Line data Source code
1 : #include "fd_sched.h"
2 : #include "../../disco/fd_disco_base.h" /* for FD_MAX_TXN_PER_SLOT */
3 : #include "../../flamenco/runtime/fd_runtime.h" /* for fd_runtime_load_txn_address_lookup_tables */
4 :
5 : #include "../../flamenco/runtime/sysvar/fd_sysvar_slot_hashes.h" /* for ALUTs */
6 :
7 :
8 0 : #define FD_SCHED_MAX_DEPTH (FD_RDISP_MAX_DEPTH>>2)
9 : #define FD_SCHED_MAX_STAGING_LANES_LOG (2)
10 : #define FD_SCHED_MAX_STAGING_LANES (1UL<<FD_SCHED_MAX_STAGING_LANES_LOG)
11 :
12 : /* 64 ticks per slot, and a single gigantic microblock containing min
13 : size transactions. */
14 : FD_STATIC_ASSERT( FD_MAX_TXN_PER_SLOT_SHRED==((FD_SHRED_DATA_PAYLOAD_MAX_PER_SLOT-65UL*sizeof(fd_microblock_hdr_t))/FD_TXN_MIN_SERIALIZED_SZ), max_txn_per_slot_shred );
15 :
16 : /* We size the buffer to be able to hold residual data from the previous
17 : FEC set that only becomes parseable after the next FEC set is
18 : ingested, as well as the incoming FEC set. The largest minimally
19 : parseable unit of data is a transaction. So that much data may
20 : straddle FEC set boundaries. Other minimally parseable units of data
21 : include the microblock header and the microblock count within a
22 : batch. */
23 0 : #define FD_SCHED_MAX_PAYLOAD_PER_FEC (FD_STORE_DATA_MAX)
24 : #define FD_SCHED_MAX_FEC_BUF_SZ (FD_SCHED_MAX_PAYLOAD_PER_FEC+FD_TXN_MTU)
25 : FD_STATIC_ASSERT( FD_TXN_MTU>=sizeof(fd_microblock_hdr_t), resize buffer for residual data );
26 : FD_STATIC_ASSERT( FD_TXN_MTU>=sizeof(ulong), resize buffer for residual data );
27 :
28 0 : #define FD_SCHED_MAGIC (0xace8a79c181f89b6UL) /* echo -n "fd_sched_v0" | sha512sum | head -c 16 */
29 :
30 0 : #define FD_SCHED_PARSER_OK (0)
31 0 : #define FD_SCHED_PARSER_AGAIN_LATER (1)
32 0 : #define FD_SCHED_PARSER_BAD_BLOCK (2)
33 :
34 :
35 : /* Structs. */
36 :
37 : #define SET_NAME txn_bitset
38 : #define SET_MAX FD_SCHED_MAX_DEPTH
39 : #include "../../util/tmpl/fd_set.c"
40 :
41 : struct fd_sched_block {
42 : ulong slot;
43 : ulong parent_slot;
44 : ulong parent_idx; /* Index of the parent in the pool. */
45 : ulong child_idx; /* Index of the left-child in the pool. */
46 : ulong sibling_idx; /* Index of the right-sibling in the pool. */
47 :
48 : /* Counters. */
49 : uint txn_parsed_cnt;
50 : /* txn_queued_cnt = txn_parsed_cnt-txn_in_flight_cnt-txn_done_cnt */
51 : uint txn_exec_in_flight_cnt;
52 : uint txn_exec_done_cnt;
53 : uint txn_sigverify_in_flight_cnt;
54 : uint txn_sigverify_done_cnt;
55 : uint txn_done_cnt; /* A transaction is considered done when all types of tasks associated with it are done. */
56 : ulong txn_idx[ FD_MAX_TXN_PER_SLOT ]; /* Indexed by parse order. */
57 : uint shred_cnt;
58 :
59 : /* Parser state. */
60 : uchar txn[ FD_TXN_MAX_SZ ] __attribute__((aligned(alignof(fd_txn_t))));
61 : fd_hash_t poh; /* Latest PoH hash we've seen from the ingested FEC sets. */
62 : ulong mblks_rem; /* Number of microblocks remaining in the current batch. */
63 : ulong txns_rem; /* Number of transactions remaining in the current microblock. */
64 : fd_acct_addr_t aluts[ 256 ]; /* Resolve ALUT accounts into this buffer for more parallelism. */
65 : uint fec_buf_sz; /* Size of the fec_buf in bytes. */
66 : uint fec_buf_soff; /* Starting offset into fec_buf for unparsed transactions. */
67 : uint fec_eob:1; /* FEC end-of-batch: set if the last FEC set in the batch is being
68 : ingested. */
69 : uint fec_sob:1; /* FEC start-of-batch: set if the parser expects to be receiving a new
70 : batch. */
71 :
72 : /* Block state. */
73 : uint fec_eos:1; /* FEC end-of-stream: set if the last FEC set in the block has been
74 : ingested. */
75 : uint rooted:1; /* Set if the block is rooted. */
76 : uint dying:1; /* Set if the block has been abandoned and no transactions should be
77 : scheduled from it. */
78 : uint in_sched:1; /* Set if the block is being tracked by the scheduler. */
79 : uint in_rdisp:1; /* Set if the block is being tracked by the dispatcher, either as staged
80 : or unstaged. */
81 : uint block_start_signaled:1; /* Set if the start-of-block sentinel has been dispatched. */
82 : uint block_end_signaled:1; /* Set if the end-of-block sentinel has been dispatched. */
83 : uint block_start_done:1; /* Set if the start-of-block processing has been completed. */
84 : uint block_end_done:1; /* Set if the end-of-block processing has been completed. */
85 : uint staged:1; /* Set if the block is in a dispatcher staging lane; a staged block is
86 : tracked by the dispatcher. */
87 : ulong staging_lane; /* Ignored if staged==0. */
88 : ulong luf_depth; /* Depth of longest unstaged fork starting from this node; only
89 : stageable unstaged descendants are counted. */
90 : uchar fec_buf[ FD_SCHED_MAX_FEC_BUF_SZ ]; /* The previous FEC set could have some residual data that only becomes
91 : parseable after the next FEC set is ingested. */
92 : };
93 : typedef struct fd_sched_block fd_sched_block_t;
94 :
95 : FD_STATIC_ASSERT( sizeof(fd_hash_t)==sizeof(((fd_microblock_hdr_t *)0)->hash), unexpected poh hash size );
96 :
97 :
98 : struct fd_sched_metrics {
99 : uint block_added_cnt;
100 : uint block_added_staged_cnt;
101 : uint block_added_unstaged_cnt;
102 : uint block_added_dead_ood_cnt;
103 : uint block_removed_cnt;
104 : uint block_abandoned_cnt;
105 : uint block_bad_cnt;
106 : uint block_promoted_cnt;
107 : uint block_demoted_cnt;
108 : uint deactivate_no_child_cnt;
109 : uint deactivate_no_txn_cnt;
110 : uint deactivate_pruned_cnt;
111 : uint deactivate_abandoned_cnt;
112 : uint lane_switch_cnt;
113 : uint lane_promoted_cnt;
114 : uint lane_demoted_cnt;
115 : uint alut_success_cnt;
116 : uint alut_serializing_cnt;
117 : uint txn_abandoned_parsed_cnt;
118 : uint txn_abandoned_exec_done_cnt;
119 : uint txn_abandoned_done_cnt;
120 : uint txn_max_in_flight_cnt;
121 : ulong txn_weighted_in_flight_cnt;
122 : ulong txn_weighted_in_flight_tickcount;
123 : ulong txn_none_in_flight_tickcount;
124 : ulong txn_parsed_cnt;
125 : ulong txn_exec_done_cnt;
126 : ulong txn_sigverify_done_cnt;
127 : ulong txn_done_cnt;
128 : ulong bytes_ingested_cnt;
129 : ulong bytes_ingested_unparsed_cnt;
130 : ulong bytes_dropped_cnt;
131 : ulong fec_cnt;
132 : };
133 : typedef struct fd_sched_metrics fd_sched_metrics_t;
134 :
135 : struct fd_sched {
136 : fd_sched_metrics_t metrics[ 1 ];
137 : ulong canary; /* == FD_SCHED_MAGIC */
138 : ulong block_cnt_max; /* Immutable. */
139 : long txn_in_flight_last_tick;
140 : ulong root_idx;
141 : fd_rdisp_t * rdisp;
142 : ulong active_bank_idx; /* Index of the actively replayed block, or ULONG_MAX if no block is
143 : actively replayed; has to have a transaction to dispatch; staged
144 : blocks that have no transactions to dispatch are not eligible for
145 : being active. */
146 : ulong staged_bitset; /* Bit i set if staging lane i is occupied. */
147 : ulong staged_head_bank_idx[ FD_SCHED_MAX_STAGING_LANES ]; /* Head of the linear chain in each staging lane, ignored if bit i is
148 : not set in the bitset. */
149 : ulong txn_pool_free_cnt;
150 : fd_txn_p_t txn_pool[ FD_SCHED_MAX_DEPTH ];
151 : ulong txn_to_bank_idx[ FD_SCHED_MAX_DEPTH ]; /* Index of the bank that the txn belongs to. */
152 : txn_bitset_t exec_done_set[ txn_bitset_word_cnt ]; /* Indexed by txn_idx. */
153 : txn_bitset_t sigverify_done_set[ txn_bitset_word_cnt ]; /* Indexed by txn_idx. */
154 : fd_sched_block_t * block_pool; /* Just a flat array. */
155 : };
156 : typedef struct fd_sched fd_sched_t;
157 :
158 :
159 : /* Internal helpers. */
160 :
161 : static void
162 : add_block( fd_sched_t * sched,
163 : ulong bank_idx,
164 : ulong parent_bank_idx );
165 :
166 : FD_WARN_UNUSED static int
167 : fd_sched_parse( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx );
168 :
169 : FD_WARN_UNUSED static int
170 : fd_sched_parse_txn( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx );
171 :
172 : static void
173 : try_activate_block( fd_sched_t * sched );
174 :
175 : static void
176 : check_or_set_active_block( fd_sched_t * sched );
177 :
178 : static void
179 : subtree_abandon( fd_sched_t * sched, fd_sched_block_t * block );
180 :
181 : static void
182 : maybe_switch_block( fd_sched_t * sched, ulong bank_idx );
183 :
184 : FD_FN_UNUSED static ulong
185 : find_and_stage_longest_unstaged_fork( fd_sched_t * sched, int lane_idx );
186 :
187 : static ulong
188 : compute_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx );
189 :
190 : static ulong
191 : stage_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx, int lane_idx );
192 :
193 : static inline fd_sched_block_t *
194 0 : block_pool_ele( fd_sched_t * sched, ulong idx ) {
195 0 : FD_TEST( idx<sched->block_cnt_max || idx==ULONG_MAX );
196 0 : return idx==ULONG_MAX ? NULL : sched->block_pool+idx;
197 0 : }
198 :
199 : FD_FN_UNUSED static inline int
200 0 : block_is_void( fd_sched_block_t * block ) {
201 0 : /* We've seen everything in the block and no transaction got parsed
