd_neigh4_hmap_query_test( neigh_query ) ) ) {
604 0 : ctx->metrics.tx_neigh_fail_cnt++;
605 0 : return 0;
606 0 : }
607 :
608 0 : return 1;
609 0 : }
610 :
611 : /* before_frag is called when a new metadata descriptor for a TX job is
612 : found. This callback determines whether this net tile is responsible
613 : for the TX job. If so, it prepares the TX op for the during_frag and
614 : after_frag callbacks. */
615 :
616 : static inline int
617 : before_frag( fd_net_ctx_t * ctx,
618 : ulong in_idx,
619 : ulong seq,
620 0 : ulong sig ) {
621 0 : (void)in_idx; (void)seq;
622 :
623 : /* Find interface index of next packet */
624 0 : ulong proto = fd_disco_netmux_sig_proto( sig );
625 0 : if( FD_UNLIKELY( proto!=DST_PROTO_OUTGOING ) ) return 1;
626 :
627 : /* Load balance TX */
628 0 : uint net_tile_cnt = ctx->net_tile_cnt;
629 0 : uint hash = (uint)fd_disco_netmux_sig_hash( sig );
630 0 : uint target_idx = hash % net_tile_cnt;
631 :
632 0 : uint dst_ip = fd_disco_netmux_sig_dst_ip( sig );
633 0 : if( FD_UNLIKELY( !net_tx_route( ctx, dst_ip ) ) ) {
634 0 : return 1; /* metrics incremented by net_tx_route */
635 0 : }
636 :
637 0 : uint net_tile_id = ctx->net_tile_id;
638 0 : uint xsk_idx = ctx->tx_op.xsk_idx;
639 :
640 0 : if( ctx->tx_op.use_gre ) {
641 0 : uint inner_src_ip = ctx->tx_op.src_ip;
642 0 : if( FD_UNLIKELY( !inner_src_ip ) ) {
643 0 : ctx->metrics.tx_route_fail_cnt++;
644 0 : return 1;
645 0 : }
646 :
647 : /* Find the MAC addrs for the eth hdr, and src ip for outer ip4 hdr if not found in netdev tbl */
648 0 : ctx->tx_op.src_ip = 0;
649 0 : ctx->tx_op.use_gre = 0;
650 0 : if( FD_UNLIKELY( !net_tx_route( ctx, ctx->tx_op.gre_outer_dst_ip ) ) ) {
651 0 : return 1; /* metrics incremented by net_tx_route */
652 0 : }
653 0 : if( ctx->tx_op.use_gre==1 ) {
654 : /* Only one layer of tunnelling supported */
655 0 : ctx->metrics.tx_route_fail_cnt++;
656 0 : return 1;
657 0 : }
658 0 : if( !ctx->tx_op.gre_outer_src_ip ) {
659 0 : ctx->tx_op.gre_outer_src_ip = ctx->tx_op.src_ip;
660 0 : }
661 0 : ctx->tx_op.use_gre = 1; /* indicate to during_frag to use GRE header */
662 0 : ctx->tx_op.src_ip = inner_src_ip;
663 :
664 0 : xsk_idx = XSK_IDX_MAIN;
665 0 : }
666 :
667 0 : if( FD_UNLIKELY( xsk_idx>=ctx->xsk_cnt ) ) {
668 : /* Packet does not route to an XDP interface */
669 0 : ctx->metrics.tx_no_xdp_cnt++;
670 0 : return 1;
671 0 : }
672 :
673 0 : if( xsk_idx==XSK_IDX_LO ) target_idx = 0; /* loopback always targets tile 0 */
674 :
675 : /* Skip if another net tile is responsible for this packet */
676 :
677 0 : if( net_tile_id!=target_idx ) return 1; /* ignore */
678 :
679 : /* Skip if TX is blocked */
680 :
681 0 : if( FD_UNLIKELY( !net_tx_ready( ctx, xsk_idx ) ) ) {
682 0 : ctx->metrics.tx_full_fail_cnt++;
683 0 : return 1;
684 0 : }
685 :
686 : /* Allocate buffer for receive */
687 :
688 0 : fd_net_free_ring_t * free = &ctx->free_tx;
689 0 : ulong alloc_seq = free->cons;
690 0 : void * frame = (void *)free->queue[ alloc_seq % free->depth ];
691 0 : free->cons = fd_seq_inc( alloc_seq, 1UL );
692 :
693 0 : ctx->tx_op.frame = frame;
694 :
695 0 : return 0; /* continue */
696 0 : }
697 :
698 : /* during_frag is called when before_frag has committed to transmit an
699 : outgoing packet. */
700 :
701 : static inline void
702 : during_frag( fd_net_ctx_t * ctx,
703 : ulong in_idx,
704 : ulong seq FD_PARAM_UNUSED,
705 : ulong sig FD_PARAM_UNUSED,
706 : ulong chunk,
707 : ulong sz,
708 0 : ulong ctl FD_PARAM_UNUSED ) {
709 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz>FD_NET_MTU ) )
710 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 ));
711 :
712 0 : if( FD_UNLIKELY( sz<( sizeof(fd_eth_hdr_t)+sizeof(fd_ip4_hdr_t) ) ) )
713 0 : FD_LOG_ERR(( "packet too small %lu (in_idx=%lu)", sz, in_idx ));
714 :
715 0 : if( FD_UNLIKELY( sz>FD_ETH_PAYLOAD_MAX ) )
716 0 : FD_LOG_ERR(( "packet too big %lu (in_idx=%lu)", sz, in_idx ));
717 :
718 0 : void * frame = ctx->tx_op.frame;
719 0 : if( FD_UNLIKELY( (ulong)frame < (ulong)ctx->umem_frame0 ) )
720 0 : FD_LOG_ERR(( "frame %p out of bounds (below %p)", frame, (void *)ctx->umem_frame0 ));
721 0 : ulong umem_off = (ulong)frame - (ulong)ctx->umem_frame0;
722 0 : if( FD_UNLIKELY( (ulong)umem_off > (ulong)ctx->umem_sz ) )
723 0 : FD_LOG_ERR(( "frame %p out of bounds (beyond %p)", frame, (void *)ctx->umem_sz ));
724 :
725 : /* Speculatively copy frame into XDP buffer */
726 0 : uchar const * src = fd_chunk_to_laddr_const( ctx->in[ in_idx ].mem, chunk );
727 :
728 0 : if( ctx->tx_op.use_gre ) {
729 : /* Discard the ethernet hdr from src. Copy the rest to where the inner ip4_hdr is.
