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 "../../util/math/fd_stat.h" /* for sorted search */
6 : #include "../../disco/fd_disco_base.h" /* for FD_MAX_TXN_PER_SLOT */
7 : #include "../../flamenco/accdb/fd_accdb_impl_v1.h"
8 : #include "../../flamenco/runtime/fd_runtime.h" /* for fd_runtime_load_txn_address_lookup_tables */
9 : #include "../../flamenco/runtime/sysvar/fd_sysvar_slot_hashes.h" /* for ALUTs */
10 :
11 0 : #define FD_SCHED_MAX_DEPTH (FD_RDISP_MAX_DEPTH>>2)
12 0 : #define FD_SCHED_MAX_STAGING_LANES_LOG (2)
13 0 : #define FD_SCHED_MAX_STAGING_LANES (1UL<<FD_SCHED_MAX_STAGING_LANES_LOG)
14 : #define FD_SCHED_MAX_EXEC_TILE_CNT (64UL)
15 0 : #define FD_SCHED_MAX_PRINT_BUF_SZ (2UL<<20)
16 :
17 : /* 64 ticks per slot, and a single gigantic microblock containing min
18 : size transactions. */
19 : 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 );
20 :
21 : /* We size the buffer to be able to hold residual data from the previous
22 : FEC set that only becomes parseable after the next FEC set is
23 : ingested, as well as the incoming FEC set. The largest minimally
24 : parseable unit of data is a transaction. So that much data may
25 : straddle FEC set boundaries. Other minimally parseable units of data
26 : include the microblock header and the microblock count within a
27 : batch. */
28 0 : #define FD_SCHED_MAX_PAYLOAD_PER_FEC (FD_STORE_DATA_MAX)
29 : #define FD_SCHED_MAX_FEC_BUF_SZ (FD_SCHED_MAX_PAYLOAD_PER_FEC+FD_TXN_MTU)
30 : FD_STATIC_ASSERT( FD_TXN_MTU>=sizeof(fd_microblock_hdr_t), resize buffer for residual data );
31 : FD_STATIC_ASSERT( FD_TXN_MTU>=sizeof(ulong), resize buffer for residual data );
32 :
33 0 : #define FD_SCHED_MAX_TXN_PER_FEC ((FD_SCHED_MAX_PAYLOAD_PER_FEC-1UL)/FD_TXN_MIN_SERIALIZED_SZ+1UL) /* 478 */
34 :
35 0 : #define FD_SCHED_MAGIC (0xace8a79c181f89b6UL) /* echo -n "fd_sched_v0" | sha512sum | head -c 16 */
36 :
37 0 : #define FD_SCHED_PARSER_OK (0)
38 0 : #define FD_SCHED_PARSER_AGAIN_LATER (1)
39 0 : #define FD_SCHED_PARSER_BAD_BLOCK (2)
40 :
41 :
42 : /* Structs. */
43 :
44 : #define SET_NAME txn_bitset
45 : #define SET_MAX FD_SCHED_MAX_DEPTH
46 : #include "../../util/tmpl/fd_set.c"
47 :
48 : struct fd_sched_block {
49 : ulong slot;
50 : ulong parent_slot;
51 : ulong parent_idx; /* Index of the parent in the pool. */
52 : ulong child_idx; /* Index of the left-child in the pool. */
53 : ulong sibling_idx; /* Index of the right-sibling in the pool. */
54 :
55 : /* Counters. */
56 : uint txn_parsed_cnt;
57 : /* txn_queued_cnt = txn_parsed_cnt-txn_in_flight_cnt-txn_done_cnt */
58 : uint txn_exec_in_flight_cnt;
59 : uint txn_exec_done_cnt;
60 : uint txn_sigverify_in_flight_cnt;
61 : uint txn_sigverify_done_cnt;
62 : uint txn_done_cnt; /* A transaction is considered done when all types of tasks associated with it are done. */
63 : ulong txn_pool_max_popcnt; /* Peak transaction pool occupancy during the time this block was replaying. */
64 : ulong block_pool_max_popcnt; /* Peak block pool occupancy. */
65 : uint shred_cnt;
66 : uint fec_cnt;
67 : ulong txn_idx[ FD_MAX_TXN_PER_SLOT ]; /* Indexed by parse order. */
68 :
69 : /* Parser state. */
70 : uchar txn[ FD_TXN_MAX_SZ ] __attribute__((aligned(alignof(fd_txn_t))));
71 : fd_hash_t poh; /* Latest PoH hash we've seen from the ingested FEC sets. */
72 : ulong mblks_rem; /* Number of microblocks remaining in the current batch. */
73 : ulong txns_rem; /* Number of transactions remaining in the current microblock. */
74 : fd_acct_addr_t aluts[ 256 ]; /* Resolve ALUT accounts into this buffer for more parallelism. */
75 : uint fec_buf_sz; /* Size of the fec_buf in bytes. */
76 : uint fec_buf_soff; /* Starting offset into fec_buf for unparsed transactions. */
77 : uint fec_buf_boff; /* Byte offset into raw block data of the first byte currently in fec_buf */
78 : uint fec_eob:1; /* FEC end-of-batch: set if the last FEC set in the batch is being
79 : ingested. */
80 : uint fec_sob:1; /* FEC start-of-batch: set if the parser expects to be receiving a new
81 : batch. */
82 :
83 : /* Block state. */
84 : uint fec_eos:1; /* FEC end-of-stream: set if the last FEC set in the block has been
85 : ingested. */
86 : uint rooted:1; /* Set if the block is rooted. */
87 : uint dying:1; /* Set if the block has been abandoned and no transactions should be
88 : scheduled from it. */
89 : uint in_sched:1; /* Set if the block is being tracked by the scheduler. */
90 : uint in_rdisp:1; /* Set if the block is being tracked by the dispatcher, either as staged
91 : or unstaged. */
92 : uint block_start_signaled:1; /* Set if the start-of-block sentinel has been dispatched. */
93 : uint block_end_signaled:1; /* Set if the end-of-block sentinel has been dispatched. */
94 : uint block_start_done:1; /* Set if the start-of-block processing has been completed. */
95 : uint block_end_done:1; /* Set if the end-of-block processing has been completed. */
96 : uint staged:1; /* Set if the block is in a dispatcher staging lane; a staged block is
97 : tracked by the dispatcher. */
98 : ulong staging_lane; /* Ignored if staged==0. */
99 : ulong luf_depth; /* Depth of longest unstaged fork starting from this node; only
100 : stageable unstaged descendants are counted. */
101 : uchar fec_buf[ FD_SCHED_MAX_FEC_BUF_SZ ]; /* The previous FEC set could have some residual data that only becomes
102 : parseable after the next FEC set is ingested. */
103 : uint shred_blk_offs[ FD_SHRED_BLK_MAX ]; /* The byte offsets into block data of ingested shreds */
104 : };
105 : typedef struct fd_sched_block fd_sched_block_t;
106 :
107 : FD_STATIC_ASSERT( sizeof(fd_hash_t)==sizeof(((fd_microblock_hdr_t *)0)->hash), unexpected poh hash size );
108 :
109 :
110 : struct fd_sched_metrics {
111 : uint block_added_cnt;
112 : uint block_added_staged_cnt;
113 : uint block_added_unstaged_cnt;
114 : uint block_added_dead_ood_cnt;
115 : uint block_removed_cnt;
116 : uint block_abandoned_cnt;
117 : uint block_bad_cnt;
118 : uint block_promoted_cnt;
119 : uint block_demoted_cnt;
120 : uint deactivate_no_child_cnt;
121 : uint deactivate_no_txn_cnt;
122 : uint deactivate_pruned_cnt;
123 : uint deactivate_abandoned_cnt;
124 : uint lane_switch_cnt;
125 : uint lane_promoted_cnt;
126 : uint lane_demoted_cnt;
127 : uint alut_success_cnt;
128 : uint alut_serializing_cnt;
129 : uint txn_abandoned_parsed_cnt;
130 : uint txn_abandoned_exec_done_cnt;
131 : uint txn_abandoned_done_cnt;
132 : uint txn_max_in_flight_cnt;
133 : ulong txn_weighted_in_flight_cnt;
134 : ulong txn_weighted_in_flight_tickcount;
135 : ulong txn_none_in_flight_tickcount;
136 : ulong txn_parsed_cnt;
137 : ulong txn_exec_done_cnt;
138 : ulong txn_sigverify_done_cnt;
139 : ulong txn_done_cnt;
140 : ulong bytes_ingested_cnt;
141 : ulong bytes_ingested_unparsed_cnt;
142 : ulong bytes_dropped_cnt;
143 : ulong fec_cnt;
144 : };
145 : typedef struct fd_sched_metrics fd_sched_metrics_t;
146 :
147 : struct fd_sched {
148 : char print_buf[ FD_SCHED_MAX_PRINT_BUF_SZ ];
149 : ulong print_buf_sz;
150 : fd_sched_metrics_t metrics[ 1 ];
151 : ulong canary; /* == FD_SCHED_MAGIC */
152 : ulong block_cnt_max; /* Immutable. */
153 : ulong exec_cnt; /* Immutable. */
154 : long txn_in_flight_last_tick;
155 : ulong root_idx;
156 : fd_rdisp_t * rdisp;
157 : ulong txn_exec_ready_bitset[ 1 ];
158 : ulong sigverify_ready_bitset[ 1 ];
