Line data Source code
1 : #include <stdio.h> /* for vsnprintf */
2 : #include <stdarg.h> /* for va_list */
3 :
4 : #include "fd_sched.h"
5 : #include "fd_execrp.h" /* for poh hash value */
6 : #include "../../util/math/fd_stat.h" /* for sorted search */
7 : #include "../../disco/fd_disco_base.h" /* for FD_MAX_TXN_PER_SLOT */
8 : #include "../../disco/fd_txn_p.h"
9 : #include "../../disco/metrics/fd_metrics.h" /* for fd_metrics_convert_seconds_to_ticks and etc. */
10 : #include "../../disco/pack/fd_chkdup.h"
11 : #include "../../disco/shred/fd_shredder.h" /* FD_SHREDDER_CHAINED_FEC_SET_PAYLOAD_SZ */
12 : #include "../../discof/poh/fd_poh.h" /* for MAX_SKIPPED_TICKS */
13 : #include "../../flamenco/runtime/fd_runtime.h" /* for fd_runtime_load_txn_address_lookup_tables */
14 : #include "../../flamenco/runtime/fd_system_ids.h"
15 : #include "../../flamenco/runtime/sysvar/fd_sysvar_slot_hashes.h" /* for ALUTs */
16 :
17 105 : #define FD_SCHED_MAX_STAGING_LANES_LOG (2)
18 105 : #define FD_SCHED_MAX_STAGING_LANES (1UL<<FD_SCHED_MAX_STAGING_LANES_LOG)
19 : #define FD_SCHED_MAX_EXEC_TILE_CNT (64UL)
20 186 : #define FD_SCHED_MAX_PRINT_BUF_SZ (2UL<<20)
21 :
22 : #define FD_SCHED_MAX_MBLK_PER_SLOT (MAX_SKIPPED_TICKS)
23 9 : #define FD_SCHED_MAX_POH_HASHES_PER_TASK (4096UL) /* This seems to be the sweet spot. */
24 :
25 : /* 64 ticks per slot, and a single gigantic microblock containing min
26 : size transactions. */
27 : 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 );
28 :
29 : /* We size the buffer to be able to hold residual data from the previous
30 : FEC set that only becomes parseable after the next FEC set is
31 : ingested, as well as the incoming FEC set. The largest minimally
32 : parseable unit of data is a transaction. So that much data may
33 : straddle FEC set boundaries. Other minimally parseable units of data
34 : include the microblock header and the microblock count within a
35 : batch. */
36 6 : #define FD_SCHED_MAX_PAYLOAD_PER_FEC (63985UL)
37 : FD_STATIC_ASSERT( FD_SCHED_MAX_PAYLOAD_PER_FEC>=FD_SHREDDER_CHAINED_FEC_SET_PAYLOAD_SZ, bump sched fec size bound ); /* FD_SHREDDER_CHAINED_FEC_SET_PAYLOAD_SZ is the bound we want, but older ledgers have larger FEC sets. */
38 : #define FD_SCHED_MAX_FEC_BUF_SZ (FD_SCHED_MAX_PAYLOAD_PER_FEC+FD_TXN_MTU)
39 : FD_STATIC_ASSERT( FD_TXN_MTU>=sizeof(fd_microblock_hdr_t), resize buffer for residual data );
40 : FD_STATIC_ASSERT( FD_TXN_MTU>=sizeof(ulong), resize buffer for residual data );
41 :
42 3 : #define FD_SCHED_MAX_TXN_PER_FEC ((FD_SCHED_MAX_PAYLOAD_PER_FEC-1UL)/FD_TXN_MIN_SERIALIZED_SZ+1UL) /* 478 */
43 3 : #define FD_SCHED_MAX_MBLK_PER_FEC ((FD_SCHED_MAX_PAYLOAD_PER_FEC-1UL)/sizeof(fd_microblock_hdr_t)+1UL) /* 1334 */
44 :
45 : FD_STATIC_ASSERT( FD_SCHED_MIN_DEPTH>=FD_SCHED_MAX_TXN_PER_FEC, limits );
46 : FD_STATIC_ASSERT( FD_SCHED_MAX_DEPTH<=FD_RDISP_MAX_DEPTH, limits );
47 :
48 12 : #define FD_SCHED_MAGIC (0xace8a79c181f89b6UL) /* echo -n "fd_sched_v0" | sha512sum | head -c 16 */
49 :
50 9 : #define FD_SCHED_OK (0)
51 15 : #define FD_SCHED_AGAIN_LATER (1)
52 0 : #define FD_SCHED_BAD_BLOCK (2)
53 :
54 :
55 : /* Structs. */
56 :
57 : struct fd_sched_mblk {
58 : ulong start_txn_idx; /* inclusive parse idx */
59 : ulong end_txn_idx; /* non-inclusive parse idx */
60 : ulong curr_txn_idx; /* next txn to mixin, parse idx */
61 : ulong hashcnt; /* number of pure hashes, excluding final mixin */
62 : ulong curr_hashcnt;
63 : fd_hash_t end_hash[ 1 ];
64 : fd_hash_t curr_hash[ 1 ];
65 : uint curr_sig_cnt;
66 : uint next;
67 : int is_tick;
68 : };
69 : typedef struct fd_sched_mblk fd_sched_mblk_t;
70 :
71 : #define SLIST_NAME mblk_slist
72 : #define SLIST_ELE_T fd_sched_mblk_t
73 9 : #define SLIST_IDX_T uint
74 24 : #define SLIST_NEXT next
75 : #include "../../util/tmpl/fd_slist.c"
76 :
77 : #define SET_NAME txn_bitset
78 : #define SET_MAX FD_SCHED_MAX_DEPTH
79 : #include "../../util/tmpl/fd_set.c"
80 :
81 : struct fd_sched_block {
82 : ulong slot;
83 : ulong parent_slot;
84 : ulong parent_idx; /* Index of the parent in the pool. */
85 : ulong child_idx; /* Index of the left-child in the pool. */
86 : ulong sibling_idx; /* Index of the right-sibling in the pool. */
87 :
88 : /* Counters. */
89 : uint txn_parsed_cnt;
90 : /* txn_queued_cnt = txn_parsed_cnt-txn_in_flight_cnt-txn_done_cnt */
91 : uint txn_exec_in_flight_cnt;
92 : uint txn_exec_done_cnt;
93 : uint txn_sigverify_in_flight_cnt;
94 : uint txn_sigverify_done_cnt;
95 : uint poh_hashing_in_flight_cnt;
96 : uint poh_hashing_done_cnt;
97 : uint poh_hash_cmp_done_cnt; /* poh_hashing_done_cnt==poh_hash_cmp_done_cnt+len(mixin_in_progress) */
98 : uint txn_done_cnt; /* A transaction is considered done when all types of tasks associated with it are done. */
99 : uint shred_cnt;
100 : uint mblk_cnt; /* Total number of microblocks, including ticks and non ticks.
101 : mblk_cnt==len(unhashed)+len(hashing_in_progress)+hashing_in_flight_cnt+len(mixin_in_progress)+hash_cmp_done_cnt */
102 : uint mblk_tick_cnt; /* Total number of tick microblocks. */
103 : uint mblk_freed_cnt; /* This is ==hash_cmp_done_cnt in most cases, except for aborted
104 : blocks, where the freed cnt will catch up to mblk_cnt and surpass
105 : hash_cmp_done_cnt when the block is reaped. */
106 : uint mblk_unhashed_cnt; /* ==len(unhashed) */
107 : ulong hashcnt; /* How many hashes this block wants replay to do. A mixin/record counts as one hash. */
108 : ulong txn_pool_max_popcnt; /* Peak transaction pool occupancy during the time this block was replaying. */
109 : ulong mblk_pool_max_popcnt; /* Peak mblk pool occupancy. */
110 : ulong block_pool_max_popcnt; /* Peak block pool occupancy. */
111 : ulong txn_idx[ FD_MAX_TXN_PER_SLOT ]; /* Indexed by parse order. */
112 :
113 : /* PoH verify. */
114 : fd_hash_t poh_hash[ 1 ]; /* running end_hash of last parsed mblk */
115 : int last_mblk_is_tick;
116 : mblk_slist_t mblks_unhashed[ 1 ]; /* A microblock, once parsed out, is in one of these queues. It
117 : generally progresses from unhashed to hashing to mixin. When a
118 : microblock is being hashed/in-flight, it'll be transiently out of
119 : any of the queues. Once a microblock progresses through all stages
120 : of work, it'll be immediately freed. */
121 : mblk_slist_t mblks_hashing_in_progress[ 1 ];
122 : mblk_slist_t mblks_mixin_in_progress[ 1 ];
123 : uchar bmtree_mem[ FD_BMTREE_COMMIT_FOOTPRINT(0) ] __attribute__((aligned(FD_BMTREE_COMMIT_ALIGN)));
124 : fd_bmtree_commit_t * bmtree;
125 : ulong tick_hashcnt_wmk; /* All ticks in a valid block must accumulate the same number of
126 : hashes since the previous tick, or since block start, and this hash
127 : count must match hashes_per_tick for the block. */
128 : ulong curr_tick_hashcnt; /* Starts at 0, accumulates hashcnt, resets to 0 on the next tick. */
129 : ulong tick_height; /* Block is built off of a parent block with this many ticks. */
130 : ulong max_tick_height; /* Block should end with precisely this many ticks. */
131 : ulong hashes_per_tick; /* Fixed per block, feature gated, known after bank clone. */
132 : int inconsistent_hashes_per_tick;
133 : int zero_hash_tick;
134 :
135 : /* Parser state. */
136 : uchar txn[ FD_TXN_MAX_SZ ] __attribute__((aligned(alignof(fd_txn_t))));
137 : ulong mblks_rem; /* Number of microblocks remaining in the current batch. */
138 : ulong txns_rem; /* Number of transactions remaining in the current microblock. */
139 : uint fec_buf_sz; /* Size of the fec_buf in bytes. */
140 : uint fec_buf_soff; /* Starting offset into fec_buf for unparsed transactions. */
141 : uint fec_buf_boff; /* Byte offset into raw block data of the first byte currently in fec_buf */
142 : uint fec_eob:1; /* FEC end-of-batch: set if the last FEC set in the batch is being
143 : ingested. */
144 : uint fec_sob:1; /* FEC start-of-batch: set if the parser expects to be parsing out a
145 : batch header. */
146 :
147 : /* Block state. */
148 : uint fec_eos:1; /* FEC end-of-stream: set if the last FEC set in the block has been
149 : ingested. */
150 : uint rooted:1; /* Set if the block is rooted. */
151 : uint dying:1; /* Set if the block has been abandoned and no transactions should be
152 : scheduled from it. */
153 : uint refcnt:1; /* Starts at 1 when the block is added, set to 0 if caller has been
154 : informed to decrement refcnt for sched. */
155 : uint in_sched:1; /* Set if the block is being tracked by the scheduler. */
156 : uint in_rdisp:1; /* Set if the block is being tracked by the dispatcher, either as staged
157 : or unstaged. */
158 : uint block_start_signaled:1; /* Set if the start-of-block sentinel has been dispatched. */
159 : uint block_end_signaled:1; /* Set if the end-of-block sentinel has been dispatched. */
160 : uint block_start_done:1; /* Set if the start-of-block processing has been completed. */
161 : uint block_end_done:1; /* Set if the end-of-block processing has been completed. */
162 : uint staged:1; /* Set if the block is in a dispatcher staging lane; a staged block is
163 : tracked by the dispatcher. */
164 : ulong staging_lane; /* Ignored if staged==0. */
165 : ulong luf_depth; /* Depth of longest unstaged fork starting from this node; only
166 : stageable unstaged descendants are counted. */
167 : uchar fec_buf[ FD_SCHED_MAX_FEC_BUF_SZ ]; /* The previous FEC set could have some residual data that only becomes
168 : parseable after the next FEC set is ingested. */
169 : uint shred_blk_offs[ FD_SHRED_BLK_MAX ]; /* The byte offsets into block data of ingested shreds */
170 : };
171 : typedef struct fd_sched_block fd_sched_block_t;
172 :
173 : FD_STATIC_ASSERT( sizeof(fd_hash_t)==sizeof(((fd_microblock_hdr_t *)0)->hash), unexpected poh hash size );
174 :
175 :
176 : struct fd_sched_metrics {
177 : uint block_added_cnt;
178 : uint block_added_staged_cnt;
179 : uint block_added_unstaged_cnt;
180 : uint block_added_dead_ood_cnt;
181 : uint block_removed_cnt;
182 : uint block_abandoned_cnt;
183 : uint block_bad_cnt;
184 : uint block_promoted_cnt;
185 : uint block_demoted_cnt;
186 : uint deactivate_no_child_cnt;
187 : uint deactivate_no_txn_cnt;
188 : uint deactivate_pruned_cnt;
189 : uint deactivate_abandoned_cnt;
190 : uint lane_switch_cnt;
191 : uint lane_promoted_cnt;
192 : uint lane_demoted_cnt;
193 : uint fork_observed_cnt;
194 : uint alut_success_cnt;
195 : uint alut_serializing_cnt;
196 : uint txn_abandoned_parsed_cnt;
197 : uint txn_abandoned_exec_done_cnt;
198 : uint txn_abandoned_done_cnt;
199 : uint txn_max_in_flight_cnt;
200 : ulong txn_weighted_in_flight_cnt;
201 : ulong txn_weighted_in_flight_tickcount;
202 : ulong txn_none_in_flight_tickcount;
203 : ulong txn_parsed_cnt;
204 : ulong txn_exec_done_cnt;
205 : ulong txn_sigverify_done_cnt;
206 : ulong txn_mixin_done_cnt;
207 : ulong txn_done_cnt;
208 : ulong mblk_parsed_cnt;
209 : ulong mblk_poh_hashed_cnt;
210 : ulong mblk_poh_done_cnt;
211 : ulong bytes_ingested_cnt;
212 : ulong bytes_ingested_unparsed_cnt;
213 : ulong bytes_dropped_cnt;
214 : ulong fec_cnt;
215 : };
216 : typedef struct fd_sched_metrics fd_sched_metrics_t;
217 :
218 : #define DEQUE_NAME ref_q
219 9 : #define DEQUE_T ulong
220 : #include "../../util/tmpl/fd_deque_dynamic.c"
221 :
222 : struct fd_sched {
223 : fd_acct_addr_t aluts[ 256 ]; /* Resolve ALUT accounts into this buffer for more parallelism. */
224 : char print_buf[ FD_SCHED_MAX_PRINT_BUF_SZ ];
225 : ulong print_buf_sz;
226 : fd_chkdup_t chkdup[ 1 ];
227 : fd_sched_metrics_t metrics[ 1 ];
228 : ulong canary; /* == FD_SCHED_MAGIC */
229 : ulong depth; /* Immutable. */
230 : ulong block_cnt_max; /* Immutable. */
231 : ulong exec_cnt; /* Immutable. */
232 : int bypass_poh_verify; /* Test/fuzz: skip the PoH end_hash compare in maybe_mixin. */
233 : int bypass_alut_resolution; /* Test/fuzz: skip ALUT resolution (no accdb). */
234 : long txn_in_flight_last_tick;
235 : long next_ready_last_tick;
236 : ulong next_ready_last_bank_idx;
237 : ulong root_idx;
238 : fd_rdisp_t * rdisp;
239 : ulong txn_exec_ready_bitset[ 1 ];
240 : ulong sigverify_ready_bitset[ 1 ];
241 : ulong poh_ready_bitset[ 1 ];
242 : ulong active_bank_idx; /* Index of the actively replayed block, or ULONG_MAX if no block is
243 : actively replayed; has to have a transaction to dispatch; staged
244 : blocks that have no transactions to dispatch are not eligible for
245 : being active. */
246 : ulong last_active_bank_idx;
247 : ulong staged_bitset; /* Bit i set if staging lane i is occupied. */
248 : ulong staged_head_bank_idx[ FD_SCHED_MAX_STAGING_LANES ]; /* Head of the linear chain in each staging lane, ignored if bit i is
249 : not set in the bitset. */
250 : ulong staged_popcnt_wmk;
251 : ulong txn_pool_free_cnt;
252 : fd_txn_p_t * txn_pool; /* Just a flat array. */
253 : fd_sched_txn_info_t * txn_info_pool; /* Just a flat array. */
254 : fd_sched_mblk_t * mblk_pool; /* Just a flat array. */
255 : ulong mblk_pool_free_cnt;
256 : uint mblk_pool_free_head;
257 : ulong tile_to_bank_idx[ FD_SCHED_MAX_EXEC_TILE_CNT ]; /* Index of the bank that the exec tile is executing against. */
258 : txn_bitset_t exec_done_set[ txn_bitset_word_cnt ]; /* Indexed by txn_idx. */
259 : txn_bitset_t sigverify_done_set[ txn_bitset_word_cnt ]; /* Indexed by txn_idx. */
260 : txn_bitset_t poh_mixin_done_set[ txn_bitset_word_cnt ]; /* Indexed by txn_idx. */
261 : fd_sched_block_t * block_pool; /* Just a flat array. */
262 : ulong block_pool_popcnt;
263 : ulong * ref_q;
264 : };
265 : typedef struct fd_sched fd_sched_t;
266 :
267 :
268 : /* Internal helpers. */
269 :
270 : static int
271 : verify_ticks_eager( fd_sched_block_t * block );
272 :
273 : static int
274 : verify_ticks_final( fd_sched_block_t * block );
275 :
276 : static void
277 : add_block( fd_sched_t * sched,
278 : ulong bank_idx,
279 : ulong parent_bank_idx );
280 :
281 : FD_WARN_UNUSED static int
282 : fd_sched_parse( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx );
283 :
284 : FD_WARN_UNUSED static int
285 : fd_sched_parse_txn( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx );
286 :
287 : static void
288 : dispatch_sigverify( fd_sched_t * sched, fd_sched_block_t * block, ulong bank_idx, int exec_tile_idx, fd_sched_task_t * out );
289 :
290 : static void
291 : dispatch_poh( fd_sched_t * sched, fd_sched_block_t * block, ulong bank_idx, int exec_tile_idx, fd_sched_task_t * out );
292 :
293 : FD_WARN_UNUSED static int
294 : maybe_mixin( fd_sched_t * sched, fd_sched_block_t * block );
295 :
296 : static void
297 9 : free_mblk( fd_sched_t * sched, fd_sched_block_t * block, uint mblk_idx ) {
298 9 : sched->mblk_pool[ mblk_idx ].next = sched->mblk_pool_free_head;
299 9 : sched->mblk_pool_free_head = mblk_idx;
300 9 : sched->mblk_pool_free_cnt++;
301 9 : block->mblk_freed_cnt++;
302 9 : }
303 :
304 : static void
305 0 : free_mblk_slist( fd_sched_t * sched, fd_sched_block_t * block, mblk_slist_t * list ) {
306 0 : while( !mblk_slist_is_empty( list, sched->mblk_pool ) ) {
307 0 : uint idx = (uint)mblk_slist_idx_pop_head( list, sched->mblk_pool );
308 0 : free_mblk( sched, block, idx );
309 0 : }
310 0 : }
311 :
312 : static void
313 : try_activate_block( fd_sched_t * sched );
314 :
315 : static void
316 : check_or_set_active_block( fd_sched_t * sched );
317 :
318 : static void
319 : subtree_abandon( fd_sched_t * sched, fd_sched_block_t * block );
320 :
321 : static void
322 : subtree_release_refcnt( fd_sched_t * sched, fd_sched_block_t * block );
323 :
324 : static void
325 : subtree_prune( fd_sched_t * sched, ulong bank_idx, ulong except_idx );
326 :
327 : static void
328 : maybe_switch_block( fd_sched_t * sched, ulong bank_idx );
329 :
330 : FD_FN_UNUSED static ulong
331 : find_and_stage_longest_unstaged_fork( fd_sched_t * sched, int lane_idx );
332 :
333 : static ulong
334 : compute_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx );
335 :
336 : static ulong
337 : stage_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx, int lane_idx );
338 :
339 : static int
340 : lane_is_demotable( fd_sched_t * sched, int lane_idx );
341 :
342 : static ulong
343 : demote_lane( fd_sched_t * sched, int lane_idx );
344 :
345 : static inline fd_sched_block_t *
346 747 : block_pool_ele( fd_sched_t * sched, ulong idx ) {
347 747 : FD_TEST( idx<sched->block_cnt_max || idx==ULONG_MAX );
348 747 : return idx==ULONG_MAX ? NULL : sched->block_pool+idx;
349 747 : }
350 :
351 : FD_FN_UNUSED static inline int
352 0 : block_is_void( fd_sched_block_t * block ) {
353 0 : /* We've seen everything in the block and no transaction got parsed
354 0 : out. */
355 0 : return block->fec_eos && block->txn_parsed_cnt==0;
356 0 : }
357 :
358 : static inline int
359 3 : block_should_signal_end( fd_sched_block_t * block ) {
360 : /* Under the current policy of eager synchronous PoH mixin, hashing
361 : done plus fec_eos imply that all mixins have been done. */
362 3 : if( FD_UNLIKELY( !( !block->fec_eos || ((block->mblk_cnt==block->poh_hashing_done_cnt&&block->mblk_cnt==block->poh_hash_cmp_done_cnt)||block->mblk_cnt!=block->poh_hashing_done_cnt) ) ) ) FD_LOG_CRIT(( "invariant violation: slot %lu fec_eos %d mblk_cnt %u poh_hashing_done_cnt %u poh_hash_cmp_done_cnt %u", block->slot, block->fec_eos, block->mblk_cnt, block->poh_hashing_done_cnt, block->poh_hash_cmp_done_cnt ));
363 3 : return block->fec_eos && block->txn_parsed_cnt==block->txn_done_cnt && block->mblk_cnt==block->poh_hashing_done_cnt && block->block_start_done && !block->block_end_signaled;
364 3 : }
365 :
366 : static inline int
367 153 : block_will_signal_end( fd_sched_block_t * block ) {
368 153 : return block->fec_eos && !block->block_end_signaled;
369 153 : }
370 :
371 : /* Is there something known to be dispatchable in the block? This is an
372 : important liveness property. A block that doesn't contain any known
373 : dispatchable tasks will be deactivated or demoted. */
374 : static inline int
375 222 : block_is_dispatchable( fd_sched_block_t * block ) {
376 222 : ulong exec_queued_cnt = block->txn_parsed_cnt-block->txn_exec_in_flight_cnt-block->txn_exec_done_cnt;
377 222 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
378 222 : ulong poh_queued_cnt = block->mblk_cnt-block->poh_hashing_in_flight_cnt-block->poh_hashing_done_cnt;
379 222 : return exec_queued_cnt>0UL ||
380 222 : sigverify_queued_cnt>0UL ||
381 222 : poh_queued_cnt>0UL ||
382 222 : !block->block_start_signaled ||
383 222 : block_will_signal_end( block );
384 222 : }
385 :
386 : static inline int
387 45 : block_is_in_flight( fd_sched_block_t * block ) {
388 45 : return block->txn_exec_in_flight_cnt || block->txn_sigverify_in_flight_cnt || block->poh_hashing_in_flight_cnt || (block->block_end_signaled && !block->block_end_done);
389 45 : }
390 :
391 : static inline int
392 384 : block_is_done( fd_sched_block_t * block ) {
393 384 : return block->fec_eos && block->txn_parsed_cnt==block->txn_done_cnt && block->mblk_cnt==block->poh_hash_cmp_done_cnt && block->block_start_done && block->block_end_done;
394 384 : }
395 :
396 : static inline int
397 273 : block_is_stageable( fd_sched_block_t * block ) {
398 273 : int rv = !block_is_done( block ) && !block->dying;
399 273 : if( FD_UNLIKELY( rv && !block->in_rdisp ) ) {
400 : /* Invariant: stageable blocks may be currently staged or unstaged,
401 : but must be in the dispatcher either way. When a block
402 : transitions to DONE, it will be immediately removed from the
403 : dispatcher. When a block transitions to DYING, it will be
404 : eventually abandoned from the dispatcher. */
405 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",
406 0 : block->txn_parsed_cnt, block->txn_done_cnt, (uint)block->fec_eos, block->slot, block->parent_slot ));
407 0 : }
408 273 : return rv;
409 273 : }
410 :
411 : static inline int
412 102 : block_is_promotable( fd_sched_block_t * block ) {
413 102 : return block_is_stageable( block ) && block_is_dispatchable( block ) && !block->staged;
414 102 : }
415 :
416 : static inline int
417 12 : block_is_demotable( fd_sched_block_t * block ) {
418 : /* A block can only be demoted from rdisp if it is empty, meaning no
419 : PENDING, READY, or DISPATCHED transactions. This is equivalent to
420 : having no in-flight transactions (DISPATCHED) and no queued
421 : transactions (PENDING or READY). This function actually implements
422 : a stronger requirement. We consider a block demotable only if
423 : there are no in-flight or queued tasks of any kind. */
424 12 : return !block_is_in_flight( block ) && !block_is_dispatchable( block ) && block->staged;
425 12 : }
426 :
427 : static inline int
428 147 : block_is_activatable( fd_sched_block_t * block ) {
429 147 : return block_is_stageable( block ) && block_is_dispatchable( block ) && block->staged;
430 147 : }
431 :
432 : static inline int
433 63 : block_should_deactivate( fd_sched_block_t * block ) {
434 : /* We allow a grace period, during which a block has nothing to
435 : dispatch, but has something in-flight. The block is allowed to
436 : stay activated and ingest FEC sets during this time. The block
437 : will be deactivated if there's still nothing to dispatch by the
438 : time all in-flight tasks are completed. */
439 63 : return !block_is_activatable( block ) && !block_is_in_flight( block );
440 63 : }
441 :
442 : static inline int
443 9 : block_is_prunable( fd_sched_block_t * block ) {
444 9 : return !block->in_rdisp && !block_is_in_flight( block );
445 9 : }
446 :
447 : static inline ulong
448 84 : block_to_idx( fd_sched_t * sched, fd_sched_block_t * block ) { return (ulong)(block-sched->block_pool); }
449 :
450 : __attribute__((format(printf,2,3)))
451 : static void
452 : fd_sched_printf( fd_sched_t * sched,
453 : char const * fmt,
454 93 : ... ) {
455 93 : va_list ap;
456 93 : ulong len;
457 93 : va_start( ap, fmt );
