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 : #include "../fd_net_tile_private.h"
7 :
8 : #include <errno.h>
9 : #include <fcntl.h>
10 : #include <net/if.h>
11 : #include <netinet/in.h>
12 : #include <sys/socket.h> /* MSG_DONTWAIT needed before importing the net seccomp filter */
13 : #include <linux/if_xdp.h>
14 : #include <sys/epoll.h>
15 :
16 : #include "../../metrics/fd_metrics.h"
17 : #include "../../netlink/fd_netlink_tile.h" /* neigh4_solicit */
18 : #include "../../waker/fd_waker.h"
19 : #include "../../topo/fd_topo.h"
20 :
21 : #include "../../../waltz/ip/fd_fib4.h"
22 : #include "../../../waltz/ip/fd_iproute.h"
23 : #include "../../../waltz/neigh/fd_neigh4_map.h"
24 : #include "../../../waltz/mib/fd_netdev_tbl.h"
25 : #include "../../../waltz/xdp/fd_xdp_redirect_user.h" /* fd_xsk_activate */
26 : #include "../../../waltz/xdp/fd_xsk.h"
27 : #include "../../../util/log/fd_dtrace.h"
28 : #include "../../../util/net/fd_eth.h"
29 : #include "../../../util/net/fd_ip4.h"
30 : #include "../../../util/net/fd_gre.h"
31 : #include "../../../util/pod/fd_pod_format.h"
32 :
33 : #include <unistd.h>
34 : #include <linux/if.h> /* struct ifreq */
35 : #include <sys/ioctl.h>
36 : #include <linux/if_arp.h>
37 : #include <linux/rtnetlink.h>
38 : #include <linux/futex.h>
39 :
40 : #define FD_SOCKADDR_IN_SZ sizeof(struct sockaddr_in)
41 : #include "generated/fd_xdp_tile_seccomp.h"
42 :
43 :
44 : /* FD_XDP_STATS_INTERVAL_NS controls the XDP stats refresh interval.
45 : This should be lower than the interval at which the metrics tile
46 : collects metrics. */
47 :
48 0 : #define FD_XDP_STATS_INTERVAL_NS (11e6) /* 11ms */
49 :
50 : /* XSK_IDX_{MAIN,LO} are the hardcoded XSK indices in ctx->xsk[ ... ].
51 : Only net tile 0 has XSK_IDX_LO, all net tiles have XSK_IDX_MAIN. */
52 :
53 0 : #define XSK_IDX_MAIN 0
54 0 : #define XSK_IDX_LO 1
55 :
56 : /* XSK 'busy_poll_usecs' value (max amount of time
57 : spent spinning in a NAPI poll before returning back to
58 : userspace if the processing budget hasn't already ran out).
59 :
60 : 64us chosen based on napibusy configuration tested and
61 : shown in https://lwn.net/Articles/997491/ Linux patch
62 : cover letter. Chosen over fullbusy since for Firedancer
63 : the values of fullbusy are unnecessarily high and could
64 : cause some extra latency to regular non-Firedancer traffic.*/
65 0 : #define PREFBUSY_TIME_BUDGET_MICROS (64L)
66 :
67 : /* PREFBUSY_RX_BUDGET is the NAPI RX processing budget (max num RX
68 : packets that the NIC driver can move from the hw rings into the
69 : XSK rings per poll).
70 :
71 : Default RX budget used by NIC drivers is 64, therefore
72 : it is safest to use 64 in prefbusy polling. Also reduces
73 : TX starvation risks as the TX budget set by the NIC driver
74 : is also generally 64. */
75 0 : #define PREFBUSY_RX_BUDGET (64L)
76 :
77 : /* Min time between each prefbusy poll. Necessary to avoid a no RX
78 : scenario livelocking TX with overly frequent sendto calls, given
79 : prefbusy polls whenever the RX queue is empty.
80 :
81 : Value chosen based on experimentation on ixgbe, mlx5 and i40e as well
82 : as on varying CPUs and clock speeds. Too low -> lower max TX
83 : throughput when RX is very low. Too high -> lower max RX and TX throughput. */
84 0 : #define PREFBUSY_MIN_INTERVAL_NS (5e3) /* 5us */
85 :
86 : /* Max time since last prefbusy poll before a prefbusy poll is
87 : forced (has been read that polling can sometimes resolve a stall).
88 :
89 : Exact value again not particularly important as this is just extra
90 : protection against stalls which have not been observed in testing but
91 : are still good to protect against since there is no cost to doing so.
92 :
93 : Value of 150us chosen since it is easily large enough to not interfere
94 : with standard prefbusy runtime unless there is a serious problem. */
95 0 : #define PREFBUSY_STALL_TIMEOUT_NS (150e3) /* 150us */
96 :
97 : /* fd_net_flusher_t controls the pacing of XDP sendto calls for flushing
98 : TX batches. In the 'wakeup' XDP mode, no TX occurs unless the net
99 : tile wakes up the kernel periodically using the sendto() syscall.
100 : If sendto() is called too frequently, time is wasted on context
101 : switches. If sendto() is called not often enough, packets are
102 : delayed or dropped. sendto() calls make almost no guarantees how
103 : much packets are sent out, nor do they indicate when the kernel
104 : finishes a wakeup call (asynchronously dispatched). The net tile
105 : thus uses a myriad of flush triggers that were tested for best
106 : performance. */
107 :
108 : struct fd_net_flusher {
109 :
110 : /* Packets that were enqueued after the last sendto() wakeup are
111 : considered "pending". If there are more than pending_wmark packets
112 : pending, a wakeup is dispatched. Thus, this dispatch trigger is
113 : proportional to packet rate, but does not trigger if I/O is seldom. */
114 : ulong pending_cnt;
115 : ulong pending_wmark;
116 :
117 : /* Sometimes, packets are not flushed out even after a sendto()
118 : wakeup. This can result in the tail of a burst getting delayed or
119 : overrun. If more than tail_flush_backoff ticks pass since the last
120 : sendto() wakeup and there are still unacknowledged packets in the
121 : TX ring, issues another wakeup. Only used by "softirq" poll mode. */
122 : long next_tail_flush_ticks;
123 : long tail_flush_backoff;
124 :
125 : /* When the most recent prefbusy poll was. */
126 : long prefbusy_last_poll_ticks;
127 : /* Min time between each prefbusy poll. */
128 : long prefbusy_min_interval_ticks;
129 : /* Max time since last prefbusy poll before a prefbusy poll is
130 : forced (has been read that polling can sometimes resolve a stall). */
131 : long prefbusy_stall_timeout_ticks;
132 : };
133 :
134 : typedef struct fd_net_flusher fd_net_flusher_t;
135 :
136 : FD_PROTOTYPES_BEGIN
137 :
138 : /* fd_net_flusher_inc marks a new packet as enqueued. */
139 :
140 : static inline void
141 : fd_net_flusher_inc( fd_net_flusher_t * flusher,
142 0 : long now ) {
143 0 : flusher->pending_cnt++;
144 0 : long next_flush = now + flusher->tail_flush_backoff;
145 0 : flusher->next_tail_flush_ticks = fd_long_min( flusher->next_tail_flush_ticks, next_flush );
146 0 : }
147 :
148 : /* fd_net_flusher_check returns 1 if a sendto() wakeup should be issued
149 : immediately. now is a recent fd_tickcount() value.