202 0 : out. */
203 0 : return block->fec_eos && block->txn_parsed_cnt==0;
204 0 : }
205 :
206 : static inline int
207 0 : block_should_signal_end( fd_sched_block_t * block ) {
208 0 : return block->fec_eos && block->txn_parsed_cnt==block->txn_done_cnt && block->block_start_done && !block->block_end_signaled;
209 0 : }
210 :
211 : static inline int
212 0 : block_will_signal_end( fd_sched_block_t * block ) {
213 0 : return block->fec_eos && !block->block_end_signaled;
214 0 : }
215 :
216 : /* Is there something known to be dispatchable in the block? This is an
217 : important liveness property. A block that doesn't contain any known
218 : dispatchable tasks will be deactivated or demoted. */
219 : static inline int
220 0 : block_is_dispatchable( fd_sched_block_t * block ) {
221 0 : ulong exec_queued_cnt = block->txn_parsed_cnt-block->txn_exec_in_flight_cnt-block->txn_exec_done_cnt;
222 0 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
223 0 : return exec_queued_cnt>0UL ||
224 0 : sigverify_queued_cnt>0UL ||
225 0 : !block->block_start_signaled ||
226 0 : block_will_signal_end( block );
227 0 : }
228 :
229 : static inline int
230 0 : block_is_in_flight( fd_sched_block_t * block ) {
231 0 : return block->txn_exec_in_flight_cnt || block->txn_sigverify_in_flight_cnt || (block->block_end_signaled && !block->block_end_done);
232 0 : }
233 :
234 : static inline int
235 0 : block_is_done( fd_sched_block_t * block ) {
236 0 : return block->fec_eos && block->txn_parsed_cnt==block->txn_done_cnt && block->block_start_done && block->block_end_done;
237 0 : }
238 :
239 : static inline int
240 0 : block_is_stageable( fd_sched_block_t * block ) {
241 0 : int rv = !block_is_done( block ) && !block->dying;
242 0 : if( FD_UNLIKELY( rv && !block->in_rdisp ) ) {
243 : /* Invariant: stageable blocks may be currently staged or unstaged,
244 : but must be in the dispatcher either way. When a block
245 : transitions to DONE, it will be immediately removed from the
246 : dispatcher. When a block transitions to DYING, it will be
247 : eventually abandoned from the dispatcher. */
248 0 : FD_LOG_CRIT(( "invariant violation: stageable block->in_rdisp==0, txn_parsed_cnt %u, txn_done_cnt %u, fec_eos %u,, slot %lu, parent slot %lu",
249 0 : block->txn_parsed_cnt, block->txn_done_cnt, (uint)block->fec_eos, block->slot, block->parent_slot ));
250 0 : }
251 0 : return rv;
252 0 : }
253 :
254 : static inline int
255 0 : block_is_promotable( fd_sched_block_t * block ) {
256 0 : return block_is_stageable( block ) && block_is_dispatchable( block ) && !block->staged;
257 0 : }
258 :
259 : static inline int
260 0 : block_is_activatable( fd_sched_block_t * block ) {
261 0 : return block_is_stageable( block ) && block_is_dispatchable( block ) && block->staged;
262 0 : }
263 :
264 : static inline int
265 0 : block_should_deactivate( fd_sched_block_t * block ) {
266 : /* We allow a grace period, during which a block has nothing to
267 : dispatch, but has something in-flight. The block is allowed to
268 : stay activated and ingest FEC sets during this time. The block
269 : will be deactivated if there's still nothing to dispatch by the
270 : time all in-flight tasks are completed. */
271 0 : return !block_is_activatable( block ) && !block_is_in_flight( block );
272 0 : }
273 :
274 : FD_FN_UNUSED static void
275 0 : print_block( fd_sched_block_t * block ) {
276 0 : FD_LOG_NOTICE(( "block slot %lu, parent_slot %lu, staged %d (lane %lu), dying %d, in_rdisp %d, fec_eos %d, rooted %d, block_start_signaled %d, block_end_signaled %d, block_start_done %d, block_end_done %d, txn_parsed_cnt %u, txn_exec_in_flight_cnt %u, txn_exec_done_cnt %u, txn_sigverify_in_flight_cnt %u, txn_sigverify_done_cnt %u, txn_done_cnt %u, shred_cnt %u",
277 0 : block->slot, block->parent_slot, block->staged, block->staging_lane, block->dying, block->in_rdisp, block->fec_eos, block->rooted, block->block_start_signaled, block->block_end_signaled, block->block_start_done, block->block_end_done, block->txn_parsed_cnt, block->txn_exec_in_flight_cnt, block->txn_exec_done_cnt, block->txn_sigverify_in_flight_cnt, block->txn_sigverify_done_cnt, block->txn_done_cnt, block->shred_cnt ));
278 0 : }
279 :
280 : FD_FN_UNUSED static void
281 0 : print_block_and_parent( fd_sched_t * sched, fd_sched_block_t * block ) {
282 0 : print_block( block );
283 0 : fd_sched_block_t * parent = block_pool_ele( sched, block->parent_idx );
284 0 : if( FD_LIKELY( parent ) ) print_block( parent );
285 0 : }
286 :
287 : FD_FN_UNUSED static void
288 0 : print_metrics( fd_sched_t * sched ) {
289 0 : FD_LOG_NOTICE(( "metrics: block_added_cnt %u, block_added_staged_cnt %u, block_added_unstaged_cnt %u, block_added_dead_ood_cnt %u, block_removed_cnt %u, block_abandoned_cnt %u, block_promoted_cnt %u, block_demoted_cnt %u, deactivate_no_child_cnt %u, deactivate_no_txn_cnt %u, deactivate_pruned_cnt %u, deactivate_abandoned_cnt %u, lane_switch_cnt %u, lane_promoted_cnt %u, lane_demoted_cnt %u, alut_success_cnt %u, alut_serializing_cnt %u, txn_abandoned_parsed_cnt %u, txn_abandoned_done_cnt %u, txn_max_in_flight_cnt %u, txn_weighted_in_flight_cnt %lu, txn_weighted_in_flight_tickcount %lu, txn_none_in_flight_tickcount %lu, txn_parsed_cnt %lu, txn_exec_done_cnt %lu, txn_sigverify_done_cnt %lu, txn_done_cnt %lu, bytes_ingested_cnt %lu, bytes_ingested_unparsed_cnt %lu, bytes_dropped_cnt %lu, fec_cnt %lu",
290 0 : sched->metrics->block_added_cnt, sched->metrics->block_added_staged_cnt, sched->metrics->block_added_unstaged_cnt, sched->metrics->block_added_dead_ood_cnt, sched->metrics->block_removed_cnt, sched->metrics->block_abandoned_cnt, sched->metrics->block_promoted_cnt, sched->metrics->block_demoted_cnt, sched->metrics->deactivate_no_child_cnt, sched->metrics->deactivate_no_txn_cnt, sched->metrics->deactivate_pruned_cnt, sched->metrics->deactivate_abandoned_cnt, sched->metrics->lane_switch_cnt, sched->metrics->lane_promoted_cnt, sched->metrics->lane_demoted_cnt, sched->metrics->alut_success_cnt, sched->metrics->alut_serializing_cnt, sched->metrics->txn_abandoned_parsed_cnt, sched->metrics->txn_abandoned_done_cnt, sched->metrics->txn_max_in_flight_cnt, sched->metrics->txn_weighted_in_flight_cnt, sched->metrics->txn_weighted_in_flight_tickcount, sched->metrics->txn_none_in_flight_tickcount, sched->metrics->txn_parsed_cnt, sched->metrics->txn_exec_done_cnt, sched->metrics->txn_sigverify_done_cnt, sched->metrics->txn_done_cnt, sched->metrics->bytes_ingested_cnt, sched->metrics->bytes_ingested_unparsed_cnt, sched->metrics->bytes_dropped_cnt, sched->metrics->fec_cnt ));
291 0 : }
292 :
293 : FD_FN_UNUSED static void
294 0 : print_sched( fd_sched_t * sched ) {
295 0 : FD_LOG_NOTICE(( "sched canary 0x%lx, block_cnt_max %lu, root_idx %lu, active_idx %lu, staged_bitset %lu, staged_head_idx[0] %lu, staged_head_idx[1] %lu, staged_head_idx[2] %lu, staged_head_idx[3] %lu, txn_pool_free_cnt %lu",
296 0 : sched->canary, sched->block_cnt_max, sched->root_idx, sched->active_bank_idx, sched->staged_bitset, sched->staged_head_bank_idx[ 0 ], sched->staged_head_bank_idx[ 1 ], sched->staged_head_bank_idx[ 2 ], sched->staged_head_bank_idx[ 3 ], sched->txn_pool_free_cnt ));
297 0 : }
298 :
299 : FD_FN_UNUSED static void
300 0 : print_all( fd_sched_t * sched, fd_sched_block_t * block ) {
301 0 : print_metrics( sched );
302 0 : print_sched( sched );
303 0 : print_block_and_parent( sched, block );
304 0 : }
305 :
306 :
307 : /* Public functions. */
308 :
309 0 : ulong fd_sched_align( void ) {
310 0 : return fd_ulong_max( alignof(fd_sched_t),
311 0 : fd_ulong_max( fd_rdisp_align(),
312 0 : fd_ulong_max( alignof(fd_sched_block_t), 64UL ))); /* Minimally cache line aligned. */
313 0 : }
314 :
315 : ulong
316 0 : fd_sched_footprint( ulong block_cnt_max ) {
317 0 : ulong l = FD_LAYOUT_INIT;
318 0 : l = FD_LAYOUT_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
319 0 : l = FD_LAYOUT_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( FD_SCHED_MAX_DEPTH, block_cnt_max ) ); /* dispatcher */
320 0 : l = FD_LAYOUT_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) ); /* block pool */
321 0 : return FD_LAYOUT_FINI( l, fd_sched_align() );
322 0 : }
323 :
324 : void *
325 0 : fd_sched_new( void * mem, ulong block_cnt_max ) {
326 0 : FD_SCRATCH_ALLOC_INIT( l, mem );
327 0 : fd_sched_t * sched = FD_SCRATCH_ALLOC_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
328 0 : void * _rdisp = FD_SCRATCH_ALLOC_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( FD_SCHED_MAX_DEPTH, block_cnt_max ) );
329 0 : void * _bpool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) );
330 0 : FD_SCRATCH_ALLOC_FINI( l, fd_sched_align() );
331 :
332 0 : ulong seed = ((ulong)fd_tickcount()) ^ FD_SCHED_MAGIC;
333 0 : fd_rdisp_new( _rdisp, FD_SCHED_MAX_DEPTH, block_cnt_max, seed );
334 :
335 0 : fd_sched_block_t * bpool = (fd_sched_block_t *)_bpool;
336 0 : for( ulong i=0; i<block_cnt_max; i++ ) {
