Line data Source code
1 : #include "fd_stem.h"
2 :
3 : /* fd_stem provides services to multiplex multiple streams of input
4 : fragments and present them to a mix of reliable and unreliable
5 : consumers as though they were generated by multiple different
6 : multi-stream producers. The code can be included to generate
7 : a definition of stem_run which can be called as a tile main run
8 : loop.
9 :
10 : The template supports various callback functions which can be
11 : defined like #define STEM_CALLBACK_BEFORE_FRAG before_frag to
12 : tune the behavior of the stem_run loop. The callbacks are:
13 :
14 : SHOULD_SHUTDOWN
15 : It is called at the beginning of each iteration of the stem run loop,
16 : and if it returns non-zero, the stem will exit the run loop and
17 : return from the stem_run function. This is useful for shutting down
18 : the tile.
19 :
20 : DURING_HOUSEKEEPING
21 : Is called during the housekeeping routine, which happens infrequently
22 : on a schedule determined by the stem (based on the lazy parameter,
23 : see fd_tempo.h for more information). It is appropriate to do
24 : slightly expensive things here that wouldn't be OK to do in the main
25 : loop, like updating sequence numbers that are shared with other tiles
26 : (e.g. synchronization information), or sending batched information
27 : somewhere. The ctx is a user-provided context object from when the
28 : stem was initialized.
29 :
30 : METRICS_WRITE
31 : By convention, tiles may wish to accumulate high traffic metrics
32 : locally so they don't cause a lot of cache coherency traffic, and
33 : then periodically publish them to external observers. This callback
34 : is here to support that use case. It occurs infrequently during the
35 : housekeeping loop, and is called inside a compiler fence to ensure
36 : the writes do not get reordered, which may be important for observers
37 : or monitoring tools. The ctx is a user-provided context object from
38 : when the stem tile was initialized.
39 :
40 : BEFORE_CREDIT
41 : Is called every iteration of the stem run loop, whether there is a
42 : new frag ready to receive or not. This callback is also still
43 : invoked even if the stem is backpressured and cannot read any new
44 : fragments while waiting for downstream consumers to catch up. This
45 : callback is useful for things that need to occur even if no new frags
46 : are being handled. For example, servicing network connections could
47 : happen here. The ctx is a user-provided context object from when the
48 : stem tile was initialized. The stem is the stem which is invoking
49 : this callback. The stem should only be used for calling
50 : fd_stem_publish to publish a fragment to downstream consumers.
51 :
52 : The charge_busy argument is 0 by default, and should be set to 1 if
53 : the before_credit function is doing work that should be accounted for
54 : as part of the tiles busy indicator.
55 :
56 : AFTER_CREDIT
57 : Is called every iteration of the stem run loop, whether there is a
58 : new frag ready to receive or not, except in cases where the stem is
59 : backpressured by a downstream consumer and would not be able to
60 : publish. The callback might be used for publishing new fragments to
61 : downstream consumers in the main loop which are not in response to an
62 : incoming fragment. For example, code that collects incoming
63 : fragments over a period of 1 second and joins them together before
64 : publishing a large block fragment downstream, would publish the block
65 : here. The ctx is a user-provided context object from when the stem
66 : tile was initialized. The stem is the stem which is invoking this
67 : callback. The stem should only be used for calling fd_stem_publish to
68 : publish a fragment to downstream consumers.
69 :
70 : The opt_poll_in argument determines if the stem should proceed with
71 : checking for new fragments to consumer, or should `continue` the main
72 : stem loop to do credit checking again. This could be used if the
73 : after_credit function publishes, and the flow control needs to be
74 : checked again. By default, opt_poll_in is true and the stem will
75 : poll for fragments right away without rerunning the loop or checking
76 : for credits.
77 :
78 : The charge_busy argument is 0 by default, and should be set to 1 if
79 : the after_credit function is doing work that should be accounted for
80 : as part of the tiles busy indicator.
81 :
82 : BEFORE_FRAG
83 : Is called immediately whenever a new fragment has been detected that
84 : was published by an upstream producer. The signature and sequence
85 : number (sig and seq) provided as arguments are read atomically from
86 : shared memory, so must both match each other from the published
87 : fragment (aka. they will not be torn or partially overwritten).
88 : in_idx is an index in [0, num_ins) indicating which producer
89 : published the fragment. No fragment data has been read yet here, nor
90 : has other metadata, for example the size or timestamps of the
91 : fragment. Mainly this callback is useful for deciding whether to
92 : filter the fragment based on its signature. If the return value is
93 : non-zero, the frag will be skipped completely, no fragment data will
94 : be read, and the in will be advanced so that we now wait for the next
95 : fragment. If the return value is -1, then the frag is returned back
96 : to the message queue and will be reprocessed. The ctx is a
97 : user-provided context object from when the stem tile was initialized.
98 :
99 : DURING_FRAG
100 : Is called after the stem has received a new frag from an in, but
101 : before the stem has checked that it was overrun. This callback is
102 : not invoked if the stem is backpressured, as it would not try and
103 : read a frag from an in in the first place (instead, leaving it on the
104 : in mcache to backpressure the upstream producer). in_idx will be the
105 : index of the in that the frag was received from. If the producer of
106 : the frags is respecting flow control, it is safe to read frag data in
107 : any of the callbacks, but it is suggested to copy or read frag data
108 : within this callback, as if the producer does not respect flow
109 : control, the frag may be torn or corrupt due to an overrun by the
110 : reader. If the frag being read from has been overwritten while this
111 : callback is running, the frag will be ignored and the stem will not
112 : call the after_frag function. Instead it will recover from the
113 : overrun and continue with new frags. This function cannot fail. The
114 : ctx is a user-provided context object from when the stem tile was
115 : initialized. seq, sig, chunk, and sz are the respective fields from
116 : the mcache fragment that was received. If the producer is not
117 : respecting flow control, these may be corrupt or torn and should not
118 : be trusted, except for seq which is read atomically.
