Line data Source code
1 : #include "../tiles.h"
2 :
3 : #include "generated/fd_pack_tile_seccomp.h"
4 :
5 : #include "../../util/pod/fd_pod_format.h"
6 : #include "../../discof/replay/fd_replay_tile.h" // layering violation
7 : #include "../fd_txn_m.h"
8 : #include "../keyguard/fd_keyload.h"
9 : #include "../keyguard/fd_keyswitch.h"
10 : #include "../keyguard/fd_keyguard.h"
11 : #include "../keyguard/fd_keyguard_client.h"
12 : #include "../metrics/fd_metrics.h"
13 : #include "../pack/fd_pack.h"
14 : #include "../pack/fd_pack_cost.h"
15 : #include "../pack/fd_pack_pacing.h"
16 :
17 : #include <string.h>
18 :
19 : /* fd_pack is responsible for taking verified transactions, and
20 : arranging them into "microblocks" (groups) of transactions to
21 : be executed serially. It can try to do clever things so that
22 : multiple microblocks can execute in parallel, if they don't
23 : write to the same accounts. */
24 :
25 0 : #define IN_KIND_RESOLV (0UL)
26 0 : #define IN_KIND_POH (1UL)
27 0 : #define IN_KIND_EXECLE (2UL)
28 0 : #define IN_KIND_SIGN (3UL)
29 0 : #define IN_KIND_REPLAY (4UL)
30 0 : #define IN_KIND_EXECUTED_TXN (5UL)
31 :
32 : /* Pace microblocks, but only slightly. This helps keep performance
33 : more stable. This limit is 2,000 microblocks/second/execle. At
34 : MAX_TXN_PER_MICROBLOCK transactions/microblock, that's
35 : 2000*MAX_TXN_PER_MICROBLOCK txn/sec/execle. */
36 0 : #define MICROBLOCK_DURATION_NS (0L)
37 :
38 : /* There are 151 accepted blockhashes, but those don't include skips.
39 : This check is neither precise nor accurate, but just good enough.
40 : The execle tile does the final check. We give a little margin for a
41 : few percent skip rate. */
42 0 : #define TRANSACTION_LIFETIME_SLOTS 160UL
43 :
44 : /* Time is normally a long, but pack expects a ulong. Add -LONG_MIN to
45 : the time values so that LONG_MIN maps to 0, LONG_MAX maps to
46 : ULONG_MAX, and everything in between maps linearly with a slope of 1.
47 : Just subtracting LONG_MIN results in signed integer overflow, which
48 : is U.B. */
49 : #define TIME_OFFSET 0x8000000000000000UL
50 : FD_STATIC_ASSERT( (ulong)LONG_MIN+TIME_OFFSET==0UL, time_offset );
51 : FD_STATIC_ASSERT( (ulong)LONG_MAX+TIME_OFFSET==ULONG_MAX, time_offset );
52 :
53 : /* 1.6 M cost units, enough for 1 max size transaction */
54 : const ulong CUS_PER_MICROBLOCK = 1600000UL;
55 :
56 : #define SMALL_MICROBLOCKS 1
57 :
58 : #if SMALL_MICROBLOCKS
59 : const float VOTE_FRACTION = 1.0f; /* schedule all available votes first */
60 0 : #define EFFECTIVE_TXN_PER_MICROBLOCK 1UL
61 : #else
62 : const float VOTE_FRACTION = 0.75f; /* TODO: Is this the right value? */
63 : #define EFFECTIVE_TXN_PER_MICROBLOCK MAX_TXN_PER_MICROBLOCK
64 : #endif
65 :
66 : #if !SMALL_MICROBLOCKS
67 : /* There's overhead associated with each microblock the execle tile
68 : tries to execute it, so the optimal strategy is not to produce a
69 : microblock with a single transaction as soon as we receive it.
70 : Basically, if we have less than MAX_TXN_PER_MICROBLOCK transactions,
71 : we want to wait a little to see if we receive additional transactions
72 : before we schedule a microblock. We can model the optimum amount of
73 : time to wait, but the equation is complicated enough that we want to
74 : compute it before compile time. wait_duration[i] for i in
75 : [0, MAX_TXN_PER_MICROBLOCK] gives the time in nanoseconds pack should
76 : wait after receiving its most recent transaction before scheduling if
77 : it has i transactions available. Unsurprisingly,
78 : wait_duration[MAX_TXN_PER_MICROBLOCK] is 0. wait_duration[0] is
79 : ULONG_MAX, so we'll always wait if we have 0 transactions. */
80 : FD_IMPORT( wait_duration, "src/disco/pack/pack_delay.bin", ulong, 6, "" );
81 : #endif
82 :
83 :
84 :
85 : #if FD_PACK_USE_EXTRA_STORAGE
86 : /* When we are done being leader for a slot and we are leader in the
87 : very next slot, it can still take some time to transition. This is
88 : because the bank has to be finalized, a hash calculated, and various
89 : other things done in the replay stage to create the new child bank.
90 :
91 : During that time, pack cannot send transactions to execles so it
92 : needs to be able to buffer. Typically, these so called "leader
93 : transitions" are short (<15 millis), so a low value here would
94 : suffice. However, in some cases when there is memory pressure on the
95 : NUMA node or when the operating system context switches relevant
96 : threads out, it can take significantly longer.
97 :
98 : To prevent drops in these cases and because we assume execles are
99 : fast enough to drain this buffer once we do become leader, we set
100 : this buffer size to be quite large. */
101 :
102 : #define DEQUE_NAME extra_txn_deq
103 : #define DEQUE_T fd_txn_e_t
104 : #define DEQUE_MAX (128UL*1024UL)
105 : #include "../../../../util/tmpl/fd_deque.c"
106 :
107 : #endif
108 :
109 : /* Sync with src/app/shared/fd_config.c */
110 0 : #define FD_PACK_STRATEGY_PERF 0
111 0 : #define FD_PACK_STRATEGY_BALANCED 1
112 :
113 : static char const * const schedule_strategy_strings[2] = { "PRF", "BAL" };
114 :
115 :
116 : typedef struct {
117 : fd_acct_addr_t commission_pubkey[1];
118 : ulong commission;
119 : } block_builder_info_t;
120 :
121 : typedef struct {
122 : long time;
123 : ulong sched_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_CNT ];
124 : } fd_pack_sched_results_snap_t;
125 :
126 : typedef struct {
127 : fd_wksp_t * mem;
128 : ulong chunk0;
129 : ulong wmark;
130 : } fd_pack_in_ctx_t;
131 :
132 : typedef struct {
133 : fd_pack_t * pack;
134 : fd_txn_e_t * cur_spot;
135 : int is_bundle; /* is the current transaction a bundle */
136 :
137 : uchar executed_txn_sig[ 64UL ];
138 : uchar txn_committed;
139 :
140 : /* One of the FD_PACK_STRATEGY_* values defined above */
141 : int strategy;
142 :
143 : /* The value passed to fd_pack_new, etc. */
144 : ulong max_pending_transactions;
145 :
146 : /* The leader slot we are currently packing for, or ULONG_MAX if we
147 : are not the leader. */
148 : ulong leader_slot;
149 : void const * leader_bank;
150 : ulong leader_bank_idx;
151 : ulong leader_bank_seq;
152 :
153 : fd_became_leader_t _became_leader[1];
154 :
155 : /* The number of microblocks we have packed for the current leader
156 : slot. Will always be <= slot_max_microblocks. We must track
157 : this so that when we are done we can tell the PoH tile how many
158 : microblocks to expect in the slot. */
159 : ulong slot_microblock_cnt;
160 :
161 : /* Counter which increments when we've finished packing for a slot */
162 : uint pack_idx;
163 :
164 : ulong pack_txn_cnt; /* total num transactions packed since startup */
165 :
166 : /* The maximum number of microblocks that can be packed in this slot.
167 : Provided by the PoH tile when we become leader.*/
168 : ulong slot_max_microblocks;
169 :
170 : /* Cap (in bytes) of the amount of transaction data we produce in each
171 : block to avoid hitting the shred limits. See where this is set for
172 : more explanation. */
173 : ulong slot_max_data;
174 : int larger_shred_limits_per_block;
175 :
176 : /* Consensus critical slot cost limits. */
177 : struct {
178 : ulong slot_max_cost;
179 : ulong slot_max_vote_cost;
180 : ulong slot_max_write_cost_per_acct;
181 : ulong slot_max_allocated_data_per_block;
182 : ulong slot_max_data_shreds;
183 : } limits;
184 :
185 : /* If drain_execle is non-zero, then the pack tile must wait until all
186 : execle are idle before scheduling any more microblocks. This is
187 : primarily helpful in irregular leader transitions, e.g. while being
188 : leader for slot N, we switch forks to a slot M (!=N+1) in which we
189 : are also leader. We don't want to execute microblocks for
190 : different slots concurrently. */
191 : int drain_execle;
192 :
193 : /* Updated during housekeeping and used only for checking if the
194 : leader slot has ended. Might be off by one housekeeping duration,
195 : but that should be small relative to a slot duration. */
196 : long approx_wallclock_ns;
197 :
198 : /* approx_tickcount is updated in during_housekeeping() with
199 : fd_tickcount() and will match approx_wallclock_ns. This is done
200 : because we need to include an accurate nanosecond timestamp in
201 : every fd_txn_p_t but don't want to have to call the expensive
202 : fd_log_wallclock() in in the critical path. We can use
203 : fd_tempo_tick_per_ns() to convert from ticks to nanoseconds over
204 : small periods of time. */
205 : long approx_tickcount;
206 :
207 : fd_rng_t * rng;
208 :
209 : uint rng_seed;
210 : ulong rng_idx;
211 :
212 : /* The end wallclock time of the leader slot we are currently packing
213 : for, if we are currently packing for a slot.*/
214 : long slot_end_ns;
215 :
216 : /* The current dynamic upper bound on total microblocks for this slot.
217 : Monotonically decreasing over the slot lifetime. */
218 : ulong slot_dynamic_max_microblocks;
219 :
220 : /* Set by during_housekeeping when the dynamic bound drops below
221 : slot_max_microblocks. Consumed by after_credit which publishes
222 : the updated bound to POH over the pack_poh link. */
223 : int pending_reduce_mb_bound;
224 :
225 : /* pacer and ticks_per_ns are used for pacing CUs through the slot,
226 : i.e. deciding when to schedule a microblock given the number of CUs
227 : that have been consumed so far. pacer is an opaque pacing object,
228 : which is initialized when the pack tile is packing a slot.