458 93 : int ret = vsnprintf( sched->print_buf+sched->print_buf_sz,
459 93 : FD_SCHED_MAX_PRINT_BUF_SZ-sched->print_buf_sz,
460 93 : fmt, ap );
461 93 : va_end( ap );
462 93 : len = fd_ulong_if( ret<0, 0UL, fd_ulong_min( (ulong)ret, FD_SCHED_MAX_PRINT_BUF_SZ-sched->print_buf_sz-1UL ) );
463 93 : sched->print_buf[ sched->print_buf_sz+len ] = '\0';
464 93 : sched->print_buf_sz += len;
465 93 : }
466 :
467 : FD_FN_UNUSED static void
468 0 : print_histogram( fd_sched_t * sched, fd_histf_t * hist, ulong converter, char * title ) {
469 0 : fd_sched_printf( sched, " +---------------------+----------------------+--------------+\n" );
470 0 : fd_sched_printf( sched, " | %-19s | | Count |\n", title );
471 0 : fd_sched_printf( sched, " +---------------------+----------------------+--------------+\n" );
472 0 :
473 0 : ulong total_count = 0;
474 0 : for( ulong i=0UL; i<fd_histf_bucket_cnt( hist ); i++ ) {
475 0 : total_count += fd_histf_cnt( hist, i );
476 0 : }
477 0 :
478 0 : for( ulong i=0UL; i< fd_histf_bucket_cnt( hist ); i++ ) {
479 0 : ulong bucket_count = fd_histf_cnt( hist, i );
480 0 :
481 0 : char * lt_str;
482 0 : char lt_buf[ 64 ];
483 0 : if( FD_UNLIKELY( i==fd_histf_bucket_cnt( hist )-1UL ) ) {
484 0 : lt_str = "+Inf";
485 0 : } else {
486 0 : ulong edge = fd_histf_right( hist, i );
487 0 : if( converter==FD_METRICS_CONVERTER_NANOSECONDS ) {
488 0 : edge = fd_metrics_convert_ticks_to_nanoseconds( edge-1UL );
489 0 : FD_TEST( fd_cstr_printf_check( lt_buf, sizeof( lt_buf ), NULL, "<= %lu nanos", edge ) );
490 0 : } else if( converter==FD_METRICS_CONVERTER_NONE ) {
491 0 : FD_TEST( fd_cstr_printf_check( lt_buf, sizeof( lt_buf ), NULL, "<= %lu", edge-1UL ) );
492 0 : }
493 0 : lt_str = lt_buf;
494 0 : }
495 0 :
496 0 : /* Create visual bar - scale to max 20 characters. */
497 0 : char bar_buf[ 22 ];
498 0 : if( bucket_count>0UL && total_count>0UL ) {
499 0 : ulong bar_length = (bucket_count*20UL)/total_count;
500 0 : if( !bar_length ) bar_length = 1;
501 0 : for( ulong j=0UL; j<bar_length; j++ ) { bar_buf[ j ] = '*'; }
502 0 : bar_buf[ bar_length ] = '\0';
503 0 : } else {
504 0 : bar_buf[ 0 ] = '\0';
505 0 : }
506 0 :
507 0 : fd_sched_printf( sched, " | %19s | %-20s | %12lu |\n", lt_str, bar_buf, bucket_count );
508 0 : }
509 0 : }
510 :
511 : FD_FN_UNUSED static void
512 0 : print_block_metrics( fd_sched_t * sched, fd_sched_block_t * block ) {
513 0 : fd_sched_printf( sched, "block idx %lu, block slot %lu, parent_slot %lu, fec_eos %d, rooted %d, txn_parsed_cnt %u, txn_exec_done_cnt %u, txn_sigverify_done_cnt %u, poh_hashing_done_cnt %u, poh_hash_cmp_done_cnt %u, txn_done_cnt %u, shred_cnt %u, mblk_cnt %u, mblk_freed_cnt %u, mblk_tick_cnt %u, mblk_unhashed_cnt %u, hashcnt %lu, txn_pool_max_popcnt %lu/%lu, mblk_pool_max_popcnt %lu/%lu, block_pool_max_popcnt %lu/%lu, mblks_rem %lu, txns_rem %lu, fec_buf_sz %u, fec_buf_boff %u, fec_buf_soff %u, fec_eob %d, fec_sob %d\n",
514 0 : block_to_idx( sched, block ), block->slot, block->parent_slot, block->fec_eos, block->rooted, block->txn_parsed_cnt, block->txn_exec_done_cnt, block->txn_sigverify_done_cnt, block->poh_hashing_done_cnt, block->poh_hash_cmp_done_cnt, block->txn_done_cnt, block->shred_cnt, block->mblk_cnt, block->mblk_freed_cnt, block->mblk_tick_cnt, block->mblk_unhashed_cnt, block->hashcnt, block->txn_pool_max_popcnt, sched->depth, block->mblk_pool_max_popcnt, sched->depth, block->block_pool_max_popcnt, sched->block_cnt_max, block->mblks_rem, block->txns_rem, block->fec_buf_sz, block->fec_buf_boff, block->fec_buf_soff, block->fec_eob, block->fec_sob );
515 0 : }
516 :
517 : FD_FN_UNUSED static void
518 57 : print_block_debug( fd_sched_t * sched, fd_sched_block_t * block ) {
519 57 : fd_sched_printf( sched, "block idx %lu, 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, poh_hashing_in_flight_cnt %u, poh_hashing_done_cnt %u, poh_hash_cmp_done_cnt %u, txn_done_cnt %u, shred_cnt %u, mblk_cnt %u, mblk_freed_cnt %u, mblk_tick_cnt %u, mblk_unhashed_cnt %u, hashcnt %lu, txn_pool_max_popcnt %lu/%lu, mblk_pool_max_popcnt %lu/%lu, block_pool_max_popcnt %lu/%lu, tick_hashcnt_wmk %lu, curr_tick_hashcnt %lu, hashes_per_tick %lu, mblks_rem %lu, txns_rem %lu, fec_buf_sz %u, fec_buf_boff %u, fec_buf_soff %u, fec_eob %d, fec_sob %d\n",
520 57 : block_to_idx( sched, block ), 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->poh_hashing_in_flight_cnt, block->poh_hashing_done_cnt, block->poh_hash_cmp_done_cnt, block->txn_done_cnt, block->shred_cnt, block->mblk_cnt, block->mblk_freed_cnt, block->mblk_tick_cnt, block->mblk_unhashed_cnt, block->hashcnt, block->txn_pool_max_popcnt, sched->depth, block->mblk_pool_max_popcnt, sched->depth, block->block_pool_max_popcnt, sched->block_cnt_max, block->tick_hashcnt_wmk, block->curr_tick_hashcnt, block->hashes_per_tick, block->mblks_rem, block->txns_rem, block->fec_buf_sz, block->fec_buf_boff, block->fec_buf_soff, block->fec_eob, block->fec_sob );
521 57 : }
522 :
523 : FD_FN_UNUSED static void
524 9 : print_block_and_parent( fd_sched_t * sched, fd_sched_block_t * block ) {
525 9 : print_block_debug( sched, block );
526 9 : fd_sched_block_t * parent = block_pool_ele( sched, block->parent_idx );
527 9 : if( FD_LIKELY( parent ) ) print_block_debug( sched, parent );
528 9 : }
529 :
530 : FD_FN_UNUSED static void
531 18 : print_metrics( fd_sched_t * sched ) {
532 18 : fd_sched_printf( sched, "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_bad_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, fork_observed_cnt %u, alut_success_cnt %u, alut_serializing_cnt %u, txn_abandoned_parsed_cnt %u, txn_abandoned_exec_done_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_mixin_done_cnt %lu, txn_done_cnt %lu, mblk_parsed_cnt %lu, mblk_poh_hashed_cnt %lu, mblk_poh_done_cnt %lu, bytes_ingested_cnt %lu, bytes_ingested_unparsed_cnt %lu, bytes_dropped_cnt %lu, fec_cnt %lu\n",
533 18 : 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_bad_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->fork_observed_cnt, sched->metrics->alut_success_cnt, sched->metrics->alut_serializing_cnt, sched->metrics->txn_abandoned_parsed_cnt, sched->metrics->txn_abandoned_exec_done_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_mixin_done_cnt, sched->metrics->txn_done_cnt, sched->metrics->mblk_parsed_cnt, sched->metrics->mblk_poh_hashed_cnt, sched->metrics->mblk_poh_done_cnt, sched->metrics->bytes_ingested_cnt, sched->metrics->bytes_ingested_unparsed_cnt, sched->metrics->bytes_dropped_cnt, sched->metrics->fec_cnt );
534 :
535 18 : }
536 :
537 : FD_FN_UNUSED static void
538 18 : print_sched( fd_sched_t * sched ) {
539 18 : fd_sched_printf( sched, "sched canary 0x%lx, exec_cnt %lu, root_idx %lu, txn_exec_ready_bitset[ 0 ] 0x%lx, sigverify_ready_bitset[ 0 ] 0x%lx, poh_ready_bitset[ 0 ] 0x%lx, 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, staged_popcnt_wmk %lu, txn_pool_free_cnt %lu/%lu, block_pool_popcnt %lu/%lu\n",
540 18 : sched->canary, sched->exec_cnt, sched->root_idx, sched->txn_exec_ready_bitset[ 0 ], sched->sigverify_ready_bitset[ 0 ], sched->poh_ready_bitset[ 0 ], 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->staged_popcnt_wmk, sched->txn_pool_free_cnt, sched->depth, sched->block_pool_popcnt, sched->block_cnt_max );
541 18 : fd_sched_block_t * active_block = block_pool_ele( sched, sched->active_bank_idx );
542 18 : if( active_block ) print_block_debug( sched, active_block );
543 90 : for( int l=0; l<(int)FD_SCHED_MAX_STAGING_LANES; l++ ) {
544 72 : if( fd_ulong_extract_bit( sched->staged_bitset, l ) ) {
545 27 : fd_sched_block_t * block = block_pool_ele( sched, sched->staged_head_bank_idx[ l ] );
546 27 : print_block_debug( sched, block );
547 27 : }
548 72 : }
549 18 : }
550 :
551 : FD_FN_UNUSED static void
552 9 : print_all( fd_sched_t * sched, fd_sched_block_t * block ) {
553 9 : print_metrics( sched );
554 9 : print_sched( sched );
555 9 : print_block_and_parent( sched, block );
556 9 : }
557 :
558 : static void
559 9 : handle_bad_block( fd_sched_t * sched, fd_sched_block_t * block ) {
560 9 : sched->print_buf_sz = 0UL;
561 9 : print_all( sched, block );
562 9 : FD_LOG_DEBUG(( "%s", sched->print_buf ));
563 9 : subtree_abandon( sched, block );
564 9 : sched->metrics->block_bad_cnt++;
565 9 : check_or_set_active_block( sched );
566 9 : }
567 :
568 :
569 : /* Public functions. */
570 :
571 : ulong
572 144 : fd_sched_align( void ) {
573 144 : return fd_ulong_max( alignof(fd_sched_t),
574 144 : fd_ulong_max( fd_rdisp_align(),
575 144 : fd_ulong_max( alignof(fd_sched_block_t), 64UL ))); /* Minimally cache line aligned. */
576 144 : }
577 :
578 : ulong
579 : fd_sched_footprint( ulong depth,
580 12 : ulong block_cnt_max ) {
581 12 : if( FD_UNLIKELY( depth<FD_SCHED_MIN_DEPTH || depth>FD_SCHED_MAX_DEPTH ) ) return 0UL; /* bad depth */
582 12 : if( FD_UNLIKELY( !block_cnt_max ) ) return 0UL; /* bad block_cnt_max */
583 12 : if( FD_UNLIKELY( depth>UINT_MAX-1UL ) ) return 0UL; /* mblk_pool use uint as pointers */
584 :
585 12 : ulong l = FD_LAYOUT_INIT;
586 12 : l = FD_LAYOUT_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
587 12 : l = FD_LAYOUT_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( depth, block_cnt_max ) ); /* dispatcher */
588 12 : l = FD_LAYOUT_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) ); /* block pool */
589 12 : l = FD_LAYOUT_APPEND( l, ref_q_align(), ref_q_footprint( block_cnt_max ) );
590 12 : l = FD_LAYOUT_APPEND( l, alignof(fd_txn_p_t), depth*sizeof(fd_txn_p_t) ); /* txn_pool */
591 12 : l = FD_LAYOUT_APPEND( l, alignof(fd_sched_txn_info_t), depth*sizeof(fd_sched_txn_info_t) ); /* txn_info_pool */
592 12 : l = FD_LAYOUT_APPEND( l, alignof(fd_sched_mblk_t), depth*sizeof(fd_sched_mblk_t) ); /* mblk_pool */
593 12 : return FD_LAYOUT_FINI( l, fd_sched_align() );
594 12 : }
595 :
596 : void *
597 : fd_sched_new( void * mem,
598 : fd_rng_t * rng,
599 : ulong depth,
600 : ulong block_cnt_max,
601 12 : ulong exec_cnt ) {
602 :
603 12 : if( FD_UNLIKELY( !mem ) ) {
604 0 : FD_LOG_WARNING(( "NULL mem" ));
605 0 : return NULL;
606 0 : }
607 :
608 12 : if( FD_UNLIKELY( !rng ) ) {
609 0 : FD_LOG_WARNING(( "NULL rng" ));
610 0 : return NULL;
611 0 : }
612 :
613 12 : if( FD_UNLIKELY( !fd_ulong_is_aligned( (ulong)mem, fd_sched_align() ) ) ) {
614 0 : FD_LOG_WARNING(( "misaligned mem (%p)", mem ));
615 0 : return NULL;
616 0 : }
617 :
618 12 : if( FD_UNLIKELY( depth<FD_SCHED_MIN_DEPTH || depth>FD_SCHED_MAX_DEPTH ) ) {
619 0 : FD_LOG_WARNING(( "bad depth (%lu)", depth ));
620 0 : return NULL;
621 0 : }
622 :
623 12 : if( FD_UNLIKELY( !block_cnt_max ) ) {
624 0 : FD_LOG_WARNING(( "bad block_cnt_max (%lu)", block_cnt_max ));
625 0 : return NULL;
626 0 : }
627 :
628 12 : if( FD_UNLIKELY( depth>UINT_MAX-1UL ) ) {
629 0 : FD_LOG_WARNING(( "bad depth (%lu)", depth ));
630 0 : return NULL;
631 0 : }
632 :
633 12 : if( FD_UNLIKELY( !exec_cnt || exec_cnt>FD_SCHED_MAX_EXEC_TILE_CNT ) ) {
634 0 : FD_LOG_WARNING(( "bad exec_cnt (%lu)", exec_cnt ));
635 0 : return NULL;
636 0 : }
637 :
638 12 : FD_SCRATCH_ALLOC_INIT( l, mem );
639 12 : fd_sched_t * sched = FD_SCRATCH_ALLOC_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
640 12 : void * _rdisp = FD_SCRATCH_ALLOC_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( depth, block_cnt_max ) );
641 12 : void * _bpool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) );
642 12 : void * _ref_q = FD_SCRATCH_ALLOC_APPEND( l, ref_q_align(), ref_q_footprint( block_cnt_max ) );
643 12 : fd_txn_p_t * _txn_pool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_txn_p_t), depth*sizeof(fd_txn_p_t) );
644 12 : fd_sched_txn_info_t * _txn_info_pool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_txn_info_t), depth*sizeof(fd_sched_txn_info_t) );
645 12 : fd_sched_mblk_t * _mblk_pool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_mblk_t), depth*sizeof(fd_sched_mblk_t) );
646 12 : FD_SCRATCH_ALLOC_FINI( l, fd_sched_align() );
647 :
648 12 : sched->txn_pool = _txn_pool;
649 12 : sched->txn_info_pool = _txn_info_pool;
650 12 : sched->mblk_pool = _mblk_pool;
651 :
652 12 : fd_rdisp_new( _rdisp, depth, block_cnt_max, fd_rng_ulong( rng ) );
653 :
654 12 : fd_sched_block_t * bpool = (fd_sched_block_t *)_bpool;
655 72 : for( ulong i=0; i<block_cnt_max; i++ ) {
656 60 : bpool[ i ].in_sched = 0;
657 60 : mblk_slist_new( bpool[ i ].mblks_unhashed );
658 60 : mblk_slist_new( bpool[ i ].mblks_hashing_in_progress );
659 60 : mblk_slist_new( bpool[ i ].mblks_mixin_in_progress );
660 60 : }
661 :
662 12 : FD_TEST( fd_chkdup_new( sched->chkdup, rng ) );
663 :
664 12 : fd_memset( sched->metrics, 0, sizeof(fd_sched_metrics_t) );
665 12 : sched->txn_in_flight_last_tick = LONG_MAX;
666 12 : sched->next_ready_last_tick = LONG_MAX;
667 12 : sched->next_ready_last_bank_idx = ULONG_MAX;
668 :
669 12 : sched->canary = FD_SCHED_MAGIC;
670 12 : sched->depth = depth;
671 12 : sched->block_cnt_max = block_cnt_max;
672 12 : sched->exec_cnt = exec_cnt;
673 12 : sched->bypass_poh_verify = 0;
674 12 : sched->bypass_alut_resolution = 0;
675 12 : sched->root_idx = ULONG_MAX;
676 12 : sched->active_bank_idx = ULONG_MAX;
677 12 : sched->last_active_bank_idx = ULONG_MAX;
678 12 : sched->staged_bitset = 0UL;
679 12 : sched->staged_popcnt_wmk = 0UL;
680 :
681 12 : sched->txn_exec_ready_bitset[ 0 ] = fd_ulong_mask_lsb( (int)exec_cnt );
682 12 : sched->sigverify_ready_bitset[ 0 ] = fd_ulong_mask_lsb( (int)exec_cnt );
683 12 : sched->poh_ready_bitset[ 0 ] = fd_ulong_mask_lsb( (int)exec_cnt );
684 :
685 12 : sched->txn_pool_free_cnt = depth-1UL; /* -1 because index 0 is unusable as a sentinel reserved by the dispatcher */
686 :
687 6144 : for( ulong i=0UL; i<depth-1UL; i++ ) sched->mblk_pool[ i ].next = (uint)(i+1UL);
688 12 : sched->mblk_pool[ depth-1UL ].next = UINT_MAX;
689 12 : sched->mblk_pool_free_head = 0U;
690 12 : sched->mblk_pool_free_cnt = depth;
691 :
692 12 : txn_bitset_new( sched->exec_done_set );
693 12 : txn_bitset_new( sched->sigverify_done_set );
694 12 : txn_bitset_new( sched->poh_mixin_done_set );
695 :
696 12 : sched->block_pool_popcnt = 0UL;
697 :
698 12 : ref_q_new( _ref_q, block_cnt_max );
699 :
700 12 : return sched;
701 12 : }
702 :
703 : fd_sched_t *
704 12 : fd_sched_join( void * mem ) {
705 :
706 12 : if( FD_UNLIKELY( !mem ) ) {
707 0 : FD_LOG_WARNING(( "NULL mem" ));
708 0 : return NULL;
709 0 : }
710 :
711 12 : fd_sched_t * sched = (fd_sched_t *)mem;
712 12 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
713 12 : ulong depth = sched->depth;
714 12 : ulong block_cnt_max = sched->block_cnt_max;
715 :
716 12 : FD_SCRATCH_ALLOC_INIT( l, mem );
717 12 : /* */ FD_SCRATCH_ALLOC_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
718 12 : void * _rdisp = FD_SCRATCH_ALLOC_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( depth, block_cnt_max ) );
719 12 : void * _bpool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) );
720 12 : void * _ref_q = FD_SCRATCH_ALLOC_APPEND( l, ref_q_align(), ref_q_footprint( block_cnt_max ) );
721 12 : /* */ FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_txn_p_t), depth*sizeof(fd_txn_p_t) );
722 12 : /* */ FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_txn_info_t), depth*sizeof(fd_sched_txn_info_t) );
723 12 : /* */ FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_mblk_t), depth*sizeof(fd_sched_mblk_t) );
724 12 : FD_SCRATCH_ALLOC_FINI( l, fd_sched_align() );
725 :
726 12 : sched->rdisp = fd_rdisp_join( _rdisp );
727 12 : sched->ref_q = ref_q_join( _ref_q );
728 12 : sched->block_pool = _bpool;
729 :
730 72 : for( ulong i=0; i<block_cnt_max; i++ ) {
731 60 : mblk_slist_join( sched->block_pool[ i ].mblks_unhashed );
732 60 : mblk_slist_join( sched->block_pool[ i ].mblks_hashing_in_progress );
733 60 : mblk_slist_join( sched->block_pool[ i ].mblks_mixin_in_progress );
734 60 : }
735 :
736 12 : txn_bitset_join( sched->exec_done_set );
737 12 : txn_bitset_join( sched->sigverify_done_set );
738 12 : txn_bitset_join( sched->poh_mixin_done_set );
739 :
740 12 : return sched;
741 12 : }
742 :
743 : int
744 24 : fd_sched_fec_can_ingest( fd_sched_t * sched, fd_sched_fec_t * fec ) {
745 24 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
746 24 : FD_TEST( fec->bank_idx<sched->block_cnt_max );
747 24 : FD_TEST( fec->parent_bank_idx<sched->block_cnt_max );
748 :
749 24 : if( FD_UNLIKELY( fec->fec->data_sz>FD_SCHED_MAX_PAYLOAD_PER_FEC ) ) {
750 0 : sched->print_buf_sz = 0UL;
751 0 : print_metrics( sched );
752 0 : print_sched( sched );
753 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
754 0 : FD_LOG_CRIT(( "invalid FEC set: fec->data_sz %lu, slot %lu, parent slot %lu", fec->fec->data_sz, fec->slot, fec->parent_slot ));
755 0 : }
756 :
757 24 : ulong fec_buf_sz = 0UL;
758 24 : fd_sched_block_t * block = block_pool_ele( sched, fec->bank_idx );
759 24 : if( FD_LIKELY( !fec->is_first_in_block ) ) {
760 0 : fec_buf_sz += block->fec_buf_sz-block->fec_buf_soff;
761 24 : } else {
762 : /* No residual data as this is a fresh new block. */
763 24 : }
764 : /* Addition is safe and won't overflow because we checked the FEC set
765 : size above. */
766 24 : fec_buf_sz += fec->fec->data_sz;
767 : /* Assuming every transaction is min size, do we have enough free
768 : entries in the txn pool? For a more precise txn count, we would
769 : have to do some parsing. */
770 24 : return sched->txn_pool_free_cnt>=fec_buf_sz/FD_TXN_MIN_SERIALIZED_SZ && sched->mblk_pool_free_cnt>=fec_buf_sz/sizeof(fd_microblock_hdr_t);
771 24 : }
772 :
773 : ulong
774 3 : fd_sched_can_ingest_cnt( fd_sched_t * sched ) {
775 3 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
776 : /* Worst case, we need one byte from the incoming data to extract a
777 : transaction out of the residual data, and the rest of the incoming
778 : data contributes toward min sized transactions. */
779 3 : return fd_ulong_min( sched->txn_pool_free_cnt/FD_SCHED_MAX_TXN_PER_FEC, sched->mblk_pool_free_cnt/FD_SCHED_MAX_MBLK_PER_FEC );
780 3 : }
781 :
782 : int
783 27 : fd_sched_is_drained( fd_sched_t * sched ) {
784 27 : int nothing_inflight = sched->exec_cnt==(ulong)fd_ulong_popcnt( sched->txn_exec_ready_bitset[ 0 ]&sched->sigverify_ready_bitset[ 0 ]&sched->poh_ready_bitset[ 0 ] );
785 27 : int nothing_queued = sched->active_bank_idx==ULONG_MAX;
786 27 : return nothing_inflight && nothing_queued;
787 27 : }
788 :
789 : FD_WARN_UNUSED int
790 : fd_sched_fec_ingest( fd_sched_t * sched,
791 24 : fd_sched_fec_t * fec ) {
792 24 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
793 24 : FD_TEST( fec->bank_idx<sched->block_cnt_max );
794 24 : FD_TEST( fec->parent_bank_idx<sched->block_cnt_max );
795 24 : FD_TEST( ref_q_empty( sched->ref_q ) );
796 :
797 24 : fd_sched_block_t * block = block_pool_ele( sched, fec->bank_idx );
798 :
799 24 : if( FD_UNLIKELY( fec->fec->data_sz>FD_SCHED_MAX_PAYLOAD_PER_FEC ) ) {
800 0 : sched->print_buf_sz = 0UL;
801 0 : print_all( sched, block );
802 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
803 0 : FD_LOG_CRIT(( "invalid FEC set: fec->data_sz %lu, slot %lu, parent slot %lu", fec->fec->data_sz, fec->slot, fec->parent_slot ));
804 0 : }
805 :
806 24 : sched->metrics->fec_cnt++;
807 :
808 24 : if( FD_UNLIKELY( fec->is_first_in_block ) ) {
809 : /* This is a new block. */
810 24 : add_block( sched, fec->bank_idx, fec->parent_bank_idx );
811 24 : block->slot = fec->slot;
812 24 : block->parent_slot = fec->parent_slot;
813 :
814 24 : if( FD_UNLIKELY( block->dying ) ) {
815 : /* The child of a dead block is also dead. We added it to our
816 : fork tree just so we could track an entire lineage of dead
817 : children and propagate the dead property to the entire lineage,
818 : in case there were frags for more than one dead children
819 : in-flight at the time the parent was abandoned. That being
820 : said, we shouldn't need to add the dead child to the
821 : dispatcher. */
822 0 : sched->metrics->block_added_dead_ood_cnt++;
823 :
824 : /* Release the refcnt right away since we don't intend to replay
825 : it at all. We do this so its bank does not linger and stall
826 : root advancement until the next root notify. */
827 0 : subtree_release_refcnt( sched, block );
828 :
829 : /* Ignore the FEC set for a dead block. */
830 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
831 0 : return 0;
832 0 : }
833 :
834 : /* Try to find a staging lane for this block. */
835 24 : int alloc_lane = 0;
836 24 : fd_sched_block_t * parent_block = block_pool_ele( sched, fec->parent_bank_idx );
837 24 : if( FD_LIKELY( parent_block->staged ) ) {
838 : /* Parent is staged. So see if we can continue down the same
839 : staging lane. */
840 0 : ulong staging_lane = parent_block->staging_lane;
841 0 : ulong child_idx = parent_block->child_idx;
842 0 : while( child_idx!=ULONG_MAX ) {
843 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
844 0 : if( child->staged && child->staging_lane==staging_lane ) {
845 : /* Found a child on the same lane. So we're done. */
846 0 : staging_lane = FD_RDISP_UNSTAGED;
847 0 : break;
848 0 : }
849 0 : child_idx = child->sibling_idx;
850 0 : }
851 : /* No child is staged on the same lane as the parent. So stage
852 : this block. This is the common case. */
853 0 : if( FD_LIKELY( staging_lane!=FD_RDISP_UNSTAGED ) ) {
854 0 : block->in_rdisp = 1;
855 0 : block->staged = 1;
856 0 : block->staging_lane = staging_lane;
857 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, staging_lane );
858 0 : sched->metrics->block_added_cnt++;
859 0 : sched->metrics->block_added_staged_cnt++;
860 0 : FD_LOG_DEBUG(( "block %lu:%lu entered lane %lu: add", block->slot, fec->bank_idx, staging_lane ));
861 0 : } else {
862 0 : alloc_lane = 1;
863 0 : }
864 24 : } else {
865 24 : if( block_is_stageable( parent_block ) ) {
866 : /* Parent is unstaged but stageable. So let's be unstaged too.