150 : If tx_ring_empty!=0 then the kernel is caught up with the net tile
151 : on the XDP TX ring. (Otherwise, the kernel is behind the net tile) */
152 :
153 : static inline int
154 : fd_net_flusher_check( fd_net_flusher_t * flusher,
155 : long now,
156 0 : int tx_ring_empty ) {
157 0 : int flush_level = flusher->pending_cnt >= flusher->pending_wmark;
158 0 : int flush_timeout = now >= flusher->next_tail_flush_ticks;
159 0 : int flush = flush_level || flush_timeout;
160 0 : if( !flush ) return 0;
161 0 : if( FD_UNLIKELY( tx_ring_empty ) ) {
162 : /* Flush requested but caught up */
163 0 : flusher->pending_cnt = 0UL;
164 0 : flusher->next_tail_flush_ticks = LONG_MAX;
165 0 : return 0;
166 0 : }
167 0 : return 1;
168 0 : }
169 :
170 : /* fd_net_flusher_wakeup signals a sendto() wakeup was done. now is a
171 : recent fd_tickcount() value. */
172 :
173 : static inline void
174 : fd_net_flusher_wakeup( fd_net_flusher_t * flusher,
175 0 : long now ) {
176 0 : flusher->pending_cnt = 0UL;
177 0 : flusher->next_tail_flush_ticks = now + flusher->tail_flush_backoff;
178 0 : }
179 :
180 : FD_PROTOTYPES_END
181 :
182 : /* fd_net_free_ring is a FIFO queue that stores pointers to free XDP TX
183 : frames. */
184 :
185 : struct fd_net_free_ring {
186 : ulong prod;
187 : ulong cons;
188 : ulong depth;
189 : ulong * queue;
190 : };
191 : typedef struct fd_net_free_ring fd_net_free_ring_t;
192 :
193 : typedef struct {
194 : fd_net_tile_t net;
195 :
196 : /* An "XSK" is an AF_XDP socket */
197 : uint xsk_cnt;
198 : fd_xsk_t xsk[ 2 ];
199 : int prog_link_fds[ 2 ];
200 :
201 : /* UMEM frame region within dcache */
202 : void * umem; /* Start of UMEM */
203 : ulong umem_sz; /* Size of UMEM */
204 :
205 : /* Details pertaining to an inflight send op */
206 : struct {
207 : uint xsk_idx;
208 : void * frame;
209 : } tx_op;
210 :
211 : /* Round-robin cycle service operations */
212 : uint rr_idx;
213 :
214 : /* Ring tracking free packet buffers */
215 : fd_net_free_ring_t free_tx;
216 :
217 : uchar src_mac_addr[6];
218 :
219 : /* XDP stats refresh timer */
220 : long xdp_stats_interval_ticks;
221 : long next_xdp_stats_refresh;
222 :
223 : /* TX flush timers */
224 : fd_net_flusher_t tx_flusher[2]; /* one per XSK */
225 : ulong waker_client_idx;
226 : ulong * waker_fseq;
227 :
228 : ulong netdev_seq; /* netdev_shared seqlock value of the net.router.netdev_tbl copy */
229 :
230 : struct {
231 : ulong rx_fill_blocked_cnt;
232 : long rx_busy_cnt;
233 : long rx_idle_cnt;
234 :
235 : ulong tx_submit_cnt;
236 : ulong tx_full_fail_cnt;
237 : long tx_busy_cnt;
238 : long tx_idle_cnt;
239 :
240 : ulong xsk_tx_wakeup_cnt;
241 : ulong xsk_rx_wakeup_cnt;
242 : } metrics;
243 : } fd_net_ctx_t;
244 :
245 : fd_fib4_t *
246 : fd_xdp_tile_fib4_join( fd_fib4_t * out,
247 : fd_topo_t const * topo,
248 : fd_topo_tile_t const * tile,
249 0 : int main_table ) {
250 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
251 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
252 0 : (void)FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
253 0 : (void)FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth*sizeof(ulong) );
254 0 : (void)FD_SCRATCH_ALLOC_APPEND( l, fd_netdev_tbl_align(), fd_netdev_tbl_footprint( NETDEV_MAX, BOND_MASTER_MAX ) );
255 0 : void * local_mem = FD_SCRATCH_ALLOC_APPEND( l, fd_fib4_align(), fd_fib4_footprint( tile->xdp.route_max, tile->xdp.route_peer_max ) );
256 0 : void * main_mem = FD_SCRATCH_ALLOC_APPEND( l, fd_fib4_align(), fd_fib4_footprint( tile->xdp.route_max, tile->xdp.route_peer_max ) );
257 0 : return fd_fib4_join( out, main_table ? main_mem : local_mem );
258 0 : }
259 :
260 : FD_FN_CONST static inline ulong
261 0 : scratch_align( void ) {
262 0 : return 4096UL;
263 0 : }
264 :
265 : FD_FN_PURE static inline ulong
266 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
267 0 : ulong l = FD_LAYOUT_INIT;
268 0 : l = FD_LAYOUT_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
269 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );
270 0 : l = FD_LAYOUT_APPEND( l, fd_netdev_tbl_align(), fd_netdev_tbl_footprint( NETDEV_MAX, BOND_MASTER_MAX ) );
271 0 : for( ulong i=0UL; i<2UL; i++ ) l = FD_LAYOUT_APPEND( l, fd_fib4_align(), fd_fib4_footprint( tile->xdp.route_max, tile->xdp.route_peer_max ) );
272 0 : return FD_LAYOUT_FINI( l, scratch_align() );
273 0 : }
274 :
275 : static void
276 0 : metrics_write( fd_net_ctx_t * ctx ) {
277 0 : FD_MCNT_SET( NET, PKT_RX, ctx->net.metrics.rx_pkt_cnt );
278 0 : FD_MCNT_SET( NET, PKT_RX_BYTES, ctx->net.metrics.rx_bytes_total );
279 0 : FD_MCNT_ENUM_COPY( NET, PKT_RX_DROPPED, ctx->net.metrics.rx_drop_cnt );
280 0 : FD_MCNT_SET( NET, GRE_PKT_RX, ctx->net.metrics.rx_gre_cnt );
281 0 : FD_MCNT_SET( NET, VLAN_PKT_RX, ctx->net.metrics.rx_vlan_cnt );
282 0 : FD_MCNT_SET( NET, RX_FILL_RING_STALLED, ctx->metrics.rx_fill_blocked_cnt );
283 0 : FD_MGAUGE_SET( NET, RX_BUFFER_USED, (ulong)fd_long_max( ctx->metrics.rx_busy_cnt, 0L ) );
284 0 : FD_MGAUGE_SET( NET, RX_BUFFER_FREE, (ulong)fd_long_max( ctx->metrics.rx_idle_cnt, 0L ) );
285 0 : FD_MGAUGE_SET( NET, TX_BUFFER_USED, (ulong)fd_long_max( ctx->metrics.tx_busy_cnt, 0L ) );
286 0 : FD_MGAUGE_SET( NET, TX_BUFFER_FREE, (ulong)fd_long_max( ctx->metrics.tx_idle_cnt, 0L ) );
287 :
288 0 : FD_MCNT_SET( NET, PKT_TX_SUBMITTED, ctx->metrics.tx_submit_cnt );
289 0 : FD_MCNT_SET( NET, PKT_TX_COMPLETED, ctx->net.metrics.tx_pkt_cnt );
290 0 : FD_MCNT_SET( NET, PKT_TX_SUBMITTED_BYTES, ctx->net.metrics.tx_bytes_total );
291 0 : ctx->net.metrics.tx_drop_cnt[ FD_METRICS_ENUM_PKT_TX_DROP_REASON_V_RING_FULL_IDX ] = ctx->metrics.tx_full_fail_cnt;
292 0 : FD_MCNT_ENUM_COPY( NET, PKT_TX_DROPPED, ctx->net.metrics.tx_drop_cnt );
293 0 : FD_MCNT_ENUM_COPY( NET, PKT_TX_ROUTED, ctx->net.metrics.tx_route_result_cnt );
294 0 : FD_MCNT_ENUM_COPY( NET, PKT_TX_SLOW_PATH_ROUTED, ctx->net.metrics.tx_slow_path_reason_cnt );
295 :
296 0 : FD_MCNT_SET( NET, XSK_SYSCALL_TX, ctx->metrics.xsk_tx_wakeup_cnt );
297 0 : FD_MCNT_SET( NET, XSK_SYSCALL_RX, ctx->metrics.xsk_rx_wakeup_cnt );
298 :
299 0 : FD_MCNT_SET( NET, GRE_PKT_TX_SUBMITTED, ctx->net.metrics.tx_gre_cnt );
300 0 : FD_MCNT_SET( NET, PKT_TX_SLOW_PATH_SUBMITTED, ctx->net.metrics.tx_slow_path_submit_cnt );
301 : /* fd_fib4_cnt includes the synthetic throw route at index zero. */
302 0 : FD_MGAUGE_SET( NET, ROUTE_COUNT_LOCAL, fd_ulong_sat_sub( fd_fib4_cnt( ctx->net.router.fib_local ), 1UL ) );
303 0 : FD_MGAUGE_SET( NET, ROUTE_COUNT_MAIN, fd_ulong_sat_sub( fd_fib4_cnt( ctx->net.router.fib_main ), 1UL ) );
304 0 : }
305 :
306 : struct xdp_statistics_v0 {
307 : __u64 rx_dropped; /* Dropped for other reasons */
308 : __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
309 : __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
310 : };
311 :
312 : struct xdp_statistics_v1 {
313 : __u64 rx_dropped; /* Dropped for other reasons */
314 : __u64 rx_invalid_descs; /* Dropped due to invalid descriptor */
315 : __u64 tx_invalid_descs; /* Dropped due to invalid descriptor */
316 : __u64 rx_ring_full; /* Dropped due to rx ring being full */
317 : __u64 rx_fill_ring_empty_descs; /* Failed to retrieve item from fill ring */
318 : __u64 tx_ring_empty_descs; /* Failed to retrieve item from tx ring */
319 : };
320 :
321 : static void
322 0 : poll_xdp_statistics( fd_net_ctx_t * ctx ) {
323 0 : struct xdp_statistics_v1 stats = {0};
324 0 : ulong xsk_cnt = ctx->xsk_cnt;
325 0 : for( ulong j=0UL; j<xsk_cnt; j++ ) {
326 0 : struct xdp_statistics_v1 sub_stats = {0};
327 0 : uint optlen = (uint)sizeof(struct xdp_statistics_v1);
328 0 : if( FD_UNLIKELY( -1==getsockopt( ctx->xsk[ j ].xsk_fd, SOL_XDP, XDP_STATISTICS, &sub_stats, &optlen ) ) )
329 0 : FD_LOG_ERR(( "getsockopt(SOL_XDP, XDP_STATISTICS) failed: %s", strerror( errno ) ));
330 0 : if( FD_UNLIKELY( optlen!=sizeof(struct xdp_statistics_v0) &&
331 0 : optlen!=sizeof(struct xdp_statistics_v1) ) ) {
332 0 : FD_LOG_ERR(( "getsockopt(SOL_XDP, XDP_STATISTICS) returned unexpected size %u", optlen ));
333 0 : }
334 0 : stats.rx_dropped += sub_stats.rx_dropped;
335 0 : stats.rx_invalid_descs += sub_stats.rx_invalid_descs;
336 0 : stats.tx_invalid_descs += sub_stats.tx_invalid_descs;
337 0 : stats.rx_ring_full += sub_stats.rx_ring_full;
338 0 : stats.rx_fill_ring_empty_descs += sub_stats.rx_fill_ring_empty_descs;
339 0 : stats.tx_ring_empty_descs += sub_stats.tx_ring_empty_descs;
340 0 : }
341 :
342 0 : FD_MCNT_SET( NET, XDP_RX_OTHER_DROPPED, stats.rx_dropped );
343 0 : FD_MCNT_SET( NET, XDP_RX_INVALID_DESCRIPTOR_DROPPED, stats.rx_invalid_descs );
344 0 : FD_MCNT_SET( NET, XDP_TX_INVALID_DESCRIPTOR_DROPPED, stats.tx_invalid_descs );
345 0 : FD_MCNT_SET( NET, XDP_RX_RING_FULL_DROPPED, stats.rx_ring_full );
346 0 : FD_MCNT_SET( NET, XDP_RX_FILL_RING_EMPTY, stats.rx_fill_ring_empty_descs );
347 0 : FD_MCNT_SET( NET, XDP_TX_RING_EMPTY, stats.tx_ring_empty_descs );
348 0 : }
349 :
350 : /* net_is_fatal_xdp_error returns 1 if the given errno returned by an
351 : XDP API indicates a non-recoverable error code. The net tile should
352 : crash if it sees such an error so the problem does not go undetected.