229 : ticks_per_ns is the cached value from tempo. */
230 : fd_pack_pacing_t pacer[1];
231 : double ticks_per_ns;
232 :
233 : /* last_successful_insert stores the tickcount of the last
234 : successful transaction insert. */
235 : long last_successful_insert;
236 :
237 : /* highest_observed_slot stores the highest slot number we've seen
238 : from any transaction coming from the resolv tile. When this
239 : increases, we expire old transactions. */
240 : ulong highest_observed_slot;
241 :
242 : /* microblock_duration_ns, and wait_duration
243 : respectively scaled to be in ticks instead of nanoseconds */
244 : ulong microblock_duration_ticks;
245 : #if !SMALL_MICROBLOCKS
246 : ulong wait_duration_ticks[ MAX_TXN_PER_MICROBLOCK+1UL ];
247 : #endif
248 :
249 : #if FD_PACK_USE_EXTRA_STORAGE
250 : /* In addition to the available transactions that pack knows about, we
251 : also store a larger ring buffer for handling cases when pack is
252 : full. This is an fd_deque. */
253 : fd_txn_e_t * extra_txn_deq;
254 : int insert_to_extra; /* whether the last insert was into pack or the extra deq */
255 : #endif
256 :
257 : fd_pack_in_ctx_t in[ 32 ];
258 : int in_kind[ 32 ];
259 :
260 : ulong execle_cnt;
261 : ulong execle_idle_bitset; /* bit i is 1 if we've observed *execle_current[i]==execle_expect[i] */
262 : int poll_cursor; /* in [0, execle_cnt), the next execle to poll */
263 : int use_consumed_cus;
264 : long skip_cnt;
265 : ulong * execle_current[ FD_PACK_MAX_EXECLE_TILES ];
266 : ulong execle_expect[ FD_PACK_MAX_EXECLE_TILES ];
267 : /* execle_ready_at[x] means don't check execle x until tickcount is at
268 : least execle_ready_at[x]. */
269 : long execle_ready_at[ FD_PACK_MAX_EXECLE_TILES ];
270 :
271 : fd_wksp_t * execle_out_mem;
272 : ulong execle_out_chunk0;
273 : ulong execle_out_wmark;
274 : ulong execle_out_chunk;
275 :
276 : fd_wksp_t * poh_out_mem;
277 : ulong poh_out_chunk0;
278 : ulong poh_out_wmark;
279 : ulong poh_out_chunk;
280 :
281 : ulong insert_result[ FD_PACK_INSERT_RETVAL_CNT ];
282 : fd_histf_t schedule_duration[ 1 ];
283 : fd_histf_t no_sched_duration[ 1 ];
284 : fd_histf_t insert_duration [ 1 ];
285 : fd_histf_t complete_duration[ 1 ];
286 :
287 : struct {
288 : uint metric_state;
289 : long metric_state_begin;
290 : long metric_timing[ 16 ];
291 : };
292 :
293 : /* last_sched_metrics is a snapshot of the schedule outcome counters
294 : during the last schedule which included at least one successful
295 : outcome. */
296 : fd_pack_sched_results_snap_t last_sched_metrics[ 1 ];
297 :
298 : /* last_sched_metrics is a snapshot of the schedule outcome counters
299 : at the last start-of-leader-block event. */
300 : fd_pack_sched_results_snap_t start_block_sched_metrics[ 1 ];
301 :
302 : struct {
303 : ulong id;
304 : ulong txn_cnt;
305 : ulong txn_received;
306 : ulong min_blockhash_slot;
307 : fd_txn_e_t * _txn[ FD_PACK_MAX_TXN_PER_BUNDLE ];
308 : fd_txn_e_t * const * bundle; /* points to _txn when non-NULL */
309 : } current_bundle[1];
310 :
311 : block_builder_info_t blk_engine_cfg[1];
312 :
313 : struct {
314 : int enabled;
315 : int ib_inserted; /* in this slot */
316 : fd_acct_addr_t vote_pubkey[1];
317 : fd_acct_addr_t identity_pubkey[1];
318 : fd_bundle_crank_gen_t gen[1];
319 : fd_acct_addr_t tip_receiver_owner[1];
320 : ulong epoch;
321 : fd_bundle_crank_tip_payment_config_t prev_config[1]; /* as of start of slot, then updated */
322 : uchar recent_blockhash[32];
323 : fd_ed25519_sig_t last_sig[1];
324 :
325 : fd_keyswitch_t * keyswitch;
326 : fd_keyguard_client_t keyguard_client[1];
327 :
328 : ulong metrics[4];
329 : } crank[1];
330 :
331 :
332 : /* Used between during_frag and after_frag */
333 : ulong pending_rebate_sz;
334 : union{ fd_pack_rebate_t rebate[1]; uchar footprint[USHORT_MAX]; } rebate[1];
335 : } fd_pack_ctx_t;
336 :
337 0 : #define BUNDLE_META_SZ 40UL
338 : FD_STATIC_ASSERT( sizeof(block_builder_info_t)==BUNDLE_META_SZ, blk_engine_cfg );
339 :
340 0 : #define FD_PACK_METRIC_STATE_TRANSACTIONS 0
341 0 : #define FD_PACK_METRIC_STATE_EXECLES 1
342 0 : #define FD_PACK_METRIC_STATE_LEADER 2
343 0 : #define FD_PACK_METRIC_STATE_MICROBLOCKS 3
344 :
345 : /* Updates one component of the metric state. If the state has changed,
346 : records the change. */
347 : static inline void
348 : update_metric_state( fd_pack_ctx_t * ctx,
349 : long effective_as_of,
350 : int type,
351 0 : int status ) {
352 0 : uint current_state = fd_uint_insert_bit( ctx->metric_state, type, status );
353 0 : if( FD_UNLIKELY( current_state!=ctx->metric_state ) ) {
354 0 : ctx->metric_timing[ ctx->metric_state ] += effective_as_of - ctx->metric_state_begin;
355 0 : ctx->metric_state_begin = effective_as_of;
356 0 : ctx->metric_state = current_state;
357 0 : }
358 0 : }
359 :
360 : static inline void
361 0 : remove_ib( fd_pack_ctx_t * ctx ) {
362 : /* It's likely the initializer bundle is long scheduled, but we want to
363 : try deleting it just in case. */
364 0 : if( FD_UNLIKELY( ctx->crank->enabled & ctx->crank->ib_inserted ) ) {
365 0 : ulong deleted = fd_pack_delete_transaction( ctx->pack, (fd_ed25519_sig_t const *)ctx->crank->last_sig );
366 0 : FD_MCNT_INC( PACK, TXN_DELETED, deleted );
367 0 : }
368 0 : ctx->crank->ib_inserted = 0;
369 0 : }
370 :
371 :
372 : FD_FN_CONST static inline ulong
373 0 : scratch_align( void ) {
374 0 : return 4096UL;
375 0 : }
376 :
377 : FD_FN_PURE static inline ulong
378 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
379 0 : fd_pack_limits_t limits[1] = {{
380 0 : .max_cost_per_block = tile->pack.larger_max_cost_per_block ? LARGER_MAX_COST_PER_BLOCK : FD_PACK_MAX_COST_PER_BLOCK_UPPER_BOUND,
381 0 : .max_vote_cost_per_block = FD_PACK_MAX_VOTE_COST_PER_BLOCK_UPPER_BOUND,
382 0 : .max_write_cost_per_acct = FD_PACK_MAX_WRITE_COST_PER_ACCT_UPPER_BOUND,
383 0 : .max_data_bytes_per_block = tile->pack.larger_shred_limits_per_block ? LARGER_MAX_DATA_PER_BLOCK : FD_PACK_MAX_DATA_PER_BLOCK,
384 0 : .max_txn_per_microblock = EFFECTIVE_TXN_PER_MICROBLOCK,
385 0 : .max_microblocks_per_block = (ulong)UINT_MAX, /* Limit not known yet */
386 0 : .max_allocated_data_per_block = FD_PACK_MAX_ALLOCATED_DATA_PER_BLOCK,
387 0 : }};
388 :
389 0 : ulong l = FD_LAYOUT_INIT;
390 0 : l = FD_LAYOUT_APPEND( l, alignof( fd_pack_ctx_t ), sizeof( fd_pack_ctx_t ) );
391 0 : l = FD_LAYOUT_APPEND( l, fd_rng_align(), fd_rng_footprint() );
392 0 : l = FD_LAYOUT_APPEND( l, fd_pack_align(), fd_pack_footprint( tile->pack.max_pending_transactions,
393 0 : BUNDLE_META_SZ,
394 0 : tile->pack.execle_tile_count,
395 0 : limits ) );
396 : #if FD_PACK_USE_EXTRA_STORAGE
397 : l = FD_LAYOUT_APPEND( l, extra_txn_deq_align(), extra_txn_deq_footprint() );
398 : #endif
399 0 : return FD_LAYOUT_FINI( l, scratch_align() );
400 0 : }
401 :
402 : static inline void
403 : log_end_block_metrics( fd_pack_ctx_t * ctx,
404 : long now,
405 : char const * reason,
406 0 : ulong cus_consumed_in_block ) {
407 0 : #define DELTA( m ) (fd_metrics_tl[ MIDX(COUNTER, PACK, TXN_SCHEDULED_##m) ] - ctx->last_sched_metrics->sched_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_##m##_IDX ])
408 0 : #define AVAIL( m ) (fd_metrics_tl[ MIDX(GAUGE, PACK, TXN_AVAILABLE_##m) ])
409 0 : FD_LOG_INFO(( "pack_end_block(slot=%lu,%s,%lx,ticks_since_last_schedule=%ld,reasons=%lu,%lu,%lu,%lu,%lu,%lu,%lu,%lu;remaining=%lu+%lu+%lu+%lu;smallest=%lu;cus=%lu->%lu)",
410 0 : ctx->leader_slot, reason, ctx->execle_idle_bitset, now-ctx->last_sched_metrics->time,
411 0 : DELTA( TAKEN ), DELTA( CU_LIMIT ), DELTA( FAST_PATH ), DELTA( BYTE_LIMIT ), DELTA( ALLOC_LIMIT ), DELTA( WRITE_COST ), DELTA( SLOW_PATH ), DELTA( DEFER_SKIP ),
412 0 : AVAIL(REGULAR), AVAIL(VOTES), AVAIL(BUNDLES), AVAIL(CONFLICTING),
413 0 : (fd_metrics_tl[ MIDX(GAUGE, PACK, TXN_PENDING_SMALLEST_CU) ]),
414 0 : (cus_consumed_in_block),
415 0 : (fd_metrics_tl[ MIDX(GAUGE, PACK, BLOCK_CU_CONSUMED) ])
416 0 : ));
417 0 : #undef AVAIL
418 0 : #undef DELTA
419 0 : }
420 :
421 : static inline void
422 0 : get_done_packing( fd_pack_ctx_t * ctx, fd_done_packing_t * done_packing, int reason ) {
423 0 : done_packing->microblocks_in_slot = ctx->slot_microblock_cnt;
424 0 : done_packing->end_slot_reason = reason;
425 0 : fd_pack_get_block_limits( ctx->pack, done_packing->limits_usage, done_packing->limits );
426 :
427 0 : #define DELTA( mem, m ) (fd_metrics_tl[ MIDX(COUNTER, PACK, TXN_SCHEDULED_##m) ] - ctx->mem->sched_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_##m##_IDX ])
428 0 : done_packing->block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_TAKEN_IDX ] = DELTA( start_block_sched_metrics, TAKEN );
429 0 : done_packing->block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_CU_LIMIT_IDX ] = DELTA( start_block_sched_metrics, CU_LIMIT );
430 0 : done_packing->block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_FAST_PATH_IDX ] = DELTA( start_block_sched_metrics, FAST_PATH );
431 0 : done_packing->block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_BYTE_LIMIT_IDX ] = DELTA( start_block_sched_metrics, BYTE_LIMIT );
432 0 : done_packing->block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_ALLOC_LIMIT_IDX ] = DELTA( start_block_sched_metrics, ALLOC_LIMIT );
433 0 : done_packing->block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_WRITE_COST_IDX ] = DELTA( start_block_sched_metrics, WRITE_COST );
434 0 : done_packing->block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_SLOW_PATH_IDX ] = DELTA( start_block_sched_metrics, SLOW_PATH );
435 0 : done_packing->block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_DEFER_SKIP_IDX ] = DELTA( start_block_sched_metrics, DEFER_SKIP );
436 :
437 0 : done_packing->end_block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_TAKEN_IDX ] = DELTA( last_sched_metrics, TAKEN );
438 0 : done_packing->end_block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_CU_LIMIT_IDX ] = DELTA( last_sched_metrics, CU_LIMIT );
439 0 : done_packing->end_block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_FAST_PATH_IDX ] = DELTA( last_sched_metrics, FAST_PATH );
440 0 : done_packing->end_block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_BYTE_LIMIT_IDX ] = DELTA( last_sched_metrics, BYTE_LIMIT );
441 0 : done_packing->end_block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_ALLOC_LIMIT_IDX ] = DELTA( last_sched_metrics, ALLOC_LIMIT );
442 0 : done_packing->end_block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_WRITE_COST_IDX ] = DELTA( last_sched_metrics, WRITE_COST );
443 0 : done_packing->end_block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_SLOW_PATH_IDX ] = DELTA( last_sched_metrics, SLOW_PATH );
444 0 : done_packing->end_block_results[ FD_METRICS_ENUM_PACK_TXN_SCHEDULE_V_DEFER_SKIP_IDX ] = DELTA( last_sched_metrics, DEFER_SKIP );
445 0 : #undef DELTA
446 :
447 0 : fd_pack_get_pending_smallest( ctx->pack, done_packing->pending_smallest, done_packing->pending_votes_smallest );
448 0 : }
449 :
450 : static inline void
451 0 : metrics_write( fd_pack_ctx_t * ctx ) {
452 0 : FD_MCNT_ENUM_COPY( PACK, TXN_INSERTED, ctx->insert_result );
453 0 : FD_MCNT_ENUM_COPY( PACK, STATE_DURATION_NANOS, ((ulong*)ctx->metric_timing) );
454 0 : FD_MCNT_ENUM_COPY( PACK, BUNDLE_CRANK_RESULT, ctx->crank->metrics );
455 0 : FD_MHIST_COPY( PACK, SCHEDULE_MICROBLOCK_DURATION_SECONDS, ctx->schedule_duration );
456 0 : FD_MHIST_COPY( PACK, NO_SCHEDULE_MICROBLOCK_DURATION_SECONDS, ctx->no_sched_duration );
457 0 : FD_MHIST_COPY( PACK, INSERT_TRANSACTION_DURATION_SECONDS, ctx->insert_duration );
458 0 : FD_MHIST_COPY( PACK, COMPLETE_MICROBLOCK_DURATION_SECONDS, ctx->complete_duration );
459 :
460 0 : fd_pack_metrics_write( ctx->pack );
461 0 : }
462 :
463 : /* compute_dynamic_max_microblocks: Computes the upper bound on total
464 : microblocks based on remaining time and bank count.
465 :
466 : The basic idea here is that if there is 1ms left in the slot, we
467 : don't expect to schedule 130k microblocks. We can reduce our
468 : remaining budget which allows POH to advance hashing a bit further
469 : and avoid having to do a lot of hashing after the slot ends.
470 :
471 : The fastest we can execute a transaction is about 1us. With n
472 : execle tiles, that means we can execute at most n txn/us. If we have
473 : k ms left in the block, only reserve up to k*n*1000 microblocks. */
474 :
475 : static inline ulong
476 0 : compute_dynamic_max_microblocks( fd_pack_ctx_t * ctx ) {
477 0 : long now = ctx->approx_wallclock_ns + (long)((double)(fd_tickcount() - ctx->approx_tickcount) / ctx->ticks_per_ns);
478 0 : long end = ctx->slot_end_ns;
479 :
480 : /* If the slot has ended, don't reserve any more microblocks. */
481 0 : if( FD_UNLIKELY( now>=end ) ) return ctx->slot_microblock_cnt;
482 :
483 : /* remaining_ns * n / 1000 = (remaining_ns/1e6 ms) * n * 1000.