159 : ulong active_bank_idx; /* Index of the actively replayed block, or ULONG_MAX if no block is
160 : actively replayed; has to have a transaction to dispatch; staged
161 : blocks that have no transactions to dispatch are not eligible for
162 : being active. */
163 : ulong staged_bitset; /* Bit i set if staging lane i is occupied. */
164 : ulong staged_head_bank_idx[ FD_SCHED_MAX_STAGING_LANES ]; /* Head of the linear chain in each staging lane, ignored if bit i is
165 : not set in the bitset. */
166 : ulong txn_pool_free_cnt;
167 : fd_txn_p_t txn_pool[ FD_SCHED_MAX_DEPTH ];
168 : ulong tile_to_bank_idx[ FD_SCHED_MAX_EXEC_TILE_CNT ]; /* Index of the bank that the exec tile is executing against. */
169 : txn_bitset_t exec_done_set[ txn_bitset_word_cnt ]; /* Indexed by txn_idx. */
170 : txn_bitset_t sigverify_done_set[ txn_bitset_word_cnt ]; /* Indexed by txn_idx. */
171 : fd_sched_block_t * block_pool; /* Just a flat array. */
172 : ulong block_pool_popcnt;
173 : };
174 : typedef struct fd_sched fd_sched_t;
175 :
176 :
177 : /* Internal helpers. */
178 :
179 : static void
180 : add_block( fd_sched_t * sched,
181 : ulong bank_idx,
182 : ulong parent_bank_idx );
183 :
184 : FD_WARN_UNUSED static int
185 : fd_sched_parse( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx );
186 :
187 : FD_WARN_UNUSED static int
188 : fd_sched_parse_txn( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx );
189 :
190 : static void
191 : try_activate_block( fd_sched_t * sched );
192 :
193 : static void
194 : check_or_set_active_block( fd_sched_t * sched );
195 :
196 : static void
197 : subtree_abandon( fd_sched_t * sched, fd_sched_block_t * block );
198 :
199 : static void
200 : maybe_switch_block( fd_sched_t * sched, ulong bank_idx );
201 :
202 : FD_FN_UNUSED static ulong
203 : find_and_stage_longest_unstaged_fork( fd_sched_t * sched, int lane_idx );
204 :
205 : static ulong
206 : compute_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx );
207 :
208 : static ulong
209 : stage_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx, int lane_idx );
210 :
211 : static inline fd_sched_block_t *
212 0 : block_pool_ele( fd_sched_t * sched, ulong idx ) {
213 0 : FD_TEST( idx<sched->block_cnt_max || idx==ULONG_MAX );
214 0 : return idx==ULONG_MAX ? NULL : sched->block_pool+idx;
215 0 : }
216 :
217 : FD_FN_UNUSED static inline int
218 0 : block_is_void( fd_sched_block_t * block ) {
219 0 : /* We've seen everything in the block and no transaction got parsed
220 0 : out. */
221 0 : return block->fec_eos && block->txn_parsed_cnt==0;
222 0 : }
223 :
224 : static inline int
225 0 : block_should_signal_end( fd_sched_block_t * block ) {
226 0 : return block->fec_eos && block->txn_parsed_cnt==block->txn_done_cnt && block->block_start_done && !block->block_end_signaled;
227 0 : }
228 :
229 : static inline int
230 0 : block_will_signal_end( fd_sched_block_t * block ) {
231 0 : return block->fec_eos && !block->block_end_signaled;
232 0 : }
233 :
234 : /* Is there something known to be dispatchable in the block? This is an
235 : important liveness property. A block that doesn't contain any known
236 : dispatchable tasks will be deactivated or demoted. */
237 : static inline int
238 0 : block_is_dispatchable( fd_sched_block_t * block ) {
239 0 : ulong exec_queued_cnt = block->txn_parsed_cnt-block->txn_exec_in_flight_cnt-block->txn_exec_done_cnt;
240 0 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
241 0 : return exec_queued_cnt>0UL ||
242 0 : sigverify_queued_cnt>0UL ||
243 0 : !block->block_start_signaled ||
244 0 : block_will_signal_end( block );
245 0 : }
246 :
247 : static inline int
248 0 : block_is_in_flight( fd_sched_block_t * block ) {
249 0 : return block->txn_exec_in_flight_cnt || block->txn_sigverify_in_flight_cnt || (block->block_end_signaled && !block->block_end_done);
250 0 : }
251 :
252 : static inline int
253 0 : block_is_done( fd_sched_block_t * block ) {
254 0 : return block->fec_eos && block->txn_parsed_cnt==block->txn_done_cnt && block->block_start_done && block->block_end_done;
255 0 : }
256 :
257 : static inline int
258 0 : block_is_stageable( fd_sched_block_t * block ) {
259 0 : int rv = !block_is_done( block ) && !block->dying;
260 0 : if( FD_UNLIKELY( rv && !block->in_rdisp ) ) {
261 : /* Invariant: stageable blocks may be currently staged or unstaged,
262 : but must be in the dispatcher either way. When a block
263 : transitions to DONE, it will be immediately removed from the
264 : dispatcher. When a block transitions to DYING, it will be
265 : eventually abandoned from the dispatcher. */
266 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",
267 0 : block->txn_parsed_cnt, block->txn_done_cnt, (uint)block->fec_eos, block->slot, block->parent_slot ));
268 0 : }
269 0 : return rv;
270 0 : }
271 :
272 : static inline int
273 0 : block_is_promotable( fd_sched_block_t * block ) {
274 0 : return block_is_stageable( block ) && block_is_dispatchable( block ) && !block->staged;
275 0 : }
276 :
277 : static inline int
278 0 : block_is_activatable( fd_sched_block_t * block ) {
279 0 : return block_is_stageable( block ) && block_is_dispatchable( block ) && block->staged;
280 0 : }
281 :
282 : static inline int
283 0 : block_should_deactivate( fd_sched_block_t * block ) {
284 : /* We allow a grace period, during which a block has nothing to
285 : dispatch, but has something in-flight. The block is allowed to
286 : stay activated and ingest FEC sets during this time. The block
287 : will be deactivated if there's still nothing to dispatch by the
288 : time all in-flight tasks are completed. */
289 0 : return !block_is_activatable( block ) && !block_is_in_flight( block );
290 0 : }
291 :
292 : static inline ulong
293 0 : block_to_idx( fd_sched_t * sched, fd_sched_block_t * block ) { return (ulong)(block-sched->block_pool); }
294 :
295 : __attribute__((format(printf,2,3)))
296 : static void
297 : fd_sched_printf( fd_sched_t * sched,
298 : char const * fmt,
299 0 : ... ) {
300 0 : va_list ap;
301 0 : ulong len;
302 0 : va_start( ap, fmt );
303 0 : int ret = vsnprintf( sched->print_buf+sched->print_buf_sz,
304 0 : FD_SCHED_MAX_PRINT_BUF_SZ-sched->print_buf_sz,
305 0 : fmt, ap );
306 0 : va_end( ap );
307 0 : len = fd_ulong_if( ret<0, 0UL, fd_ulong_min( (ulong)ret, FD_SCHED_MAX_PRINT_BUF_SZ-sched->print_buf_sz-1UL ) );
308 0 : sched->print_buf[ sched->print_buf_sz+len ] = '\0';
309 0 : sched->print_buf_sz += len;
310 0 : }
311 :
312 : FD_FN_UNUSED static void
313 0 : print_block_metrics( fd_sched_t * sched, fd_sched_block_t * block ) {
314 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, txn_done_cnt %u, shred_cnt %u, fec_cnt %u, txn_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",
315 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->txn_done_cnt, block->shred_cnt, block->fec_cnt, block->txn_pool_max_popcnt, FD_SCHED_MAX_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 );
316 0 : }
317 :
318 : FD_FN_UNUSED static void
319 0 : print_block_debug( fd_sched_t * sched, fd_sched_block_t * block ) {
320 0 : 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, txn_done_cnt %u, shred_cnt %u, fec_cnt %u, txn_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",
321 0 : 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->txn_done_cnt, block->shred_cnt, block->fec_cnt, block->txn_pool_max_popcnt, FD_SCHED_MAX_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 );
