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 "../fd_net_tile.h"
6 :
7 : #include <errno.h>
8 : #include <fcntl.h>
9 : #include <net/if.h>
10 : #include <netinet/in.h>
11 : #include <sys/socket.h> /* MSG_DONTWAIT needed before importing the net seccomp filter */
12 : #include <linux/if_xdp.h>
13 :
14 : #include "../fd_net_common.h"
15 : #include "../../metrics/fd_metrics.h"
16 : #include "../../netlink/fd_netlink_tile.h" /* neigh4_solicit */
17 : #include "../../topo/fd_topo.h"
18 :
19 : #include "../../../waltz/ip/fd_fib4.h"
20 : #include "../../../waltz/neigh/fd_neigh4_map.h"
21 : #include "../../../waltz/mib/fd_netdev_tbl.h"
22 : #include "../../../waltz/xdp/fd_xdp_redirect_user.h" /* fd_xsk_activate */
23 : #include "../../../waltz/xdp/fd_xsk.h"
24 : #include "../../../util/log/fd_dtrace.h"
25 : #include "../../../util/net/fd_eth.h"
26 : #include "../../../util/net/fd_ip4.h"
27 : #include "../../../util/net/fd_gre.h"
28 : #include "../../../util/pod/fd_pod_format.h"
29 :
30 : #include <unistd.h>
31 : #include <linux/if.h> /* struct ifreq */
32 : #include <sys/ioctl.h>
33 : #include <linux/if_arp.h>
34 :
35 : #include "generated/fd_xdp_tile_seccomp.h"
36 :
37 : /* MAX_NET_INS controls the max number of TX links that a net tile can
38 : serve. */
39 :
40 : #define MAX_NET_INS (32UL)
41 :
42 : /* FD_XDP_STATS_INTERVAL_NS controls the XDP stats refresh interval.
43 : This should be lower than the interval at which the metrics tile
44 : collects metrics. */
45 :
46 0 : #define FD_XDP_STATS_INTERVAL_NS (11e6) /* 11ms */
47 :
48 : /* XSK_IDX_{MAIN,LO} are the hardcoded XSK indices in ctx->xsk[ ... ].
49 : Only net tile 0 has XSK_IDX_LO, all net tiles have XSK_IDX_MAIN. */
50 :
51 30 : #define XSK_IDX_MAIN 0
52 18 : #define XSK_IDX_LO 1
53 :
54 : /* XSK 'busy_poll_usecs' value (max amount of time
55 : spent spinning in a NAPI poll before returning back to
56 : userspace if the processing budget hasn't already ran out).
57 :
58 : 64us chosen based on napibusy configuration tested and
59 : shown in https://lwn.net/Articles/997491/ Linux patch
60 : cover letter. Chosen over fullbusy since for Firedancer
61 : the values of fullbusy are unnecessarily high and could
62 : cause some extra latency to regular non-Firedancer traffic.*/
63 0 : #define PREFBUSY_TIME_BUDGET_MICROS (64L)
64 :
65 : /* PREFBUSY_RX_BUDGET is the NAPI RX processing budget (max num RX
66 : packets that the NIC driver can move from the hw rings into the
67 : XSK rings per poll).
68 :
69 : Default RX budget used by NIC drivers is 64, therefore
70 : it is safest to use 64 in prefbusy polling. Also reduces
71 : TX starvation risks as the TX budget set by the NIC driver
72 : is also generally 64. */
73 0 : #define PREFBUSY_RX_BUDGET (64L)
74 :
75 : /* Min time between each prefbusy poll. Necessary to avoid a no RX
76 : scenario livelocking TX with overly frequent sendto calls, given
77 : prefbusy polls whenever the RX queue is empty.
78 :
79 : Value chosen based on experimentation on ixgbe, mlx5 and i40e as well
80 : as on varying CPUs and clock speeds. Too low -> lower max TX
81 : throughput when RX is very low. Too high -> lower max RX and TX throughput. */
82 0 : #define PREFBUSY_MIN_INTERVAL_NS (5e3) /* 5us */
83 :
84 : /* Max time since last prefbusy poll before a prefbusy poll is
85 : forced (has been read that polling can sometimes resolve a stall).
86 :
87 : Exact value again not particularly important as this is just extra
88 : protection against stalls which have not been observed in testing but
89 : are still good to protect against since there is no cost to doing so.
90 :
91 : Value of 150us chosen since it is easily large enough to not interfere
92 : with standard prefbusy runtime unless there is a serious problem. */
93 0 : #define PREFBUSY_STALL_TIMEOUT_NS (150e3) /* 150us */
94 :
95 : /* MAX_GRE_CNT is the maximum number of GRE tunnels the XDP tile will
96 : monitor. If a packet comes in with a source IP that doesn't match
97 : the endpoint of one of the first MAX_GRE_CNT tunnels (in the order
98 : the OS enumerates them), it will be dropped. This is limited for
99 : performance reasons. */
100 111 : #define MAX_GRE_CNT 4UL
101 :
102 : /* fd_net_in_ctx_t contains consumer information for an incoming tango
103 : link. It is used as part of the TX path. */
104 :
105 : typedef struct {
106 : fd_wksp_t * mem;
107 : ulong chunk0;
108 : ulong wmark;
109 : } fd_net_in_ctx_t;
110 :
111 : /* fd_net_out_ctx_t contains publisher information for a link to a
112 : downstream app tile. It is used as part of the RX path. */
113 :
114 : typedef struct {
115 : fd_frag_meta_t * mcache;
116 : ulong * sync;
117 : ulong depth;
118 : ulong seq;
119 : } fd_net_out_ctx_t;
120 :
121 : /* fd_net_flusher_t controls the pacing of XDP sendto calls for flushing
122 : TX batches. In the 'wakeup' XDP mode, no TX occurs unless the net
123 : tile wakes up the kernel periodically using the sendto() syscall.
124 : If sendto() is called too frequently, time is wasted on context
125 : switches. If sendto() is called not often enough, packets are
126 : delayed or dropped. sendto() calls make almost no guarantees how
127 : much packets are sent out, nor do they indicate when the kernel
128 : finishes a wakeup call (asynchronously dispatched). The net tile
129 : thus uses a myraid of flush triggers that were tested for best
130 : performance. */
131 :
132 : struct fd_net_flusher {
133 :
134 : /* Packets that were enqueued after the last sendto() wakeup are
135 : considered "pending". If there are more than pending_wmark packets
136 : pending, a wakeup is dispatched. Thus, this dispatch trigger is
137 : proportional to packet rate, but does not trigger if I/O is seldom. */
138 : ulong pending_cnt;
139 : ulong pending_wmark;
140 :
141 : /* Sometimes, packets are not flushed out even after a sendto()
142 : wakeup. This can result in the tail of a burst getting delayed or
143 : overrun. If more than tail_flush_backoff ticks pass since the last
144 : sendto() wakeup and there are still unacknowledged packets in the
145 : TX ring, issues another wakeup. Only used by "softirq" poll mode. */
146 : long next_tail_flush_ticks;
147 : long tail_flush_backoff;
148 :
149 : /* When the most recent prefbusy poll was. */
150 : long prefbusy_last_poll_ticks;
151 : /* Min time between each prefbusy poll. */
152 : long prefbusy_min_interval_ticks;
153 : /* Max time since last prefbusy poll before a prefbusy poll is
154 : forced (has been read that polling can sometimes resolve a stall). */
155 : long prefbusy_stall_timeout_ticks;
156 : };
157 :
158 : typedef struct fd_net_flusher fd_net_flusher_t;
159 :
160 : FD_PROTOTYPES_BEGIN
161 :
162 : /* fd_net_flusher_inc marks a new packet as enqueued. */
163 :
164 : static inline void
165 : fd_net_flusher_inc( fd_net_flusher_t * flusher,
166 18 : long now ) {
167 18 : flusher->pending_cnt++;
168 18 : long next_flush = now + flusher->tail_flush_backoff;
169 18 : flusher->next_tail_flush_ticks = fd_long_min( flusher->next_tail_flush_ticks, next_flush );
170 18 : }
171 :
172 : /* fd_net_flusher_check returns 1 if a sendto() wakeup should be issued
173 : immediately. now is a recent fd_tickcount() value.
174 : If tx_ring_empty==0 then the kernel is caught up with the net tile
175 : on the XDP TX ring. (Otherwise, the kernel is behind the net tile) */
176 :
177 : static inline int
178 : fd_net_flusher_check( fd_net_flusher_t * flusher,
179 : long now,
180 27 : int tx_ring_empty ) {
181 27 : int flush_level = flusher->pending_cnt >= flusher->pending_wmark;
182 27 : int flush_timeout = now >= flusher->next_tail_flush_ticks;
183 27 : int flush = flush_level || flush_timeout;
184 27 : if( !flush ) return 0;
185 27 : if( FD_UNLIKELY( tx_ring_empty ) ) {
186 : /* Flush requested but caught up */
187 3 : flusher->pending_cnt = 0UL;
188 3 : flusher->next_tail_flush_ticks = LONG_MAX;
189 3 : return 0;
190 3 : }
191 24 : return 1;
192 27 : }
193 :
194 : /* fd_net_flusher_wakeup signals a sendto() wakeup was done. now is a
195 : recent fd_tickcount() value. */
196 :
197 : static inline void
198 : fd_net_flusher_wakeup( fd_net_flusher_t * flusher,
199 24 : long now ) {
200 24 : flusher->pending_cnt = 0UL;
201 24 : flusher->next_tail_flush_ticks = now + flusher->tail_flush_backoff;
202 24 : }
203 :
204 : FD_PROTOTYPES_END
205 :
206 : /* fd_net_free_ring is a FIFO queue that stores pointers to free XDP TX
207 : frames. */
208 :
209 : struct fd_net_free_ring {
210 : ulong prod;
211 : ulong cons;
212 : ulong depth;
213 : ulong * queue;
214 : };
215 : typedef struct fd_net_free_ring fd_net_free_ring_t;
216 :
217 : typedef struct {
218 : /* An "XSK" is an AF_XDP socket */
219 : uint xsk_cnt;
220 : fd_xsk_t xsk[ 2 ];
221 : int prog_link_fds[ 2 ];
222 : uint if_virt;
223 :
224 : /* UMEM frame region within dcache */
225 : void * umem; /* Start of UMEM */
226 : ulong umem_sz; /* Size of UMEM */
227 :
228 : /* UMEM chunk region within workspace */
229 : uint umem_chunk0; /* Lowest allowed chunk number */
230 : uint umem_wmark; /* Highest allowed chunk number */
231 :
232 : /* All net tiles are subscribed to the same TX links. (These are
233 : incoming links from app tiles asking the net tile to send out packets)
234 : The net tiles "take turns" doing TX jobs based on the L3+L4 dst hash.