119 :
120 : RETURNABLE_FRAG
121 : Is called after the stem has received a new frag from an in, and
122 : assumes that the stem cannot be overrun. This special callback can
123 : instruct the stem not to advance the input sequence number, and
124 : instead return the fragment to the stem to be processed again. This
125 : is useful for processing partial data from fragments without copying
126 : it. This callback is unsafe in general contexts, since it assumes
127 : that the frag will not be overwritten while the callback is running,
128 : and that the frag data is valid throughout the function call. It
129 : should only be used when the stem is guaranteed to not be overrun.
130 : This callback is not invoked if the stem is backpressured, as it
131 : would not try and read a frag from an in in the first place (instead,
132 : leaving it on the in mcache to backpressure the upstream producer).
133 : in_idx will be the index of the in that the frag was received from.
134 : seq, sig, chunk, and sz are the respective fields from the mcache
135 : fragment that was received. tsorig and tspub are the timestamps of
136 : the fragment that was received, and are read atomically from shared
137 : memory, so must both match each other from the published fragment
138 : (aka. they will not be torn or partially overwritten). The ctx is a
139 : user-provided context object from when the stem tile was initialized.
140 : The callback should return 1 if the fragment was not fully processed
141 : and should be returned to the stem for further processing, or 0 if
142 : the fragment was fully processed and the consumer link should be
143 : advanced.
144 :
145 : AFTER_FRAG
146 : Is called immediately after the DURING_FRAG, along with an additional
147 : check that the reader was not overrun while handling the frag. If
148 : the reader was overrun, the frag is abandoned and this function is
149 : not called. This callback is not invoked if the stem is
150 : backpressured, as it would not read a frag in the first place.
151 : in_idx will be the index of the in that the frag was received from.
152 : You should not read the frag data directly here, as it might still
153 : get overrun, instead it should be copied out of the frag during the
154 : read callback if needed later. This function cannot fail. The ctx is
155 : a user-provided context object from when the stem tile was
156 : initialized. stem should only be used for calling fd_stem_publish to
157 : publish a fragment to downstream consumers. seq is the sequence
158 : number of the fragment that was read from the input mcache. sig,
159 : chunk, sz, tsorig, and tspub are the respective fields from the
160 : mcache fragment that was received. If the producer is not respecting
161 : flow control, these may be corrupt or torn and should not be trusted.
162 :
163 : AFTER_POLL_OVERRUN
164 : Is called when an overrun is detected while polling for new frags.
165 : This callback is not called when an overrun is detected in
166 : during_frag. */
167 :
168 : #if !FD_HAS_ALLOCA
169 : #error "fd_stem requires alloca"
170 : #endif
171 :
172 : #include "../topo/fd_topo.h"
173 : #include "../metrics/fd_metrics.h"
174 : #include "../../tango/fd_tango.h"
175 :
176 : #ifndef STEM_NAME
177 : #define STEM_NAME stem
178 : #endif
179 0 : #define STEM_(n) FD_EXPAND_THEN_CONCAT3(STEM_NAME,_,n)
180 :
181 : #ifndef STEM_BURST
182 : #error "STEM_BURST must be defined"
183 : #endif
184 :
185 : #ifndef STEM_CALLBACK_CONTEXT_TYPE
186 : #error "STEM_CALLBACK_CONTEXT_TYPE must be defined"
187 : #endif
188 :
189 : #ifndef STEM_LAZY
190 0 : #define STEM_LAZY (0L)
191 : #endif
192 :
193 0 : #define STEM_SHUTDOWN_SEQ (ULONG_MAX-1UL)
194 :
195 : static inline void
196 0 : STEM_(in_update)( fd_stem_tile_in_t * in ) {
197 0 : fd_fseq_update( in->fseq, in->seq );
198 :
199 0 : volatile ulong * metrics = fd_metrics_link_in( fd_metrics_base_tl, in->idx );
200 :
201 0 : uint * accum = in->accum;
202 0 : ulong a0 = (ulong)accum[0]; ulong a1 = (ulong)accum[1]; ulong a2 = (ulong)accum[2];
203 0 : ulong a3 = (ulong)accum[3]; ulong a4 = (ulong)accum[4]; ulong a5 = (ulong)accum[5];
204 0 : FD_COMPILER_MFENCE();
205 0 : metrics[0] += a0; metrics[1] += a1; metrics[2] += a2;
206 0 : metrics[3] += a3; metrics[4] += a4; metrics[5] += a5;
207 0 : FD_COMPILER_MFENCE();
208 0 : accum[0] = 0U; accum[1] = 0U; accum[2] = 0U;
209 0 : accum[3] = 0U; accum[4] = 0U; accum[5] = 0U;
210 0 : }
211 :
212 : FD_FN_PURE static inline ulong
213 0 : STEM_(scratch_align)( void ) {
214 0 : return FD_STEM_SCRATCH_ALIGN;
215 0 : }
216 :
217 : FD_FN_PURE static inline ulong
218 : STEM_(scratch_footprint)( ulong in_cnt,
219 : ulong out_cnt,
220 0 : ulong cons_cnt ) {
221 0 : ulong l = FD_LAYOUT_INIT;