322 0 : }
323 :
324 : FD_FN_UNUSED static void
325 0 : print_block_and_parent( fd_sched_t * sched, fd_sched_block_t * block ) {
326 0 : print_block_debug( sched, block );
327 0 : fd_sched_block_t * parent = block_pool_ele( sched, block->parent_idx );
328 0 : if( FD_LIKELY( parent ) ) print_block_debug( sched, parent );
329 0 : }
330 :
331 : FD_FN_UNUSED static void
332 0 : print_metrics( fd_sched_t * sched ) {
333 0 : 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, 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_done_cnt %lu, bytes_ingested_cnt %lu, bytes_ingested_unparsed_cnt %lu, bytes_dropped_cnt %lu, fec_cnt %lu\n",
334 0 : sched->metrics->block_added_cnt, sched->metrics->block_added_staged_cnt, sched->metrics->block_added_unstaged_cnt, sched->metrics->block_added_dead_ood_cnt, sched->metrics->block_removed_cnt, sched->metrics->block_abandoned_cnt, sched->metrics->block_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->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_done_cnt, sched->metrics->bytes_ingested_cnt, sched->metrics->bytes_ingested_unparsed_cnt, sched->metrics->bytes_dropped_cnt, sched->metrics->fec_cnt );
335 :
336 0 : }
337 :
338 : FD_FN_UNUSED static void
339 0 : print_sched( fd_sched_t * sched ) {
340 0 : 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, active_idx %lu, staged_bitset %lu, staged_head_idx[0] %lu, staged_head_idx[1] %lu, staged_head_idx[2] %lu, staged_head_idx[3] %lu, txn_pool_free_cnt %lu/%lu, block_pool_popcnt %lu/%lu\n",
341 0 : sched->canary, sched->exec_cnt, sched->root_idx, sched->txn_exec_ready_bitset[ 0 ], sched->sigverify_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->txn_pool_free_cnt, FD_SCHED_MAX_DEPTH, sched->block_pool_popcnt, sched->block_cnt_max );
342 0 : fd_sched_block_t * active_block = block_pool_ele( sched, sched->active_bank_idx );
343 0 : if( active_block ) print_block_debug( sched, active_block );
344 0 : for( int l=0; l<(int)FD_SCHED_MAX_STAGING_LANES; l++ ) {
345 0 : if( fd_ulong_extract_bit( sched->staged_bitset, l ) ) {
346 0 : fd_sched_block_t * block = block_pool_ele( sched, sched->staged_head_bank_idx[ l ] );
347 0 : print_block_debug( sched, block );
348 0 : }
349 0 : }
350 0 : }
351 :
352 : FD_FN_UNUSED static void
353 0 : print_all( fd_sched_t * sched, fd_sched_block_t * block ) {
354 0 : print_metrics( sched );
355 0 : print_sched( sched );
356 0 : print_block_and_parent( sched, block );
357 0 : }
358 :
359 :
360 : /* Public functions. */
361 :
362 0 : ulong fd_sched_align( void ) {
363 0 : return fd_ulong_max( alignof(fd_sched_t),
364 0 : fd_ulong_max( fd_rdisp_align(),
365 0 : fd_ulong_max( alignof(fd_sched_block_t), 64UL ))); /* Minimally cache line aligned. */
366 0 : }
367 :
368 : ulong
369 0 : fd_sched_footprint( ulong block_cnt_max ) {
370 0 : ulong l = FD_LAYOUT_INIT;
371 0 : l = FD_LAYOUT_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
372 0 : l = FD_LAYOUT_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( FD_SCHED_MAX_DEPTH, block_cnt_max ) ); /* dispatcher */
373 0 : l = FD_LAYOUT_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) ); /* block pool */
374 0 : return FD_LAYOUT_FINI( l, fd_sched_align() );
375 0 : }
376 :
377 : void *
378 0 : fd_sched_new( void * mem, ulong block_cnt_max, ulong exec_cnt ) {
379 0 : FD_TEST( exec_cnt && exec_cnt<=FD_SCHED_MAX_EXEC_TILE_CNT );
380 :
381 0 : FD_SCRATCH_ALLOC_INIT( l, mem );
382 0 : fd_sched_t * sched = FD_SCRATCH_ALLOC_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
383 0 : void * _rdisp = FD_SCRATCH_ALLOC_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( FD_SCHED_MAX_DEPTH, block_cnt_max ) );
384 0 : void * _bpool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) );
385 0 : FD_SCRATCH_ALLOC_FINI( l, fd_sched_align() );
386 :
387 0 : ulong seed = ((ulong)fd_tickcount()) ^ FD_SCHED_MAGIC;
388 0 : fd_rdisp_new( _rdisp, FD_SCHED_MAX_DEPTH, block_cnt_max, seed );
389 :
390 0 : fd_sched_block_t * bpool = (fd_sched_block_t *)_bpool;
391 0 : for( ulong i=0; i<block_cnt_max; i++ ) {
392 0 : bpool[ i ].in_sched = 0;
393 0 : }
394 :
395 0 : fd_memset( sched->metrics, 0, sizeof(fd_sched_metrics_t) );
396 0 : sched->txn_in_flight_last_tick = LONG_MAX;
397 :
398 0 : sched->canary = FD_SCHED_MAGIC;
399 0 : sched->block_cnt_max = block_cnt_max;
400 0 : sched->exec_cnt = exec_cnt;
401 0 : sched->root_idx = ULONG_MAX;
402 0 : sched->active_bank_idx = ULONG_MAX;
403 0 : sched->staged_bitset = 0UL;
404 :
405 0 : sched->txn_exec_ready_bitset[ 0 ] = fd_ulong_mask_lsb( (int)exec_cnt );
406 0 : sched->sigverify_ready_bitset[ 0 ] = fd_ulong_mask_lsb( (int)exec_cnt );
407 :
408 0 : sched->txn_pool_free_cnt = FD_SCHED_MAX_DEPTH-1UL; /* -1 because index 0 is unusable as a sentinel reserved by the dispatcher */
409 :
410 0 : txn_bitset_new( sched->exec_done_set );
411 0 : txn_bitset_new( sched->sigverify_done_set );
412 :
413 0 : sched->block_pool_popcnt = 0UL;
414 :
415 0 : return sched;
416 0 : }
417 :
418 : fd_sched_t *
419 0 : fd_sched_join( void * mem, ulong block_cnt_max ) {
420 0 : fd_sched_t * sched = (fd_sched_t *)mem;
421 :
422 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
423 0 : FD_TEST( sched->block_cnt_max==block_cnt_max );
424 :
425 0 : FD_SCRATCH_ALLOC_INIT( l, mem );
426 0 : /* */ FD_SCRATCH_ALLOC_APPEND( l, fd_sched_align(), sizeof(fd_sched_t) );
427 0 : void * _rdisp = FD_SCRATCH_ALLOC_APPEND( l, fd_rdisp_align(), fd_rdisp_footprint( FD_SCHED_MAX_DEPTH, block_cnt_max ) );
428 0 : void * _bpool = FD_SCRATCH_ALLOC_APPEND( l, alignof(fd_sched_block_t), block_cnt_max*sizeof(fd_sched_block_t) );
429 0 : FD_SCRATCH_ALLOC_FINI( l, fd_sched_align() );
430 :
431 0 : sched->rdisp = fd_rdisp_join( _rdisp );
432 0 : sched->block_pool = _bpool;
433 :
434 0 : txn_bitset_join( sched->exec_done_set );
435 0 : txn_bitset_join( sched->sigverify_done_set );
436 :
437 0 : return sched;
438 0 : }
439 :
440 : int
441 0 : fd_sched_fec_can_ingest( fd_sched_t * sched, fd_sched_fec_t * fec ) {
442 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
443 0 : FD_TEST( fec->bank_idx<sched->block_cnt_max );
444 0 : FD_TEST( fec->parent_bank_idx<sched->block_cnt_max );
445 :
446 0 : if( FD_UNLIKELY( fec->fec->data_sz>FD_SCHED_MAX_PAYLOAD_PER_FEC ) ) {
447 0 : sched->print_buf_sz = 0UL;
448 0 : print_metrics( sched );
449 0 : print_sched( sched );
450 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
451 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 ));
452 0 : }
453 :
454 0 : ulong fec_buf_sz = 0UL;
455 0 : fd_sched_block_t * block = block_pool_ele( sched, fec->bank_idx );
456 0 : if( FD_LIKELY( !fec->is_first_in_block ) ) {
457 0 : fec_buf_sz += block->fec_buf_sz-block->fec_buf_soff;
458 0 : } else {
459 : /* No residual data as this is a fresh new block. */
460 0 : }
461 : /* Addition is safe and won't overflow because we checked the FEC set
462 : size above. */
463 0 : fec_buf_sz += fec->fec->data_sz;
464 : /* Assuming every transaction is min size, do we have enough free
465 : entries in the txn pool? For a more precise txn count, we would
466 : have to do some parsing. */
467 0 : return sched->txn_pool_free_cnt>=fec_buf_sz/FD_TXN_MIN_SERIALIZED_SZ;
468 0 : }
469 :
470 : int
471 : fd_sched_can_ingest( fd_sched_t * sched,