235 : net_tile_id is the index of the current interface, net_tile_cnt is the
236 : total amount of interfaces. */
237 : uint net_tile_id;
238 : uint net_tile_cnt;
239 :
240 : /* Details pertaining to an inflight send op */
241 : struct {
242 : uint xsk_idx;
243 : void * frame;
244 : uchar mac_addrs[12]; /* First 12 bytes of Ethernet header */
245 : uint src_ip; /* src_ip in net order */
246 :
247 : uint use_gre; /* The tx packet will be GRE-encapsulated */
248 : uint gre_outer_src_ip; /* For GRE: Outer iphdr's src_ip in net order */
249 : uint gre_outer_dst_ip; /* For GRE: Outer iphdr's dst_ip in net order */
250 : } tx_op;
251 :
252 : /* Round-robin cycle serivce operations */
253 : uint rr_idx;
254 :
255 : /* Ring tracking free packet buffers */
256 : fd_net_free_ring_t free_tx;
257 :
258 : uchar src_mac_addr[6];
259 : uint default_address;
260 :
261 : uint bind_address;
262 : ushort shred_listen_port;
263 : ushort quic_transaction_listen_port;
264 : ushort legacy_transaction_listen_port;
265 : ushort gossip_listen_port;
266 : ushort repair_client_listen_port;
267 : ushort repair_serve_listen_port;
268 : ushort txsend_src_port;
269 :
270 : ulong in_cnt;
271 : fd_net_in_ctx_t in[ MAX_NET_INS ];
272 :
273 : fd_net_out_ctx_t quic_out[1];
274 : fd_net_out_ctx_t shred_out[1];
275 : fd_net_out_ctx_t gossvf_out[1];
276 : fd_net_out_ctx_t repair_out[1];
277 : fd_net_out_ctx_t txsend_out[1];
278 :
279 : fd_net_out_ctx_t rserve_out[1];
280 : int rserve_enabled;
281 :
282 : /* XDP stats refresh timer */
283 : long xdp_stats_interval_ticks;
284 : long next_xdp_stats_refresh;
285 :
286 : /* TX flush timers */
287 : fd_net_flusher_t tx_flusher[2]; /* one per XSK */
288 :
289 : /* Route and neighbor tables */
290 : fd_fib4_t fib_local[1];
291 : fd_fib4_t fib_main[1];
292 : fd_neigh4_hmap_t neigh4[1];
293 : fd_netlink_neigh4_solicit_link_t neigh4_solicit[1];
294 :
295 : /* Netdev table */
296 : fd_netdev_tbl_join_t netdev_tbl; /* local copy in scratch (hot path) */
297 : fd_netdev_tbl_join_t netdev_shared; /* shared table in netbase (seqlock protected) */
298 : uint gre_tunnel_ip[MAX_GRE_CNT]; /* 0 means unused */
299 :
300 : struct {
301 : ulong rx_pkt_cnt;
302 : ulong rx_bytes_total;
303 : ulong rx_src_addr_invalid_cnt;
304 : ulong rx_undersz_cnt;
305 : ulong rx_fill_blocked_cnt;
306 : ulong rx_backp_cnt;
307 : long rx_busy_cnt;
308 : long rx_idle_cnt;
309 :
310 : ulong tx_submit_cnt;
311 : ulong tx_complete_cnt;
312 : ulong tx_bytes_total;
313 : ulong tx_route_fail_cnt;
314 : ulong tx_no_xdp_cnt;
315 : ulong tx_neigh_fail_cnt;
316 : ulong tx_full_fail_cnt;
317 : long tx_busy_cnt;
318 : long tx_idle_cnt;
319 :
320 : ulong xsk_tx_wakeup_cnt;
321 : ulong xsk_rx_wakeup_cnt;
322 :
323 : ulong rx_gre_cnt;
324 : ulong rx_gre_ignored_cnt;
325 : ulong rx_gre_inv_pkt_cnt;
326 : ulong tx_gre_cnt;
327 : ulong tx_gre_route_fail_cnt;
328 : } metrics;
329 : } fd_net_ctx_t;
330 :
331 : FD_FN_CONST static inline ulong
332 9 : scratch_align( void ) {
333 9 : return 4096UL;
334 9 : }
335 :
336 : FD_FN_PURE static inline ulong
337 3 : scratch_footprint( fd_topo_tile_t const * tile ) {
338 3 : ulong l = FD_LAYOUT_INIT;
339 3 : l = FD_LAYOUT_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
340 3 : l = FD_LAYOUT_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );
341 3 : l = FD_LAYOUT_APPEND( l, fd_netdev_tbl_align(), fd_netdev_tbl_footprint( NETDEV_MAX, BOND_MASTER_MAX ) );
342 3 : return FD_LAYOUT_FINI( l, scratch_align() );
343 3 : }
344 :
345 : static void
346 0 : metrics_write( fd_net_ctx_t * ctx ) {
347 0 : FD_MCNT_SET( NET, PKT_RX, ctx->metrics.rx_pkt_cnt );
348 0 : FD_MCNT_SET( NET, PKT_RX_BYTES, ctx->metrics.rx_bytes_total );
349 0 : FD_MCNT_SET( NET, PKT_RX_UNDERSIZE, ctx->metrics.rx_undersz_cnt );
350 0 : FD_MCNT_SET( NET, PKT_RX_FILL_RING_FULL, ctx->metrics.rx_fill_blocked_cnt );
351 0 : FD_MCNT_SET( NET, PKT_RX_BACKPRESSURE, ctx->metrics.rx_backp_cnt );
352 0 : FD_MGAUGE_SET( NET, RX_BUFFER_BUSY, (ulong)fd_long_max( ctx->metrics.rx_busy_cnt, 0L ) );
353 0 : FD_MGAUGE_SET( NET, RX_BUFFER_IDLE, (ulong)fd_long_max( ctx->metrics.rx_idle_cnt, 0L ) );
354 0 : FD_MGAUGE_SET( NET, TX_BUFFER_BUSY, (ulong)fd_long_max( ctx->metrics.tx_busy_cnt, 0L ) );
355 0 : FD_MGAUGE_SET( NET, TX_BUFFER_IDLE, (ulong)fd_long_max( ctx->metrics.tx_idle_cnt, 0L ) );
356 :
357 0 : FD_MCNT_SET( NET, PKT_TX_SUBMITTED, ctx->metrics.tx_submit_cnt );
358 0 : FD_MCNT_SET( NET, PKT_TX_COMPLETED, ctx->metrics.tx_complete_cnt );
359 0 : FD_MCNT_SET( NET, PKT_TX_BYTES, ctx->metrics.tx_bytes_total );
360 0 : FD_MCNT_SET( NET, PKT_TX_NO_ROUTE, ctx->metrics.tx_route_fail_cnt );
361 0 : FD_MCNT_SET( NET, PKT_TX_NO_NEIGHBOR, ctx->metrics.tx_neigh_fail_cnt );
362 0 : FD_MCNT_SET( NET, PKT_TX_RING_FULL, ctx->metrics.tx_full_fail_cnt );
363 :
364 0 : FD_MCNT_SET( NET, XSK_SYSCALL_TX, ctx->metrics.xsk_tx_wakeup_cnt );
365 0 : FD_MCNT_SET( NET, XSK_SYSCALL_RX, ctx->metrics.xsk_rx_wakeup_cnt );
366 :
367 0 : FD_MCNT_SET( NET, GRE_PKT_RX, ctx->metrics.rx_gre_cnt );
368 0 : FD_MCNT_SET( NET, GRE_PKT_RX_INVALID, ctx->metrics.rx_gre_inv_pkt_cnt );
369 0 : FD_MCNT_SET( NET, GRE_PKT_RX_IGNORED, ctx->metrics.rx_gre_ignored_cnt );
370 0 : FD_MCNT_SET( NET, GRE_PKT_TX_SUBMITTED, ctx->metrics.tx_gre_cnt );
371 0 : FD_MCNT_SET( NET, GRE_PKT_TX_NO_ROUTE, ctx->metrics.tx_gre_route_fail_cnt );
372 0 : FD_MCNT_SET( NET, PKT_RX_SRC_INVALID, ctx->metrics.rx_src_addr_invalid_cnt );
373 0 : }
374 :
375 : struct xdp_statistics_v0 {
376 : __u64 rx_dropped; /* Dropped for other reasons */
377 : __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
378 : __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
379 : };
380 :
381 : struct xdp_statistics_v1 {
382 : __u64 rx_dropped; /* Dropped for other reasons */
383 : __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
384 : __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
385 : __u64 rx_ring_full; /* Dropped due to rx ring being full */
386 : __u64 rx_fill_ring_empty_descs; /* Failed to retrieve item from fill ring */
387 : __u64 tx_ring_empty_descs; /* Failed to retrieve item from tx ring */
388 : };
389 :
390 : static void
391 0 : poll_xdp_statistics( fd_net_ctx_t * ctx ) {
392 0 : struct xdp_statistics_v1 stats = {0};
393 0 : ulong xsk_cnt = ctx->xsk_cnt;
394 0 : for( ulong j=0UL; j<xsk_cnt; j++ ) {
395 0 : struct xdp_statistics_v1 sub_stats = {0};
396 0 : uint optlen = (uint)sizeof(struct xdp_statistics_v1);
397 0 : if( FD_UNLIKELY( -1==getsockopt( ctx->xsk[ j ].xsk_fd, SOL_XDP, XDP_STATISTICS, &sub_stats, &optlen ) ) )
398 0 : FD_LOG_ERR(( "getsockopt(SOL_XDP, XDP_STATISTICS) failed: %s", strerror( errno ) ));
399 0 : if( FD_UNLIKELY( optlen!=sizeof(struct xdp_statistics_v0) &&
400 0 : optlen!=sizeof(struct xdp_statistics_v1) ) ) {
401 0 : FD_LOG_ERR(( "getsockopt(SOL_XDP, XDP_STATISTICS) returned unexpected size %u", optlen ));
402 0 : }
403 0 : stats.rx_dropped += sub_stats.rx_dropped;
404 0 : stats.rx_invalid_descs += sub_stats.rx_invalid_descs;
405 0 : stats.tx_invalid_descs += sub_stats.tx_invalid_descs;
406 0 : stats.rx_ring_full += sub_stats.rx_ring_full;
407 0 : stats.rx_fill_ring_empty_descs += sub_stats.rx_fill_ring_empty_descs;
408 0 : stats.tx_ring_empty_descs += sub_stats.tx_ring_empty_descs;
409 0 : }
410 :
411 0 : FD_MCNT_SET( NET, XDP_RX_OTHER_DROPPED, stats.rx_dropped );
412 0 : FD_MCNT_SET( NET, XDP_RX_INVALID_DESCRIPTOR, stats.rx_invalid_descs );
413 0 : FD_MCNT_SET( NET, XDP_TX_INVALID_DESCRIPTOR, stats.tx_invalid_descs );
414 0 : FD_MCNT_SET( NET, XDP_RX_RING_FULL, stats.rx_ring_full );
415 0 : FD_MCNT_SET( NET, XDP_RX_FILL_RING_EMPTY, stats.rx_fill_ring_empty_descs );
416 0 : FD_MCNT_SET( NET, XDP_TX_RING_EMPTY, stats.tx_ring_empty_descs );
417 0 : }
418 :
419 : /* net_is_fatal_xdp_error returns 1 if the given errno returned by an
420 : XDP API indicates a non-recoverable error code. The net tile should
421 : crash if it sees such an error so the problem does not go undetected.
422 : Otherwise, returns 0. */
423 :
424 : static int
425 0 : net_is_fatal_xdp_error( int err ) {
426 0 : return err==ESOCKTNOSUPPORT || err==EOPNOTSUPP || err==EINVAL ||
427 0 : err==EPERM;
428 0 : }
429 :
430 : /* net_gre_tunnel_ip fills ctx->gre_tunnel_ip. The first gre_tunnel_cnt
431 : entries will be populated with the IP address of the GRE tunnel peer
432 : for the first gre_tunnel_cnt untagged GRE tunnels, and the rest of
433 : the entries will be set to 0, where gre_tunnel_cnt = min(MAX_GRE_CNT,
434 : the number of untagged GRE tunnels). Returns gre_tunnel_cnt. */
435 :
436 : static ulong
437 6 : net_gre_tunnel_ip( fd_net_ctx_t * ctx ) {
438 6 : fd_netdev_t * dev_tbl = ctx->netdev_tbl.dev_tbl;
439 6 : ushort dev_cnt = ctx->netdev_tbl.hdr->dev_cnt;
440 :
441 6 : ulong gre_tunnel_cnt = 0UL;
442 6 : memset( ctx->gre_tunnel_ip, '\0', MAX_GRE_CNT*sizeof(uint) );
443 21 : for( ushort if_idx = 0; (if_idx<dev_cnt) & (gre_tunnel_cnt<MAX_GRE_CNT); if_idx++ ) {
444 15 : fd_netdev_t const * dev = dev_tbl+if_idx;
445 15 : if( dev->dev_type==ARPHRD_IPGRE && dev->gre_dst_ip ) ctx->gre_tunnel_ip[ gre_tunnel_cnt++ ] = dev->gre_dst_ip;
446 15 : }
447 6 : return gre_tunnel_cnt;
448 6 : }
449 :
450 :
451 : /* net_tx_ready returns 1 if we can submit a job to this TX ring, and 0 otherwise.