867 : This is not only a policy decision to be lazy and not promote
868 : the parent at the moment, but also an important invariant
869 : that we maintain for deadlock freeness in the face of staging
870 : lane shortage. See the comments in lane eviction for how
871 : this invariant is relevant. */
872 0 : block->in_rdisp = 1;
873 0 : block->staged = 0;
874 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, FD_RDISP_UNSTAGED );
875 0 : sched->metrics->block_added_cnt++;
876 0 : sched->metrics->block_added_unstaged_cnt++;
877 0 : FD_LOG_DEBUG(( "block %lu:%lu entered lane unstaged: add", block->slot, fec->bank_idx ));
878 24 : } else {
879 24 : alloc_lane = 1;
880 24 : }
881 24 : }
882 24 : if( FD_UNLIKELY( alloc_lane ) ) {
883 : /* We weren't able to inherit the parent's staging lane. So try
884 : to find a new staging lane. */
885 24 : if( FD_LIKELY( sched->staged_bitset!=fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) ) ) { /* Optimize for lane available. */
886 21 : int lane_idx = fd_ulong_find_lsb( ~sched->staged_bitset );
887 21 : if( FD_UNLIKELY( lane_idx>=(int)FD_SCHED_MAX_STAGING_LANES ) ) {
888 0 : FD_LOG_CRIT(( "invariant violation: lane_idx %d, sched->staged_bitset %lx",
889 0 : lane_idx, sched->staged_bitset ));
890 0 : }
891 21 : sched->staged_bitset = fd_ulong_set_bit( sched->staged_bitset, lane_idx );
892 21 : sched->staged_head_bank_idx[ lane_idx ] = fec->bank_idx;
893 21 : sched->staged_popcnt_wmk = fd_ulong_max( sched->staged_popcnt_wmk, (ulong)fd_ulong_popcnt( sched->staged_bitset ) );
894 21 : block->in_rdisp = 1;
895 21 : block->staged = 1;
896 21 : block->staging_lane = (ulong)lane_idx;
897 21 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, block->staging_lane );
898 21 : sched->metrics->block_added_cnt++;
899 21 : sched->metrics->block_added_staged_cnt++;
900 21 : FD_LOG_DEBUG(( "block %lu:%lu entered lane %lu: add", block->slot, fec->bank_idx, block->staging_lane ));
901 21 : } else {
902 : /* No lanes available. */
903 3 : block->in_rdisp = 1;
904 3 : block->staged = 0;
905 3 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, FD_RDISP_UNSTAGED );
906 3 : sched->metrics->block_added_cnt++;
907 3 : sched->metrics->block_added_unstaged_cnt++;
908 3 : FD_LOG_DEBUG(( "block %lu:%lu entered lane unstaged: add", block->slot, fec->bank_idx ));
909 3 : }
910 24 : }
911 24 : }
912 :
913 24 : block->txn_pool_max_popcnt = fd_ulong_max( block->txn_pool_max_popcnt, sched->depth - sched->txn_pool_free_cnt - 1UL );
914 24 : block->mblk_pool_max_popcnt = fd_ulong_max( block->mblk_pool_max_popcnt, sched->depth - sched->mblk_pool_free_cnt );
915 24 : block->block_pool_max_popcnt = fd_ulong_max( block->block_pool_max_popcnt, sched->block_pool_popcnt );
916 :
917 24 : if( FD_UNLIKELY( block->dying ) ) {
918 : /* Ignore the FEC set for a dead block. */
919 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
920 0 : return 1;
921 0 : }
922 :
923 24 : if( FD_UNLIKELY( !block->in_rdisp ) ) {
924 : /* Invariant: block must be in the dispatcher at this point. */
925 0 : sched->print_buf_sz = 0UL;
926 0 : print_all( sched, block );
927 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
928 0 : FD_LOG_CRIT(( "invariant violation: block->in_rdisp==0, slot %lu, parent slot %lu",
929 0 : block->slot, block->parent_slot ));
930 0 : }
931 :
932 24 : if( FD_UNLIKELY( block->fec_eos ) ) {
933 : /* This means something is wrong upstream. We're getting more FEC
934 : sets for a block that has already ended, or so we were told. */
935 0 : sched->print_buf_sz = 0UL;
936 0 : print_all( sched, block );
937 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
938 0 : FD_LOG_CRIT(( "invariant violation: block->fec_eos set but getting more FEC sets, slot %lu, parent slot %lu", fec->slot, fec->parent_slot ));
939 0 : }
940 24 : if( FD_UNLIKELY( block->fec_eob ) ) {
941 : /* If the previous FEC set ingestion and parse was successful,
942 : block->fec_eob should be cleared. The fact that fec_eob is set
943 : means that the previous batch didn't parse properly. So this is
944 : a bad block. We should refuse to replay down the fork. */
945 0 : FD_LOG_INFO(( "bad block: failed to parse, slot %lu, parent slot %lu", fec->slot, fec->parent_slot ));
946 0 : handle_bad_block( sched, block );
947 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
948 0 : return 0;
949 0 : }
950 24 : if( FD_UNLIKELY( block->child_idx!=ULONG_MAX ) ) {
951 : /* This means something is wrong upstream. FEC sets are not being
952 : delivered in replay order. We got a child block FEC set before
953 : this block was completely delivered. */
954 0 : sched->print_buf_sz = 0UL;
955 0 : print_all( sched, block );
956 0 : fd_sched_block_t * child_block = block_pool_ele( sched, block->child_idx );
957 0 : print_block_debug( sched, child_block );
958 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
959 0 : FD_LOG_CRIT(( "invariant violation: block->child_idx %lu, slot %lu, parent slot %lu", block->child_idx, fec->slot, fec->parent_slot ));
960 0 : }
961 :
962 24 : FD_TEST( block->fec_buf_sz>=block->fec_buf_soff );
963 24 : if( FD_LIKELY( block->fec_buf_sz>block->fec_buf_soff ) ) {
964 : /* If there is residual data from the previous FEC set within the
965 : same batch, we move it to the beginning of the buffer and append
966 : the new FEC set. */
967 0 : memmove( block->fec_buf, block->fec_buf+block->fec_buf_soff, block->fec_buf_sz-block->fec_buf_soff );
968 0 : }
969 24 : block->fec_buf_boff += block->fec_buf_soff;
970 24 : block->fec_buf_sz -= block->fec_buf_soff;
971 24 : block->fec_buf_soff = 0;
972 : /* Addition is safe and won't overflow because we checked the FEC
973 : set size above. */
974 24 : if( FD_UNLIKELY( block->fec_buf_sz+fec->fec->data_sz>FD_SCHED_MAX_FEC_BUF_SZ ) ) {
975 : /* The residual carried over from the previous parse is bounded: it
976 : can only be a partial transaction or a microblock/batch header
977 : that straddles a FEC set boundary. Any trailing bytes past the
978 : last microblock within the same batch are discarded by the
979 : parser, so they never contribute to the residual. So residual <=
980 : max(sizeof(ulong), sizeof(fd_microblock_hdr_t), FD_TXN_MTU), and
981 : the buffer is sized to always fit the residual plus a single FEC
982 : set. Otherwise, it's a bad block. Instead of crashing, we
983 : should refuse to replay down the fork. */
984 0 : FD_LOG_INFO(( "bad block: fec_buf_sz %u, fec->data_sz %lu, slot %lu, parent slot %lu", block->fec_buf_sz, fec->fec->data_sz, fec->slot, fec->parent_slot ));
985 0 : handle_bad_block( sched, block );
986 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
987 0 : return 0;
988 0 : }
989 :
990 : /* Append the new FEC set to the end of the buffer. */
991 24 : fd_memcpy( block->fec_buf+block->fec_buf_sz, fec->data, fec->fec->data_sz );
992 24 : block->fec_buf_sz += (uint)fec->fec->data_sz;
993 24 : sched->metrics->bytes_ingested_cnt += fec->fec->data_sz;
994 :
995 24 : block->fec_eob = fec->is_last_in_batch;
996 24 : block->fec_eos = fec->is_last_in_block;
997 :
998 24 : ulong block_sz = block->shred_cnt>0 ? block->shred_blk_offs[ block->shred_cnt-1 ] : 0UL;
999 48 : for( ulong i=0; i<fec->shred_cnt; i++ ) {
1000 24 : if( FD_LIKELY( i<32UL ) ) {
1001 24 : block->shred_blk_offs[ block->shred_cnt++ ] = (uint)block_sz + fec->fec->shred_offs[ i ];
1002 24 : } else if( FD_UNLIKELY( i!=fec->shred_cnt-1UL ) ) {
1003 : /* We don't track shred boundaries after 32 shreds, assume they're
1004 : sized uniformly */
1005 0 : ulong num_overflow_shreds = fec->shred_cnt-32UL;
1006 0 : ulong overflow_idx = i-32UL;
1007 0 : ulong overflow_data_sz = fec->fec->data_sz-fec->fec->shred_offs[ 31 ];
1008 0 : block->shred_blk_offs[ block->shred_cnt++ ] = (uint)block_sz + fec->fec->shred_offs[ 31 ] + (uint)(overflow_data_sz / num_overflow_shreds * (overflow_idx + 1UL));
1009 0 : } else {
1010 0 : block->shred_blk_offs[ block->shred_cnt++ ] = (uint)block_sz + (uint)fec->fec->data_sz;
1011 0 : }
1012 24 : }
1013 :
1014 24 : int err = fd_sched_parse( sched, block, fec->alut_ctx );
1015 :
1016 24 : if( FD_UNLIKELY( err==FD_SCHED_BAD_BLOCK ) ) {
1017 0 : handle_bad_block( sched, block );
1018 0 : sched->metrics->bytes_dropped_cnt += block->fec_buf_sz-block->fec_buf_soff;
1019 0 : return 0;
1020 0 : }
1021 :
1022 24 : if( FD_UNLIKELY( (fec->is_last_in_batch||fec->is_last_in_block) && (block->txns_rem||block->mblks_rem||block->fec_eob) ) ) {
1023 : /* A malformed block that fails to parse out exactly as many
1024 : transactions and microblocks as it should.
1025 :
1026 : Upon getting a last-in-batch FEC, everything in the ongoing batch
1027 : should completely parse, and the eob flag should be reset. There
1028 : should be no go around for a last-in-batch FEC. */
1029 0 : FD_LOG_INFO(( "bad block: bytes_rem %u, txns_rem %lu, mblks_rem %lu, fec_eob %d, slot %lu, parent slot %lu", block->fec_buf_sz-block->fec_buf_soff, block->txns_rem, block->mblks_rem, block->fec_eob, block->slot, block->parent_slot ));
1030 0 : handle_bad_block( sched, block );
1031 0 : return 0;
1032 0 : }
1033 :
1034 24 : if( FD_UNLIKELY( block->fec_eos && !block->last_mblk_is_tick ) ) {
1035 : /* The last microblock should be a tick.
1036 :
1037 : Note that this early parse-time detection could cause us to throw
1038 : a slightly different error from Agave, in the case that there are
1039 : too few ticks, since the tick count check precedes the trailing
1040 : entry check in Agave. That being said, ultimately a
1041 : TRAILING_ENTRY renders a block invalid, regardless of anything
1042 : else. */
1043 0 : FD_LOG_INFO(( "bad block: TRAILING_ENTRY, slot %lu, parent slot %lu, mblk_cnt %u", block->slot, block->parent_slot, block->mblk_cnt ));
1044 0 : handle_bad_block( sched, block );
1045 0 : return 0;
1046 0 : }
1047 :
1048 : /* We just received a FEC set, which may have made all transactions in
1049 : a partially parsed microblock available. If this were a malformed
1050 : block that ends in a non-tick microblock, there's not going to be a
1051 : hashing task from the missing ending tick to drain the mixin queue.
1052 : So we try to drain the mixin queue right here. Another option is
1053 : to drain it at dispatch time, when we are about to dispatch the end
1054 : of block signal, right before the check for whether block should
1055 : end. */
1056 24 : int mixin_res;
1057 24 : while( (mixin_res=maybe_mixin( sched, block )) ) {
1058 0 : if( FD_UNLIKELY( mixin_res==-1 ) ) {
1059 0 : handle_bad_block( sched, block );
1060 0 : return 0;
1061 0 : }
1062 0 : FD_TEST( mixin_res==1||mixin_res==2 );
1063 0 : }
1064 :
1065 : /* Check if we need to set the active block. */
1066 24 : check_or_set_active_block( sched );
1067 :
1068 24 : return 1;
1069 24 : }
1070 :
1071 : ulong
1072 45 : fd_sched_task_next_ready( fd_sched_t * sched, fd_sched_task_t * out ) {
1073 45 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1074 45 : FD_TEST( ref_q_empty( sched->ref_q ) );
1075 :
1076 45 : ulong exec_ready_bitset0 = sched->txn_exec_ready_bitset[ 0 ];
1077 45 : ulong exec_fully_ready_bitset = sched->sigverify_ready_bitset[ 0 ] & sched->poh_ready_bitset[ 0 ] & exec_ready_bitset0;
1078 45 : if( FD_UNLIKELY( !exec_fully_ready_bitset ) ) {
1079 : /* Early exit if no exec tiles available. */
1080 0 : return 0UL;
1081 0 : }
1082 :
1083 45 : if( FD_UNLIKELY( sched->active_bank_idx==ULONG_MAX ) ) {
1084 : /* No need to try activating a block. If we're in this state,
1085 : there's truly nothing to execute. We will activate something
1086 : when we ingest a FEC set with transactions. */
1087 9 : return 0UL;
1088 9 : }
1089 :
1090 36 : out->task_type = FD_SCHED_TT_NULL;
1091 :
1092 : /* We could in theory reevaluate staging lane allocation here and do
1093 : promotion/demotion as needed. It's a policy decision to minimize
1094 : fork churn for now and just execute down the same active fork. */
1095 :
1096 36 : ulong bank_idx = sched->active_bank_idx;
1097 36 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1098 36 : if( FD_UNLIKELY( block_should_deactivate( block ) ) ) {
1099 0 : sched->print_buf_sz = 0UL;
1100 0 : print_all( sched, block );
1101 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
1102 0 : FD_LOG_CRIT(( "invariant violation: active block %lu:%lu is not activatable nor has anything in-flight", block->slot, sched->active_bank_idx ));
1103 0 : }
1104 :
1105 36 : block->txn_pool_max_popcnt = fd_ulong_max( block->txn_pool_max_popcnt, sched->depth - sched->txn_pool_free_cnt - 1UL );
1106 36 : block->mblk_pool_max_popcnt = fd_ulong_max( block->mblk_pool_max_popcnt, sched->depth - sched->mblk_pool_free_cnt );
1107 36 : block->block_pool_max_popcnt = fd_ulong_max( block->block_pool_max_popcnt, sched->block_pool_popcnt );
1108 :
1109 36 : if( FD_UNLIKELY( !block->block_start_signaled ) ) {
1110 24 : out->task_type = FD_SCHED_TT_BLOCK_START;
1111 24 : out->block_start->bank_idx = bank_idx;
1112 24 : out->block_start->parent_bank_idx = block->parent_idx;
1113 24 : out->block_start->slot = block->slot;
1114 24 : block->block_start_signaled = 1;
1115 24 : sched->next_ready_last_tick = fd_tickcount();
1116 24 : sched->next_ready_last_bank_idx = bank_idx;
1117 24 : return 1UL;
1118 24 : }
1119 :
1120 12 : ulong exec_tile_idx0 = fd_ulong_if( !!exec_fully_ready_bitset, (ulong)fd_ulong_find_lsb( exec_fully_ready_bitset ), ULONG_MAX );
1121 12 : ulong exec_queued_cnt = block->txn_parsed_cnt-block->txn_exec_in_flight_cnt-block->txn_exec_done_cnt;
1122 12 : if( FD_LIKELY( exec_queued_cnt>0UL && fd_ulong_popcnt( exec_fully_ready_bitset ) ) ) { /* Optimize for no fork switching. */
1123 : /* Transaction execution has the highest priority. Current mainnet
1124 : block times are very much dominated by critical path transaction
1125 : execution. To achieve the fastest block replay speed, we can't
1126 : afford to make any mistake in critical path dispatching. Any
1127 : deviation from perfect critical path dispatching is basically
1128 : irrecoverable. As such, we try to keep all the exec tiles busy
1129 : with transaction execution, but we allow at most one transaction
1130 : to be in-flight per exec tile. This is to ensure that whenever a
1131 : critical path transaction completes, we have at least one exec
1132 : tile, e.g. the one that just completed said transaction, readily
1133 : available to continue executing down the critical path. */
1134 0 : out->txn_exec->txn_idx = fd_rdisp_get_next_ready( sched->rdisp, bank_idx );
1135 0 : if( FD_UNLIKELY( out->txn_exec->txn_idx==0UL ) ) {
1136 : /* There are transactions queued but none ready for execution.
1137 : This implies that there must be in-flight transactions on whose
1138 : completion the queued transactions depend. So we return and
1139 : wait for those in-flight transactions to retire. This is a
1140 : policy decision to execute as much as we can down the current
1141 : fork. */
1142 0 : if( FD_UNLIKELY( !block->txn_exec_in_flight_cnt ) ) {
1143 0 : sched->print_buf_sz = 0UL;
1144 0 : print_all( sched, block );
1145 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
1146 0 : FD_LOG_CRIT(( "invariant violation: no ready transaction found but block->txn_exec_in_flight_cnt==0" ));
1147 0 : }
1148 :
1149 : /* Next up are PoH tasks. Same dispatching policy as sigverify
1150 : tasks. */
1151 0 : ulong poh_ready_bitset = exec_fully_ready_bitset;
1152 0 : ulong poh_hashing_queued_cnt = block->mblk_cnt-block->poh_hashing_in_flight_cnt-block->poh_hashing_done_cnt;
1153 0 : if( FD_LIKELY( poh_hashing_queued_cnt>0UL && fd_ulong_popcnt( poh_ready_bitset )>fd_int_if( block->txn_exec_in_flight_cnt>0U, 0, 1 ) ) ) {
1154 0 : dispatch_poh( sched, block, bank_idx, fd_ulong_find_lsb( poh_ready_bitset ), out );
1155 0 : sched->next_ready_last_tick = fd_tickcount();
1156 0 : sched->next_ready_last_bank_idx = bank_idx;
1157 0 : return 1UL;
1158 0 : }
1159 :
1160 : /* Dispatch more sigverify tasks only if at least one exec tile is
1161 : executing transactions or completely idle. Allow at most one
1162 : sigverify task in-flight per tile, and only dispatch to
1163 : completely idle tiles. */
1164 0 : ulong sigverify_ready_bitset = exec_fully_ready_bitset;
1165 0 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
1166 0 : if( FD_LIKELY( sigverify_queued_cnt>0UL && fd_ulong_popcnt( sigverify_ready_bitset )>fd_int_if( block->txn_exec_in_flight_cnt>0U, 0, 1 ) ) ) {
1167 0 : dispatch_sigverify( sched, block, bank_idx, fd_ulong_find_lsb( sigverify_ready_bitset ), out );
1168 0 : sched->next_ready_last_tick = sched->txn_info_pool[ out->txn_sigverify->txn_idx ].tick_sigverify_disp = fd_tickcount();
1169 0 : sched->next_ready_last_bank_idx = bank_idx;
1170 0 : return 1UL;
1171 0 : }
1172 0 : return 0UL;
1173 0 : }
1174 0 : out->task_type = FD_SCHED_TT_TXN_EXEC;
1175 0 : out->txn_exec->bank_idx = bank_idx;
1176 0 : out->txn_exec->slot = block->slot;
1177 0 : out->txn_exec->exec_idx = exec_tile_idx0;
1178 0 : FD_TEST( out->txn_exec->exec_idx!=ULONG_MAX );
1179 :
1180 0 : long now = fd_tickcount();
1181 0 : ulong delta = (ulong)(now-sched->txn_in_flight_last_tick);
1182 0 : ulong txn_exec_busy_cnt = sched->exec_cnt-(ulong)fd_ulong_popcnt( exec_ready_bitset0 );
1183 0 : sched->metrics->txn_none_in_flight_tickcount += fd_ulong_if( txn_exec_busy_cnt==0UL && sched->txn_in_flight_last_tick!=LONG_MAX, delta, 0UL );
1184 0 : sched->metrics->txn_weighted_in_flight_tickcount += fd_ulong_if( txn_exec_busy_cnt!=0UL, delta, 0UL );
1185 0 : sched->metrics->txn_weighted_in_flight_cnt += delta*txn_exec_busy_cnt;
1186 0 : sched->txn_in_flight_last_tick = now;
1187 :
1188 0 : sched->txn_info_pool[ out->txn_exec->txn_idx ].tick_exec_disp = now;
1189 :
1190 0 : sched->txn_exec_ready_bitset[ 0 ] = fd_ulong_clear_bit( exec_ready_bitset0, (int)exec_tile_idx0);
1191 0 : sched->tile_to_bank_idx[ exec_tile_idx0 ] = bank_idx;
1192 :
1193 0 : block->txn_exec_in_flight_cnt++;
1194 0 : sched->metrics->txn_max_in_flight_cnt = fd_uint_max( sched->metrics->txn_max_in_flight_cnt, block->txn_exec_in_flight_cnt );
1195 :
1196 0 : if( FD_UNLIKELY( (~sched->txn_exec_ready_bitset[ 0 ])&(~sched->sigverify_ready_bitset[ 0 ])&(~sched->poh_ready_bitset[ 0 ])&fd_ulong_mask_lsb( (int)sched->exec_cnt ) ) ) FD_LOG_CRIT(( "invariant violation: txn_exec_ready_bitset 0x%lx sigverify_ready_bitset 0x%lx poh_ready_bitset 0x%lx", sched->txn_exec_ready_bitset[ 0 ], sched->sigverify_ready_bitset[ 0 ], sched->poh_ready_bitset[ 0 ] ));
1197 0 : ulong total_exec_busy_cnt = sched->exec_cnt-(ulong)fd_ulong_popcnt( sched->txn_exec_ready_bitset[ 0 ]&sched->sigverify_ready_bitset[ 0 ]&sched->poh_ready_bitset[ 0 ] );
1198 0 : if( FD_UNLIKELY( block->txn_exec_in_flight_cnt+block->txn_sigverify_in_flight_cnt+block->poh_hashing_in_flight_cnt!=total_exec_busy_cnt ) ) {
1199 : /* Ideally we'd simply assert that the two sides of the equation
1200 : are equal. But abandoned blocks throw a wrench into this. We
1201 : allow abandoned blocks to have in-flight transactions that are
1202 : naturally drained while we try to dispatch from another block.