472 0 : ulong fec_cnt ) {
473 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
474 : /* Worst case, we need one byte from the incoming data to extract a
475 : transaction out of the residual data, and the rest of the incoming
476 : data contributes toward min sized transactions. */
477 0 : return sched->txn_pool_free_cnt>=(FD_SCHED_MAX_TXN_PER_FEC*fec_cnt);
478 0 : }
479 :
480 : FD_WARN_UNUSED int
481 0 : fd_sched_fec_ingest( fd_sched_t * sched, fd_sched_fec_t * fec ) {
482 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
483 0 : FD_TEST( fec->bank_idx<sched->block_cnt_max );
484 0 : FD_TEST( fec->parent_bank_idx<sched->block_cnt_max );
485 :
486 0 : fd_sched_block_t * block = block_pool_ele( sched, fec->bank_idx );
487 :
488 0 : if( FD_UNLIKELY( fec->fec->data_sz>FD_SCHED_MAX_PAYLOAD_PER_FEC ) ) {
489 0 : sched->print_buf_sz = 0UL;
490 0 : print_all( sched, block );
491 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
492 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 ));
493 0 : }
494 :
495 0 : if( FD_UNLIKELY( fec->is_first_in_block ) ) {
496 : /* This is a new block. */
497 0 : add_block( sched, fec->bank_idx, fec->parent_bank_idx );
498 0 : block->slot = fec->slot;
499 0 : block->parent_slot = fec->parent_slot;
500 :
501 0 : if( FD_UNLIKELY( block->dying ) ) {
502 : /* The child of a dead block is also dead. We added it to our
503 : fork tree just so we could track an entire lineage of dead
504 : children and propagate the dead property to the entire lineage,
505 : in case there were frags for more than one dead children
506 : in-flight at the time the parent was abandoned. That being
507 : said, we shouldn't need to add the dead child to the
508 : dispatcher. */
509 0 : sched->metrics->block_added_dead_ood_cnt++;
510 :
511 : /* Ignore the FEC set for a dead block. */
512 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
513 0 : return 1;
514 0 : }
515 :
516 : /* Try to find a staging lane for this block. */
517 0 : int alloc_lane = 0;
518 0 : fd_sched_block_t * parent_block = block_pool_ele( sched, fec->parent_bank_idx );
519 0 : if( FD_LIKELY( parent_block->staged ) ) {
520 : /* Parent is staged. So see if we can continue down the same
521 : staging lane. */
522 0 : ulong staging_lane = parent_block->staging_lane;
523 0 : ulong child_idx = parent_block->child_idx;
524 0 : while( child_idx!=ULONG_MAX ) {
525 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
526 0 : if( child->staged && child->staging_lane==staging_lane ) {
527 : /* Found a child on the same lane. So we're done. */
528 0 : staging_lane = FD_RDISP_UNSTAGED;
529 0 : break;
530 0 : }
531 0 : child_idx = child->sibling_idx;
532 0 : }
533 : /* No child is staged on the same lane as the parent. So stage
534 : this block. This is the common case. */
535 0 : if( FD_LIKELY( staging_lane!=FD_RDISP_UNSTAGED ) ) {
536 0 : block->in_rdisp = 1;
537 0 : block->staged = 1;
538 0 : block->staging_lane = staging_lane;
539 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, staging_lane );
540 0 : sched->metrics->block_added_cnt++;
541 0 : sched->metrics->block_added_staged_cnt++;
542 0 : } else {
543 0 : alloc_lane = 1;
544 0 : }
545 0 : } else {
546 0 : if( block_is_stageable( parent_block ) ) {
547 : /* Parent is unstaged but stageable. So let's be unstaged too.
548 : This is a policy decision to be lazy and not promote parent
549 : at the moment. */
550 0 : block->in_rdisp = 1;
551 0 : block->staged = 0;
552 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, FD_RDISP_UNSTAGED );
553 0 : sched->metrics->block_added_cnt++;
554 0 : sched->metrics->block_added_unstaged_cnt++;
555 0 : } else {
556 0 : alloc_lane = 1;
557 0 : }
558 0 : }
559 0 : if( FD_UNLIKELY( alloc_lane ) ) {
560 : /* We weren't able to inherit the parent's staging lane. So try
561 : to find a new staging lane. */
562 0 : if( FD_LIKELY( sched->staged_bitset!=fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) ) ) { /* Optimize for lane available. */
563 0 : int lane_idx = fd_ulong_find_lsb( ~sched->staged_bitset );
564 0 : if( FD_UNLIKELY( lane_idx>=(int)FD_SCHED_MAX_STAGING_LANES ) ) {
565 0 : FD_LOG_CRIT(( "invariant violation: lane_idx %d, sched->staged_bitset %lx",
566 0 : lane_idx, sched->staged_bitset ));
567 0 : }
568 0 : sched->staged_bitset = fd_ulong_set_bit( sched->staged_bitset, lane_idx );
569 0 : sched->staged_head_bank_idx[ lane_idx ] = fec->bank_idx;
570 0 : block->in_rdisp = 1;
571 0 : block->staged = 1;
572 0 : block->staging_lane = (ulong)lane_idx;
573 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, block->staging_lane );
574 0 : sched->metrics->block_added_cnt++;
575 0 : sched->metrics->block_added_staged_cnt++;
576 0 : } else {
577 : /* No lanes available. */
578 0 : block->in_rdisp = 1;
579 0 : block->staged = 0;
580 0 : fd_rdisp_add_block( sched->rdisp, fec->bank_idx, FD_RDISP_UNSTAGED );
581 0 : sched->metrics->block_added_cnt++;
582 0 : sched->metrics->block_added_unstaged_cnt++;
583 0 : }
584 0 : }
585 0 : }
586 :
587 0 : block->txn_pool_max_popcnt = fd_ulong_max( block->txn_pool_max_popcnt, FD_SCHED_MAX_DEPTH-sched->txn_pool_free_cnt );
588 0 : block->block_pool_max_popcnt = fd_ulong_max( block->block_pool_max_popcnt, sched->block_pool_popcnt );
589 :
590 0 : if( FD_UNLIKELY( block->dying ) ) {
591 : /* Ignore the FEC set for a dead block. */
592 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
593 0 : return 1;
594 0 : }
595 :
596 0 : if( FD_UNLIKELY( !block->in_rdisp ) ) {
597 : /* Invariant: block must be in the dispatcher at this point. */
598 0 : sched->print_buf_sz = 0UL;
599 0 : print_all( sched, block );
600 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
601 0 : FD_LOG_CRIT(( "invariant violation: block->in_rdisp==0, slot %lu, parent slot %lu",
602 0 : block->slot, block->parent_slot ));
603 0 : }
604 :
605 0 : if( FD_UNLIKELY( block->fec_eos ) ) {
606 : /* This means something is wrong upstream. We're getting more FEC
607 : sets for a block that has already ended, or so we were told. */
608 0 : sched->print_buf_sz = 0UL;
609 0 : print_all( sched, block );
610 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
611 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 ));
612 0 : }
613 0 : if( FD_UNLIKELY( block->fec_eob && fec->is_last_in_batch ) ) {
614 : /* If the previous FEC set ingestion and parse was successful,
615 : block->fec_eob should be cleared. The fact that fec_eob is set
616 : means that the previous batch didn't parse properly. So this is
617 : a bad block. We should refuse to replay down the fork. */
618 0 : FD_LOG_INFO(( "bad block: failed to parse, slot %lu, parent slot %lu", fec->slot, fec->parent_slot ));
619 0 : sched->print_buf_sz = 0UL;
620 0 : print_all( sched, block );
621 0 : FD_LOG_DEBUG(( "%s", sched->print_buf ));
622 0 : subtree_abandon( sched, block );
623 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
624 0 : sched->metrics->block_bad_cnt++;
625 0 : check_or_set_active_block( sched );
626 0 : return 0;
627 0 : }
628 0 : if( FD_UNLIKELY( block->child_idx!=ULONG_MAX ) ) {
629 : /* This means something is wrong upstream. FEC sets are not being
630 : delivered in replay order. We got a child block FEC set before
631 : this block was completely delivered. */
632 0 : sched->print_buf_sz = 0UL;
633 0 : print_all( sched, block );