484 :
485 : Overflow: remaining_ns is at most ~4e8, n at most 64, so
486 : remaining_ns * n is at most ~2.6e10 << 1.8e19. */
487 :
488 0 : ulong remaining_ns = (ulong)(end - now);
489 0 : ulong n = ctx->execle_cnt;
490 0 : ulong cnt = ctx->slot_microblock_cnt;
491 0 : ulong R = ctx->slot_max_microblocks - cnt;
492 0 : ulong can_execute = remaining_ns * n / 1000UL;
493 :
494 0 : return cnt + fd_ulong_min( R, can_execute );
495 0 : }
496 :
497 : static inline void
498 0 : during_housekeeping( fd_pack_ctx_t * ctx ) {
499 0 : ctx->approx_wallclock_ns = fd_log_wallclock();
500 0 : ctx->approx_tickcount = fd_tickcount();
501 :
502 0 : if( FD_UNLIKELY( ctx->crank->enabled && fd_keyswitch_state_query( ctx->crank->keyswitch )==FD_KEYSWITCH_STATE_SWITCH_PENDING ) ) {
503 0 : fd_memcpy( ctx->crank->identity_pubkey, ctx->crank->keyswitch->bytes, 32UL );
504 0 : fd_keyswitch_state( ctx->crank->keyswitch, FD_KEYSWITCH_STATE_COMPLETED );
505 0 : }
506 :
507 0 : if( FD_LIKELY( ctx->leader_slot!=ULONG_MAX ) ) {
508 0 : ulong raw = compute_dynamic_max_microblocks( ctx );
509 0 : ulong prev = ctx->slot_dynamic_max_microblocks;
510 :
511 : /* Enforce monotonically decreasing. This ensures pack's bound is
512 : always smaller than POH's. */
513 0 : ctx->slot_dynamic_max_microblocks = fd_ulong_min( raw, prev );
514 :
515 : /* If the bound decreased, we must update pack's internal scheduling
516 : limit. Otherwise, pack could schedule a microblock that exceeds
517 : the remaining capacity from the tile's and poh's perspectives
518 : (e.g. pack might produce a 5-transaction microblock when only 4
519 : microblocks worth of space remain). */
520 0 : if( FD_UNLIKELY( ctx->slot_dynamic_max_microblocks < prev ) ) {
521 0 : fd_pack_limits_t limits[1];
522 0 : limits->max_cost_per_block = ctx->limits.slot_max_cost;
523 0 : limits->max_data_bytes_per_block = ctx->slot_max_data;
524 0 : limits->max_microblocks_per_block = ctx->slot_dynamic_max_microblocks;
525 0 : limits->max_vote_cost_per_block = ctx->limits.slot_max_vote_cost;
526 0 : limits->max_write_cost_per_acct = ctx->limits.slot_max_write_cost_per_acct;
527 0 : limits->max_txn_per_microblock = ULONG_MAX; /* unused */
528 0 : limits->max_allocated_data_per_block = ctx->limits.slot_max_allocated_data_per_block;
529 0 : fd_pack_set_block_limits( ctx->pack, limits );
530 :
531 0 : ctx->pending_reduce_mb_bound = 1; /* publish bound decrease */
532 0 : }
533 0 : }
534 0 : }
535 :
536 : static inline void
537 : before_credit( fd_pack_ctx_t * ctx,
538 : fd_stem_context_t * stem,
539 0 : int * charge_busy ) {
540 0 : (void)stem;
541 :
542 0 : if( FD_UNLIKELY( (ctx->cur_spot!=NULL) & !ctx->is_bundle ) ) {
543 0 : *charge_busy = 1;
544 :
545 : /* If we were overrun while processing a frag from an in, then
546 : cur_spot is left dangling and not cleaned up, so clean it up here
547 : (by returning the slot to the pool of free slots). If the last
548 : transaction was a bundle, then we don't want to return it. When
549 : we try to process the first transaction in the next bundle, we'll
550 : see we never got the full bundle and cancel the whole last
551 : bundle, returning all the storage to the pool. */
552 : #if FD_PACK_USE_EXTRA_STORAGE
553 : if( FD_LIKELY( !ctx->insert_to_extra ) ) fd_pack_insert_txn_cancel( ctx->pack, ctx->cur_spot );
554 : else extra_txn_deq_remove_tail( ctx->extra_txn_deq );
555 : #else
556 0 : fd_pack_insert_txn_cancel( ctx->pack, ctx->cur_spot );
557 0 : #endif
558 0 : ctx->cur_spot = NULL;
559 0 : }
560 0 : }
561 :
562 : #if FD_PACK_USE_EXTRA_STORAGE
563 : /* insert_from_extra: helper method to pop the transaction at the head
564 : off the extra txn deque and insert it into pack. Requires that
565 : ctx->extra_txn_deq is non-empty, but it's okay to call it if pack is
566 : full. Returns the result of fd_pack_insert_txn_fini. */
567 : static inline int
568 : insert_from_extra( fd_pack_ctx_t * ctx ) {
569 : fd_txn_e_t * spot = fd_pack_insert_txn_init( ctx->pack );
570 : fd_txn_e_t const * insert = extra_txn_deq_peek_head( ctx->extra_txn_deq );
571 : fd_txn_t const * insert_txn = TXN(insert->txnp);
572 : fd_memcpy( spot->txnp->payload, insert->txnp->payload, insert->txnp->payload_sz );
573 : fd_memcpy( TXN(spot->txnp), insert_txn, fd_txn_footprint( insert_txn->instr_cnt, insert_txn->addr_table_lookup_cnt ) );
574 : fd_memcpy( spot->alt_accts, insert->alt_accts, insert_txn->addr_table_adtl_cnt*sizeof(fd_acct_addr_t) );
575 : spot->txnp->payload_sz = insert->txnp->payload_sz;
576 : spot->txnp->source_tpu = insert->txnp->source_tpu;
577 : spot->txnp->source_ipv4 = insert->txnp->source_ipv4;
578 : spot->txnp->scheduler_arrival_time_nanos = insert->txnp->scheduler_arrival_time_nanos;
579 : spot->txnp->first_seen_nanos = insert->txnp->first_seen_nanos;
580 : extra_txn_deq_remove_head( ctx->extra_txn_deq );
581 :
582 : ulong blockhash_slot = insert->txnp->blockhash_slot;
583 :
584 : ulong deleted;
585 : long insert_duration = -fd_tickcount();
586 : int result = fd_pack_insert_txn_fini( ctx->pack, spot, blockhash_slot, &deleted );
587 : insert_duration += fd_tickcount();
588 :
589 : FD_MCNT_INC( PACK, TXN_DELETED, deleted );
590 : ctx->insert_result[ result + FD_PACK_INSERT_RETVAL_OFF ]++;
591 : fd_histf_sample( ctx->insert_duration, (ulong)insert_duration );
592 : FD_MCNT_INC( PACK, TXN_EXTRA_RETRIEVED, 1UL );
593 : return result;
594 : }
595 : #endif
596 :
597 : static inline void
598 : after_credit( fd_pack_ctx_t * ctx,
599 : fd_stem_context_t * stem,
600 : int * opt_poll_in FD_PARAM_UNUSED,
601 0 : int * charge_busy ) {
602 0 : if( FD_UNLIKELY( (ctx->skip_cnt--)>0L ) ) return; /* It would take ages for this to hit LONG_MIN */
603 :
604 0 : long now = fd_tickcount();
605 :
606 0 : int pacing_execle_cnt = (int)fd_pack_pacing_enabled_bank_cnt( ctx->pacer, now );
607 :
608 0 : ulong execle_cnt = ctx->execle_cnt;
609 :
610 :
611 : /* If any execle are busy, check one of the busy ones see if it is
612 : still busy. */
613 0 : if( FD_LIKELY( ctx->execle_idle_bitset!=fd_ulong_mask_lsb( (int)execle_cnt ) ) ) {
614 0 : int poll_cursor = ctx->poll_cursor;
615 0 : ulong busy_bitset = (~ctx->execle_idle_bitset) & fd_ulong_mask_lsb( (int)execle_cnt );
616 :
617 : /* Suppose execle_cnt is 4 and idle_bitset looks something like this
618 : (pretending it's a uchar):
619 : 0000 1001
620 : ^ busy cursor is 1
621 : Then busy_bitset is
622 : 0000 0110
623 : Rotate it right by 2 bits
624 : 1000 0001
625 : Find lsb returns 0, so busy cursor remains 2, and we poll
626 : execle 2.
627 :
628 : If instead idle_bitset were
629 : 0000 1110
630 : ^
631 : The rotated version would be
632 : 0100 0000
633 : Find lsb will return 6, so busy cursor would be set to 0, and
634 : we'd poll execle 0, which is the right one. */
635 0 : poll_cursor++;
636 0 : poll_cursor = (poll_cursor + fd_ulong_find_lsb( fd_ulong_rotate_right( busy_bitset, (poll_cursor&63) ) )) & 63;
637 :
638 0 : if( FD_UNLIKELY(
639 : /* if microblock duration is 0, bypass the execle_ready_at check
640 : to avoid a potential cache miss. Can't use an ifdef here
641 : because FD_UNLIKELY is a macro, but the compiler should
642 : eliminate the check easily. */
643 0 : ( (MICROBLOCK_DURATION_NS==0L) || (ctx->execle_ready_at[poll_cursor]<now) ) &&
644 0 : (fd_fseq_query( ctx->execle_current[poll_cursor] )==ctx->execle_expect[poll_cursor]) ) ) {
645 0 : *charge_busy = 1;
646 0 : ctx->execle_idle_bitset |= 1UL<<poll_cursor;
647 :
648 0 : long complete_duration = -fd_tickcount();
649 0 : int completed = fd_pack_microblock_complete( ctx->pack, (ulong)poll_cursor );
650 0 : complete_duration += fd_tickcount();
651 0 : if( FD_LIKELY( completed ) ) fd_histf_sample( ctx->complete_duration, (ulong)complete_duration );
652 0 : }
653 :
654 0 : ctx->poll_cursor = poll_cursor;
655 0 : }
656 :
657 :
658 : /* If we time out on our slot, then stop being leader. This can only
659 : happen in the first after_credit after a housekeeping. */
660 0 : if( FD_UNLIKELY( ctx->approx_wallclock_ns>=ctx->slot_end_ns && ctx->leader_slot!=ULONG_MAX ) ) {
661 0 : *charge_busy = 1;
662 :
663 0 : fd_done_packing_t * done_packing = fd_chunk_to_laddr( ctx->poh_out_mem, ctx->poh_out_chunk );
664 0 : get_done_packing( ctx, done_packing, FD_PACK_END_SLOT_REASON_TIME ); /* needs to be called before fd_pack_end_block */
665 0 : fd_pack_end_block( ctx->pack );
666 0 : fd_pack_get_top_writers( ctx->pack, done_packing->limits_usage->top_writers ); /* needs to be called after fd_pack_end_block */
667 :
668 0 : fd_stem_publish( stem, 1UL, fd_disco_execle_sig( ctx->leader_slot, ctx->pack_idx ), ctx->poh_out_chunk, sizeof(fd_done_packing_t), 0UL, 0UL, fd_frag_meta_ts_comp( fd_tickcount() ) );
669 0 : ctx->poh_out_chunk = fd_dcache_compact_next( ctx->poh_out_chunk, sizeof(fd_done_packing_t), ctx->poh_out_chunk0, ctx->poh_out_wmark );
670 0 : ctx->pack_idx++;
671 :
672 0 : log_end_block_metrics( ctx, now, "time", done_packing->limits_usage->block_cost );
673 0 : ctx->drain_execle = 1;
674 0 : ctx->leader_slot = ULONG_MAX;
675 0 : ctx->slot_microblock_cnt = 0UL;
676 0 : remove_ib( ctx );
677 :
678 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_LEADER, 0 );
679 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_EXECLES, 0 );
680 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_MICROBLOCKS, 0 );
681 0 : return;
682 0 : }
683 :
684 : /* Am I leader? If not, see about inserting at most one transaction
685 : from extra storage. It's important not to insert too many
686 : transactions here, or we won't end up servicing dedup_pack enough.
687 : If extra storage is empty or pack is full, do nothing. */
688 0 : if( FD_UNLIKELY( ctx->leader_slot==ULONG_MAX ) ) {
689 : #if FD_PACK_USE_EXTRA_STORAGE
690 : if( FD_UNLIKELY( !extra_txn_deq_empty( ctx->extra_txn_deq ) &&
691 : fd_pack_avail_txn_cnt( ctx->pack )<ctx->max_pending_transactions ) ) {
692 : *charge_busy = 1;
693 :
694 : int result = insert_from_extra( ctx );
695 : if( FD_LIKELY( result>=0 ) ) ctx->last_successful_insert = now;
696 : }
697 : #endif
698 0 : return;
699 0 : }
700 :
701 : /* Am I in drain mode? If so, check if I can exit it */
702 0 : if( FD_UNLIKELY( ctx->drain_execle ) ) {
703 0 : if( FD_LIKELY( ctx->execle_idle_bitset==fd_ulong_mask_lsb( (int)execle_cnt ) ) ) {
704 0 : ctx->drain_execle = 0;
705 :
706 : /* Pack notifies poh when execle are drained so that poh can
707 : relinquish pack's ownership over the slot execle (by decrementing
708 : its Arc). We do this by sending a FD_PACK_MSG_DONE_DRAINING
709 : sig over the pack_poh mcache.