337 0 : bpool[ i ].in_sched = 0;
338 0 : }
339 :
340 0 : fd_memset( sched->metrics, 0, sizeof(fd_sched_metrics_t) );
341 0 : sched->txn_in_flight_last_tick = LONG_MAX;
342 :
343 0 : sched->canary = FD_SCHED_MAGIC;
344 0 : sched->block_cnt_max = block_cnt_max;
345 0 : sched->root_idx = ULONG_MAX;
346 0 : sched->active_bank_idx = ULONG_MAX;
347 0 : sched->staged_bitset = 0UL;
348 :
349 0 : sched->txn_pool_free_cnt = FD_SCHED_MAX_DEPTH-1UL; /* -1 because index 0 is unusable as a sentinel reserved by the dispatcher */
350 :
351 0 : txn_bitset_new( sched->exec_done_set );
352 0 : txn_bitset_new( sched->sigverify_done_set );
353 :
354 0 : return sched;
355 0 : }
356 :
357 : fd_sched_t *
358 0 : fd_sched_join( void * mem, ulong block_cnt_max ) {
359 0 : fd_sched_t * sched = (fd_sched_t *)mem;
360 :
361 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
362 0 : FD_TEST( sched->block_cnt_max==block_cnt_max );
363 :
364 0 : FD_SCRATCH_ALLOC_INIT( l, mem );
365 0 : /* */ FD_SCRATCH_ALLOC_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
366 0 : void * _rdisp = FD_SCRATCH_ALLOC_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( FD_SCHED_MAX_DEPTH, block_cnt_max ) );
367 0 : void * _bpool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) );
368 0 : FD_SCRATCH_ALLOC_FINI( l, fd_sched_align() );
369 :
370 0 : sched->rdisp = fd_rdisp_join( _rdisp );
371 0 : sched->block_pool = _bpool;
372 :
373 0 : txn_bitset_join( sched->exec_done_set );
374 0 : txn_bitset_join( sched->sigverify_done_set );
375 :
376 0 : return sched;
377 0 : }
378 :
379 : int
380 0 : fd_sched_fec_can_ingest( fd_sched_t * sched, fd_sched_fec_t * fec ) {
381 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
382 0 : FD_TEST( fec->bank_idx<sched->block_cnt_max );
383 0 : FD_TEST( fec->parent_bank_idx<sched->block_cnt_max );
384 :
385 0 : if( FD_UNLIKELY( fec->fec->data_sz>FD_SCHED_MAX_PAYLOAD_PER_FEC ) ) {
386 0 : FD_LOG_CRIT(( "invalid FEC set: fec->data_sz %lu, fec->mr %s, slot %lu, parent slot %lu",
387 0 : fec->fec->data_sz, FD_BASE58_ENC_32_ALLOCA( fec->fec->key.mr.hash ), fec->slot, fec->parent_slot ));
388 0 : }
389 :
390 0 : ulong fec_buf_sz = 0UL;
391 0 : fd_sched_block_t * block = block_pool_ele( sched, fec->bank_idx );
392 0 : if( FD_LIKELY( !fec->is_first_in_block ) ) {
393 0 : fec_buf_sz += block->fec_buf_sz-block->fec_buf_soff;
394 0 : } else {
395 : /* No residual data as this is a fresh new block. */
396 0 : }
397 : /* Addition is safe and won't overflow because we checked the FEC set
398 : size above. */
399 0 : fec_buf_sz += fec->fec->data_sz;
400 : /* Assuming every transaction is min size, do we have enough free
401 : entries in the txn pool? For a more precise txn count, we would
402 : have to do some parsing. */
403 0 : return sched->txn_pool_free_cnt>=fec_buf_sz/FD_TXN_MIN_SERIALIZED_SZ;
404 0 : }
405 :
406 : int
407 0 : fd_sched_can_ingest( fd_sched_t * sched ) {
408 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
409 : /* Worst case, we need one byte from the incoming data to extract a
410 : transaction out of the residual data, and the rest of the incoming
411 : data contributes toward min sized transactions. */
412 0 : ulong txn_cnt = (FD_SCHED_MAX_PAYLOAD_PER_FEC-1UL)/FD_TXN_MIN_SERIALIZED_SZ+1UL; /* 478 */
413 0 : return sched->txn_pool_free_cnt>=txn_cnt;
414 0 : }
415 :
416 : FD_WARN_UNUSED int
417 0 : fd_sched_fec_ingest( fd_sched_t * sched, fd_sched_fec_t * fec ) {
418 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
419 0 : FD_TEST( fec->bank_idx<sched->block_cnt_max );
420 0 : FD_TEST( fec->parent_bank_idx<sched->block_cnt_max );
421 :
422 0 : if( FD_UNLIKELY( fec->fec->data_sz>FD_SCHED_MAX_PAYLOAD_PER_FEC ) ) {
423 0 : FD_LOG_CRIT(( "invalid FEC set: fec->data_sz %lu, fec->mr %s, slot %lu, parent slot %lu",
424 0 : fec->fec->data_sz, FD_BASE58_ENC_32_ALLOCA( fec->fec->key.mr.hash ), fec->slot, fec->parent_slot ));
425 0 : }
426 :
427 0 : fd_sched_block_t * block = block_pool_ele( sched, fec->bank_idx );
428 0 : if( FD_UNLIKELY( fec->is_first_in_block ) ) {
429 : /* This is a new block. */
430 0 : add_block( sched, fec->bank_idx, fec->parent_bank_idx );
431 0 : block->slot = fec->slot;
432 0 : block->parent_slot = fec->parent_slot;
433 :
434 0 : if( FD_UNLIKELY( block->dying ) ) {
435 : /* The child of a dead block is also dead. We added it to our
436 : fork tree just so we could track an entire lineage of dead
437 : children and propagate the dead property to the entire lineage,
438 : in case there were frags for more than one dead children
439 : in-flight at the time the parent was abandoned. That being
440 : said, we shouldn't need to add the dead child to the
441 : dispatcher. */
442 0 : sched->metrics->block_added_dead_ood_cnt++;
443 :
444 : /* Ignore the FEC set for a dead block. */
445 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
446 0 : return 1;
447 0 : }
448 :
449 : /* Try to find a staging lane for this block. */
450 0 : int alloc_lane = 0;
451 0 : fd_sched_block_t * parent_block = block_pool_ele( sched, fec->parent_bank_idx );
452 0 : if( FD_LIKELY( parent_block->staged ) ) {
453 : /* Parent is staged. So see if we can continue down the same
454 : staging lane. */
455 0 : ulong staging_lane = parent_block->staging_lane;
456 0 : ulong child_idx = parent_block->child_idx;
457 0 : while( child_idx!=ULONG_MAX ) {
458 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
459 0 : if( child->staged && child->staging_lane==staging_lane ) {
460 : /* Found a child on the same lane. So we're done. */
461 0 : staging_lane = FD_RDISP_UNSTAGED;
462 0 : break;
463 0 : }
464 0 : child_idx = child->sibling_idx;
465 0 : }
466 : /* No child is staged on the same lane as the parent. So stage
467 : this block. This is the common case. */
468 0 : if( FD_LIKELY( staging_lane!=FD_RDISP_UNSTAGED ) ) {
469 0 : block->in_rdisp = 1;
470 0 : block->staged = 1;
471 0 : block->staging_lane = staging_lane;
472 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, staging_lane );
473 0 : sched->metrics->block_added_cnt++;
474 0 : sched->metrics->block_added_staged_cnt++;
475 0 : } else {
476 0 : alloc_lane = 1;
477 0 : }
478 0 : } else {
479 0 : if( block_is_stageable( parent_block ) ) {
480 : /* Parent is unstaged but stageable. So let's be unstaged too.
481 : This is a policy decision to be lazy and not promote parent
482 : at the moment. */
483 0 : block->in_rdisp = 1;
484 0 : block->staged = 0;
485 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, FD_RDISP_UNSTAGED );
486 0 : sched->metrics->block_added_cnt++;
487 0 : sched->metrics->block_added_unstaged_cnt++;
488 0 : } else {
489 0 : alloc_lane = 1;
490 0 : }
491 0 : }
492 0 : if( FD_UNLIKELY( alloc_lane ) ) {
493 : /* We weren't able to inherit the parent's staging lane. So try
494 : to find a new staging lane. */
495 0 : if( FD_LIKELY( sched->staged_bitset!=fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) ) ) { /* Optimize for lane available. */
496 0 : int lane_idx = fd_ulong_find_lsb( ~sched->staged_bitset );
497 0 : if( FD_UNLIKELY( lane_idx>=(int)FD_SCHED_MAX_STAGING_LANES ) ) {
498 0 : FD_LOG_CRIT(( "invariant violation: lane_idx %d, sched->staged_bitset %lx",
499 0 : lane_idx, sched->staged_bitset ));
500 0 : }
501 0 : sched->staged_bitset = fd_ulong_set_bit( sched->staged_bitset, lane_idx );
502 0 : sched->staged_head_bank_idx[ lane_idx ] = fec->bank_idx;
503 0 : block->in_rdisp = 1;
504 0 : block->staged = 1;
505 0 : block->staging_lane = (ulong)lane_idx;
506 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, block->staging_lane );
507 0 : sched->metrics->block_added_cnt++;
508 0 : sched->metrics->block_added_staged_cnt++;
509 0 : } else {
510 : /* No lanes available. */
511 0 : block->in_rdisp = 1;
512 0 : block->staged = 0;
513 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, FD_RDISP_UNSTAGED );
514 0 : sched->metrics->block_added_cnt++;
515 0 : sched->metrics->block_added_unstaged_cnt++;
516 0 : }
517 0 : }
518 0 : }
519 :
520 0 : if( FD_UNLIKELY( block->dying ) ) {
521 : /* Ignore the FEC set for a dead block. */
522 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
523 0 : return 1;
524 0 : }
525 :
526 0 : if( FD_UNLIKELY( !block->in_rdisp ) ) {
527 : /* Invariant: block must be in the dispatcher at this point. */
528 0 : FD_LOG_CRIT(( "invariant violation: block->in_rdisp==0, slot %lu, parent slot %lu",
529 0 : block->slot, block->parent_slot ));
530 0 : }
531 :
532 0 : if( FD_UNLIKELY( block->fec_eos ) ) {
533 : /* This means something is wrong upstream. */
534 0 : FD_LOG_CRIT(( "invariant violation: block->fec_eos set but getting more FEC sets fec->mr %s, slot %lu, parent slot %lu",
535 0 : FD_BASE58_ENC_32_ALLOCA( fec->fec->key.mr.hash ), fec->slot, fec->parent_slot ));
536 0 : }
537 0 : if( FD_UNLIKELY( block->fec_eob && fec->is_last_in_batch ) ) {
538 : /* This means the previous batch didn't parse properly. So this is