452 : Reasons for block include:
453 : - No TX buffer is available (free ring empty)
454 : - TX ring is full
455 :
456 : tx_ring: pointer to the XDP TX ring
457 : free_ring: pointer to the free TX ring */
458 :
459 : static int
460 : net_tx_ready( fd_xdp_ring_t * tx_ring,
461 42 : fd_net_free_ring_t * free_ring ) {
462 42 : if( FD_UNLIKELY( free_ring->prod == free_ring->cons ) ) return 0; /* drop - no free buffers */
463 36 : if( FD_UNLIKELY( fd_xdp_ring_full( tx_ring ) ) ) return 0; /* drop - tx ring full */
464 33 : return 1;
465 36 : }
466 :
467 : /* net_rx_wakeup triggers xsk_recvmsg to run in the kernel. Needs to be
468 : called periodically in order to receive packets. */
469 :
470 : static void
471 : net_rx_wakeup( fd_net_ctx_t * ctx,
472 : fd_xsk_t * xsk,
473 0 : int * charge_busy ) {
474 0 : FD_VOLATILE( *xsk->ring_rx.cons ) = xsk->ring_rx.cached_cons; /* write-back local copies to fseqs */
475 0 : FD_VOLATILE( *xsk->ring_fr.prod ) = xsk->ring_fr.cached_prod;
476 0 : if( !fd_xsk_rx_need_wakeup( xsk ) ) return;
477 0 : *charge_busy = 1;
478 0 : struct msghdr _ignored[ 1 ] = { 0 };
479 0 : if( FD_UNLIKELY( -1==recvmsg( xsk->xsk_fd, _ignored, MSG_DONTWAIT ) ) ) {
480 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
481 0 : FD_LOG_ERR(( "xsk recvmsg failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
482 0 : }
483 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
484 0 : long ts = fd_log_wallclock();
485 0 : if( ts > xsk->log_suppress_until_ns ) {
486 0 : FD_LOG_WARNING(( "xsk recvmsg failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
487 0 : xsk->log_suppress_until_ns = ts + (long)1e9;
488 0 : }
489 0 : }
490 0 : }
491 0 : ctx->metrics.xsk_rx_wakeup_cnt++;
492 0 : }
493 :
494 : /* net_tx_wakeup triggers xsk_sendmsg to run in the kernel. Needs to be
495 : called periodically in order to transmit packets. Should only be called
496 : if there are unconsumed packets in Tx ring. */
497 :
498 : static void
499 : net_tx_wakeup( fd_net_ctx_t * ctx,
500 : fd_xsk_t * xsk,
501 24 : int * charge_busy ) {
502 24 : FD_VOLATILE( *xsk->ring_tx.prod ) = xsk->ring_tx.cached_prod; /* write-back local copies to fseqs */
503 24 : FD_VOLATILE( *xsk->ring_cr.cons ) = xsk->ring_cr.cached_cons;
504 24 : if( !fd_xsk_tx_need_wakeup( xsk ) ) return;
505 0 : *charge_busy = 1;
506 0 : if( FD_UNLIKELY( -1==sendto( xsk->xsk_fd, NULL, 0, MSG_DONTWAIT, NULL, 0 ) ) ) {
507 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
508 0 : FD_LOG_ERR(( "xsk sendto failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
509 0 : }
510 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
511 0 : long ts = fd_log_wallclock();
512 0 : if( ts > xsk->log_suppress_until_ns ) {
513 0 : FD_LOG_WARNING(( "xsk sendto failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
514 0 : xsk->log_suppress_until_ns = ts + (long)1e9;
515 0 : }
516 0 : }
517 0 : }
518 0 : ctx->metrics.xsk_tx_wakeup_cnt++;
519 0 : }
520 :
521 : /* net_tx_periodic_wakeup does a timer based xsk_sendmsg wakeup. */
522 :
523 : static inline int
524 : net_tx_periodic_wakeup( fd_net_ctx_t * ctx,
525 : uint xsk_idx,
526 : long now,
527 27 : int * charge_busy ) {
528 27 : fd_xdp_ring_t * tx_ring = &ctx->xsk[ xsk_idx ].ring_tx;
529 27 : int tx_ring_empty = fd_xdp_ring_empty( tx_ring, FD_XDP_RING_ROLE_PROD );
530 27 : if( fd_net_flusher_check( ctx->tx_flusher+xsk_idx, now, tx_ring_empty ) ) {
531 24 : net_tx_wakeup( ctx, &ctx->xsk[ xsk_idx ], charge_busy );
532 24 : fd_net_flusher_wakeup( ctx->tx_flusher+xsk_idx, now );
533 24 : }
534 27 : return 0;
535 27 : }
536 :
537 : static void
538 0 : during_housekeeping( fd_net_ctx_t * ctx ) {
539 0 : long now = fd_tickcount();
540 0 : if( FD_LIKELY( !fd_seqlock_locked_hint( &ctx->netdev_shared.hdr->seqlock ) ) ) {
541 0 : fd_netdev_tbl_copy( &ctx->netdev_tbl, &ctx->netdev_shared );
542 0 : }
543 0 : net_gre_tunnel_ip( ctx );
544 :
545 0 : ctx->metrics.rx_busy_cnt = 0UL;
546 0 : ctx->metrics.rx_idle_cnt = 0UL;
547 0 : ctx->metrics.tx_busy_cnt = 0UL;
548 0 : ctx->metrics.tx_idle_cnt = fd_seq_diff( ctx->free_tx.prod, ctx->free_tx.cons );
549 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
550 0 : fd_xsk_t * xsk = &ctx->xsk[ j ];
551 0 : FD_COMPILER_MFENCE();
552 : /* Write back local copies to fseqs that we own */
553 0 : FD_VOLATILE( *xsk->ring_fr.prod ) = xsk->ring_fr.cached_prod;
554 0 : FD_VOLATILE( *xsk->ring_rx.cons ) = xsk->ring_rx.cached_cons;
555 0 : FD_VOLATILE( *xsk->ring_tx.prod ) = xsk->ring_tx.cached_prod;
556 0 : FD_VOLATILE( *xsk->ring_cr.cons ) = xsk->ring_cr.cached_cons;
557 :
558 : /* Refresh kernel-owned seq numbers for accurate stats */
559 0 : xsk->ring_fr.cached_cons = FD_VOLATILE_CONST( *xsk->ring_fr.cons );
560 0 : xsk->ring_rx.cached_prod = FD_VOLATILE_CONST( *xsk->ring_rx.prod );
561 0 : xsk->ring_tx.cached_cons = FD_VOLATILE_CONST( *xsk->ring_tx.cons );
562 0 : xsk->ring_cr.cached_prod = FD_VOLATILE_CONST( *xsk->ring_cr.prod );
563 :
564 0 : FD_COMPILER_MFENCE();
565 0 : ctx->metrics.rx_busy_cnt += (long)(int)( xsk->ring_rx.cached_prod - xsk->ring_rx.cached_cons );
566 0 : ctx->metrics.rx_idle_cnt += (long)(int)( xsk->ring_fr.cached_prod - xsk->ring_fr.cached_cons );
567 0 : ctx->metrics.tx_busy_cnt += (long)(int)( xsk->ring_tx.cached_prod - xsk->ring_tx.cached_cons );
568 0 : ctx->metrics.tx_busy_cnt += (long)(int)( xsk->ring_cr.cached_prod - xsk->ring_cr.cached_cons );
569 0 : }
570 :
571 0 : if( now > ctx->next_xdp_stats_refresh ) {
572 0 : ctx->next_xdp_stats_refresh = now + ctx->xdp_stats_interval_ticks;
573 0 : poll_xdp_statistics( ctx );
574 0 : }
575 0 : }
576 :
577 :
578 : /* net_tx_route resolves the xsk index, src ip address, src MAC address, and
579 : dst MAC address. Returns 1 on success, 0 on failure.
580 : On success, tx_op->{xsk_idx,src_ip,mac_addrs} is set, and if the dst_ip
581 : belongs to a GRE interface, is_gre_inf will set to 1 and
582 : tx_op->{gre_outer_src_ip, gre_outer_dst_ip} will be loaded from the netdev
583 : table. is_gre_inf is set to 0 if dst_ip doesn't belong to a GRE interface. */
584 :
585 : static int
586 : net_tx_route( fd_net_ctx_t * ctx,
587 : uint dst_ip,
588 33 : uint * is_gre_inf ) {
589 :
590 : /* Route lookup */
591 :
592 33 : fd_fib4_hop_t hop[2] = {0};
593 33 : hop[0] = fd_fib4_lookup( ctx->fib_local, dst_ip, 0UL );
594 33 : hop[1] = fd_fib4_lookup( ctx->fib_main, dst_ip, 0UL );
595 33 : fd_fib4_hop_t const * next_hop = fd_fib4_hop_or( hop+0, hop+1 );
596 :
597 33 : uint rtype = next_hop->rtype;
598 33 : uint if_idx = next_hop->if_idx;
599 33 : uint ip4_src = next_hop->ip4_src;
600 :
601 33 : if( FD_UNLIKELY( rtype==FD_FIB4_RTYPE_LOCAL ) ) {
602 0 : rtype = FD_FIB4_RTYPE_UNICAST;
603 0 : if_idx = 1;
604 0 : }
605 :
606 33 : if( FD_UNLIKELY( rtype!=FD_FIB4_RTYPE_UNICAST ) ) {
607 0 : ctx->metrics.tx_route_fail_cnt++;
608 0 : return 0;
609 0 : }
610 :
611 33 : fd_netdev_t * netdev = fd_netdev_tbl_query( &ctx->netdev_tbl, if_idx );
612 33 : if( !netdev ) {
613 3 : ctx->metrics.tx_route_fail_cnt++;
614 3 : return 0;
615 3 : }
616 :
617 30 : ip4_src = fd_uint_if( !!ctx->bind_address, ctx->bind_address, ip4_src );
618 30 : ctx->tx_op.src_ip = ip4_src;
619 30 : ctx->tx_op.xsk_idx = UINT_MAX;
620 :
621 30 : FD_TEST( is_gre_inf );
622 30 : *is_gre_inf = 0;
623 30 : if( netdev->dev_type==ARPHRD_LOOPBACK ) {
624 : /* Set Ethernet src and dst address to 00:00:00:00:00:00 */
625 0 : memset( ctx->tx_op.mac_addrs, 0, 12UL );
626 0 : ctx->tx_op.xsk_idx = XSK_IDX_LO;
627 : /* Set preferred src address to 127.0.0.1 if no bind address is set */
628 0 : if( !ctx->tx_op.src_ip ) ctx->tx_op.src_ip = FD_IP4_ADDR( 127,0,0,1 );
629 0 : return 1;
630 30 : } else if( netdev->dev_type==ARPHRD_IPGRE ) {
631 : /* skip MAC addrs lookup for GRE inner dst ip */
632 12 : if( netdev->gre_src_ip ) ctx->tx_op.gre_outer_src_ip = netdev->gre_src_ip;
633 12 : ctx->tx_op.gre_outer_dst_ip = netdev->gre_dst_ip;
634 12 : *is_gre_inf = 1;
635 12 : return 1;
636 12 : }
637 :
638 18 : if( FD_UNLIKELY( netdev->dev_type!=ARPHRD_ETHER ) ) return 0; // drop
639 :
640 18 : if( FD_UNLIKELY( if_idx!=ctx->if_virt ) ) {
641 0 : ctx->metrics.tx_no_xdp_cnt++;
642 0 : return 0;
643 0 : }
644 18 : ctx->tx_op.xsk_idx = XSK_IDX_MAIN;
645 :
646 : /* Neighbor resolve */
647 18 : uint neigh_ip = next_hop->ip4_gw;
648 18 : if( !neigh_ip ) neigh_ip = dst_ip;
649 :
650 18 : fd_neigh4_entry_t neigh[1];
651 18 : int neigh_res = fd_neigh4_hmap_query_entry( ctx->neigh4, neigh_ip, neigh );
652 18 : if( FD_UNLIKELY( neigh_res!=FD_MAP_SUCCESS ) ) {
653 : /* Neighbor not found */
654 0 : fd_netlink_neigh4_solicit( ctx->neigh4_solicit, neigh_ip, if_idx, fd_frag_meta_ts_comp( fd_tickcount() ) );
655 0 : ctx->metrics.tx_neigh_fail_cnt++;
656 0 : return 0;
657 0 : }
658 18 : if( FD_UNLIKELY( neigh->state != FD_NEIGH4_STATE_ACTIVE ) ) {
659 0 : ctx->metrics.tx_neigh_fail_cnt++;
660 0 : return 0;
661 0 : }
662 18 : ip4_src = fd_uint_if( !ip4_src, ctx->default_address, ip4_src );
663 18 : ctx->tx_op.src_ip = ip4_src;
664 18 : memcpy( ctx->tx_op.mac_addrs+0, neigh->mac_addr, 6 );
665 18 : memcpy( ctx->tx_op.mac_addrs+6, netdev->mac_addr, 6 );
666 :
667 18 : return 1;
668 18 : }
669 :
670 : /* before_frag is called when a new metadata descriptor for a TX job is
671 : found. This callback determines whether this net tile is responsible
672 : for the TX job. If so, it prepares the TX op for the during_frag and
673 : after_frag callbacks. */
674 :
675 : static inline int
676 : before_frag( fd_net_ctx_t * ctx,
677 : ulong in_idx,
678 : ulong seq,
679 18 : ulong sig ) {
680 18 : (void)in_idx; (void)seq;
681 :
682 : /* Find interface index of next packet */
683 18 : ulong proto = fd_disco_netmux_sig_proto( sig );
684 18 : if( FD_UNLIKELY( proto!=DST_PROTO_OUTGOING ) ) return 1;
685 :
686 : /* Load balance TX */
687 18 : uint net_tile_cnt = ctx->net_tile_cnt;
688 18 : uint hash = (uint)fd_disco_netmux_sig_hash( sig );
689 18 : uint target_idx = hash % net_tile_cnt;
690 18 : uint net_tile_id = ctx->net_tile_id;
691 18 : uint dst_ip = fd_disco_netmux_sig_ip( sig );
692 :
693 : /* Skip if another net tile is responsible for this packet.