710 :
711 : TODO: This is only needed for Frankendancer, not Firedancer,
712 : which manages bank lifetime different. */
713 0 : fd_stem_publish( stem, 1UL, FD_PACK_MSG_DONE_DRAINING, 0UL, 0UL, 0UL, 0UL, fd_frag_meta_ts_comp( fd_tickcount() ) );
714 0 : *charge_busy = 1;
715 0 : return;
716 0 : } else {
717 0 : return;
718 0 : }
719 0 : }
720 :
721 0 : if( FD_UNLIKELY( ctx->pending_reduce_mb_bound ) ) {
722 0 : ctx->pending_reduce_mb_bound = 0;
723 0 : ulong * dst = fd_chunk_to_laddr( ctx->poh_out_mem, ctx->poh_out_chunk );
724 0 : *dst = ctx->slot_dynamic_max_microblocks;
725 0 : fd_stem_publish( stem, 1UL, FD_PACK_MSG_REDUCE_MB_BOUND, ctx->poh_out_chunk, sizeof(ulong), 0UL, 0UL, fd_frag_meta_ts_comp( fd_tickcount() ) );
726 0 : ctx->poh_out_chunk = fd_dcache_compact_next( ctx->poh_out_chunk, sizeof(ulong), ctx->poh_out_chunk0, ctx->poh_out_wmark );
727 0 : *charge_busy = 1;
728 0 : return;
729 0 : }
730 :
731 : /* Have I sent the max allowed microblocks? Nothing to do. */
732 0 : if( FD_UNLIKELY( ctx->slot_microblock_cnt>=ctx->slot_dynamic_max_microblocks ) ) return;
733 :
734 : /* Do I have enough transactions and/or have I waited enough time? */
735 : #if !SMALL_MICROBLOCKS
736 : if( FD_UNLIKELY( (ulong)(now-ctx->last_successful_insert) <
737 : ctx->wait_duration_ticks[ fd_ulong_min( fd_pack_avail_txn_cnt( ctx->pack ), MAX_TXN_PER_MICROBLOCK ) ] ) ) {
738 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_TRANSACTIONS, 0 );
739 : return;
740 : }
741 : #endif
742 :
743 0 : int any_ready = 0;
744 0 : int any_scheduled = 0;
745 :
746 0 : *charge_busy = 1;
747 :
748 0 : if( FD_LIKELY( ctx->crank->enabled ) ) {
749 0 : block_builder_info_t const * top_meta = fd_pack_peek_bundle_meta( ctx->pack );
750 0 : if( FD_UNLIKELY( top_meta ) ) {
751 : /* Have bundles, in a reasonable state to crank. */
752 :
753 0 : fd_txn_e_t * _bundle[ 1UL ];
754 0 : fd_txn_e_t * const * bundle = fd_pack_insert_bundle_init( ctx->pack, _bundle, 1UL );
755 :
756 0 : ulong txn_sz = fd_bundle_crank_generate( ctx->crank->gen, ctx->crank->prev_config, top_meta->commission_pubkey,
757 0 : ctx->crank->identity_pubkey, ctx->crank->tip_receiver_owner, ctx->crank->epoch, top_meta->commission,
758 0 : bundle[0]->txnp->payload, TXN( bundle[0]->txnp ) );
759 :
760 0 : if( FD_LIKELY( txn_sz==0UL ) ) { /* Everything in good shape! */
761 0 : fd_pack_insert_bundle_cancel( ctx->pack, bundle, 1UL );
762 0 : fd_pack_set_initializer_bundles_ready( ctx->pack );
763 0 : ctx->crank->metrics[ 0 ]++; /* BUNDLE_CRANK_RESULT_NOT_NEEDED */
764 0 : }
765 0 : else if( FD_LIKELY( txn_sz<ULONG_MAX ) ) {
766 0 : bundle[0]->txnp->payload_sz = (ushort)txn_sz;
767 0 : bundle[0]->txnp->source_tpu = FD_TXN_M_TPU_SOURCE_BUNDLE;
768 0 : bundle[0]->txnp->source_ipv4 = 0; /* not applicable */
769 0 : bundle[0]->txnp->scheduler_arrival_time_nanos = ctx->approx_wallclock_ns + (long)((double)(fd_tickcount() - ctx->approx_tickcount) / ctx->ticks_per_ns);
770 0 : bundle[0]->txnp->first_seen_nanos = bundle[0]->txnp->scheduler_arrival_time_nanos;
771 0 : memcpy( bundle[0]->txnp->payload+TXN(bundle[0]->txnp)->recent_blockhash_off, ctx->crank->recent_blockhash, 32UL );
772 :
773 0 : fd_keyguard_client_sign( ctx->crank->keyguard_client, bundle[0]->txnp->payload+1UL,
774 0 : bundle[0]->txnp->payload+65UL, txn_sz-65UL, FD_KEYGUARD_SIGN_TYPE_ED25519 );
775 :
776 0 : memcpy( ctx->crank->last_sig, bundle[0]->txnp->payload+1UL, 64UL );
777 :
778 0 : ctx->crank->ib_inserted = 1;
779 0 : ulong deleted;
780 0 : int retval = fd_pack_insert_bundle_fini( ctx->pack, bundle, 1UL, ctx->leader_slot-1UL, 1, NULL, &deleted );
781 0 : FD_MCNT_INC( PACK, TXN_DELETED, deleted );
782 0 : ctx->insert_result[ retval + FD_PACK_INSERT_RETVAL_OFF ]++;
783 0 : if( FD_UNLIKELY( retval<0 ) ) {
784 0 : ctx->crank->metrics[ 3 ]++; /* BUNDLE_CRANK_RESULT_INSERTION_FAILED */
785 0 : FD_LOG_WARNING(( "inserting initializer bundle returned %i", retval ));
786 0 : } else {
787 : /* Update the cached copy of the on-chain state. This seems a
788 : little dangerous, since we're updating it as if the bundle
789 : succeeded without knowing if that's true, but here's why
790 : it's safe:
791 :
792 : From now until we get the rebate call for this initializer
793 : bundle (which lets us know if it succeeded or failed), pack
794 : will be in [Pending] state, which means peek_bundle_meta
795 : will return NULL, so we won't read this state.
796 :
797 : Then, if the initializer bundle failed, we'll go into
798 : [Failed] IB state until the end of the block, which will
799 : cause top_meta to remain NULL so we don't read these values
800 : again.
801 :
802 : Otherwise, the initializer bundle succeeded, which means
803 : that these are the right values to use. */
804 0 : fd_bundle_crank_apply( ctx->crank->gen, ctx->crank->prev_config, top_meta->commission_pubkey,
805 0 : ctx->crank->tip_receiver_owner, ctx->crank->epoch, top_meta->commission );
806 0 : ctx->crank->metrics[ 1 ]++; /* BUNDLE_CRANK_RESULT_INSERTED */
807 0 : }
808 0 : } else {
809 : /* Already logged a warning in this case */
810 0 : fd_pack_insert_bundle_cancel( ctx->pack, bundle, 1UL );
811 0 : ctx->crank->metrics[ 2 ]++; /* BUNDLE_CRANK_RESULT_CREATION_FAILED' */
812 0 : }
813 0 : }
814 0 : }
815 :
816 : /* Try to schedule the next microblock. */
817 0 : if( FD_LIKELY( ctx->execle_idle_bitset ) ) { /* Optimize for schedule */
818 0 : any_ready = 1;
819 :
820 0 : int i = fd_ulong_find_lsb( ctx->execle_idle_bitset );
821 :
822 0 : int flags;
823 :
824 0 : switch( ctx->strategy ) {
825 0 : default:
826 0 : case FD_PACK_STRATEGY_PERF:
827 0 : flags = FD_PACK_SCHEDULE_VOTE | FD_PACK_SCHEDULE_BUNDLE | FD_PACK_SCHEDULE_TXN;
828 0 : break;
829 0 : case FD_PACK_STRATEGY_BALANCED:
830 : /* We want to exempt votes from pacing, so we always allow
831 : scheduling votes. It doesn't really make much sense to pace
832 : bundles, because they get scheduled in FIFO order. However,
833 : we keep pacing for normal transactions. For example, if
834 : pacing_execle_cnt is 0, then pack won't schedule normal
835 : transactions to any execle tile. */
836 0 : flags = FD_PACK_SCHEDULE_VOTE | fd_int_if( i==0, FD_PACK_SCHEDULE_BUNDLE, 0 )
837 0 : | fd_int_if( i<pacing_execle_cnt, FD_PACK_SCHEDULE_TXN, 0 );
838 0 : break;
839 0 : }
840 :
841 0 : fd_txn_e_t * microblock_dst = fd_chunk_to_laddr( ctx->execle_out_mem, ctx->execle_out_chunk );
842 0 : long schedule_duration = -fd_tickcount();
843 0 : ulong schedule_cnt = fd_pack_schedule_next_microblock( ctx->pack, CUS_PER_MICROBLOCK, VOTE_FRACTION, (ulong)i, flags, microblock_dst );
844 0 : schedule_duration += fd_tickcount();
845 0 : fd_histf_sample( (schedule_cnt>0UL) ? ctx->schedule_duration : ctx->no_sched_duration, (ulong)schedule_duration );
846 :
847 0 : if( FD_LIKELY( schedule_cnt ) ) {
848 0 : any_scheduled = 1;
849 0 : long now2 = fd_tickcount();
850 0 : ulong tsorig = (ulong)fd_frag_meta_ts_comp( now ); /* A bound on when we observed execle was idle */
851 0 : ulong tspub = (ulong)fd_frag_meta_ts_comp( now2 );
852 0 : ulong chunk = ctx->execle_out_chunk;
853 0 : ulong msg_sz = schedule_cnt*sizeof(fd_txn_e_t);
854 0 : fd_microblock_execle_trailer_t * trailer = (fd_microblock_execle_trailer_t*)(microblock_dst+schedule_cnt);
855 0 : trailer->bank = ctx->leader_bank;
856 0 : trailer->bank_idx = ctx->leader_bank_idx;
857 0 : trailer->bank_seq = ctx->leader_bank_seq;
858 0 : trailer->microblock_idx = ctx->slot_microblock_cnt;
859 0 : trailer->pack_idx = ctx->pack_idx;
860 0 : trailer->pack_txn_idx = ctx->pack_txn_cnt;
861 0 : trailer->is_bundle = !!(microblock_dst->txnp->flags & FD_TXN_P_FLAGS_BUNDLE);
862 :
863 : /* When sending MAX_TXN_PER_MICROBLOCK transactions as fd_txn_e_t
864 : to execle, there must be room for the trailer at the end. */
865 0 : FD_STATIC_ASSERT( MAX_TXN_PER_MICROBLOCK*sizeof(fd_txn_e_t)+sizeof(fd_microblock_execle_trailer_t)<=MAX_MICROBLOCK_SZ, pack_execle_mtu );
866 :
867 0 : ulong sig = fd_disco_poh_sig( ctx->leader_slot, POH_PKT_TYPE_MICROBLOCK, (ulong)i );
868 0 : fd_stem_publish( stem, 0UL, sig, chunk, msg_sz+sizeof(fd_microblock_execle_trailer_t), 0UL, tsorig, tspub );
869 0 : ctx->execle_expect[ i ] = stem->seqs[0]-1UL;
870 0 : ctx->execle_ready_at[i] = now2 + (long)ctx->microblock_duration_ticks;
871 0 : ctx->execle_out_chunk = fd_dcache_compact_next( ctx->execle_out_chunk, msg_sz+sizeof(fd_microblock_execle_trailer_t), ctx->execle_out_chunk0, ctx->execle_out_wmark );
872 0 : ctx->slot_microblock_cnt += fd_ulong_if( trailer->is_bundle, schedule_cnt, 1UL );
873 0 : ctx->pack_idx += fd_uint_if( trailer->is_bundle, (uint)schedule_cnt, 1U );
874 0 : ctx->pack_txn_cnt += schedule_cnt;
875 :
876 0 : ctx->execle_idle_bitset = fd_ulong_pop_lsb( ctx->execle_idle_bitset );
877 0 : ctx->skip_cnt = (long)schedule_cnt * fd_long_if( ctx->use_consumed_cus, (long)execle_cnt/2L, 1L );
878 0 : fd_pack_pacing_update_consumed_cus( ctx->pacer, fd_pack_current_block_cost( ctx->pack ), now2 );
879 :
880 0 : ctx->last_sched_metrics->time = now2;
881 0 : fd_pack_get_sched_metrics( ctx->pack, ctx->last_sched_metrics->sched_results );
882 :
883 : /* If we're using CU rebates, then we have one in for each execle
884 : in addition to the two normal ones. We want to skip schedule
885 : attempts for (execle_cnt + 1) link polls after a successful
886 : schedule attempt. */
887 0 : fd_long_store_if( ctx->use_consumed_cus, &(ctx->skip_cnt), (long)(ctx->execle_cnt + 1) );
888 0 : }
889 0 : }