222 0 : l = FD_LAYOUT_APPEND( l, alignof(fd_stem_tile_in_t), in_cnt*sizeof(fd_stem_tile_in_t) ); /* in */
223 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong), out_cnt*sizeof(ulong) ); /* cr_avail */
224 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong), out_cnt*sizeof(ulong) ); /* out_depth */
225 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong), out_cnt*sizeof(ulong) ); /* out_seq */
226 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong const *), cons_cnt*sizeof(ulong const *) ); /* cons_fseq */
227 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong *), cons_cnt*sizeof(ulong *) ); /* cons_slow */
228 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong), cons_cnt*sizeof(ulong) ); /* cons_out */
229 0 : l = FD_LAYOUT_APPEND( l, alignof(ulong), cons_cnt*sizeof(ulong) ); /* cons_seq */
230 0 : const ulong event_cnt = in_cnt + 1UL + cons_cnt;
231 0 : l = FD_LAYOUT_APPEND( l, alignof(ushort), event_cnt*sizeof(ushort) ); /* event_map */
232 0 : return FD_LAYOUT_FINI( l, STEM_(scratch_align)() );
233 0 : }
234 :
235 : static inline void
236 : STEM_(run1)( ulong in_cnt,
237 : fd_frag_meta_t const ** in_mcache,
238 : ulong ** in_fseq,
239 : ulong out_cnt,
240 : fd_frag_meta_t ** out_mcache,
241 : ulong cons_cnt,
242 : ulong * _cons_out,
243 : ulong ** _cons_fseq,
244 : ulong burst,
245 : long lazy,
246 : fd_rng_t * rng,
247 : void * scratch,
248 0 : STEM_CALLBACK_CONTEXT_TYPE * ctx ) {
249 : /* in frag stream state */
250 0 : ulong in_seq; /* current position in input poll sequence, in [0,in_cnt) */
251 0 : fd_stem_tile_in_t * in; /* in[in_seq] for in_seq in [0,in_cnt) has information about input fragment stream currently at
252 : position in_seq in the in_idx polling sequence. The ordering of this array is continuously
253 : shuffled to avoid lighthousing effects in the output fragment stream at extreme fan-in and load */
254 :
255 : /* out frag stream state */
256 0 : ulong * out_depth; /* ==fd_mcache_depth( out_mcache[out_idx] ) for out_idx in [0, out_cnt) */
257 0 : ulong * out_seq; /* next mux frag sequence number to publish for out_idx in [0, out_cnt) ]*/
258 :
259 : /* out flow control state */
260 0 : ulong * cr_avail; /* number of flow control credits available to publish downstream across all outs */
261 0 : ulong min_cr_avail; /* minimum number of flow control credits available to publish downstream */
262 0 : ulong const ** cons_fseq; /* cons_fseq[cons_idx] for cons_idx in [0,cons_cnt) is where to receive fctl credits from consumers */
263 0 : ulong ** cons_slow; /* cons_slow[cons_idx] for cons_idx in [0,cons_cnt) is where to accumulate slow events */
264 0 : ulong * cons_out; /* cons_out[cons_idx] for cons_idx in [0,cons_ct) is which out the consumer consumes from */
265 0 : ulong * cons_seq; /* cons_seq [cons_idx] is the most recent observation of cons_fseq[cons_idx] */
266 :
267 : /* housekeeping state */
268 0 : ulong event_cnt; /* ==in_cnt+cons_cnt+1, total number of housekeeping events */
269 0 : ulong event_seq; /* current position in housekeeping event sequence, in [0,event_cnt) */
270 0 : ushort * event_map; /* current mapping of event_seq to event idx, event_map[ event_seq ] is next event to process */
271 0 : ulong async_min; /* minimum number of ticks between processing a housekeeping event, positive integer power of 2 */
272 :
273 : /* performance metrics */
274 0 : ulong metric_in_backp; /* is the run loop currently backpressured by one or more of the outs, in [0,1] */
275 0 : ulong metric_backp_cnt; /* Accumulates number of transitions of tile to backpressured between housekeeping events */
276 :
277 0 : ulong metric_regime_ticks[9]; /* How many ticks the tile has spent in each regime */
278 :
279 0 : if( FD_UNLIKELY( !scratch ) ) FD_LOG_ERR(( "NULL scratch" ));
280 0 : if( FD_UNLIKELY( !fd_ulong_is_aligned( (ulong)scratch, STEM_(scratch_align)() ) ) ) FD_LOG_ERR(( "misaligned scratch" ));
281 :
282 : /* in_backp==1, backp_cnt==0 indicates waiting for initial credits,
283 : cleared during first housekeeping if credits available */
284 0 : metric_in_backp = 1UL;
285 0 : metric_backp_cnt = 0UL;
286 0 : memset( metric_regime_ticks, 0, sizeof( metric_regime_ticks ) );
287 :
288 : /* in frag stream init */
289 :
290 0 : in_seq = 0UL; /* First in to poll */
291 :
292 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
293 0 : in = (fd_stem_tile_in_t *)FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_stem_tile_in_t), in_cnt*sizeof(fd_stem_tile_in_t) );
294 :
295 0 : if( FD_UNLIKELY( !!in_cnt && !in_mcache ) ) FD_LOG_ERR(( "NULL in_mcache" ));
296 0 : if( FD_UNLIKELY( !!in_cnt && !in_fseq ) ) FD_LOG_ERR(( "NULL in_fseq" ));
297 0 : if( FD_UNLIKELY( in_cnt > UINT_MAX ) ) FD_LOG_ERR(( "in_cnt too large" ));