1203 : In such cases, the total number of in-flight transactions
1204 : should include the abandoned blocks too. The contract is that
1205 : blocks with in-flight transactions cannot be abandoned or
1206 : demoted from rdisp. So a dying block has to be the head of one
1207 : of the staging lanes. */
1208 : // FIXME This contract no longer true if we implement immediate
1209 : // demotion of abandoned blocks.
1210 0 : ulong total_in_flight = 0UL;
1211 0 : for( int l=0; l<(int)FD_SCHED_MAX_STAGING_LANES; l++ ) {
1212 0 : if( fd_ulong_extract_bit( sched->staged_bitset, l ) ) {
1213 0 : fd_sched_block_t * staged_block = block_pool_ele( sched, sched->staged_head_bank_idx[ l ] );
1214 0 : if( FD_UNLIKELY( block_is_in_flight( staged_block )&&!(staged_block==block||staged_block->dying) ) ) {
1215 0 : sched->print_buf_sz = 0UL;
1216 0 : print_all( sched, staged_block );
1217 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
1218 0 : FD_LOG_CRIT(( "invariant violation: in-flight block is neither active nor dying" ));
1219 0 : }
1220 0 : total_in_flight += staged_block->txn_exec_in_flight_cnt;
1221 0 : total_in_flight += staged_block->txn_sigverify_in_flight_cnt;
1222 0 : total_in_flight += staged_block->poh_hashing_in_flight_cnt;
1223 0 : }
1224 0 : }
1225 0 : if( FD_UNLIKELY( total_in_flight!=total_exec_busy_cnt ) ) {
1226 0 : sched->print_buf_sz = 0UL;
1227 0 : print_all( sched, block );
1228 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
1229 0 : FD_LOG_CRIT(( "invariant violation: total_in_flight %lu != total_exec_busy_cnt %lu", total_in_flight, total_exec_busy_cnt ));
1230 0 : }
1231 0 : FD_LOG_DEBUG(( "exec_busy_cnt %lu checks out", total_exec_busy_cnt ));
1232 0 : }
1233 0 : sched->next_ready_last_tick = now;
1234 0 : sched->next_ready_last_bank_idx = bank_idx;
1235 0 : return 1UL;
1236 0 : }
1237 :
1238 : /* At this point txn_queued_cnt==0 */
1239 :
1240 : /* Next up are PoH tasks. Same dispatching policy as sigverify. */
1241 12 : ulong poh_ready_bitset = exec_fully_ready_bitset;
1242 12 : ulong poh_hashing_queued_cnt = block->mblk_cnt-block->poh_hashing_in_flight_cnt-block->poh_hashing_done_cnt;
1243 12 : if( FD_LIKELY( poh_hashing_queued_cnt>0UL && fd_ulong_popcnt( poh_ready_bitset )>fd_int_if( block->fec_eos||block->txn_exec_in_flight_cnt>0U||sched->exec_cnt==1UL, 0, 1 ) ) ) {
1244 9 : dispatch_poh( sched, block, bank_idx, fd_ulong_find_lsb( poh_ready_bitset ), out );
1245 9 : sched->next_ready_last_tick = fd_tickcount();
1246 9 : sched->next_ready_last_bank_idx = bank_idx;
1247 9 : return 1UL;
1248 9 : }
1249 :
1250 : /* Try to dispatch a sigverify task, but leave one exec tile idle for
1251 : critical path execution, unless there's not going to be any more
1252 : real transactions for the critical path. In the degenerate case of
1253 : only one exec tile, keep it busy. */
1254 3 : ulong sigverify_ready_bitset = exec_fully_ready_bitset;
1255 3 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
1256 3 : if( FD_LIKELY( sigverify_queued_cnt>0UL && fd_ulong_popcnt( sigverify_ready_bitset )>fd_int_if( block->fec_eos||block->txn_exec_in_flight_cnt>0U||sched->exec_cnt==1UL, 0, 1 ) ) ) {
1257 0 : dispatch_sigverify( sched, block, bank_idx, fd_ulong_find_lsb( sigverify_ready_bitset ), out );
1258 0 : sched->next_ready_last_tick = sched->txn_info_pool[ out->txn_sigverify->txn_idx ].tick_sigverify_disp = fd_tickcount();
1259 0 : sched->next_ready_last_bank_idx = bank_idx;
1260 0 : return 1UL;
1261 0 : }
1262 :
1263 3 : if( FD_UNLIKELY( block_should_signal_end( block ) ) ) {
1264 3 : FD_TEST( block->block_start_signaled );
1265 3 : if( FD_UNLIKELY( verify_ticks_final( block ) ) ) {
1266 : /* Tick verification can't be done at parse time (except for
1267 : TRAILING_ENTRY), because we may not know the expected number of
1268 : hashes yet. It can't be driven by transaction dispatch or
1269 : completion, because the block may be empty. Similary, it can't
1270 : be driven by PoH hashing, because a bad block may simply not
1271 : have any microblocks. */
1272 3 : handle_bad_block( sched, block );
1273 3 : out->task_type = FD_SCHED_TT_MARK_DEAD;
1274 3 : out->mark_dead->bank_idx = bank_idx;
1275 3 : sched->next_ready_last_tick = fd_tickcount();
1276 3 : sched->next_ready_last_bank_idx = bank_idx;
1277 3 : return 1UL;
1278 3 : }
1279 0 : out->task_type = FD_SCHED_TT_BLOCK_END;
1280 0 : out->block_end->bank_idx = bank_idx;
1281 0 : block->block_end_signaled = 1;
1282 0 : FD_TEST( block->refcnt );
1283 0 : block->refcnt = 0;
1284 0 : FD_TEST( ref_q_avail( sched->ref_q ) );
1285 0 : ref_q_push_tail( sched->ref_q, bank_idx );
1286 0 : sched->next_ready_last_tick = fd_tickcount();
1287 0 : sched->next_ready_last_bank_idx = bank_idx;
1288 0 : return 1UL;
1289 0 : }
1290 :
1291 : /* Nothing queued for the active block. If we haven't received all
1292 : the FEC sets for it, then return and wait for more FEC sets, while
1293 : there are in-flight transactions. This is a policy decision to
1294 : minimize fork churn and allow for executing down the current fork
1295 : as much as we can. If we have received all the FEC sets for it,
1296 : then we'd still like to return and wait for the in-flight
1297 : transactions to retire, before switching to a different block.
1298 :
1299 : Either way, there should be in-flight transactions. We deactivate
1300 : the active block the moment we exhausted transactions from it.
1301 :
1302 : We don't assert block_is_in_flight() here because there might be
1303 : in-flight tasks from dying blocks that are preventing us from
1304 : dispatching anything momentarily. So we assert the global exec
1305 : tile busy count. This doesn't lose too much assertion coverage
1306 : since dying blocks are few and far between. */
1307 0 : ulong total_exec_busy_cnt = sched->exec_cnt-(ulong)fd_ulong_popcnt( exec_fully_ready_bitset );
1308 0 : if( FD_UNLIKELY( !total_exec_busy_cnt ) ) {
1309 0 : sched->print_buf_sz = 0UL;
1310 0 : print_all( sched, block );
1311 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
1312 0 : FD_LOG_CRIT(( "invariant violation: expected in-flight tasks but none found" ));
1313 0 : }
1314 :
1315 0 : return 0UL;
1316 0 : }
1317 :
1318 : int
1319 33 : fd_sched_task_done( fd_sched_t * sched, ulong task_type, ulong txn_idx, ulong exec_idx, void * data ) {
1320 33 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1321 :
1322 33 : ulong bank_idx = ULONG_MAX;
1323 33 : switch( task_type ) {
1324 24 : case FD_SCHED_TT_BLOCK_START:
1325 24 : case FD_SCHED_TT_BLOCK_END: {
1326 24 : (void)txn_idx;
1327 24 : (void)data;
1328 24 : bank_idx = sched->active_bank_idx;
1329 24 : break;
1330 24 : }
1331 0 : case FD_SCHED_TT_TXN_EXEC:
1332 0 : case FD_SCHED_TT_TXN_SIGVERIFY: {
1333 0 : (void)data;
1334 0 : FD_TEST( txn_idx < sched->depth );
1335 0 : bank_idx = sched->tile_to_bank_idx[ exec_idx ];
1336 0 : break;
1337 0 : }
1338 9 : case FD_SCHED_TT_POH_HASH: {
1339 9 : (void)txn_idx;
1340 9 : bank_idx = sched->tile_to_bank_idx[ exec_idx ];
1341 9 : break;
1342 0 : }
1343 0 : default: FD_LOG_CRIT(( "unsupported task_type %lu", task_type ));
1344 33 : }
1345 33 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1346 :
1347 33 : if( FD_UNLIKELY( !block->in_sched ) ) {
1348 0 : FD_LOG_CRIT(( "invariant violation: block->in_sched==0, block %lu:%lu, parent slot %lu",
1349 0 : block->slot, bank_idx, block->parent_slot ));
1350 0 : }
1351 33 : if( FD_UNLIKELY( !block->staged ) ) {
1352 : /* Invariant: only staged blocks can have in-flight transactions. */
1353 0 : FD_LOG_CRIT(( "invariant violation: block->staged==0, block %lu:%lu, parent slot %lu",
1354 0 : block->slot, bank_idx, block->parent_slot ));
1355 0 : }
1356 33 : if( FD_UNLIKELY( !block->in_rdisp ) ) {
1357 : /* Invariant: staged blocks must be in the dispatcher. */
1358 0 : FD_LOG_CRIT(( "invariant violation: block->in_rdisp==0, block %lu:%lu, parent slot %lu",
1359 0 : block->slot, bank_idx, block->parent_slot ));
1360 0 : }
1361 :
1362 33 : block->txn_pool_max_popcnt = fd_ulong_max( block->txn_pool_max_popcnt, sched->depth - sched->txn_pool_free_cnt - 1UL );
1363 33 : block->mblk_pool_max_popcnt = fd_ulong_max( block->mblk_pool_max_popcnt, sched->depth - sched->mblk_pool_free_cnt );
1364 33 : block->block_pool_max_popcnt = fd_ulong_max( block->block_pool_max_popcnt, sched->block_pool_popcnt );
1365 :
1366 33 : int exec_tile_idx = (int)exec_idx;
1367 :
1368 33 : switch( task_type ) {
1369 24 : case FD_SCHED_TT_BLOCK_START: {
1370 24 : FD_TEST( !block->block_start_done );
1371 24 : block->block_start_done = 1;
1372 24 : break;
1373 24 : }
1374 0 : case FD_SCHED_TT_BLOCK_END: {
1375 : /* It may seem redundant to be invoking task_done() on these
1376 : somewhat fake tasks. But these are necessary to drive state
1377 : transition for empty blocks or slow blocks. */
1378 0 : FD_TEST( !block->block_end_done );
1379 0 : block->block_end_done = 1;
1380 0 : sched->print_buf_sz = 0UL;
1381 0 : print_block_metrics( sched, block );
1382 0 : FD_LOG_DEBUG(( "block %lu:%lu replayed fully: %s", block->slot, bank_idx, sched->print_buf ));
1383 0 : break;
1384 0 : }
1385 0 : case FD_SCHED_TT_TXN_EXEC: {
1386 0 : long now = fd_tickcount();
1387 0 : ulong delta = (ulong)(now-sched->txn_in_flight_last_tick);
1388 0 : ulong txn_exec_busy_cnt = sched->exec_cnt-(ulong)fd_ulong_popcnt( sched->txn_exec_ready_bitset[ 0 ] );
1389 0 : sched->metrics->txn_weighted_in_flight_tickcount += delta;
1390 0 : sched->metrics->txn_weighted_in_flight_cnt += delta*txn_exec_busy_cnt;
1391 0 : sched->txn_in_flight_last_tick = now;
1392 :
1393 0 : sched->txn_info_pool[ txn_idx ].tick_exec_done = now;
1394 :
1395 0 : block->txn_exec_done_cnt++;
1396 0 : block->txn_exec_in_flight_cnt--;
1397 0 : FD_TEST( !fd_ulong_extract_bit( sched->txn_exec_ready_bitset[ 0 ], exec_tile_idx ) );
1398 0 : sched->txn_exec_ready_bitset[ 0 ] = fd_ulong_set_bit( sched->txn_exec_ready_bitset[ 0 ], exec_tile_idx );
1399 0 : sched->metrics->txn_exec_done_cnt++;
1400 0 : txn_bitset_insert( sched->exec_done_set, txn_idx );
1401 0 : sched->txn_info_pool[ txn_idx ].flags |= FD_SCHED_TXN_EXEC_DONE;
1402 0 : if( txn_bitset_test( sched->sigverify_done_set, txn_idx ) && txn_bitset_test( sched->poh_mixin_done_set, txn_idx ) ) {
1403 : /* Release the txn_idx if all tasks on it are done. This is
1404 : guaranteed to only happen once per transaction because
1405 : whichever one completed first would not release. */
1406 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 1 );
1407 0 : sched->txn_pool_free_cnt++;
1408 0 : block->txn_done_cnt++;
1409 0 : sched->metrics->txn_done_cnt++;
1410 0 : } else {
1411 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 0 );
1412 0 : }
1413 0 : break;
1414 0 : }
1415 0 : case FD_SCHED_TT_TXN_SIGVERIFY: {
1416 0 : sched->txn_info_pool[ txn_idx ].tick_sigverify_done = fd_tickcount();
1417 0 : block->txn_sigverify_done_cnt++;
1418 0 : block->txn_sigverify_in_flight_cnt--;
1419 0 : FD_TEST( !fd_ulong_extract_bit( sched->sigverify_ready_bitset[ 0 ], exec_tile_idx ) );
1420 0 : sched->sigverify_ready_bitset[ 0 ] = fd_ulong_set_bit( sched->sigverify_ready_bitset[ 0 ], exec_tile_idx );
1421 0 : sched->metrics->txn_sigverify_done_cnt++;
1422 0 : txn_bitset_insert( sched->sigverify_done_set, txn_idx );
1423 0 : sched->txn_info_pool[ txn_idx ].flags |= FD_SCHED_TXN_SIGVERIFY_DONE;
1424 0 : if( txn_bitset_test( sched->exec_done_set, txn_idx ) && txn_bitset_test( sched->poh_mixin_done_set, txn_idx ) ) {
1425 : /* Release the txn_idx if all tasks on it are done. This is
1426 : guaranteed to only happen once per transaction because
1427 : whichever one completed first would not release. */
1428 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 1 );
1429 0 : sched->txn_pool_free_cnt++;
1430 0 : block->txn_done_cnt++;
1431 0 : sched->metrics->txn_done_cnt++;
1432 0 : }
1433 0 : break;
1434 0 : }
1435 9 : case FD_SCHED_TT_POH_HASH: {
1436 9 : block->poh_hashing_in_flight_cnt--;
1437 9 : FD_TEST( !fd_ulong_extract_bit( sched->poh_ready_bitset[ 0 ], exec_tile_idx ) );
1438 9 : sched->poh_ready_bitset[ 0 ] = fd_ulong_set_bit( sched->poh_ready_bitset[ 0 ], exec_tile_idx );
1439 9 : fd_execrp_poh_hash_done_msg_t * msg = fd_type_pun( data );
1440 9 : fd_sched_mblk_t * mblk = sched->mblk_pool+msg->mblk_idx;
1441 9 : mblk->curr_hashcnt += msg->hashcnt;
1442 9 : memcpy( mblk->curr_hash, msg->hash, sizeof(fd_hash_t) );
1443 9 : ulong hashcnt_todo = mblk->hashcnt-mblk->curr_hashcnt;
1444 9 : if( !hashcnt_todo ) {
1445 9 : block->poh_hashing_done_cnt++;
1446 9 : sched->metrics->mblk_poh_hashed_cnt++;
1447 9 : if( FD_LIKELY( !mblk->is_tick ) ) {
1448 : /* This is not a tick. Enqueue for mixin. */
1449 0 : mblk_slist_idx_push_tail( block->mblks_mixin_in_progress, msg->mblk_idx, sched->mblk_pool );
1450 9 : } else {
1451 : /* This is a tick. No need to mixin. Check the hash value
1452 : right away. */
1453 9 : block->poh_hash_cmp_done_cnt++;
1454 9 : sched->metrics->mblk_poh_done_cnt++;
1455 9 : free_mblk( sched, block, (uint)msg->mblk_idx );
1456 9 : if( FD_UNLIKELY( memcmp( mblk->curr_hash, mblk->end_hash, sizeof(fd_hash_t) ) ) ) {
1457 0 : FD_BASE58_ENCODE_32_BYTES( mblk->curr_hash->hash, our_str );
1458 0 : FD_BASE58_ENCODE_32_BYTES( mblk->end_hash->hash, ref_str );
1459 0 : FD_LOG_INFO(( "bad block: poh hash mismatch on mblk %lu, ours %s, claimed %s, hashcnt %lu, is_tick, slot %lu, parent slot %lu", msg->mblk_idx, our_str, ref_str, mblk->hashcnt, block->slot, block->parent_slot ));
1460 0 : handle_bad_block( sched, block );
1461 0 : return -1;
1462 0 : }
1463 9 : }
1464 : /* Try to drain the mixin queue. */
1465 9 : int mixin_res;
1466 9 : while( (mixin_res=maybe_mixin( sched, block )) ) {
1467 0 : if( FD_UNLIKELY( mixin_res==-1 ) ) {
1468 0 : handle_bad_block( sched, block );
1469 0 : return -1;
1470 0 : }
1471 0 : FD_TEST( mixin_res==1||mixin_res==2 );
1472 0 : }
1473 9 : } else {
1474 0 : mblk_slist_idx_push_tail( block->mblks_hashing_in_progress, msg->mblk_idx, sched->mblk_pool );
1475 0 : }
1476 9 : if( FD_UNLIKELY( verify_ticks_eager( block ) ) ) {
1477 6 : handle_bad_block( sched, block );
1478 6 : return -1;
1479 6 : }
1480 3 : break;
1481 9 : }
1482 33 : }
1483 :
1484 27 : if( FD_UNLIKELY( block->dying && !block_is_in_flight( block ) ) ) {
1485 0 : if( FD_UNLIKELY( sched->active_bank_idx==bank_idx ) ) {
1486 0 : FD_LOG_CRIT(( "invariant violation: active block %lu:%lu shouldn't be dying, parent slot %lu",
1487 0 : block->slot, bank_idx, block->parent_slot ));
1488 0 : }
1489 0 : FD_LOG_DEBUG(( "dying block %lu:%lu drained", block->slot, bank_idx ));
1490 0 : subtree_abandon( sched, block );
1491 0 : try_activate_block( sched );
1492 0 : return 0;
1493 0 : }
1494 :
1495 27 : if( FD_UNLIKELY( !block->dying && sched->active_bank_idx!=bank_idx ) ) {
1496 : /* Block is not dead. So we should be actively replaying it. */
1497 0 : fd_sched_block_t * active_block = block_pool_ele( sched, sched->active_bank_idx );
1498 0 : FD_LOG_CRIT(( "invariant violation: sched->active_bank_idx %lu, slot %lu, parent slot %lu, bank_idx %lu, slot %lu, parent slot %lu",
1499 0 : sched->active_bank_idx, active_block->slot, active_block->parent_slot,
1500 0 : bank_idx, block->slot, block->parent_slot ));
1501 0 : }
1502 :
1503 27 : maybe_switch_block( sched, bank_idx );
1504 :
1505 27 : return 0;
1506 27 : }
1507 :
1508 : void
1509 0 : fd_sched_block_abandon( fd_sched_t * sched, ulong bank_idx ) {
1510 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1511 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1512 :
1513 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1514 0 : if( FD_UNLIKELY( !block->in_sched ) ) {
1515 0 : FD_LOG_CRIT(( "invariant violation: block->in_sched==0, block %lu:%lu, parent slot %lu",
1516 0 : block->slot, bank_idx, block->parent_slot ));
1517 0 : }
1518 :
1519 0 : FD_LOG_INFO(( "abandoning block %lu:%lu", block->slot, bank_idx ));
1520 0 : sched->print_buf_sz = 0UL;
1521 0 : print_all( sched, block );
1522 0 : FD_LOG_DEBUG(( "%s", sched->print_buf ));
1523 :
1524 0 : subtree_abandon( sched, block );
1525 0 : try_activate_block( sched );
1526 0 : }
1527 :
1528 : void
1529 0 : fd_sched_cancel( fd_sched_t * sched, ulong bank_idx ) {
1530 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1531 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1532 :
1533 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1534 0 : if( FD_UNLIKELY( !block->in_sched ) ) return;
1535 0 : FD_LOG_INFO(( "canceling subtree at block %lu:%lu", block->slot, bank_idx ));
1536 0 : fd_sched_block_t * parent = block_pool_ele( sched, block->parent_idx );
1537 0 : if( FD_LIKELY( parent ) ) {
1538 : /* Splice the block out of its parent's children list. */
1539 0 : ulong block_idx = block_to_idx( sched, block );
1540 0 : ulong * idx_p = &parent->child_idx;
1541 0 : while( *idx_p!=block_idx ) {
1542 0 : idx_p = &(block_pool_ele( sched, *idx_p )->sibling_idx);
1543 0 : }
1544 0 : *idx_p = block->sibling_idx;
1545 0 : }
1546 0 : subtree_prune( sched, bank_idx, ULONG_MAX );
1547 0 : }
1548 :
1549 : void
1550 12 : fd_sched_block_add_done( fd_sched_t * sched, ulong bank_idx, ulong parent_bank_idx, ulong slot ) {
1551 12 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1552 12 : FD_TEST( bank_idx<sched->block_cnt_max );
1553 :
1554 12 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1555 12 : add_block( sched, bank_idx, parent_bank_idx );
1556 12 : block->slot = slot;
1557 12 : block->fec_eos = 1;
1558 12 : block->block_start_signaled = 1;
1559 12 : block->block_end_signaled = 1;
1560 12 : block->block_start_done = 1;
1561 12 : block->block_end_done = 1;
1562 12 : block->refcnt = 0;
1563 12 : if( FD_LIKELY( parent_bank_idx!=ULONG_MAX ) ) {
1564 0 : fd_sched_block_t * parent_block = block_pool_ele( sched, parent_bank_idx );
1565 0 : block->parent_slot = parent_block->slot;
1566 0 : }
1567 12 : if( FD_UNLIKELY( parent_bank_idx==ULONG_MAX ) ) {
1568 : /* Assumes that a NULL parent implies the snapshot slot. */
1569 12 : block->parent_slot = ULONG_MAX;
1570 12 : block->rooted = 1;
1571 12 : sched->root_idx = bank_idx;
1572 12 : }
1573 12 : }
1574 :
1575 : void
1576 0 : fd_sched_advance_root( fd_sched_t * sched, ulong root_idx ) {
1577 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1578 0 : FD_TEST( root_idx<sched->block_cnt_max );
1579 0 : FD_TEST( sched->root_idx<sched->block_cnt_max );
1580 0 : FD_TEST( ref_q_empty( sched->ref_q ) );
1581 :
1582 0 : fd_sched_block_t * new_root = block_pool_ele( sched, root_idx );
1583 0 : fd_sched_block_t * old_root = block_pool_ele( sched, sched->root_idx );
1584 0 : if( FD_UNLIKELY( !old_root->rooted ) ) {
1585 0 : FD_LOG_CRIT(( "invariant violation: old_root is not rooted, slot %lu, parent slot %lu",
1586 0 : old_root->slot, old_root->parent_slot ));
1587 0 : }
1588 :
1589 : /* Early exit if the new root is the same as the old root. */
1590 0 : if( FD_UNLIKELY( root_idx==sched->root_idx ) ) {
1591 0 : FD_LOG_INFO(( "new root is the same as the old root, slot %lu, parent slot %lu",
1592 0 : new_root->slot, new_root->parent_slot ));
1593 0 : return;
1594 0 : }
1595 :
1596 0 : subtree_prune( sched, sched->root_idx, root_idx );
1597 :
1598 0 : new_root->parent_idx = ULONG_MAX;
1599 0 : sched->root_idx = root_idx;
1600 0 : }
1601 :
1602 : void
1603 0 : fd_sched_root_notify( fd_sched_t * sched, ulong root_idx ) {
1604 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1605 0 : FD_TEST( root_idx<sched->block_cnt_max );
1606 0 : FD_TEST( sched->root_idx<sched->block_cnt_max );
1607 0 : FD_TEST( ref_q_empty( sched->ref_q ) );
1608 :
1609 0 : fd_sched_block_t * block = block_pool_ele( sched, root_idx );