890 :
891 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_EXECLES, any_ready );
892 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_MICROBLOCKS, any_scheduled );
893 0 : now = fd_tickcount();
894 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_TRANSACTIONS, fd_pack_avail_txn_cnt( ctx->pack )>0 );
895 :
896 : #if FD_PACK_USE_EXTRA_STORAGE
897 : if( FD_UNLIKELY( !extra_txn_deq_empty( ctx->extra_txn_deq ) ) ) {
898 : /* Don't start pulling from the extra storage until the available
899 : transaction count drops below half. */
900 : ulong avail_space = (ulong)fd_long_max( 0L, (long)(ctx->max_pending_transactions>>1)-(long)fd_pack_avail_txn_cnt( ctx->pack ) );
901 : ulong qty_to_insert = fd_ulong_min( 10UL, fd_ulong_min( extra_txn_deq_cnt( ctx->extra_txn_deq ), avail_space ) );
902 : int any_successes = 0;
903 : for( ulong i=0UL; i<qty_to_insert; i++ ) any_successes |= (0<=insert_from_extra( ctx ));
904 : if( FD_LIKELY( any_successes ) ) ctx->last_successful_insert = now;
905 : }
906 : #endif
907 :
908 : /* Did we send the maximum allowed microblocks? Then end the slot. */
909 0 : if( FD_UNLIKELY( ctx->slot_microblock_cnt==ctx->slot_dynamic_max_microblocks )) {
910 0 : *charge_busy = 1;
911 :
912 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_LEADER, 0 );
913 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_EXECLES, 0 );
914 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_MICROBLOCKS, 0 );
915 : /* The pack object also does this accounting and increases this
916 : metric, but we end the slot early so won't see it unless we also
917 : increment it here. */
918 0 : FD_MCNT_INC( PACK, MICROBLOCK_PER_BLOCK_LIMIT_REACHED, 1UL );
919 :
920 0 : fd_done_packing_t * done_packing = fd_chunk_to_laddr( ctx->poh_out_mem, ctx->poh_out_chunk );
921 0 : get_done_packing( ctx, done_packing, FD_PACK_END_SLOT_REASON_MICROBLOCK );
922 0 : fd_pack_end_block( ctx->pack );
923 0 : fd_pack_get_top_writers( ctx->pack, done_packing->limits_usage->top_writers );
924 :
925 0 : fd_stem_publish( stem, 1UL, fd_disco_execle_sig( ctx->leader_slot, ctx->pack_idx ), ctx->poh_out_chunk, sizeof(fd_done_packing_t), 0UL, 0UL, fd_frag_meta_ts_comp( fd_tickcount() ) );
926 0 : ctx->poh_out_chunk = fd_dcache_compact_next( ctx->poh_out_chunk, sizeof(fd_done_packing_t), ctx->poh_out_chunk0, ctx->poh_out_wmark );
927 0 : ctx->pack_idx++;
928 :
929 0 : log_end_block_metrics( ctx, now, "microblock", done_packing->limits_usage->block_cost );
930 0 : ctx->drain_execle = 1;
931 0 : ctx->leader_slot = ULONG_MAX;
932 0 : ctx->slot_microblock_cnt = 0UL;
933 0 : remove_ib( ctx );
934 :
935 0 : return;
936 0 : }
937 0 : }
938 :
939 :
940 : /* At this point, we have started receiving frag seq with details in
941 : mline at time now. Speculatively process it here. */
942 :
943 : static inline void
944 : during_frag( fd_pack_ctx_t * ctx,
945 : ulong in_idx,
946 : ulong seq FD_PARAM_UNUSED,
947 : ulong sig,
948 : ulong chunk,
949 : ulong sz,
950 0 : ulong ctl FD_PARAM_UNUSED ) {
951 :
952 0 : uchar const * dcache_entry = fd_chunk_to_laddr_const( ctx->in[ in_idx ].mem, chunk );
953 :
954 0 : switch( ctx->in_kind[ in_idx ] ) {
955 0 : case IN_KIND_REPLAY: {
956 0 : if( FD_LIKELY( sig==REPLAY_SIG_TXN_EXECUTED ) ) {
957 0 : fd_replay_txn_executed_t const * txn_executed = fd_type_pun_const( dcache_entry );
958 0 : ctx->txn_committed = !!txn_executed->is_committable;
959 0 : if( FD_UNLIKELY( !txn_executed->is_committable ) ) return;
960 0 : memcpy( ctx->executed_txn_sig, fd_txn_get_signatures( TXN(txn_executed->txn), txn_executed->txn->payload ), FD_TXN_SIGNATURE_SZ );
961 0 : return;
962 0 : }
963 0 : if( FD_LIKELY( sig!=REPLAY_SIG_BECAME_LEADER ) ) return;
964 :
965 : /* There was a leader transition. Handle it. */
966 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz!=sizeof(fd_became_leader_t) ) )
967 0 : FD_LOG_ERR(( "chunk %lu %lu corrupt, not in range [%lu,%lu]", chunk, sz, ctx->in[ in_idx ].chunk0, ctx->in[ in_idx ].wmark ));
968 :
969 0 : fd_memcpy( ctx->_became_leader, dcache_entry, sizeof(fd_became_leader_t) );
970 0 : return;
971 0 : }
972 0 : case IN_KIND_POH: {
973 : /* Not interested in stamped microblocks, only leader updates. */
974 0 : if( fd_disco_poh_sig_pkt_type( sig )!=POH_PKT_TYPE_BECAME_LEADER ) return;
975 :
976 : /* There was a leader transition. Handle it. */
977 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz!=sizeof(fd_became_leader_t) ) )
978 0 : FD_LOG_ERR(( "chunk %lu %lu corrupt, not in range [%lu,%lu]", chunk, sz, ctx->in[ in_idx ].chunk0, ctx->in[ in_idx ].wmark ));
979 :
980 0 : fd_memcpy( ctx->_became_leader, dcache_entry, sizeof(fd_became_leader_t) );
981 0 : return;
982 0 : }
983 0 : case IN_KIND_EXECLE: {
984 0 : FD_TEST( ctx->use_consumed_cus );
985 : /* For a previous slot */
986 0 : if( FD_UNLIKELY( sig!=ctx->leader_slot ) ) return;
987 :
988 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz<FD_PACK_REBATE_MIN_SZ
989 0 : || sz>FD_PACK_REBATE_MAX_SZ ) )
990 0 : FD_LOG_ERR(( "chunk %lu %lu corrupt, not in range [%lu,%lu]", chunk, sz, ctx->in[ in_idx ].chunk0, ctx->in[ in_idx ].wmark ));
991 :
992 0 : ctx->pending_rebate_sz = sz;
993 0 : fd_memcpy( ctx->rebate, dcache_entry, sz );
994 0 : return;
995 0 : }
996 0 : case IN_KIND_RESOLV: {
997 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz>FD_TPU_RESOLVED_MTU ) )
998 0 : FD_LOG_ERR(( "chunk %lu %lu corrupt, not in range [%lu,%lu]", chunk, sz, ctx->in[ in_idx ].chunk0, ctx->in[ in_idx ].wmark ));
999 :
1000 0 : fd_txn_m_t * txnm = (fd_txn_m_t *)dcache_entry;
1001 0 : ulong payload_sz = txnm->payload_sz;
1002 0 : ulong txn_t_sz = txnm->txn_t_sz;
1003 0 : uint source_ipv4 = txnm->source_ipv4;
1004 0 : uchar source_tpu = txnm->source_tpu;
1005 0 : FD_TEST( payload_sz<=FD_TPU_MTU );
1006 0 : FD_TEST( txn_t_sz <=FD_TXN_MAX_SZ );
1007 0 : fd_txn_t * txn = fd_txn_m_txn_t( txnm );
1008 :
1009 0 : ulong addr_table_sz = 32UL*txn->addr_table_adtl_cnt;
1010 0 : FD_TEST( addr_table_sz<=32UL*FD_TXN_ACCT_ADDR_MAX );
1011 :
1012 0 : if( FD_UNLIKELY( (ctx->leader_slot==ULONG_MAX) & (sig>ctx->highest_observed_slot) ) ) {
1013 : /* Using the resolv tile's knowledge of the current slot is a bit
1014 : of a hack, since we don't get any info if there are no
1015 : transactions and we're not leader. We're actually in exactly
1016 : the case where that's okay though. The point of calling
1017 : expire_before long before we become leader is so that we don't
1018 : drop new but low-fee-paying transactions when pack is clogged
1019 : with expired but high-fee-paying transactions. That can only
1020 : happen if we are getting transactions. */
1021 0 : ctx->highest_observed_slot = sig;
1022 0 : ulong exp_cnt = fd_pack_expire_before( ctx->pack, fd_ulong_max( ctx->highest_observed_slot, TRANSACTION_LIFETIME_SLOTS )-TRANSACTION_LIFETIME_SLOTS );
1023 0 : FD_MCNT_INC( PACK, TXN_EXPIRED, exp_cnt );
1024 0 : }
1025 :
1026 :
1027 0 : ulong bundle_id = txnm->block_engine.bundle_id;
1028 0 : if( FD_UNLIKELY( bundle_id ) ) {
1029 0 : ctx->is_bundle = 1;
1030 0 : if( FD_LIKELY( bundle_id!=ctx->current_bundle->id ) ) {
1031 0 : if( FD_UNLIKELY( ctx->current_bundle->bundle ) ) {
1032 0 : FD_MCNT_INC( PACK, TXN_PARTIAL_BUNDLE, ctx->current_bundle->txn_received );
1033 0 : fd_pack_insert_bundle_cancel( ctx->pack, ctx->current_bundle->bundle, ctx->current_bundle->txn_cnt );
1034 0 : }
1035 0 : ctx->current_bundle->id = bundle_id;
1036 0 : ctx->current_bundle->txn_cnt = txnm->block_engine.bundle_txn_cnt;
1037 0 : ctx->current_bundle->min_blockhash_slot = ULONG_MAX;
1038 0 : ctx->current_bundle->txn_received = 0UL;
1039 :
1040 0 : if( FD_UNLIKELY( ctx->current_bundle->txn_cnt==0UL ) ) {
1041 0 : FD_MCNT_INC( PACK, TXN_PARTIAL_BUNDLE, 1UL );
1042 0 : ctx->current_bundle->id = 0UL;
1043 0 : return;
1044 0 : }
1045 0 : ctx->blk_engine_cfg->commission = txnm->block_engine.commission;
1046 0 : memcpy( ctx->blk_engine_cfg->commission_pubkey->b, txnm->block_engine.commission_pubkey, 32UL );
1047 :
1048 0 : ctx->current_bundle->bundle = fd_pack_insert_bundle_init( ctx->pack, ctx->current_bundle->_txn, ctx->current_bundle->txn_cnt );
1049 0 : }
1050 0 : ctx->cur_spot = ctx->current_bundle->bundle[ ctx->current_bundle->txn_received ];
1051 0 : ctx->current_bundle->min_blockhash_slot = fd_ulong_min( ctx->current_bundle->min_blockhash_slot, sig );
1052 0 : } else {
1053 0 : ctx->is_bundle = 0;
1054 : #if FD_PACK_USE_EXTRA_STORAGE
1055 : if( FD_LIKELY( ctx->leader_slot!=ULONG_MAX || fd_pack_avail_txn_cnt( ctx->pack )<ctx->max_pending_transactions ) ) {
1056 : ctx->cur_spot = fd_pack_insert_txn_init( ctx->pack );
1057 : ctx->insert_to_extra = 0;
1058 : } else {
1059 : if( FD_UNLIKELY( extra_txn_deq_full( ctx->extra_txn_deq ) ) ) {
1060 : extra_txn_deq_remove_head( ctx->extra_txn_deq );
1061 : FD_MCNT_INC( PACK, TXN_EXTRA_DROPPED, 1UL );
1062 : }
1063 : ctx->cur_spot = extra_txn_deq_peek_tail( extra_txn_deq_insert_tail( ctx->extra_txn_deq ) );
1064 : /* We want to store the current time in cur_spot so that we can
1065 : track its expiration better. We just stash it in the CU
1066 : fields, since those aren't important right now. */
1067 : ctx->cur_spot->txnp->blockhash_slot = sig;
1068 : ctx->insert_to_extra = 1;
1069 : FD_MCNT_INC( PACK, TXN_EXTRA_INSERTED, 1UL );
1070 : }
1071 : #else
1072 0 : ctx->cur_spot = fd_pack_insert_txn_init( ctx->pack );
1073 0 : #endif
1074 0 : }
1075 :
1076 : /* We get transactions from the resolv tile.