634 0 : fd_sched_block_t * child_block = block_pool_ele( sched, block->child_idx );
635 0 : print_block_debug( sched, child_block );
636 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
637 0 : FD_LOG_CRIT(( "invariant violation: block->child_idx %lu, slot %lu, parent slot %lu", block->child_idx, fec->slot, fec->parent_slot ));
638 0 : }
639 :
640 0 : FD_TEST( block->fec_buf_sz>=block->fec_buf_soff );
641 0 : if( FD_LIKELY( block->fec_buf_sz>block->fec_buf_soff ) ) {
642 : /* If there is residual data from the previous FEC set within the
643 : same batch, we move it to the beginning of the buffer and append
644 : the new FEC set. */
645 0 : memmove( block->fec_buf, block->fec_buf+block->fec_buf_soff, block->fec_buf_sz-block->fec_buf_soff );
646 0 : }
647 0 : block->fec_buf_boff += block->fec_buf_soff;
648 0 : block->fec_buf_sz -= block->fec_buf_soff;
649 0 : block->fec_buf_soff = 0;
650 : /* Addition is safe and won't overflow because we checked the FEC
651 : set size above. */
652 0 : if( FD_UNLIKELY( block->fec_buf_sz+fec->fec->data_sz>FD_SCHED_MAX_FEC_BUF_SZ ) ) {
653 : /* In a conformant block, there shouldn't be more than a
654 : transaction's worth of residual data left over from the previous
655 : FEC set within the same batch. So if this condition doesn't
656 : hold, it's a bad block. Instead of crashing, we should refuse to
657 : replay down the fork. */
658 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 ));
659 0 : sched->print_buf_sz = 0UL;
660 0 : print_all( sched, block );
661 0 : FD_LOG_DEBUG(( "%s", sched->print_buf ));
662 0 : subtree_abandon( sched, block );
663 0 : sched->metrics->bytes_dropped_cnt += fec->fec->data_sz;
664 0 : sched->metrics->block_bad_cnt++;
665 0 : check_or_set_active_block( sched );
666 0 : return 0;
667 0 : }
668 :
669 : /* Append the new FEC set to the end of the buffer. */
670 0 : fd_memcpy( block->fec_buf+block->fec_buf_sz, fec->fec->data, fec->fec->data_sz );
671 0 : block->fec_buf_sz += (uint)fec->fec->data_sz;
672 0 : sched->metrics->bytes_ingested_cnt += fec->fec->data_sz;
673 :
674 0 : block->fec_eob = fec->is_last_in_batch;
675 0 : block->fec_eos = fec->is_last_in_block;
676 :
677 0 : ulong block_sz = block->shred_cnt>0 ? block->shred_blk_offs[ block->shred_cnt-1 ] : 0UL;
678 0 : for( ulong i=0; i<fec->shred_cnt; i++ ) {
679 0 : if( FD_LIKELY( i<32UL ) ) {
680 0 : block->shred_blk_offs[ block->shred_cnt++ ] = (uint)block_sz + fec->fec->block_offs[ i ];
681 0 : } else if( FD_UNLIKELY( i!=fec->shred_cnt-1UL ) ) {
682 : /* We don't track shred boundaries after 32 shreds, assume they're
683 : sized uniformly */
684 0 : ulong num_overflow_shreds = fec->shred_cnt-32UL;
685 0 : ulong overflow_idx = i-32UL;
686 0 : ulong overflow_data_sz = fec->fec->data_sz-fec->fec->block_offs[ 31 ];
687 0 : block->shred_blk_offs[ block->shred_cnt++ ] = (uint)block_sz + fec->fec->block_offs[ 31 ] + (uint)(overflow_data_sz / num_overflow_shreds * (overflow_idx + 1UL));
688 0 : } else {
689 0 : block->shred_blk_offs[ block->shred_cnt++ ] = (uint)block_sz + (uint)fec->fec->data_sz;
690 0 : }
691 0 : }
692 :
693 0 : int err = fd_sched_parse( sched, block, fec->alut_ctx );
694 :
695 0 : block->fec_cnt++;
696 0 : sched->metrics->fec_cnt++;
697 :
698 0 : if( FD_UNLIKELY( err==FD_SCHED_PARSER_BAD_BLOCK ) ) {
699 0 : FD_LOG_INFO(( "bad block: slot %lu, parent slot %lu", block->slot, block->parent_slot ));
700 0 : sched->print_buf_sz = 0UL;
701 0 : print_all( sched, block );
702 0 : FD_LOG_DEBUG(( "%s", sched->print_buf ));
703 0 : subtree_abandon( sched, block );
704 0 : sched->metrics->bytes_dropped_cnt += block->fec_buf_sz-block->fec_buf_soff;
705 0 : sched->metrics->block_bad_cnt++;
706 0 : check_or_set_active_block( sched );
707 0 : return 0;
708 0 : }
709 :
710 : /* Check if we need to set the active block. */
711 0 : check_or_set_active_block( sched );
712 :
713 0 : return 1;
714 0 : }
715 :
716 : ulong
717 0 : fd_sched_task_next_ready( fd_sched_t * sched, fd_sched_task_t * out ) {
718 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
719 :
720 0 : ulong exec_ready_bitset0 = sched->txn_exec_ready_bitset[ 0 ];
721 0 : ulong exec_fully_ready_bitset = sched->sigverify_ready_bitset[ 0 ] & exec_ready_bitset0;
722 0 : if( FD_UNLIKELY( !exec_fully_ready_bitset ) ) {
723 : /* Early exit if no exec tiles available. */
724 0 : return 0UL;
725 0 : }
726 :
727 0 : if( FD_UNLIKELY( sched->active_bank_idx==ULONG_MAX ) ) {
728 : /* No need to try activating a block. If we're in this state,
729 : there's truly nothing to execute. We will activate something
730 : when we ingest a FEC set with transactions. */
731 0 : return 0UL;
732 0 : }
733 :
734 0 : out->task_type = FD_SCHED_TT_NULL;
735 :
736 : /* We could in theory reevaluate staging lane allocation here and do
737 : promotion/demotion as needed. It's a policy decision to minimize
738 : fork churn for now and just execute down the same active fork. */
739 :
740 0 : ulong bank_idx = sched->active_bank_idx;
741 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
742 0 : if( FD_UNLIKELY( block_should_deactivate( block ) ) ) {
743 0 : sched->print_buf_sz = 0UL;
744 0 : print_all( sched, block );
745 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
746 0 : FD_LOG_CRIT(( "invariant violation: active_bank_idx %lu is not activatable nor has anything in-flight", sched->active_bank_idx ));
747 0 : }
748 :
749 0 : block->txn_pool_max_popcnt = fd_ulong_max( block->txn_pool_max_popcnt, FD_SCHED_MAX_DEPTH-sched->txn_pool_free_cnt );
750 0 : block->block_pool_max_popcnt = fd_ulong_max( block->block_pool_max_popcnt, sched->block_pool_popcnt );
751 :
752 0 : if( FD_UNLIKELY( !block->block_start_signaled ) ) {
753 0 : out->task_type = FD_SCHED_TT_BLOCK_START;
754 0 : out->block_start->bank_idx = bank_idx;
755 0 : out->block_start->parent_bank_idx = block->parent_idx;
756 0 : out->block_start->slot = block->slot;
757 0 : block->block_start_signaled = 1;
758 0 : return 1UL;
759 0 : }
760 :
761 0 : ulong exec_tile_idx0 = fd_ulong_if( !!exec_fully_ready_bitset, (ulong)fd_ulong_find_lsb( exec_fully_ready_bitset ), ULONG_MAX );
762 0 : ulong exec_queued_cnt = block->txn_parsed_cnt-block->txn_exec_in_flight_cnt-block->txn_exec_done_cnt;
763 0 : if( FD_LIKELY( exec_queued_cnt>0UL && fd_ulong_popcnt( exec_fully_ready_bitset ) ) ) { /* Optimize for no fork switching. */
764 : /* Transaction execution has the highest priority. Current mainnet
765 : block times are very much dominated by critical path transaction
766 : execution. To achieve the fastest block replay speed, we can't
767 : afford to make any mistake in critical path dispatching. Any
768 : deviation from perfect critical path dispatching is basically
769 : irrecoverable. As such, we try to keep all the exec tiles busy
770 : with transaction execution, but we allow at most one transaction
771 : to be in-flight per exec tile. This is to ensure that whenever a
772 : critical path transaction completes, we have at least one exec
773 : tile, e.g. the one that just completed said transaction, readily
774 : available to continue executing down the critical path. */
775 0 : out->txn_exec->txn_idx = fd_rdisp_get_next_ready( sched->rdisp, bank_idx );
776 0 : if( FD_UNLIKELY( out->txn_exec->txn_idx==0UL ) ) {
777 : /* There are transactions queued but none ready for execution.