539 : a bad block. We should refuse to replay down the fork. */
540 0 : FD_LOG_INFO(( "bad block: block->fec_eob set but getting another FEC set that is last in batch fec->mr %s, slot %lu, parent slot %lu",
541 0 : FD_BASE58_ENC_32_ALLOCA( fec->fec->key.mr.hash ), fec->slot, fec->parent_slot ));
542 0 : subtree_abandon( sched, block );
543 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
544 0 : check_or_set_active_block( sched );
545 0 : sched->metrics->block_bad_cnt++;
546 0 : return 0;
547 0 : }
548 0 : if( FD_UNLIKELY( block->child_idx!=ULONG_MAX ) ) {
549 : /* This means something is wrong upstream. FEC sets are not being
550 : delivered in replay order. */
551 0 : FD_LOG_CRIT(( "invariant violation: block->child_idx %lu, fec->mr %s, slot %lu, parent slot %lu",
552 0 : block->child_idx, FD_BASE58_ENC_32_ALLOCA( fec->fec->key.mr.hash ), fec->slot, fec->parent_slot ));
553 0 : }
554 :
555 0 : FD_TEST( block->fec_buf_sz>=block->fec_buf_soff );
556 0 : if( FD_LIKELY( block->fec_buf_sz>block->fec_buf_soff ) ) {
557 : /* If there is residual data from the previous FEC set within the
558 : same batch, we move it to the beginning of the buffer and append
559 : the new FEC set. */
560 0 : memmove( block->fec_buf, block->fec_buf+block->fec_buf_soff, block->fec_buf_sz-block->fec_buf_soff );
561 0 : }
562 0 : block->fec_buf_sz -= block->fec_buf_soff;
563 0 : block->fec_buf_soff = 0;
564 : /* Addition is safe and won't overflow because we checked the FEC
565 : set size above. */
566 0 : if( FD_UNLIKELY( block->fec_buf_sz-block->fec_buf_soff+fec->fec->data_sz>FD_SCHED_MAX_FEC_BUF_SZ ) ) {
567 : /* In a conformant block, there shouldn't be more than a
568 : transaction's worth of residual data left over from the previous
569 : FEC set within the same batch. So if this condition doesn't
570 : hold, it's a bad block. Instead of crashing, we should refuse to
571 : replay down the fork. */
572 0 : FD_LOG_INFO(( "bad block: fec_buf_sz %u, fec_buf_soff %u, fec->data_sz %lu, fec->mr %s, slot %lu, parent slot %lu",
573 0 : block->fec_buf_sz, block->fec_buf_soff, fec->fec->data_sz, FD_BASE58_ENC_32_ALLOCA( fec->fec->key.mr.hash ), fec->slot, fec->parent_slot ));
574 0 : subtree_abandon( sched, block );
575 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
576 0 : sched->metrics->block_bad_cnt++;
577 0 : check_or_set_active_block( sched );
578 0 : return 0;
579 0 : }
580 :
581 0 : block->shred_cnt += fec->shred_cnt;
582 0 : sched->metrics->fec_cnt++;
583 :
584 : /* Append the new FEC set to the end of the buffer. */
585 0 : fd_memcpy( block->fec_buf+block->fec_buf_sz, fec->fec->data, fec->fec->data_sz );
586 0 : block->fec_buf_sz += (uint)fec->fec->data_sz;
587 0 : sched->metrics->bytes_ingested_cnt += fec->fec->data_sz;
588 :
589 0 : block->fec_eob = fec->is_last_in_batch;
590 0 : block->fec_eos = fec->is_last_in_block;
591 :
592 0 : int err = fd_sched_parse( sched, block, fec->alut_ctx );
593 0 : if( FD_UNLIKELY( err==FD_SCHED_PARSER_BAD_BLOCK ) ) {
594 0 : FD_LOG_INFO(( "bad block: slot %lu, parent slot %lu, txn_parsed_cnt %u, txn_done_cnt %u",
595 0 : fec->slot, fec->parent_slot, block->txn_parsed_cnt, block->txn_done_cnt ));
596 0 : subtree_abandon( sched, block );
597 0 : sched->metrics->bytes_dropped_cnt += block->fec_buf_sz-block->fec_buf_soff;
598 0 : sched->metrics->block_bad_cnt++;
599 0 : check_or_set_active_block( sched );
600 0 : return 0;
601 0 : }
602 :
603 : /* Check if we need to set the active block. */
604 0 : check_or_set_active_block( sched );
605 :
606 0 : return 1;
607 0 : }
608 :
609 : ulong
610 0 : fd_sched_task_next_ready( fd_sched_t * sched, fd_sched_task_t * out, ulong exec_tile_cnt ) {
611 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
612 :
613 0 : if( FD_UNLIKELY( sched->active_bank_idx==ULONG_MAX ) ) {
614 : /* No need to try activating a block. If we're in this state,
615 : there's truly nothing to execute. We will activate something
616 : when we ingest a FEC set with transactions. */
617 0 : return 0UL;
618 0 : }
619 :
620 0 : out->task_type = FD_SCHED_TT_NULL;
621 :
622 : /* We could in theory reevaluate staging lane allocation here and do
623 : promotion/demotion as needed. It's a policy decision to minimize
624 : fork churn for now and just execute down the same active fork. */
625 :
626 0 : ulong bank_idx = sched->active_bank_idx;
627 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
628 0 : if( FD_UNLIKELY( block_should_deactivate( block ) ) ) {
629 0 : print_all( sched, block );
630 0 : FD_LOG_CRIT(( "invariant violation: active_bank_idx %lu is not activatable nor has anything in-flight", sched->active_bank_idx ));
631 0 : }
632 :
633 0 : if( FD_UNLIKELY( !block->block_start_signaled ) ) {
634 0 : out->task_type = FD_SCHED_TT_BLOCK_START;
635 0 : out->block_start->bank_idx = bank_idx;
636 0 : out->block_start->parent_bank_idx = block->parent_idx;
637 0 : out->block_start->slot = block->slot;
638 0 : block->block_start_signaled = 1;
639 0 : return 1UL;
640 0 : }
641 :
642 0 : ulong exec_queued_cnt = block->txn_parsed_cnt-block->txn_exec_in_flight_cnt-block->txn_exec_done_cnt;
643 0 : if( FD_LIKELY( exec_queued_cnt>0UL ) ) { /* Optimize for no fork switching. */
644 0 : out->txn_exec->txn_idx = fd_rdisp_get_next_ready( sched->rdisp, bank_idx );
645 0 : if( FD_UNLIKELY( out->txn_exec->txn_idx==0UL ) ) {
646 : /* There are transactions queued but none ready for execution.
647 : This implies that there must be in-flight transactions on whose
648 : completion the queued transactions depend. So we return and
649 : wait for those in-flight transactions to retire. This is a
650 : policy decision to execute as much as we can down the current
651 : fork. */
652 0 : if( FD_UNLIKELY( !block->txn_exec_in_flight_cnt ) ) {
653 0 : print_all( sched, block );
654 0 : FD_LOG_CRIT(( "invariant violation: no ready transaction found but block->txn_exec_in_flight_cnt==0" ));
655 0 : }
656 0 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
657 0 : if( FD_LIKELY( sigverify_queued_cnt>0UL && exec_tile_cnt>1UL ) ) { /* Leave one exec tile idle for critical path execution. */
658 : /* Dispatch transactions for sigverify in parse order. */
659 0 : out->task_type = FD_SCHED_TT_TXN_SIGVERIFY;
660 0 : out->txn_sigverify->bank_idx = bank_idx;
661 0 : out->txn_sigverify->txn_idx = block->txn_idx[ block->txn_sigverify_done_cnt+block->txn_sigverify_in_flight_cnt ];
662 0 : block->txn_sigverify_in_flight_cnt++;
663 0 : return 1UL;
664 0 : }
665 0 : return 0UL;
666 0 : }
667 0 : out->task_type = FD_SCHED_TT_TXN_EXEC;
668 0 : out->txn_exec->bank_idx = bank_idx;
669 0 : out->txn_exec->slot = block->slot;
670 :
671 0 : long now = fd_tickcount();
672 0 : ulong delta = (ulong)(now-sched->txn_in_flight_last_tick);
673 0 : sched->metrics->txn_none_in_flight_tickcount += fd_ulong_if( block->txn_exec_in_flight_cnt==0U && sched->txn_in_flight_last_tick!=LONG_MAX, delta, 0UL );
674 0 : sched->metrics->txn_weighted_in_flight_tickcount += fd_ulong_if( block->txn_exec_in_flight_cnt!=0U, delta, 0UL );
675 0 : sched->metrics->txn_weighted_in_flight_cnt += delta*block->txn_exec_in_flight_cnt;
676 0 : sched->txn_in_flight_last_tick = now;
677 :
678 0 : block->txn_exec_in_flight_cnt++;
679 0 : sched->metrics->txn_max_in_flight_cnt = fd_uint_max( sched->metrics->txn_max_in_flight_cnt, block->txn_exec_in_flight_cnt );
680 0 : return 1UL;
681 0 : }
682 :
683 : /* At this point txn_queued_cnt==0 */
684 :
685 : /* Try to dispatch a sigverify task, but leave one exec tile idle for
686 : critical path execution, unless there's not going to be any more
687 : real transactions for the critical path. */
688 0 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
689 0 : if( FD_LIKELY( sigverify_queued_cnt>0UL && exec_tile_cnt>fd_ulong_if( block->fec_eos, 0UL, 1UL ) ) ) {
690 : /* Dispatch transactions for sigverify in parse order. */
691 0 : out->task_type = FD_SCHED_TT_TXN_SIGVERIFY;
692 0 : out->txn_sigverify->txn_idx = block->txn_idx[ block->txn_sigverify_done_cnt+block->txn_sigverify_in_flight_cnt ];
693 0 : out->txn_sigverify->bank_idx = bank_idx;
694 0 : block->txn_sigverify_in_flight_cnt++;
695 0 : return 1UL;
696 0 : }
697 :
698 0 : if( FD_UNLIKELY( block_should_signal_end( block ) ) ) {
699 0 : FD_TEST( block->block_start_signaled );
700 0 : out->task_type = FD_SCHED_TT_BLOCK_END;
701 0 : out->block_end->bank_idx = bank_idx;
702 0 : block->block_end_signaled = 1;
703 0 : return 1UL;
704 0 : }
705 :
706 : /* Nothing queued for the active block. If we haven't received all
707 : the FEC sets for it, then return and wait for more FEC sets, while
708 : there are in-flight transactions. This is a policy decision to
709 : minimize fork churn and allow for executing down the current fork
710 : as much as we can. If we have received all the FEC sets for it,
711 : then we'd still like to return and wait for the in-flight
712 : transactions to retire, before switching to a different block.