298 0 : for( ulong in_idx=0UL; in_idx<in_cnt; in_idx++ ) {
299 :
300 0 : if( FD_UNLIKELY( !in_mcache[ in_idx ] ) ) FD_LOG_ERR(( "NULL in_mcache[%lu]", in_idx ));
301 0 : if( FD_UNLIKELY( !in_fseq [ in_idx ] ) ) FD_LOG_ERR(( "NULL in_fseq[%lu]", in_idx ));
302 :
303 0 : fd_stem_tile_in_t * this_in = &in[ in_idx ];
304 :
305 0 : this_in->mcache = in_mcache[ in_idx ];
306 0 : this_in->fseq = in_fseq [ in_idx ];
307 :
308 0 : ulong depth = fd_mcache_depth( this_in->mcache );
309 0 : if( FD_UNLIKELY( depth > UINT_MAX ) ) FD_LOG_ERR(( "in_mcache[%lu] too deep", in_idx ));
310 0 : this_in->depth = (uint)depth;
311 0 : this_in->idx = (uint)in_idx;
312 0 : this_in->seq = 0UL;
313 0 : this_in->mline = this_in->mcache + fd_mcache_line_idx( this_in->seq, this_in->depth );
314 :
315 0 : this_in->accum[0] = 0U; this_in->accum[1] = 0U; this_in->accum[2] = 0U;
316 0 : this_in->accum[3] = 0U; this_in->accum[4] = 0U; this_in->accum[5] = 0U;
317 0 : }
318 :
319 : /* out frag stream init */
320 :
321 0 : cr_avail = (ulong *)FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), out_cnt*sizeof(ulong) );
322 0 : min_cr_avail = 0UL;
323 :
324 0 : out_depth = (ulong *)FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), out_cnt*sizeof(ulong) );
325 0 : out_seq = (ulong *)FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), out_cnt*sizeof(ulong) );
326 :
327 0 : ulong cr_max = fd_ulong_if( !out_cnt, 128UL, ULONG_MAX );
328 :
329 0 : for( ulong out_idx=0UL; out_idx<out_cnt; out_idx++ ) {
330 :
331 0 : if( FD_UNLIKELY( !out_mcache[ out_idx ] ) ) FD_LOG_ERR(( "NULL out_mcache[%lu]", out_idx ));
332 :
333 0 : out_depth[ out_idx ] = fd_mcache_depth( out_mcache[ out_idx ] );
334 0 : out_seq[ out_idx ] = 0UL;
335 :
336 0 : cr_max = fd_ulong_min( cr_max, out_depth[ out_idx ] );
337 0 : cr_avail[ out_idx ] = out_depth[ out_idx ];
338 0 : }
339 :
340 0 : cons_fseq = (ulong const **)FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong const *), cons_cnt*sizeof(ulong const *) );
341 0 : cons_slow = (ulong **) FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong *), cons_cnt*sizeof(ulong *) );
342 0 : cons_out = (ulong *) FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), cons_cnt*sizeof(ulong) );
343 0 : cons_seq = (ulong *) FD_SCRATCH_ALLOC_APPEND( l, alignof(ulong), cons_cnt*sizeof(ulong) );
344 :
345 0 : if( FD_UNLIKELY( !!cons_cnt && !_cons_fseq ) ) FD_LOG_ERR(( "NULL cons_fseq" ));
346 0 : for( ulong cons_idx=0UL; cons_idx<cons_cnt; cons_idx++ ) {
347 0 : if( FD_UNLIKELY( !_cons_fseq[ cons_idx ] ) ) FD_LOG_ERR(( "NULL cons_fseq[%lu]", cons_idx ));
348 0 : cons_fseq[ cons_idx ] = _cons_fseq[ cons_idx ];
349 0 : cons_out [ cons_idx ] = _cons_out [ cons_idx ];
350 0 : cons_slow[ cons_idx ] = (ulong*)(fd_metrics_link_out( fd_metrics_base_tl, cons_idx ) + FD_METRICS_COUNTER_LINK_SLOW_COUNT_OFF);
351 0 : cons_seq [ cons_idx ] = fd_fseq_query( _cons_fseq[ cons_idx ] );
352 0 : }
353 :
354 : /* housekeeping init */
355 :
356 0 : if( lazy<=0L ) lazy = fd_tempo_lazy_default( cr_max );
357 0 : FD_LOG_INFO(( "Configuring housekeeping (lazy %li ns)", lazy ));
358 :
359 : /* Initialize the initial event sequence to immediately update
360 : cr_avail on the first run loop iteration and then update all the
361 : ins accordingly. */
362 :
363 0 : event_cnt = in_cnt + 1UL + cons_cnt;
364 0 : event_map = (ushort *)FD_SCRATCH_ALLOC_APPEND( l, alignof(ushort), event_cnt*sizeof(ushort) );
365 0 : event_seq = 0UL; event_map[ event_seq++ ] = (ushort)cons_cnt;
366 0 : for( ulong in_idx=0UL; in_idx< in_cnt; in_idx++ ) event_map[ event_seq++ ] = (ushort)(in_idx+cons_cnt+1UL);
367 0 : for( ulong cons_idx=0UL; cons_idx<cons_cnt; cons_idx++ ) event_map[ event_seq++ ] = (ushort)cons_idx;
368 0 : event_seq = 0UL;
369 :
370 0 : async_min = fd_tempo_async_min( lazy, event_cnt, (float)fd_tempo_tick_per_ns( NULL ) );
371 0 : if( FD_UNLIKELY( !async_min ) ) FD_LOG_ERR(( "bad lazy %lu %lu", (ulong)lazy, event_cnt ));
372 :
373 0 : FD_LOG_INFO(( "Running stem, cr_max = %lu", cr_max ));
374 0 : FD_MGAUGE_SET( TILE, STATUS, 1UL );
375 0 : long then = fd_tickcount();
376 0 : long now = then;
377 0 : for(;;) {
378 :
379 : #ifdef STEM_CALLBACK_SHOULD_SHUTDOWN
380 0 : if( FD_UNLIKELY( STEM_CALLBACK_SHOULD_SHUTDOWN( ctx ) ) ) break;
381 0 : #endif
382 :
383 : /* Do housekeeping at a low rate in the background */
384 :
385 0 : ulong housekeeping_ticks = 0UL;
386 0 : if( FD_UNLIKELY( (now-then)>=0L ) ) {
387 0 : ulong event_idx = (ulong)event_map[ event_seq ];
388 :
389 : /* Do the next async event. event_idx:
390 : <out_cnt - receive credits from out event_idx
391 : ==out_cnt - housekeeping
392 : >out_cnt - send credits to in event_idx - out_cnt - 1.