778 : This implies that there must be in-flight transactions on whose
779 : completion the queued transactions depend. So we return and
780 : wait for those in-flight transactions to retire. This is a
781 : policy decision to execute as much as we can down the current
782 : fork. */
783 0 : if( FD_UNLIKELY( !block->txn_exec_in_flight_cnt ) ) {
784 0 : sched->print_buf_sz = 0UL;
785 0 : print_all( sched, block );
786 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
787 0 : FD_LOG_CRIT(( "invariant violation: no ready transaction found but block->txn_exec_in_flight_cnt==0" ));
788 0 : }
789 :
790 : /* Dispatch more sigverify tasks only if at least one exec tile is
791 : executing transactions or completely idle. Allow at most one
792 : sigverify task in-flight per tile, and only dispatch to
793 : completely idle tiles. */
794 0 : ulong sigverify_ready_bitset = exec_fully_ready_bitset;
795 0 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
796 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 ) ) ) {
797 : /* Dispatch transactions for sigverify in parse order. */
798 0 : int exec_tile_idx_sigverify = fd_ulong_find_lsb( sigverify_ready_bitset );
799 0 : out->task_type = FD_SCHED_TT_TXN_SIGVERIFY;
800 0 : out->txn_sigverify->bank_idx = bank_idx;
801 0 : out->txn_sigverify->txn_idx = block->txn_idx[ block->txn_sigverify_done_cnt+block->txn_sigverify_in_flight_cnt ];
802 0 : out->txn_sigverify->exec_idx = (ulong)exec_tile_idx_sigverify;
803 0 : sched->sigverify_ready_bitset[ 0 ] = fd_ulong_clear_bit( sched->sigverify_ready_bitset[ 0 ], exec_tile_idx_sigverify );
804 0 : sched->tile_to_bank_idx[ exec_tile_idx_sigverify ] = bank_idx;
805 0 : block->txn_sigverify_in_flight_cnt++;
806 0 : if( FD_UNLIKELY( (~sched->txn_exec_ready_bitset[ 0 ])&(~sched->sigverify_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", sched->txn_exec_ready_bitset[ 0 ], sched->sigverify_ready_bitset[ 0 ] ));
807 0 : return 1UL;
808 0 : }
809 0 : return 0UL;
810 0 : }
811 0 : out->task_type = FD_SCHED_TT_TXN_EXEC;
812 0 : out->txn_exec->bank_idx = bank_idx;
813 0 : out->txn_exec->slot = block->slot;
814 0 : out->txn_exec->exec_idx = exec_tile_idx0;
815 0 : FD_TEST( out->txn_exec->exec_idx!=ULONG_MAX );
816 :
817 0 : long now = fd_tickcount();
818 0 : ulong delta = (ulong)(now-sched->txn_in_flight_last_tick);
819 0 : ulong txn_exec_busy_cnt = sched->exec_cnt-(ulong)fd_ulong_popcnt( exec_ready_bitset0 );
820 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 );
821 0 : sched->metrics->txn_weighted_in_flight_tickcount += fd_ulong_if( txn_exec_busy_cnt!=0UL, delta, 0UL );
822 0 : sched->metrics->txn_weighted_in_flight_cnt += delta*txn_exec_busy_cnt;
823 0 : sched->txn_in_flight_last_tick = now;
824 :
825 0 : sched->txn_exec_ready_bitset[ 0 ] = fd_ulong_clear_bit( exec_ready_bitset0, (int)exec_tile_idx0);
826 0 : sched->tile_to_bank_idx[ exec_tile_idx0 ] = bank_idx;
827 :
828 0 : block->txn_exec_in_flight_cnt++;
829 0 : sched->metrics->txn_max_in_flight_cnt = fd_uint_max( sched->metrics->txn_max_in_flight_cnt, block->txn_exec_in_flight_cnt );
830 :
831 0 : ulong total_exec_busy_cnt = sched->exec_cnt-(ulong)fd_ulong_popcnt( sched->txn_exec_ready_bitset[ 0 ]&sched->sigverify_ready_bitset[ 0 ] );
832 0 : if( FD_UNLIKELY( (~sched->txn_exec_ready_bitset[ 0 ])&(~sched->sigverify_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", sched->txn_exec_ready_bitset[ 0 ], sched->sigverify_ready_bitset[ 0 ] ));
833 0 : if( FD_UNLIKELY( block->txn_exec_in_flight_cnt+block->txn_sigverify_in_flight_cnt!=total_exec_busy_cnt ) ) {
834 : /* Ideally we'd simply assert that the two sides of the equation
835 : are equal. But abandoned blocks throw a wrench into this. We
836 : allow abandoned blocks to have in-flight transactions that are
837 : naturally drained while we try to dispatch from another block.
838 : In such cases, the total number of in-flight transactions
839 : should include the abandoned blocks too. The contract is that
840 : blocks with in-flight transactions cannot be abandoned or
841 : demoted from rdisp. So a dying block has to be the head of one
842 : of the staging lanes. */
843 0 : ulong total_in_flight = 0UL;
844 0 : for( int l=0; l<(int)FD_SCHED_MAX_STAGING_LANES; l++ ) {
845 0 : if( fd_ulong_extract_bit( sched->staged_bitset, l ) ) {
846 0 : fd_sched_block_t * staged_block = block_pool_ele( sched, sched->staged_head_bank_idx[ l ] );
847 0 : if( FD_UNLIKELY( block_is_in_flight( staged_block )&&!(staged_block==block||staged_block->dying) ) ) {
848 0 : sched->print_buf_sz = 0UL;
849 0 : print_all( sched, staged_block );
850 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
851 0 : FD_LOG_CRIT(( "invariant violation: in-flight block is neither active nor dying" ));
852 0 : }
853 0 : total_in_flight += staged_block->txn_exec_in_flight_cnt;
854 0 : total_in_flight += staged_block->txn_sigverify_in_flight_cnt;
855 0 : }
856 0 : }
857 0 : if( FD_UNLIKELY( total_in_flight!=total_exec_busy_cnt ) ) {
858 0 : sched->print_buf_sz = 0UL;
859 0 : print_all( sched, block );
860 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
861 0 : FD_LOG_CRIT(( "invariant violation: total_in_flight %lu != total_exec_busy_cnt %lu", total_in_flight, total_exec_busy_cnt ));
862 0 : }
863 0 : FD_LOG_DEBUG(( "exec_busy_cnt %lu checks out", total_exec_busy_cnt ));
864 0 : }
865 0 : return 1UL;
866 0 : }
867 :
868 : /* At this point txn_queued_cnt==0 */
869 :
870 : /* Try to dispatch a sigverify task, but leave one exec tile idle for
871 : critical path execution, unless there's not going to be any more
872 : real transactions for the critical path. In the degenerate case of
873 : only one exec tile, keep it busy. */
874 0 : ulong sigverify_ready_bitset = exec_fully_ready_bitset;
875 0 : ulong sigverify_queued_cnt = block->txn_parsed_cnt-block->txn_sigverify_in_flight_cnt-block->txn_sigverify_done_cnt;
876 0 : 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 ) ) ) {
877 : /* Dispatch transactions for sigverify in parse order. */
878 0 : int exec_tile_idx_sigverify = fd_ulong_find_lsb( sigverify_ready_bitset );
879 0 : out->task_type = FD_SCHED_TT_TXN_SIGVERIFY;
880 0 : out->txn_sigverify->txn_idx = block->txn_idx[ block->txn_sigverify_done_cnt+block->txn_sigverify_in_flight_cnt ];
881 0 : out->txn_sigverify->bank_idx = bank_idx;
882 0 : out->txn_sigverify->exec_idx = (ulong)exec_tile_idx_sigverify;
883 0 : sched->sigverify_ready_bitset[ 0 ] = fd_ulong_clear_bit( sched->sigverify_ready_bitset[ 0 ], exec_tile_idx_sigverify );
884 0 : sched->tile_to_bank_idx[ exec_tile_idx_sigverify ] = bank_idx;
885 0 : block->txn_sigverify_in_flight_cnt++;
886 0 : if( FD_UNLIKELY( (~sched->txn_exec_ready_bitset[ 0 ])&(~sched->sigverify_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", sched->txn_exec_ready_bitset[ 0 ], sched->sigverify_ready_bitset[ 0 ] ));
887 0 : return 1UL;
888 0 : }
889 :
890 0 : if( FD_UNLIKELY( block_should_signal_end( block ) ) ) {
891 0 : FD_TEST( block->block_start_signaled );
892 0 : out->task_type = FD_SCHED_TT_BLOCK_END;
893 0 : out->block_end->bank_idx = bank_idx;
894 0 : block->block_end_signaled = 1;
895 0 : return 1UL;
896 0 : }
897 :
898 : /* Nothing queued for the active block. If we haven't received all
899 : the FEC sets for it, then return and wait for more FEC sets, while
900 : there are in-flight transactions. This is a policy decision to
901 : minimize fork churn and allow for executing down the current fork
902 : as much as we can. If we have received all the FEC sets for it,
903 : then we'd still like to return and wait for the in-flight
904 : transactions to retire, before switching to a different block.