694 : Fast path for net tiles other than net_tile 0. */
695 :
696 18 : if( net_tile_id!=0 && net_tile_id!=target_idx ) return 1; /* ignore */
697 :
698 :
699 18 : ctx->tx_op.use_gre = 0;
700 18 : ctx->tx_op.gre_outer_dst_ip = 0;
701 18 : ctx->tx_op.gre_outer_src_ip = 0;
702 18 : uint is_gre_inf = 0;
703 :
704 18 : if( FD_UNLIKELY( !net_tx_route( ctx, dst_ip, &is_gre_inf ) ) ) {
705 0 : return 1; /* metrics incremented by net_tx_route */
706 0 : }
707 :
708 18 : uint xsk_idx = ctx->tx_op.xsk_idx;
709 :
710 18 : if( is_gre_inf ) {
711 12 : uint inner_src_ip = ctx->tx_op.src_ip;
712 12 : if( FD_UNLIKELY( !inner_src_ip ) ) {
713 0 : ctx->metrics.tx_gre_route_fail_cnt++;
714 0 : return 1;
715 0 : }
716 : /* Find the MAC addrs for the eth hdr, and src ip for outer ip4 hdr if not found in netdev tbl */
717 12 : ctx->tx_op.src_ip = 0;
718 12 : is_gre_inf = 0;
719 12 : if( FD_UNLIKELY( !net_tx_route( ctx, ctx->tx_op.gre_outer_dst_ip, &is_gre_inf ) ) ) {
720 0 : ctx->metrics.tx_gre_route_fail_cnt++;
721 0 : return 1;
722 0 : }
723 12 : if( is_gre_inf ) {
724 : /* Only one layer of tunnelling supported */
725 0 : ctx->metrics.tx_gre_route_fail_cnt++;
726 0 : return 1;
727 0 : }
728 12 : if( !ctx->tx_op.gre_outer_src_ip ) {
729 6 : ctx->tx_op.gre_outer_src_ip = ctx->tx_op.src_ip;
730 6 : }
731 12 : ctx->tx_op.use_gre = 1; /* indicate to during_frag to use GRE header */
732 12 : ctx->tx_op.src_ip = inner_src_ip;
733 12 : xsk_idx = XSK_IDX_MAIN;
734 12 : }
735 :
736 18 : if( FD_UNLIKELY( xsk_idx>=ctx->xsk_cnt ) ) {
737 : /* Packet does not route to an XDP interface */
738 0 : ctx->metrics.tx_no_xdp_cnt++;
739 0 : return 1;
740 0 : }
741 :
742 18 : if( xsk_idx==XSK_IDX_LO ) target_idx = 0; /* loopback always targets tile 0 */
743 :
744 : /* Skip if another net tile is responsible for this packet */
745 :
746 18 : if( net_tile_id!=target_idx ) return 1; /* ignore */
747 :
748 : /* Skip if TX is blocked */
749 :
750 18 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
751 18 : fd_net_free_ring_t * free = &ctx->free_tx;
752 18 : if( FD_UNLIKELY( !net_tx_ready( &xsk->ring_tx, free ) ) ) {
753 0 : ctx->metrics.tx_full_fail_cnt++;
754 0 : return 1;
755 0 : }
756 :
757 : /* Allocate buffer for receive */
758 18 : ulong alloc_seq = free->cons;
759 18 : void * frame = (void *)free->queue[ alloc_seq % free->depth ];
760 18 : free->cons = fd_seq_inc( alloc_seq, 1UL );
761 :
762 18 : ctx->tx_op.frame = frame;
763 :
764 18 : return 0; /* continue */
765 18 : }
766 :
767 : /* during_frag is called when before_frag has committed to transmit an
768 : outgoing packet. */
769 :
770 : static inline void
771 : during_frag( fd_net_ctx_t * ctx,
772 : ulong in_idx,
773 : ulong seq FD_PARAM_UNUSED,
774 : ulong sig FD_PARAM_UNUSED,
775 : ulong chunk,
776 : ulong sz,
777 18 : ulong ctl FD_PARAM_UNUSED ) {
778 18 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz>FD_NET_MTU ) )
779 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 ));
780 :
781 18 : if( FD_UNLIKELY( sz<( sizeof(fd_eth_hdr_t)+sizeof(fd_ip4_hdr_t) ) ) )
782 0 : FD_LOG_ERR(( "packet too small %lu (in_idx=%lu)", sz, in_idx ));
783 :
784 18 : if( FD_UNLIKELY( sz>FD_ETH_PAYLOAD_MAX ) )
785 0 : FD_LOG_ERR(( "packet too big %lu (in_idx=%lu)", sz, in_idx ));
786 :
787 18 : void * frame = ctx->tx_op.frame;
788 18 : if( FD_UNLIKELY( (ulong)frame < (ulong)ctx->umem ) )
789 0 : FD_LOG_ERR(( "frame %p out of bounds (below %p)", frame, (void *)ctx->umem ));
790 18 : ulong umem_off = (ulong)frame - (ulong)ctx->umem;
791 18 : if( FD_UNLIKELY( (ulong)umem_off > (ulong)ctx->umem_sz ) )
792 0 : FD_LOG_ERR(( "frame %p out of bounds (beyond %p)", frame, (void *)ctx->umem_sz ));
793 :
794 : /* Speculatively copy frame into XDP buffer */
795 18 : uchar const * src = fd_chunk_to_laddr_const( ctx->in[ in_idx ].mem, chunk );
796 :
797 18 : if( ctx->tx_op.use_gre ) {
798 : /* Discard the ethernet hdr from src. Copy the rest to where the inner ip4_hdr is.
799 : Safe from overflow: FD_ETH_PAYLOAD_MAX + header overhead < frame size (2048UL) */
800 12 : ulong overhead = sizeof(fd_eth_hdr_t) + sizeof(fd_ip4_hdr_t) + sizeof(fd_gre_hdr_t);
801 12 : fd_memcpy( (void *)( (ulong)ctx->tx_op.frame + overhead ), src + sizeof(fd_eth_hdr_t), sz - sizeof(fd_eth_hdr_t) );
802 12 : } else {
803 6 : fd_memcpy( ctx->tx_op.frame, src, sz );
804 6 : }
805 18 : }
806 :
807 : /* after_frag is called when the during_frag memcpy was _not_ overrun. */
808 :
809 : static void
810 : after_frag( fd_net_ctx_t * ctx,
811 : ulong in_idx,
812 : ulong seq,
813 : ulong sig,
814 : ulong sz,
815 : ulong tsorig,
816 : ulong tspub,
817 18 : fd_stem_context_t * stem ) {
818 18 : (void)in_idx; (void)seq; (void)sig; (void)tsorig; (void)tspub; (void)stem;
819 :
820 : /* Current send operation */
821 :
822 18 : uchar * frame = ctx->tx_op.frame;
823 18 : uint xsk_idx = ctx->tx_op.xsk_idx;
824 :
825 : /* Select Ethernet addresses */
826 18 : memcpy( frame, ctx->tx_op.mac_addrs, 12 );
827 :
828 18 : uchar * iphdr = frame + sizeof(fd_eth_hdr_t);
829 :
830 18 : if( ctx->tx_op.use_gre ) {
831 :
832 : /* For GRE packets, the ethertype will always be FD_ETH_HDR_TYPE_IP. outer source ip can't be 0 */
833 12 : if( FD_UNLIKELY( ctx->tx_op.gre_outer_src_ip==0 ) ) {
834 0 : ctx->metrics.tx_gre_route_fail_cnt++;
835 0 : return;
836 0 : }
837 :
838 : /* Write the last two bytes for eth_hdr */
839 12 : FD_STORE( ushort, frame+12, fd_ushort_bswap( FD_ETH_HDR_TYPE_IP ) );
840 :
841 12 : uchar * outer_iphdr = frame + sizeof(fd_eth_hdr_t);
842 12 : uchar * gre_hdr = outer_iphdr + sizeof(fd_ip4_hdr_t);
843 12 : uchar * inner_iphdr = gre_hdr + sizeof(fd_gre_hdr_t);
844 :
845 : /* outer hdr + gre hdr + inner net_tot_len */
846 12 : ushort outer_net_tot_len = (ushort)( sizeof(fd_ip4_hdr_t) + sizeof(fd_gre_hdr_t) + fd_ushort_bswap( ( (fd_ip4_hdr_t *)inner_iphdr )->net_tot_len ) );
847 :
848 : /* Construct outer ip header */
849 12 : fd_ip4_hdr_t ip4_outer = (fd_ip4_hdr_t) {
850 12 : .verihl = FD_IP4_VERIHL( 4,5 ),
851 12 : .tos = 0,
852 12 : .net_tot_len = fd_ushort_bswap( outer_net_tot_len ),
853 12 : .net_id = 0,
854 12 : .net_frag_off = fd_ushort_bswap( FD_IP4_HDR_FRAG_OFF_DF ),
855 12 : .ttl = 64,
856 12 : .protocol = FD_IP4_HDR_PROTOCOL_GRE,
857 12 : .check = 0,
858 12 : .saddr = ctx->tx_op.gre_outer_src_ip,
859 12 : .daddr = ctx->tx_op.gre_outer_dst_ip,
860 12 : };
861 12 : ip4_outer.check = fd_ip4_hdr_check_fast( &ip4_outer );
862 12 : FD_STORE( fd_ip4_hdr_t, outer_iphdr, ip4_outer );
863 :
864 : /* Construct gre header */
865 12 : fd_gre_hdr_t gre_hdr_ = {
866 12 : .flags_version = FD_GRE_HDR_FLG_VER_BASIC,
867 12 : .protocol = fd_ushort_bswap( FD_ETH_HDR_TYPE_IP )
868 12 : };
869 12 : FD_STORE( fd_gre_hdr_t, gre_hdr, gre_hdr_ );
870 :
871 12 : iphdr = inner_iphdr;
872 12 : sz = sizeof(fd_eth_hdr_t) + outer_net_tot_len;
873 12 : xsk_idx = 0;
874 12 : }
875 :
876 : /* Construct (inner) ip header */
877 18 : uint ihl = FD_IP4_GET_LEN( *(fd_ip4_hdr_t *)iphdr );
878 18 : uint ver = FD_IP4_GET_VERSION( *(fd_ip4_hdr_t *)iphdr );
879 18 : uint ip4_saddr = FD_LOAD( uint, iphdr+12 );
880 18 : ushort ethertype = FD_LOAD( ushort, frame+12 );
881 :
882 18 : if( FD_UNLIKELY( ethertype!=fd_ushort_bswap( FD_ETH_HDR_TYPE_IP ) ) ) {
883 0 : FD_LOG_CRIT(( "in link %lu attempted to send packet with invalid ethertype %04x",
884 0 : in_idx, fd_ushort_bswap( ethertype ) ));
885 0 : }
886 :
887 18 : if( ver!=0x4 ) {
888 0 : ctx->metrics.tx_route_fail_cnt++; // Not an IPv4 packet. drop
889 0 : return;
890 0 : }
891 :
892 18 : if( ip4_saddr==0 ) {
893 18 : if( FD_UNLIKELY( ctx->tx_op.src_ip==0 ||
894 18 : ihl<sizeof(fd_ip4_hdr_t) ||
895 18 : (sizeof(fd_eth_hdr_t)+ihl)>sz ) ) {
896 : /* Outgoing IPv4 packet with unknown src IP or invalid IHL */
897 : /* FIXME should select first IPv4 address of device table here */
898 0 : ctx->metrics.tx_route_fail_cnt++;
899 0 : return;
900 0 : }
901 : /* Recompute checksum after changing header */
902 18 : FD_STORE( uint, iphdr+12, ctx->tx_op.src_ip );
903 18 : FD_STORE( ushort, iphdr+10, 0 );
904 18 : FD_STORE( ushort, iphdr+10, fd_ip4_hdr_check( iphdr ) );
905 18 : }
906 :
907 : /* Submit packet TX job
908 :
909 : Invariant for ring_tx: prod-cons<length
910 : (This invariant breaks if any other packet is sent over this ring
911 : between before_frag and this point, e.g. send_arp_probe.) */
912 :
913 18 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
914 18 : fd_xdp_ring_t * tx_ring = &xsk->ring_tx;
915 18 : uint tx_seq = tx_ring->cached_prod;
916 18 : uint tx_mask = tx_ring->depth - 1U;
917 18 : xsk->ring_tx.packet_ring[ tx_seq&tx_mask ] = (struct xdp_desc) {
918 18 : .addr = (ulong)frame - (ulong)ctx->umem,
919 18 : .len = (uint)sz,
920 18 : .options = 0
921 18 : };
922 :
923 : /* Frame is now owned by kernel. Clear tx_op. */
924 18 : ctx->tx_op.frame = NULL;
925 :
926 : /* Register newly enqueued packet */
927 18 : tx_ring->cached_prod = tx_seq+1U;
928 18 : ctx->metrics.tx_submit_cnt++;
929 18 : ctx->metrics.tx_bytes_total += sz;
930 18 : if( ctx->tx_op.use_gre ) ctx->metrics.tx_gre_cnt++;
931 18 : fd_net_flusher_inc( ctx->tx_flusher+xsk_idx, fd_tickcount() );
932 18 : }
933 :
934 : /* net_rx_packet is called when a new Ethernet frame is available.