1610 0 : fd_sched_block_t * old_root = block_pool_ele( sched, sched->root_idx );
1611 0 : if( FD_UNLIKELY( !old_root->rooted ) ) {
1612 0 : FD_LOG_CRIT(( "invariant violation: old_root is not rooted, slot %lu, parent slot %lu",
1613 0 : old_root->slot, old_root->parent_slot ));
1614 0 : }
1615 :
1616 : /* Early exit if the new root is the same as the old root. */
1617 0 : if( FD_UNLIKELY( root_idx==sched->root_idx ) ) {
1618 0 : FD_LOG_INFO(( "new root is the same as the old root, slot %lu, parent slot %lu",
1619 0 : block->slot, block->parent_slot ));
1620 0 : return;
1621 0 : }
1622 :
1623 : /* Mark every node from the new root up through its parents to the
1624 : old root as being rooted. */
1625 0 : fd_sched_block_t * curr = block;
1626 0 : fd_sched_block_t * prev = NULL;
1627 0 : while( curr ) {
1628 0 : if( FD_UNLIKELY( !block_is_done( curr ) ) ) {
1629 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is not done, slot %lu, parent slot %lu",
1630 0 : curr->slot, curr->parent_slot ));
1631 0 : }
1632 0 : if( FD_UNLIKELY( curr->dying ) ) {
1633 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is dying, slot %lu, parent slot %lu",
1634 0 : curr->slot, curr->parent_slot ));
1635 0 : }
1636 0 : if( FD_UNLIKELY( curr->staged ) ) {
1637 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is staged, slot %lu, parent slot %lu",
1638 0 : curr->slot, curr->parent_slot ));
1639 0 : }
1640 0 : if( FD_UNLIKELY( curr->in_rdisp ) ) {
1641 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is in the dispatcher, slot %lu, parent slot %lu",
1642 0 : curr->slot, curr->parent_slot ));
1643 0 : }
1644 0 : curr->rooted = 1;
1645 0 : prev = curr;
1646 0 : curr = block_pool_ele( sched, curr->parent_idx );
1647 0 : }
1648 :
1649 : /* If we didn't reach the old root, the new root is not a descendant. */
1650 0 : if( FD_UNLIKELY( prev!=old_root ) ) {
1651 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",
1652 0 : block->slot, block->parent_slot, old_root->slot, old_root->parent_slot ));
1653 0 : }
1654 :
1655 0 : ulong old_active_bank_idx = sched->active_bank_idx;
1656 :
1657 : /* Now traverse from old root towards new root, and abandon all
1658 : minority forks. */
1659 0 : curr = old_root;
1660 0 : while( curr && curr->rooted && curr!=block ) { /* curr!=block to avoid abandoning good forks. */
1661 0 : fd_sched_block_t * rooted_child_block = NULL;
1662 0 : ulong child_idx = curr->child_idx;
1663 0 : while( child_idx!=ULONG_MAX ) {
1664 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1665 0 : child_idx = child->sibling_idx;
1666 0 : if( child->rooted ) {
1667 0 : rooted_child_block = child;
1668 0 : } else {
1669 : /* This is a minority fork. */
1670 0 : ulong abandoned_cnt = sched->metrics->block_abandoned_cnt;
1671 0 : subtree_abandon( sched, child );
1672 0 : abandoned_cnt = sched->metrics->block_abandoned_cnt-abandoned_cnt;
1673 0 : if( FD_UNLIKELY( abandoned_cnt ) ) FD_LOG_DEBUG(( "abandoned %lu blocks on minority fork starting at block %lu:%lu", abandoned_cnt, child->slot, block_to_idx( sched, child ) ));
1674 0 : }
1675 0 : }
1676 0 : curr = rooted_child_block;
1677 0 : }
1678 :
1679 : /* If the active block got abandoned, we need to reset it. */
1680 0 : if( sched->active_bank_idx==ULONG_MAX ) {
1681 0 : sched->metrics->deactivate_pruned_cnt += fd_uint_if( old_active_bank_idx!=ULONG_MAX, 1U, 0U );
1682 0 : try_activate_block( sched );
1683 0 : }
1684 0 : }
1685 :
1686 : ulong
1687 48 : fd_sched_pruned_block_next( fd_sched_t * sched ) {
1688 48 : if( !ref_q_empty( sched->ref_q ) ) {
1689 9 : ulong bank_idx = ref_q_pop_head( sched->ref_q );
1690 9 : return bank_idx;
1691 9 : }
1692 39 : return ULONG_MAX;
1693 48 : }
1694 :
1695 : void
1696 24 : fd_sched_set_poh_params( fd_sched_t * sched, ulong bank_idx, ulong tick_height, ulong max_tick_height, ulong hashes_per_tick, fd_hash_t const * start_poh ) {
1697 24 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1698 24 : FD_TEST( bank_idx<sched->block_cnt_max );
1699 24 : FD_TEST( max_tick_height>tick_height );
1700 24 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1701 24 : block->tick_height = tick_height;
1702 24 : block->max_tick_height = max_tick_height;
1703 24 : block->hashes_per_tick = hashes_per_tick;
1704 : #if FD_SCHED_SKIP_POH
1705 : /* No-op. */
1706 : (void)start_poh;
1707 : #else
1708 24 : if( FD_LIKELY( block->mblk_cnt ) ) {
1709 : /* Fix up the first mblk's curr_hash. */
1710 9 : FD_TEST( block->mblk_unhashed_cnt );
1711 9 : FD_TEST( !mblk_slist_is_empty( block->mblks_unhashed, sched->mblk_pool ) );
1712 9 : FD_TEST( !block->mblk_freed_cnt );
1713 9 : fd_sched_mblk_t * first_mblk = sched->mblk_pool + mblk_slist_idx_peek_head( block->mblks_unhashed, sched->mblk_pool );
1714 9 : memcpy( first_mblk->curr_hash, start_poh, sizeof(fd_hash_t) );
1715 15 : } else {
1716 15 : memcpy( block->poh_hash, start_poh, sizeof(fd_hash_t) );
1717 15 : }
1718 24 : #endif
1719 24 : }
1720 :
1721 : void
1722 0 : fd_sched_set_bypass_poh_verify( fd_sched_t * sched, int bypass_poh_verify ) {
1723 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1724 0 : sched->bypass_poh_verify = !!bypass_poh_verify;
1725 0 : }
1726 :
1727 : void
1728 0 : fd_sched_set_bypass_alut_resolution( fd_sched_t * sched, int bypass_alut_resolution ) {
1729 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1730 0 : sched->bypass_alut_resolution = !!bypass_alut_resolution;
1731 0 : }
1732 :
1733 : fd_txn_p_t *
1734 0 : fd_sched_get_txn( fd_sched_t * sched, ulong txn_idx ) {
1735 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1736 0 : if( FD_UNLIKELY( txn_idx>=sched->depth ) ) {
1737 0 : return NULL;
1738 0 : }
1739 0 : return sched->txn_pool+txn_idx;
1740 0 : }
1741 :
1742 : fd_sched_txn_info_t *
1743 0 : fd_sched_get_txn_info( fd_sched_t * sched, ulong txn_idx ) {
1744 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1745 0 : if( FD_UNLIKELY( txn_idx>=sched->depth ) ) {
1746 0 : return NULL;
1747 0 : }
1748 0 : return sched->txn_info_pool+txn_idx;
1749 0 : }
1750 :
1751 : fd_hash_t *
1752 0 : fd_sched_get_poh( fd_sched_t * sched, ulong bank_idx ) {
1753 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1754 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1755 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1756 0 : FD_TEST( block->fec_eos );
1757 0 : FD_TEST( block->mblk_cnt );
1758 0 : return block->poh_hash;
1759 0 : }
1760 :
1761 : uint
1762 0 : fd_sched_get_shred_cnt( fd_sched_t * sched, ulong bank_idx ) {
1763 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1764 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1765 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1766 0 : return block->shred_cnt;
1767 0 : }
1768 :
1769 : void
1770 0 : fd_sched_metrics_write( fd_sched_t * sched ) {
1771 0 : FD_MGAUGE_SET( REPLAY, SCHED_ACTIVE_BANK_INDEX, sched->active_bank_idx );
1772 0 : FD_MGAUGE_SET( REPLAY, SCHED_LAST_DISPATCH_BANK_INDEX, sched->next_ready_last_bank_idx );
1773 0 : FD_MGAUGE_SET( REPLAY, SCHED_LAST_DISPATCH_TIMESTAMP_NANOS, fd_ulong_if( sched->next_ready_last_tick!=LONG_MAX, (ulong)sched->next_ready_last_tick, ULONG_MAX ) );
1774 0 : FD_MGAUGE_SET( REPLAY, SCHED_STAGING_LANE_OCCUPIED, (ulong)fd_ulong_popcnt( sched->staged_bitset ) );
1775 0 : FD_MGAUGE_SET( REPLAY, SCHED_STAGING_LANE_OCCUPIED_WATERMARK, sched->staged_popcnt_wmk );
1776 0 : ulong staging_lane_head_bank_idx[ FD_METRICS_ENUM_STAGING_LANE_CNT ];
1777 0 : for( ulong lane=0UL; lane<FD_METRICS_ENUM_STAGING_LANE_CNT; lane++ ) {
1778 0 : staging_lane_head_bank_idx[ lane ] = fd_ulong_if( fd_ulong_extract_bit( sched->staged_bitset, (int)lane ), sched->staged_head_bank_idx[ lane ], ULONG_MAX );
1779 0 : }
1780 0 : FD_MGAUGE_ENUM_COPY( REPLAY, SCHED_STAGING_LANE_HEAD_BANK_INDEX, staging_lane_head_bank_idx );
1781 0 : FD_MGAUGE_SET( REPLAY, SCHED_TXN_POOL_OCCUPIED, sched->depth-sched->txn_pool_free_cnt-1UL );
1782 0 : FD_MGAUGE_SET( REPLAY, SCHED_TXN_POOL_SIZE, sched->depth-1UL );
1783 0 : FD_MGAUGE_SET( REPLAY, SCHED_MICROBLOCK_POOL_OCCUPIED, sched->depth-sched->mblk_pool_free_cnt );
1784 0 : FD_MGAUGE_SET( REPLAY, SCHED_MICROBLOCK_POOL_SIZE, sched->depth );
1785 0 : FD_MGAUGE_SET( REPLAY, SCHED_BLOCK_POOL_OCCUPIED, sched->block_pool_popcnt );
1786 0 : FD_MGAUGE_SET( REPLAY, SCHED_BLOCK_POOL_SIZE, sched->block_cnt_max );
1787 :
1788 0 : ulong sched_block_added[ FD_METRICS_ENUM_SCHED_BLOCK_STAGING_CNT ];
1789 0 : sched_block_added[ FD_METRICS_ENUM_SCHED_BLOCK_STAGING_V_STAGED_IDX ] = sched->metrics->block_added_staged_cnt;
1790 0 : sched_block_added[ FD_METRICS_ENUM_SCHED_BLOCK_STAGING_V_UNSTAGED_IDX ] = sched->metrics->block_added_unstaged_cnt;
1791 0 : FD_MCNT_ENUM_COPY( REPLAY, SCHED_BLOCK_ADDED, sched_block_added );
1792 0 : FD_MCNT_SET( REPLAY, SCHED_BLOCK_REPLAYED, sched->metrics->block_removed_cnt );
1793 0 : FD_MCNT_SET( REPLAY, SCHED_BLOCK_ABANDONED, sched->metrics->block_abandoned_cnt );
1794 0 : FD_MCNT_SET( REPLAY, SCHED_BLOCK_REJECTED, sched->metrics->block_bad_cnt );
1795 0 : FD_MCNT_SET( REPLAY, SCHED_BLOCK_PROMOTED, sched->metrics->block_promoted_cnt );
1796 0 : FD_MCNT_SET( REPLAY, SCHED_BLOCK_DEMOTED, sched->metrics->block_demoted_cnt );
1797 0 : ulong sched_deactivate[ FD_METRICS_ENUM_SCHED_DEACTIVATE_REASON_CNT ];
1798 0 : sched_deactivate[ FD_METRICS_ENUM_SCHED_DEACTIVATE_REASON_V_NO_CHILD_IDX ] = sched->metrics->deactivate_no_child_cnt;
1799 0 : sched_deactivate[ FD_METRICS_ENUM_SCHED_DEACTIVATE_REASON_V_NO_WORK_IDX ] = sched->metrics->deactivate_no_txn_cnt;
1800 0 : sched_deactivate[ FD_METRICS_ENUM_SCHED_DEACTIVATE_REASON_V_ABANDONED_IDX ] = sched->metrics->deactivate_abandoned_cnt;
1801 0 : sched_deactivate[ FD_METRICS_ENUM_SCHED_DEACTIVATE_REASON_V_MINORITY_IDX ] = sched->metrics->deactivate_pruned_cnt;
1802 0 : FD_MCNT_ENUM_COPY( REPLAY, SCHED_DEACTIVATE, sched_deactivate );
1803 0 : FD_MCNT_SET( REPLAY, SCHED_LANE_SWITCHED, sched->metrics->lane_switch_cnt );
1804 0 : FD_MCNT_SET( REPLAY, SCHED_LANE_PROMOTED, sched->metrics->lane_promoted_cnt );
1805 0 : FD_MCNT_SET( REPLAY, SCHED_LANE_DEMOTED, sched->metrics->lane_demoted_cnt );
1806 0 : FD_MCNT_SET( REPLAY, SCHED_FORK_OBSERVED, sched->metrics->fork_observed_cnt );
1807 0 : ulong sched_alut[ FD_METRICS_ENUM_SCHED_ALUT_RESULT_CNT ];
1808 0 : sched_alut[ FD_METRICS_ENUM_SCHED_ALUT_RESULT_V_SUCCESS_IDX ] = sched->metrics->alut_success_cnt;
1809 0 : sched_alut[ FD_METRICS_ENUM_SCHED_ALUT_RESULT_V_FAILED_IDX ] = sched->metrics->alut_serializing_cnt;
1810 0 : FD_MCNT_ENUM_COPY( REPLAY, SCHED_ALUT, sched_alut );
1811 0 : FD_MCNT_SET( REPLAY, SCHED_TXN_PARSED_ABANDONED, sched->metrics->txn_abandoned_parsed_cnt );
1812 0 : FD_MCNT_SET( REPLAY, SCHED_TXN_EXECUTED_ABANDONED, sched->metrics->txn_abandoned_exec_done_cnt );
1813 0 : FD_MCNT_SET( REPLAY, SCHED_TXN_DONE_ABANDONED, sched->metrics->txn_abandoned_done_cnt );
1814 0 : FD_MCNT_SET( REPLAY, SCHED_WEIGHTED_IN_FLIGHT, sched->metrics->txn_weighted_in_flight_cnt );
1815 0 : FD_MCNT_SET( REPLAY, SCHED_WEIGHTED_IN_FLIGHT_DURATION_NANOS, sched->metrics->txn_weighted_in_flight_tickcount );
1816 0 : FD_MCNT_SET( REPLAY, SCHED_NONE_IN_FLIGHT_DURATION_NANOS, sched->metrics->txn_none_in_flight_tickcount );
1817 0 : FD_MCNT_SET( REPLAY, SCHED_TXN_PARSED, sched->metrics->txn_parsed_cnt );
1818 0 : FD_MCNT_SET( REPLAY, SCHED_TXN_EXECUTED, sched->metrics->txn_exec_done_cnt );
1819 0 : FD_MCNT_SET( REPLAY, SCHED_TXN_SIGNATURE_VERIFIED, sched->metrics->txn_sigverify_done_cnt );
1820 0 : FD_MCNT_SET( REPLAY, SCHED_TXN_POH_MIXED, sched->metrics->txn_mixin_done_cnt );
1821 0 : FD_MCNT_SET( REPLAY, SCHED_TXN_DONE, sched->metrics->txn_done_cnt );
1822 0 : FD_MCNT_SET( REPLAY, SCHED_MICROBLOCK_PARSED, sched->metrics->mblk_parsed_cnt );
1823 0 : FD_MCNT_SET( REPLAY, SCHED_MICROBLOCK_HASHED, sched->metrics->mblk_poh_hashed_cnt );
1824 0 : FD_MCNT_SET( REPLAY, SCHED_MICROBLOCK_DONE, sched->metrics->mblk_poh_done_cnt );
1825 0 : FD_MCNT_SET( REPLAY, SCHED_BYTES_INGESTED, sched->metrics->bytes_ingested_cnt );
1826 0 : FD_MCNT_SET( REPLAY, SCHED_BYTES_INGESTED_PADDING, sched->metrics->bytes_ingested_unparsed_cnt );
1827 0 : FD_MCNT_SET( REPLAY, SCHED_BYTES_DROPPED, sched->metrics->bytes_dropped_cnt );
1828 0 : FD_MCNT_SET( REPLAY, SCHED_FEC_INGESTED, sched->metrics->fec_cnt );
1829 0 : }
1830 :
1831 : char *
1832 9 : fd_sched_get_state_cstr( fd_sched_t * sched ) {
1833 9 : sched->print_buf_sz = 0UL;
1834 9 : print_metrics( sched );
1835 9 : print_sched( sched );
1836 9 : return sched->print_buf;
1837 9 : }
1838 :
1839 12 : void * fd_sched_leave ( fd_sched_t * sched ) { return sched; }
1840 12 : void * fd_sched_delete( void * mem ) { return mem; }
1841 :
1842 :
1843 : /* Internal helpers. */
1844 :
1845 : static void
1846 : add_block( fd_sched_t * sched,
1847 : ulong bank_idx,
1848 36 : ulong parent_bank_idx ) {
1849 36 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1850 36 : FD_TEST( !block->in_sched );
1851 36 : sched->block_pool_popcnt++;
1852 :
1853 36 : block->txn_parsed_cnt = 0U;
1854 36 : block->txn_exec_in_flight_cnt = 0U;
1855 36 : block->txn_exec_done_cnt = 0U;
1856 36 : block->txn_sigverify_in_flight_cnt = 0U;
1857 36 : block->txn_sigverify_done_cnt = 0U;
1858 36 : block->poh_hashing_in_flight_cnt = 0U;
1859 36 : block->poh_hashing_done_cnt = 0U;
1860 36 : block->poh_hash_cmp_done_cnt = 0U;
1861 36 : block->txn_done_cnt = 0U;
1862 36 : block->shred_cnt = 0U;
1863 36 : block->mblk_cnt = 0U;
1864 36 : block->mblk_freed_cnt = 0U;
1865 36 : block->mblk_tick_cnt = 0U;
1866 36 : block->mblk_unhashed_cnt = 0U;
1867 36 : block->hashcnt = 0UL;
1868 36 : block->txn_pool_max_popcnt = sched->depth - sched->txn_pool_free_cnt - 1UL;
1869 36 : block->mblk_pool_max_popcnt = sched->depth - sched->mblk_pool_free_cnt;
1870 36 : block->block_pool_max_popcnt = sched->block_pool_popcnt;
1871 :
1872 36 : mblk_slist_remove_all( block->mblks_unhashed, sched->mblk_pool );
1873 36 : mblk_slist_remove_all( block->mblks_hashing_in_progress, sched->mblk_pool );
1874 36 : mblk_slist_remove_all( block->mblks_mixin_in_progress, sched->mblk_pool );
1875 36 : block->last_mblk_is_tick = 0;
1876 36 : block->tick_hashcnt_wmk = 0UL;
1877 36 : block->curr_tick_hashcnt = 0UL;
1878 36 : block->tick_height = ULONG_MAX;
1879 36 : block->max_tick_height = ULONG_MAX;
1880 36 : block->hashes_per_tick = ULONG_MAX;
1881 36 : block->inconsistent_hashes_per_tick = 0;
1882 36 : block->zero_hash_tick = 0;
1883 :
1884 36 : block->mblks_rem = 0UL;
1885 36 : block->txns_rem = 0UL;
1886 36 : block->fec_buf_sz = 0U;
1887 36 : block->fec_buf_boff = 0U;
1888 36 : block->fec_buf_soff = 0U;
1889 36 : block->fec_eob = 0;
1890 36 : block->fec_sob = 1;
1891 :
1892 36 : block->fec_eos = 0;
1893 36 : block->rooted = 0;
1894 36 : block->dying = 0;
1895 36 : block->refcnt = 1;
1896 36 : block->in_sched = 1;
1897 36 : block->in_rdisp = 0;
1898 36 : block->block_start_signaled = 0;
1899 36 : block->block_end_signaled = 0;
1900 36 : block->block_start_done = 0;
1901 36 : block->block_end_done = 0;
1902 36 : block->staged = 0;
1903 :
1904 36 : block->luf_depth = 0UL;
1905 :
1906 : /* New leaf node, no child, no sibling. */
1907 36 : block->child_idx = ULONG_MAX;
1908 36 : block->sibling_idx = ULONG_MAX;
1909 36 : block->parent_idx = ULONG_MAX;
1910 :
1911 36 : if( FD_UNLIKELY( parent_bank_idx==ULONG_MAX ) ) {
1912 12 : return;
1913 12 : }
1914 :
1915 : /* node->parent link */
1916 24 : fd_sched_block_t * parent_block = block_pool_ele( sched, parent_bank_idx );
1917 24 : block->parent_idx = parent_bank_idx;
1918 :
1919 : /* parent->node and sibling->node links */
1920 24 : ulong child_idx = bank_idx;
1921 24 : if( FD_LIKELY( parent_block->child_idx==ULONG_MAX ) ) { /* Optimize for no forking. */
1922 12 : parent_block->child_idx = child_idx;
1923 12 : } else {
1924 12 : fd_sched_block_t * curr_block = block_pool_ele( sched, parent_block->child_idx );
1925 30 : while( curr_block->sibling_idx!=ULONG_MAX ) {
1926 18 : curr_block = block_pool_ele( sched, curr_block->sibling_idx );
1927 18 : }
1928 12 : curr_block->sibling_idx = child_idx;
1929 12 : sched->metrics->fork_observed_cnt++;
1930 12 : }
1931 :
1932 24 : if( FD_UNLIKELY( parent_block->dying ) ) {
1933 0 : block->dying = 1;
1934 0 : }
1935 24 : }
1936 :
1937 : /* Agave invokes verify_ticks() anywhere between once per slot and once
1938 : per entry batch, before tranactions are parsed or dispatched for
1939 : execution. We can't do quite the same thing due to out-of-order
1940 : scheduling and the fact that we allow parsing to run well ahead of
1941 : block boundaries. Out-of-order scheduling is good, so is overlapping
1942 : parsing with execution. The easiest thing for us would be to just
1943 : delay verify_ticks() wholesale till the end of a slot, except that
1944 : this opens us up to bogus tick and hash counts, potentially causing
1945 : runaway consumption of compute cycles. Of all the checks that are
1946 : performed in verify_ticks(), two types are relevant to mitigating
1947 : this risk. One is constraining the number of ticks, and the other is
1948 : constraining the number of hashes per tick. So we implement these
1949 : checks here, and perform them on the fly as eagerly as possible.
1950 :
1951 : Returns 0 on success. */
1952 : static int
1953 9 : verify_ticks_eager( fd_sched_block_t * block ) {
1954 9 : FD_TEST( block->hashes_per_tick!=ULONG_MAX ); /* PoH params initialized. */
1955 :
1956 9 : if( FD_UNLIKELY( block->mblk_tick_cnt+block->tick_height>block->max_tick_height ) ) {
1957 3 : FD_LOG_INFO(( "bad block: TOO_MANY_TICKS, slot %lu, parent slot %lu, tick_cnt %u, tick_height %lu, max_tick_height %lu", block->slot, block->parent_slot, block->mblk_tick_cnt, block->tick_height, block->max_tick_height ));
1958 3 : return -1;
1959 3 : }
1960 6 : if( FD_UNLIKELY( block->hashes_per_tick>1UL && block->zero_hash_tick ) ) {
1961 0 : FD_LOG_INFO(( "bad block: INVALID_TICK_HASH_COUNT, slot %lu, parent slot %lu, has at least one zero hash tick", block->slot, block->parent_slot ));
1962 0 : return -1;
1963 0 : }
1964 6 : if( FD_UNLIKELY( block->hashes_per_tick>1UL && block->mblk_tick_cnt && (block->hashes_per_tick!=block->tick_hashcnt_wmk||block->inconsistent_hashes_per_tick) ) ) {
1965 3 : FD_LOG_INFO(( "bad block: INVALID_TICK_HASH_COUNT, slot %lu, parent slot %lu, expected %lu, got %lu", block->slot, block->parent_slot, block->hashes_per_tick, block->tick_hashcnt_wmk ));
1966 3 : return -1;
1967 3 : }
1968 3 : if( FD_UNLIKELY( block->hashes_per_tick>1UL && block->curr_tick_hashcnt>block->hashes_per_tick ) ) { /* >1 to ignore low power hashing or no hashing cases */
1969 : /* We couldn't really check this at parse time because we may not
1970 : have the expected hashes per tick value yet. We couldn't delay
1971 : this till after all PoH hashing is done, because this would be a
1972 : DoS vector. This can't be merged with the above check, because a
1973 : malformed block might not end with a tick. As in, a block might
1974 : end with a non-tick microblock with a high hashcnt. Note that
1975 : checking the hashcnt between ticks transitively places an upper
1976 : bound on the hashcnt of individual microblocks, thus mitigating
1977 : the DoS vector. */
1978 0 : FD_LOG_INFO(( "bad block: INVALID_TICK_HASH_COUNT, slot %lu, parent slot %lu, observed cumulative tick_hashcnt %lu, expected %lu", block->slot, block->parent_slot, block->curr_tick_hashcnt, block->hashes_per_tick ));
1979 0 : return -1;
1980 0 : }
1981 :
1982 3 : return 0;
1983 3 : }
1984 :
1985 : /* https://github.com/anza-xyz/agave/blob/v3.0.6/ledger/src/blockstore_processor.rs#L1057
1986 :
1987 : The only check we don't do here is TRAILING_ENTRY, which can be done
1988 : independently when we parse the final FEC set of a block.