713 :
714 : Either way, there should be in-flight transactions. We deactivate
715 : the active block the moment we exhausted transactions from it. */
716 0 : if( FD_UNLIKELY( !block_is_in_flight( block ) ) ) {
717 0 : print_all( sched, block );
718 0 : FD_LOG_CRIT(( "invariant violation: expected in-flight transactions but none" ));
719 0 : }
720 :
721 0 : return 0UL;
722 0 : }
723 :
724 : void
725 0 : fd_sched_task_done( fd_sched_t * sched, ulong task_type, ulong txn_idx ) {
726 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
727 :
728 0 : ulong bank_idx = ULONG_MAX;
729 0 : switch( task_type ) {
730 0 : case FD_SCHED_TT_BLOCK_START:
731 0 : case FD_SCHED_TT_BLOCK_END: {
732 0 : (void)txn_idx;
733 0 : bank_idx = sched->active_bank_idx;
734 0 : break;
735 0 : }
736 0 : case FD_SCHED_TT_TXN_EXEC:
737 0 : case FD_SCHED_TT_TXN_SIGVERIFY: {
738 0 : FD_TEST( txn_idx<FD_SCHED_MAX_DEPTH );
739 0 : bank_idx = sched->txn_to_bank_idx[ txn_idx ];
740 0 : break;
741 0 : }
742 0 : default: FD_LOG_CRIT(( "unsupported task_type %lu", task_type ));
743 0 : }
744 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
745 :
746 0 : if( FD_UNLIKELY( !block->staged ) ) {
747 : /* Invariant: only staged blocks can have in-flight transactions. */
748 0 : FD_LOG_CRIT(( "invariant violation: block->staged==0, slot %lu, parent slot %lu",
749 0 : block->slot, block->parent_slot ));
750 0 : }
751 0 : if( FD_UNLIKELY( !block->in_rdisp ) ) {
752 : /* Invariant: staged blocks must be in the dispatcher. */
753 0 : FD_LOG_CRIT(( "invariant violation: block->in_rdisp==0, slot %lu, parent slot %lu",
754 0 : block->slot, block->parent_slot ));
755 0 : }
756 :
757 0 : switch( task_type ) {
758 0 : case FD_SCHED_TT_BLOCK_START: {
759 0 : FD_TEST( !block->block_start_done );
760 0 : block->block_start_done = 1;
761 0 : break;
762 0 : }
763 0 : case FD_SCHED_TT_BLOCK_END: {
764 : /* It may seem redundant to be invoking task_done() on these
765 : somewhat fake tasks. But these are necessary to drive state
766 : transition for empty blocks or slow blocks. */
767 0 : FD_TEST( !block->block_end_done );
768 0 : block->block_end_done = 1;
769 0 : break;
770 0 : }
771 0 : case FD_SCHED_TT_TXN_EXEC: {
772 0 : long now = fd_tickcount();
773 0 : ulong delta = (ulong)(now-sched->txn_in_flight_last_tick);
774 0 : sched->metrics->txn_weighted_in_flight_tickcount += delta;
775 0 : sched->metrics->txn_weighted_in_flight_cnt += delta*block->txn_exec_in_flight_cnt;
776 0 : sched->txn_in_flight_last_tick = now;
777 :
778 0 : block->txn_exec_done_cnt++;
779 0 : block->txn_exec_in_flight_cnt--;
780 0 : sched->metrics->txn_exec_done_cnt++;
781 0 : txn_bitset_insert( sched->exec_done_set, txn_idx );
782 0 : if( txn_bitset_test( sched->exec_done_set, txn_idx ) && txn_bitset_test( sched->sigverify_done_set, txn_idx ) ) {
783 : /* Release txn_idx if both exec and sigverify are done. This is
784 : guaranteed to only happen once per transaction because
785 : whichever one completed first would not release. */
786 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 1 );
787 0 : sched->txn_pool_free_cnt++;
788 0 : block->txn_done_cnt++;
789 0 : sched->metrics->txn_done_cnt++;
790 0 : } else {
791 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 0 );
792 0 : }
793 0 : break;
794 0 : }
795 0 : case FD_SCHED_TT_TXN_SIGVERIFY: {
796 0 : block->txn_sigverify_done_cnt++;
797 0 : block->txn_sigverify_in_flight_cnt--;
798 0 : sched->metrics->txn_sigverify_done_cnt++;
799 0 : txn_bitset_insert( sched->sigverify_done_set, txn_idx );
800 0 : if( txn_bitset_test( sched->exec_done_set, txn_idx ) && txn_bitset_test( sched->sigverify_done_set, txn_idx ) ) {
801 : /* Release txn_idx if both exec and sigverify are done. This is
802 : guaranteed to only happen once per transaction because
803 : whichever one completed first would not release. */
804 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 1 );
805 0 : sched->txn_pool_free_cnt++;
806 0 : block->txn_done_cnt++;
807 0 : sched->metrics->txn_done_cnt++;
808 0 : }
809 0 : break;
810 0 : }
811 0 : }
812 :
813 0 : if( FD_UNLIKELY( block->dying && !block_is_in_flight( block ) ) ) {
814 0 : if( FD_UNLIKELY( sched->active_bank_idx==bank_idx ) ) {
815 0 : FD_LOG_CRIT(( "invariant violation: active block shouldn't be dying, bank_idx %lu, slot %lu, parent slot %lu",
816 0 : bank_idx, block->slot, block->parent_slot ));
817 0 : }
818 0 : subtree_abandon( sched, block );
819 0 : return;
820 0 : }
821 :
822 0 : if( FD_UNLIKELY( !block->dying && sched->active_bank_idx!=bank_idx ) ) {
823 : /* Block is not dead. So we should be actively replaying it. */
824 0 : fd_sched_block_t * active_block = block_pool_ele( sched, sched->active_bank_idx );
825 0 : FD_LOG_CRIT(( "invariant violation: sched->active_bank_idx %lu, slot %lu, parent slot %lu, bank_idx %lu, slot %lu, parent slot %lu",
826 0 : sched->active_bank_idx, active_block->slot, active_block->parent_slot,
827 0 : bank_idx, block->slot, block->parent_slot ));
828 0 : }
829 :
830 0 : maybe_switch_block( sched, bank_idx );
831 0 : }
832 :
833 : void
834 0 : fd_sched_block_abandon( fd_sched_t * sched, ulong bank_idx ) {
835 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
836 0 : FD_TEST( bank_idx<sched->block_cnt_max );
837 :
838 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
839 0 : if( FD_UNLIKELY( bank_idx!=sched->active_bank_idx ) ) {
840 : /* Invariant: abandoning should only be performed on actively
841 : replayed blocks. We impose this requirement on the caller
842 : because the dispatcher expects blocks to be abandoned in the same
843 : order that they were added, and having this requirement makes it
844 : easier to please the dispatcher. */
845 0 : print_all( sched, block );
846 0 : FD_LOG_CRIT(( "invariant violation: active_bank_idx %lu, bank_idx %lu, slot %lu, parent slot %lu",
847 0 : sched->active_bank_idx, bank_idx, block->slot, block->parent_slot ));
848 0 : }
849 :
850 0 : subtree_abandon( sched, block );
851 :
852 : /* Reset the active block. */
853 0 : sched->active_bank_idx = ULONG_MAX;
854 0 : sched->metrics->deactivate_abandoned_cnt++;
855 0 : try_activate_block( sched );
856 0 : }
857 :
858 : void
859 0 : fd_sched_block_add_done( fd_sched_t * sched, ulong bank_idx, ulong parent_bank_idx ) {
860 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
861 0 : FD_TEST( bank_idx<sched->block_cnt_max );
862 :
863 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
864 0 : add_block( sched, bank_idx, parent_bank_idx );
865 0 : block->txn_parsed_cnt = UINT_MAX;
866 0 : block->txn_exec_done_cnt = UINT_MAX;
867 0 : block->txn_sigverify_done_cnt = UINT_MAX;
868 0 : block->txn_done_cnt = UINT_MAX;
869 0 : block->fec_eos = 1;
870 0 : block->block_start_signaled = 1;
871 0 : block->block_end_signaled = 1;
872 0 : block->block_start_done = 1;
873 0 : block->block_end_done = 1;
874 0 : if( FD_UNLIKELY( parent_bank_idx==ULONG_MAX ) ) {
875 : /* Assumes that a NULL parent implies the snapshot slot. */
876 0 : sched->root_idx = bank_idx;
877 0 : block->rooted = 1;
878 0 : }
879 0 : }
880 :
881 : void
882 0 : fd_sched_advance_root( fd_sched_t * sched, ulong root_idx ) {
883 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
884 0 : FD_TEST( root_idx<sched->block_cnt_max );
885 0 : FD_TEST( sched->root_idx<sched->block_cnt_max );
886 :
887 0 : fd_sched_block_t * new_root = block_pool_ele( sched, root_idx );
888 0 : fd_sched_block_t * old_root = block_pool_ele( sched, sched->root_idx );
889 0 : if( FD_UNLIKELY( !old_root->rooted ) ) {
890 0 : FD_LOG_CRIT(( "invariant violation: old_root is not rooted, slot %lu, parent slot %lu",
891 0 : old_root->slot, old_root->parent_slot ));
892 0 : }
893 :
894 : /* Early exit if the new root is the same as the old root. */
895 0 : if( FD_UNLIKELY( root_idx==sched->root_idx ) ) {
896 0 : FD_LOG_INFO(( "new root is the same as the old root, slot %lu, parent slot %lu",
897 0 : new_root->slot, new_root->parent_slot ));
898 0 : return;
899 0 : }
900 :
901 0 : fd_sched_block_t * head = old_root;
902 0 : head->parent_idx = ULONG_MAX;
903 0 : fd_sched_block_t * tail = head;
904 :
905 0 : while( head ) {
906 0 : FD_TEST( head->in_sched );
907 0 : head->in_sched = 0;
908 0 : ulong child_idx = head->child_idx;
909 0 : while( child_idx!=ULONG_MAX ) {
910 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
911 : /* Add children to be visited. We abuse the parent_idx field to
912 : link up the next block to visit. */
913 0 : if( child!=new_root ) {
914 0 : tail->parent_idx = child_idx;
915 0 : tail = child;
916 0 : tail->parent_idx = ULONG_MAX;
917 0 : }
918 0 : child_idx = child->sibling_idx;
919 0 : }
920 :
921 : /* Prune the current block. We will never publish halfway into a
922 : staging lane, because anything on the rooted fork should have
923 : finished replaying gracefully and be out of the dispatcher. In
924 : fact, anything that we are publishing away should be out of the
925 : dispatcher at this point. And there should be no more in-flight
926 : transactions. */
927 0 : if( FD_UNLIKELY( block_is_in_flight( head ) ) ) {
928 0 : FD_LOG_CRIT(( "invariant violation: block has transactions in flight (%u exec %u sigverify), slot %lu, parent slot %lu",
929 0 : head->txn_exec_in_flight_cnt, head->txn_sigverify_in_flight_cnt, head->slot, head->parent_slot ));
930 0 : }
931 0 : if( FD_UNLIKELY( head->in_rdisp ) ) {
932 : /* We should have removed it from the dispatcher when we were
933 : notified of the new root, or when in-flight transactions were
934 : drained. */
935 0 : FD_LOG_CRIT(( "invariant violation: block is in the dispatcher, slot %lu, parent slot %lu",
936 0 : head->slot, head->parent_slot ));
937 0 : }
938 0 : fd_sched_block_t * next = block_pool_ele( sched, head->parent_idx );
939 0 : head = next;
940 0 : }
941 :
942 0 : new_root->parent_idx = ULONG_MAX;
943 0 : sched->root_idx = root_idx;
944 0 : }
945 :
946 : void
947 0 : fd_sched_root_notify( fd_sched_t * sched, ulong root_idx ) {