353 : Otherwise, returns 0. */
354 :
355 : static int
356 0 : net_is_fatal_xdp_error( int err ) {
357 0 : return err==ESOCKTNOSUPPORT || err==EOPNOTSUPP || err==EINVAL ||
358 0 : err==EPERM;
359 0 : }
360 :
361 : /* net_tx_ready returns 1 if we can submit a job to this TX ring, and 0 otherwise.
362 : Reasons for block include:
363 : - No TX buffer is available (free ring empty)
364 : - TX ring is full
365 :
366 : tx_ring: pointer to the XDP TX ring
367 : free_ring: pointer to the free TX ring */
368 :
369 : static int
370 : net_tx_ready( fd_xdp_ring_t * tx_ring,
371 0 : fd_net_free_ring_t * free_ring ) {
372 0 : if( FD_UNLIKELY( free_ring->prod == free_ring->cons ) ) return 0; /* drop - no free buffers */
373 0 : if( FD_UNLIKELY( fd_xdp_ring_full( tx_ring ) ) ) return 0; /* drop - tx ring full */
374 0 : return 1;
375 0 : }
376 :
377 : /* net_rx_wakeup triggers xsk_recvmsg to run in the kernel. Needs to be
378 : called periodically in order to receive packets. */
379 :
380 : static void
381 : net_rx_wakeup( fd_net_ctx_t * ctx,
382 : fd_xsk_t * xsk,
383 0 : int * charge_busy ) {
384 : /* write-back local copies to fseqs */
385 0 : __atomic_store_n( xsk->ring_rx.cons, xsk->ring_rx.cached_cons, __ATOMIC_RELEASE );
386 0 : __atomic_store_n( xsk->ring_fr.prod, xsk->ring_fr.cached_prod, __ATOMIC_RELEASE );
387 0 : if( !fd_xsk_rx_need_wakeup( xsk ) ) return;
388 0 : *charge_busy = 1;
389 0 : struct msghdr _ignored[ 1 ] = { 0 };
390 0 : if( FD_UNLIKELY( -1==recvmsg( xsk->xsk_fd, _ignored, MSG_DONTWAIT ) ) ) {
391 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
392 0 : FD_LOG_ERR(( "xsk recvmsg failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
393 0 : }
394 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
395 0 : long ts = fd_log_wallclock();
396 0 : if( ts > xsk->log_suppress_until_ns ) {
397 0 : FD_LOG_WARNING(( "xsk recvmsg failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
398 0 : xsk->log_suppress_until_ns = ts + (long)1e9;
399 0 : }
400 0 : }
401 0 : }
402 0 : ctx->metrics.xsk_rx_wakeup_cnt++;
403 0 : }
404 :
405 : /* net_tx_wakeup triggers xsk_sendmsg to run in the kernel. Needs to be
406 : called periodically in order to transmit packets. Should only be called
407 : if there are unconsumed packets in Tx ring. */
408 :
409 : static void
410 : net_tx_wakeup( fd_net_ctx_t * ctx,
411 : fd_xsk_t * xsk,
412 0 : int * charge_busy ) {
413 : /* write-back local copies to fseqs */
414 0 : __atomic_store_n( xsk->ring_tx.prod, xsk->ring_tx.cached_prod, __ATOMIC_RELEASE );
415 0 : __atomic_store_n( xsk->ring_cr.cons, xsk->ring_cr.cached_cons, __ATOMIC_RELEASE );
416 0 : if( !fd_xsk_tx_need_wakeup( xsk ) ) return;
417 0 : *charge_busy = 1;
418 0 : if( FD_UNLIKELY( -1==sendto( xsk->xsk_fd, NULL, 0, MSG_DONTWAIT, NULL, 0 ) ) ) {
419 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
420 0 : FD_LOG_ERR(( "xsk sendto failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
421 0 : }
422 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
423 0 : long ts = fd_log_wallclock();
424 0 : if( ts > xsk->log_suppress_until_ns ) {
425 0 : FD_LOG_WARNING(( "xsk sendto failed xsk_fd=%d (%i-%s)", xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
426 0 : xsk->log_suppress_until_ns = ts + (long)1e9;
427 0 : }
428 0 : }
429 0 : }
430 0 : ctx->metrics.xsk_tx_wakeup_cnt++;
431 0 : }
432 :
433 : /* net_tx_periodic_wakeup does a timer based xsk_sendmsg wakeup. */
434 :
435 : static inline int
436 : net_tx_periodic_wakeup( fd_net_ctx_t * ctx,
437 : uint xsk_idx,
438 : long now,
439 0 : int * charge_busy ) {
440 0 : fd_xdp_ring_t * tx_ring = &ctx->xsk[ xsk_idx ].ring_tx;
441 0 : int tx_ring_empty = fd_xdp_ring_empty( tx_ring, FD_XDP_RING_ROLE_PROD );
442 0 : if( fd_net_flusher_check( ctx->tx_flusher+xsk_idx, now, tx_ring_empty ) ) {
443 0 : net_tx_wakeup( ctx, &ctx->xsk[ xsk_idx ], charge_busy );
444 0 : fd_net_flusher_wakeup( ctx->tx_flusher+xsk_idx, now );
445 0 : }
446 0 : return 0;
447 0 : }
448 :
449 : static void
450 0 : during_housekeeping( fd_net_ctx_t * ctx ) {
451 0 : long now = fd_tickcount();
452 0 : if( FD_UNLIKELY( fd_netdev_tbl_refresh( &ctx->net.router.netdev_tbl, &ctx->net.router.netdev_shared, &ctx->netdev_seq ) ) ) {
453 0 : fd_net_gre_tunnels_refresh( &ctx->net, &ctx->net.router.netdev_tbl );
454 0 : }
455 :
456 0 : ctx->metrics.rx_busy_cnt = 0UL;
457 0 : ctx->metrics.rx_idle_cnt = 0UL;
458 0 : ctx->metrics.tx_idle_cnt = fd_seq_diff( ctx->free_tx.prod, ctx->free_tx.cons );
459 0 : ctx->metrics.tx_busy_cnt = (long)ctx->free_tx.depth - ctx->metrics.tx_idle_cnt;
460 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
461 0 : fd_xsk_t * xsk = &ctx->xsk[ j ];
462 0 : FD_COMPILER_MFENCE();
463 : /* Write back local copies to fseqs that we own */
464 0 : __atomic_store_n( xsk->ring_fr.prod, xsk->ring_fr.cached_prod, __ATOMIC_RELEASE );
465 0 : __atomic_store_n( xsk->ring_rx.cons, xsk->ring_rx.cached_cons, __ATOMIC_RELEASE );
466 0 : __atomic_store_n( xsk->ring_tx.prod, xsk->ring_tx.cached_prod, __ATOMIC_RELEASE );
467 0 : __atomic_store_n( xsk->ring_cr.cons, xsk->ring_cr.cached_cons, __ATOMIC_RELEASE );
468 :
469 : /* Refresh kernel-owned seq numbers for accurate stats */
470 0 : xsk->ring_fr.cached_cons = __atomic_load_n( xsk->ring_fr.cons, __ATOMIC_ACQUIRE );
471 0 : xsk->ring_rx.cached_prod = __atomic_load_n( xsk->ring_rx.prod, __ATOMIC_ACQUIRE );
472 0 : xsk->ring_tx.cached_cons = __atomic_load_n( xsk->ring_tx.cons, __ATOMIC_ACQUIRE );
473 0 : xsk->ring_cr.cached_prod = __atomic_load_n( xsk->ring_cr.prod, __ATOMIC_ACQUIRE );
474 :
475 0 : FD_COMPILER_MFENCE();
476 0 : ctx->metrics.rx_busy_cnt += (long)(int)( xsk->ring_rx.cached_prod - xsk->ring_rx.cached_cons );
477 0 : ctx->metrics.rx_idle_cnt += (long)(int)( xsk->ring_fr.cached_prod - xsk->ring_fr.cached_cons );
478 0 : }
479 :
480 0 : if( now > ctx->next_xdp_stats_refresh ) {
481 0 : ctx->next_xdp_stats_refresh = now + ctx->xdp_stats_interval_ticks;
482 0 : poll_xdp_statistics( ctx );
483 0 : }
484 0 : }
485 :
486 :
487 : /* before_frag is called when a new metadata descriptor for a TX job is
488 : found. This callback determines whether this net tile is responsible
489 : for the TX job. If so, it prepares the TX op for the during_frag and
490 : after_frag callbacks. */
491 :
492 : static inline int
493 : before_frag( fd_net_ctx_t * ctx,
494 : ulong in_idx,
495 : ulong seq,
496 0 : ulong sig ) {
497 0 : (void)seq;
498 :
499 : /* Short circuit TX logic for route updates */
500 0 : if( FD_UNLIKELY( ctx->net.in_kind[ in_idx ]==FD_NET_IN_KIND_IPROUTE ) ) return 0;
501 :
502 0 : if( !fd_net_tx_select( &ctx->net, sig ) ) return 1;
503 :
504 0 : fd_net_tx_route_t * route = &ctx->net.tx_route;
505 0 : uint xsk_idx = route->use_loopback && route->if_idx!=ctx->net.router.if_virt ? XSK_IDX_LO : XSK_IDX_MAIN;
506 0 : ctx->tx_op.xsk_idx = xsk_idx;
507 :
508 0 : if( FD_UNLIKELY( xsk_idx>=ctx->xsk_cnt ) ) {
509 : /* Packet does not route to an XDP interface, so send it on net.sock */
510 0 : ctx->net.metrics.tx_slow_path_reason_cnt[ FD_METRICS_ENUM_PKT_TX_SLOW_PATH_REASON_V_UNSUPPORTED_INTERFACE_IDX ]++;
511 0 : route->routing_failure = 1;
512 0 : xsk_idx = XSK_IDX_MAIN;
513 0 : ctx->tx_op.xsk_idx = XSK_IDX_MAIN;