1989 :
1990 : Returns 0 on success. */
1991 : static int
1992 3 : verify_ticks_final( fd_sched_block_t * block ) {
1993 3 : FD_TEST( block->fec_eos );
1994 :
1995 3 : if( FD_UNLIKELY( block->mblk_tick_cnt+block->tick_height<block->max_tick_height ) ) {
1996 3 : FD_LOG_INFO(( "bad block: TOO_FEW_TICKS, slot %lu, parent slot %lu, tick_cnt %u, tick_height %lu, max_tick_height %lu", block->slot, block->parent_slot, block->mblk_tick_cnt, block->tick_height, block->max_tick_height ));
1997 3 : return -1;
1998 3 : }
1999 :
2000 0 : return verify_ticks_eager( block );
2001 3 : }
2002 :
2003 : int
2004 : fd_sched_block_verify_ticks( fd_sched_t * sched,
2005 : ulong bank_idx,
2006 : ulong tick_height,
2007 : ulong max_tick_height,
2008 0 : ulong hashes_per_tick ) {
2009 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
2010 0 : FD_TEST( bank_idx<sched->block_cnt_max );
2011 0 : FD_TEST( max_tick_height>tick_height );
2012 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
2013 : /* Set the tick window directly; skip fd_sched_set_poh_params
2014 : PoH fixup, unneeded here and unsafe to repeat per FEC set. */
2015 0 : block->tick_height = tick_height;
2016 0 : block->max_tick_height = max_tick_height;
2017 0 : block->hashes_per_tick = hashes_per_tick;
2018 : /* fec_eos => slot fully ingested, so the TOO_FEW check is meaningful
2019 : too. */
2020 0 : return block->fec_eos ? verify_ticks_final( block ) : verify_ticks_eager( block );
2021 0 : }
2022 :
2023 39 : #define CHECK( cond ) do { \
2024 39 : if( FD_UNLIKELY( !(cond) ) ) { \
2025 15 : return FD_SCHED_AGAIN_LATER; \
2026 15 : } \
2027 39 : } while( 0 )
2028 :
2029 : /* CHECK that it is safe to read at least n more bytes. */
2030 39 : #define CHECK_LEFT( n ) CHECK( (n)<=(block->fec_buf_sz-block->fec_buf_soff) )
2031 :
2032 : /* Consume as much as possible from the buffer. By the end of this
2033 : function, there could be residual data left over in the buffer. That
2034 : residual has to be either a partially received microblock header or a
2035 : transaction that straddles a FEC set boundary. These bytes are kept
2036 : and will be concatenated with the next FEC set. This residual is
2037 : bounded.
2038 :
2039 : Trailing bytes within a batch are dropped immediately. These
2040 : trailing bytes can span arbitrarily many FEC sets. */
2041 : FD_WARN_UNUSED static int
2042 24 : fd_sched_parse( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx ) {
2043 48 : while( 1 ) {
2044 48 : while( block->txns_rem>0UL ) {
2045 0 : int err;
2046 0 : if( FD_UNLIKELY( (err=fd_sched_parse_txn( sched, block, alut_ctx ))!=FD_SCHED_OK ) ) {
2047 0 : return err;
2048 0 : }
2049 0 : }
2050 48 : if( block->txns_rem==0UL && block->mblks_rem>0UL ) {
2051 15 : if( FD_UNLIKELY( block->mblk_cnt>=FD_SCHED_MAX_MBLK_PER_SLOT ) ) {
2052 : /* A valid block shouldn't contain more than this amount of
2053 : microblocks. */
2054 0 : FD_LOG_INFO(( "bad block: slot %lu, parent slot %lu, mblk_cnt %u (%u ticks) >= %lu", block->slot, block->parent_slot, block->mblk_cnt, block->mblk_tick_cnt, FD_SCHED_MAX_MBLK_PER_SLOT ));
2055 0 : return FD_SCHED_BAD_BLOCK;
2056 0 : }
2057 :
2058 15 : CHECK_LEFT( sizeof(fd_microblock_hdr_t) );
2059 15 : fd_microblock_hdr_t * hdr = (fd_microblock_hdr_t *)fd_type_pun( block->fec_buf+block->fec_buf_soff );
2060 15 : block->fec_buf_soff += (uint)sizeof(fd_microblock_hdr_t);
2061 :
2062 15 : if( FD_UNLIKELY( hdr->txn_cnt>fd_ulong_sat_sub( FD_MAX_TXN_PER_SLOT, block->txn_parsed_cnt ) ) ) {
2063 0 : FD_LOG_INFO(( "bad block: illegally many transactions specified in microblock header in slot %lu, parent slot %lu, txn_parsed_cnt %u, hdr->txn_cnt %lu", block->slot, block->parent_slot, block->txn_parsed_cnt, hdr->txn_cnt ));
2064 0 : return FD_SCHED_BAD_BLOCK;
2065 0 : }
2066 15 : if( FD_UNLIKELY( hdr->hash_cnt>fd_ulong_sat_sub( FD_RUNTIME_MAX_HASHES_PER_TICK, block->curr_tick_hashcnt ) ) ) {
2067 0 : FD_LOG_INFO(( "bad block: slot %lu, parent slot %lu, curr_tick_hashcnt %lu, hdr->hash_cnt %lu", block->slot, block->parent_slot, block->curr_tick_hashcnt, hdr->hash_cnt ));
2068 0 : return FD_SCHED_BAD_BLOCK;
2069 0 : }
2070 :
2071 15 : block->mblks_rem--;
2072 15 : block->txns_rem = hdr->txn_cnt;
2073 :
2074 : /* One might think that every microblock needs to have at least
2075 : one hash, otherwise the block should be considered invalid. A
2076 : vanilla validator certainly produces microblocks that conform
2077 : to this. But a modded validator could in theory produce
2078 : zero-hash microblocks. Agave's replay stage will happily take
2079 : those microblocks. The Agave implementation-defined way of
2080 : doing PoH verify is as follows:
2081 :
2082 : For a tick microblock, do the same number of hashes as
2083 : specified by the microblock. Zero hashes are not allowed.
2084 :
2085 : For a transaction microblock, if the number of hashes specified
2086 : by the microblock is <= 1, then do zero pure hashes, and simply
2087 : do a mixin/record. Otherwise, do (number of hashes-1) amount
2088 : of pure hashing, and then do a mixin. However, note that for
2089 : the purposes of tick_verify, the number of hashes specified by
2090 : the microblock is taken verbatim.
2091 :
2092 : An additional constraint is that Agave expects non-tick
2093 : microblocks to leave the cumulative tick hash count at least
2094 : one away from hashes_per_tick at the end of a batch:
2095 : https://github.com/anza-xyz/agave/blob/v4.0.0-rc.0/entry/src/entry.rs#L672
2096 :
2097 : Observe that this is not a net new constraint. Since ticks are
2098 : not allowed to have zero hashes, and ticks have to end at
2099 : exactly hashes_per_tick, this check is redundant.
2100 :
2101 :
2102 : Some additional references to Agave:
2103 :
2104 : On the consumer side, non-tick microblocks can have a zero hash
2105 : count, and the mixin will happen anyways:
2106 : https://github.com/anza-xyz/agave/blob/v4.0.0-rc.0/entry/src/entry.rs#L326
2107 : https://github.com/anza-xyz/agave/blob/v4.0.0-rc.0/entry/src/entry.rs#L542
2108 :
2109 : Ticks cannot have a zero hash count:
2110 : https://github.com/anza-xyz/agave/blob/v4.1.1/entry/src/entry.rs#L684
2111 :
2112 : On the producer side, Agave reserves at least one hash for the
2113 : tick, so an Agave produced tick would satisfy the verifier:
2114 : https://github.com/anza-xyz/agave/blob/v4.0.0-rc.0/entry/src/poh.rs#L78
2115 : https://github.com/anza-xyz/agave/blob/v4.0.0-rc.0/entry/src/poh.rs#L101 */
2116 15 : block->curr_tick_hashcnt = fd_ulong_sat_add( hdr->hash_cnt, block->curr_tick_hashcnt ); /* For tick_verify, take the number of hashes verbatim. */
2117 15 : sched->metrics->mblk_parsed_cnt++;
2118 15 : if( FD_UNLIKELY( !hdr->txn_cnt ) ) {
2119 : /* This is a tick microblock. */
2120 15 : if( FD_UNLIKELY( block->mblk_tick_cnt && block->tick_hashcnt_wmk!=block->curr_tick_hashcnt ) ) {
2121 3 : block->inconsistent_hashes_per_tick = 1;
2122 3 : if( FD_LIKELY( block->hashes_per_tick!=ULONG_MAX && block->hashes_per_tick>1UL ) ) {
2123 : /* >1 to ignore low power hashing or hashing disabled */
2124 0 : FD_LOG_INFO(( "bad block: INVALID_TICK_HASH_COUNT, slot %lu, parent slot %lu, tick idx %u, tick_hashcnt_wmk %lu, curr hashcnt %lu, hashes_per_tick %lu", block->slot, block->parent_slot, block->mblk_tick_cnt, block->tick_hashcnt_wmk, block->curr_tick_hashcnt, block->hashes_per_tick ));
2125 0 : return FD_SCHED_BAD_BLOCK;
2126 0 : }
2127 3 : }
2128 15 : if( FD_UNLIKELY( !hdr->hash_cnt ) ) {
2129 0 : block->zero_hash_tick = 1;
2130 0 : if( FD_LIKELY( block->hashes_per_tick!=ULONG_MAX && block->hashes_per_tick>1UL ) ) {
2131 0 : FD_LOG_INFO(( "bad block: INVALID_TICK_HASH_COUNT, slot %lu, parent slot %lu, tick idx %u has zero hashes", block->slot, block->parent_slot, block->mblk_tick_cnt ));
2132 0 : return FD_SCHED_BAD_BLOCK;
2133 0 : }
2134 0 : }
2135 15 : block->tick_hashcnt_wmk = fd_ulong_max( block->curr_tick_hashcnt, block->tick_hashcnt_wmk );
2136 15 : block->curr_tick_hashcnt = 0UL;
2137 15 : block->mblk_tick_cnt++;
2138 15 : }
2139 :
2140 15 : FD_TEST( sched->mblk_pool_free_cnt ); /* can_ingest should have guaranteed sufficient free capacity. */
2141 15 : uint mblk_idx = sched->mblk_pool_free_head;
2142 15 : sched->mblk_pool_free_head = sched->mblk_pool[ mblk_idx ].next;
2143 15 : sched->mblk_pool_free_cnt--;
2144 :
2145 15 : fd_sched_mblk_t * mblk = sched->mblk_pool+mblk_idx;
2146 15 : mblk->start_txn_idx = block->txn_parsed_cnt;
2147 15 : mblk->end_txn_idx = mblk->start_txn_idx+hdr->txn_cnt;
2148 15 : mblk->curr_txn_idx = mblk->start_txn_idx;
2149 15 : mblk->hashcnt = fd_ulong_sat_sub( hdr->hash_cnt, fd_ulong_if( !hdr->txn_cnt, 0UL, 1UL ) ); /* For pure hashing, implement saturating sub for non-tick microblocks. */
2150 15 : mblk->curr_hashcnt = 0UL;
2151 15 : mblk->curr_sig_cnt = 0U;
2152 15 : mblk->is_tick = !hdr->txn_cnt;
2153 15 : memcpy( mblk->end_hash, hdr->hash, sizeof(fd_hash_t) );
2154 15 : memcpy( mblk->curr_hash, block->poh_hash, sizeof(fd_hash_t) );
2155 :
2156 : /* Update block tracking. */
2157 15 : block->hashcnt += mblk->hashcnt+fd_ulong_if( !hdr->txn_cnt, 0UL, 1UL );
2158 15 : memcpy( block->poh_hash, hdr->hash, sizeof(fd_hash_t) );
2159 15 : block->last_mblk_is_tick = mblk->is_tick;
2160 15 : block->mblk_cnt++;
2161 :
2162 : #if FD_SCHED_SKIP_POH
2163 : block->poh_hashing_done_cnt++;
2164 : block->poh_hash_cmp_done_cnt++;
2165 : free_mblk( sched, block, mblk_idx );
2166 : #else
2167 15 : mblk_slist_idx_push_tail( block->mblks_unhashed, mblk_idx, sched->mblk_pool );
2168 15 : block->mblk_unhashed_cnt++;
2169 15 : #endif
2170 15 : continue;
2171 15 : }
2172 33 : if( block->txns_rem==0UL && block->mblks_rem==0UL && block->fec_sob ) {
2173 24 : CHECK_LEFT( sizeof(ulong) );
2174 9 : FD_TEST( block->fec_buf_soff==0U );
2175 9 : block->mblks_rem = FD_LOAD( ulong, block->fec_buf );
2176 9 : block->fec_buf_soff += (uint)sizeof(ulong);
2177 :
2178 9 : if( FD_UNLIKELY( block->mblks_rem>fd_ulong_sat_sub( FD_SCHED_MAX_MBLK_PER_SLOT, block->mblk_cnt ) ) ) {
2179 0 : FD_LOG_INFO(( "bad block: slot %lu, parent slot %lu, mblk_cnt %u (%u ticks), hdr->mblk_cnt %lu >= %lu", block->slot, block->parent_slot, block->mblk_cnt, block->mblk_tick_cnt, block->mblks_rem, FD_SCHED_MAX_MBLK_PER_SLOT ));
2180 0 : return FD_SCHED_BAD_BLOCK;
2181 0 : }
2182 :
2183 9 : block->fec_sob = 0;
2184 9 : if( FD_UNLIKELY( !block->mblks_rem ) ) {
2185 0 : FD_LOG_INFO(( "bad block: slot %lu, parent slot %lu, mblk_cnt %u (%u ticks), empty batch detected", block->slot, block->parent_slot, block->mblk_cnt, block->mblk_tick_cnt ));
2186 0 : return FD_SCHED_BAD_BLOCK;
2187 0 : }
2188 9 : continue;
2189 9 : }
2190 9 : if( block->txns_rem==0UL && block->mblks_rem==0UL ) {
2191 9 : break;
2192 9 : }
2193 9 : }
2194 9 : if( !block->fec_sob && block->txns_rem==0UL && block->mblks_rem==0UL ) {
2195 : /* All microblocks announced by the current batch header have been
2196 : parsed out. Anything still in the buffer must be trailing bytes.
2197 : Drop them now because trailing bytes can span many FEC sets and
2198 : accumulating them would overflow the parse buffer. Any followup
2199 : FEC sets within the same batch will simply be discarded
2200 : wholesale. */
2201 9 : sched->metrics->bytes_ingested_unparsed_cnt += block->fec_buf_sz-block->fec_buf_soff;
2202 9 : block->fec_buf_boff += block->fec_buf_sz;
2203 9 : block->fec_buf_soff = 0U;
2204 9 : block->fec_buf_sz = 0U;
2205 9 : }
2206 9 : if( block->fec_eob ) {
2207 9 : block->fec_sob = 1;
2208 9 : block->fec_eob = 0;
2209 9 : }
2210 9 : return FD_SCHED_OK;
2211 24 : }
2212 :
2213 : FD_WARN_UNUSED static int
2214 0 : fd_sched_parse_txn( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx ) {
2215 0 : fd_txn_t * txn = fd_type_pun( block->txn );
2216 :
2217 0 : uchar * payload = block->fec_buf+block->fec_buf_soff;
2218 0 : ulong pay_sz = 0UL;
2219 0 : ulong txn_sz = fd_txn_parse_core( payload,
2220 0 : fd_ulong_min( FD_TXN_MTU, block->fec_buf_sz-block->fec_buf_soff ),
2221 0 : txn,
2222 0 : NULL,
2223 0 : &pay_sz );
2224 :
2225 0 : if( FD_UNLIKELY( !pay_sz || !txn_sz ) ) {
2226 : /* Can't parse out a full transaction. */
2227 0 : return FD_SCHED_AGAIN_LATER;
2228 0 : }
2229 :
2230 0 : if( FD_UNLIKELY( block->txn_parsed_cnt>=FD_MAX_TXN_PER_SLOT ) ) {
2231 : /* The block contains more transactions than a valid block would.
2232 : Mark the block dead instead of keep processing it. */
2233 0 : FD_LOG_INFO(( "bad block: illegally many transactions in slot %lu, parent slot %lu, txn_parsed_cnt %u", block->slot, block->parent_slot, block->txn_parsed_cnt ));
2234 0 : return FD_SCHED_BAD_BLOCK;
2235 0 : }
2236 :
2237 0 : ulong imm_cnt = fd_txn_account_cnt( txn, FD_TXN_ACCT_CAT_IMM );
2238 0 : ulong alt_cnt = fd_txn_account_cnt( txn, FD_TXN_ACCT_CAT_ALT );
2239 :
2240 : /* Try to expand ALUTs. */
2241 0 : int serializing = 0;
2242 0 : if( alt_cnt>0UL ) {
2243 0 : if( FD_UNLIKELY( sched->bypass_alut_resolution ) ) {
2244 : /* test/fuzz: no accdb to query, so treat ALUT txns as serializing. */
2245 0 : serializing = 1;
2246 0 : } else {
2247 : /* Copy the slot hashes sysvar out and release the read BEFORE
2248 : resolving the ALTs. fd_runtime_load_txn_address_lookup_tables
2249 : issues its own fd_accdb_read_one per lookup table, and the accdb
2250 : acquire state is a single non-nestable flag — holding this read
2251 : open across that call would trip the IDLE assertion in
2252 : fd_accdb_acquire. The slot_hashes view aliases the record data,
2253 : so it must view the copy, not the released record. */
2254 0 : static uchar slot_hashes_buf[ FD_SYSVAR_SLOT_HASHES_BINCODE_SZ ];
2255 0 : ulong slot_hashes_sz = 0UL;
2256 0 : int have_slot_hashes = 0;
2257 0 : fd_acc_t ro = fd_accdb_read_one( alut_ctx->accdb, alut_ctx->fork_id, fd_sysvar_slot_hashes_id.uc );
2258 0 : if( FD_LIKELY( ro.lamports && ro.data_len<=sizeof(slot_hashes_buf) ) ) {
2259 0 : fd_memcpy( slot_hashes_buf, ro.data, ro.data_len );
2260 0 : slot_hashes_sz = ro.data_len;
2261 0 : have_slot_hashes = 1;
2262 0 : }
2263 0 : fd_accdb_unread_one( alut_ctx->accdb, &ro );
2264 :
2265 0 : fd_slot_hashes_t slot_hashes_view[1];
2266 0 : if( FD_LIKELY( have_slot_hashes &&
2267 0 : fd_sysvar_slot_hashes_view( slot_hashes_view, slot_hashes_buf, slot_hashes_sz ) ) ) {
2268 0 : serializing = !!fd_runtime_load_txn_address_lookup_tables( txn, payload, alut_ctx->accdb, alut_ctx->fork_id, alut_ctx->els, slot_hashes_view, sched->aluts );
2269 0 : sched->metrics->alut_success_cnt += (uint)!serializing;
2270 0 : } else {
2271 0 : serializing = 1;
2272 0 : }
2273 0 : }
2274 0 : }
2275 :
2276 : /* Transactions should not have duplicate accounts.