730 : Safe from overflow: FD_ETH_PAYLOAD_MAX + header overhead < frame size (2048UL) */
731 0 : ulong overhead = sizeof(fd_eth_hdr_t) + sizeof(fd_ip4_hdr_t) + sizeof(fd_gre_hdr_t);
732 0 : fd_memcpy( (void *)( (ulong)ctx->tx_op.frame + overhead ), src + sizeof(fd_eth_hdr_t), sz - sizeof(fd_eth_hdr_t) );
733 0 : } else {
734 0 : fd_memcpy( ctx->tx_op.frame, src, sz );
735 0 : }
736 0 : }
737 :
738 : /* after_frag is called when the during_frag memcpy was _not_ overrun. */
739 :
740 : static void
741 : after_frag( fd_net_ctx_t * ctx,
742 : ulong in_idx,
743 : ulong seq,
744 : ulong sig,
745 : ulong sz,
746 : ulong tsorig,
747 : ulong tspub,
748 0 : fd_stem_context_t * stem ) {
749 0 : (void)in_idx; (void)seq; (void)sig; (void)tsorig; (void)tspub; (void)stem;
750 :
751 : /* Current send operation */
752 :
753 0 : uchar * frame = ctx->tx_op.frame;
754 0 : uint xsk_idx = ctx->tx_op.xsk_idx;
755 :
756 : /* Select Ethernet addresses */
757 0 : memcpy( frame, ctx->tx_op.mac_addrs, 12 );
758 :
759 0 : uchar * iphdr = frame + sizeof(fd_eth_hdr_t);
760 :
761 0 : if( ctx->tx_op.use_gre ) {
762 :
763 : /* For GRE packets, the ethertype will always be FD_ETH_HDR_TYPE_IP. outer source ip can't be 0 */
764 0 : if( FD_UNLIKELY( ctx->tx_op.gre_outer_src_ip==0 ) ) {
765 0 : ctx->metrics.tx_route_fail_cnt++;
766 0 : return;
767 0 : }
768 :
769 : /* Write the last two bytes for eth_hdr */
770 0 : FD_STORE( ushort, frame+12, fd_ushort_bswap( FD_ETH_HDR_TYPE_IP ) );
771 :
772 0 : uchar * outer_iphdr = frame + sizeof(fd_eth_hdr_t);
773 0 : uchar * gre_hdr = outer_iphdr + sizeof(fd_ip4_hdr_t);
774 0 : uchar * inner_iphdr = gre_hdr + sizeof(fd_gre_hdr_t);
775 :
776 : /* outer hdr + gre hdr + inner net_tot_len */
777 0 : ushort outer_net_tot_len = sizeof(fd_ip4_hdr_t) + sizeof(fd_gre_hdr_t) + fd_ushort_bswap( ( (fd_ip4_hdr_t *)inner_iphdr )->net_tot_len );
778 :
779 : /* Construct outer ip header */
780 0 : FD_STORE( uchar, outer_iphdr, 0x45 ); // outer_iphdr->verihl = 0x45;
781 0 : FD_STORE( ushort, outer_iphdr+2, fd_ushort_bswap( outer_net_tot_len ) );
782 0 : FD_STORE( uchar, outer_iphdr+8, 64 ); // ttl = 64
783 0 : FD_STORE( uchar, outer_iphdr+9, FD_IP4_HDR_PROTOCOL_GRE );
784 0 : FD_STORE( uint, outer_iphdr+12, ctx->tx_op.gre_outer_src_ip );
785 0 : FD_STORE( uint, outer_iphdr+16, ctx->tx_op.gre_outer_dst_ip );
786 : /* Compute checksum for the outer hdr */
787 0 : FD_STORE( ushort, outer_iphdr+10, 0 );
788 0 : FD_STORE( ushort, outer_iphdr+10, fd_ip4_hdr_check_fast( outer_iphdr ) );
789 :
790 : /* Construct gre header */
791 0 : FD_STORE( ushort, gre_hdr, 0 );
792 0 : FD_STORE( ushort, gre_hdr+2, fd_ushort_bswap( FD_ETH_HDR_TYPE_IP ) );
793 :
794 0 : iphdr = inner_iphdr;
795 0 : sz = sizeof(fd_eth_hdr_t) + outer_net_tot_len;
796 0 : xsk_idx = 0;
797 0 : }
798 :
799 : /* Construct (inner) ip header */
800 0 : uint ihl = FD_IP4_GET_LEN( *(fd_ip4_hdr_t *)iphdr );
801 0 : uint ip4_saddr = FD_LOAD( uint, iphdr+12 );
802 0 : ushort ethertype = FD_LOAD( ushort, frame+12 );
803 :
804 0 : if( ethertype==fd_ushort_bswap( FD_ETH_HDR_TYPE_IP ) && ip4_saddr==0 ) {
805 0 : if( FD_UNLIKELY( ctx->tx_op.src_ip==0 ||
806 0 : ihl<sizeof(fd_ip4_hdr_t) ||
807 0 : (sizeof(fd_eth_hdr_t)+ihl)>sz ) ) {
808 : /* Outgoing IPv4 packet with unknown src IP or invalid IHL */
809 : /* FIXME should select first IPv4 address of device table here */
810 0 : ctx->metrics.tx_route_fail_cnt++;
811 0 : return;
812 0 : }
813 : /* Recompute checksum after changing header */
814 0 : FD_STORE( uint, iphdr+12, ctx->tx_op.src_ip );
815 0 : FD_STORE( ushort, iphdr+10, 0 );
816 0 : FD_STORE( ushort, iphdr+10, fd_ip4_hdr_check( iphdr ) );
817 0 : }
818 :
819 : /* Submit packet TX job
820 :
821 : Invariant for ring_tx: prod-cons<length
822 : (This invariant breaks if any other packet is sent over this ring
823 : between before_frag and this point, e.g. send_arp_probe.) */
824 :
825 :
826 0 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
827 0 : fd_xdp_ring_t * tx_ring = &xsk->ring_tx;
828 0 : uint tx_seq = FD_VOLATILE_CONST( *tx_ring->prod );
829 0 : uint tx_mask = tx_ring->depth - 1U;
830 0 : xsk->ring_tx.packet_ring[ tx_seq&tx_mask ] = (struct xdp_desc) {
831 0 : .addr = (ulong)frame - (ulong)ctx->umem_frame0,
832 0 : .len = (uint)sz,
833 0 : .options = 0
834 0 : };
835 :
836 : /* Frame is now owned by kernel. Clear tx_op. */
837 0 : ctx->tx_op.frame = NULL;
838 :
839 : /* Register newly enqueued packet */
840 0 : FD_VOLATILE( *xsk->ring_tx.prod ) = tx_ring->cached_prod = tx_seq+1U;
841 0 : ctx->metrics.tx_submit_cnt++;
842 0 : ctx->metrics.tx_bytes_total += sz;
843 0 : fd_net_flusher_inc( ctx->tx_flusher+xsk_idx, fd_tickcount() );
844 :
845 0 : }
846 :
847 : /* net_rx_packet is called when a new Ethernet frame is available.