935 : Attempts to copy out the frame to a downstream tile. */
936 :
937 : static void
938 : net_rx_packet( fd_net_ctx_t * ctx,
939 : ulong umem_off,
940 : ulong sz,
941 27 : uint * freed_chunk ) {
942 :
943 27 : if( FD_UNLIKELY( sz<sizeof(fd_eth_hdr_t)+sizeof(fd_ip4_hdr_t)+sizeof(fd_udp_hdr_t) ) ) {
944 0 : FD_DTRACE_PROBE( net_tile_err_rx_undersz );
945 0 : ctx->metrics.rx_undersz_cnt++;
946 0 : return;
947 0 : }
948 :
949 27 : uchar * packet = (uchar *)ctx->umem + umem_off;
950 27 : uchar const * packet_end = packet + sz;
951 27 : fd_ip4_hdr_t * iphdr = (fd_ip4_hdr_t *)(packet + sizeof(fd_eth_hdr_t));
952 :
953 27 : if( FD_UNLIKELY( ((fd_eth_hdr_t *)packet)->net_type!=fd_ushort_bswap( FD_ETH_HDR_TYPE_IP ) ) ) return;
954 :
955 27 : int is_packet_gre = 0;
956 : /* Discard the GRE overhead (outer iphdr and gre hdr) */
957 27 : if( iphdr->protocol == FD_IP4_HDR_PROTOCOL_GRE ) {
958 15 : if( FD_UNLIKELY( !ctx->gre_tunnel_ip[0] ) ) { /* if the first entry is 0, they all are */
959 0 : ctx->metrics.rx_gre_ignored_cnt++;
960 0 : return;
961 0 : }
962 15 : ulong gre_ipver = FD_IP4_GET_VERSION( *iphdr );
963 15 : ulong gre_iplen = FD_IP4_GET_LEN( *iphdr );
964 15 : if( FD_UNLIKELY( gre_ipver!=0x4 || gre_iplen<20 ) ) {
965 0 : FD_DTRACE_PROBE( net_tile_err_rx_noip );
966 0 : ctx->metrics.rx_gre_inv_pkt_cnt++; /* drop IPv6 packets */
967 0 : return;
968 0 : }
969 :
970 15 : int found = 0;
971 75 : for( ulong i=0UL; i<MAX_GRE_CNT; i++ ) found |= (iphdr->saddr==ctx->gre_tunnel_ip[i]);
972 15 : if( FD_UNLIKELY( (!found) | (iphdr->saddr==0U) ) ) {
973 3 : ctx->metrics.rx_src_addr_invalid_cnt++;
974 3 : return;
975 3 : }
976 :
977 12 : ulong overhead = gre_iplen + sizeof(fd_gre_hdr_t);
978 12 : if( FD_UNLIKELY( (uchar *)iphdr+overhead+sizeof(fd_ip4_hdr_t)>packet_end ) ) {
979 0 : FD_DTRACE_PROBE( net_tile_err_rx_undersz );
980 0 : ctx->metrics.rx_undersz_cnt++; /* inner ip4 header invalid */
981 0 : return;
982 0 : }
983 :
984 : /* The new iphdr is where the inner iphdr was. Copy over the eth_hdr */
985 12 : iphdr = (fd_ip4_hdr_t *)((uchar *)iphdr + overhead);
986 12 : uchar * new_packet = (uchar *)iphdr - sizeof(fd_eth_hdr_t);
987 12 : fd_memcpy( new_packet, packet, sizeof(fd_eth_hdr_t) );
988 12 : sz -= overhead;
989 12 : packet = new_packet;
990 12 : umem_off = (ulong)( packet - (uchar *)ctx->umem );
991 12 : is_packet_gre = 1;
992 12 : }
993 :
994 : /* Translate packet to UMEM frame index */
995 24 : ulong chunk = ctx->umem_chunk0 + (umem_off>>FD_CHUNK_LG_SZ);
996 24 : ulong ctl = umem_off & 0x3fUL;
997 :
998 : /* Filter for UDP/IPv4 packets. */
999 24 : ulong ipver = FD_IP4_GET_VERSION( *iphdr );
1000 24 : ulong iplen = FD_IP4_GET_LEN ( *iphdr );
1001 24 : if( FD_UNLIKELY( ipver!=0x4 || iplen<20 ||
1002 24 : iphdr->protocol!=FD_IP4_HDR_PROTOCOL_UDP ) ) {
1003 0 : FD_DTRACE_PROBE( net_tile_err_rx_noip );
1004 0 : ctx->metrics.rx_undersz_cnt++; /* drop IPv6 packets */
1005 0 : return;
1006 0 : }
1007 :
1008 24 : uchar const * udp = (uchar *)iphdr + iplen;
1009 24 : if( FD_UNLIKELY( udp+sizeof(fd_udp_hdr_t) > packet_end ) ) {
1010 0 : FD_DTRACE_PROBE( net_tile_err_rx_undersz );
1011 0 : ctx->metrics.rx_undersz_cnt++;
1012 0 : return;
1013 0 : }
1014 :
1015 24 : fd_udp_hdr_t const * udp_hdr = (fd_udp_hdr_t const *)udp;
1016 24 : ulong const udp_sz = fd_ushort_bswap( udp_hdr->net_len );
1017 24 : if( FD_UNLIKELY( (udp_sz<sizeof(fd_udp_hdr_t)) | (udp+udp_sz>packet_end) ) ) {
1018 6 : FD_DTRACE_PROBE( net_tile_err_rx_undersz );
1019 6 : ctx->metrics.rx_undersz_cnt++;
1020 6 : return;
1021 6 : }
1022 :
1023 : /* Extract IP dest addr and UDP src/dest port */
1024 18 : uint ip_srcaddr = iphdr->saddr;
1025 18 : ushort udp_srcport = fd_ushort_bswap( udp_hdr->net_sport );
1026 18 : ushort udp_dstport = fd_ushort_bswap( udp_hdr->net_dport );
1027 :
1028 18 : if( FD_UNLIKELY( fd_ip4_addr_is_mcast( ip_srcaddr ) ) ) {
1029 0 : ctx->metrics.rx_src_addr_invalid_cnt++;
1030 0 : return;
1031 0 : }
1032 :
1033 18 : FD_DTRACE_PROBE_4( net_tile_pkt_rx, ip_srcaddr, udp_srcport, udp_dstport, sz );
1034 :
1035 : /* Route packet to downstream tile */
1036 18 : ushort proto;
1037 18 : fd_net_out_ctx_t * out;
1038 18 : if( FD_UNLIKELY( udp_dstport==ctx->shred_listen_port ) ) {
1039 18 : proto = DST_PROTO_SHRED;
1040 18 : out = ctx->shred_out;
1041 18 : } else if( FD_UNLIKELY( udp_dstport==ctx->quic_transaction_listen_port ) ) {
1042 0 : proto = DST_PROTO_TPU_QUIC;
1043 0 : out = ctx->quic_out;
1044 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->legacy_transaction_listen_port ) ) {
1045 0 : proto = DST_PROTO_TPU_UDP;
1046 0 : out = ctx->quic_out;
1047 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->gossip_listen_port ) ) {
1048 0 : proto = DST_PROTO_GOSSIP;
1049 0 : out = ctx->gossvf_out;
1050 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->repair_client_listen_port ) ) {
1051 0 : proto = DST_PROTO_REPAIR;
1052 0 : if( FD_UNLIKELY( sz == REPAIR_PING_SZ ) ) out = ctx->repair_out; /* ping-pong */
1053 0 : else out = ctx->shred_out;
1054 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->repair_serve_listen_port ) ) {
1055 0 : if( FD_UNLIKELY( !ctx->rserve_enabled ) ) return;
1056 0 : proto = DST_PROTO_RSERVE;
1057 0 : out = ctx->rserve_out;
1058 0 : } else if( FD_UNLIKELY( udp_dstport==ctx->txsend_src_port ) ) {
1059 0 : proto = DST_PROTO_SEND;
1060 0 : out = ctx->txsend_out;
1061 0 : } else {
1062 0 : FD_LOG_ERR(( "Firedancer received a UDP packet on port %hu which was not expected. "
1063 0 : "Only the following ports should be configured to forward packets: "
1064 0 : "%hu, %hu, %hu, %hu, %hu, %hu (excluding any 0 ports, which can be ignored)."
1065 0 : "Please report this error to Firedancer maintainers.",
1066 0 : udp_dstport,
1067 0 : ctx->shred_listen_port,
1068 0 : ctx->quic_transaction_listen_port,
1069 0 : ctx->legacy_transaction_listen_port,
1070 0 : ctx->gossip_listen_port,
1071 0 : ctx->repair_client_listen_port,
1072 0 : ctx->repair_serve_listen_port ));
1073 0 : }
1074 :
1075 : /* tile can decide how to partition based on src ip addr and src port */
1076 18 : ulong sig = fd_disco_netmux_sig( ip_srcaddr, udp_srcport, ip_srcaddr, proto, 14UL+8UL+iplen );
1077 :
1078 : /* Peek the mline for an old frame */
1079 18 : fd_frag_meta_t * mline = out->mcache + fd_mcache_line_idx( out->seq, out->depth );
1080 18 : *freed_chunk = mline->chunk;
1081 :
1082 : /* Overwrite the mline with the new frame */
1083 18 : ulong tspub = (ulong)fd_frag_meta_ts_comp( fd_tickcount() );
1084 18 : # if FD_HAS_AVX
1085 18 : fd_mcache_publish_avx( out->mcache, out->depth, out->seq, sig, chunk, sz, ctl, 0, tspub );
1086 : # else
1087 : fd_mcache_publish( out->mcache, out->depth, out->seq, sig, chunk, sz, ctl, 0, tspub );
1088 : # endif
1089 :
1090 : /* Wind up for the next iteration */
1091 18 : out->seq = fd_seq_inc( out->seq, 1UL );
1092 :
1093 18 : if( is_packet_gre ) ctx->metrics.rx_gre_cnt++;
1094 18 : ctx->metrics.rx_pkt_cnt++;
1095 18 : ctx->metrics.rx_bytes_total += sz;
1096 18 : }
1097 :
1098 : /* net_comp_event is called when an XDP TX frame is free again. */
1099 :
1100 : static void
1101 : net_comp_event( fd_net_ctx_t * ctx,
1102 : fd_xsk_t * xsk,
1103 0 : uint comp_seq ) {
1104 :
1105 : /* Locate the incoming frame */
1106 :
1107 0 : fd_xdp_ring_t * comp_ring = &xsk->ring_cr;
1108 0 : uint comp_mask = comp_ring->depth - 1U;
1109 0 : ulong frame = FD_VOLATILE_CONST( comp_ring->frame_ring[ comp_seq&comp_mask ] );
1110 0 : ulong const frame_mask = FD_NET_MTU - 1UL;
1111 0 : FD_STATIC_ASSERT( FD_ULONG_IS_POW2( FD_NET_MTU ), "FD_NET_MTU must be a power of two" );
1112 0 : if( FD_UNLIKELY( frame+FD_NET_MTU > ctx->umem_sz ) ) {
1113 0 : FD_LOG_ERR(( "Bounds check failed: frame=0x%lx umem_sz=0x%lx",
1114 0 : frame, (ulong)ctx->umem_sz ));
1115 0 : }
1116 :
1117 : /* Check if we have space to return the freed frame */
1118 :
1119 0 : fd_net_free_ring_t * free = &ctx->free_tx;
1120 0 : ulong free_prod = free->prod;
1121 0 : ulong free_mask = free->depth - 1UL;
1122 0 : ulong free_cons = free->cons;
1123 0 : long free_cnt = fd_seq_diff( free_prod, free_cons );
1124 0 : FD_TEST( free_prod >= free_cons );
1125 0 : if( FD_UNLIKELY( free_cnt>=(long)free->depth ) ) return; /* blocked */
1126 :
1127 0 : free->queue[ free_prod&free_mask ] = (ulong)ctx->umem + (frame & (~frame_mask));
1128 0 : free->prod = fd_seq_inc( free_prod, 1UL );
1129 :
1130 : /* Wind up for next iteration */
1131 :
1132 0 : comp_ring->cached_cons = comp_seq+1U;
1133 0 : ctx->metrics.tx_complete_cnt++;
1134 0 : }
1135 :
1136 : /* net_rx_event is called when a new XDP RX frame is available. Calls
1137 : net_rx_packet, then returns the packet back to the kernel via the fill
1138 : ring. */
1139 :
1140 : static void
1141 : net_rx_event( fd_net_ctx_t * ctx,
1142 : fd_xsk_t * xsk,
1143 27 : uint rx_seq ) {
1144 : /* Locate the incoming frame */
1145 :
1146 27 : fd_xdp_ring_t * rx_ring = &xsk->ring_rx;
1147 27 : uint rx_mask = rx_ring->depth - 1U;
1148 27 : struct xdp_desc frame = FD_VOLATILE_CONST( rx_ring->packet_ring[ rx_seq&rx_mask ] );
1149 :
1150 27 : if( FD_UNLIKELY( frame.len>FD_NET_MTU ) )
1151 0 : FD_LOG_ERR(( "received a UDP packet with a too large payload (%u)", frame.len ));
1152 :
1153 : /* Check if we have space in the fill ring to free the frame */
1154 :
1155 27 : fd_xdp_ring_t * fill_ring = &xsk->ring_fr;
1156 27 : if( FD_UNLIKELY( fd_xdp_ring_full( fill_ring ) ) ) {
1157 0 : ctx->metrics.rx_fill_blocked_cnt++;
1158 0 : return; /* blocked */
1159 0 : }
1160 :
1161 : /* Pass it to the receive handler */