948 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
949 0 : FD_TEST( root_idx<sched->block_cnt_max );
950 0 : FD_TEST( sched->root_idx<sched->block_cnt_max );
951 :
952 0 : fd_sched_block_t * block = block_pool_ele( sched, root_idx );
953 0 : fd_sched_block_t * old_root = block_pool_ele( sched, sched->root_idx );
954 0 : if( FD_UNLIKELY( !old_root->rooted ) ) {
955 0 : FD_LOG_CRIT(( "invariant violation: old_root is not rooted, slot %lu, parent slot %lu",
956 0 : old_root->slot, old_root->parent_slot ));
957 0 : }
958 :
959 : /* Early exit if the new root is the same as the old root. */
960 0 : if( FD_UNLIKELY( root_idx==sched->root_idx ) ) {
961 0 : FD_LOG_INFO(( "new root is the same as the old root, slot %lu, parent slot %lu",
962 0 : block->slot, block->parent_slot ));
963 0 : return;
964 0 : }
965 :
966 : /* Mark every node from the new root up through its parents to the
967 : old root as being rooted. */
968 0 : fd_sched_block_t * curr = block;
969 0 : fd_sched_block_t * prev = NULL;
970 0 : while( curr ) {
971 0 : if( FD_UNLIKELY( !block_is_done( curr ) ) ) {
972 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is not done, slot %lu, parent slot %lu",
973 0 : curr->slot, curr->parent_slot ));
974 0 : }
975 0 : if( FD_UNLIKELY( curr->dying ) ) {
976 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is dying, slot %lu, parent slot %lu",
977 0 : curr->slot, curr->parent_slot ));
978 0 : }
979 0 : if( FD_UNLIKELY( curr->staged ) ) {
980 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is staged, slot %lu, parent slot %lu",
981 0 : curr->slot, curr->parent_slot ));
982 0 : }
983 0 : if( FD_UNLIKELY( curr->in_rdisp ) ) {
984 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is in the dispatcher, slot %lu, parent slot %lu",
985 0 : curr->slot, curr->parent_slot ));
986 0 : }
987 0 : curr->rooted = 1;
988 0 : prev = curr;
989 0 : curr = block_pool_ele( sched, curr->parent_idx );
990 0 : }
991 :
992 : /* If we didn't reach the old root, the new root is not a descendant. */
993 0 : if( FD_UNLIKELY( prev!=old_root ) ) {
994 0 : FD_LOG_CRIT(( "invariant violation: new root is not a descendant of old root, new root slot %lu, parent slot %lu, old root slot %lu, parent slot %lu",
995 0 : block->slot, block->parent_slot, old_root->slot, old_root->parent_slot ));
996 0 : }
997 :
998 0 : ulong old_active_bank_idx = sched->active_bank_idx;
999 :
1000 : /* Now traverse from old root towards new root, and abandon all
1001 : minority forks. */
1002 0 : curr = old_root;
1003 0 : while( curr && curr->rooted && curr!=block ) { /* curr!=block to avoid abandoning good forks. */
1004 0 : fd_sched_block_t * rooted_child_block = NULL;
1005 0 : ulong child_idx = curr->child_idx;
1006 0 : while( child_idx!=ULONG_MAX ) {
1007 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1008 0 : if( child->rooted ) {
1009 0 : rooted_child_block = child;
1010 0 : } else {
1011 : /* This is a minority fork. */
1012 0 : subtree_abandon( sched, child );
1013 0 : }
1014 0 : child_idx = child->sibling_idx;
1015 0 : }
1016 0 : curr = rooted_child_block;
1017 0 : }
1018 :
1019 : /* If the active block got abandoned, we need to reset it. */
1020 0 : if( sched->active_bank_idx==ULONG_MAX ) {
1021 0 : sched->metrics->deactivate_pruned_cnt += fd_uint_if( old_active_bank_idx!=ULONG_MAX, 1U, 0U );
1022 0 : try_activate_block( sched );
1023 0 : }
1024 0 : }
1025 :
1026 : fd_txn_p_t *
1027 0 : fd_sched_get_txn( fd_sched_t * sched, ulong txn_idx ) {
1028 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1029 0 : if( FD_UNLIKELY( txn_idx>=FD_SCHED_MAX_DEPTH ) ) {
1030 0 : return NULL;
1031 0 : }
1032 0 : return sched->txn_pool+txn_idx;
1033 0 : }
1034 :
1035 : fd_hash_t *
1036 0 : fd_sched_get_poh( fd_sched_t * sched, ulong bank_idx ) {
1037 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1038 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1039 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1040 0 : return &block->poh;
1041 0 : }
1042 :
1043 : uint
1044 0 : fd_sched_get_shred_cnt( fd_sched_t * sched, ulong bank_idx ) {
1045 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1046 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1047 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1048 0 : return block->shred_cnt;
1049 0 : }
1050 :
1051 0 : void * fd_sched_leave ( fd_sched_t * sched ) { return sched; }
1052 0 : void * fd_sched_delete( void * mem ) { return mem; }
1053 :
1054 :
1055 : /* Internal helpers. */
1056 :
1057 : static void
1058 : add_block( fd_sched_t * sched,
1059 : ulong bank_idx,
1060 0 : ulong parent_bank_idx ) {
1061 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1062 0 : FD_TEST( !block->in_sched );
1063 :
1064 0 : block->txn_parsed_cnt = 0U;
1065 0 : block->txn_exec_in_flight_cnt = 0U;
1066 0 : block->txn_exec_done_cnt = 0U;
1067 0 : block->txn_sigverify_in_flight_cnt = 0U;
1068 0 : block->txn_sigverify_done_cnt = 0U;
1069 0 : block->txn_done_cnt = 0U;
1070 0 : block->shred_cnt = 0U;
1071 :
1072 0 : block->mblks_rem = 0UL;
1073 0 : block->txns_rem = 0UL;
1074 0 : block->fec_buf_sz = 0U;
1075 0 : block->fec_buf_soff = 0U;
1076 0 : block->fec_eob = 0;
1077 0 : block->fec_sob = 1;
1078 :
1079 0 : block->fec_eos = 0;
1080 0 : block->rooted = 0;
1081 0 : block->dying = 0;
1082 0 : block->in_sched = 1;
1083 0 : block->in_rdisp = 0;
1084 0 : block->block_start_signaled = 0;
1085 0 : block->block_end_signaled = 0;
1086 0 : block->block_start_done = 0;
1087 0 : block->block_end_done = 0;
1088 0 : block->staged = 0;
1089 :
1090 0 : block->luf_depth = 0UL;
1091 :
1092 : /* New leaf node, no child, no sibling. */
1093 0 : block->child_idx = ULONG_MAX;
1094 0 : block->sibling_idx = ULONG_MAX;
1095 0 : block->parent_idx = ULONG_MAX;
1096 :
1097 0 : if( FD_UNLIKELY( parent_bank_idx==ULONG_MAX ) ) {
1098 0 : return;
1099 0 : }
1100 :
1101 : /* node->parent link */
1102 0 : fd_sched_block_t * parent_block = block_pool_ele( sched, parent_bank_idx );
1103 0 : block->parent_idx = parent_bank_idx;
1104 :
1105 : /* parent->node and sibling->node links */
1106 0 : ulong child_idx = bank_idx;
1107 0 : if( FD_LIKELY( parent_block->child_idx==ULONG_MAX ) ) { /* Optimize for no forking. */
1108 0 : parent_block->child_idx = child_idx;
1109 0 : } else {
1110 0 : fd_sched_block_t * curr_block = block_pool_ele( sched, parent_block->child_idx );
1111 0 : while( curr_block->sibling_idx!=ULONG_MAX ) {
1112 0 : curr_block = block_pool_ele( sched, curr_block->sibling_idx );
1113 0 : }
1114 0 : curr_block->sibling_idx = child_idx;
1115 0 : }
1116 :
1117 0 : if( FD_UNLIKELY( parent_block->dying ) ) {
1118 0 : block->dying = 1;
1119 0 : }
1120 0 : }
1121 :
1122 0 : #define CHECK( cond ) do { \
1123 0 : if( FD_UNLIKELY( !(cond) ) ) { \
1124 0 : return FD_SCHED_PARSER_AGAIN_LATER; \
1125 0 : } \
1126 0 : } while( 0 )
1127 :
1128 : /* CHECK that it is safe to read at least n more bytes. */
1129 0 : #define CHECK_LEFT( n ) CHECK( (n)<=(block->fec_buf_sz-block->fec_buf_soff) )
1130 :
1131 : /* Consume as much as possible from the buffer. By the end of this
1132 : function, we will either have residual data that is unparseable only
1133 : because it is a batch that straddles FEC set boundaries, or we will
1134 : have reached the end of a batch. In the former case, any remaining
1135 : bytes should be concatenated with the next FEC set for further
1136 : parsing. In the latter case, any remaining bytes should be thrown
1137 : away. */
1138 : FD_WARN_UNUSED static int
1139 0 : fd_sched_parse( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx ) {
1140 0 : while( 1 ) {
1141 0 : while( block->txns_rem>0UL ) {
1142 0 : int err;
1143 0 : if( FD_UNLIKELY( (err=fd_sched_parse_txn( sched, block, alut_ctx ))!=FD_SCHED_PARSER_OK ) ) {
1144 0 : return err;
1145 0 : }
1146 0 : }
1147 0 : if( block->txns_rem==0UL && block->mblks_rem>0UL ) {
1148 0 : CHECK_LEFT( sizeof(fd_microblock_hdr_t) );
1149 0 : fd_microblock_hdr_t * hdr = (fd_microblock_hdr_t *)fd_type_pun( block->fec_buf+block->fec_buf_soff );
1150 0 : block->fec_buf_soff += (uint)sizeof(fd_microblock_hdr_t);
1151 :
1152 0 : memcpy( block->poh.hash, hdr->hash, sizeof(block->poh.hash) );
1153 0 : block->txns_rem = hdr->txn_cnt;
1154 0 : block->mblks_rem--;
1155 0 : continue;
1156 0 : }
1157 0 : if( block->txns_rem==0UL && block->mblks_rem==0UL && block->fec_sob ) {
1158 0 : CHECK_LEFT( sizeof(ulong) );
1159 0 : FD_TEST( block->fec_buf_soff==0U );
1160 0 : block->mblks_rem = FD_LOAD( ulong, block->fec_buf );
1161 0 : block->fec_buf_soff += (uint)sizeof(ulong);
1162 : /* FIXME what happens if someone sends us mblks_rem==0UL here? */
1163 :
1164 0 : block->fec_sob = 0;
1165 0 : continue;
1166 0 : }
1167 0 : if( block->txns_rem==0UL && block->mblks_rem==0UL ) {
1168 0 : break;
1169 0 : }
1170 0 : }
1171 0 : if( block->fec_eob ) {
1172 : /* Ignore trailing bytes at the end of a batch. */
1173 0 : sched->metrics->bytes_ingested_unparsed_cnt += block->fec_buf_sz-block->fec_buf_soff;
1174 0 : block->fec_buf_soff = 0U;
1175 0 : block->fec_buf_sz = 0U;
1176 0 : block->fec_sob = 1;
1177 0 : block->fec_eob = 0;
1178 0 : }
1179 0 : return FD_SCHED_PARSER_OK;
1180 0 : }
1181 :
1182 : FD_WARN_UNUSED static int
1183 0 : fd_sched_parse_txn( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx ) {
1184 0 : fd_txn_t * txn = fd_type_pun( block->txn );
1185 :
1186 0 : ulong pay_sz = 0UL;
1187 0 : ulong txn_sz = fd_txn_parse_core( block->fec_buf+block->fec_buf_soff,
1188 0 : fd_ulong_min( FD_TXN_MTU, block->fec_buf_sz-block->fec_buf_soff ),
1189 0 : txn,
1190 0 : NULL,
1191 0 : &pay_sz );
1192 :
1193 0 : if( FD_UNLIKELY( !pay_sz || !txn_sz ) ) {
1194 : /* Can't parse out a full transaction. */
1195 0 : return FD_SCHED_PARSER_AGAIN_LATER;
1196 0 : }
1197 :
1198 0 : if( FD_UNLIKELY( block->txn_parsed_cnt>=FD_MAX_TXN_PER_SLOT ) ) {
1199 : /* The block contains more transactions than a valid block would.