848 : Attempts to copy out the frame to a downstream tile. */
849 :
850 : static void
851 : net_rx_packet( fd_net_ctx_t * ctx,
852 : ulong umem_off,
853 : ulong sz,
854 0 : uint * freed_chunk ) {
855 :
856 0 : uchar * packet = (uchar *)ctx->umem_frame0 + umem_off;
857 0 : uchar const * packet_end = packet + sz;
858 0 : uchar * iphdr = packet + 14U;
859 :
860 : /* Discard the GRE overhead (outer iphdr and gre hdr) */
861 0 : if( iphdr[9] == FD_IP4_HDR_PROTOCOL_GRE ) {
862 :
863 0 : if( FD_UNLIKELY( ctx->has_gre_interface==0 ) ) return; // drop
864 :
865 0 : ulong overhead = FD_IP4_GET_LEN( *(fd_ip4_hdr_t *)iphdr ) + sizeof(fd_gre_hdr_t);
866 :
867 : /* The new iphdr is where the inner iphdr was */
868 0 : iphdr += overhead;
869 :
870 : /* Copy over the eth_hdr */
871 0 : fd_eth_hdr_t * new_packet = (fd_eth_hdr_t *)( (char *)iphdr - sizeof(fd_eth_hdr_t) );
872 0 : fd_memcpy( new_packet, packet, sizeof(fd_eth_hdr_t) );
873 :
874 0 : sz -= overhead;
875 0 : packet = (uchar *)new_packet;
876 0 : umem_off = (ulong)( packet - (uchar *)ctx->umem_frame0 );
877 0 : }
878 :
879 :
880 : /* Translate packet to UMEM frame index */
881 0 : ulong chunk = ctx->umem_chunk0 + (umem_off>>FD_CHUNK_LG_SZ);
882 0 : ulong ctl = umem_off & 0x3fUL;
883 :
884 :
885 : /* Filter for UDP/IPv4 packets. Test for ethtype and ipproto in 1
886 : branch */
887 0 : uint test_ethip = ( (uint)packet[12] << 16u ) | ( (uint)packet[13] << 8u ) | (uint)packet[23];
888 0 : if( FD_UNLIKELY( test_ethip!=0x080011 ) ) {
889 0 : FD_LOG_ERR(( "Firedancer received a packet from the XDP program that was either "
890 0 : "not an IPv4 packet, or not a UDP packet. It is likely your XDP program "
891 0 : "is not configured correctly." ));
892 0 : }
893 :
894 : /* IPv4 is variable-length, so lookup IHL to find start of UDP */
895 0 : uint iplen = FD_IP4_GET_LEN( *(fd_ip4_hdr_t *)iphdr );
896 0 : uchar const * udp = iphdr + iplen;
897 :
898 : /* Ignore if UDP header is too short */
899 0 : if( FD_UNLIKELY( udp+8U > packet_end ) ) {
900 0 : FD_DTRACE_PROBE( net_tile_err_rx_undersz );
901 0 : ctx->metrics.rx_undersz_cnt++;
902 0 : return;
903 0 : }
904 :
905 : /* Extract IP dest addr and UDP src/dest port */
906 0 : uint ip_srcaddr = *(uint *)( iphdr+12UL );
907 0 : ushort udp_srcport = fd_ushort_bswap( *(ushort *)( udp+0UL ) );
908 0 : ushort udp_dstport = fd_ushort_bswap( *(ushort *)( udp+2UL ) );
909 :
910 0 : FD_DTRACE_PROBE_4( net_tile_pkt_rx, ip_srcaddr, udp_srcport, udp_dstport, sz );
911 :
912 : /* Route packet to downstream tile */
913 0 : ushort proto;
914 0 : fd_net_out_ctx_t * out;
915 0 : if( FD_UNLIKELY( udp_dstport==ctx->shred_listen_port ) ) {
916 0 : proto = DST_PROTO_SHRED;
917 0 : out = ctx->shred_out;
918 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->quic_transaction_listen_port ) ) {
919 0 : proto = DST_PROTO_TPU_QUIC;
920 0 : out = ctx->quic_out;
921 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->legacy_transaction_listen_port ) ) {
922 0 : proto = DST_PROTO_TPU_UDP;
923 0 : out = ctx->quic_out;
924 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->gossip_listen_port ) ) {
925 0 : proto = DST_PROTO_GOSSIP;
926 0 : out = ctx->gossip_out;
927 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->repair_intake_listen_port ) ) {
928 0 : proto = DST_PROTO_REPAIR;
929 0 : if( FD_UNLIKELY( sz == REPAIR_PING_SZ ) ) out = ctx->repair_out; /* ping-pong */
930 0 : else out = ctx->shred_out;
931 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->repair_serve_listen_port ) ) {
932 0 : proto = DST_PROTO_REPAIR;
933 0 : out = ctx->repair_out;
934 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->send_src_port ) ) {
935 0 : proto = DST_PROTO_SEND;
936 0 : out = ctx->send_out;
937 0 : } else {
938 :
939 0 : FD_LOG_ERR(( "Firedancer received a UDP packet on port %hu which was not expected. "
940 0 : "Only the following ports should be configured to forward packets: "
941 0 : "%hu, %hu, %hu, %hu, %hu, %hu (excluding any 0 ports, which can be ignored)."