393 : Branch hints and order are optimized for the case:
394 : out_cnt >~ in_cnt >~ 1. */
395 :
396 0 : if( FD_LIKELY( event_idx<cons_cnt ) ) { /* cons fctl for cons cons_idx */
397 0 : ulong cons_idx = event_idx;
398 :
399 : /* Receive flow control credits from this out. */
400 0 : cons_seq[ cons_idx ] = fd_fseq_query( cons_fseq[ cons_idx ] );
401 :
402 0 : } else if( FD_LIKELY( event_idx>cons_cnt ) ) { /* in fctl for in in_idx */
403 0 : ulong in_idx = event_idx - cons_cnt - 1UL;
404 :
405 : /* Send flow control credits and drain flow control diagnostics
406 : for in_idx. */
407 :
408 0 : STEM_(in_update)( &in[ in_idx ] );
409 :
410 0 : } else { /* event_idx==cons_cnt, housekeeping event */
411 :
412 : /* Update metrics counters to external viewers */
413 0 : FD_COMPILER_MFENCE();
414 0 : FD_MGAUGE_SET( TILE, HEARTBEAT, (ulong)now );
415 0 : FD_MGAUGE_SET( TILE, IN_BACKPRESSURE, metric_in_backp );
416 0 : FD_MCNT_INC ( TILE, BACKPRESSURE_COUNT, metric_backp_cnt );
417 0 : FD_MCNT_ENUM_COPY( TILE, REGIME_DURATION_NANOS, metric_regime_ticks );
418 : #ifdef STEM_CALLBACK_METRICS_WRITE
419 0 : STEM_CALLBACK_METRICS_WRITE( ctx );
420 : #endif
421 0 : FD_COMPILER_MFENCE();
422 0 : metric_backp_cnt = 0UL;
423 :
424 : /* Receive flow control credits */
425 0 : if( FD_LIKELY( min_cr_avail<cr_max ) ) {
426 0 : ulong slowest_cons = ULONG_MAX;
427 0 : min_cr_avail = cr_max;
428 0 : for( ulong out_idx=0; out_idx<out_cnt; out_idx++ ) {
429 0 : cr_avail[ out_idx ] = out_depth[ out_idx ];
430 0 : }
431 :
432 0 : for( ulong cons_idx=0UL; cons_idx<cons_cnt; cons_idx++ ) {
433 0 : ulong out_idx = cons_out[ cons_idx ];
434 0 : ulong cons_cr_avail = (ulong)fd_long_max( (long)out_depth[ out_idx ]-fd_long_max( fd_seq_diff( out_seq[ out_idx ], cons_seq[ cons_idx ] ), 0L ), 0L );
435 :
436 : /* If a reliable consumer exits, they can set the credit
437 : return fseq to STEM_SHUTDOWN_SEQ to indicate they are no
438 : longer actively consuming. */
439 0 : cons_cr_avail = fd_ulong_if( cons_seq[ cons_idx ]==STEM_SHUTDOWN_SEQ, out_depth[ out_idx ], cons_cr_avail );
440 0 : slowest_cons = fd_ulong_if( cons_cr_avail<min_cr_avail, cons_idx, slowest_cons );
441 :
442 0 : cr_avail[ out_idx ] = fd_ulong_min( cr_avail[ out_idx ], cons_cr_avail );
443 0 : min_cr_avail = fd_ulong_min( cons_cr_avail, min_cr_avail );
444 0 : }
445 :
446 : /* See notes above about use of quasi-atomic diagnostic accum */
447 0 : if( FD_LIKELY( slowest_cons!=ULONG_MAX ) ) {
448 0 : FD_COMPILER_MFENCE();
449 0 : (*cons_slow[ slowest_cons ]) += metric_in_backp;
450 0 : FD_COMPILER_MFENCE();
451 0 : }
452 0 : }
453 :
454 : #ifdef STEM_CALLBACK_DURING_HOUSEKEEPING
455 0 : STEM_CALLBACK_DURING_HOUSEKEEPING( ctx );
456 : #else
457 : (void)ctx;
458 : #endif
459 0 : }
460 :
461 : /* Select which event to do next (randomized round robin) and
462 : reload the housekeeping timer. */
463 :
464 0 : event_seq++;
465 0 : if( FD_UNLIKELY( event_seq>=event_cnt ) ) {
466 0 : event_seq = 0UL;
467 :
468 : /* Randomize the order of event processing for the next event
469 : event_cnt events to avoid lighthousing effects causing input
470 : credit starvation at extreme fan in/fan out, extreme in load
471 : and high credit return laziness. */
472 :
473 0 : ulong swap_idx = (ulong)fd_rng_uint_roll( rng, (uint)event_cnt );
474 0 : ushort map_tmp = event_map[ swap_idx ];
475 0 : event_map[ swap_idx ] = event_map[ 0 ];
476 0 : event_map[ 0 ] = map_tmp;
477 :
478 : /* We also do the same with the ins to prevent there being a
479 : correlated order frag origins from different inputs
480 : downstream at extreme fan in and extreme in load. */
481 :
482 0 : if( FD_LIKELY( in_cnt>1UL ) ) {
483 0 : swap_idx = (ulong)fd_rng_uint_roll( rng, (uint)in_cnt );
484 0 : fd_stem_tile_in_t in_tmp;
485 0 : in_tmp = in[ swap_idx ];
486 0 : in[ swap_idx ] = in[ 0 ];
487 0 : in[ 0 ] = in_tmp;
488 0 : }
489 0 : }
490 :
491 : /* Reload housekeeping timer */
492 0 : then = now + (long)fd_tempo_async_reload( rng, async_min );
493 0 : long next = fd_tickcount();
494 0 : housekeeping_ticks = (ulong)(next - now);
495 0 : now = next;
496 0 : }
497 :
498 : #if defined(STEM_CALLBACK_BEFORE_CREDIT) || defined(STEM_CALLBACK_AFTER_CREDIT) || defined(STEM_CALLBACK_AFTER_FRAG) || defined(STEM_CALLBACK_RETURNABLE_FRAG)
499 : fd_stem_context_t stem = {
500 : .mcaches = out_mcache,
501 : .depths = out_depth,