905 :
906 : Either way, there should be in-flight transactions. We deactivate
907 : the active block the moment we exhausted transactions from it. */
908 0 : if( FD_UNLIKELY( !block_is_in_flight( block ) ) ) {
909 0 : sched->print_buf_sz = 0UL;
910 0 : print_all( sched, block );
911 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
912 0 : FD_LOG_CRIT(( "invariant violation: expected in-flight transactions but none" ));
913 0 : }
914 :
915 0 : return 0UL;
916 0 : }
917 :
918 : void
919 0 : fd_sched_task_done( fd_sched_t * sched, ulong task_type, ulong txn_idx, ulong exec_idx ) {
920 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
921 :
922 0 : ulong bank_idx = ULONG_MAX;
923 0 : switch( task_type ) {
924 0 : case FD_SCHED_TT_BLOCK_START:
925 0 : case FD_SCHED_TT_BLOCK_END: {
926 0 : (void)txn_idx;
927 0 : bank_idx = sched->active_bank_idx;
928 0 : break;
929 0 : }
930 0 : case FD_SCHED_TT_TXN_EXEC:
931 0 : case FD_SCHED_TT_TXN_SIGVERIFY: {
932 0 : FD_TEST( txn_idx<FD_SCHED_MAX_DEPTH );
933 0 : bank_idx = sched->tile_to_bank_idx[ exec_idx ];
934 0 : break;
935 0 : }
936 0 : default: FD_LOG_CRIT(( "unsupported task_type %lu", task_type ));
937 0 : }
938 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
939 :
940 0 : if( FD_UNLIKELY( !block->staged ) ) {
941 : /* Invariant: only staged blocks can have in-flight transactions. */
942 0 : FD_LOG_CRIT(( "invariant violation: block->staged==0, slot %lu, parent slot %lu",
943 0 : block->slot, block->parent_slot ));
944 0 : }
945 0 : if( FD_UNLIKELY( !block->in_rdisp ) ) {
946 : /* Invariant: staged blocks must be in the dispatcher. */
947 0 : FD_LOG_CRIT(( "invariant violation: block->in_rdisp==0, slot %lu, parent slot %lu",
948 0 : block->slot, block->parent_slot ));
949 0 : }
950 :
951 0 : block->txn_pool_max_popcnt = fd_ulong_max( block->txn_pool_max_popcnt, FD_SCHED_MAX_DEPTH-sched->txn_pool_free_cnt );
952 0 : block->block_pool_max_popcnt = fd_ulong_max( block->block_pool_max_popcnt, sched->block_pool_popcnt );
953 :
954 0 : int exec_tile_idx = (int)exec_idx;
955 :
956 0 : switch( task_type ) {
957 0 : case FD_SCHED_TT_BLOCK_START: {
958 0 : FD_TEST( !block->block_start_done );
959 0 : block->block_start_done = 1;
960 0 : break;
961 0 : }
962 0 : case FD_SCHED_TT_BLOCK_END: {
963 : /* It may seem redundant to be invoking task_done() on these
964 : somewhat fake tasks. But these are necessary to drive state
965 : transition for empty blocks or slow blocks. */
966 0 : FD_TEST( !block->block_end_done );
967 0 : block->block_end_done = 1;
968 0 : break;
969 0 : }
970 0 : case FD_SCHED_TT_TXN_EXEC: {
971 0 : long now = fd_tickcount();
972 0 : ulong delta = (ulong)(now-sched->txn_in_flight_last_tick);
973 0 : ulong txn_exec_busy_cnt = sched->exec_cnt-(ulong)fd_ulong_popcnt( sched->txn_exec_ready_bitset[ 0 ] );
974 0 : sched->metrics->txn_weighted_in_flight_tickcount += delta;
975 0 : sched->metrics->txn_weighted_in_flight_cnt += delta*txn_exec_busy_cnt;
976 0 : sched->txn_in_flight_last_tick = now;
977 :
978 0 : block->txn_exec_done_cnt++;
979 0 : block->txn_exec_in_flight_cnt--;
980 0 : sched->metrics->txn_exec_done_cnt++;
981 0 : txn_bitset_insert( sched->exec_done_set, txn_idx );
982 0 : if( txn_bitset_test( sched->exec_done_set, txn_idx ) && txn_bitset_test( sched->sigverify_done_set, txn_idx ) ) {
983 : /* Release txn_idx if both exec and sigverify are done. This is
984 : guaranteed to only happen once per transaction because
985 : whichever one completed first would not release. */
986 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 1 );
987 0 : sched->txn_pool_free_cnt++;
988 0 : block->txn_done_cnt++;
989 0 : sched->metrics->txn_done_cnt++;
990 0 : } else {
991 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 0 );
992 0 : }
993 :
994 0 : FD_TEST( !fd_ulong_extract_bit( sched->txn_exec_ready_bitset[ 0 ], exec_tile_idx ) );
995 0 : sched->txn_exec_ready_bitset[ 0 ] = fd_ulong_set_bit( sched->txn_exec_ready_bitset[ 0 ], exec_tile_idx );
996 0 : break;
997 0 : }
998 0 : case FD_SCHED_TT_TXN_SIGVERIFY: {
999 0 : block->txn_sigverify_done_cnt++;
1000 0 : block->txn_sigverify_in_flight_cnt--;
1001 0 : sched->metrics->txn_sigverify_done_cnt++;
1002 0 : txn_bitset_insert( sched->sigverify_done_set, txn_idx );
1003 0 : if( txn_bitset_test( sched->exec_done_set, txn_idx ) && txn_bitset_test( sched->sigverify_done_set, txn_idx ) ) {
1004 : /* Release txn_idx if both exec and sigverify are done. This is
1005 : guaranteed to only happen once per transaction because
1006 : whichever one completed first would not release. */
1007 0 : fd_rdisp_complete_txn( sched->rdisp, txn_idx, 1 );
1008 0 : sched->txn_pool_free_cnt++;
1009 0 : block->txn_done_cnt++;
1010 0 : sched->metrics->txn_done_cnt++;
1011 0 : }
1012 :
1013 0 : FD_TEST( !fd_ulong_extract_bit( sched->sigverify_ready_bitset[ 0 ], exec_tile_idx ) );
1014 0 : sched->sigverify_ready_bitset[ 0 ] = fd_ulong_set_bit( sched->sigverify_ready_bitset[ 0 ], exec_tile_idx );
1015 0 : break;
1016 0 : }
1017 0 : }
1018 :
1019 0 : if( FD_UNLIKELY( block->dying && !block_is_in_flight( block ) ) ) {
1020 0 : if( FD_UNLIKELY( sched->active_bank_idx==bank_idx ) ) {
1021 0 : FD_LOG_CRIT(( "invariant violation: active block shouldn't be dying, bank_idx %lu, slot %lu, parent slot %lu",
1022 0 : bank_idx, block->slot, block->parent_slot ));
1023 0 : }
1024 0 : FD_LOG_DEBUG(( "dying block %lu drained", block->slot ));
1025 0 : subtree_abandon( sched, block );
1026 0 : return;
1027 0 : }
1028 :
1029 0 : if( FD_UNLIKELY( !block->dying && sched->active_bank_idx!=bank_idx ) ) {
1030 : /* Block is not dead. So we should be actively replaying it. */
1031 0 : fd_sched_block_t * active_block = block_pool_ele( sched, sched->active_bank_idx );
1032 0 : FD_LOG_CRIT(( "invariant violation: sched->active_bank_idx %lu, slot %lu, parent slot %lu, bank_idx %lu, slot %lu, parent slot %lu",
1033 0 : sched->active_bank_idx, active_block->slot, active_block->parent_slot,
1034 0 : bank_idx, block->slot, block->parent_slot ));
1035 0 : }
1036 :
1037 0 : maybe_switch_block( sched, bank_idx );
1038 0 : }
1039 :
1040 : void
1041 0 : fd_sched_block_abandon( fd_sched_t * sched, ulong bank_idx ) {
1042 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1043 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1044 :
1045 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1046 0 : if( FD_UNLIKELY( bank_idx!=sched->active_bank_idx ) ) {
1047 : /* Invariant: abandoning should only be performed on actively
1048 : replayed blocks. We impose this requirement on the caller
1049 : because the dispatcher expects blocks to be abandoned in the same
1050 : order that they were added, and having this requirement makes it
1051 : easier to please the dispatcher. */
1052 0 : sched->print_buf_sz = 0UL;
1053 0 : print_all( sched, block );
1054 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
1055 0 : FD_LOG_CRIT(( "invariant violation: active_bank_idx %lu, bank_idx %lu, slot %lu, parent slot %lu",
1056 0 : sched->active_bank_idx, bank_idx, block->slot, block->parent_slot ));
1057 0 : }
1058 :
1059 0 : subtree_abandon( sched, block );
1060 :
1061 : /* Reset the active block. */
1062 0 : FD_LOG_DEBUG(( "reset active_bank_idx %lu", sched->active_bank_idx ));
1063 0 : sched->active_bank_idx = ULONG_MAX;
1064 0 : sched->metrics->deactivate_abandoned_cnt++;
1065 0 : FD_LOG_INFO(( "block %lu abandoned", block->slot ));
1066 0 : sched->print_buf_sz = 0UL;
1067 0 : print_all( sched, block );
1068 0 : FD_LOG_DEBUG(( "%s", sched->print_buf ));
1069 0 : try_activate_block( sched );
1070 0 : }
1071 :
1072 : void
1073 0 : fd_sched_block_add_done( fd_sched_t * sched, ulong bank_idx, ulong parent_bank_idx, ulong slot ) {
1074 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1075 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1076 :
1077 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1078 0 : add_block( sched, bank_idx, parent_bank_idx );
1079 0 : block->slot = slot;
1080 0 : block->txn_parsed_cnt = UINT_MAX;
1081 0 : block->txn_exec_done_cnt = UINT_MAX;
1082 0 : block->txn_sigverify_done_cnt = UINT_MAX;