942 0 : "Please report this error to Firedancer maintainers.",
943 0 : udp_dstport,
944 0 : ctx->shred_listen_port,
945 0 : ctx->quic_transaction_listen_port,
946 0 : ctx->legacy_transaction_listen_port,
947 0 : ctx->gossip_listen_port,
948 0 : ctx->repair_intake_listen_port,
949 0 : ctx->repair_serve_listen_port ));
950 0 : }
951 :
952 : /* tile can decide how to partition based on src ip addr and src port */
953 0 : ulong sig = fd_disco_netmux_sig( ip_srcaddr, udp_srcport, 0U, proto, 14UL+8UL+iplen );
954 :
955 : /* Peek the mline for an old frame */
956 0 : fd_frag_meta_t * mline = out->mcache + fd_mcache_line_idx( out->seq, out->depth );
957 0 : *freed_chunk = mline->chunk;
958 :
959 : /* Overwrite the mline with the new frame */
960 0 : ulong tspub = (ulong)fd_frag_meta_ts_comp( fd_tickcount() );
961 0 : fd_mcache_publish( out->mcache, out->depth, out->seq, sig, chunk, sz, ctl, 0, tspub );
962 :
963 : /* Wind up for the next iteration */
964 0 : out->seq = fd_seq_inc( out->seq, 1UL );
965 :
966 0 : ctx->metrics.rx_pkt_cnt++;
967 0 : ctx->metrics.rx_bytes_total += sz;
968 0 : }
969 :
970 : /* net_comp_event is called when an XDP TX frame is free again. */
971 :
972 : static void
973 : net_comp_event( fd_net_ctx_t * ctx,
974 : fd_xsk_t * xsk,
975 0 : uint comp_seq ) {
976 :
977 : /* Locate the incoming frame */
978 :
979 0 : fd_xdp_ring_t * comp_ring = &xsk->ring_cr;
980 0 : uint comp_mask = comp_ring->depth - 1U;
981 0 : ulong frame = FD_VOLATILE_CONST( comp_ring->frame_ring[ comp_seq&comp_mask ] );
982 0 : ulong const frame_mask = FD_NET_MTU - 1UL;
983 0 : if( FD_UNLIKELY( frame+FD_NET_MTU > ctx->umem_sz ) ) {
984 0 : FD_LOG_ERR(( "Bounds check failed: frame=0x%lx umem_sz=0x%lx",
985 0 : frame, (ulong)ctx->umem_sz ));
986 0 : }
987 :
988 : /* Check if we have space to return the freed frame */
989 :
990 0 : fd_net_free_ring_t * free = &ctx->free_tx;
991 0 : ulong free_prod = free->prod;
992 0 : ulong free_mask = free->depth - 1UL;
993 0 : long free_cnt = fd_seq_diff( free_prod, free->cons );
994 0 : if( FD_UNLIKELY( free_cnt>=(long)free->depth ) ) return; /* blocked */
995 :
996 0 : free->queue[ free_prod&free_mask ] = (ulong)ctx->umem_frame0 + (frame & (~frame_mask));
997 0 : free->prod = fd_seq_inc( free_prod, 1UL );
998 :
999 : /* Wind up for next iteration */
1000 :
1001 0 : FD_VOLATILE( *comp_ring->cons ) = comp_ring->cached_cons = comp_seq+1U;
1002 :
1003 0 : ctx->metrics.tx_complete_cnt++;
1004 :
1005 0 : }
1006 :
1007 : /* net_rx_event is called when a new XDP RX frame is available. Calls
1008 : net_rx_packet, then returns the packet back to the kernel via the fill
1009 : ring. */
1010 :
1011 : static void
1012 : net_rx_event( fd_net_ctx_t * ctx,
1013 : fd_xsk_t * xsk,
1014 0 : uint rx_seq ) {
1015 : /* Locate the incoming frame */
1016 :
1017 0 : fd_xdp_ring_t * rx_ring = &xsk->ring_rx;
1018 0 : uint rx_mask = rx_ring->depth - 1U;
1019 0 : struct xdp_desc frame = FD_VOLATILE_CONST( rx_ring->packet_ring[ rx_seq&rx_mask ] );
1020 :
1021 0 : if( FD_UNLIKELY( frame.len>FD_NET_MTU ) )
1022 0 : FD_LOG_ERR(( "received a UDP packet with a too large payload (%u)", frame.len ));
1023 :
1024 : /* Check if we have space in the fill ring to free the frame */
1025 :
1026 0 : fd_xdp_ring_t * fill_ring = &xsk->ring_fr;
1027 0 : uint fill_depth = fill_ring->depth;
1028 0 : uint fill_mask = fill_depth-1U;
1029 0 : ulong frame_mask = FD_NET_MTU - 1UL;
1030 0 : uint fill_prod = FD_VOLATILE_CONST( *fill_ring->prod );
1031 0 : uint fill_cons = FD_VOLATILE_CONST( *fill_ring->cons );
1032 :
1033 0 : if( FD_UNLIKELY( (int)(fill_prod-fill_cons) >= (int)fill_depth ) ) {
1034 0 : ctx->metrics.rx_fill_blocked_cnt++;
1035 0 : return; /* blocked */
1036 0 : }
1037 :
1038 : /* Pass it to the receive handler */
1039 :
1040 0 : uint freed_chunk = UINT_MAX;
1041 0 : net_rx_packet( ctx, frame.addr, frame.len, &freed_chunk );
1042 :
1043 0 : FD_COMPILER_MFENCE();
1044 0 : FD_VOLATILE( *rx_ring->cons ) = rx_ring->cached_cons = rx_seq+1U;
1045 :
1046 : /* If this mcache publish shadowed a previous publish, mark the old
1047 : frame as free. */
1048 :
1049 0 : if( FD_LIKELY( freed_chunk!=UINT_MAX ) ) {
1050 0 : if( FD_UNLIKELY( ( freed_chunk < ctx->umem_chunk0 ) |
1051 0 : ( freed_chunk > ctx->umem_wmark ) ) ) {
1052 0 : FD_LOG_ERR(( "mcache corruption detected: chunk=%u chunk0=%u wmark=%u",
1053 0 : freed_chunk, ctx->umem_chunk0, ctx->umem_wmark ));
1054 0 : }
1055 0 : ulong freed_off = (freed_chunk - ctx->umem_chunk0)<<FD_CHUNK_LG_SZ;
1056 0 : fill_ring->frame_ring[ fill_prod&fill_mask ] = freed_off & (~frame_mask);
1057 0 : FD_VOLATILE( *fill_ring->prod ) = fill_ring->cached_prod = fill_prod+1U;
1058 0 : }
1059 :
1060 0 : }
1061 :
1062 : /* before_credit is called every loop iteration. */
1063 :
1064 : static void
1065 : before_credit( fd_net_ctx_t * ctx,
1066 : fd_stem_context_t * stem,
1067 0 : int * charge_busy ) {
1068 0 : (void)stem;
1069 : /* A previous send attempt was overrun. A corrupt copy of the packet was
1070 : placed into an XDP frame, but the frame was not yet submitted to the
1071 : TX ring. Return the tx buffer to the free list. */
1072 :
1073 0 : if( ctx->tx_op.frame ) {
1074 0 : *charge_busy = 1;
1075 0 : fd_net_free_ring_t * free = &ctx->free_tx;
1076 0 : ulong alloc_seq = free->prod;
1077 0 : free->queue[ alloc_seq % free->depth ] = (ulong)ctx->tx_op.frame;
1078 0 : free->prod = fd_seq_inc( alloc_seq, 1UL );