1077 : The transactions should have been parsed and verified. */
1078 0 : FD_MCNT_INC( PACK, TXN_NORMAL_RX, 1UL );
1079 :
1080 0 : fd_memcpy( ctx->cur_spot->txnp->payload, fd_txn_m_payload( txnm ), payload_sz );
1081 0 : fd_memcpy( TXN(ctx->cur_spot->txnp), txn, txn_t_sz );
1082 0 : fd_memcpy( ctx->cur_spot->alt_accts, fd_txn_m_alut( txnm ), addr_table_sz );
1083 0 : ctx->cur_spot->txnp->scheduler_arrival_time_nanos = ctx->approx_wallclock_ns + (long)((double)(fd_tickcount() - ctx->approx_tickcount) / ctx->ticks_per_ns);
1084 0 : ctx->cur_spot->txnp->first_seen_nanos = txnm->first_seen_nanos;
1085 0 : ctx->cur_spot->txnp->payload_sz = payload_sz;
1086 0 : ctx->cur_spot->txnp->source_ipv4 = source_ipv4;
1087 0 : ctx->cur_spot->txnp->source_tpu = source_tpu;
1088 :
1089 0 : break;
1090 0 : }
1091 0 : case IN_KIND_EXECUTED_TXN: {
1092 0 : FD_TEST( sz==64UL );
1093 0 : fd_memcpy( ctx->executed_txn_sig, dcache_entry, sz );
1094 0 : break;
1095 0 : }
1096 0 : }
1097 0 : }
1098 :
1099 :
1100 : /* After the transaction has been fully received, and we know we were
1101 : not overrun while reading it, insert it into pack. */
1102 :
1103 : static inline void
1104 : after_frag( fd_pack_ctx_t * ctx,
1105 : ulong in_idx,
1106 : ulong seq,
1107 : ulong sig,
1108 : ulong sz,
1109 : ulong tsorig,
1110 : ulong tspub,
1111 0 : fd_stem_context_t * stem ) {
1112 0 : (void)seq;
1113 0 : (void)sz;
1114 0 : (void)tsorig;
1115 0 : (void)tspub;
1116 0 : (void)stem;
1117 :
1118 0 : long now = fd_tickcount();
1119 :
1120 0 : ulong leader_slot = ULONG_MAX;
1121 0 : switch( ctx->in_kind[ in_idx ] ) {
1122 0 : case IN_KIND_REPLAY:
1123 0 : if( FD_LIKELY( sig==REPLAY_SIG_TXN_EXECUTED && ctx->txn_committed ) ) {
1124 0 : ulong deleted = fd_pack_delete_transaction( ctx->pack, fd_type_pun( ctx->executed_txn_sig ) );
1125 0 : FD_MCNT_INC( PACK, TXN_ALREADY_EXECUTED, deleted );
1126 0 : }
1127 0 : if( FD_UNLIKELY( sig!=REPLAY_SIG_BECAME_LEADER ) ) return;
1128 0 : leader_slot = ctx->_became_leader->slot;
1129 :
1130 0 : ctx->start_block_sched_metrics->time = now;
1131 0 : fd_pack_get_sched_metrics( ctx->pack, ctx->start_block_sched_metrics->sched_results );
1132 0 : break;
1133 0 : case IN_KIND_POH:
1134 0 : if( fd_disco_poh_sig_pkt_type( sig )!=POH_PKT_TYPE_BECAME_LEADER ) return;
1135 0 : leader_slot = fd_disco_poh_sig_slot( sig );
1136 0 : break;
1137 0 : default:
1138 0 : break;
1139 0 : }
1140 :
1141 0 : switch( ctx->in_kind[ in_idx ] ) {
1142 0 : case IN_KIND_REPLAY:
1143 0 : case IN_KIND_POH: {
1144 0 : long now_ticks = fd_tickcount();
1145 0 : long now_ns = fd_log_wallclock();
1146 :
1147 0 : if( FD_UNLIKELY( ctx->leader_slot!=ULONG_MAX ) ) {
1148 0 : fd_done_packing_t * done_packing = fd_chunk_to_laddr( ctx->poh_out_mem, ctx->poh_out_chunk );
1149 0 : get_done_packing( ctx, done_packing, FD_PACK_END_SLOT_REASON_LEADER_SWITCH );
1150 0 : fd_pack_end_block( ctx->pack );
1151 0 : fd_pack_get_top_writers( ctx->pack, done_packing->limits_usage->top_writers );
1152 :
1153 0 : fd_stem_publish( stem, 1UL, fd_disco_execle_sig( ctx->leader_slot, ctx->pack_idx ), ctx->poh_out_chunk, sizeof(fd_done_packing_t), 0UL, 0UL, fd_frag_meta_ts_comp( fd_tickcount() ) );
1154 0 : ctx->poh_out_chunk = fd_dcache_compact_next( ctx->poh_out_chunk, sizeof(fd_done_packing_t), ctx->poh_out_chunk0, ctx->poh_out_wmark );
1155 0 : ctx->pack_idx++;
1156 :
1157 0 : FD_LOG_WARNING(( "switching to slot %lu while packing for slot %lu. Draining execle tiles.", leader_slot, ctx->leader_slot ));
1158 0 : log_end_block_metrics( ctx, now_ticks, "switch", done_packing->limits_usage->block_cost );
1159 0 : ctx->drain_execle = 1;
1160 0 : ctx->leader_slot = ULONG_MAX;
1161 0 : ctx->slot_microblock_cnt = 0UL;
1162 0 : remove_ib( ctx );
1163 0 : }
1164 0 : ctx->leader_slot = leader_slot;
1165 :
1166 0 : ulong exp_cnt = fd_pack_expire_before( ctx->pack, fd_ulong_max( ctx->leader_slot, TRANSACTION_LIFETIME_SLOTS )-TRANSACTION_LIFETIME_SLOTS );
1167 0 : FD_MCNT_INC( PACK, TXN_EXPIRED, exp_cnt );
1168 :
1169 0 : ctx->leader_bank = ctx->_became_leader->bank;
1170 0 : ctx->leader_bank_idx = ctx->_became_leader->bank_idx;
1171 0 : ctx->leader_bank_seq = ctx->_became_leader->bank_seq;
1172 0 : ctx->slot_max_microblocks = ctx->_became_leader->max_microblocks_in_slot;
1173 :
1174 0 : ulong base_max_data = ctx->larger_shred_limits_per_block ? LARGER_MAX_DATA_PER_BLOCK : FD_PACK_MAX_DATA_PER_BLOCK;
1175 0 : if( FD_LIKELY( !ctx->larger_shred_limits_per_block ) ) {
1176 : /* Compute base_max_data to ensure that we don't overflow
1177 : slot_max_data_shreds. See FD_SHRED_BATCH_BLOCK_DATA_SZ_MAX in
1178 : fd_shred_batch.h. Some of the terms are
1179 : based on the worst-case number of FEC sets in a block, which
1180 : scales with the slot time reductions:
1181 : - pad_ohead = per-batch padding (OHEAD_PAD in fd_shred_batch.h)
1182 : - hdr_ohead = per-batch header (OHEAD_HDR in fd_shred_batch.h)
1183 : - reg_ohead = only used for the last batch
1184 : (OHEAD_REG in fd_shred_batch.h) */
1185 0 : ulong fec_set_cnt = ctx->_became_leader->limits.slot_max_data_shreds/32UL;
1186 0 : ulong pad_ohead = ( fec_set_cnt/2UL )*FD_SHREDDER_CHAINED_FEC_SET_PAYLOAD_SZ;
1187 0 : ulong hdr_ohead = fec_set_cnt*8UL;
1188 0 : ulong reg_ohead = 8192UL;
1189 0 : FD_TEST( fec_set_cnt >= 2UL ); /* guard against underflow */
1190 0 : ulong fec_data = ( fec_set_cnt - 2UL )*FD_SHREDDER_CHAINED_FEC_SET_PAYLOAD_SZ;
1191 0 : FD_TEST( fec_data >= pad_ohead+hdr_ohead+reg_ohead );
1192 0 : base_max_data = fec_data - pad_ohead - hdr_ohead - reg_ohead;
1193 0 : }
1194 : /* Reserve some space in the block for ticks */
1195 0 : ctx->slot_max_data = base_max_data
1196 0 : - 48UL*(ctx->_became_leader->ticks_per_slot+ctx->_became_leader->total_skipped_ticks);
1197 :
1198 0 : ctx->limits.slot_max_cost = ctx->_became_leader->limits.slot_max_cost;
1199 0 : ctx->limits.slot_max_vote_cost = ctx->_became_leader->limits.slot_max_vote_cost;
1200 0 : ctx->limits.slot_max_write_cost_per_acct = ctx->_became_leader->limits.slot_max_write_cost_per_acct;
1201 0 : ctx->limits.slot_max_allocated_data_per_block = ctx->_became_leader->limits.slot_max_allocated_data_per_block;
1202 0 : ctx->limits.slot_max_data_shreds = ctx->_became_leader->limits.slot_max_data_shreds;
1203 :
1204 : /* ticks_per_ns is probably relatively stable over 400ms, but not
1205 : over several hours, so we need to compute the slot duration in
1206 : milliseconds first and then convert to ticks. This doesn't need
1207 : to be super accurate, but we don't want it to vary wildly. */
1208 0 : long end_ticks = now_ticks + (long)((double)fd_long_max( ctx->_became_leader->slot_end_ns - now_ns, 1L )*ctx->ticks_per_ns);
1209 : /* We may still get overrun, but then we'll never use this and just
1210 : reinitialize it the next time when we actually become leader. */
1211 0 : fd_pack_pacing_init( ctx->pacer, now_ticks, end_ticks, (float)ctx->ticks_per_ns, ctx->limits.slot_max_cost );
1212 :
1213 0 : if( FD_UNLIKELY( ctx->crank->enabled ) ) {
1214 : /* If we get overrun, we'll just never use these values, but the
1215 : old values aren't really useful either. */
1216 0 : ctx->crank->epoch = ctx->_became_leader->epoch;
1217 0 : *(ctx->crank->prev_config) = *(ctx->_became_leader->bundle->config);
1218 0 : memcpy( ctx->crank->recent_blockhash, ctx->_became_leader->bundle->last_blockhash, 32UL );
1219 0 : memcpy( ctx->crank->tip_receiver_owner, ctx->_became_leader->bundle->tip_receiver_owner, 32UL );
1220 0 : }
1221 :
1222 0 : FD_LOG_INFO(( "pack_became_leader(slot=%lu,ends_at=%ld)", ctx->leader_slot, ctx->_became_leader->slot_end_ns ));
1223 :
1224 0 : update_metric_state( ctx, fd_tickcount(), FD_PACK_METRIC_STATE_LEADER, 1 );
1225 :
1226 0 : ctx->slot_end_ns = ctx->_became_leader->slot_end_ns;
1227 0 : ctx->slot_dynamic_max_microblocks = ctx->slot_max_microblocks;
1228 0 : ctx->pending_reduce_mb_bound = 0;
1229 0 : fd_pack_limits_t limits[ 1 ];
1230 0 : limits->max_cost_per_block = ctx->limits.slot_max_cost;
1231 0 : limits->max_data_bytes_per_block = ctx->slot_max_data;
1232 0 : limits->max_microblocks_per_block = ctx->slot_max_microblocks;
1233 0 : limits->max_vote_cost_per_block = ctx->limits.slot_max_vote_cost;
1234 0 : limits->max_write_cost_per_acct = ctx->limits.slot_max_write_cost_per_acct;
1235 0 : limits->max_txn_per_microblock = ULONG_MAX; /* unused */
1236 0 : limits->max_allocated_data_per_block = ctx->limits.slot_max_allocated_data_per_block;
1237 0 : fd_pack_set_block_limits( ctx->pack, limits );
1238 0 : fd_pack_pacing_update_consumed_cus( ctx->pacer, fd_pack_current_block_cost( ctx->pack ), now );
1239 :
1240 0 : break;
1241 0 : }
1242 0 : case IN_KIND_EXECLE: {
1243 : /* For a previous slot */
1244 0 : if( FD_UNLIKELY( sig!=ctx->leader_slot ) ) return;
1245 :
1246 0 : fd_pack_rebate_cus( ctx->pack, ctx->rebate->rebate );
1247 0 : ctx->pending_rebate_sz = 0UL;
1248 0 : fd_pack_pacing_update_consumed_cus( ctx->pacer, fd_pack_current_block_cost( ctx->pack ), now );
1249 0 : break;
1250 0 : }
1251 0 : case IN_KIND_RESOLV: {
1252 : /* Normal transaction case */
1253 : #if FD_PACK_USE_EXTRA_STORAGE
1254 : if( FD_LIKELY( !ctx->insert_to_extra ) ) {
1255 : #else
1256 0 : if( 1 ) {
1257 0 : #endif
1258 0 : if( FD_UNLIKELY( ctx->is_bundle ) ) {
1259 0 : if( FD_UNLIKELY( ctx->current_bundle->txn_cnt==0UL ) ) return;
1260 0 : if( FD_UNLIKELY( ++(ctx->current_bundle->txn_received)==ctx->current_bundle->txn_cnt ) ) {
1261 0 : ulong deleted;
1262 0 : long insert_duration = -fd_tickcount();
1263 0 : int result = fd_pack_insert_bundle_fini( ctx->pack, ctx->current_bundle->bundle, ctx->current_bundle->txn_cnt, ctx->current_bundle->min_blockhash_slot, 0, ctx->blk_engine_cfg, &deleted );
1264 0 : insert_duration += fd_tickcount();
1265 0 : FD_MCNT_INC( PACK, TXN_DELETED, deleted );
1266 0 : ctx->insert_result[ result + FD_PACK_INSERT_RETVAL_OFF ] += ctx->current_bundle->txn_received;