1200 : Mark the block dead instead of keep processing it. */
1201 0 : return FD_SCHED_PARSER_BAD_BLOCK;
1202 0 : }
1203 :
1204 : /* Try to expand ALUTs. */
1205 0 : int has_aluts = txn->transaction_version==FD_TXN_V0 && txn->addr_table_adtl_cnt>0;
1206 0 : int serializing = 0;
1207 0 : if( has_aluts ) {
1208 : /* FIXME: statically size out slot hashes decode footprint. */
1209 0 : FD_SPAD_FRAME_BEGIN( alut_ctx->runtime_spad ) {
1210 0 : fd_slot_hashes_global_t const * slot_hashes_global = fd_sysvar_slot_hashes_read( alut_ctx->funk, alut_ctx->xid, alut_ctx->runtime_spad );
1211 0 : if( FD_LIKELY( slot_hashes_global ) ) {
1212 0 : fd_slot_hash_t * slot_hash = deq_fd_slot_hash_t_join( (uchar *)slot_hashes_global + slot_hashes_global->hashes_offset );
1213 0 : serializing = !!fd_runtime_load_txn_address_lookup_tables( txn, block->fec_buf+block->fec_buf_soff, alut_ctx->funk, alut_ctx->xid, alut_ctx->els, slot_hash, block->aluts );
1214 0 : sched->metrics->alut_success_cnt += (uint)!serializing;
1215 0 : } else {
1216 0 : serializing = 1;
1217 0 : }
1218 0 : } FD_SPAD_FRAME_END;
1219 0 : }
1220 :
1221 0 : ulong bank_idx = (ulong)(block-sched->block_pool);
1222 0 : ulong txn_idx = fd_rdisp_add_txn( sched->rdisp, bank_idx, txn, block->fec_buf+block->fec_buf_soff, serializing ? NULL : block->aluts, serializing );
1223 0 : FD_TEST( txn_idx!=0UL );
1224 0 : sched->metrics->txn_parsed_cnt++;
1225 0 : sched->metrics->alut_serializing_cnt += (uint)serializing;
1226 0 : sched->txn_pool_free_cnt--;
1227 0 : fd_txn_p_t * txn_p = sched->txn_pool + txn_idx;
1228 0 : txn_p->payload_sz = pay_sz;
1229 0 : fd_memcpy( txn_p->payload, block->fec_buf+block->fec_buf_soff, pay_sz );
1230 0 : fd_memcpy( TXN(txn_p), txn, txn_sz );
1231 0 : sched->txn_to_bank_idx[ txn_idx ] = bank_idx;
1232 0 : txn_bitset_remove( sched->exec_done_set, txn_idx );
1233 0 : txn_bitset_remove( sched->sigverify_done_set, txn_idx );
1234 0 : block->txn_idx[ block->txn_parsed_cnt ] = txn_idx;
1235 0 : block->fec_buf_soff += (uint)pay_sz;
1236 0 : block->txn_parsed_cnt++;
1237 : #if FD_SCHED_SKIP_SIGVERIFY
1238 : txn_bitset_insert( sched->sigverify_done_set, txn_idx );
1239 : block->txn_sigverify_done_cnt++;
1240 : #endif
1241 0 : block->txns_rem--;
1242 0 : return FD_SCHED_PARSER_OK;
1243 0 : }
1244 :
1245 : #undef CHECK
1246 : #undef CHECK_LEFT
1247 :
1248 : static void
1249 0 : try_activate_block( fd_sched_t * sched ) {
1250 :
1251 : /* See if there are any allocated staging lanes that we can activate
1252 : for scheduling ... */
1253 0 : ulong staged_bitset = sched->staged_bitset;
1254 0 : while( staged_bitset ) {
1255 0 : int lane_idx = fd_ulong_find_lsb( staged_bitset );
1256 0 : staged_bitset = fd_ulong_pop_lsb( staged_bitset );
1257 :
1258 0 : ulong head_idx = sched->staged_head_bank_idx[ lane_idx ];
1259 0 : fd_sched_block_t * head_block = block_pool_ele( sched, head_idx );
1260 0 : fd_sched_block_t * parent_block = block_pool_ele( sched, head_block->parent_idx );
1261 0 : if( FD_UNLIKELY( parent_block->dying ) ) {
1262 : /* Invariant: no child of a dying block should be staged. */
1263 0 : FD_LOG_CRIT(( "invariant violation: staged_head_bank_idx %lu, slot %lu, parent slot %lu on lane %d has parent_block->dying set, slot %lu, parent slot %lu",
1264 0 : head_idx, head_block->slot, head_block->parent_slot, lane_idx, parent_block->slot, parent_block->parent_slot ));
1265 0 : }
1266 : //FIXME: restore this invariant check when we have immediate demotion of dying blocks
1267 : // if( FD_UNLIKELY( head_block->dying ) ) {
1268 : // /* Invariant: no dying block should be staged. */
1269 : // FD_LOG_CRIT(( "invariant violation: staged_head_bank_idx %lu, slot %lu, prime %lu on lane %u has head_block->dying set",
1270 : // head_idx, (ulong)head_block->block_id.slot, (ulong)head_block->block_id.prime, lane_idx ));
1271 : // }
1272 0 : if( block_is_done( parent_block ) && block_is_activatable( head_block ) ) {
1273 : /* ... Yes, on this staging lane the parent block is done. So we
1274 : can switch to the staged child. */
1275 0 : sched->active_bank_idx = head_idx;
1276 0 : sched->metrics->lane_switch_cnt++;
1277 0 : return;
1278 0 : }
1279 0 : }
1280 :
1281 : /* ... No, promote unstaged blocks. */
1282 0 : ulong root_idx = sched->root_idx;
1283 0 : if( FD_UNLIKELY( root_idx==ULONG_MAX ) ) {
1284 0 : FD_LOG_CRIT(( "invariant violation: root_idx==ULONG_MAX indicating fd_sched is unintialized" ));
1285 0 : }
1286 : /* Find and stage the longest stageable unstaged fork. This is a
1287 : policy decision. */
1288 0 : ulong depth = compute_longest_unstaged_fork( sched, root_idx );
1289 0 : if( FD_LIKELY( depth>0UL ) ) {
1290 0 : if( FD_UNLIKELY( sched->staged_bitset==fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) ) ) {
1291 : /* No more staging lanes available. All of them are occupied by
1292 : slow squatters. Demote one of them. */
1293 : //FIXME implement this
1294 0 : FD_LOG_CRIT(( "unimplemented" ));
1295 0 : sched->metrics->lane_demoted_cnt++;
1296 : // sched->metrics->block_demoted_cnt++; for every demoted block
1297 0 : }
1298 0 : FD_TEST( sched->staged_bitset!=fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) );
1299 0 : int lane_idx = fd_ulong_find_lsb( ~sched->staged_bitset );
1300 0 : if( FD_UNLIKELY( lane_idx>=(int)FD_SCHED_MAX_STAGING_LANES ) ) {
1301 0 : FD_LOG_CRIT(( "invariant violation: lane_idx %d, sched->staged_bitset %lx",
1302 0 : lane_idx, sched->staged_bitset ));
1303 0 : }
1304 0 : ulong head_bank_idx = stage_longest_unstaged_fork( sched, root_idx, lane_idx );
1305 0 : if( FD_UNLIKELY( head_bank_idx==ULONG_MAX ) ) {
1306 : /* We found a promotable fork depth>0. This should not happen. */
1307 0 : FD_LOG_CRIT(( "invariant violation: head_bank_idx==ULONG_MAX" ));
1308 0 : }
1309 : /* We don't bother with promotion unless the block is immediately
1310 : dispatchable. So it's okay to set the active block here. */
1311 0 : sched->active_bank_idx = head_bank_idx;
1312 0 : return;
1313 0 : }
1314 : /* No unstaged blocks to promote. So we're done. Yay. */
1315 0 : }
1316 :
1317 : static void
1318 0 : check_or_set_active_block( fd_sched_t * sched ) {
1319 0 : if( FD_UNLIKELY( sched->active_bank_idx==ULONG_MAX ) ) {
1320 0 : try_activate_block( sched );
1321 0 : } else {
1322 0 : fd_sched_block_t * active_block = block_pool_ele( sched, sched->active_bank_idx );
1323 0 : if( FD_UNLIKELY( block_should_deactivate( active_block ) ) ) {
1324 0 : print_all( sched, active_block );
1325 0 : FD_LOG_CRIT(( "invariant violation: should have been deactivated" ));
1326 0 : }
1327 0 : }
1328 0 : }
1329 :
1330 : /* It's safe to call this function more than once on the same block. */
1331 : static void
1332 0 : subtree_abandon( fd_sched_t * sched, fd_sched_block_t * block ) {
1333 0 : if( FD_UNLIKELY( block->rooted ) ) {
1334 0 : FD_LOG_CRIT(( "invariant violation: rooted block should not be abandoned, slot %lu, parent slot %lu",
1335 0 : block->slot, block->parent_slot ));
1336 0 : }
1337 : /* All minority fork nodes pass through this function eventually. So
1338 : this is a good point to check per-node invariants for minority
1339 : forks. */
1340 0 : if( FD_UNLIKELY( block->staged && !block->in_rdisp ) ) {
1341 0 : FD_LOG_CRIT(( "invariant violation: staged block is not in the dispatcher, slot %lu, parent slot %lu",
1342 0 : block->slot, block->parent_slot ));
1343 0 : }
1344 :
1345 : /* Setting the flag is non-optional and can happen more than once. */
1346 0 : block->dying = 1;
1347 :
1348 : /* Removal from dispatcher should only happen once. */
1349 0 : if( block->in_rdisp ) {
1350 0 : fd_sched_block_t * parent = block_pool_ele( sched, block->parent_idx );
1351 0 : if( FD_UNLIKELY( !parent ) ) {
1352 : /* Only the root has no parent. Abandon should never be called on
1353 : the root. So any block we are trying to abandon should have a
1354 : parent. */
1355 0 : FD_LOG_CRIT(( "invariant violation: parent not found slot %lu, parent slot %lu",
1356 0 : block->slot, block->parent_slot ));
1357 0 : }
1358 :
1359 : /* The dispatcher expects blocks to be abandoned in the same order
1360 : that they were added on each lane. There are no requirements on
1361 : the order of abandoning if two blocks are not on the same lane,
1362 : or if a block is unstaged. This means that in general we
1363 : shouldn't abandon a child block if the parent hasn't been
1364 : abandoned yet, if and only if they are on the same lane. So wait
1365 : until we can abandon the parent, and then descend down the fork
1366 : tree to ensure orderly abandoning. */
1367 0 : int in_order = !parent->in_rdisp || /* parent is not in the dispatcher */
1368 0 : !parent->staged || /* parent is in the dispatcher but not staged */
1369 0 : !block->staged || /* parent is in the dispatcher and staged but this block is unstaged */
1370 0 : block->staging_lane!=parent->staging_lane; /* this block is on a different staging lane than its parent */
1371 :
1372 : /* We inform the dispatcher of an abandon only when there are no
1373 : more in-flight transactions. Otherwise, if the dispatcher
1374 : recycles the same txn_id that was just abandoned, and we receive
1375 : completion of an in-flight transaction whose txn_id was just
1376 : recycled, we would basically be aliasing the same txn_id and end
1377 : up indexing into txn_to_bank_idx[] that is already overwritten
1378 : with new blocks. */
1379 0 : int abandon = in_order && block->txn_exec_in_flight_cnt==0 && block->txn_sigverify_in_flight_cnt==0;
1380 :
1381 0 : if( abandon ) {
1382 0 : block->in_rdisp = 0;
1383 0 : fd_rdisp_abandon_block( sched->rdisp, (ulong)(block-sched->block_pool) );
1384 0 : sched->txn_pool_free_cnt += block->txn_parsed_cnt-block->txn_done_cnt; /* in_flight_cnt==0 */
1385 0 : sched->metrics->block_abandoned_cnt++;
1386 0 : sched->metrics->txn_abandoned_parsed_cnt += block->txn_parsed_cnt;
1387 0 : sched->metrics->txn_abandoned_exec_done_cnt += block->txn_exec_done_cnt;
1388 0 : sched->metrics->txn_abandoned_done_cnt += block->txn_done_cnt;
1389 :
1390 : /* Now release the staging lane. */
1391 : //FIXME when demote supports non-empty blocks, we should demote
1392 : //the block from the lane unconditionally and immediately,
1393 : //regardles of whether it's safe to abandon or not. So a block
1394 : //would go immediately from staged to unstaged and eventually to
1395 : //abandoned.