1079 0 : ctx->tx_op.frame = NULL;
1080 0 : }
1081 :
1082 : /* Check if new packets are available or if TX frames are free again
1083 : (Round-robin through sockets) */
1084 :
1085 0 : uint rr_idx = ctx->rr_idx;
1086 0 : fd_xsk_t * rr_xsk = &ctx->xsk[ rr_idx ];
1087 :
1088 0 : net_tx_periodic_wakeup( ctx, rr_idx, fd_tickcount(), charge_busy );
1089 :
1090 0 : uint rx_cons = rr_xsk->ring_rx.cached_cons;
1091 0 : uint rx_prod = FD_VOLATILE_CONST( *rr_xsk->ring_rx.prod );
1092 0 : if( rx_cons!=rx_prod ) {
1093 0 : *charge_busy = 1;
1094 0 : rr_xsk->ring_rx.cached_prod = rx_prod;
1095 0 : net_rx_event( ctx, rr_xsk, rx_cons );
1096 0 : } else {
1097 0 : net_rx_wakeup( ctx, rr_xsk, charge_busy );
1098 0 : ctx->rr_idx++;
1099 0 : ctx->rr_idx = fd_uint_if( ctx->rr_idx>=ctx->xsk_cnt, 0, ctx->rr_idx );
1100 0 : }
1101 :
1102 0 : uint comp_cons = FD_VOLATILE_CONST( *rr_xsk->ring_cr.cons );
1103 0 : uint comp_prod = FD_VOLATILE_CONST( *rr_xsk->ring_cr.prod );
1104 0 : if( comp_cons!=comp_prod ) {
1105 0 : *charge_busy = 1;
1106 0 : rr_xsk->ring_cr.cached_prod = comp_prod;
1107 0 : net_comp_event( ctx, rr_xsk, comp_cons );
1108 0 : }
1109 :
1110 0 : }
1111 :
1112 : /* net_xsk_bootstrap assigns UMEM frames to the FILL ring. */
1113 :
1114 : static ulong
1115 : net_xsk_bootstrap( fd_net_ctx_t * ctx,
1116 : uint xsk_idx,
1117 0 : ulong frame_off ) {
1118 0 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
1119 :
1120 0 : ulong const frame_sz = FD_NET_MTU;
1121 0 : ulong const fr_depth = ctx->xsk[ xsk_idx ].ring_fr.depth/2UL;
1122 :
1123 0 : fd_xdp_ring_t * fill = &xsk->ring_fr;
1124 0 : uint fill_prod = fill->cached_prod;
1125 0 : for( ulong j=0UL; j<fr_depth; j++ ) {
1126 0 : fill->frame_ring[ j ] = frame_off;
1127 0 : frame_off += frame_sz;
1128 0 : }
1129 0 : FD_VOLATILE( *fill->prod ) = fill->cached_prod = fill_prod + (uint)fr_depth;
1130 :
1131 0 : return frame_off;
1132 0 : }
1133 :
1134 : /* FIXME source MAC address from netlnk tile instead */
1135 :
1136 : static void
1137 : interface_addrs( const char * interface,
1138 : uchar * mac,
1139 0 : uint * ip4_addr ) {
1140 0 : int fd = socket( AF_INET, SOCK_DGRAM, 0 );
1141 0 : struct ifreq ifr;
1142 0 : ifr.ifr_addr.sa_family = AF_INET;
1143 :
1144 0 : strncpy( ifr.ifr_name, interface, IFNAMSIZ );
1145 0 : if( FD_UNLIKELY( ioctl( fd, SIOCGIFHWADDR, &ifr ) ) )
1146 0 : FD_LOG_ERR(( "could not get MAC address of interface `%s`: (%i-%s)", interface, errno, fd_io_strerror( errno ) ));
1147 0 : fd_memcpy( mac, ifr.ifr_hwaddr.sa_data, 6 );
1148 :
1149 0 : if( FD_UNLIKELY( ioctl( fd, SIOCGIFADDR, &ifr ) ) )
1150 0 : FD_LOG_ERR(( "could not get IP address of interface `%s`: (%i-%s)", interface, errno, fd_io_strerror( errno ) ));
1151 0 : *ip4_addr = ((struct sockaddr_in *)fd_type_pun( &ifr.ifr_addr ))->sin_addr.s_addr;
1152 :
1153 0 : if( FD_UNLIKELY( close(fd) ) )
1154 0 : FD_LOG_ERR(( "could not close socket (%i-%s)", errno, fd_io_strerror( errno ) ));
1155 0 : }
1156 :
1157 : /* privileged_init does the following initialization steps:
1158 :
1159 : - Create an AF_XDP socket
1160 : - Map XDP metadata rings
1161 : - Register UMEM data region with socket
1162 : - Insert AF_XDP socket into xsk_map
1163 :
1164 : Net tile 0 also runs fd_xdp_install and repeats the above step for
1165 : the loopback device. (Unless the main interface is already loopback)
1166 :
1167 : Kernel object references:
1168 :
1169 : BPF_LINK file descriptor
1170 : |
1171 : +-> XDP program installation on NIC
1172 : | |
1173 : | +-> XDP program <-- BPF_PROG file descriptor (prog_fd)
1174 : |
1175 : +-> XSKMAP object <-- BPF_MAP file descriptor (xsk_map) */
1176 :
1177 : FD_FN_UNUSED static void
1178 : privileged_init( fd_topo_t * topo,
1179 0 : fd_topo_tile_t * tile ) {
1180 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1181 :
1182 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1183 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
1184 0 : ulong * free_tx = FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );;
1185 :
1186 0 : fd_memset( ctx, 0, sizeof(fd_net_ctx_t) );
1187 :
1188 0 : uint if_idx = if_nametoindex( tile->xdp.interface );
1189 0 : if( FD_UNLIKELY( !if_idx ) ) FD_LOG_ERR(( "if_nametoindex(%s) failed", tile->xdp.interface ));
1190 :
1191 0 : interface_addrs( tile->xdp.interface, ctx->src_mac_addr, &ctx->default_address );
1192 :
1193 : /* Load up dcache containing UMEM */
1194 :
1195 0 : void * const dcache_mem = fd_topo_obj_laddr( topo, tile->net.umem_dcache_obj_id );
1196 0 : void * const umem_dcache = fd_dcache_join( dcache_mem );
1197 0 : ulong const umem_dcache_data_sz = fd_dcache_data_sz( umem_dcache );
1198 0 : ulong const umem_frame_sz = 2048UL;
1199 :
1200 : /* Left shrink UMEM region to be 4096 byte aligned */
1201 :
1202 0 : void * const umem_frame0 = (void *)fd_ulong_align_up( (ulong)umem_dcache, 4096UL );
1203 0 : ulong umem_sz = umem_dcache_data_sz - ((ulong)umem_frame0 - (ulong)umem_dcache);
1204 0 : umem_sz = fd_ulong_align_dn( umem_sz, umem_frame_sz );
1205 :
1206 : /* Derive chunk bounds */
1207 :
1208 0 : void * const umem_base = fd_wksp_containing( dcache_mem );
1209 0 : ulong const umem_chunk0 = ( (ulong)umem_frame0 - (ulong)umem_base )>>FD_CHUNK_LG_SZ;
1210 0 : ulong const umem_wmark = umem_chunk0 + ( ( umem_sz-umem_frame_sz )>>FD_CHUNK_LG_SZ );
1211 :
1212 0 : if( FD_UNLIKELY( umem_chunk0>UINT_MAX || umem_wmark>UINT_MAX || umem_chunk0>umem_wmark ) ) {