514 0 : }
515 :
516 : /* Skip if TX is blocked */
517 0 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
518 0 : fd_net_free_ring_t * free = &ctx->free_tx;
519 0 : if( FD_UNLIKELY( !net_tx_ready( &xsk->ring_tx, free ) ) ) {
520 0 : ctx->metrics.tx_full_fail_cnt++;
521 0 : return 1;
522 0 : }
523 :
524 : /* Allocate buffer for receive */
525 0 : ulong alloc_seq = free->cons;
526 0 : void * frame = (void *)free->queue[ alloc_seq & (free->depth-1UL) ];
527 0 : free->cons = fd_seq_inc( alloc_seq, 1UL );
528 :
529 0 : ctx->tx_op.frame = frame;
530 0 : fd_net_tx_route_result_increment( &ctx->net );
531 :
532 0 : return 0; /* continue */
533 0 : }
534 :
535 : /* during_frag stages route updates or copies an outgoing packet. */
536 :
537 : static inline void
538 : during_frag( fd_net_ctx_t * ctx,
539 : ulong in_idx,
540 : ulong seq FD_PARAM_UNUSED,
541 : ulong sig FD_PARAM_UNUSED,
542 : ulong chunk,
543 : ulong sz,
544 0 : ulong ctl FD_PARAM_UNUSED ) {
545 0 : fd_net_in_frag_validate( &ctx->net, in_idx, chunk, sz );
546 :
547 0 : if( FD_UNLIKELY( fd_net_iproute_msg_stage( &ctx->net, in_idx, chunk, sz ) ) ) return;
548 :
549 0 : void * frame = ctx->tx_op.frame;
550 0 : if( FD_UNLIKELY( (ulong)frame < (ulong)ctx->umem ) )
551 0 : FD_LOG_ERR(( "frame %p out of bounds (below %p)", frame, (void *)ctx->umem ));
552 0 : ulong umem_off = (ulong)frame - (ulong)ctx->umem;
553 0 : if( FD_UNLIKELY( (ulong)umem_off > (ulong)ctx->umem_sz ) )
554 0 : FD_LOG_ERR(( "frame %p out of bounds (beyond %p)", frame, (void *)ctx->umem_sz ));
555 :
556 : /* Speculatively copy frame into buffer from in link */
557 0 : fd_net_tx_pkt_cpy( &ctx->net, ctx->tx_op.frame, chunk, sz, in_idx );
558 0 : }
559 :
560 : /* after_frag is called when the during_frag memcpy was _not_ overrun. */
561 :
562 : static void
563 : after_frag( fd_net_ctx_t * ctx,
564 : ulong in_idx,
565 : ulong seq,
566 : ulong sig,
567 : ulong sz,
568 : ulong tsorig,
569 : ulong tspub,
570 0 : fd_stem_context_t * stem ) {
571 0 : (void)seq; (void)sig; (void)tsorig; (void)tspub; (void)stem;
572 :
573 0 : if( FD_UNLIKELY( ctx->net.in_kind[ in_idx ]==FD_NET_IN_KIND_IPROUTE ) ) {
574 0 : fd_iproute_msg_t const * msg = &ctx->net.iproute_msg;
575 0 : if( msg->op==FD_IPROUTE_OP_FLUSH ) {
576 0 : fd_fib4_clear( ctx->net.router.fib_local );
577 0 : fd_fib4_clear( ctx->net.router.fib_main );
578 0 : return;
579 0 : }
580 0 : fd_fib4_t * fib;
581 0 : if( msg->table_id==RT_TABLE_LOCAL ) fib = ctx->net.router.fib_local;
582 0 : else if( msg->table_id==RT_TABLE_MAIN ) fib = ctx->net.router.fib_main;
583 0 : else return;
584 0 : if( msg->op==FD_IPROUTE_OP_UPSERT && FD_UNLIKELY( !fd_fib4_insert( fib, msg->dst_addr, msg->prefix, msg->prio, &msg->hop ) ) ) {
585 0 : FD_LOG_WARNING(( "route update dropped: route table full (increase [tiles.netlink].max_routes or [tiles.netlink].max_peer_routes)" ));
586 0 : if( FD_UNLIKELY( ctx->net.kind_id==0UL ) ) {
587 0 : fd_stem_publish( stem, ctx->net.router.netlnk_out_idx, FD_NETLINK_ROUTE4_SYNC_SIG, 0UL, 0UL, 0UL, 0UL, fd_frag_meta_ts_comp( fd_tickcount() ) );
588 0 : }
589 0 : }
590 0 : else if( msg->op==FD_IPROUTE_OP_DELETE ) fd_fib4_remove( fib, msg->dst_addr, msg->prefix, msg->prio );
591 0 : return;
592 0 : }
593 :
594 : /* Current send operation */
595 :
596 0 : uchar * frame = ctx->tx_op.frame;
597 0 : uint xsk_idx = ctx->tx_op.xsk_idx;
598 :
599 0 : if( FD_UNLIKELY( !fd_net_tx_pkt_prep( &ctx->net, frame, &sz, in_idx ) ) ) return;
600 :
601 : /* Packets that failed userland routing are sent on net.sock starting
602 : at the IPv4 header, and the kernel builds the L2 header. The send
603 : is synchronous, so the frame goes straight back to free_tx, and a
604 : sent packet is counted as both submitted and completed (there is
605 : no XSK completion event for it). */
606 0 : if( FD_UNLIKELY( ctx->net.tx_route.routing_failure ) ) {
607 0 : int ok = fd_net_tx_sock_send( &ctx->net, frame, sz );
608 0 : fd_net_free_ring_t * free = &ctx->free_tx;
609 0 : ulong free_idx = free->prod & (free->depth-1UL);
610 0 : free->queue[ free_idx ] = (ulong)frame;
611 0 : free->prod = fd_seq_inc( free->prod, 1UL );
612 0 : ctx->tx_op.frame = NULL;
613 0 : if( FD_LIKELY( ok ) ) {
614 0 : ctx->metrics.tx_submit_cnt++;
615 0 : ctx->net.metrics.tx_pkt_cnt++;
616 0 : ctx->net.metrics.tx_bytes_total += sz;
617 0 : ctx->net.metrics.tx_gre_cnt += (ulong)ctx->net.tx_route.use_gre;
618 0 : }
619 0 : return;
620 0 : }
621 :
622 : /* Submit packet TX job
623 :
624 : Invariant for ring_tx: prod-cons<length
625 : (This invariant breaks if any other packet is sent over this ring
626 : between before_frag and this point.) */
627 :
628 0 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
629 0 : fd_xdp_ring_t * tx_ring = &xsk->ring_tx;
630 0 : uint tx_seq = tx_ring->cached_prod;
631 0 : uint tx_mask = tx_ring->depth - 1U;
632 0 : xsk->ring_tx.packet_ring[ tx_seq&tx_mask ] = (struct xdp_desc) {
633 0 : .addr = (ulong)frame - (ulong)ctx->umem,
634 0 : .len = (uint)sz,
635 0 : .options = 0
636 0 : };
637 :
638 : /* Frame is now owned by kernel. Clear tx_op. */
639 0 : ctx->tx_op.frame = NULL;
640 :
641 : /* Register newly enqueued packet */
642 0 : tx_ring->cached_prod = tx_seq+1U;
643 0 : ctx->metrics.tx_submit_cnt++;
644 0 : ctx->net.metrics.tx_bytes_total += sz;
645 0 : if( ctx->net.tx_route.use_gre ) ctx->net.metrics.tx_gre_cnt++;
646 0 : fd_net_flusher_inc( ctx->tx_flusher+xsk_idx, fd_tickcount() );
647 0 : }
648 :
649 : /* net_comp_event is called when an XDP TX frame is free again. */
650 :
651 : static void
652 : net_comp_event( fd_net_ctx_t * ctx,
653 : fd_xsk_t * xsk,
654 0 : uint comp_seq ) {
655 :
656 : /* Locate the incoming frame */
657 :
658 0 : fd_xdp_ring_t * comp_ring = &xsk->ring_cr;
659 0 : uint comp_mask = comp_ring->depth - 1U;
660 0 : ulong frame = FD_VOLATILE_CONST( comp_ring->frame_ring[ comp_seq&comp_mask ] );
661 0 : ulong const frame_mask = FD_NET_MTU - 1UL;
662 0 : FD_STATIC_ASSERT( FD_ULONG_IS_POW2( FD_NET_MTU ), "FD_NET_MTU must be a power of two" );
663 0 : if( FD_UNLIKELY( frame+FD_NET_MTU > ctx->umem_sz ) ) {
664 0 : FD_LOG_ERR(( "Bounds check failed: frame=0x%lx umem_sz=0x%lx",
665 0 : frame, (ulong)ctx->umem_sz ));
666 0 : }
667 :
668 : /* Check if we have space to return the freed frame */
669 :
670 0 : fd_net_free_ring_t * free = &ctx->free_tx;
671 0 : ulong free_prod = free->prod;
672 0 : ulong free_mask = free->depth - 1UL;
673 0 : ulong free_cons = free->cons;
674 0 : long free_cnt = fd_seq_diff( free_prod, free_cons );
675 0 : FD_TEST( free_prod >= free_cons );
676 0 : if( FD_UNLIKELY( free_cnt>=(long)free->depth ) ) return; /* blocked */
677 :
678 0 : free->queue[ free_prod&free_mask ] = (ulong)ctx->umem + (frame & (~frame_mask));
679 0 : free->prod = fd_seq_inc( free_prod, 1UL );
680 :
681 : /* Wind up for next iteration */
682 :
683 0 : comp_ring->cached_cons = comp_seq+1U;
684 0 : ctx->net.metrics.tx_pkt_cnt++;
685 0 : }
686 :
687 : static int
688 0 : prevent_park( fd_net_ctx_t * ctx ) {
689 0 : for( uint i=0U; i<ctx->xsk_cnt; i++ ) {
690 0 : fd_xsk_t * xsk = &ctx->xsk[ i ];
691 0 : if( FD_UNLIKELY( xsk->prefbusy_poll_enabled ) ) return 1;
692 0 : if( FD_UNLIKELY( !fd_xdp_ring_empty( &xsk->ring_rx, FD_XDP_RING_ROLE_CONS ) ) ) return 1;