2277 : https://github.com/anza-xyz/agave/blob/v3.1.11/ledger/src/blockstore_processor.rs#L778-L790 */
2278 0 : fd_acct_addr_t const * imms = fd_txn_get_acct_addrs( txn, payload );
2279 0 : fd_acct_addr_t * alts = (!alt_cnt||serializing) ? NULL : sched->aluts;
2280 0 : alt_cnt = alts ? alt_cnt : 0UL;
2281 0 : if( FD_UNLIKELY( fd_chkdup_check( sched->chkdup, imms, imm_cnt, alts, alt_cnt ) ) ) {
2282 0 : FD_LOG_INFO(( "bad block: duplicate accounts in slot %lu, parent slot %lu, txn_parsed_cnt %u", block->slot, block->parent_slot, block->txn_parsed_cnt ));
2283 0 : return FD_SCHED_BAD_BLOCK;
2284 0 : }
2285 :
2286 0 : ulong bank_idx = (ulong)(block-sched->block_pool);
2287 0 : ulong txn_idx = fd_rdisp_add_txn( sched->rdisp, bank_idx, txn, payload, alts, serializing );
2288 0 : FD_TEST( txn_idx!=0UL );
2289 0 : sched->metrics->txn_parsed_cnt++;
2290 0 : sched->metrics->alut_serializing_cnt += (uint)serializing;
2291 0 : sched->txn_pool_free_cnt--;
2292 0 : fd_txn_p_t * txn_p = sched->txn_pool + txn_idx;
2293 0 : txn_p->payload_sz = pay_sz;
2294 :
2295 0 : txn_p->start_shred_idx = (ushort)fd_sort_up_uint_split( block->shred_blk_offs, block->shred_cnt, block->fec_buf_boff+block->fec_buf_soff );
2296 0 : txn_p->start_shred_idx = fd_ushort_if( txn_p->start_shred_idx>0U, (ushort)(txn_p->start_shred_idx-1U), txn_p->start_shred_idx );
2297 0 : txn_p->end_shred_idx = (ushort)fd_sort_up_uint_split( block->shred_blk_offs, block->shred_cnt, block->fec_buf_boff+block->fec_buf_soff+(uint)pay_sz );
2298 :
2299 0 : fd_memcpy( txn_p->payload, payload, pay_sz );
2300 0 : fd_memcpy( TXN(txn_p), txn, txn_sz );
2301 0 : txn_bitset_remove( sched->exec_done_set, txn_idx );
2302 0 : txn_bitset_remove( sched->sigverify_done_set, txn_idx );
2303 0 : txn_bitset_remove( sched->poh_mixin_done_set, txn_idx );
2304 0 : sched->txn_info_pool[ txn_idx ].flags = 0UL;
2305 0 : sched->txn_info_pool[ txn_idx ].txn_err = 0;
2306 0 : sched->txn_info_pool[ txn_idx ].tick_parsed = fd_tickcount();
2307 0 : sched->txn_info_pool[ txn_idx ].tick_sigverify_disp = LONG_MAX;
2308 0 : sched->txn_info_pool[ txn_idx ].tick_sigverify_done = LONG_MAX;
2309 0 : sched->txn_info_pool[ txn_idx ].tick_exec_disp = LONG_MAX;
2310 0 : sched->txn_info_pool[ txn_idx ].tick_exec_done = LONG_MAX;
2311 0 : sched->txn_info_pool[ txn_idx ].index_in_slot = block->txn_parsed_cnt;
2312 0 : block->txn_idx[ block->txn_parsed_cnt ] = txn_idx;
2313 0 : block->fec_buf_soff += (uint)pay_sz;
2314 0 : block->txn_parsed_cnt++;
2315 : #if FD_SCHED_SKIP_SIGVERIFY
2316 : txn_bitset_insert( sched->sigverify_done_set, txn_idx );
2317 : block->txn_sigverify_done_cnt++;
2318 : #endif
2319 : #if FD_SCHED_SKIP_POH
2320 : txn_bitset_insert( sched->poh_mixin_done_set, txn_idx );
2321 : #endif
2322 0 : block->txns_rem--;
2323 0 : return FD_SCHED_OK;
2324 0 : }
2325 :
2326 : #undef CHECK
2327 : #undef CHECK_LEFT
2328 :
2329 : static void
2330 0 : dispatch_sigverify( fd_sched_t * sched, fd_sched_block_t * block, ulong bank_idx, int exec_tile_idx, fd_sched_task_t * out ) {
2331 : /* Dispatch transactions for sigverify in parse order. */
2332 0 : out->task_type = FD_SCHED_TT_TXN_SIGVERIFY;
2333 0 : out->txn_sigverify->bank_idx = bank_idx;
2334 0 : out->txn_sigverify->txn_idx = block->txn_idx[ block->txn_sigverify_done_cnt+block->txn_sigverify_in_flight_cnt ];
2335 0 : out->txn_sigverify->exec_idx = (ulong)exec_tile_idx;
2336 0 : sched->sigverify_ready_bitset[ 0 ] = fd_ulong_clear_bit( sched->sigverify_ready_bitset[ 0 ], exec_tile_idx );
2337 0 : sched->tile_to_bank_idx[ exec_tile_idx ] = bank_idx;
2338 0 : block->txn_sigverify_in_flight_cnt++;
2339 0 : if( FD_UNLIKELY( (~sched->txn_exec_ready_bitset[ 0 ])&(~sched->sigverify_ready_bitset[ 0 ])&(~sched->poh_ready_bitset[ 0 ])&fd_ulong_mask_lsb( (int)sched->exec_cnt ) ) ) FD_LOG_CRIT(( "invariant violation: txn_exec_ready_bitset 0x%lx sigverify_ready_bitset 0x%lx poh_ready_bitset 0x%lx", sched->txn_exec_ready_bitset[ 0 ], sched->sigverify_ready_bitset[ 0 ], sched->poh_ready_bitset[ 0 ] ));
2340 0 : }
2341 :
2342 : /* Assumes there is a PoH task available for dispatching. */
2343 : static void
2344 9 : dispatch_poh( fd_sched_t * sched, fd_sched_block_t * block, ulong bank_idx, int exec_tile_idx, fd_sched_task_t * out ) {
2345 9 : fd_sched_mblk_t * mblk = NULL;
2346 9 : uint mblk_idx;
2347 9 : if( FD_LIKELY( !mblk_slist_is_empty( block->mblks_hashing_in_progress, sched->mblk_pool ) ) ) {
2348 : /* There's a PoH task in progress, just continue working on that. */
2349 0 : mblk_idx = (uint)mblk_slist_idx_pop_head( block->mblks_hashing_in_progress, sched->mblk_pool );
2350 0 : mblk = sched->mblk_pool+mblk_idx;
2351 9 : } else {
2352 : /* No in progress PoH task, so start a new one. */
2353 9 : FD_TEST( block->mblk_unhashed_cnt );
2354 9 : mblk_idx = (uint)mblk_slist_idx_pop_head( block->mblks_unhashed, sched->mblk_pool );
2355 9 : mblk = sched->mblk_pool+mblk_idx;
2356 9 : block->mblk_unhashed_cnt--;
2357 9 : }
2358 9 : out->task_type = FD_SCHED_TT_POH_HASH;
2359 9 : out->poh_hash->bank_idx = bank_idx;
2360 9 : out->poh_hash->mblk_idx = mblk_idx;
2361 9 : out->poh_hash->exec_idx = (ulong)exec_tile_idx;
2362 9 : ulong hashcnt_todo = mblk->hashcnt-mblk->curr_hashcnt;
2363 9 : out->poh_hash->hashcnt = fd_ulong_min( hashcnt_todo, FD_SCHED_MAX_POH_HASHES_PER_TASK );
2364 9 : memcpy( out->poh_hash->hash, mblk->curr_hash, sizeof(fd_hash_t) );
2365 9 : sched->poh_ready_bitset[ 0 ] = fd_ulong_clear_bit( sched->poh_ready_bitset[ 0 ], exec_tile_idx );
2366 9 : sched->tile_to_bank_idx[ exec_tile_idx ] = bank_idx;
2367 9 : block->poh_hashing_in_flight_cnt++;
2368 9 : if( FD_UNLIKELY( (~sched->txn_exec_ready_bitset[ 0 ])&(~sched->sigverify_ready_bitset[ 0 ])&(~sched->poh_ready_bitset[ 0 ])&fd_ulong_mask_lsb( (int)sched->exec_cnt ) ) ) FD_LOG_CRIT(( "invariant violation: txn_exec_ready_bitset 0x%lx sigverify_ready_bitset 0x%lx poh_ready_bitset 0x%lx", sched->txn_exec_ready_bitset[ 0 ], sched->sigverify_ready_bitset[ 0 ], sched->poh_ready_bitset[ 0 ] ));
2369 9 : }
2370 :
2371 : /* Does up to one transaction mixin. Returns 1 if one mixin was done, 2
2372 : if that mixin also completed a microblock, 0 if no transaction mixin
2373 : was available, -1 if there is a PoH verify error. */
2374 : FD_WARN_UNUSED static int
2375 33 : maybe_mixin( fd_sched_t * sched, fd_sched_block_t * block ) {
2376 33 : if( FD_UNLIKELY( mblk_slist_is_empty( block->mblks_mixin_in_progress, sched->mblk_pool ) ) ) return 0;
2377 0 : FD_TEST( block->poh_hashing_done_cnt-block->poh_hash_cmp_done_cnt>0 );
2378 :
2379 : /* The microblock we would like to do mixin on is at the head of the
2380 : queue. It may have had some mixin, it may have never had any
2381 : mixin. In the case of the former, we should continue to mixin the
2382 : same head microblock until it's done, lest the per-block bmtree
2383 : gets clobbered when we start a new one. */
2384 0 : ulong mblk_idx = mblk_slist_idx_pop_head( block->mblks_mixin_in_progress, sched->mblk_pool );
2385 0 : fd_sched_mblk_t * mblk = sched->mblk_pool+mblk_idx;
2386 :
2387 0 : if( FD_UNLIKELY( mblk->end_txn_idx>block->txn_parsed_cnt ) ) {
2388 : /* A partially parsed microblock is by definition at the end of the
2389 : FEC stream. If such a microblock is in progress, there should be
2390 : no other microblock in this block so far that hasn't been
2391 : dispatched, because microblocks are dispatched in parse order. */
2392 0 : if( FD_UNLIKELY( block->mblk_unhashed_cnt ) ) {
2393 0 : sched->print_buf_sz = 0UL;
2394 0 : print_all( sched, block );
2395 0 : FD_LOG_CRIT(( "invariant violation end_txn_idx %lu: %s", mblk->end_txn_idx, sched->print_buf ));
2396 0 : }
2397 :
2398 : /* If we've decided to start mixin on a partially parsed microblock,
2399 : there better be nothing else in-progress. Otherwise, they might
2400 : clobber the per-block bmtree for mixin. */
2401 0 : if( FD_UNLIKELY( mblk->curr_txn_idx!=mblk->start_txn_idx && (block->poh_hashing_in_flight_cnt||!mblk_slist_is_empty( block->mblks_hashing_in_progress, sched->mblk_pool )||!mblk_slist_is_empty( block->mblks_mixin_in_progress, sched->mblk_pool )) ) ) {
2402 0 : sched->print_buf_sz = 0UL;
2403 0 : print_all( sched, block );
2404 0 : FD_LOG_CRIT(( "invariant violation end_txn_idx %lu start_txn_idx %lu curr_txn_idx %lu: %s", mblk->end_txn_idx, mblk->start_txn_idx, mblk->curr_txn_idx, sched->print_buf ));
2405 0 : }
2406 0 : }
2407 :
2408 : /* Very rarely, we've finished hashing, but not all transactions in
2409 : the microblock have been parsed out. This can happen if we haven't
2410 : received all the FEC sets for this microblock. We can't yet fully
2411 : mixin the microblock. So we'll stick it back into the end of the
2412 : queue, and try to see if there's a fully parsed microblock.
2413 : Unless, there's truly nothing else to mixin. Then we would start
2414 : mixin with the partially parsed microblock. We do this because the
2415 : txn pool is meant to be an OOO scheduling window not tied to
2416 : max_live_slots sizing requirements, so there shouldn't be a way for
2417 : external input to tie up txn pool entries for longer than
2418 : necessary. */
2419 0 : if( FD_UNLIKELY( mblk->curr_txn_idx>=block->txn_parsed_cnt || /* Nothing more to mixin for this microblock. */
2420 0 : (mblk->end_txn_idx>block->txn_parsed_cnt && /* There is something to mixin, but the microblock isn't fully parsed yet ... */
2421 0 : mblk->curr_txn_idx==mblk->start_txn_idx && /* ... and we haven't started mixin on it yet ... */
2422 0 : (block->poh_hashing_in_flight_cnt || /* ... and another microblock is in-progress and might preempt this microblock and clobber the bmtree, so we shouldn't start the partial microblock just yet. */
2423 0 : !mblk_slist_is_empty( block->mblks_hashing_in_progress, sched->mblk_pool ) ||
2424 0 : !mblk_slist_is_empty( block->mblks_mixin_in_progress, sched->mblk_pool ))) ) ) {
2425 0 : mblk_slist_idx_push_tail( block->mblks_mixin_in_progress, mblk_idx, sched->mblk_pool );
2426 :
2427 : /* No other microblock in the mixin queue. */
2428 0 : if( FD_UNLIKELY( block->poh_hashing_done_cnt-block->poh_hash_cmp_done_cnt==1 ) ) return 0;
2429 :
2430 : /* At this point, there's at least one more microblock in the mixin
2431 : queue we could try. It's a predecessor (in parse order) that
2432 : finished hashing later than the partially parsed microblock at
2433 : the head of the mixin queue. */
2434 :
2435 : /* It should never clobber the bmtree for a microblock that has had some mixin done on it. */
2436 0 : if( FD_UNLIKELY( mblk->curr_txn_idx!=mblk->start_txn_idx ) ) {
2437 0 : sched->print_buf_sz = 0UL;
2438 0 : print_all( sched, block );
2439 0 : FD_LOG_CRIT(( "invariant violation curr_txn_idx %lu start_txn_idx %lu: %s", mblk->curr_txn_idx, mblk->start_txn_idx, sched->print_buf ));
2440 0 : }
2441 :
2442 0 : mblk_idx = mblk_slist_idx_pop_head( block->mblks_mixin_in_progress, sched->mblk_pool );
2443 0 : mblk = sched->mblk_pool+mblk_idx;
2444 :
2445 : /* It should be a fresh microblock for mixin. */
2446 0 : FD_TEST( mblk->curr_txn_idx==mblk->start_txn_idx );
2447 : /* Invariant: at any given point in time, there can be at most one
2448 : microblock that hasn't been fully parsed yet, due to the nature
2449 : of sequential parsing. So this microblock has to be fully
2450 : parsed. */
2451 0 : FD_TEST( mblk->end_txn_idx<=block->txn_parsed_cnt );
2452 0 : }
2453 :
2454 0 : FD_TEST( mblk->curr_txn_idx<mblk->end_txn_idx );
2455 :
2456 : /* Now mixin. */
2457 0 : if( FD_LIKELY( mblk->curr_txn_idx==mblk->start_txn_idx ) ) block->bmtree = fd_bmtree_commit_init( block->bmtree_mem, 32UL, 1UL, 0UL ); /* Optimize for single-transaction microblocks, which are the majority. */
2458 :
2459 0 : ulong txn_gidx = block->txn_idx[ mblk->curr_txn_idx ];
2460 0 : fd_txn_p_t * _txn = sched->txn_pool+txn_gidx;
2461 0 : fd_txn_t * txn = TXN(_txn);
2462 0 : for( ulong j=0; j<txn->signature_cnt; j++ ) {
2463 0 : fd_bmtree_node_t node[ 1 ];
2464 0 : fd_bmtree_hash_leaf( node, _txn->payload+txn->signature_off+FD_TXN_SIGNATURE_SZ*j, 64UL, 1UL );
2465 0 : fd_bmtree_commit_append( block->bmtree, node, 1UL );
2466 0 : mblk->curr_sig_cnt++;
2467 0 : }
2468 :
2469 : /* Release the txn_idx. */
2470 0 : txn_bitset_insert( sched->poh_mixin_done_set, txn_gidx );
2471 0 : sched->metrics->txn_mixin_done_cnt++;
2472 0 : if( txn_bitset_test( sched->exec_done_set, txn_gidx ) && txn_bitset_test( sched->sigverify_done_set, txn_gidx ) ) {
2473 0 : fd_rdisp_complete_txn( sched->rdisp, txn_gidx, 1 );
2474 0 : sched->txn_pool_free_cnt++;
2475 0 : block->txn_done_cnt++;
2476 0 : sched->metrics->txn_done_cnt++;
2477 0 : }
2478 :
2479 0 : mblk->curr_txn_idx++;
2480 0 : int rv = 2;
2481 0 : if( FD_LIKELY( mblk->curr_txn_idx==mblk->end_txn_idx ) ) {
2482 : /* Ready to compute the final hash for this microblock. */
2483 0 : block->poh_hash_cmp_done_cnt++;
2484 0 : sched->metrics->mblk_poh_done_cnt++;
2485 0 : uchar * root = fd_bmtree_commit_fini( block->bmtree );
2486 0 : uchar mixin_buf[ 64 ];
2487 0 : fd_memcpy( mixin_buf, mblk->curr_hash, 32UL );
2488 0 : fd_memcpy( mixin_buf+32UL, root, 32UL );
2489 0 : fd_sha256_hash( mixin_buf, 64UL, mblk->curr_hash );
2490 0 : free_mblk( sched, block, (uint)mblk_idx );
2491 : /* Bypass PoH verification for fuzzing/testing throughput. */
2492 0 : if( FD_UNLIKELY( !sched->bypass_poh_verify && memcmp( mblk->curr_hash, mblk->end_hash, sizeof(fd_hash_t) ) ) ) {
2493 0 : FD_BASE58_ENCODE_32_BYTES( mblk->curr_hash->hash, our_str );
2494 0 : FD_BASE58_ENCODE_32_BYTES( mblk->end_hash->hash, ref_str );
2495 0 : FD_LOG_INFO(( "bad block: poh hash mismatch on mblk %lu, ours %s, claimed %s, hashcnt %lu, txns [%lu,%lu), %u sigs, slot %lu, parent slot %lu", mblk_idx, our_str, ref_str, mblk->hashcnt, mblk->start_txn_idx, mblk->end_txn_idx, mblk->curr_sig_cnt, block->slot, block->parent_slot ));
2496 0 : return -1;
2497 0 : }
2498 0 : } else {
2499 : /* There are more transactions to mixin in this microblock. */
2500 0 : mblk_slist_idx_push_head( block->mblks_mixin_in_progress, mblk_idx, sched->mblk_pool );
2501 0 : rv = 1;
2502 0 : }
2503 :
2504 0 : return rv;
2505 0 : }
2506 :
2507 : static void
2508 48 : try_activate_block( fd_sched_t * sched ) {
2509 : /* Early return if there's already an active block. */
2510 48 : if( FD_LIKELY( sched->active_bank_idx!=ULONG_MAX ) ) return;
2511 :
2512 : /* See if there are any allocated staging lanes that we can activate
2513 : for scheduling ... */
2514 48 : ulong staged_bitset = sched->staged_bitset;
2515 111 : while( staged_bitset ) {
2516 84 : int lane_idx = fd_ulong_find_lsb( staged_bitset );
2517 84 : staged_bitset = fd_ulong_pop_lsb( staged_bitset );
2518 :
2519 84 : ulong head_idx = sched->staged_head_bank_idx[ lane_idx ];
2520 84 : fd_sched_block_t * head_block = block_pool_ele( sched, head_idx );
2521 84 : fd_sched_block_t * parent_block = block_pool_ele( sched, head_block->parent_idx );
2522 : //FIXME: restore these invariant checks when we have immediate demotion of dying blocks
2523 : //Today, dying blocks can remain staged if they have in-flight transactions.
2524 : // if( FD_UNLIKELY( parent_block->dying ) ) {
2525 : // /* Invariant: no child of a dying block should be staged. */
2526 : // 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",
2527 : // head_idx, head_block->slot, head_block->parent_slot, lane_idx, parent_block->slot, parent_block->parent_slot ));
2528 : // }
2529 : // if( FD_UNLIKELY( head_block->dying ) ) {
2530 : // /* Invariant: no dying block should be staged. */
2531 : // FD_LOG_CRIT(( "invariant violation: staged_head_bank_idx %lu, slot %lu, prime %lu on lane %u has head_block->dying set",
2532 : // head_idx, (ulong)head_block->block_id.slot, (ulong)head_block->block_id.prime, lane_idx ));
2533 : // }
2534 84 : if( block_is_done( parent_block ) && block_is_activatable( head_block ) ) {
2535 : /* ... Yes, on this staging lane the parent block is done. So we
2536 : can activate the staged child. */
2537 21 : if( FD_UNLIKELY( head_idx!=sched->last_active_bank_idx ) ) { /* Unlikely because only possible under forking or on slot boundary. */
2538 21 : if( FD_UNLIKELY( sched->last_active_bank_idx!=head_block->parent_idx ) ) { /* Forking is rare. */
2539 21 : FD_LOG_DEBUG(( "activating block %lu:%lu: lane switch to %d", head_block->slot, head_idx, lane_idx ));
2540 21 : sched->metrics->lane_switch_cnt++;
2541 21 : } else {
2542 0 : FD_LOG_DEBUG(( "activating block %lu:%lu: lane %d waking up on slot boundary", head_block->slot, head_idx, lane_idx ));
2543 0 : }
2544 21 : }
2545 21 : sched->active_bank_idx = head_idx;
2546 21 : return;
2547 21 : }
2548 84 : }
2549 :
2550 : /* ... No, promote unstaged blocks. */
2551 27 : ulong root_idx = sched->root_idx;
2552 27 : if( FD_UNLIKELY( root_idx==ULONG_MAX ) ) {
2553 0 : FD_LOG_CRIT(( "invariant violation: root_idx==ULONG_MAX indicating fd_sched is uninitialized" ));
2554 0 : }
2555 : /* Find and stage the longest stageable unstaged fork. This is a
2556 : policy decision. */
2557 27 : ulong depth = compute_longest_unstaged_fork( sched, root_idx );
2558 27 : if( FD_LIKELY( depth>0UL ) ) {
2559 3 : if( FD_UNLIKELY( sched->staged_bitset==fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) ) ) {
2560 : /* No more staging lanes available. All of them are occupied by
2561 : slow squatters. Only empty blocks can be demoted, and so
2562 : blocks with in-flight transactions, including dying in-flight
2563 : blocks, shouldn't be demoted. We demote all demotable lanes.
2564 : Demotion isn't all that expensive, since demotable blocks have
2565 : no transactions in them. If a demoted block proves to be
2566 : active still, it'll naturally promote back into a staging lane.
2567 :
2568 : In fact, all lanes should be demotable at this point. None of
2569 : the lanes have anything dispatchable, otherwise we would have
2570 : simply activated one of the dispatchable lanes. None of the
2571 : lanes have anything in-flight either, as we allow for a grace
2572 : period while something is in-flight, before we deactivate any
2573 : block. In principle, we could get rid of the grace period and
2574 : deactivate right away. In that case, it's okay if nothing is
2575 : demotable at the moment, as that simply implies that all lanes
2576 : have in-flight tasks. We would get another chance to try to
2577 : demote when the last in-flight task on any lane completes.
2578 :
2579 : Another interesting side effect of the current dispatching and
2580 : lane switching policy is that each lane should have exactly one
2581 : block in it at this point. A parent block by definition can't
2582 : be partially ingested. Any parent block that is fully ingested
2583 : and dispatchable would have made the lane dispatchable, and we
2584 : wouldn't be here. Any parent that is fully ingested and fully
2585 : dispatched would be fully done after the grace period. So
2586 : there could only be one block per lane, and it is
2587 : simultaneously the head and the tail of the lane.
2588 :
2589 : A note on why this whole thing does not deadlock:
2590 :
2591 : One might reasonably wonder what happens if all the lanes are
2592 : non-empty, non-dead, but for some reason couldn't be activated
2593 : for dispatching. We would deadlock in this case, as no lane
2594 : dispatches to the point of being demotable, and no unstaged
2595 : block can be promoted. Such is not in fact possible. The only
2596 : way a dispatchable lane can be ineligible for activation is if
2597 : it has a parent block that isn't done yet. So a deadlock
2598 : happens when this parent block, or any of its dispatchable
2599 : ancestors, is unstaged. An important invariant we maintain is
2600 : that a staged block can't have an unstaged stageable parent.
2601 : This invariant, by induction, gives us the guarantee that at
2602 : least one of the lanes can be activated. */
2603 15 : for( int l=0; l<(int)FD_SCHED_MAX_STAGING_LANES; l++ ) {
2604 : /* We would be able to assert that all lanes are demotable,
2605 : except that abandoned blocks are given no grace period for
2606 : deactivation. So there could be lanes transiently occupied
2607 : by dying blocks that are neither demotable (due to in-flight
2608 : tasks) nor activatable (due to being dying). If
2609 : rdisp_demote() supported non-empty blocks, then we could
2610 : probably restore the assertion. */
2611 12 : if( FD_UNLIKELY( !lane_is_demotable( sched, l ) ) ) continue;
2612 12 : ulong demoted_cnt = demote_lane( sched, l );
2613 12 : if( FD_UNLIKELY( demoted_cnt!=1UL ) ) {
2614 0 : FD_LOG_CRIT(( "invariant violation: %lu blocks demoted from lane %d, expected 1 demotion", demoted_cnt, l ));
2615 0 : }
2616 12 : sched->metrics->lane_demoted_cnt++;
2617 12 : }
2618 : /* We weren't able to successfully demote anything. This is
2619 : likely because all lanes are occupied by dying blocks with
2620 : in-flight tasks. We would get another chance to try to demote
2621 : when the last in-flight task on any lane completes. */
2622 3 : if( FD_UNLIKELY( sched->staged_bitset==fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) ) ) return;
2623 3 : }
2624 3 : FD_TEST( sched->staged_bitset!=fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) );
2625 3 : int lane_idx = fd_ulong_find_lsb( ~sched->staged_bitset );
2626 3 : if( FD_UNLIKELY( lane_idx>=(int)FD_SCHED_MAX_STAGING_LANES ) ) {
2627 0 : FD_LOG_CRIT(( "invariant violation: lane_idx %d, sched->staged_bitset %lx",
2628 0 : lane_idx, sched->staged_bitset ));
2629 0 : }
2630 3 : ulong head_bank_idx = stage_longest_unstaged_fork( sched, root_idx, lane_idx );
2631 3 : if( FD_UNLIKELY( head_bank_idx==ULONG_MAX ) ) {
2632 : /* We found a promotable fork depth>0. This should not happen. */
2633 0 : FD_LOG_CRIT(( "invariant violation: head_bank_idx==ULONG_MAX" ));
2634 0 : }
2635 : /* We don't bother with promotion unless the block is immediately
2636 : dispatchable. So it's okay to set the active block here. This
2637 : doesn't cause out-of-order block replay because any parent block
2638 : must be fully done. If the parent block were dead, this fork
2639 : would be marked dead too and ineligible for promotion. If the
2640 : parent block were not dead and not done and staged, we wouldn't
2641 : be trying to promote an unstaged fork. If the parent block were
2642 : not dead and not done and unstaged, it would've been part of this
2643 : unstaged fork. */
2644 3 : fd_sched_block_t * head_block = block_pool_ele( sched, head_bank_idx );
2645 3 : FD_LOG_DEBUG(( "activating block %lu:%lu: unstaged promotion to lane %d", head_block->slot, head_bank_idx, lane_idx ));
2646 3 : sched->active_bank_idx = head_bank_idx;
2647 3 : return;
2648 3 : }
2649 : /* No unstaged blocks to promote. So we're done. Yay. */
2650 27 : }
2651 :
2652 : static void
2653 33 : check_or_set_active_block( fd_sched_t * sched ) {
2654 33 : if( FD_UNLIKELY( sched->active_bank_idx==ULONG_MAX ) ) {
2655 33 : try_activate_block( sched );
2656 33 : } else {
2657 0 : fd_sched_block_t * active_block = block_pool_ele( sched, sched->active_bank_idx );
2658 0 : if( FD_UNLIKELY( block_should_deactivate( active_block ) ) ) {
2659 0 : sched->print_buf_sz = 0UL;
2660 0 : print_all( sched, active_block );
2661 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
2662 0 : FD_LOG_CRIT(( "invariant violation: should have been deactivated" ));
2663 0 : }
2664 0 : }
2665 33 : }
2666 :
2667 : /* This function has two main jobs:
2668 : - Mark everything on the fork tree dying.
2669 : - Take blocks out of rdisp if possible. */
2670 : static void
2671 9 : subtree_mark_and_maybe_prune_rdisp( fd_sched_t * sched, fd_sched_block_t * block ) {
2672 9 : if( FD_UNLIKELY( block->rooted ) ) {
2673 0 : FD_LOG_CRIT(( "invariant violation: rooted block should not be abandoned, slot %lu, parent slot %lu",
2674 0 : block->slot, block->parent_slot ));
2675 0 : }
2676 : /* All minority fork nodes pass through this function eventually. So
2677 : this is a good point to check per-node invariants for minority
2678 : forks. */
2679 9 : if( FD_UNLIKELY( block->staged && !block->in_rdisp ) ) {
2680 0 : FD_LOG_CRIT(( "invariant violation: staged block is not in the dispatcher, slot %lu, parent slot %lu",
2681 0 : block->slot, block->parent_slot ));
2682 0 : }
2683 :
2684 : /* Setting the flag is non-optional and can happen more than once. */
2685 9 : block->dying = 1;
2686 :
2687 : /* Removal from dispatcher should only happen once. */
2688 9 : if( block->in_rdisp ) {
2689 9 : fd_sched_block_t * parent = block_pool_ele( sched, block->parent_idx );
2690 9 : if( FD_UNLIKELY( !parent ) ) {
2691 : /* Only the root has no parent. Abandon should never be called on
2692 : the root. So any block we are trying to abandon should have a
2693 : parent. */
2694 0 : FD_LOG_CRIT(( "invariant violation: parent not found slot %lu, parent slot %lu",
2695 0 : block->slot, block->parent_slot ));
2696 0 : }
2697 :
2698 : /* The dispatcher expects blocks to be abandoned in the same order
2699 : that they were added on each lane. There are no requirements on
2700 : the order of abandoning if two blocks are not on the same lane,
2701 : or if a block is unstaged. This means that in general we
2702 : shouldn't abandon a child block if the parent hasn't been
2703 : abandoned yet, if and only if they are on the same lane. So wait
2704 : until we can abandon the parent, and then descend down the fork
2705 : tree to ensure orderly abandoning. */
2706 9 : int in_order = !parent->in_rdisp || /* parent is not in the dispatcher */
2707 9 : !parent->staged || /* parent is in the dispatcher but not staged */
2708 9 : !block->staged || /* parent is in the dispatcher and staged but this block is unstaged */
2709 9 : block->staging_lane!=parent->staging_lane; /* this block is on a different staging lane than its parent */
2710 :
2711 9 : if( FD_UNLIKELY( in_order && block->staged && sched->active_bank_idx==sched->staged_head_bank_idx[ block->staging_lane ] && sched->active_bank_idx!=ULONG_MAX ) ) {
2712 9 : FD_TEST( block_pool_ele( sched, sched->active_bank_idx )==block );
2713 9 : FD_LOG_DEBUG(( "reset active_bank_idx %lu: abandon", sched->active_bank_idx ));
2714 9 : sched->last_active_bank_idx = sched->active_bank_idx;
2715 9 : sched->active_bank_idx = ULONG_MAX;
2716 9 : sched->metrics->deactivate_abandoned_cnt++;
2717 9 : }
2718 :
2719 : /* We inform the dispatcher of an abandon only when there are no
2720 : more in-flight transactions. Otherwise, if the dispatcher
2721 : recycles the same txn_id that was just abandoned, and we receive
2722 : completion of an in-flight transaction whose txn_id was just
2723 : recycled. */
2724 : // FIXME The recycling might be fine now that we no longer use
2725 : // txn_id to index into anything. We might be able to just drop
2726 : // txn_id on abandoned blocks. Though would this leak transaction
2727 : // content if the txn_id is recycled?