1162 :
1163 27 : uint freed_chunk = (uint)( ctx->umem_chunk0 + (frame.addr>>FD_CHUNK_LG_SZ) );
1164 27 : net_rx_packet( ctx, frame.addr, frame.len, &freed_chunk );
1165 27 : FD_COMPILER_MFENCE();
1166 27 : rx_ring->cached_cons = rx_seq+1U;
1167 :
1168 : /* Every RX operation returns one frame to the FILL ring. If the
1169 : packet was forwarded to a downstream ring, the newly shadowed frame
1170 : is returned. Otherwise, the frame just received is returned. */
1171 :
1172 27 : if( FD_UNLIKELY( ( freed_chunk < ctx->umem_chunk0 ) |
1173 27 : ( freed_chunk > ctx->umem_wmark ) ) ) {
1174 0 : FD_LOG_CRIT(( "mcache corruption detected: chunk=%u chunk0=%u wmark=%u",
1175 0 : freed_chunk, ctx->umem_chunk0, ctx->umem_wmark ));
1176 0 : }
1177 :
1178 27 : FD_STATIC_ASSERT( FD_ULONG_IS_POW2( FD_NET_MTU ), "FD_NET_MTU must be a power of two" );
1179 27 : uint fill_prod = fill_ring->cached_prod;
1180 27 : uint fill_mask = (fill_ring->depth)-1U;
1181 27 : ulong frame_mask = FD_NET_MTU - 1UL;
1182 27 : ulong freed_off = (freed_chunk - ctx->umem_chunk0)<<FD_CHUNK_LG_SZ;
1183 27 : fill_ring->frame_ring[ fill_prod&fill_mask ] = freed_off & (~frame_mask);
1184 27 : fill_ring->cached_prod = fill_prod+1U;
1185 27 : }
1186 :
1187 : static void
1188 : before_credit_softirq( fd_net_ctx_t * ctx,
1189 : int * charge_busy,
1190 : uint rr_idx,
1191 27 : fd_xsk_t * rr_xsk ) {
1192 :
1193 27 : net_tx_periodic_wakeup( ctx, rr_idx, fd_tickcount(), charge_busy );
1194 :
1195 : /* Fire RX event if we have RX desc avail */
1196 27 : if( !fd_xdp_ring_empty( &rr_xsk->ring_rx, FD_XDP_RING_ROLE_CONS ) ) {
1197 27 : *charge_busy = 1;
1198 27 : net_rx_event( ctx, rr_xsk, rr_xsk->ring_rx.cached_cons );
1199 27 : } else {
1200 0 : net_rx_wakeup( ctx, rr_xsk, charge_busy );
1201 :
1202 : /* Iterate onto the next NAPI queue. */
1203 0 : ctx->rr_idx++;
1204 0 : ctx->rr_idx = fd_uint_if( ctx->rr_idx>=ctx->xsk_cnt, 0, ctx->rr_idx );
1205 0 : }
1206 27 : }
1207 :
1208 : static int
1209 : net_prefbusy_poll_ready( fd_xsk_t * rr_xsk,
1210 : fd_net_flusher_t * flusher,
1211 0 : long now ) {
1212 :
1213 0 : if( FD_LIKELY( now < ( flusher->prefbusy_last_poll_ticks + flusher->prefbusy_min_interval_ticks ) ) ) return 0;
1214 0 : if( FD_UNLIKELY( now > ( flusher->prefbusy_last_poll_ticks + flusher->prefbusy_stall_timeout_ticks ) ) ) return 1;
1215 :
1216 0 : int rx_empty = fd_xdp_ring_empty( &rr_xsk->ring_rx, FD_XDP_RING_ROLE_CONS );
1217 :
1218 0 : return rx_empty;
1219 0 : }
1220 :
1221 : static void
1222 : net_prefbusy_poll_flush( fd_net_flusher_t * flusher,
1223 0 : long now ) {
1224 0 : flusher->prefbusy_last_poll_ticks = now;
1225 0 : }
1226 :
1227 : static void
1228 : before_credit_prefbusy( fd_net_ctx_t * ctx,
1229 : int * charge_busy,
1230 : uint rr_idx,
1231 0 : fd_xsk_t * rr_xsk ) {
1232 :
1233 0 : fd_net_flusher_t * flusher = ctx->tx_flusher+rr_idx;
1234 0 : if( FD_UNLIKELY( net_prefbusy_poll_ready( rr_xsk, flusher, fd_tickcount() ) ) ) {
1235 : /* NAPI needs to be polled to process new TX from
1236 : Firedancer's net tile and process new RX from the NIC. */
1237 :
1238 0 : FD_VOLATILE( *rr_xsk->ring_tx.prod ) = rr_xsk->ring_tx.cached_prod; /* write-back local copies to fseqs */
1239 0 : FD_VOLATILE( *rr_xsk->ring_cr.cons ) = rr_xsk->ring_cr.cached_cons;
1240 0 : FD_VOLATILE( *rr_xsk->ring_rx.cons ) = rr_xsk->ring_rx.cached_cons;
1241 0 : FD_VOLATILE( *rr_xsk->ring_fr.prod ) = rr_xsk->ring_fr.cached_prod;
1242 :
1243 0 : if( FD_UNLIKELY( -1==sendto( rr_xsk->xsk_fd, NULL, 0, MSG_DONTWAIT, NULL, 0 ) ) ) {
1244 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
1245 0 : FD_LOG_ERR(( "xsk sendto failed xsk_fd=%d (%i-%s)", rr_xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
1246 0 : }
1247 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
1248 0 : long ts = fd_log_wallclock();
1249 0 : if( ts > rr_xsk->log_suppress_until_ns ) {
1250 0 : FD_LOG_WARNING(( "xsk sendto failed xsk_fd=%d (%i-%s)", rr_xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
1251 0 : rr_xsk->log_suppress_until_ns = ts + (long)1e9;
1252 0 : }
1253 0 : }
1254 0 : }
1255 : /* Since xsk sendmsg in prefbusy mode drives both rx and tx, both are incremented */
1256 0 : ctx->metrics.xsk_tx_wakeup_cnt++;
1257 0 : ctx->metrics.xsk_rx_wakeup_cnt++;
1258 :
1259 0 : net_prefbusy_poll_flush( flusher, fd_tickcount() );
1260 0 : }
1261 :
1262 : /* Process new RX from xsk ring if there is any. */
1263 0 : if( !fd_xdp_ring_empty( &rr_xsk->ring_rx, FD_XDP_RING_ROLE_CONS ) ) {
1264 0 : *charge_busy = 1;
1265 0 : net_rx_event( ctx, rr_xsk, rr_xsk->ring_rx.cached_cons );
1266 0 : }
1267 : /* Iterate onto the next NAPI queue. */
1268 0 : ctx->rr_idx++;
1269 0 : ctx->rr_idx = fd_uint_if( ctx->rr_idx>=ctx->xsk_cnt, 0, ctx->rr_idx );
1270 0 : }
1271 :
1272 : /* before_credit is called every loop iteration. */
1273 :
1274 : static void
1275 : before_credit( fd_net_ctx_t * ctx,
1276 : fd_stem_context_t * stem,
1277 27 : int * charge_busy ) {
1278 27 : (void)stem;
1279 : /* A previous send attempt was overrun. A corrupt copy of the packet was
1280 : placed into an XDP frame, but the frame was not yet submitted to the
1281 : TX ring. Return the tx buffer to the free list. */
1282 :
1283 27 : if( ctx->tx_op.frame ) {
1284 0 : *charge_busy = 1;
1285 0 : fd_net_free_ring_t * free = &ctx->free_tx;
1286 0 : ulong alloc_seq = free->prod;
1287 0 : free->queue[ alloc_seq % free->depth ] = (ulong)ctx->tx_op.frame;
1288 0 : free->prod = fd_seq_inc( alloc_seq, 1UL );
1289 0 : ctx->tx_op.frame = NULL;
1290 0 : }
1291 :
1292 : /* Check if new packets are available or if TX frames are free again
1293 : (Round-robin through sockets) */
1294 :
1295 27 : uint rr_idx = ctx->rr_idx;
1296 27 : fd_xsk_t * rr_xsk = &ctx->xsk[ rr_idx ];
1297 :
1298 27 : if( FD_LIKELY( !rr_xsk->prefbusy_poll_enabled ) ) {
1299 : /* Default poll mode which relies on irqs and wakeups */
1300 27 : before_credit_softirq( ctx, charge_busy, rr_idx, rr_xsk );
1301 27 : } else {
1302 0 : before_credit_prefbusy( ctx, charge_busy, rr_idx, rr_xsk );
1303 0 : }
1304 :
1305 : /* Fire comp event if we have comp desc avail */
1306 27 : if( !fd_xdp_ring_empty( &rr_xsk->ring_cr, FD_XDP_RING_ROLE_CONS ) ) {
1307 0 : *charge_busy = 1;
1308 0 : net_comp_event( ctx, rr_xsk, rr_xsk->ring_cr.cached_cons );
1309 0 : }
1310 27 : }
1311 :
1312 : /* net_xsk_bootstrap assigns UMEM frames to the FILL ring. */
1313 :
1314 : static ulong
1315 : net_xsk_bootstrap( fd_net_ctx_t * ctx,
1316 : uint xsk_idx,
1317 0 : ulong frame_off ) {
1318 0 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
1319 :
1320 0 : ulong const frame_sz = FD_NET_MTU;
1321 0 : ulong const fr_depth = ctx->xsk[ xsk_idx ].ring_fr.depth/2UL;
1322 :
1323 0 : fd_xdp_ring_t * fill = &xsk->ring_fr;
1324 0 : uint fill_prod = fill->cached_prod;
1325 0 : for( ulong j=0UL; j<fr_depth; j++ ) {
1326 0 : fill->frame_ring[ j ] = frame_off;
1327 0 : frame_off += frame_sz;
1328 0 : }
1329 0 : FD_VOLATILE( *fill->prod ) = fill->cached_prod = fill_prod + (uint)fr_depth;
1330 :
1331 0 : return frame_off;
1332 0 : }
1333 :
1334 : /* FIXME source MAC address from netlnk tile instead */
1335 :
1336 : static void
1337 : interface_addrs( const char * interface,
1338 : uchar * mac,
1339 0 : uint * ip4_addr ) {
1340 0 : int fd = socket( AF_INET, SOCK_DGRAM, 0 );
1341 0 : struct ifreq ifr;
1342 0 : ifr.ifr_addr.sa_family = AF_INET;
1343 :
1344 0 : strncpy( ifr.ifr_name, interface, IFNAMSIZ );
1345 0 : if( FD_UNLIKELY( ioctl( fd, SIOCGIFHWADDR, &ifr ) ) )
1346 0 : FD_LOG_ERR(( "could not get MAC address of interface `%s`: (%i-%s)", interface, errno, fd_io_strerror( errno ) ));
1347 0 : fd_memcpy( mac, ifr.ifr_hwaddr.sa_data, 6 );
1348 :
1349 0 : if( FD_UNLIKELY( ioctl( fd, SIOCGIFADDR, &ifr ) ) )
1350 0 : FD_LOG_ERR(( "could not get IP address of interface `%s`: (%i-%s)", interface, errno, fd_io_strerror( errno ) ));
1351 0 : *ip4_addr = ((struct sockaddr_in *)fd_type_pun( &ifr.ifr_addr ))->sin_addr.s_addr;
1352 :
1353 0 : if( FD_UNLIKELY( close(fd) ) )
1354 0 : FD_LOG_ERR(( "could not close socket (%i-%s)", errno, fd_io_strerror( errno ) ));
1355 0 : }
1356 :
1357 : /* privileged_init does the following initialization steps:
1358 :
1359 : - Create an AF_XDP socket
1360 : - Map XDP metadata rings
1361 : - Register UMEM data region with socket
1362 : - Insert AF_XDP socket into xsk_map
1363 :
1364 : Net tile 0 also runs fd_xdp_install and repeats the above step for
1365 : the loopback device. (Unless the main interface is already loopback)
1366 :
1367 : Kernel object references:
1368 :
1369 : BPF_LINK file descriptor
1370 : |
1371 : +-> XDP program installation on NIC
1372 : | |
1373 : | +-> XDP program <-- BPF_PROG file descriptor (prog_fd)
1374 : |
1375 : +-> XSKMAP object <-- BPF_MAP file descriptor (xsk_map) */
1376 :
1377 : FD_FN_UNUSED static void
1378 : privileged_init( fd_topo_t const * topo,
1379 0 : fd_topo_tile_t const * tile ) {
1380 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1381 :
1382 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1383 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
1384 0 : ulong * free_tx = FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );;
1385 :
1386 0 : fd_memset( ctx, 0, sizeof(fd_net_ctx_t) );
1387 :
1388 0 : interface_addrs( tile->xdp.if_virt, ctx->src_mac_addr, &ctx->default_address );
1389 0 : ctx->if_virt = if_nametoindex( tile->xdp.if_virt ); FD_TEST( ctx->if_virt );
1390 :
1391 : /* Load up dcache containing UMEM */
1392 :
1393 0 : void * const dcache_mem = fd_topo_obj_laddr( topo, tile->net.umem_dcache_obj_id );