693 0 : if( FD_UNLIKELY( !fd_xdp_ring_empty( &xsk->ring_cr, FD_XDP_RING_ROLE_CONS ) ) ) return 1;
694 0 : if( FD_UNLIKELY( ctx->tx_flusher[ i ].pending_cnt ) ) return 1;
695 0 : if( FD_UNLIKELY( !fd_xdp_ring_empty( &xsk->ring_tx, FD_XDP_RING_ROLE_PROD ) ) ) return 1; /* the flusher paces the kicks */
696 0 : }
697 0 : return 0;
698 0 : }
699 :
700 : /* net_rx_event is called when a new XDP RX frame is available. Calls
701 : fd_net_rx_pkt, then returns the packet back to the kernel via the fill
702 : ring. */
703 :
704 : static void
705 : net_rx_event( fd_net_ctx_t * ctx,
706 : fd_stem_context_t * stem,
707 : fd_xsk_t * xsk,
708 0 : uint rx_seq ) {
709 : /* Locate the incoming frame */
710 :
711 0 : fd_xdp_ring_t * rx_ring = &xsk->ring_rx;
712 0 : uint rx_mask = rx_ring->depth - 1U;
713 0 : struct xdp_desc frame = FD_VOLATILE_CONST( rx_ring->packet_ring[ rx_seq&rx_mask ] );
714 :
715 0 : if( FD_UNLIKELY( frame.len>FD_NET_MTU ) )
716 0 : FD_LOG_ERR(( "received a UDP packet with a too large payload (%u)", frame.len ));
717 :
718 : /* Check if we have space in the fill ring to free the frame */
719 :
720 0 : fd_xdp_ring_t * fill_ring = &xsk->ring_fr;
721 0 : if( FD_UNLIKELY( fd_xdp_ring_full( fill_ring ) ) ) {
722 0 : ctx->metrics.rx_fill_blocked_cnt++;
723 0 : return; /* blocked */
724 0 : }
725 :
726 : /* Pass it to the receive handler */
727 :
728 0 : ulong const chunk = ctx->net.pkt_buf_chunk0 + (frame.addr >> FD_CHUNK_LG_SZ);
729 0 : ulong const tspub = (ulong)fd_frag_meta_ts_comp( fd_tickcount() );
730 0 : ulong freed_chunk;
731 0 : fd_net_rx_pkt( &ctx->net, stem, chunk, frame.addr & (FD_CHUNK_SZ-1UL), frame.len, tspub, &freed_chunk, NULL );
732 :
733 0 : FD_COMPILER_MFENCE();
734 0 : rx_ring->cached_cons = rx_seq+1U;
735 :
736 : /* Every RX operation returns one frame to the FILL ring. If the
737 : packet was forwarded to a downstream ring, the newly shadowed frame
738 : is returned. Otherwise, the frame just received is returned. */
739 :
740 0 : FD_STATIC_ASSERT( FD_ULONG_IS_POW2( FD_NET_MTU ), "FD_NET_MTU must be a power of two" );
741 : /* Packets can start partway into a UMEM frame, so align down to
742 : the frame start. */
743 0 : ulong const frame_sz_chunks = FD_NET_MTU>>FD_CHUNK_LG_SZ;
744 0 : freed_chunk = ctx->net.pkt_buf_chunk0 + fd_ulong_align_dn( freed_chunk-ctx->net.pkt_buf_chunk0, frame_sz_chunks );
745 0 : if( FD_UNLIKELY( freed_chunk<ctx->net.pkt_buf_chunk0 || freed_chunk>ctx->net.pkt_buf_wmark ) ) {
746 0 : FD_LOG_CRIT(( "mcache corruption detected: buffer chunk %lu out of bounds [%lu,%lu]",
747 0 : freed_chunk, ctx->net.pkt_buf_chunk0, ctx->net.pkt_buf_wmark ));
748 0 : }
749 0 : ulong const freed_off = (freed_chunk-ctx->net.pkt_buf_chunk0)<<FD_CHUNK_LG_SZ;
750 :
751 0 : uint fill_prod = fill_ring->cached_prod;
752 0 : uint fill_mask = (fill_ring->depth)-1U;
753 0 : fill_ring->frame_ring[ fill_prod&fill_mask ] = freed_off;
754 0 : fill_ring->cached_prod = fill_prod+1U;
755 0 : }
756 :
757 : static void
758 : before_credit_softirq( fd_net_ctx_t * ctx,
759 : fd_stem_context_t * stem,
760 : int * charge_busy,
761 : uint rr_idx,
762 0 : fd_xsk_t * rr_xsk ) {
763 :
764 0 : net_tx_periodic_wakeup( ctx, rr_idx, fd_tickcount(), charge_busy );
765 :
766 : /* Fire RX event if we have RX desc avail */
767 0 : if( !fd_xdp_ring_empty( &rr_xsk->ring_rx, FD_XDP_RING_ROLE_CONS ) ) {
768 0 : *charge_busy = 1;
769 0 : net_rx_event( ctx, stem, rr_xsk, rr_xsk->ring_rx.cached_cons );
770 0 : } else {
771 0 : net_rx_wakeup( ctx, rr_xsk, charge_busy );
772 :
773 : /* Iterate onto the next NAPI queue. */
774 0 : ctx->rr_idx++;
775 0 : ctx->rr_idx = fd_uint_if( ctx->rr_idx>=ctx->xsk_cnt, 0, ctx->rr_idx );
776 0 : }
777 0 : }
778 :
779 : static int
780 : net_prefbusy_poll_ready( fd_xsk_t * rr_xsk,
781 : fd_net_flusher_t * flusher,
782 0 : long now ) {
783 :
784 0 : if( FD_LIKELY( now < ( flusher->prefbusy_last_poll_ticks + flusher->prefbusy_min_interval_ticks ) ) ) return 0;
785 0 : if( FD_UNLIKELY( now > ( flusher->prefbusy_last_poll_ticks + flusher->prefbusy_stall_timeout_ticks ) ) ) return 1;
786 :
787 0 : int rx_empty = fd_xdp_ring_empty( &rr_xsk->ring_rx, FD_XDP_RING_ROLE_CONS );
788 :
789 0 : return rx_empty;
790 0 : }
791 :
792 : static void
793 : net_prefbusy_poll_flush( fd_net_flusher_t * flusher,
794 0 : long now ) {
795 0 : flusher->prefbusy_last_poll_ticks = now;
796 0 : }
797 :
798 : static void
799 : before_credit_prefbusy( fd_net_ctx_t * ctx,
800 : fd_stem_context_t * stem,
801 : int * charge_busy,
802 : uint rr_idx,
803 0 : fd_xsk_t * rr_xsk ) {
804 :
805 0 : fd_net_flusher_t * flusher = ctx->tx_flusher+rr_idx;
806 0 : if( FD_UNLIKELY( net_prefbusy_poll_ready( rr_xsk, flusher, fd_tickcount() ) ) ) {
807 : /* NAPI needs to be polled to process new TX from
808 : Firedancer's net tile and process new RX from the NIC. */
809 :
810 : /* write-back local copies to fseqs */
811 0 : __atomic_store_n( rr_xsk->ring_tx.prod, rr_xsk->ring_tx.cached_prod, __ATOMIC_RELEASE );
812 0 : __atomic_store_n( rr_xsk->ring_cr.cons, rr_xsk->ring_cr.cached_cons, __ATOMIC_RELEASE );
813 0 : __atomic_store_n( rr_xsk->ring_rx.cons, rr_xsk->ring_rx.cached_cons, __ATOMIC_RELEASE );
814 0 : __atomic_store_n( rr_xsk->ring_fr.prod, rr_xsk->ring_fr.cached_prod, __ATOMIC_RELEASE );
815 :
816 0 : if( FD_UNLIKELY( -1==sendto( rr_xsk->xsk_fd, NULL, 0, MSG_DONTWAIT, NULL, 0 ) ) ) {
817 0 : if( FD_UNLIKELY( net_is_fatal_xdp_error( errno ) ) ) {
818 0 : FD_LOG_ERR(( "xsk sendto failed xsk_fd=%d (%i-%s)", rr_xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
819 0 : }
820 0 : if( FD_UNLIKELY( errno!=EAGAIN ) ) {
821 0 : long ts = fd_log_wallclock();
822 0 : if( ts > rr_xsk->log_suppress_until_ns ) {
823 0 : FD_LOG_WARNING(( "xsk sendto failed xsk_fd=%d (%i-%s)", rr_xsk->xsk_fd, errno, fd_io_strerror( errno ) ));
824 0 : rr_xsk->log_suppress_until_ns = ts + (long)1e9;
825 0 : }
826 0 : }
827 0 : }
828 : /* Since xsk sendmsg in prefbusy mode drives both rx and tx, both are incremented */
829 0 : ctx->metrics.xsk_tx_wakeup_cnt++;
830 0 : ctx->metrics.xsk_rx_wakeup_cnt++;
831 :
832 0 : net_prefbusy_poll_flush( flusher, fd_tickcount() );
833 0 : }
834 :
835 : /* Process new RX from xsk ring if there is any. */
836 0 : if( !fd_xdp_ring_empty( &rr_xsk->ring_rx, FD_XDP_RING_ROLE_CONS ) ) {
837 0 : *charge_busy = 1;
838 0 : net_rx_event( ctx, stem, rr_xsk, rr_xsk->ring_rx.cached_cons );
839 0 : }
840 : /* Iterate onto the next NAPI queue. */
841 0 : ctx->rr_idx++;
842 0 : ctx->rr_idx = fd_uint_if( ctx->rr_idx>=ctx->xsk_cnt, 0, ctx->rr_idx );
843 0 : }
844 :
845 : static void
846 : before_credit( fd_net_ctx_t * ctx,
847 : fd_stem_context_t * stem,
848 0 : int * charge_busy ) {
849 0 : fd_net_router_solicit( &ctx->net.router, stem );
850 :
851 : /* A previous send attempt was overrun. A corrupt copy of the packet was
852 : placed into an XDP frame, but the frame was not yet submitted to the
853 : TX ring. Return the tx buffer to the free list. */
854 :
855 0 : if( ctx->tx_op.frame ) {
856 0 : *charge_busy = 1;
857 0 : fd_net_free_ring_t * free = &ctx->free_tx;
858 0 : ulong alloc_seq = free->prod;
859 0 : free->queue[ alloc_seq & (free->depth-1UL) ] = (ulong)ctx->tx_op.frame;