2728 : // Note that subtree pruning from sched isn't dependent on the
2729 : // in-flight check being present here, as is_prunable already checks
2730 : // for in-flight==0.
2731 9 : int abandon = in_order && !block_is_in_flight( block );
2732 :
2733 9 : if( abandon ) {
2734 9 : block->in_rdisp = 0;
2735 9 : fd_rdisp_abandon_block( sched->rdisp, (ulong)(block-sched->block_pool) );
2736 9 : sched->txn_pool_free_cnt += block->txn_parsed_cnt-block->txn_done_cnt; /* in_flight_cnt==0 */
2737 :
2738 9 : sched->metrics->block_abandoned_cnt++;
2739 9 : sched->metrics->txn_abandoned_parsed_cnt += block->txn_parsed_cnt;
2740 9 : sched->metrics->txn_abandoned_exec_done_cnt += block->txn_exec_done_cnt;
2741 9 : sched->metrics->txn_abandoned_done_cnt += block->txn_done_cnt;
2742 :
2743 : //FIXME when demote supports non-empty blocks, we should demote
2744 : //the block from the lane unconditionally and immediately,
2745 : //regardles of whether it's safe to abandon or not. So a block
2746 : //would go immediately from staged to unstaged and eventually to
2747 : //abandoned.
2748 9 : if( FD_LIKELY( block->staged ) ) {
2749 9 : FD_LOG_DEBUG(( "block %lu:%lu exited lane %lu: abandon", block->slot, block_to_idx( sched, block ), block->staging_lane ));
2750 9 : block->staged = 0;
2751 : /* Now release the staging lane. This will release the lane as
2752 : soon as we abandon the head block on a lane. Technically a
2753 : release should only happen when we remove the tail block on a
2754 : lane. This is fine though. The way we abandon guarantees by
2755 : induction that an entire lane will be abandoned. Only the
2756 : head block on a lane can possibly have in-flight
2757 : transactions, and so once a head block becomes eligible for
2758 : abandoning, the entire lane all the way to the tail block,
2759 : will be eligible. */
2760 9 : sched->staged_bitset = fd_ulong_clear_bit( sched->staged_bitset, (int)block->staging_lane );
2761 9 : sched->staged_head_bank_idx[ block->staging_lane ] = ULONG_MAX;
2762 9 : }
2763 9 : }
2764 9 : }
2765 :
2766 : /* Abandon the entire fork chaining off of this block. */
2767 9 : ulong child_idx = block->child_idx;
2768 9 : while( child_idx!=ULONG_MAX ) {
2769 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
2770 0 : subtree_mark_and_maybe_prune_rdisp( sched, child );
2771 0 : child_idx = child->sibling_idx;
2772 0 : }
2773 9 : }
2774 :
2775 : /* This function tells sched that the subtree is no longer worth
2776 : replaying. It can happen as a result of one of the following.
2777 : - Bad block at head of lane.
2778 : - Bad block at tail of lane.
2779 : - Root notify.
2780 :
2781 : It's safe to call this function more than once on the same block. */
2782 : static void
2783 9 : subtree_abandon( fd_sched_t * sched, fd_sched_block_t * block ) {
2784 9 : subtree_mark_and_maybe_prune_rdisp( sched, block );
2785 : /* We do not gate refcnt releases on the entire subtree being
2786 : prunable. In the root notify case there can be in-flight tasks in
2787 : the middle of a subtree, and gating on whole-subtree prunability
2788 : would defer the release of an otherwise prunable block, e.g. an
2789 : inactive sibling of an in-flight block, to a later abandon such as
2790 : a subsequent root notify. That could stall bank reclamation.
2791 :
2792 : The sched fork tree still stays in sync with the banks tree. This
2793 : is to prevent attacks targeting lifetime discrepancy. So while we
2794 : know that the subtree can be pruned from sched at this point, we
2795 : only release refcnts here and stop short of actually pruning, which
2796 : remains solely initiated by banks. */
2797 9 : subtree_release_refcnt( sched, block );
2798 9 : }
2799 :
2800 : /* Release the bank refcnt for every block in the subtree that sched is
2801 : done with. This is evaluated per block. A block's refcnt is
2802 : released as soon as that block individually qualifies, independent of
2803 : whether its siblings or descendants still have in-flight tasks. This
2804 : is safe because the actual pruning is initiated separately by banks,
2805 : and gated on the whole subtree's refcnts reaching zero. Releasing
2806 : refcnts eagerly per block get us there sooner. Blocks that are still
2807 : in-flight will be released when their last task drains. This
2808 : function is idempotent. */
2809 : static void
2810 9 : subtree_release_refcnt( fd_sched_t * sched, fd_sched_block_t * block ) {
2811 9 : FD_TEST( block->in_sched );
2812 9 : if( block->refcnt && block_is_prunable( block ) ) {
2813 9 : FD_TEST( block->dying ); /* The happy path releases the refcnt on full replay. Only bad blocks end up here. */
2814 9 : FD_TEST( !block->block_end_done ); /* Implied by an outstanding refcnt. The happy path releases the refcnt on full replay. */
2815 9 : block->refcnt = 0;
2816 9 : FD_TEST( ref_q_avail( sched->ref_q ) );
2817 9 : ref_q_push_tail( sched->ref_q, block_to_idx( sched, block ) );
2818 9 : if( FD_LIKELY( block->block_start_done ) ) FD_LOG_DEBUG(( "block %lu:%lu replayed partially, releasing refcnt without full replay", block->slot, block_to_idx( sched, block ) ));
2819 0 : else FD_LOG_DEBUG(( "block %lu:%lu replayed nothing, releasing refcnt without any replay", block->slot, block_to_idx( sched, block ) ));
2820 9 : }
2821 :
2822 9 : ulong child_idx = block->child_idx;
2823 9 : while( child_idx!=ULONG_MAX ) {
2824 0 : fd_sched_block_t * child_block = block_pool_ele( sched, child_idx );
2825 0 : subtree_release_refcnt( sched, child_block );
2826 0 : child_idx = child_block->sibling_idx;
2827 0 : }
2828 9 : }
2829 :
2830 : static void
2831 0 : subtree_prune( fd_sched_t * sched, ulong bank_idx, ulong except_idx ) {
2832 0 : fd_sched_block_t * head = block_pool_ele( sched, bank_idx );
2833 0 : head->parent_idx = ULONG_MAX;
2834 0 : fd_sched_block_t * tail = head;
2835 :
2836 0 : while( head ) {
2837 0 : FD_TEST( !head->refcnt );
2838 0 : FD_TEST( head->in_sched );
2839 0 : head->in_sched = 0;
2840 :
2841 0 : ulong child_idx = head->child_idx;
2842 0 : while( child_idx!=ULONG_MAX ) {
2843 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
2844 : /* Add children to be visited. We abuse the parent_idx field to
2845 : link up the next block to visit. */
2846 0 : if( child_idx!=except_idx ) {
2847 0 : tail->parent_idx = child_idx;
2848 0 : tail = child;
2849 0 : tail->parent_idx = ULONG_MAX;
2850 0 : }
2851 0 : child_idx = child->sibling_idx;
2852 0 : }
2853 :
2854 : /* Prune the current block. We will never publish halfway into a
2855 : staging lane, because anything that we are publishing away should
2856 : be out of the dispatcher at this point, much less staged.
2857 : - Anything on the rooted fork should have finished replaying
2858 : gracefully and be out of the dispatcher.
2859 : - Anything on minority forks should have been marked dying and
2860 : be taken out of the dispatcher.
2861 :
2862 : There should be no more in-flight tasks either. Prunes are
2863 : initiated by banks, and the refcnt on the bank won't drop to zero
2864 : unless all in-flight tasks have been drained. So the fact that
2865 : we're pruning implies that the bank thinks there's nothing more
2866 : in-flight. */
2867 0 : if( FD_UNLIKELY( block_is_in_flight( head ) ) ) {
2868 0 : FD_LOG_CRIT(( "invariant violation: block has tasks in flight (%u exec %u sigverify %u poh), slot %lu, parent slot %lu",
2869 0 : head->txn_exec_in_flight_cnt, head->txn_sigverify_in_flight_cnt, head->poh_hashing_in_flight_cnt, head->slot, head->parent_slot ));
2870 0 : }
2871 0 : if( FD_UNLIKELY( head->in_rdisp ) ) {
2872 : /* We should have removed it from the dispatcher when we were
2873 : notified of the new root, or when in-flight transactions were
2874 : drained. */
2875 0 : FD_LOG_CRIT(( "invariant violation: block is in the dispatcher, slot %lu, parent slot %lu", head->slot, head->parent_slot ));
2876 0 : }
2877 :
2878 : /* Return remaining mblk descriptors to the shared pool. */
2879 0 : free_mblk_slist( sched, head, head->mblks_unhashed );
2880 0 : free_mblk_slist( sched, head, head->mblks_hashing_in_progress );
2881 0 : free_mblk_slist( sched, head, head->mblks_mixin_in_progress );
2882 :
2883 0 : if( FD_UNLIKELY( !head->block_end_done ) ) {
2884 0 : sched->print_buf_sz = 0UL;
2885 0 : print_block_metrics( sched, head );
2886 0 : if( FD_LIKELY( head->block_start_done ) ) FD_LOG_DEBUG(( "block %lu:%lu replayed partially, pruning without full replay: %s", head->slot, block_to_idx( sched, head ), sched->print_buf ));
2887 0 : else FD_LOG_DEBUG(( "block %lu:%lu replayed nothing, pruning without any replay: %s", head->slot, block_to_idx( sched, head ), sched->print_buf ));
2888 0 : }
2889 :
2890 0 : sched->block_pool_popcnt--;
2891 :
2892 0 : fd_sched_block_t * next = block_pool_ele( sched, head->parent_idx );
2893 :
2894 : /* We don't have to clear the indices here since no one should be
2895 : accessing them. Defensive programming. */
2896 0 : head->parent_idx = ULONG_MAX;
2897 0 : head->child_idx = ULONG_MAX;
2898 0 : head->sibling_idx = ULONG_MAX;
2899 :
2900 0 : head = next;
2901 0 : }
2902 0 : }
2903 :
2904 : static void
2905 27 : maybe_switch_block( fd_sched_t * sched, ulong bank_idx ) {
2906 : /* This only happens rarely when there are dying in-flight blocks.
2907 : Early exit and don't let dying blocks affect replay. */
2908 27 : if( FD_UNLIKELY( bank_idx!=sched->active_bank_idx ) ) return;
2909 :
2910 27 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
2911 27 : if( FD_UNLIKELY( block_is_done( block ) ) ) {
2912 0 : fd_rdisp_remove_block( sched->rdisp, bank_idx );
2913 0 : FD_LOG_DEBUG(( "block %lu:%lu exited lane %lu: remove", block->slot, bank_idx, block->staging_lane ));
2914 0 : block->in_rdisp = 0;
2915 0 : block->staged = 0;
2916 0 : sched->metrics->block_removed_cnt++;
2917 0 : FD_LOG_DEBUG(( "reset active_bank_idx %lu: remove", sched->active_bank_idx ));
2918 0 : sched->last_active_bank_idx = sched->active_bank_idx;
2919 0 : sched->active_bank_idx = ULONG_MAX;
2920 :
2921 : /* See if there is a child block down the same staging lane. This
2922 : is a policy decision to minimize fork churn. We could in theory
2923 : reevaluate staging lane allocation here and do promotion/demotion
2924 : as needed. */
2925 0 : ulong child_idx = block->child_idx;
2926 0 : while( child_idx!=ULONG_MAX ) {
2927 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
2928 0 : if( FD_LIKELY( child->staged && child->staging_lane==block->staging_lane ) ) {
2929 : /* There is a child block down the same staging lane ... */
2930 0 : if( FD_LIKELY( !child->dying ) ) {
2931 : /* ... and the child isn't dead */
2932 0 : if( FD_UNLIKELY( !block_is_activatable( child ) ) ) {
2933 : /* ... but the child is not activatable, likely because
2934 : there are no transactions available yet. */
2935 0 : sched->metrics->deactivate_no_txn_cnt++;
2936 0 : try_activate_block( sched );
2937 0 : return;
2938 0 : }
2939 : /* ... and it's immediately dispatchable, so switch the active
2940 : block to it, and have the child inherit the head status of
2941 : the lane. This is the common case. */
2942 0 : FD_LOG_DEBUG(( "activating block %lu:%lu: child inheritance on lane %lu", child->slot, child_idx, child->staging_lane ));
2943 0 : sched->active_bank_idx = child_idx;
2944 0 : sched->staged_head_bank_idx[ block->staging_lane ] = child_idx;
2945 0 : if( FD_UNLIKELY( !fd_ulong_extract_bit( sched->staged_bitset, (int)block->staging_lane ) ) ) {
2946 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",
2947 0 : sched->staged_bitset, block->staging_lane, block->slot, block->parent_slot, child->slot, child->parent_slot ));
2948 0 : }
2949 0 : return;
2950 0 : } else {
2951 : /* ... but the child block is considered dead, likely because
2952 : the parser considers it invalid. */
2953 0 : FD_LOG_INFO(( "child block %lu is already dead", child->slot ));
2954 0 : subtree_abandon( sched, child );
2955 0 : break;
2956 0 : }
2957 0 : }
2958 0 : child_idx = child->sibling_idx;
2959 0 : }
2960 : /* There isn't a child block down the same staging lane. This is
2961 : the last block in the staging lane. Release the staging lane. */
2962 0 : sched->staged_bitset = fd_ulong_clear_bit( sched->staged_bitset, (int)block->staging_lane );
2963 0 : sched->staged_head_bank_idx[ block->staging_lane ] = ULONG_MAX;
2964 0 : sched->metrics->deactivate_no_child_cnt++;
2965 0 : try_activate_block( sched );
2966 27 : } else if( block_should_deactivate( block ) ) {
2967 : /* We exhausted the active block, but it's not fully done yet. We
2968 : are just not getting FEC sets for it fast enough. This could
2969 : happen when the network path is congested, or when the leader
2970 : simply went down. Reset the active block. */
2971 15 : sched->last_active_bank_idx = sched->active_bank_idx;
2972 15 : sched->active_bank_idx = ULONG_MAX;
2973 15 : sched->metrics->deactivate_no_txn_cnt++;
2974 15 : try_activate_block( sched );
2975 15 : }
2976 27 : }
2977 :
2978 : FD_FN_UNUSED static ulong
2979 0 : find_and_stage_longest_unstaged_fork( fd_sched_t * sched, int lane_idx ) {
2980 0 : ulong root_idx = sched->root_idx;
2981 0 :
2982 0 : if( FD_UNLIKELY( root_idx==ULONG_MAX ) ) {
2983 0 : FD_LOG_CRIT(( "invariant violation: root_idx==ULONG_MAX indicating fd_sched is uninitialized" ));
2984 0 : }
2985 0 :
2986 0 : /* First pass: compute the longest unstaged fork depth for each node
2987 0 : in the fork tree. */
2988 0 : ulong depth = compute_longest_unstaged_fork( sched, root_idx );
2989 0 :
2990 0 : /* Second pass: stage blocks on the longest unstaged fork. */
2991 0 : ulong head_bank_idx = stage_longest_unstaged_fork( sched, root_idx, lane_idx );
2992 0 :
2993 0 : if( FD_UNLIKELY( (depth>0UL && head_bank_idx==ULONG_MAX) || (depth==0UL && head_bank_idx!=ULONG_MAX) ) ) {
2994 0 : FD_LOG_CRIT(( "invariant violation: depth %lu, head_bank_idx %lu",
2995 0 : depth, head_bank_idx ));
2996 0 : }
2997 0 :
2998 0 : return head_bank_idx;
2999 0 : }
3000 :
3001 : /* Returns length of the longest stageable unstaged fork, if there is
3002 : one, and 0 otherwise. */
3003 : static ulong
3004 96 : compute_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx ) {
3005 96 : if( FD_UNLIKELY( bank_idx==ULONG_MAX ) ) {
3006 0 : FD_LOG_CRIT(( "invariant violation: bank_idx==ULONG_MAX" ));
3007 0 : }
3008 :
3009 96 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
3010 :
3011 96 : ulong max_child_depth = 0UL;
3012 96 : ulong child_idx = block->child_idx;
3013 165 : while( child_idx!=ULONG_MAX ) {
3014 69 : ulong child_depth = compute_longest_unstaged_fork( sched, child_idx );
3015 69 : if( child_depth > max_child_depth ) {
3016 3 : max_child_depth = child_depth;
3017 3 : }
3018 69 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
3019 69 : child_idx = child->sibling_idx;
3020 69 : }
3021 :
3022 96 : block->luf_depth = max_child_depth + fd_ulong_if( block_is_promotable( block ), 1UL, 0UL );
3023 96 : return block->luf_depth;
3024 96 : }
3025 :
3026 : static ulong
3027 6 : stage_longest_unstaged_fork_helper( fd_sched_t * sched, ulong bank_idx, int lane_idx ) {
3028 6 : if( FD_UNLIKELY( bank_idx==ULONG_MAX ) ) {
3029 0 : FD_LOG_CRIT(( "invariant violation: bank_idx==ULONG_MAX" ));
3030 0 : }
3031 :
3032 6 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
3033 :
3034 6 : int stage_it = fd_int_if( block_is_promotable( block ), 1, 0 );
3035 6 : ulong rv = fd_ulong_if( stage_it, bank_idx, ULONG_MAX );
3036 6 : if( FD_LIKELY( stage_it ) ) {
3037 3 : block->staged = 1;
3038 3 : block->staging_lane = (ulong)lane_idx;
3039 3 : fd_rdisp_promote_block( sched->rdisp, bank_idx, block->staging_lane );
3040 3 : sched->metrics->block_promoted_cnt++;
3041 3 : FD_LOG_DEBUG(( "block %lu:%lu entered lane %lu: promote", block->slot, bank_idx, block->staging_lane ));
3042 3 : }
3043 :
3044 : /* Base case: leaf node. */
3045 6 : if( block->child_idx==ULONG_MAX ) return rv;
3046 :
3047 3 : ulong max_depth = 0UL;
3048 3 : ulong best_child_idx = ULONG_MAX;
3049 3 : ulong child_idx = block->child_idx;
3050 18 : while( child_idx!=ULONG_MAX ) {
3051 15 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
3052 15 : if( child->luf_depth>max_depth ) {
3053 3 : max_depth = child->luf_depth;
3054 3 : best_child_idx = child_idx;
3055 3 : }
3056 15 : child_idx = child->sibling_idx;
3057 15 : }
3058 :
3059 : /* Recursively stage descendants. */
3060 3 : if( best_child_idx!=ULONG_MAX ) {
3061 3 : ulong head_bank_idx = stage_longest_unstaged_fork_helper( sched, best_child_idx, lane_idx );
3062 3 : rv = fd_ulong_if( rv!=ULONG_MAX, rv, head_bank_idx );
3063 3 : }
3064 :
3065 3 : return rv;
3066 6 : }
3067 :
3068 : /* Returns idx of head block of staged lane on success, idx_null
3069 : otherwise. */
3070 : static ulong
3071 3 : stage_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx, int lane_idx ) {
3072 3 : ulong head_bank_idx = stage_longest_unstaged_fork_helper( sched, bank_idx, lane_idx );
3073 3 : if( FD_LIKELY( head_bank_idx!=ULONG_MAX ) ) {
3074 3 : sched->metrics->lane_promoted_cnt++;
3075 3 : sched->staged_bitset = fd_ulong_set_bit( sched->staged_bitset, lane_idx );
3076 : /* No need to update staged_popcnt_wmk because the fact that there
3077 : are unstaged blocks implies we already maxed out lanes at one
3078 : point. */
3079 3 : sched->staged_head_bank_idx[ lane_idx ] = head_bank_idx;
3080 3 : }
3081 3 : return head_bank_idx;
3082 3 : }
3083 :
3084 : /* Check if an entire staging lane can be demoted. Returns 1 if all
3085 : blocks in the lane are demotable, 0 otherwise. */
3086 : static int
3087 12 : lane_is_demotable( fd_sched_t * sched, int lane_idx ) {
3088 12 : ulong bank_idx = sched->staged_head_bank_idx[ lane_idx ];
3089 :
3090 24 : while( bank_idx!=ULONG_MAX ) {
3091 12 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
3092 12 : FD_TEST( block->staged );
3093 12 : FD_TEST( block->staging_lane==(ulong)lane_idx );
3094 :
3095 12 : if( FD_UNLIKELY( !block_is_demotable( block ) ) ) {
3096 : /* Found a non-demotable block. Early exit. */
3097 0 : return 0;
3098 0 : }
3099 :
3100 : /* Find the child in the same staging lane. */
3101 12 : ulong child_idx = block->child_idx;
3102 12 : ulong next_bank_idx = ULONG_MAX;
3103 12 : while( child_idx!=ULONG_MAX ) {
3104 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
3105 0 : if( child->staged && child->staging_lane==(ulong)lane_idx ) {
3106 0 : next_bank_idx = child_idx;
3107 0 : break;
3108 0 : }
3109 0 : child_idx = child->sibling_idx;
3110 0 : }
3111 12 : bank_idx = next_bank_idx;
3112 12 : }
3113 :
3114 12 : return 1;
3115 12 : }
3116 :
3117 : /* Demote all blocks in a staging lane. Assumes that all blocks in the
3118 : lane are demotable. Returns the number of blocks demoted. */
3119 : static ulong
3120 12 : demote_lane( fd_sched_t * sched, int lane_idx ) {
3121 12 : ulong bank_idx = sched->staged_head_bank_idx[ lane_idx ];
3122 12 : uint demoted_cnt = 0U;
3123 :
3124 24 : while( bank_idx!=ULONG_MAX ) {
3125 12 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
3126 12 : FD_TEST( block->staged );
3127 12 : FD_TEST( block->staging_lane==(ulong)lane_idx );
3128 :
3129 12 : int ret = fd_rdisp_demote_block( sched->rdisp, bank_idx );
3130 12 : if( FD_UNLIKELY( ret!=0 ) ) {
3131 0 : FD_LOG_CRIT(( "fd_rdisp_demote_block failed for block %lu:%lu, lane %d", block->slot, bank_idx, lane_idx ));
3132 0 : }
3133 12 : FD_LOG_DEBUG(( "block %lu:%lu exited lane %lu: demote", block->slot, bank_idx, block->staging_lane ));
3134 12 : block->staged = 0;
3135 12 : demoted_cnt++;
3136 :
3137 : /* Find the child in the same staging lane. */
3138 12 : ulong child_idx = block->child_idx;
3139 12 : ulong next_bank_idx = ULONG_MAX;
3140 12 : while( child_idx!=ULONG_MAX ) {
3141 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
3142 0 : if( child->staged && child->staging_lane==(ulong)lane_idx ) {
3143 0 : next_bank_idx = child_idx;
3144 0 : break;
3145 0 : }
3146 0 : child_idx = child->sibling_idx;
3147 0 : }
3148 12 : bank_idx = next_bank_idx;
3149 12 : }
3150 :
3151 : /* Clear the lane. */
3152 12 : sched->staged_bitset = fd_ulong_clear_bit( sched->staged_bitset, lane_idx );
3153 12 : sched->staged_head_bank_idx[ lane_idx ] = ULONG_MAX;
3154 :
3155 12 : sched->metrics->block_demoted_cnt += demoted_cnt;
3156 12 : FD_LOG_DEBUG(( "demoted %u blocks in lane %d", demoted_cnt, lane_idx ));
3157 12 : return demoted_cnt;
3158 12 : }
|