502 : .seqs = out_seq,
503 :
504 : .cr_avail = cr_avail,
505 : .min_cr_avail = &min_cr_avail,
506 : .cr_decrement_amount = fd_ulong_if( out_cnt>0UL, 1UL, 0UL ),
507 : };
508 : #endif
509 :
510 0 : int charge_busy_before = 0;
511 : #ifdef STEM_CALLBACK_BEFORE_CREDIT
512 0 : STEM_CALLBACK_BEFORE_CREDIT( ctx, &stem, &charge_busy_before );
513 : #endif
514 :
515 : /* Check if we are backpressured. If so, count any transition into
516 : a backpressured regime and spin to wait for flow control credits
517 : to return. We don't do a fully atomic update here as it is only
518 : diagnostic and it will still be correct in the usual case where
519 : individual diagnostic counters aren't used by writers in
520 : different threads of execution. We only count the transition
521 : from not backpressured to backpressured. */
522 :
523 0 : if( FD_UNLIKELY( min_cr_avail<burst ) ) {
524 0 : metric_backp_cnt += (ulong)!metric_in_backp;
525 0 : metric_in_backp = 1UL;
526 0 : FD_SPIN_PAUSE();
527 0 : metric_regime_ticks[2] += housekeeping_ticks;
528 0 : long next = fd_tickcount();
529 0 : metric_regime_ticks[5] += (ulong)(next - now);
530 0 : now = next;
531 0 : continue;
532 0 : }
533 0 : metric_in_backp = 0UL;
534 :
535 0 : int charge_busy_after = 0;
536 : #ifdef STEM_CALLBACK_AFTER_CREDIT
537 : int poll_in = 1;
538 0 : STEM_CALLBACK_AFTER_CREDIT( ctx, &stem, &poll_in, &charge_busy_after );
539 0 : if( FD_UNLIKELY( !poll_in ) ) {
540 0 : metric_regime_ticks[1] += housekeeping_ticks;
541 0 : long next = fd_tickcount();
542 0 : metric_regime_ticks[4] += (ulong)(next - now);
543 0 : now = next;
544 0 : continue;
545 0 : }
546 0 : #endif
547 :
548 : /* Select which in to poll next (randomized round robin) */
549 :
550 0 : if( FD_UNLIKELY( !in_cnt ) ) {
551 0 : int was_busy = 0;
552 0 : was_busy |= !!charge_busy_before;
553 0 : was_busy |= !!charge_busy_after;
554 0 : metric_regime_ticks[ 0+was_busy ] += housekeeping_ticks;
555 0 : long next = fd_tickcount();
556 0 : metric_regime_ticks[ 3+was_busy ] += (ulong)(next - now);
557 0 : now = next;
558 0 : continue;
559 0 : }
560 :
561 0 : ulong prefrag_ticks = 0UL;
562 : #if defined(STEM_CALLBACK_BEFORE_CREDIT) && defined(STEM_CALLBACK_AFTER_CREDIT)
563 0 : if( FD_LIKELY( charge_busy_before || charge_busy_after ) ) {
564 : #elif defined(STEM_CALLBACK_BEFORE_CREDIT)
565 0 : if( FD_LIKELY( charge_busy_before ) ) {
566 : #elif defined(STEM_CALLBACK_AFTER_CREDIT)
567 0 : if( FD_LIKELY( charge_busy_after ) ) {
568 0 : #endif
569 :
570 : #if defined(STEM_CALLBACK_BEFORE_CREDIT) || defined(STEM_CALLBACK_AFTER_CREDIT)
571 0 : long prefrag_next = fd_tickcount();
572 0 : prefrag_ticks = (ulong)(prefrag_next - now);
573 0 : now = prefrag_next;
574 0 : }
575 : #endif
576 :
577 0 : fd_stem_tile_in_t * this_in = &in[ in_seq ];
578 0 : in_seq++;
579 0 : if( in_seq>=in_cnt ) in_seq = 0UL; /* cmov */
580 :
581 : /* Check if this in has any new fragments to mux */
582 :
583 0 : ulong this_in_seq = this_in->seq;
584 0 : fd_frag_meta_t const * this_in_mline = this_in->mline; /* Already at appropriate line for this_in_seq */
585 :
586 0 : #if FD_HAS_SSE
587 0 : __m128i seq_sig = fd_frag_meta_seq_sig_query( this_in_mline );
588 0 : ulong seq_found = fd_frag_meta_sse0_seq( seq_sig );
589 0 : ulong sig = fd_frag_meta_sse0_sig( seq_sig );
590 : #else
591 : /* Without SSE, seq and sig might be read from different frags (due
592 : to overrun), which results in a before_frag and during_frag being
593 : issued with incorrect arguments, but not after_frag. */
594 : ulong seq_found = FD_VOLATILE_CONST( this_in_mline->seq );
595 : ulong sig = FD_VOLATILE_CONST( this_in_mline->sig );
596 : #endif
597 0 : (void)sig;
598 :
599 0 : long diff = fd_seq_diff( this_in_seq, seq_found );
600 0 : if( FD_UNLIKELY( diff ) ) { /* Caught up or overrun, optimize for new frag case */
601 0 : ulong * housekeeping_regime = &metric_regime_ticks[0];
602 0 : ulong * prefrag_regime = &metric_regime_ticks[3];
603 0 : ulong * finish_regime = &metric_regime_ticks[6];
604 0 : if( FD_UNLIKELY( diff<0L ) ) { /* Overrun (impossible if in is honoring our flow control) */
605 0 : this_in->seq = seq_found; /* Resume from here (probably reasonably current, could query in mcache sync directly instead) */
606 0 : housekeeping_regime = &metric_regime_ticks[1];
607 0 : prefrag_regime = &metric_regime_ticks[4];
608 0 : finish_regime = &metric_regime_ticks[7];