860 0 : free->prod = fd_seq_inc( alloc_seq, 1UL );
861 0 : ctx->tx_op.frame = NULL;
862 0 : }
863 :
864 : /* Check if new packets are available or if TX frames are free again
865 : (Round-robin through sockets) */
866 :
867 0 : uint rr_idx = ctx->rr_idx;
868 0 : fd_xsk_t * rr_xsk = &ctx->xsk[ rr_idx ];
869 :
870 0 : if( FD_LIKELY( !rr_xsk->prefbusy_poll_enabled ) ) {
871 : /* Default poll mode which relies on irqs and wakeups */
872 0 : before_credit_softirq( ctx, stem, charge_busy, rr_idx, rr_xsk );
873 0 : } else {
874 0 : before_credit_prefbusy( ctx, stem, charge_busy, rr_idx, rr_xsk );
875 0 : }
876 :
877 : /* Fire comp event if we have comp desc avail */
878 0 : if( !fd_xdp_ring_empty( &rr_xsk->ring_cr, FD_XDP_RING_ROLE_CONS ) ) {
879 0 : *charge_busy = 1;
880 0 : net_comp_event( ctx, rr_xsk, rr_xsk->ring_cr.cached_cons );
881 0 : }
882 :
883 0 : if( FD_UNLIKELY( ctx->waker_fseq && fd_fseq_query( ctx->waker_fseq )==1UL ) ) {
884 0 : for( uint i=0U; i<ctx->xsk_cnt; i++ ) {
885 0 : if( FD_UNLIKELY( !fd_xdp_ring_empty( &ctx->xsk[ i ].ring_rx, FD_XDP_RING_ROLE_CONS ) ) ) return;
886 0 : if( FD_UNLIKELY( !fd_xdp_ring_empty( &ctx->xsk[ i ].ring_cr, FD_XDP_RING_ROLE_CONS ) ) ) return;
887 0 : }
888 0 : fd_fseq_update( ctx->waker_fseq, 0UL );
889 0 : fd_waker_client_rearm( ctx->waker_client_idx );
890 0 : }
891 0 : }
892 :
893 : /* net_xsk_bootstrap assigns UMEM frames to the FILL ring. */
894 :
895 : static ulong
896 : net_xsk_bootstrap( fd_net_ctx_t * ctx,
897 : uint xsk_idx,
898 0 : ulong frame_off ) {
899 0 : fd_xsk_t * xsk = &ctx->xsk[ xsk_idx ];
900 :
901 0 : ulong const frame_sz = FD_NET_MTU;
902 0 : ulong const fr_depth = ctx->xsk[ xsk_idx ].ring_fr.depth/2UL;
903 :
904 0 : fd_xdp_ring_t * fill = &xsk->ring_fr;
905 0 : uint fill_prod = fill->cached_prod;
906 0 : for( ulong j=0UL; j<fr_depth; j++ ) {
907 0 : fill->frame_ring[ j ] = frame_off;
908 0 : frame_off += frame_sz;
909 0 : }
910 0 : fill->cached_prod = fill_prod + (uint)fr_depth;
911 0 : __atomic_store_n( fill->prod, fill->cached_prod, __ATOMIC_RELEASE );
912 :
913 0 : return frame_off;
914 0 : }
915 :
916 : /* FIXME source MAC address from netlnk tile instead */
917 :
918 : static void
919 : interface_addrs( const char * interface,
920 : uchar * mac,
921 0 : uint * ip4_addr ) {
922 0 : int fd = socket( AF_INET, SOCK_DGRAM, 0 );
923 0 : struct ifreq ifr;
924 0 : ifr.ifr_addr.sa_family = AF_INET;
925 :
926 0 : strncpy( ifr.ifr_name, interface, IFNAMSIZ );
927 0 : if( FD_UNLIKELY( ioctl( fd, SIOCGIFHWADDR, &ifr ) ) )
928 0 : FD_LOG_ERR(( "could not get MAC address of interface `%s`: (%i-%s)", interface, errno, fd_io_strerror( errno ) ));
929 0 : fd_memcpy( mac, ifr.ifr_hwaddr.sa_data, 6 );
930 :
931 0 : if( FD_UNLIKELY( ioctl( fd, SIOCGIFADDR, &ifr ) ) )
932 0 : FD_LOG_ERR(( "could not get IP address of interface `%s`: (%i-%s)", interface, errno, fd_io_strerror( errno ) ));
933 0 : *ip4_addr = ((struct sockaddr_in *)fd_type_pun( &ifr.ifr_addr ))->sin_addr.s_addr;
934 :
935 0 : if( FD_UNLIKELY( close(fd) ) )
936 0 : FD_LOG_ERR(( "could not close socket (%i-%s)", errno, fd_io_strerror( errno ) ));
937 0 : }
938 :
939 : /* privileged_init does the following initialization steps:
940 :
941 : - Create an AF_XDP socket
942 : - Map XDP metadata rings
943 : - Register UMEM data region with socket
944 : - Insert AF_XDP socket into xsk_map
945 :
946 : Net tile 0 repeats these steps for the loopback device, unless the
947 : main interface is already loopback. XDP programs are installed before
948 : tiles start.
949 :
950 : Kernel object references:
951 :
952 : BPF_LINK file descriptor
953 : |
954 : +-> XDP program installation on NIC
955 : | |
956 : | +-> XDP program <-- BPF_PROG file descriptor (prog_fd)
957 : |
958 : +-> XSKMAP object <-- BPF_MAP file descriptor (xsk_map) */
959 :
960 : FD_FN_UNUSED static void
961 : privileged_init( fd_topo_t const * topo,
962 0 : fd_topo_tile_t const * tile ) {
963 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
964 :
965 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
966 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
967 0 : ulong * free_tx = FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );;
968 :
969 0 : fd_memset( ctx, 0, sizeof(fd_net_ctx_t) );
970 :
971 0 : interface_addrs( tile->xdp.if_virt, ctx->src_mac_addr, &ctx->net.router.default_address );
972 0 : ctx->net.router.if_virt = if_nametoindex( tile->xdp.if_virt ); FD_TEST( ctx->net.router.if_virt );
973 :
974 : /* Load up dcache containing UMEM */
975 :
976 0 : void * const dcache_mem = fd_topo_obj_laddr( topo, tile->net.umem_dcache_obj_id );
977 0 : void * const umem = fd_dcache_join( dcache_mem );
978 0 : ulong const umem_dcache_data_sz = fd_dcache_data_sz( umem );
979 0 : ulong const umem_frame_sz = 2048UL;
980 0 : ulong const umem_sz = fd_ulong_align_dn( umem_dcache_data_sz, umem_frame_sz );
981 :
982 : /* Derive chunk bounds */
983 :
984 0 : void * const umem_base = fd_wksp_containing( dcache_mem );
985 0 : ulong const umem_chunk0 = ( (ulong)umem - (ulong)umem_base )>>FD_CHUNK_LG_SZ;
986 0 : ulong const umem_wmark = umem_chunk0 + ( ( umem_sz-umem_frame_sz )>>FD_CHUNK_LG_SZ );
987 :
988 0 : if( FD_UNLIKELY( umem_chunk0>UINT_MAX || umem_wmark>UINT_MAX || umem_chunk0>umem_wmark ) ) {
989 0 : FD_LOG_ERR(( "Calculated invalid UMEM bounds [%lu,%lu]", umem_chunk0, umem_wmark ));
990 0 : }
991 :
992 0 : if( FD_UNLIKELY( !umem_base ) ) FD_LOG_ERR(( "UMEM dcache is not in a workspace" ));
993 :
994 0 : ctx->umem = umem;
995 0 : ctx->umem_sz = umem_sz;
996 0 : ctx->net.pkt_buf_wksp_base = umem_base;
997 0 : ctx->net.pkt_buf_chunk0 = umem_chunk0;
998 0 : ctx->net.pkt_buf_wmark = umem_wmark;
999 :
1000 0 : ctx->free_tx.queue = free_tx;
1001 0 : ctx->free_tx.depth = tile->xdp.free_ring_depth;
1002 :
1003 : /* Firedancer requires the ring be a power of two. */
1004 0 : FD_TEST( fd_ulong_is_pow2( tile->xdp.free_ring_depth ) );
1005 :
1006 : /* Create and install XSKs */
1007 :
1008 0 : uint if_phys_if_idx = if_nametoindex( tile->xdp.if_phys );
1009 0 : if( FD_UNLIKELY( !if_phys_if_idx ) ) FD_LOG_ERR(( "if_nametoindex(%s) failed", tile->xdp.if_phys ));
1010 :
1011 0 : fd_xsk_params_t params0 = {
1012 0 : .if_idx = if_phys_if_idx,
1013 0 : .if_queue_id = tile->xdp.if_queue,
1014 :
1015 : /* Some kernels produce EOPNOTSUP errors on sendto calls when
1016 : starting up without either XDP_ZEROCOPY or XDP_COPY
1017 : (e.g. 5.14.0-503.23.1.el9_5 with i40e) */
1018 0 : .bind_flags = tile->xdp.zero_copy ? XDP_ZEROCOPY : XDP_COPY,
1019 :
1020 0 : .prefbusy_time_budget_micros = PREFBUSY_TIME_BUDGET_MICROS,
1021 :
1022 0 : .prefbusy_rx_budget = PREFBUSY_RX_BUDGET,
1023 :
1024 0 : .fr_depth = tile->xdp.xdp_rx_queue_size*2,
1025 0 : .rx_depth = tile->xdp.xdp_rx_queue_size,
1026 0 : .cr_depth = tile->xdp.xdp_tx_queue_size,
1027 0 : .tx_depth = tile->xdp.xdp_tx_queue_size,
1028 :
1029 0 : .umem_addr = umem,
1030 0 : .frame_sz = umem_frame_sz,
1031 0 : .umem_sz = umem_sz,
1032 0 : };
1033 :
1034 0 : fd_cstr_ncpy( params0.poll_mode, tile->xdp.poll_mode, sizeof(params0.poll_mode) );
1035 :
1036 : /* Re-derive XDP file descriptors */
1037 :
1038 0 : fd_xdp_fds_t xdp_fds[ FD_TOPO_XDP_FDS_MAX ];