1267 0 : fd_histf_sample( ctx->insert_duration, (ulong)insert_duration );
1268 0 : ctx->current_bundle->bundle = NULL;
1269 0 : }
1270 0 : } else {
1271 0 : ulong blockhash_slot = sig;
1272 0 : ulong deleted;
1273 0 : long insert_duration = -fd_tickcount();
1274 0 : int result = fd_pack_insert_txn_fini( ctx->pack, ctx->cur_spot, blockhash_slot, &deleted );
1275 0 : insert_duration += fd_tickcount();
1276 0 : FD_MCNT_INC( PACK, TXN_DELETED, deleted );
1277 0 : ctx->insert_result[ result + FD_PACK_INSERT_RETVAL_OFF ]++;
1278 0 : fd_histf_sample( ctx->insert_duration, (ulong)insert_duration );
1279 0 : if( FD_LIKELY( result>=0 ) ) ctx->last_successful_insert = now;
1280 0 : }
1281 0 : }
1282 :
1283 0 : ctx->cur_spot = NULL;
1284 0 : break;
1285 0 : }
1286 0 : case IN_KIND_EXECUTED_TXN: {
1287 0 : ulong deleted = fd_pack_delete_transaction( ctx->pack, fd_type_pun( ctx->executed_txn_sig ) );
1288 0 : FD_MCNT_INC( PACK, TXN_ALREADY_EXECUTED, deleted );
1289 0 : break;
1290 0 : }
1291 0 : }
1292 :
1293 0 : update_metric_state( ctx, now, FD_PACK_METRIC_STATE_TRANSACTIONS, fd_pack_avail_txn_cnt( ctx->pack )>0 );
1294 0 : }
1295 :
1296 : static void
1297 : privileged_init( fd_topo_t const * topo,
1298 0 : fd_topo_tile_t const * tile ) {
1299 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1300 :
1301 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1302 0 : fd_pack_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_pack_ctx_t ), sizeof( fd_pack_ctx_t ) );
1303 :
1304 0 : if( FD_UNLIKELY( !fd_rng_secure( &ctx->rng_seed, sizeof(uint) ) ) ) {
1305 0 : FD_LOG_CRIT(( "fd_rng_secure failed" ));
1306 0 : }
1307 0 : if( FD_UNLIKELY( !fd_rng_secure( &ctx->rng_idx, sizeof(ulong) ) ) ) {
1308 0 : FD_LOG_CRIT(( "fd_rng_secure failed" ));
1309 0 : }
1310 :
1311 0 : if( FD_LIKELY( !tile->pack.bundle.enabled ) ) return;
1312 0 : if( FD_UNLIKELY( !tile->pack.bundle.vote_account_path[0] ) ) {
1313 0 : FD_LOG_WARNING(( "Disabling bundle crank because no vote account was specified" ));
1314 0 : return;
1315 0 : }
1316 :
1317 0 : if( FD_UNLIKELY( !strcmp( tile->pack.bundle.identity_key_path, "" ) ) )
1318 0 : FD_LOG_ERR(( "identity_key_path not set" ));
1319 :
1320 0 : const uchar * identity_key = fd_keyload_load( tile->pack.bundle.identity_key_path, /* pubkey only: */ 1 );
1321 0 : fd_memcpy( ctx->crank->identity_pubkey->b, identity_key, 32UL );
1322 :
1323 0 : if( FD_UNLIKELY( !fd_base58_decode_32( tile->pack.bundle.vote_account_path, ctx->crank->vote_pubkey->b ) ) ) {
1324 0 : const uchar * vote_key = fd_keyload_load( tile->pack.bundle.vote_account_path, /* pubkey only: */ 1 );
1325 0 : fd_memcpy( ctx->crank->vote_pubkey->b, vote_key, 32UL );
1326 0 : }
1327 0 : }
1328 :
1329 : static void
1330 : unprivileged_init( fd_topo_t const * topo,
1331 0 : fd_topo_tile_t const * tile ) {
1332 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
1333 :
1334 0 : if( FD_UNLIKELY( tile->pack.max_pending_transactions >= USHORT_MAX-10UL ) ) FD_LOG_ERR(( "pack tile supports up to %lu pending transactions", USHORT_MAX-11UL ));
1335 :
1336 0 : fd_pack_limits_t limits_upper[1] = {{
1337 0 : .max_cost_per_block = tile->pack.larger_max_cost_per_block ? LARGER_MAX_COST_PER_BLOCK : FD_PACK_MAX_COST_PER_BLOCK_UPPER_BOUND,
1338 0 : .max_vote_cost_per_block = FD_PACK_MAX_VOTE_COST_PER_BLOCK_UPPER_BOUND,
1339 0 : .max_write_cost_per_acct = FD_PACK_MAX_WRITE_COST_PER_ACCT_UPPER_BOUND,
1340 0 : .max_data_bytes_per_block = tile->pack.larger_shred_limits_per_block ? LARGER_MAX_DATA_PER_BLOCK : FD_PACK_MAX_DATA_PER_BLOCK,
1341 0 : .max_txn_per_microblock = EFFECTIVE_TXN_PER_MICROBLOCK,
1342 0 : .max_microblocks_per_block = (ulong)UINT_MAX, /* Limit not known yet */
1343 0 : .max_allocated_data_per_block = FD_PACK_MAX_ALLOCATED_DATA_PER_BLOCK,
1344 0 : }};
1345 :
1346 0 : ulong pack_footprint = fd_pack_footprint( tile->pack.max_pending_transactions, BUNDLE_META_SZ, tile->pack.execle_tile_count, limits_upper );
1347 :
1348 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
1349 0 : fd_pack_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_pack_ctx_t ), sizeof( fd_pack_ctx_t ) );
1350 0 : fd_rng_t * rng = fd_rng_join( fd_rng_new( FD_SCRATCH_ALLOC_APPEND( l, fd_rng_align(), fd_rng_footprint() ), ctx->rng_seed, ctx->rng_idx ) );
1351 0 : if( FD_UNLIKELY( !rng ) ) FD_LOG_ERR(( "fd_rng_new failed" ));
1352 :
1353 0 : fd_pack_limits_t limits_lower[1] = {{
1354 0 : .max_cost_per_block = tile->pack.larger_max_cost_per_block ? LARGER_MAX_COST_PER_BLOCK : FD_PACK_MAX_COST_PER_BLOCK_LOWER_BOUND,
1355 0 : .max_vote_cost_per_block = FD_PACK_MAX_VOTE_COST_PER_BLOCK_LOWER_BOUND,
1356 0 : .max_write_cost_per_acct = FD_PACK_MAX_WRITE_COST_PER_ACCT_LOWER_BOUND,
1357 0 : .max_data_bytes_per_block = tile->pack.larger_shred_limits_per_block ? LARGER_MAX_DATA_PER_BLOCK : FD_PACK_MAX_DATA_PER_BLOCK,
1358 0 : .max_txn_per_microblock = EFFECTIVE_TXN_PER_MICROBLOCK,
1359 0 : .max_microblocks_per_block = (ulong)UINT_MAX, /* Limit not known yet */
1360 0 : .max_allocated_data_per_block = FD_PACK_MAX_ALLOCATED_DATA_PER_BLOCK,
1361 0 : }};
1362 :
1363 0 : ctx->pack = fd_pack_join( fd_pack_new( FD_SCRATCH_ALLOC_APPEND( l, fd_pack_align(), pack_footprint ),
1364 0 : tile->pack.max_pending_transactions, BUNDLE_META_SZ, tile->pack.execle_tile_count,
1365 0 : limits_upper,
1366 0 : fd_type_pun_const( tile->pack.acct_blocklist ), tile->pack.acct_blocklist_cnt,
1367 0 : rng ) );
1368 0 : if( FD_UNLIKELY( !ctx->pack ) ) FD_LOG_ERR(( "fd_pack_new failed" ));
1369 :
1370 0 : fd_pack_set_block_limits( ctx->pack, limits_lower );
1371 :
1372 0 : if( FD_UNLIKELY( tile->in_cnt>32UL ) ) FD_LOG_ERR(( "Too many input links (%lu>32) to pack tile", tile->in_cnt ));
1373 :
1374 0 : FD_TEST( tile->in_cnt<sizeof( ctx->in_kind )/sizeof( ctx->in_kind[0] ) );
1375 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
1376 0 : fd_topo_link_t const * link = &topo->links[ tile->in_link_id[ i ] ];
1377 :
1378 0 : if( FD_LIKELY( !strcmp( link->name, "resolv_pack" ) ) ) ctx->in_kind[ i ] = IN_KIND_RESOLV;
1379 0 : else if( FD_LIKELY( !strcmp( link->name, "resolh_pack" ) ) ) ctx->in_kind[ i ] = IN_KIND_RESOLV;
1380 0 : else if( FD_LIKELY( !strcmp( link->name, "poh_pack" ) ) ) ctx->in_kind[ i ] = IN_KIND_POH;
1381 0 : else if( FD_LIKELY( !strcmp( link->name, "pohh_pack" ) ) ) ctx->in_kind[ i ] = IN_KIND_POH;
1382 0 : else if( FD_LIKELY( !strcmp( link->name, "bank_pack" ) ) ) ctx->in_kind[ i ] = IN_KIND_EXECLE;
1383 0 : else if( FD_LIKELY( !strcmp( link->name, "execle_pack" ) ) ) ctx->in_kind[ i ] = IN_KIND_EXECLE;
1384 0 : else if( FD_LIKELY( !strcmp( link->name, "sign_pack" ) ) ) ctx->in_kind[ i ] = IN_KIND_SIGN;
1385 0 : else if( FD_LIKELY( !strcmp( link->name, "replay_out" ) ) ) ctx->in_kind[ i ] = IN_KIND_REPLAY;
1386 0 : else if( FD_LIKELY( !strcmp( link->name, "executed_txn" ) ) ) ctx->in_kind[ i ] = IN_KIND_EXECUTED_TXN;
1387 0 : else FD_LOG_ERR(( "pack tile has unexpected input link %lu %s", i, link->name ));
1388 0 : }
1389 :
1390 0 : ulong execle_cnt = 0UL;
1391 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
1392 0 : fd_topo_tile_t const * consumer_tile = &topo->tiles[ i ];
1393 0 : if( FD_UNLIKELY( strcmp( consumer_tile->name, "execle" ) && strcmp( consumer_tile->name, "replay" ) ) ) continue;
1394 0 : for( ulong j=0UL; j<consumer_tile->in_cnt; j++ ) {
1395 0 : if( FD_UNLIKELY( consumer_tile->in_link_id[ j ]==tile->out_link_id[ 0 ] ) ) execle_cnt++;
1396 0 : }
1397 0 : }
1398 :
1399 : // if( FD_UNLIKELY( !execle_cnt ) ) FD_LOG_ERR(( "pack tile connects to no execle tiles" ));
1400 0 : if( FD_UNLIKELY( execle_cnt>FD_PACK_MAX_EXECLE_TILES ) ) FD_LOG_ERR(( "pack tile connects to too many execle tiles" ));
1401 : // if( FD_UNLIKELY( execle_cnt!=tile->pack.execle_tile_count ) ) FD_LOG_ERR(( "pack tile connects to %lu execle tiles, but tile->pack.execle_tile_count is %lu", execle_cnt, tile->pack.execle_tile_count ));
1402 :
1403 0 : FD_TEST( (tile->pack.schedule_strategy>=0) & (tile->pack.schedule_strategy<=FD_PACK_STRATEGY_BALANCED) );
1404 :
1405 0 : ctx->crank->enabled = tile->pack.bundle.enabled;
1406 0 : if( FD_UNLIKELY( tile->pack.bundle.enabled ) ) {
1407 0 : if( FD_UNLIKELY( !fd_bundle_crank_gen_init( ctx->crank->gen, (fd_acct_addr_t const *)tile->pack.bundle.tip_distribution_program_addr,
1408 0 : (fd_acct_addr_t const *)tile->pack.bundle.tip_payment_program_addr,
1409 0 : (fd_acct_addr_t const *)ctx->crank->vote_pubkey->b,
1410 0 : (fd_acct_addr_t const *)tile->pack.bundle.tip_distribution_authority,
1411 0 : schedule_strategy_strings[ tile->pack.schedule_strategy ],
1412 0 : tile->pack.bundle.commission_bps ) ) ) {
1413 0 : FD_LOG_ERR(( "constructing bundle generator failed" ));
1414 0 : }
1415 :
1416 0 : ulong sign_in_idx = fd_topo_find_tile_in_link ( topo, tile, "sign_pack", tile->kind_id );
1417 0 : ulong sign_out_idx = fd_topo_find_tile_out_link( topo, tile, "pack_sign", tile->kind_id );
1418 0 : FD_TEST( sign_in_idx!=ULONG_MAX );
1419 0 : fd_topo_link_t const * sign_in = &topo->links[ tile->in_link_id[ sign_in_idx ] ];
1420 0 : fd_topo_link_t const * sign_out = &topo->links[ tile->out_link_id[ sign_out_idx ] ];
1421 0 : if( FD_UNLIKELY( !fd_keyguard_client_join( fd_keyguard_client_new( ctx->crank->keyguard_client,
1422 0 : sign_out->mcache,
1423 0 : sign_out->dcache,
1424 0 : sign_in->mcache,
1425 0 : sign_in->dcache,
1426 0 : sign_out->mtu ) ) ) ) {
1427 0 : FD_LOG_ERR(( "failed to construct keyguard" ));
1428 0 : }
1429 : /* Initialize enough of the prev config that it produces a
1430 : transaction */
1431 0 : ctx->crank->prev_config->discriminator = 0x82ccfa1ee0aa0c9bUL;
1432 0 : ctx->crank->prev_config->tip_receiver->b[1] = 1;
1433 0 : ctx->crank->prev_config->block_builder->b[2] = 1;
1434 :
1435 0 : memset( ctx->crank->tip_receiver_owner, '\0', 32UL );
1436 0 : memset( ctx->crank->recent_blockhash, '\0', 32UL );