1213 0 : FD_LOG_ERR(( "Calculated invalid UMEM bounds [%lu,%lu]", umem_chunk0, umem_wmark ));
1214 0 : }
1215 :
1216 0 : if( FD_UNLIKELY( !umem_base ) ) FD_LOG_ERR(( "UMEM dcache is not in a workspace" ));
1217 0 : if( FD_UNLIKELY( !umem_dcache ) ) FD_LOG_ERR(( "Failed to join UMEM dcache" ));
1218 :
1219 0 : ctx->umem_frame0 = umem_frame0;
1220 0 : ctx->umem_sz = umem_sz;
1221 0 : ctx->umem_chunk0 = (uint)umem_chunk0;
1222 0 : ctx->umem_wmark = (uint)umem_wmark;
1223 :
1224 0 : ctx->free_tx.queue = free_tx;
1225 0 : ctx->free_tx.depth = tile->xdp.xdp_tx_queue_size;
1226 :
1227 : /* Create and install XSKs */
1228 :
1229 0 : fd_xsk_params_t params0 = {
1230 0 : .if_idx = if_idx,
1231 0 : .if_queue_id = (uint)tile->kind_id,
1232 :
1233 : /* Some kernels produce EOPNOTSUP errors on sendto calls when
1234 : starting up without either XDP_ZEROCOPY or XDP_COPY
1235 : (e.g. 5.14.0-503.23.1.el9_5 with i40e) */
1236 0 : .bind_flags = tile->xdp.zero_copy ? XDP_ZEROCOPY : XDP_COPY,
1237 :
1238 0 : .fr_depth = tile->xdp.xdp_rx_queue_size*2,
1239 0 : .rx_depth = tile->xdp.xdp_rx_queue_size,
1240 0 : .cr_depth = tile->xdp.xdp_tx_queue_size,
1241 0 : .tx_depth = tile->xdp.xdp_tx_queue_size,
1242 :
1243 0 : .umem_addr = umem_frame0,
1244 0 : .frame_sz = umem_frame_sz,
1245 0 : .umem_sz = umem_sz
1246 0 : };
1247 :
1248 0 : int xsk_map_fd = 123462;
1249 0 : ctx->prog_link_fds[ 0 ] = 123463;
1250 : /* Init XSK */
1251 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 ));
1252 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 ));
1253 0 : ctx->xsk_cnt = 1;
1254 :
1255 0 : if( FD_UNLIKELY( fd_sandbox_gettid()==fd_sandbox_getpid() ) ) {
1256 : /* Kind of gross.. in single threaded mode we don't want to close the xsk_map_fd
1257 : since it's shared with other net tiles. Just check for that by seeing if we
1258 : are the only thread in the process. */
1259 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 ) ));
1260 0 : }
1261 :
1262 : /* Networking tile at index 0 also binds to loopback (only queue 0 available on lo) */
1263 :
1264 0 : if( FD_UNLIKELY( strcmp( tile->xdp.interface, "lo" ) && !tile->kind_id ) ) {
1265 0 : ctx->xsk_cnt = 2;
1266 :
1267 0 : ushort udp_port_candidates[] = {
1268 0 : (ushort)tile->xdp.net.legacy_transaction_listen_port,
1269 0 : (ushort)tile->xdp.net.quic_transaction_listen_port,
1270 0 : (ushort)tile->xdp.net.shred_listen_port,
1271 0 : (ushort)tile->xdp.net.gossip_listen_port,
1272 0 : (ushort)tile->xdp.net.repair_intake_listen_port,
1273 0 : (ushort)tile->xdp.net.repair_serve_listen_port,
1274 0 : (ushort)tile->xdp.net.send_src_port
1275 0 : };
1276 :
1277 0 : uint lo_idx = if_nametoindex( "lo" );
1278 0 : if( FD_UNLIKELY( !lo_idx ) ) FD_LOG_ERR(( "if_nametoindex(lo) failed" ));
1279 :
1280 : /* FIXME move this to fd_topo_run */
1281 0 : fd_xdp_fds_t lo_fds = fd_xdp_install( lo_idx,
1282 0 : tile->net.bind_address,
1283 0 : sizeof(udp_port_candidates)/sizeof(udp_port_candidates[0]),
1284 0 : udp_port_candidates,
1285 0 : "skb" );
1286 :
1287 0 : ctx->prog_link_fds[ 1 ] = lo_fds.prog_link_fd;
1288 : /* init xsk 1 */
1289 0 : fd_xsk_params_t params1 = params0;
1290 0 : params1.if_idx = lo_idx; /* probably always 1 */
1291 0 : params1.if_queue_id = 0;
1292 0 : params1.bind_flags = 0;
1293 0 : if( FD_UNLIKELY( !fd_xsk_init( &ctx->xsk[ 1 ], ¶ms1 ) ) ) FD_LOG_ERR(( "failed to bind lo_xsk" ));
1294 0 : if( FD_UNLIKELY( !fd_xsk_activate( &ctx->xsk[ 1 ], lo_fds.xsk_map_fd ) ) ) FD_LOG_ERR(( "failed to activate lo_xsk" ));
1295 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 ) ));
1296 0 : }
1297 :
1298 0 : double tick_per_ns = fd_tempo_tick_per_ns( NULL );
1299 0 : ctx->xdp_stats_interval_ticks = (long)( FD_XDP_STATS_INTERVAL_NS * tick_per_ns );
1300 :
1301 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, 1UL );
1302 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
1303 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
1304 0 : }
1305 :
1306 : FD_FN_UNUSED static void
1307 : unprivileged_init( fd_topo_t * topo,
1308 0 : fd_topo_tile_t * tile ) {
1309 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1310 :
1311 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1312 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
1313 0 : FD_TEST( ctx->xsk_cnt!=0 );
1314 0 : FD_TEST( ctx->free_tx.queue!=NULL );
1315 0 : (void)FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );
1316 0 : ctx->netdev_buf = FD_SCRATCH_ALLOC_APPEND( l, fd_netdev_tbl_align(), ctx->netdev_buf_sz );
1317 :
1318 0 : ctx->net_tile_id = (uint)tile->kind_id;
1319 0 : ctx->net_tile_cnt = (uint)fd_topo_tile_name_cnt( topo, tile->name );
1320 :
1321 0 : ctx->bind_address = tile->net.bind_address;
1322 0 : ctx->shred_listen_port = tile->net.shred_listen_port;
1323 0 : ctx->quic_transaction_listen_port = tile->net.quic_transaction_listen_port;
1324 0 : ctx->legacy_transaction_listen_port = tile->net.legacy_transaction_listen_port;
1325 0 : ctx->gossip_listen_port = tile->net.gossip_listen_port;
1326 0 : ctx->repair_intake_listen_port = tile->net.repair_intake_listen_port;
1327 0 : ctx->repair_serve_listen_port = tile->net.repair_serve_listen_port;
1328 0 : ctx->send_src_port = tile->net.send_src_port;
1329 :
1330 : /* Put a bound on chunks we read from the input, to make sure they
1331 : are within in the data region of the workspace. */
1332 :