1396 0 : if( FD_LIKELY( block->staged ) ) {
1397 0 : block->staged = 0;
1398 0 : sched->staged_bitset = fd_ulong_clear_bit( sched->staged_bitset, (int)block->staging_lane );
1399 0 : sched->staged_head_bank_idx[ block->staging_lane ] = ULONG_MAX;
1400 0 : }
1401 0 : }
1402 :
1403 0 : if( FD_UNLIKELY( in_order && block->staged && sched->active_bank_idx==sched->staged_head_bank_idx[ block->staging_lane ] ) ) {
1404 0 : FD_TEST( block_pool_ele( sched, sched->active_bank_idx )==block );
1405 : /* Dying blocks should not be active. */
1406 0 : sched->active_bank_idx = ULONG_MAX;
1407 0 : }
1408 0 : }
1409 :
1410 : /* Abandon the entire fork chaining off of this block. */
1411 0 : ulong child_idx = block->child_idx;
1412 0 : while( child_idx!=ULONG_MAX ) {
1413 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1414 0 : subtree_abandon( sched, child );
1415 0 : child_idx = child->sibling_idx;
1416 0 : }
1417 0 : }
1418 :
1419 : static void
1420 0 : maybe_switch_block( fd_sched_t * sched, ulong bank_idx ) {
1421 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1422 0 : if( FD_UNLIKELY( block_is_done( block ) ) ) {
1423 0 : block->in_rdisp = 0;
1424 0 : block->staged = 0;
1425 0 : fd_rdisp_remove_block( sched->rdisp, bank_idx );
1426 0 : sched->metrics->block_removed_cnt++;
1427 :
1428 : /* See if there is a child block down the same staging lane. This
1429 : is a policy decision to minimize fork churn. We could in theory
1430 : reevaluate staging lane allocation here and do promotion/demotion
1431 : as needed. */
1432 0 : ulong child_idx = block->child_idx;
1433 0 : while( child_idx!=ULONG_MAX ) {
1434 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1435 0 : if( FD_LIKELY( child->staged && child->staging_lane==block->staging_lane ) ) {
1436 : /* There is a child block down the same staging lane ... */
1437 0 : if( FD_LIKELY( !child->dying ) ) {
1438 : /* ... and the child isn't dead */
1439 0 : if( FD_UNLIKELY( !block_is_activatable( child ) ) ) {
1440 : /* ... but the child is not activatable, likely because
1441 : there are no transactions available yet. */
1442 0 : sched->active_bank_idx = ULONG_MAX;
1443 0 : sched->metrics->deactivate_no_txn_cnt++;
1444 0 : try_activate_block( sched );
1445 0 : return;
1446 0 : }
1447 : /* ... and it's immediately dispatchable, so switch the active
1448 : block to it, and have the child inherit the head status of
1449 : the lane. This is the common case. */
1450 0 : sched->active_bank_idx = child_idx;
1451 0 : sched->staged_head_bank_idx[ block->staging_lane ] = child_idx;
1452 0 : if( FD_UNLIKELY( !fd_ulong_extract_bit( sched->staged_bitset, (int)block->staging_lane ) ) ) {
1453 0 : FD_LOG_CRIT(( "invariant violation: staged_bitset 0x%lx bit %lu is not set, slot %lu, parent slot %lu, child slot %lu, parent slot %lu",
1454 0 : sched->staged_bitset, block->staging_lane, block->slot, block->parent_slot, child->slot, child->parent_slot ));
1455 0 : }
1456 0 : return;
1457 0 : } else {
1458 : /* ... but the child block is considered dead, likely because
1459 : the parser considers it invalid. */
1460 0 : subtree_abandon( sched, child );
1461 0 : break;
1462 0 : }
1463 0 : }
1464 0 : child_idx = child->sibling_idx;
1465 0 : }
1466 : /* There isn't a child block down the same staging lane. This is
1467 : the last block in the staging lane. Release the staging lane. */
1468 0 : sched->staged_bitset = fd_ulong_clear_bit( sched->staged_bitset, (int)block->staging_lane );
1469 0 : sched->staged_head_bank_idx[ block->staging_lane ] = ULONG_MAX;
1470 :
1471 : /* Reset the active block. */
1472 0 : sched->active_bank_idx = ULONG_MAX;
1473 0 : sched->metrics->deactivate_no_child_cnt++;
1474 0 : try_activate_block( sched );
1475 0 : } else if( block_should_deactivate( block ) ) {
1476 : /* We exhausted the active block, but it's not fully done yet. We
1477 : are just not getting FEC sets for it fast enough. This could
1478 : happen when the network path is congested, or when the leader
1479 : simply went down. Reset the active block. */
1480 0 : sched->active_bank_idx = ULONG_MAX;
1481 0 : sched->metrics->deactivate_no_txn_cnt++;
1482 0 : try_activate_block( sched );
1483 0 : }
1484 0 : }
1485 :
1486 : FD_FN_UNUSED static ulong
1487 0 : find_and_stage_longest_unstaged_fork( fd_sched_t * sched, int lane_idx ) {
1488 0 : ulong root_idx = sched->root_idx;
1489 0 :
1490 0 : if( FD_UNLIKELY( root_idx==ULONG_MAX ) ) {
1491 0 : FD_LOG_CRIT(( "invariant violation: root_idx==ULONG_MAX indicating fd_sched is unintialized" ));
1492 0 : }
1493 0 :
1494 0 : /* First pass: compute the longest unstaged fork depth for each node
1495 0 : in the fork tree. */
1496 0 : ulong depth = compute_longest_unstaged_fork( sched, root_idx );
1497 0 :
1498 0 : /* Second pass: stage blocks on the longest unstaged fork. */
1499 0 : ulong head_bank_idx = stage_longest_unstaged_fork( sched, root_idx, lane_idx );
1500 0 :
1501 0 : if( FD_UNLIKELY( (depth>0UL && head_bank_idx==ULONG_MAX) || (depth==0UL && head_bank_idx!=ULONG_MAX) ) ) {
1502 0 : FD_LOG_CRIT(( "invariant violation: depth %lu, head_bank_idx %lu",
1503 0 : depth, head_bank_idx ));
1504 0 : }
1505 0 :
1506 0 : return head_bank_idx;
1507 0 : }
1508 :
1509 : /* Returns length of the longest stageable unstaged fork, if there is
1510 : one, and 0 otherwise. */
1511 : static ulong
1512 0 : compute_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx ) {
1513 0 : if( FD_UNLIKELY( bank_idx==ULONG_MAX ) ) {
1514 0 : FD_LOG_CRIT(( "invariant violation: bank_idx==ULONG_MAX" ));
1515 0 : }
1516 :
1517 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1518 :
1519 0 : ulong max_child_depth = 0UL;
1520 0 : ulong child_idx = block->child_idx;
1521 0 : while( child_idx!=ULONG_MAX ) {
1522 0 : ulong child_depth = compute_longest_unstaged_fork( sched, child_idx );
1523 0 : if( child_depth > max_child_depth ) {
1524 0 : max_child_depth = child_depth;
1525 0 : }
1526 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1527 0 : child_idx = child->sibling_idx;
1528 0 : }
1529 :
1530 0 : block->luf_depth = max_child_depth + fd_ulong_if( block_is_promotable( block ), 1UL, 0UL );
1531 0 : return block->luf_depth;
1532 0 : }
1533 :
1534 : static ulong
1535 0 : stage_longest_unstaged_fork_helper( fd_sched_t * sched, ulong bank_idx, int lane_idx ) {
1536 0 : if( FD_UNLIKELY( bank_idx==ULONG_MAX ) ) {
1537 0 : FD_LOG_CRIT(( "invariant violation: bank_idx==ULONG_MAX" ));
1538 0 : }
1539 :
1540 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1541 :
1542 0 : int stage_it = fd_int_if( block_is_promotable( block ), 1, 0 );
1543 0 : ulong rv = fd_ulong_if( stage_it, bank_idx, ULONG_MAX );
1544 0 : if( FD_LIKELY( stage_it ) ) {
1545 0 : block->staged = 1;
1546 0 : block->staging_lane = (ulong)lane_idx;
1547 0 : fd_rdisp_promote_block( sched->rdisp, bank_idx, block->staging_lane );
1548 0 : sched->metrics->block_promoted_cnt++;
1549 0 : }
1550 :
1551 : /* Base case: leaf node. */
1552 0 : if( block->child_idx==ULONG_MAX ) return rv;
1553 :
1554 0 : ulong max_depth = 0UL;
1555 0 : ulong best_child_idx = ULONG_MAX;
1556 0 : ulong child_idx = block->child_idx;
1557 0 : while( child_idx!=ULONG_MAX ) {
1558 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1559 0 : if( child->luf_depth>max_depth ) {
1560 0 : max_depth = child->luf_depth;
1561 0 : best_child_idx = child_idx;
1562 0 : }
1563 0 : child_idx = child->sibling_idx;
1564 0 : }
1565 :
1566 : /* Recursively stage descendants. */
1567 0 : if( best_child_idx!=ULONG_MAX ) {
1568 0 : ulong head_bank_idx = stage_longest_unstaged_fork_helper( sched, best_child_idx, lane_idx );
1569 0 : rv = fd_ulong_if( rv!=ULONG_MAX, rv, head_bank_idx );
1570 0 : }
1571 :
1572 0 : return rv;
1573 0 : }
1574 :
1575 : /* Returns idx of head block of staged lane on success, idx_null
1576 : otherwise. */
1577 : static ulong
1578 0 : stage_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx, int lane_idx ) {
1579 0 : ulong head_bank_idx = stage_longest_unstaged_fork_helper( sched, bank_idx, lane_idx );
1580 0 : if( FD_LIKELY( head_bank_idx!=ULONG_MAX ) ) {
1581 0 : sched->metrics->lane_promoted_cnt++;
1582 0 : sched->staged_bitset = fd_ulong_set_bit( sched->staged_bitset, lane_idx );
1583 0 : sched->staged_head_bank_idx[ lane_idx ] = head_bank_idx;
1584 0 : }
1585 0 : return head_bank_idx;
1586 0 : }
|