609 0 : this_in->accum[ FD_METRICS_COUNTER_LINK_OVERRUN_POLLING_COUNT_OFF ]++;
610 0 : this_in->accum[ FD_METRICS_COUNTER_LINK_OVERRUN_POLLING_FRAG_COUNT_OFF ] += (uint)(-diff);
611 :
612 : #ifdef STEM_CALLBACK_AFTER_POLL_OVERRUN
613 0 : STEM_CALLBACK_AFTER_POLL_OVERRUN( ctx );
614 : #endif
615 0 : }
616 :
617 : /* Don't bother with spin as polling multiple locations */
618 0 : *housekeeping_regime += housekeeping_ticks;
619 0 : *prefrag_regime += prefrag_ticks;
620 0 : long next = fd_tickcount();
621 0 : *finish_regime += (ulong)(next - now);
622 0 : now = next;
623 0 : continue;
624 0 : }
625 :
626 : #ifdef STEM_CALLBACK_BEFORE_FRAG
627 0 : int filter = STEM_CALLBACK_BEFORE_FRAG( ctx, (ulong)this_in->idx, seq_found, sig );
628 0 : if( FD_UNLIKELY( filter<0 ) ) {
629 0 : metric_regime_ticks[1] += housekeeping_ticks;
630 0 : metric_regime_ticks[4] += prefrag_ticks;
631 0 : long next = fd_tickcount();
632 0 : metric_regime_ticks[7] += (ulong)(next - now);
633 0 : now = next;
634 0 : continue;
635 0 : } else if( FD_UNLIKELY( filter>0 ) ) {
636 0 : this_in->accum[ FD_METRICS_COUNTER_LINK_FILTERED_COUNT_OFF ]++;
637 0 : this_in->accum[ FD_METRICS_COUNTER_LINK_FILTERED_SIZE_BYTES_OFF ] += (uint)this_in_mline->sz; /* TODO: This might be overrun ... ? Not loaded atomically */
638 :
639 : this_in_seq = fd_seq_inc( this_in_seq, 1UL );
640 : this_in->seq = this_in_seq;
641 : this_in->mline = this_in->mcache + fd_mcache_line_idx( this_in_seq, this_in->depth );
642 :
643 0 : metric_regime_ticks[1] += housekeeping_ticks;
644 0 : metric_regime_ticks[4] += prefrag_ticks;
645 0 : long next = fd_tickcount();
646 0 : metric_regime_ticks[7] += (ulong)(next - now);
647 0 : now = next;
648 0 : continue;
649 0 : }
650 0 : #endif
651 :
652 : /* We have a new fragment to mux. Try to load it. This attempt
653 : should always be successful if in producers are honoring our flow
654 : control. Since we can cheaply detect if there are
655 : misconfigurations (should be an L1 cache hit / predictable branch
656 : in the properly configured case), we do so anyway. Note that if
657 : we are on a platform where AVX is atomic, this could be replaced
658 : by a flat AVX load of the metadata and an extraction of the found
659 : sequence number for higher performance. */
660 0 : FD_COMPILER_MFENCE();
661 0 : ulong chunk = (ulong)this_in_mline->chunk; (void)chunk;
662 0 : ulong sz = (ulong)this_in_mline->sz; (void)sz;
663 0 : ulong ctl = (ulong)this_in_mline->ctl; (void)ctl;
664 0 : ulong tsorig = (ulong)this_in_mline->tsorig; (void)tsorig;
665 0 : ulong tspub = (ulong)this_in_mline->tspub; (void)tspub;
666 :
667 : #ifdef STEM_CALLBACK_DURING_FRAG
668 0 : STEM_CALLBACK_DURING_FRAG( ctx, (ulong)this_in->idx, seq_found, sig, chunk, sz, ctl );
669 : #endif
670 :
671 0 : FD_COMPILER_MFENCE();
672 0 : ulong seq_test = this_in_mline->seq;
673 0 : FD_COMPILER_MFENCE();
674 :
675 0 : if( FD_UNLIKELY( fd_seq_ne( seq_test, seq_found ) ) ) { /* Overrun while reading (impossible if this_in honoring our fctl) */
676 0 : this_in->seq = seq_test; /* Resume from here (probably reasonably current, could query in mcache sync instead) */
677 0 : fd_metrics_link_in( fd_metrics_base_tl, this_in->idx )[ FD_METRICS_COUNTER_LINK_OVERRUN_READING_COUNT_OFF ]++; /* No local accum since extremely rare, faster to use smaller cache line */
678 0 : fd_metrics_link_in( fd_metrics_base_tl, this_in->idx )[ FD_METRICS_COUNTER_LINK_OVERRUN_READING_FRAG_COUNT_OFF ] += (uint)fd_seq_diff( seq_test, seq_found ); /* No local accum since extremely rare, faster to use smaller cache line */
679 : /* Don't bother with spin as polling multiple locations */
680 0 : metric_regime_ticks[1] += housekeeping_ticks;
681 0 : metric_regime_ticks[4] += prefrag_ticks;
682 0 : long next = fd_tickcount();
683 0 : metric_regime_ticks[7] += (ulong)(next - now);
684 0 : now = next;
685 0 : continue;
686 0 : }
687 :
688 : #ifdef STEM_CALLBACK_RETURNABLE_FRAG
689 0 : int return_frag = STEM_CALLBACK_RETURNABLE_FRAG( ctx, (ulong)this_in->idx, seq_found, sig, chunk, sz, tsorig, tspub, &stem );
690 0 : if( FD_UNLIKELY( return_frag ) ) {
691 0 : metric_regime_ticks[1] += housekeeping_ticks;
692 0 : long next = fd_tickcount();
693 0 : metric_regime_ticks[4] += (ulong)(next - now);
694 0 : now = next;
695 0 : continue;
696 0 : }
697 0 : #endif
698 :
699 : #ifdef STEM_CALLBACK_AFTER_FRAG