1039 0 : uint xdp_fds_cnt = FD_TOPO_XDP_FDS_MAX;
1040 0 : fd_topo_install_xdp( topo, xdp_fds, &xdp_fds_cnt, 0U, /* dry_run */ 1 );
1041 :
1042 0 : int xsk_map_fd = -1;
1043 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
1044 0 : if( xdp_fds[ i ].if_idx==if_phys_if_idx ) {
1045 0 : xsk_map_fd = xdp_fds[ i ].xsk_map_fd;
1046 0 : ctx->prog_link_fds[ 0 ] = xdp_fds[ i ].prog_link_fd;
1047 0 : xdp_fds[ i ].prog_link_fd = -1; /* mark as used */
1048 0 : break;
1049 0 : }
1050 0 : }
1051 0 : FD_TEST( xsk_map_fd>=0 );
1052 :
1053 : /* Init XSK */
1054 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 ));
1055 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 ));
1056 0 : ctx->xsk_cnt = 1;
1057 :
1058 : /* Networking tile at index 0 also binds to loopback (only queue 0 available on lo) */
1059 :
1060 0 : if( FD_UNLIKELY( strcmp( tile->xdp.if_virt, "lo" ) && !tile->kind_id ) ) {
1061 0 : ctx->xsk_cnt = 2;
1062 :
1063 0 : uint lo_idx = if_nametoindex( "lo" );
1064 0 : if( FD_UNLIKELY( !lo_idx ) ) FD_LOG_ERR(( "if_nametoindex(lo) failed" ));
1065 :
1066 0 : int lo_xsk_map_fd = -1;
1067 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
1068 0 : if( xdp_fds[ i ].if_idx==lo_idx ) {
1069 0 : lo_xsk_map_fd = xdp_fds[ i ].xsk_map_fd;
1070 0 : ctx->prog_link_fds[ 1 ] = xdp_fds[ i ].prog_link_fd;
1071 0 : xdp_fds[ i ].prog_link_fd = -1; /* mark as used */
1072 0 : break;
1073 0 : }
1074 0 : }
1075 0 : FD_TEST( lo_xsk_map_fd>=0 );
1076 :
1077 : /* init xsk 1 */
1078 0 : fd_xsk_params_t params1 = params0;
1079 0 : params1.if_idx = lo_idx; /* probably always 1 */
1080 0 : params1.if_queue_id = 0;
1081 0 : params1.bind_flags = 0;
1082 0 : if( FD_UNLIKELY( !fd_xsk_init( &ctx->xsk[ 1 ], ¶ms1 ) ) ) FD_LOG_ERR(( "failed to bind lo_xsk" ));
1083 0 : if( FD_UNLIKELY( !fd_xsk_activate( &ctx->xsk[ 1 ], lo_xsk_map_fd ) ) ) FD_LOG_ERR(( "failed to activate lo_xsk" ));
1084 0 : }
1085 :
1086 : /* Close unused XDP fds */
1087 :
1088 0 : if( FD_UNLIKELY( fd_sandbox_gettid()==fd_sandbox_getpid() ) ) {
1089 : /* Kind of gross.. in single threaded mode we don't want to close the xsk_map_fd
1090 : since it's shared with other net tiles. Just check for that by seeing if we
1091 : are the only thread in the process. */
1092 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
1093 0 : if( -1==close( xdp_fds[ i ].xsk_map_fd ) ) {
1094 0 : FD_LOG_ERR(( "close(%d) failed (%d-%s)", xsk_map_fd, errno, fd_io_strerror( errno ) ));
1095 0 : }
1096 0 : if( xdp_fds[ i ].prog_link_fd>0 &&
1097 0 : -1==close( xdp_fds[ i ].prog_link_fd ) ) {
1098 0 : FD_LOG_ERR(( "close(%d) failed (%d-%s)", xsk_map_fd, errno, fd_io_strerror( errno ) ));
1099 0 : }
1100 0 : }
1101 0 : }
1102 :
1103 0 : ctx->waker_client_idx = tile->waker_client_idx;
1104 0 : if( FD_UNLIKELY( topo->sleep_obj_id!=ULONG_MAX ) ) {
1105 0 : FD_TEST( ctx->waker_client_idx!=ULONG_MAX );
1106 0 : for( uint i=0U; i<ctx->xsk_cnt; i++ ) {
1107 0 : struct epoll_event ev = { .events = EPOLLIN, .data.fd = ctx->xsk[ i ].xsk_fd };
1108 0 : if( FD_UNLIKELY( -1==epoll_ctl( FD_WAKER_INNER_FD( ctx->waker_client_idx ), EPOLL_CTL_ADD, ctx->xsk[ i ].xsk_fd, &ev ) ) )
1109 0 : FD_LOG_ERR(( "epoll_ctl(ADD,xsk_fd) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
1110 0 : }
1111 0 : }
1112 :
1113 0 : double tick_per_ns = fd_tempo_tick_per_ns( NULL );
1114 0 : ctx->xdp_stats_interval_ticks = (long)( FD_XDP_STATS_INTERVAL_NS * tick_per_ns );
1115 :
1116 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, scratch_align() );
1117 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
1118 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
1119 :
1120 0 : if( FD_UNLIKELY( !fd_net_sock_init( &ctx->net.sock ) ) ) {
1121 0 : FD_LOG_ERR(( "fd_net_sock_init failed (%i-%s)", errno, fd_io_strerror( errno ) ));
1122 0 : }
1123 0 : }
1124 :
1125 : static void
1126 : init_device_table( fd_net_ctx_t * ctx,
1127 : void * netdev_tbl_shm,
1128 0 : void * netdev_tbl_local ) {
1129 0 : FD_TEST( fd_netdev_tbl_join( &ctx->net.router.netdev_shared, netdev_tbl_shm ) );
1130 0 : FD_TEST( fd_netdev_tbl_new( netdev_tbl_local, NETDEV_MAX, BOND_MASTER_MAX ) );
1131 0 : FD_TEST( fd_netdev_tbl_join( &ctx->net.router.netdev_tbl, netdev_tbl_local ) );
1132 0 : ctx->netdev_seq = fd_netdev_tbl_copy( &ctx->net.router.netdev_tbl, &ctx->net.router.netdev_shared );
1133 0 : fd_net_gre_tunnels_refresh( &ctx->net, &ctx->net.router.netdev_tbl );
1134 0 : }
1135 :
1136 : FD_FN_UNUSED static void
1137 : unprivileged_init( fd_topo_t const * topo,
1138 0 : fd_topo_tile_t const * tile ) {
1139 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1140 :
1141 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1142 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_net_ctx_t), sizeof(fd_net_ctx_t) );
1143 0 : FD_TEST( ctx->xsk_cnt!=0 );
1144 0 : FD_TEST( ctx->free_tx.queue!=NULL );
1145 0 : (void)FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), tile->xdp.free_ring_depth * sizeof(ulong) );
1146 0 : void * netdev_tbl_local = FD_SCRATCH_ALLOC_APPEND( l, fd_netdev_tbl_align(), fd_netdev_tbl_footprint( NETDEV_MAX, BOND_MASTER_MAX ) );
1147 0 : void * fib_local_mem = FD_SCRATCH_ALLOC_APPEND( l, fd_fib4_align(), fd_fib4_footprint( tile->xdp.route_max, tile->xdp.route_peer_max ) );
1148 0 : void * fib_main_mem = FD_SCRATCH_ALLOC_APPEND( l, fd_fib4_align(), fd_fib4_footprint( tile->xdp.route_max, tile->xdp.route_peer_max ) );
1149 0 : FD_TEST( fd_fib4_join( ctx->net.router.fib_local, fd_fib4_new( fib_local_mem, tile->xdp.route_max, tile->xdp.route_peer_max, tile->xdp.route_peer_seed ) ) );
1150 0 : FD_TEST( fd_fib4_join( ctx->net.router.fib_main, fd_fib4_new( fib_main_mem, tile->xdp.route_max, tile->xdp.route_peer_max, tile->xdp.route_peer_seed ) ) );
1151 :
1152 0 : ctx->net.kind_id = tile->kind_id;
1153 0 : ctx->net.tile_cnt = fd_topo_tile_name_cnt( topo, tile->name );
1154 :
1155 0 : ctx->net.bind_address = tile->net.bind_address;
1156 :
1157 : /* Put a bound on chunks we read from the input, to make sure they
1158 : are within in the data region of the workspace. */
1159 :
1160 0 : if( FD_UNLIKELY( !tile->in_cnt ) ) FD_LOG_ERR(( "net tile in link cnt is zero" ));
1161 0 : if( FD_UNLIKELY( tile->in_cnt>FD_NET_IN_MAX ) ) FD_LOG_ERR(( "net tile in link cnt %lu exceeds FD_NET_IN_MAX %lu", tile->in_cnt, FD_NET_IN_MAX ));
1162 0 : FD_TEST( tile->in_cnt<=32 );
1163 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
1164 0 : fd_topo_link_t const * link = &topo->links[ tile->in_link_id[ i ] ];
1165 0 : if( !strcmp( link->name, "iproute_out" ) ) ctx->net.in_kind[i] = FD_NET_IN_KIND_IPROUTE;
1166 0 : else {
1167 0 : ctx->net.in_kind[i] = FD_NET_IN_KIND_TX;
1168 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 ));
1169 0 : }
1170 :
1171 0 : ctx->net.in_dcache_ctx[ i ].wksp_base = topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ].wksp;
1172 0 : ctx->net.in_dcache_ctx[ i ].chunk0 = fd_dcache_compact_chunk0( ctx->net.in_dcache_ctx[ i ].wksp_base, link->dcache );
1173 0 : ctx->net.in_dcache_ctx[ i ].wmark = fd_dcache_compact_wmark( ctx->net.in_dcache_ctx[ i ].wksp_base, link->dcache, link->mtu );
1174 0 : }
1175 :
1176 0 : fd_net_rx_dests_init( &ctx->net, topo, tile );
1177 :