1437 0 : memset( ctx->crank->last_sig, '\0', 64UL );
1438 0 : ctx->crank->ib_inserted = 0;
1439 0 : ctx->crank->epoch = 0UL;
1440 0 : ctx->crank->keyswitch = fd_keyswitch_join( fd_topo_obj_laddr( topo, tile->id_keyswitch_obj_id ) );
1441 0 : FD_TEST( ctx->crank->keyswitch );
1442 0 : } else {
1443 0 : memset( ctx->crank, '\0', sizeof(ctx->crank) );
1444 0 : }
1445 :
1446 :
1447 : #if FD_PACK_USE_EXTRA_STORAGE
1448 : ctx->extra_txn_deq = extra_txn_deq_join( extra_txn_deq_new( FD_SCRATCH_ALLOC_APPEND( l, extra_txn_deq_align(),
1449 : extra_txn_deq_footprint() ) ) );
1450 : #endif
1451 :
1452 0 : ctx->cur_spot = NULL;
1453 0 : ctx->is_bundle = 0;
1454 0 : ctx->strategy = tile->pack.schedule_strategy;
1455 0 : ctx->max_pending_transactions = tile->pack.max_pending_transactions;
1456 0 : ctx->leader_slot = ULONG_MAX;
1457 0 : ctx->leader_bank = NULL;
1458 0 : ctx->leader_bank_idx = ULONG_MAX;
1459 0 : ctx->leader_bank_seq = ULONG_MAX;
1460 0 : ctx->pack_idx = 0UL;
1461 0 : ctx->slot_microblock_cnt = 0UL;
1462 0 : ctx->pack_txn_cnt = 0UL;
1463 0 : ctx->slot_max_microblocks = 0UL;
1464 0 : ctx->slot_dynamic_max_microblocks = 0UL;
1465 0 : ctx->pending_reduce_mb_bound = 0;
1466 0 : ctx->slot_max_data = 0UL;
1467 0 : ctx->larger_shred_limits_per_block = tile->pack.larger_shred_limits_per_block;
1468 0 : ctx->drain_execle = 0;
1469 0 : ctx->approx_wallclock_ns = fd_log_wallclock();
1470 0 : ctx->approx_tickcount = fd_tickcount();
1471 0 : ctx->rng = rng;
1472 0 : ctx->ticks_per_ns = fd_tempo_tick_per_ns( NULL );
1473 0 : ctx->last_successful_insert = 0L;
1474 0 : ctx->highest_observed_slot = 0UL;
1475 0 : ctx->microblock_duration_ticks = (ulong)(fd_tempo_tick_per_ns( NULL )*(double)MICROBLOCK_DURATION_NS + 0.5);
1476 : #if FD_PACK_USE_EXTRA_STORAGE
1477 : ctx->insert_to_extra = 0;
1478 : #endif
1479 0 : ctx->use_consumed_cus = tile->pack.use_consumed_cus;
1480 0 : ctx->crank->enabled = tile->pack.bundle.enabled;
1481 :
1482 : #if !SMALL_MICROBLOCKS
1483 : ctx->wait_duration_ticks[ 0 ] = ULONG_MAX;
1484 : for( ulong i=1UL; i<MAX_TXN_PER_MICROBLOCK+1UL; i++ ) {
1485 : ctx->wait_duration_ticks[ i ]=(ulong)(fd_tempo_tick_per_ns( NULL )*(double)wait_duration[ i ] + 0.5);
1486 : }
1487 : #endif
1488 :
1489 0 : ctx->limits.slot_max_cost = limits_lower->max_cost_per_block;
1490 0 : ctx->limits.slot_max_vote_cost = limits_lower->max_vote_cost_per_block;
1491 0 : ctx->limits.slot_max_write_cost_per_acct = limits_lower->max_write_cost_per_acct;
1492 :
1493 0 : ctx->execle_cnt = tile->pack.execle_tile_count;
1494 0 : ctx->poll_cursor = 0;
1495 0 : ctx->skip_cnt = 0L;
1496 0 : ctx->execle_idle_bitset = fd_ulong_mask_lsb( (int)tile->pack.execle_tile_count );
1497 0 : for( ulong i=0UL; i<tile->pack.execle_tile_count; i++ ) {
1498 0 : ulong busy_obj_id = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "execle_busy.%lu", i );
1499 0 : FD_TEST( busy_obj_id!=ULONG_MAX );
1500 0 : ctx->execle_current[ i ] = fd_fseq_join( fd_topo_obj_laddr( topo, busy_obj_id ) );
1501 0 : ctx->execle_expect[ i ] = ULONG_MAX;
1502 0 : if( FD_UNLIKELY( !ctx->execle_current[ i ] ) ) FD_LOG_ERR(( "execle tile %lu has no busy flag", i ));
1503 0 : ctx->execle_ready_at[ i ] = 0L;
1504 0 : FD_TEST( ULONG_MAX==fd_fseq_query( ctx->execle_current[ i ] ) );
1505 0 : }
1506 :
1507 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
1508 0 : fd_topo_link_t const * link = &topo->links[ tile->in_link_id[ i ] ];
1509 0 : fd_topo_wksp_t const * link_wksp = &topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ];
1510 :
1511 0 : ctx->in[ i ].mem = link_wksp->wksp;
1512 0 : ctx->in[ i ].chunk0 = fd_dcache_compact_chunk0( ctx->in[ i ].mem, link->dcache );
1513 0 : ctx->in[ i ].wmark = fd_dcache_compact_wmark ( ctx->in[ i ].mem, link->dcache, link->mtu );
1514 0 : }
1515 :
1516 0 : ctx->execle_out_mem = topo->workspaces[ topo->objs[ topo->links[ tile->out_link_id[ 0 ] ].dcache_obj_id ].wksp_id ].wksp;
1517 0 : ctx->execle_out_chunk0 = fd_dcache_compact_chunk0( ctx->execle_out_mem, topo->links[ tile->out_link_id[ 0 ] ].dcache );
1518 0 : ctx->execle_out_wmark = fd_dcache_compact_wmark ( ctx->execle_out_mem, topo->links[ tile->out_link_id[ 0 ] ].dcache, topo->links[ tile->out_link_id[ 0 ] ].mtu );
1519 0 : ctx->execle_out_chunk = ctx->execle_out_chunk0;
1520 :
1521 0 : ctx->poh_out_mem = topo->workspaces[ topo->objs[ topo->links[ tile->out_link_id[ 1 ] ].dcache_obj_id ].wksp_id ].wksp;
1522 0 : ctx->poh_out_chunk0 = fd_dcache_compact_chunk0( ctx->poh_out_mem, topo->links[ tile->out_link_id[ 1 ] ].dcache );
1523 0 : ctx->poh_out_wmark = fd_dcache_compact_wmark ( ctx->poh_out_mem, topo->links[ tile->out_link_id[ 1 ] ].dcache, topo->links[ tile->out_link_id[ 1 ] ].mtu );
1524 0 : ctx->poh_out_chunk = ctx->poh_out_chunk0;
1525 :
1526 : /* Initialize metrics storage */
1527 0 : memset( ctx->insert_result, '\0', FD_PACK_INSERT_RETVAL_CNT * sizeof(ulong) );
1528 0 : fd_histf_join( fd_histf_new( ctx->schedule_duration, FD_MHIST_SECONDS_MIN( PACK, SCHEDULE_MICROBLOCK_DURATION_SECONDS ),
1529 0 : FD_MHIST_SECONDS_MAX( PACK, SCHEDULE_MICROBLOCK_DURATION_SECONDS ) ) );
1530 0 : fd_histf_join( fd_histf_new( ctx->no_sched_duration, FD_MHIST_SECONDS_MIN( PACK, NO_SCHEDULE_MICROBLOCK_DURATION_SECONDS ),
1531 0 : FD_MHIST_SECONDS_MAX( PACK, NO_SCHEDULE_MICROBLOCK_DURATION_SECONDS ) ) );
1532 0 : fd_histf_join( fd_histf_new( ctx->insert_duration, FD_MHIST_SECONDS_MIN( PACK, INSERT_TRANSACTION_DURATION_SECONDS ),
1533 0 : FD_MHIST_SECONDS_MAX( PACK, INSERT_TRANSACTION_DURATION_SECONDS ) ) );
1534 0 : fd_histf_join( fd_histf_new( ctx->complete_duration, FD_MHIST_SECONDS_MIN( PACK, COMPLETE_MICROBLOCK_DURATION_SECONDS ),
1535 0 : FD_MHIST_SECONDS_MAX( PACK, COMPLETE_MICROBLOCK_DURATION_SECONDS ) ) );
1536 0 : ctx->metric_state = 0;
1537 0 : ctx->metric_state_begin = fd_tickcount();
1538 0 : memset( ctx->metric_timing, '\0', 16*sizeof(long) );
1539 0 : memset( ctx->current_bundle, '\0', sizeof(ctx->current_bundle) );
1540 0 : memset( ctx->blk_engine_cfg, '\0', sizeof(ctx->blk_engine_cfg) );
1541 0 : memset( ctx->last_sched_metrics, '\0', sizeof(ctx->last_sched_metrics) );
1542 0 : memset( ctx->start_block_sched_metrics, '\0', sizeof(ctx->start_block_sched_metrics) );
1543 0 : memset( ctx->crank->metrics, '\0', sizeof(ctx->crank->metrics) );
1544 :
1545 0 : FD_LOG_INFO(( "packing microblocks of at most %lu transactions to %lu execle tiles using strategy %i", EFFECTIVE_TXN_PER_MICROBLOCK, tile->pack.execle_tile_count, ctx->strategy ));
1546 :
1547 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, scratch_align() );
1548 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
1549 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
1550 0 : }
1551 :
1552 : static ulong
1553 : populate_allowed_seccomp( fd_topo_t const * topo,
1554 : fd_topo_tile_t const * tile,
1555 : ulong out_cnt,
1556 0 : struct sock_filter * out ) {
1557 0 : (void)topo;
1558 0 : (void)tile;
1559 :
1560 0 : populate_sock_filter_policy_fd_pack_tile( out_cnt, out, (uint)fd_log_private_logfile_fd() );
1561 0 : return sock_filter_policy_fd_pack_tile_instr_cnt;
1562 0 : }
1563 :
1564 : static ulong
1565 : populate_allowed_fds( fd_topo_t const * topo,
1566 : fd_topo_tile_t const * tile,
1567 : ulong out_fds_cnt,
1568 0 : int * out_fds ) {
1569 0 : (void)topo;
1570 0 : (void)tile;
1571 :
1572 0 : if( FD_UNLIKELY( out_fds_cnt<2UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
1573 :
1574 0 : ulong out_cnt = 0UL;
1575 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
1576 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
1577 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
1578 0 : return out_cnt;
1579 0 : }
1580 :
1581 : /* Pack can publish a frag in the following scenarios:
1582 :
1583 : after_credit:
1584 : A. TIMED_OUT. Sets ctx->leader_slot=ULONG_MAX. return.
1585 : B. DONE_DRAINING. *after* DONE_PACKING. return.
1586 : C. REDUCE_MB_BOUND. return.
1587 : D. SCHEDULE_MB. *doesn't* return.
1588 : E. EXHAUST_MICROBLOCKS. Sets ctx->leader_slot=ULONG_MAX. return.
1589 : after_frag:
1590 : F. LEADER_SWITCH. Requires ctx->leader_slot!=ULONG_MAX
1591 :
1592 : It isn't possible to get a burst of 3, but a burst of 2 is possible
1593 : in these situations.
1594 :
1595 : C -> F
1596 : D -> F
1597 : D -> E
1598 : */
1599 0 : #define STEM_BURST (2UL)
1600 :
1601 : /* We want lazy (measured in ns) to be small enough that the producer
1602 : and the consumer never have to wait for credits. For most tango
1603 : links, we use a default worst case speed coming from 100 Gbps
1604 : Ethernet. That's not very suitable for microblocks that go from
1605 : pack to bank. Instead we manually estimate the very aggressive
1606 : 1000ns per microblock, and then reduce it further (in line with the
1607 : default lazy value computation) to ensure the random value chosen
1608 : based on this won't lead to credit return stalls. */
1609 0 : #define STEM_LAZY (128L*3000L)
1610 :
1611 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_pack_ctx_t
1612 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_pack_ctx_t)
1613 :
1614 0 : #define STEM_CALLBACK_DURING_HOUSEKEEPING during_housekeeping
1615 0 : #define STEM_CALLBACK_BEFORE_CREDIT before_credit
1616 0 : #define STEM_CALLBACK_AFTER_CREDIT after_credit
1617 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
1618 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
1619 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
1620 :
1621 : #include "../stem/fd_stem.c"
1622 :
1623 : fd_topo_run_tile_t fd_tile_pack = {
1624 : .name = "pack",
1625 : .populate_allowed_seccomp = populate_allowed_seccomp,
1626 : .populate_allowed_fds = populate_allowed_fds,
1627 : .scratch_align = scratch_align,
1628 : .scratch_footprint = scratch_footprint,
1629 : .privileged_init = privileged_init,
1630 : .unprivileged_init = unprivileged_init,
1631 : .run = stem_run,
1632 : };
|