1394 0 : void * const umem = fd_dcache_join( dcache_mem );
1395 0 : ulong const umem_dcache_data_sz = fd_dcache_data_sz( umem );
1396 0 : ulong const umem_frame_sz = 2048UL;
1397 0 : ulong const umem_sz = fd_ulong_align_dn( umem_dcache_data_sz, umem_frame_sz );
1398 :
1399 : /* Derive chunk bounds */
1400 :
1401 0 : void * const umem_base = fd_wksp_containing( dcache_mem );
1402 0 : ulong const umem_chunk0 = ( (ulong)umem - (ulong)umem_base )>>FD_CHUNK_LG_SZ;
1403 0 : ulong const umem_wmark = umem_chunk0 + ( ( umem_sz-umem_frame_sz )>>FD_CHUNK_LG_SZ );
1404 :
1405 0 : if( FD_UNLIKELY( umem_chunk0>UINT_MAX || umem_wmark>UINT_MAX || umem_chunk0>umem_wmark ) ) {
1406 0 : FD_LOG_ERR(( "Calculated invalid UMEM bounds [%lu,%lu]", umem_chunk0, umem_wmark ));
1407 0 : }
1408 :
1409 0 : if( FD_UNLIKELY( !umem_base ) ) FD_LOG_ERR(( "UMEM dcache is not in a workspace" ));
1410 :
1411 0 : ctx->umem = umem;
1412 0 : ctx->umem_sz = umem_sz;
1413 0 : ctx->umem_chunk0 = (uint)umem_chunk0;
1414 0 : ctx->umem_wmark = (uint)umem_wmark;
1415 :
1416 0 : ctx->free_tx.queue = free_tx;
1417 0 : ctx->free_tx.depth = tile->xdp.xdp_tx_queue_size;
1418 :
1419 : /* Create and install XSKs */
1420 :
1421 0 : uint if_phys_if_idx = if_nametoindex( tile->xdp.if_phys );
1422 0 : if( FD_UNLIKELY( !if_phys_if_idx ) ) FD_LOG_ERR(( "if_nametoindex(%s) failed", tile->xdp.if_phys ));
1423 :
1424 0 : fd_xsk_params_t params0 = {
1425 0 : .if_idx = if_phys_if_idx,
1426 0 : .if_queue_id = tile->xdp.if_queue,
1427 :
1428 : /* Some kernels produce EOPNOTSUP errors on sendto calls when
1429 : starting up without either XDP_ZEROCOPY or XDP_COPY
1430 : (e.g. 5.14.0-503.23.1.el9_5 with i40e) */
1431 0 : .bind_flags = tile->xdp.zero_copy ? XDP_ZEROCOPY : XDP_COPY,
1432 :
1433 0 : .prefbusy_time_budget_micros = PREFBUSY_TIME_BUDGET_MICROS,
1434 :
1435 0 : .prefbusy_rx_budget = PREFBUSY_RX_BUDGET,
1436 :
1437 0 : .fr_depth = tile->xdp.xdp_rx_queue_size*2,
1438 0 : .rx_depth = tile->xdp.xdp_rx_queue_size,
1439 0 : .cr_depth = tile->xdp.xdp_tx_queue_size,
1440 0 : .tx_depth = tile->xdp.xdp_tx_queue_size,
1441 :
1442 0 : .umem_addr = umem,
1443 0 : .frame_sz = umem_frame_sz,
1444 0 : .umem_sz = umem_sz,
1445 :
1446 0 : .core_dump = tile->xdp.xsk_core_dump,
1447 0 : };
1448 :
1449 0 : fd_cstr_ncpy( params0.poll_mode, tile->xdp.poll_mode, sizeof(params0.poll_mode) );
1450 :
1451 : /* Re-derive XDP file descriptors */
1452 :
1453 0 : fd_xdp_fds_t xdp_fds[ FD_TOPO_XDP_FDS_MAX ];
1454 0 : uint xdp_fds_cnt = FD_TOPO_XDP_FDS_MAX;
1455 0 : fd_topo_install_xdp( topo, xdp_fds, &xdp_fds_cnt, 0U, /* dry_run */ 1 );
1456 :
1457 0 : int xsk_map_fd = -1;
1458 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
1459 0 : if( xdp_fds[ i ].if_idx==if_phys_if_idx ) {
1460 0 : xsk_map_fd = xdp_fds[ i ].xsk_map_fd;
1461 0 : ctx->prog_link_fds[ 0 ] = xdp_fds[ i ].prog_link_fd;
1462 0 : xdp_fds[ i ].prog_link_fd = -1; /* mark as used */
1463 0 : break;
1464 0 : }
1465 0 : }
1466 0 : FD_TEST( xsk_map_fd>=0 );
1467 :
1468 : /* Init XSK */
1469 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 ));
1470 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 ));
1471 0 : ctx->xsk_cnt = 1;
1472 :
1473 : /* Networking tile at index 0 also binds to loopback (only queue 0 available on lo) */
1474 :
1475 0 : if( FD_UNLIKELY( strcmp( tile->xdp.if_virt, "lo" ) && !tile->kind_id ) ) {
1476 0 : ctx->xsk_cnt = 2;
1477 :
1478 0 : uint lo_idx = if_nametoindex( "lo" );
1479 0 : if( FD_UNLIKELY( !lo_idx ) ) FD_LOG_ERR(( "if_nametoindex(lo) failed" ));
1480 :
1481 0 : int lo_xsk_map_fd = -1;
1482 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
1483 0 : if( xdp_fds[ i ].if_idx==lo_idx ) {
1484 0 : lo_xsk_map_fd = xdp_fds[ i ].xsk_map_fd;
1485 0 : ctx->prog_link_fds[ 1 ] = xdp_fds[ i ].prog_link_fd;
1486 0 : xdp_fds[ i ].prog_link_fd = -1; /* mark as used */
1487 0 : break;
1488 0 : }
1489 0 : }
1490 0 : FD_TEST( lo_xsk_map_fd>=0 );
1491 :
1492 : /* init xsk 1 */
1493 0 : fd_xsk_params_t params1 = params0;
1494 0 : params1.if_idx = lo_idx; /* probably always 1 */
1495 0 : params1.if_queue_id = 0;
1496 0 : params1.bind_flags = 0;
1497 0 : if( FD_UNLIKELY( !fd_xsk_init( &ctx->xsk[ 1 ], ¶ms1 ) ) ) FD_LOG_ERR(( "failed to bind lo_xsk" ));
1498 0 : if( FD_UNLIKELY( !fd_xsk_activate( &ctx->xsk[ 1 ], lo_xsk_map_fd ) ) ) FD_LOG_ERR(( "failed to activate lo_xsk" ));
1499 0 : }
1500 :
1501 : /* Close unused XDP fds */
1502 :
1503 0 : if( FD_UNLIKELY( fd_sandbox_gettid()==fd_sandbox_getpid() ) ) {
1504 : /* Kind of gross.. in single threaded mode we don't want to close the xsk_map_fd
1505 : since it's shared with other net tiles. Just check for that by seeing if we
1506 : are the only thread in the process. */
1507 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
1508 0 : if( -1==close( xdp_fds[ i ].xsk_map_fd ) ) {
1509 0 : FD_LOG_ERR(( "close(%d) failed (%d-%s)", xsk_map_fd, errno, fd_io_strerror( errno ) ));
1510 0 : }
1511 0 : if( xdp_fds[ i ].prog_link_fd>0 &&
1512 0 : -1==close( xdp_fds[ i ].prog_link_fd ) ) {
1513 0 : FD_LOG_ERR(( "close(%d) failed (%d-%s)", xsk_map_fd, errno, fd_io_strerror( errno ) ));
1514 0 : }
1515 0 : }
1516 0 : }
1517 :
1518 0 : double tick_per_ns = fd_tempo_tick_per_ns( NULL );
1519 0 : ctx->xdp_stats_interval_ticks = (long)( FD_XDP_STATS_INTERVAL_NS * tick_per_ns );
1520 :
1521 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, scratch_align() );
1522 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
1523 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
1524 0 : }
1525 :
1526 : static void
1527 : init_device_table( fd_net_ctx_t * ctx,
1528 : void * netdev_tbl_shm,
1529 3 : void * netdev_tbl_local ) {
1530 3 : FD_TEST( fd_netdev_tbl_join( &ctx->netdev_shared, netdev_tbl_shm ) );
1531 3 : FD_TEST( fd_netdev_tbl_new( netdev_tbl_local, NETDEV_MAX, BOND_MASTER_MAX ) );
1532 3 : FD_TEST( fd_netdev_tbl_join( &ctx->netdev_tbl, netdev_tbl_local ) );
1533 3 : }
1534 :
1535 : FD_FN_UNUSED static void
1536 : unprivileged_init( fd_topo_t const * topo,
1537 0 : fd_topo_tile_t const * tile ) {
1538 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1539 :
1540 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1541 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
1542 0 : FD_TEST( ctx->xsk_cnt!=0 );
1543 0 : FD_TEST( ctx->free_tx.queue!=NULL );
1544 0 : (void)FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );
1545 0 : void * netdev_tbl_local = FD_SCRATCH_ALLOC_APPEND( l, fd_netdev_tbl_align(), fd_netdev_tbl_footprint( NETDEV_MAX, BOND_MASTER_MAX ) );
1546 :
1547 0 : ctx->net_tile_id = (uint)tile->kind_id;
1548 0 : ctx->net_tile_cnt = (uint)fd_topo_tile_name_cnt( topo, tile->name );
1549 :
1550 0 : ctx->bind_address = tile->net.bind_address;
1551 0 : ctx->shred_listen_port = tile->net.shred_listen_port;
1552 0 : ctx->quic_transaction_listen_port = tile->net.quic_transaction_listen_port;
1553 0 : ctx->legacy_transaction_listen_port = tile->net.legacy_transaction_listen_port;
1554 0 : ctx->gossip_listen_port = tile->net.gossip_listen_port;
1555 0 : ctx->repair_client_listen_port = tile->net.repair_client_listen_port;
1556 0 : ctx->repair_serve_listen_port = tile->net.repair_serve_listen_port;
1557 0 : ctx->txsend_src_port = tile->net.txsend_src_port;
1558 :
1559 : /* Put a bound on chunks we read from the input, to make sure they
1560 : are within in the data region of the workspace. */
1561 :
1562 0 : if( FD_UNLIKELY( !tile->in_cnt ) ) FD_LOG_ERR(( "net tile in link cnt is zero" ));
1563 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 ));
1564 0 : FD_TEST( tile->in_cnt<=32 );
1565 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
1566 0 : fd_topo_link_t const * link = &topo->links[ tile->in_link_id[ i ] ];
1567 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 ));
1568 :
1569 0 : ctx->in[ i ].mem = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
1570 0 : ctx->in[ i ].chunk0 = fd_dcache_compact_chunk0( ctx->in[ i ].mem, link->dcache );
1571 0 : ctx->in[ i ].wmark = fd_dcache_compact_wmark( ctx->in[ i ].mem, link->dcache, link->mtu );
1572 0 : }
1573 :
1574 0 : ctx->rserve_enabled = 0;
1575 0 : for( ulong i = 0; i < tile->out_cnt; i++ ) {
1576 0 : fd_topo_link_t const * out_link = &topo->links[ tile->out_link_id[ i ] ];
1577 0 : if( strcmp( out_link->name, "net_quic" ) == 0 ) {
1578 0 : fd_topo_link_t const * quic_out = out_link;
1579 0 : ctx->quic_out->mcache = quic_out->mcache;
1580 0 : ctx->quic_out->sync = fd_mcache_seq_laddr( ctx->quic_out->mcache );
1581 0 : ctx->quic_out->depth = fd_mcache_depth( ctx->quic_out->mcache );
1582 0 : ctx->quic_out->seq = fd_mcache_seq_query( ctx->quic_out->sync );
1583 0 : } else if( strcmp( out_link->name, "net_shred" ) == 0 ) {
1584 0 : fd_topo_link_t const * shred_out = out_link;
1585 0 : ctx->shred_out->mcache = shred_out->mcache;
1586 0 : ctx->shred_out->sync = fd_mcache_seq_laddr( ctx->shred_out->mcache );
1587 0 : ctx->shred_out->depth = fd_mcache_depth( ctx->shred_out->mcache );
1588 0 : ctx->shred_out->seq = fd_mcache_seq_query( ctx->shred_out->sync );
1589 0 : } else if( strcmp( out_link->name, "net_gossvf" ) == 0 ) {
1590 0 : fd_topo_link_t const * gossip_out = out_link;
1591 0 : ctx->gossvf_out->mcache = gossip_out->mcache;
1592 0 : ctx->gossvf_out->sync = fd_mcache_seq_laddr( ctx->gossvf_out->mcache );
1593 0 : ctx->gossvf_out->depth = fd_mcache_depth( ctx->gossvf_out->mcache );
1594 0 : ctx->gossvf_out->seq = fd_mcache_seq_query( ctx->gossvf_out->sync );
1595 0 : } else if( strcmp( out_link->name, "net_repair" ) == 0 ) {