1178 0 : ctx->net.router.netlnk_out_idx = fd_topo_find_tile_out_link( topo, tile, "net_netlnk", tile->kind_id );
1179 0 : if( FD_UNLIKELY( ctx->net.router.netlnk_out_idx==ULONG_MAX ) ) FD_LOG_ERR(( "netlink request link not found" ));
1180 0 : ctx->net.router.solicit_ip = 0U;
1181 :
1182 0 : ctx->waker_fseq = NULL;
1183 0 : if( FD_UNLIKELY( tile->is_waker_client ) ) {
1184 0 : ctx->waker_fseq = fd_fseq_join( fd_topo_obj_laddr( topo, tile->waker_fseq_obj_id ) );
1185 0 : FD_TEST( ctx->waker_fseq );
1186 0 : }
1187 :
1188 0 : for( uint j=0U; j<2U; j++ ) {
1189 0 : ctx->tx_flusher[ j ].pending_wmark = (ulong)( (double)tile->xdp.xdp_tx_queue_size * 0.7 );
1190 0 : ctx->tx_flusher[ j ].tail_flush_backoff = (long)( (double)tile->xdp.tx_flush_timeout_ns * fd_tempo_tick_per_ns( NULL ) );
1191 0 : ctx->tx_flusher[ j ].next_tail_flush_ticks = LONG_MAX;
1192 :
1193 0 : ctx->tx_flusher[ j ].prefbusy_last_poll_ticks = 0L;
1194 0 : ctx->tx_flusher[ j ].prefbusy_min_interval_ticks = (long)( PREFBUSY_MIN_INTERVAL_NS * fd_tempo_tick_per_ns( NULL ) );
1195 0 : ctx->tx_flusher[ j ].prefbusy_stall_timeout_ticks = (long)( PREFBUSY_STALL_TIMEOUT_NS * fd_tempo_tick_per_ns( NULL ) );
1196 0 : }
1197 :
1198 0 : ulong neigh4_obj_id = tile->xdp.neigh4_obj_id;
1199 0 : ulong ele_max = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "obj.%lu.ele_max", neigh4_obj_id );
1200 0 : ulong probe_max = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "obj.%lu.probe_max", neigh4_obj_id );
1201 0 : ulong seed = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "obj.%lu.seed", neigh4_obj_id );
1202 0 : if( FD_UNLIKELY( (ele_max==ULONG_MAX) | (probe_max==ULONG_MAX) | (seed==ULONG_MAX) ) )
1203 0 : FD_LOG_ERR(( "neigh4 hmap properties not set" ));
1204 0 : if( FD_UNLIKELY( !fd_neigh4_hmap_join(
1205 0 : ctx->net.router.neigh4,
1206 0 : fd_topo_obj_laddr( topo, neigh4_obj_id ),
1207 0 : ele_max,
1208 0 : probe_max,
1209 0 : seed ) ) ) {
1210 0 : FD_LOG_ERR(( "fd_neigh4_hmap_join failed" ));
1211 0 : }
1212 :
1213 0 : init_device_table( ctx, fd_topo_obj_laddr( topo, tile->xdp.netdev_tbl_obj_id ), netdev_tbl_local );
1214 :
1215 : /* Initialize TX free ring */
1216 :
1217 0 : ulong const frame_sz = 2048UL;
1218 0 : ulong frame_off = 0UL;
1219 0 : ulong const tx_depth = ctx->free_tx.depth;
1220 0 : for( ulong j=0; j<tx_depth; j++ ) {
1221 0 : ctx->free_tx.queue[ j ] = (ulong)ctx->umem + frame_off;
1222 0 : frame_off += frame_sz;
1223 0 : }
1224 0 : ctx->free_tx.prod = tx_depth;
1225 :
1226 : /* Initialize RX mcache chunks */
1227 :
1228 0 : for( ulong i=0UL; i<(tile->out_cnt); i++ ) {
1229 0 : fd_topo_link_t const * out_link = &topo->links[ tile->out_link_id[ i ] ];
1230 0 : fd_frag_meta_t * mcache = out_link->mcache;
1231 0 : for( ulong j=0UL; j<fd_mcache_depth( mcache ); j++ ) {
1232 0 : mcache[ j ].chunk = (uint)( ctx->net.pkt_buf_chunk0 + (frame_off>>FD_CHUNK_LG_SZ) );
1233 0 : frame_off += frame_sz;
1234 0 : }
1235 0 : }
1236 :
1237 : /* Initialize FILL ring */
1238 :
1239 0 : int _charge_busy = 0;
1240 0 : for( uint j=0U; j<ctx->xsk_cnt; j++ ) {
1241 0 : frame_off = net_xsk_bootstrap( ctx, j, frame_off );
1242 0 : net_rx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
1243 0 : net_tx_wakeup( ctx, &ctx->xsk[ j ], &_charge_busy );
1244 0 : }
1245 :
1246 0 : if( FD_UNLIKELY( frame_off > ctx->umem_sz ) ) {
1247 0 : FD_LOG_ERR(( "UMEM is too small" ));
1248 0 : }
1249 0 : }
1250 :
1251 : FD_FN_UNUSED static ulong
1252 : populate_allowed_seccomp( fd_topo_t const * topo,
1253 : fd_topo_tile_t const * tile,
1254 : ulong out_cnt,
1255 0 : struct sock_filter * out ) {
1256 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1257 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1258 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_net_ctx_t ), sizeof( fd_net_ctx_t ) );
1259 :
1260 : /* A bit of a hack, if there is no loopback XSK for this tile, we still need to pass
1261 : two "allow" FD arguments to the net policy, so we just make them both the same. */
1262 0 : int allow_fd2 = ctx->xsk_cnt>1UL ? ctx->xsk[ 1 ].xsk_fd : ctx->xsk[ 0 ].xsk_fd;
1263 0 : FD_TEST( ctx->xsk[ 0 ].xsk_fd >= 0 && allow_fd2 >= 0 && ctx->net.sock.sock >= 0 );
1264 :
1265 : /* not a waker client in performance mode: no fd matches, epoll_ctl denied */
1266 0 : uint epoll_inner_fd = tile->is_waker_client ? (uint)FD_WAKER_INNER_FD( tile->waker_client_idx ) : (uint)-1;
1267 0 : uint epoll_outer_fd = tile->is_waker_client ? (uint)FD_WAKER_OUTER_FD : (uint)-1;
1268 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,
1269 0 : (uint)ctx->net.sock.sock, epoll_inner_fd, epoll_outer_fd );
1270 0 : return sock_filter_policy_fd_xdp_tile_instr_cnt;
1271 0 : }
1272 :
1273 : FD_FN_UNUSED static ulong
1274 : populate_allowed_fds( fd_topo_t const * topo,
1275 : fd_topo_tile_t const * tile,
1276 : ulong out_fds_cnt,
1277 0 : int * out_fds ) {
1278 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1279 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1280 0 : fd_net_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_net_ctx_t ), sizeof( fd_net_ctx_t ) );
1281 :
1282 0 : if( FD_UNLIKELY( out_fds_cnt<9UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
1283 :
1284 0 : ulong out_cnt = 0UL;
1285 :
1286 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
1287 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
1288 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
1289 :
1290 0 : out_fds[ out_cnt++ ] = ctx->xsk[ 0 ].xsk_fd;
1291 0 : out_fds[ out_cnt++ ] = ctx->prog_link_fds[ 0 ];
1292 0 : out_fds[ out_cnt++ ] = ctx->net.sock.sock; /* routing fallback socket */
1293 0 : if( FD_LIKELY( ctx->xsk_cnt>1UL ) ) out_fds[ out_cnt++ ] = ctx->xsk[ 1 ].xsk_fd;
1294 0 : if( FD_LIKELY( ctx->xsk_cnt>1UL ) ) out_fds[ out_cnt++ ] = ctx->prog_link_fds[ 1 ];
1295 0 : if( FD_UNLIKELY( tile->is_waker_client ) ) {
1296 0 : out_fds[ out_cnt++ ] = FD_WAKER_OUTER_FD; /* waker outer epoll fd (rearm) */
1297 0 : out_fds[ out_cnt++ ] = FD_WAKER_INNER_FD( tile->waker_client_idx ); /* waker inner epoll fd */
1298 0 : }
1299 0 : return out_cnt;
1300 0 : }
1301 :
1302 0 : #define STEM_BURST (1UL)
1303 0 : #define STEM_LAZY ((ulong)30e3) /* 30 us */
1304 :
1305 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_net_ctx_t
1306 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_net_ctx_t)
1307 :
1308 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
1309 0 : #define STEM_CALLBACK_DURING_HOUSEKEEPING during_housekeeping
1310 : #define STEM_CALLBACK_PREVENT_PARK prevent_park
1311 0 : #define STEM_CALLBACK_BEFORE_CREDIT before_credit
1312 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
1313 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
1314 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
1315 :
1316 : #include "../../stem/fd_stem.c"
1317 :
1318 : #ifndef FD_TILE_TEST
1319 : fd_topo_run_tile_t fd_tile_net = {
1320 : .name = "net",
1321 : .populate_allowed_seccomp = populate_allowed_seccomp,
1322 : .populate_allowed_fds = populate_allowed_fds,
1323 : .scratch_align = scratch_align,
1324 : .scratch_footprint = scratch_footprint,
1325 : .privileged_init = privileged_init,
1326 : .unprivileged_init = unprivileged_init,
1327 : .run = stem_run,
1328 : };
1329 : #endif
|