1333 0 : if( FD_UNLIKELY( !tile->in_cnt ) ) FD_LOG_ERR(( "net tile in link cnt is zero" ));
1334 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 ));
1335 0 : FD_TEST( tile->in_cnt<=32 );
1336 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
1337 0 : fd_topo_link_t * link = &topo->links[ tile->in_link_id[ i ] ];
1338 0 : if( FD_UNLIKELY( link->mtu!=FD_NET_MTU ) ) FD_LOG_ERR(( "net tile in link %s does not have a normal MTU", link->name ));
1339 :
1340 0 : ctx->in[ i ].mem = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
1341 0 : ctx->in[ i ].chunk0 = fd_dcache_compact_chunk0( ctx->in[ i ].mem, link->dcache );
1342 0 : ctx->in[ i ].wmark = fd_dcache_compact_wmark( ctx->in[ i ].mem, link->dcache, link->mtu );
1343 0 : }
1344 :
1345 0 : for( ulong i = 0; i < tile->out_cnt; i++ ) {
1346 0 : fd_topo_link_t * out_link = &topo->links[ tile->out_link_id[ i ] ];
1347 0 : if( strcmp( out_link->name, "net_quic" ) == 0 ) {
1348 0 : fd_topo_link_t * quic_out = out_link;
1349 0 : ctx->quic_out->mcache = quic_out->mcache;
1350 0 : ctx->quic_out->sync = fd_mcache_seq_laddr( ctx->quic_out->mcache );
1351 0 : ctx->quic_out->depth = fd_mcache_depth( ctx->quic_out->mcache );
1352 0 : ctx->quic_out->seq = fd_mcache_seq_query( ctx->quic_out->sync );
1353 0 : } else if( strcmp( out_link->name, "net_shred" ) == 0 ) {
1354 0 : fd_topo_link_t * shred_out = out_link;
1355 0 : ctx->shred_out->mcache = shred_out->mcache;
1356 0 : ctx->shred_out->sync = fd_mcache_seq_laddr( ctx->shred_out->mcache );
1357 0 : ctx->shred_out->depth = fd_mcache_depth( ctx->shred_out->mcache );
1358 0 : ctx->shred_out->seq = fd_mcache_seq_query( ctx->shred_out->sync );
1359 0 : } else if( strcmp( out_link->name, "net_gossip" ) == 0 ) {
1360 0 : fd_topo_link_t * gossip_out = out_link;
1361 0 : ctx->gossip_out->mcache = gossip_out->mcache;
1362 0 : ctx->gossip_out->sync = fd_mcache_seq_laddr( ctx->gossip_out->mcache );
1363 0 : ctx->gossip_out->depth = fd_mcache_depth( ctx->gossip_out->mcache );
1364 0 : ctx->gossip_out->seq = fd_mcache_seq_query( ctx->gossip_out->sync );
1365 0 : } else if( strcmp( out_link->name, "net_repair" ) == 0 ) {
1366 0 : fd_topo_link_t * repair_out = out_link;
1367 0 : ctx->repair_out->mcache = repair_out->mcache;
1368 0 : ctx->repair_out->sync = fd_mcache_seq_laddr( ctx->repair_out->mcache );
1369 0 : ctx->repair_out->depth = fd_mcache_depth( ctx->repair_out->mcache );
1370 0 : ctx->repair_out->seq = fd_mcache_seq_query( ctx->repair_out->sync );
1371 0 : } else if( strcmp( out_link->name, "net_netlnk" ) == 0 ) {
1372 0 : fd_topo_link_t * netlink_out = out_link;
1373 0 : ctx->neigh4_solicit->mcache = netlink_out->mcache;
1374 0 : ctx->neigh4_solicit->depth = fd_mcache_depth( ctx->neigh4_solicit->mcache );
1375 0 : ctx->neigh4_solicit->seq = fd_mcache_seq_query( fd_mcache_seq_laddr( ctx->neigh4_solicit->mcache ) );
1376 0 : } else if( strcmp( out_link->name, "net_send" ) == 0 ) {
1377 0 : fd_topo_link_t * send_out = out_link;
1378 0 : ctx->send_out->mcache = send_out->mcache;
1379 0 : ctx->send_out->sync = fd_mcache_seq_laddr( ctx->send_out->mcache );
1380 0 : ctx->send_out->depth = fd_mcache_depth( ctx->send_out->mcache );
1381 0 : ctx->send_out->seq = fd_mcache_seq_query( ctx->send_out->sync );
1382 0 : } else {
1383 0 : FD_LOG_ERR(( "unrecognized out link `%s`", out_link->name ));
1384 0 : }
1385 0 : }
1386 :
1387 : /* Check if any of the tiles we set a listen port for do not have an outlink. */
1388 0 : if( FD_UNLIKELY( ctx->shred_listen_port!=0 && ctx->shred_out->mcache==NULL ) ) {
1389 0 : FD_LOG_ERR(( "shred listen port set but no out link was found" ));
1390 0 : } else if( FD_UNLIKELY( ctx->quic_transaction_listen_port!=0 && ctx->quic_out->mcache==NULL ) ) {
1391 0 : FD_LOG_ERR(( "quic transaction listen port set but no out link was found" ));
1392 0 : } else if( FD_UNLIKELY( ctx->legacy_transaction_listen_port!=0 && ctx->quic_out->mcache==NULL ) ) {
1393 0 : FD_LOG_ERR(( "legacy transaction listen port set but no out link was found" ));
1394 0 : } else if( FD_UNLIKELY( ctx->gossip_listen_port!=0 && ctx->gossip_out->mcache==NULL ) ) {
1395 0 : FD_LOG_ERR(( "gossip listen port set but no out link was found" ));
1396 0 : } else if( FD_UNLIKELY( ctx->repair_intake_listen_port!=0 && ctx->repair_out->mcache==NULL ) ) {
1397 0 : FD_LOG_ERR(( "repair intake port set but no out link was found" ));
1398 0 : } else if( FD_UNLIKELY( ctx->repair_serve_listen_port!=0 && ctx->repair_out->mcache==NULL ) ) {
1399 0 : FD_LOG_ERR(( "repair serve listen port set but no out link was found" ));
1400 0 : } else if( FD_UNLIKELY( ctx->neigh4_solicit->mcache==NULL ) ) {
1401 0 : FD_LOG_ERR(( "netlink request link not found" ));
1402 0 : } else if( FD_UNLIKELY( ctx->send_src_port!=0 && ctx->send_out->mcache==NULL ) ) {
1403 0 : FD_LOG_ERR(( "send listen port set but no out link was found" ));
1404 0 : }
1405 :
1406 0 : for( uint j=0U; j<2U; j++ ) {
1407 0 : ctx->tx_flusher[ j ].pending_wmark = (ulong)( (double)tile->xdp.xdp_tx_queue_size * 0.7 );
1408 0 : ctx->tx_flusher[ j ].tail_flush_backoff = (long)( (double)tile->xdp.tx_flush_timeout_ns * fd_tempo_tick_per_ns( NULL ) );
1409 0 : ctx->tx_flusher[ j ].next_tail_flush_ticks = LONG_MAX;
1410 0 : }
1411 :
1412 : /* Join netbase objects */
1413 0 : ctx->fib_local = fd_fib4_join( fd_topo_obj_laddr( topo, tile->xdp.fib4_local_obj_id ) );
1414 0 : ctx->fib_main = fd_fib4_join( fd_topo_obj_laddr( topo, tile->xdp.fib4_main_obj_id ) );