700 0 : STEM_CALLBACK_AFTER_FRAG( ctx, (ulong)this_in->idx, seq_found, sig, sz, tsorig, tspub, &stem );
701 0 : #endif
702 :
703 : /* Windup for the next in poll and accumulate diagnostics */
704 :
705 0 : this_in_seq = fd_seq_inc( this_in_seq, 1UL );
706 0 : this_in->seq = this_in_seq;
707 0 : this_in->mline = this_in->mcache + fd_mcache_line_idx( this_in_seq, this_in->depth );
708 :
709 0 : this_in->accum[ FD_METRICS_COUNTER_LINK_CONSUMED_COUNT_OFF ]++;
710 0 : this_in->accum[ FD_METRICS_COUNTER_LINK_CONSUMED_SIZE_BYTES_OFF ] += (uint)sz;
711 :
712 0 : metric_regime_ticks[1] += housekeeping_ticks;
713 0 : metric_regime_ticks[4] += prefrag_ticks;
714 0 : long next = fd_tickcount();
715 0 : metric_regime_ticks[7] += (ulong)(next - now);
716 0 : now = next;
717 0 : }
718 0 : }
719 :
720 : FD_FN_UNUSED static void
721 : STEM_(run)( fd_topo_t * topo,
722 0 : fd_topo_tile_t * tile ) {
723 0 : const fd_frag_meta_t * in_mcache[ FD_TOPO_MAX_LINKS ];
724 0 : ulong * in_fseq[ FD_TOPO_MAX_TILE_IN_LINKS ];
725 :
726 0 : ulong polled_in_cnt = 0UL;
727 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
728 0 : if( FD_UNLIKELY( !tile->in_link_poll[ i ] ) ) continue;
729 :
730 0 : in_mcache[ polled_in_cnt ] = topo->links[ tile->in_link_id[ i ] ].mcache;
731 0 : FD_TEST( in_mcache[ polled_in_cnt ] );
732 0 : in_fseq[ polled_in_cnt ] = tile->in_link_fseq[ i ];
733 0 : FD_TEST( in_fseq[ polled_in_cnt ] );
734 0 : polled_in_cnt += 1;
735 0 : }
736 :
737 0 : fd_frag_meta_t * out_mcache[ FD_TOPO_MAX_LINKS ];
738 0 : for( ulong i=0UL; i<tile->out_cnt; i++ ) {
739 0 : out_mcache[ i ] = topo->links[ tile->out_link_id[ i ] ].mcache;
740 0 : FD_TEST( out_mcache[ i ] );
741 0 : }
742 :
743 0 : ulong reliable_cons_cnt = 0UL;
744 0 : ulong cons_out[ FD_TOPO_MAX_LINKS ];
745 0 : ulong * cons_fseq[ FD_TOPO_MAX_LINKS ];
746 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
747 0 : fd_topo_tile_t * consumer_tile = &topo->tiles[ i ];
748 0 : for( ulong j=0UL; j<consumer_tile->in_cnt; j++ ) {
749 0 : for( ulong k=0UL; k<tile->out_cnt; k++ ) {
750 0 : if( FD_UNLIKELY( consumer_tile->in_link_id[ j ]==tile->out_link_id[ k ] && consumer_tile->in_link_reliable[ j ] ) ) {
751 0 : cons_out[ reliable_cons_cnt ] = k;
752 0 : cons_fseq[ reliable_cons_cnt ] = consumer_tile->in_link_fseq[ j ];
753 0 : FD_TEST( cons_fseq[ reliable_cons_cnt ] );
754 0 : reliable_cons_cnt++;
755 : /* Need to test this, since each link may connect to many outs,
756 : you could construct a topology which has more than this
757 : consumers of links. */
758 0 : FD_TEST( reliable_cons_cnt<FD_TOPO_MAX_LINKS );
759 0 : }
760 0 : }
761 0 : }
762 0 : }
763 :
764 0 : fd_rng_t rng[1];
765 0 : FD_TEST( fd_rng_join( fd_rng_new( rng, 0, 0UL ) ) );
766 :
767 0 : STEM_CALLBACK_CONTEXT_TYPE * ctx = (STEM_CALLBACK_CONTEXT_TYPE*)fd_ulong_align_up( (ulong)fd_topo_obj_laddr( topo, tile->tile_obj_id ), STEM_CALLBACK_CONTEXT_ALIGN );
768 :
769 0 : STEM_(run1)( polled_in_cnt,
770 0 : in_mcache,
771 0 : in_fseq,
772 0 : tile->out_cnt,
773 0 : out_mcache,
774 0 : reliable_cons_cnt,
775 0 : cons_out,
776 0 : cons_fseq,
777 0 : STEM_BURST,
778 0 : STEM_LAZY,
779 0 : rng,
780 0 : fd_alloca( FD_STEM_SCRATCH_ALIGN, STEM_(scratch_footprint)( polled_in_cnt, tile->out_cnt, reliable_cons_cnt ) ),
781 0 : ctx );
782 :
783 0 : if( FD_LIKELY( tile->allow_shutdown ) ) {
784 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
785 0 : if( FD_UNLIKELY( !tile->in_link_poll[ i ] || !tile->in_link_reliable[ i ] ) ) continue;
786 :
787 : /* Return infinite credits on any reliable consumer links so that
788 : producers now no longer expect us to consume. */
789 0 : ulong fseq_id = tile->in_link_fseq_obj_id[ i ];
790 0 : ulong * fseq = fd_fseq_join( fd_topo_obj_laddr( topo, fseq_id ) );
791 0 : FD_TEST( fseq );
792 0 : fd_fseq_update( fseq, STEM_SHUTDOWN_SEQ );
793 0 : }
794 0 : }
795 0 : }
796 :
797 : #undef STEM_NAME
798 : #undef STEM_
799 : #undef STEM_BURST
800 : #undef STEM_CALLBACK_CONTEXT_TYPE
801 : #undef STEM_LAZY
802 : #undef STEM_CALLBACK_SHOULD_SHUTDOWN
803 : #undef STEM_CALLBACK_DURING_HOUSEKEEPING
804 : #undef STEM_CALLBACK_METRICS_WRITE
805 : #undef STEM_CALLBACK_BEFORE_CREDIT
806 : #undef STEM_CALLBACK_AFTER_CREDIT
807 : #undef STEM_CALLBACK_BEFORE_FRAG
808 : #undef STEM_CALLBACK_DURING_FRAG
809 : #undef STEM_CALLBACK_RETURNABLE_FRAG
810 : #undef STEM_CALLBACK_AFTER_FRAG
|