1596 0 : fd_topo_link_t const * repair_out = out_link;
1597 0 : ctx->repair_out->mcache = repair_out->mcache;
1598 0 : ctx->repair_out->sync = fd_mcache_seq_laddr( ctx->repair_out->mcache );
1599 0 : ctx->repair_out->depth = fd_mcache_depth( ctx->repair_out->mcache );
1600 0 : ctx->repair_out->seq = fd_mcache_seq_query( ctx->repair_out->sync );
1601 0 : } else if( strcmp( out_link->name, "net_netlnk" ) == 0 ) {
1602 0 : fd_topo_link_t const * netlink_out = out_link;
1603 0 : ctx->neigh4_solicit->mcache = netlink_out->mcache;
1604 0 : ctx->neigh4_solicit->depth = fd_mcache_depth( ctx->neigh4_solicit->mcache );
1605 0 : ctx->neigh4_solicit->seq = fd_mcache_seq_query( fd_mcache_seq_laddr( ctx->neigh4_solicit->mcache ) );
1606 0 : } else if( strcmp( out_link->name, "net_txsend" ) == 0 ) {
1607 0 : fd_topo_link_t const * txsend_out = out_link;
1608 0 : ctx->txsend_out->mcache = txsend_out->mcache;
1609 0 : ctx->txsend_out->sync = fd_mcache_seq_laddr( ctx->txsend_out->mcache );
1610 0 : ctx->txsend_out->depth = fd_mcache_depth( ctx->txsend_out->mcache );
1611 0 : ctx->txsend_out->seq = fd_mcache_seq_query( ctx->txsend_out->sync );
1612 0 : } else if( strcmp( out_link->name, "net_rserve" ) == 0 ) {
1613 0 : fd_topo_link_t const * rserve_out = out_link;
1614 0 : ctx->rserve_out->mcache = rserve_out->mcache;
1615 0 : ctx->rserve_out->sync = fd_mcache_seq_laddr( ctx->rserve_out->mcache );
1616 0 : ctx->rserve_out->depth = fd_mcache_depth( ctx->rserve_out->mcache );
1617 0 : ctx->rserve_out->seq = fd_mcache_seq_query( ctx->rserve_out->sync );
1618 0 : ctx->rserve_enabled = 1;
1619 0 : } else {
1620 0 : FD_LOG_ERR(( "unrecognized out link `%s`", out_link->name ));
1621 0 : }
1622 0 : }
1623 :
1624 : /* Check if any of the tiles we set a listen port for do not have an outlink. */
1625 0 : if( FD_UNLIKELY( ctx->shred_listen_port!=0 && ctx->shred_out->mcache==NULL ) ) {
1626 0 : FD_LOG_ERR(( "shred listen port set but no out link was found" ));
1627 0 : } else if( FD_UNLIKELY( ctx->quic_transaction_listen_port!=0 && ctx->quic_out->mcache==NULL ) ) {
1628 0 : FD_LOG_ERR(( "quic transaction listen port set but no out link was found" ));
1629 0 : } else if( FD_UNLIKELY( ctx->legacy_transaction_listen_port!=0 && ctx->quic_out->mcache==NULL ) ) {
1630 0 : FD_LOG_ERR(( "legacy transaction listen port set but no out link was found" ));
1631 0 : } else if( FD_UNLIKELY( ctx->gossip_listen_port!=0 && ctx->gossvf_out->mcache==NULL ) ) {
1632 0 : FD_LOG_ERR(( "gossip listen port set but no out link was found" ));
1633 0 : } else if( FD_UNLIKELY( ctx->repair_client_listen_port!=0 && ctx->repair_out->mcache==NULL ) ) {
1634 0 : FD_LOG_ERR(( "repair intake port set but no out link was found" ));
1635 0 : } else if( FD_UNLIKELY( ctx->repair_serve_listen_port!=0 && ctx->repair_out->mcache==NULL ) ) {
1636 0 : FD_LOG_ERR(( "repair serve listen port set but no out link was found" ));
1637 0 : } else if( FD_UNLIKELY( ctx->neigh4_solicit->mcache==NULL ) ) {
1638 0 : FD_LOG_ERR(( "netlink request link not found" ));
1639 0 : } else if( FD_UNLIKELY( ctx->txsend_src_port!=0 && ctx->txsend_out->mcache==NULL ) ) {
1640 0 : FD_LOG_ERR(( "txsend listen port set but no out link was found" ));
1641 0 : }
1642 :
1643 0 : for( uint j=0U; j<2U; j++ ) {
1644 0 : ctx->tx_flusher[ j ].pending_wmark = (ulong)( (double)tile->xdp.xdp_tx_queue_size * 0.7 );
1645 0 : ctx->tx_flusher[ j ].tail_flush_backoff = (long)( (double)tile->xdp.tx_flush_timeout_ns * fd_tempo_tick_per_ns( NULL ) );
1646 0 : ctx->tx_flusher[ j ].next_tail_flush_ticks = LONG_MAX;
1647 :
1648 0 : ctx->tx_flusher[ j ].prefbusy_last_poll_ticks = 0L;
1649 0 : ctx->tx_flusher[ j ].prefbusy_min_interval_ticks = (long)( PREFBUSY_MIN_INTERVAL_NS * fd_tempo_tick_per_ns( NULL ) );
1650 0 : ctx->tx_flusher[ j ].prefbusy_stall_timeout_ticks = (long)( PREFBUSY_STALL_TIMEOUT_NS * fd_tempo_tick_per_ns( NULL ) );
1651 0 : }
1652 :
1653 : /* Join netbase objects */
1654 0 : FD_TEST( fd_fib4_join( ctx->fib_local, fd_topo_obj_laddr( topo, tile->xdp.fib4_local_obj_id ) ) );
1655 0 : FD_TEST( fd_fib4_join( ctx->fib_main, fd_topo_obj_laddr( topo, tile->xdp.fib4_main_obj_id ) ) );
1656 :
1657 0 : ulong neigh4_obj_id = tile->xdp.neigh4_obj_id;
1658 0 : ulong ele_max = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "obj.%lu.ele_max", neigh4_obj_id );
1659 0 : ulong probe_max = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "obj.%lu.probe_max", neigh4_obj_id );
1660 0 : ulong seed = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "obj.%lu.seed", neigh4_obj_id );
1661 0 : if( FD_UNLIKELY( (ele_max==ULONG_MAX) | (probe_max==ULONG_MAX) | (seed==ULONG_MAX) ) )
1662 0 : FD_LOG_ERR(( "neigh4 hmap properties not set" ));
1663 0 : if( FD_UNLIKELY( !fd_neigh4_hmap_join(
1664 0 : ctx->neigh4,
1665 0 : fd_topo_obj_laddr( topo, neigh4_obj_id ),
1666 0 : ele_max,
1667 0 : probe_max,
1668 0 : seed ) ) ) {
1669 0 : FD_LOG_ERR(( "fd_neigh4_hmap_join failed" ));
1670 0 : }
1671 :
1672 0 : init_device_table( ctx, fd_topo_obj_laddr( topo, tile->xdp.netdev_tbl_obj_id ), netdev_tbl_local );
1673 :
1674 : /* Initialize TX free ring */
1675 :
1676 0 : ulong const frame_sz = 2048UL;
1677 0 : ulong frame_off = 0UL;
1678 0 : ulong const tx_depth = ctx->free_tx.depth;
1679 0 : for( ulong j=0; j<tx_depth; j++ ) {
1680 0 : ctx->free_tx.queue[ j ] = (ulong)ctx->umem + frame_off;
1681 0 : frame_off += frame_sz;
1682 0 : }
1683 0 : ctx->free_tx.prod = tx_depth;
1684 :
1685 : /* Initialize RX mcache chunks */
1686 :
1687 0 : for( ulong i=0UL; i<(tile->out_cnt); i++ ) {
1688 0 : fd_topo_link_t const * out_link = &topo->links[ tile->out_link_id[ i ] ];
1689 0 : fd_frag_meta_t * mcache = out_link->mcache;
1690 0 : for( ulong j=0UL; j<fd_mcache_depth( mcache ); j++ ) {
1691 0 : mcache[ j ].chunk = (uint)( ctx->umem_chunk0 + (frame_off>>FD_CHUNK_LG_SZ) );
1692 0 : frame_off += frame_sz;
1693 0 : }
1694 0 : }
1695 :
1696 : /* Initialize FILL ring */
1697 :
1698 0 : int _charge_busy = 0;
1699 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
1700 0 : frame_off = net_xsk_bootstrap( ctx, j, frame_off );
1701 0 : net_rx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
1702 0 : net_tx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
1703 0 : }
1704 :
1705 0 : if( FD_UNLIKELY( frame_off > ctx->umem_sz ) ) {
1706 0 : FD_LOG_ERR(( "UMEM is too small" ));
1707 0 : }
1708 0 : }
1709 :
1710 : FD_FN_UNUSED static ulong
1711 : populate_allowed_seccomp( fd_topo_t const * topo,
1712 : fd_topo_tile_t const * tile,
1713 : ulong out_cnt,
1714 0 : struct sock_filter * out ) {
1715 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1716 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1717 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_net_ctx_t ), sizeof( fd_net_ctx_t ) );
1718 :
1719 : /* A bit of a hack, if there is no loopback XSK for this tile, we still need to pass
1720 : two "allow" FD arguments to the net policy, so we just make them both the same. */
1721 0 : int allow_fd2 = ctx->xsk_cnt>1UL ? ctx->xsk[ 1 ].xsk_fd : ctx->xsk[ 0 ].xsk_fd;
1722 0 : FD_TEST( ctx->xsk[ 0 ].xsk_fd >= 0 && allow_fd2 >= 0 );
1723 :
1724 0 : populate_sock_filter_policy_fd_xdp_tile( out_cnt, out, (uint)fd_log_private_logfile_fd(), (uint)ctx->xsk[ 0 ].xsk_fd, (uint)allow_fd2 );
1725 0 : return sock_filter_policy_fd_xdp_tile_instr_cnt;
1726 0 : }
1727 :
1728 : FD_FN_UNUSED static ulong
1729 : populate_allowed_fds( fd_topo_t const * topo,
1730 : fd_topo_tile_t const * tile,
1731 : ulong out_fds_cnt,
1732 0 : int * out_fds ) {
1733 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1734 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1735 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_net_ctx_t ), sizeof( fd_net_ctx_t ) );
1736 :
1737 0 : if( FD_UNLIKELY( out_fds_cnt<6UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
1738 :
1739 0 : ulong out_cnt = 0UL;
1740 :
1741 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
1742 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
1743 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
1744 :
1745 0 : out_fds[ out_cnt++ ] = ctx->xsk[ 0 ].xsk_fd;
1746 0 : out_fds[ out_cnt++ ] = ctx->prog_link_fds[ 0 ];
1747 0 : if( FD_LIKELY( ctx->xsk_cnt>1UL ) ) out_fds[ out_cnt++ ] = ctx->xsk[ 1 ].xsk_fd;
1748 0 : if( FD_LIKELY( ctx->xsk_cnt>1UL ) ) out_fds[ out_cnt++ ] = ctx->prog_link_fds[ 1 ];
1749 0 : return out_cnt;
1750 0 : }
1751 :
1752 0 : #define STEM_BURST (1UL)
1753 0 : #define STEM_LAZY ((ulong)30e3) /* 30 us */
1754 :
1755 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_net_ctx_t
1756 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_net_ctx_t)
1757 :
1758 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
1759 0 : #define STEM_CALLBACK_DURING_HOUSEKEEPING during_housekeeping
1760 0 : #define STEM_CALLBACK_BEFORE_CREDIT before_credit
1761 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
1762 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
1763 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
1764 :
1765 : #include "../../stem/fd_stem.c"
1766 :
1767 : #ifndef FD_TILE_TEST
1768 : fd_topo_run_tile_t fd_tile_net = {
1769 : .name = "net",
1770 : .populate_allowed_seccomp = populate_allowed_seccomp,
1771 : .populate_allowed_fds = populate_allowed_fds,
1772 : .scratch_align = scratch_align,
1773 : .scratch_footprint = scratch_footprint,
1774 : .privileged_init = privileged_init,
1775 : .unprivileged_init = unprivileged_init,
1776 : .run = stem_run,
1777 : };
1778 : #endif
|