1415 0 : if( FD_UNLIKELY( !ctx->fib_local || !ctx->fib_main ) ) FD_LOG_ERR(( "fd_fib4_join failed" ));
1416 0 : if( FD_UNLIKELY( !fd_neigh4_hmap_join(
1417 0 : ctx->neigh4,
1418 0 : fd_topo_obj_laddr( topo, tile->xdp.neigh4_obj_id ),
1419 0 : fd_topo_obj_laddr( topo, tile->xdp.neigh4_ele_obj_id ) ) ) ) {
1420 0 : FD_LOG_ERR(( "fd_neigh4_hmap_join failed" ));
1421 0 : }
1422 :
1423 : /* Allocate and join netdev buffer */
1424 0 : ctx->netdev_dbl_handle = fd_dbl_buf_join( fd_topo_obj_laddr( topo, tile->xdp.netdev_dbl_buf_obj_id ) );
1425 0 : if( FD_UNLIKELY( ctx->netdev_dbl_handle==NULL ) ) FD_LOG_ERR(( "fd_dbl_buf_join failed" )) ;
1426 0 : ctx->netdev_buf_sz = fd_netdev_tbl_footprint( NETDEV_MAX, BOND_MASTER_MAX );
1427 :
1428 : /* Load the netdev table for the first time */
1429 0 : net_load_netdev_tbl( ctx );
1430 :
1431 : /* Loop through netdev table to find if GRE interface exists */
1432 0 : ctx->has_gre_interface = net_check_gre_interface_exists( ctx );
1433 :
1434 : /* Initialize TX free ring */
1435 :
1436 0 : ulong const frame_sz = 2048UL;
1437 0 : ulong frame_off = 0UL;
1438 0 : ulong const tx_depth = ctx->free_tx.depth;
1439 0 : for( ulong j=0; j<tx_depth; j++ ) {
1440 0 : ctx->free_tx.queue[ j ] = (ulong)ctx->umem_frame0 + frame_off;
1441 0 : frame_off += frame_sz;
1442 0 : }
1443 0 : ctx->free_tx.prod = tx_depth;
1444 :
1445 : /* Initialize RX mcache chunks */
1446 :
1447 0 : for( ulong i=0UL; i<(tile->out_cnt); i++ ) {
1448 0 : fd_topo_link_t * out_link = &topo->links[ tile->out_link_id[ i ] ];
1449 0 : fd_frag_meta_t * mcache = out_link->mcache;
1450 0 : for( ulong j=0UL; j<fd_mcache_depth( mcache ); j++ ) {
1451 0 : mcache[ j ].chunk = (uint)( ctx->umem_chunk0 + (frame_off>>FD_CHUNK_LG_SZ) );
1452 0 : frame_off += frame_sz;
1453 0 : }
1454 0 : }
1455 :
1456 : /* Initialize FILL ring */
1457 :
1458 0 : int _charge_busy = 0;
1459 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
1460 0 : frame_off = net_xsk_bootstrap( ctx, j, frame_off );
1461 0 : net_rx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
1462 0 : net_tx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
1463 0 : }
1464 :
1465 0 : if( FD_UNLIKELY( frame_off > ctx->umem_sz ) ) {
1466 0 : FD_LOG_ERR(( "UMEM is too small" ));
1467 0 : }
1468 0 : }
1469 :
1470 : FD_FN_UNUSED static ulong
1471 : populate_allowed_seccomp( fd_topo_t const * topo,
1472 : fd_topo_tile_t const * tile,
1473 : ulong out_cnt,
1474 0 : struct sock_filter * out ) {
1475 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1476 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1477 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_net_ctx_t ), sizeof( fd_net_ctx_t ) );
1478 :
1479 : /* A bit of a hack, if there is no loopback XSK for this tile, we still need to pass
1480 : two "allow" FD arguments to the net policy, so we just make them both the same. */
1481 0 : int allow_fd2 = ctx->xsk_cnt>1UL ? ctx->xsk[ 1 ].xsk_fd : ctx->xsk[ 0 ].xsk_fd;
1482 0 : FD_TEST( ctx->xsk[ 0 ].xsk_fd >= 0 && allow_fd2 >= 0 );
1483 0 : populate_sock_filter_policy_xdp( out_cnt, out, (uint)fd_log_private_logfile_fd(), (uint)ctx->xsk[ 0 ].xsk_fd, (uint)allow_fd2 );
1484 0 : return sock_filter_policy_xdp_instr_cnt;
1485 0 : }
1486 :
1487 : FD_FN_UNUSED static ulong
1488 : populate_allowed_fds( fd_topo_t const * topo,
1489 : fd_topo_tile_t const * tile,
1490 : ulong out_fds_cnt,
1491 0 : int * out_fds ) {
1492 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1493 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1494 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_net_ctx_t ), sizeof( fd_net_ctx_t ) );
1495 :
1496 0 : if( FD_UNLIKELY( out_fds_cnt<6UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
1497 :
1498 0 : ulong out_cnt = 0UL;
1499 :
1500 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
1501 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
1502 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
1503 :
1504 0 : out_fds[ out_cnt++ ] = ctx->xsk[ 0 ].xsk_fd;
1505 0 : out_fds[ out_cnt++ ] = ctx->prog_link_fds[ 0 ];
1506 0 : if( FD_LIKELY( ctx->xsk_cnt>1UL ) ) out_fds[ out_cnt++ ] = ctx->xsk[ 1 ].xsk_fd;
1507 0 : if( FD_LIKELY( ctx->xsk_cnt>1UL ) ) out_fds[ out_cnt++ ] = ctx->prog_link_fds[ 1 ];
1508 0 : return out_cnt;
1509 0 : }
1510 :
1511 0 : #define STEM_BURST (1UL)
1512 0 : #define STEM_LAZY ((ulong)30e3) /* 30 us */
1513 :
1514 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_net_ctx_t
1515 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_net_ctx_t)
1516 :
1517 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
1518 0 : #define STEM_CALLBACK_DURING_HOUSEKEEPING during_housekeeping
1519 0 : #define STEM_CALLBACK_BEFORE_CREDIT before_credit
1520 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
1521 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
1522 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
1523 :
1524 : #include "../../stem/fd_stem.c"
1525 :
1526 : #ifndef FD_TILE_TEST
1527 : fd_topo_run_tile_t fd_tile_net = {
1528 : .name = "net",
1529 : .populate_allowed_seccomp = populate_allowed_seccomp,
1530 : .populate_allowed_fds = populate_allowed_fds,
1531 : .scratch_align = scratch_align,
1532 : .scratch_footprint = scratch_footprint,
1533 : .privileged_init = privileged_init,
1534 : .unprivileged_init = unprivileged_init,
1535 : .run = stem_run,
1536 : };
1537 : #endif
|