1083 0 : block->txn_done_cnt = UINT_MAX;
1084 0 : block->fec_eos = 1;
1085 0 : block->block_start_signaled = 1;
1086 0 : block->block_end_signaled = 1;
1087 0 : block->block_start_done = 1;
1088 0 : block->block_end_done = 1;
1089 0 : if( FD_LIKELY( parent_bank_idx!=ULONG_MAX ) ) {
1090 0 : fd_sched_block_t * parent_block = block_pool_ele( sched, parent_bank_idx );
1091 0 : block->parent_slot = parent_block->slot;
1092 0 : }
1093 0 : if( FD_UNLIKELY( parent_bank_idx==ULONG_MAX ) ) {
1094 : /* Assumes that a NULL parent implies the snapshot slot. */
1095 0 : block->parent_slot = ULONG_MAX;
1096 0 : block->rooted = 1;
1097 0 : sched->root_idx = bank_idx;
1098 0 : }
1099 0 : }
1100 :
1101 : void
1102 0 : fd_sched_advance_root( fd_sched_t * sched, ulong root_idx ) {
1103 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1104 0 : FD_TEST( root_idx<sched->block_cnt_max );
1105 0 : FD_TEST( sched->root_idx<sched->block_cnt_max );
1106 :
1107 0 : fd_sched_block_t * new_root = block_pool_ele( sched, root_idx );
1108 0 : fd_sched_block_t * old_root = block_pool_ele( sched, sched->root_idx );
1109 0 : if( FD_UNLIKELY( !old_root->rooted ) ) {
1110 0 : FD_LOG_CRIT(( "invariant violation: old_root is not rooted, slot %lu, parent slot %lu",
1111 0 : old_root->slot, old_root->parent_slot ));
1112 0 : }
1113 :
1114 : /* Early exit if the new root is the same as the old root. */
1115 0 : if( FD_UNLIKELY( root_idx==sched->root_idx ) ) {
1116 0 : FD_LOG_INFO(( "new root is the same as the old root, slot %lu, parent slot %lu",
1117 0 : new_root->slot, new_root->parent_slot ));
1118 0 : return;
1119 0 : }
1120 :
1121 0 : fd_sched_block_t * head = old_root;
1122 0 : head->parent_idx = ULONG_MAX;
1123 0 : fd_sched_block_t * tail = head;
1124 :
1125 0 : while( head ) {
1126 0 : FD_TEST( head->in_sched );
1127 0 : head->in_sched = 0;
1128 :
1129 0 : sched->print_buf_sz = 0UL;
1130 0 : print_block_metrics( sched, head );
1131 0 : FD_LOG_DEBUG(( "%s", sched->print_buf ));
1132 :
1133 0 : ulong child_idx = head->child_idx;
1134 0 : while( child_idx!=ULONG_MAX ) {
1135 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1136 : /* Add children to be visited. We abuse the parent_idx field to
1137 : link up the next block to visit. */
1138 0 : if( child!=new_root ) {
1139 0 : tail->parent_idx = child_idx;
1140 0 : tail = child;
1141 0 : tail->parent_idx = ULONG_MAX;
1142 0 : }
1143 0 : child_idx = child->sibling_idx;
1144 0 : }
1145 :
1146 : /* Prune the current block. We will never publish halfway into a
1147 : staging lane, because anything on the rooted fork should have
1148 : finished replaying gracefully and be out of the dispatcher. In
1149 : fact, anything that we are publishing away should be out of the
1150 : dispatcher at this point. And there should be no more in-flight
1151 : transactions. */
1152 0 : if( FD_UNLIKELY( block_is_in_flight( head ) ) ) {
1153 0 : FD_LOG_CRIT(( "invariant violation: block has transactions in flight (%u exec %u sigverify), slot %lu, parent slot %lu",
1154 0 : head->txn_exec_in_flight_cnt, head->txn_sigverify_in_flight_cnt, head->slot, head->parent_slot ));
1155 0 : }
1156 0 : if( FD_UNLIKELY( head->in_rdisp ) ) {
1157 : /* We should have removed it from the dispatcher when we were
1158 : notified of the new root, or when in-flight transactions were
1159 : drained. */
1160 0 : FD_LOG_CRIT(( "invariant violation: block is in the dispatcher, slot %lu, parent slot %lu",
1161 0 : head->slot, head->parent_slot ));
1162 0 : }
1163 0 : sched->block_pool_popcnt--;
1164 0 : fd_sched_block_t * next = block_pool_ele( sched, head->parent_idx );
1165 0 : head = next;
1166 0 : }
1167 :
1168 0 : new_root->parent_idx = ULONG_MAX;
1169 0 : sched->root_idx = root_idx;
1170 0 : }
1171 :
1172 : void
1173 0 : fd_sched_root_notify( fd_sched_t * sched, ulong root_idx ) {
1174 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1175 0 : FD_TEST( root_idx<sched->block_cnt_max );
1176 0 : FD_TEST( sched->root_idx<sched->block_cnt_max );
1177 :
1178 0 : fd_sched_block_t * block = block_pool_ele( sched, root_idx );
1179 0 : fd_sched_block_t * old_root = block_pool_ele( sched, sched->root_idx );
1180 0 : if( FD_UNLIKELY( !old_root->rooted ) ) {
1181 0 : FD_LOG_CRIT(( "invariant violation: old_root is not rooted, slot %lu, parent slot %lu",
1182 0 : old_root->slot, old_root->parent_slot ));
1183 0 : }
1184 :
1185 : /* Early exit if the new root is the same as the old root. */
1186 0 : if( FD_UNLIKELY( root_idx==sched->root_idx ) ) {
1187 0 : FD_LOG_INFO(( "new root is the same as the old root, slot %lu, parent slot %lu",
1188 0 : block->slot, block->parent_slot ));
1189 0 : return;
1190 0 : }
1191 :
1192 : /* Mark every node from the new root up through its parents to the
1193 : old root as being rooted. */
1194 0 : fd_sched_block_t * curr = block;
1195 0 : fd_sched_block_t * prev = NULL;
1196 0 : while( curr ) {
1197 0 : if( FD_UNLIKELY( !block_is_done( curr ) ) ) {
1198 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is not done, slot %lu, parent slot %lu",
1199 0 : curr->slot, curr->parent_slot ));
1200 0 : }
1201 0 : if( FD_UNLIKELY( curr->dying ) ) {
1202 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is dying, slot %lu, parent slot %lu",
1203 0 : curr->slot, curr->parent_slot ));
1204 0 : }
1205 0 : if( FD_UNLIKELY( curr->staged ) ) {
1206 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is staged, slot %lu, parent slot %lu",
1207 0 : curr->slot, curr->parent_slot ));
1208 0 : }
1209 0 : if( FD_UNLIKELY( curr->in_rdisp ) ) {
1210 0 : FD_LOG_CRIT(( "invariant violation: rooting a block that is in the dispatcher, slot %lu, parent slot %lu",
1211 0 : curr->slot, curr->parent_slot ));
1212 0 : }
1213 0 : curr->rooted = 1;
1214 0 : prev = curr;
1215 0 : curr = block_pool_ele( sched, curr->parent_idx );
1216 0 : }
1217 :
1218 : /* If we didn't reach the old root, the new root is not a descendant. */
1219 0 : if( FD_UNLIKELY( prev!=old_root ) ) {
1220 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",
1221 0 : block->slot, block->parent_slot, old_root->slot, old_root->parent_slot ));
1222 0 : }
1223 :
1224 0 : ulong old_active_bank_idx = sched->active_bank_idx;
1225 :
1226 : /* Now traverse from old root towards new root, and abandon all
1227 : minority forks. */
1228 0 : curr = old_root;
1229 0 : while( curr && curr->rooted && curr!=block ) { /* curr!=block to avoid abandoning good forks. */
1230 0 : fd_sched_block_t * rooted_child_block = NULL;
1231 0 : ulong child_idx = curr->child_idx;
1232 0 : while( child_idx!=ULONG_MAX ) {
1233 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1234 0 : if( child->rooted ) {
1235 0 : rooted_child_block = child;
1236 0 : } else {
1237 : /* This is a minority fork. */
1238 0 : FD_LOG_DEBUG(( "abandoning minority fork on block %lu", child->slot ));
1239 0 : subtree_abandon( sched, child );
1240 0 : }
1241 0 : child_idx = child->sibling_idx;
1242 0 : }
1243 0 : curr = rooted_child_block;
1244 0 : }
1245 :
1246 : /* If the active block got abandoned, we need to reset it. */
1247 0 : if( sched->active_bank_idx==ULONG_MAX ) {
1248 0 : sched->metrics->deactivate_pruned_cnt += fd_uint_if( old_active_bank_idx!=ULONG_MAX, 1U, 0U );
1249 0 : try_activate_block( sched );
1250 0 : }
1251 0 : }
1252 :
1253 : fd_txn_p_t *
1254 0 : fd_sched_get_txn( fd_sched_t * sched, ulong txn_idx ) {
1255 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1256 0 : if( FD_UNLIKELY( txn_idx>=FD_SCHED_MAX_DEPTH ) ) {
1257 0 : return NULL;
1258 0 : }
1259 0 : return sched->txn_pool+txn_idx;
1260 0 : }
1261 :
1262 : fd_hash_t *
1263 0 : fd_sched_get_poh( fd_sched_t * sched, ulong bank_idx ) {
1264 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1265 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1266 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1267 0 : return &block->poh;
1268 0 : }
1269 :
1270 : uint
1271 0 : fd_sched_get_shred_cnt( fd_sched_t * sched, ulong bank_idx ) {
1272 0 : FD_TEST( sched->canary==FD_SCHED_MAGIC );
1273 0 : FD_TEST( bank_idx<sched->block_cnt_max );
1274 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1275 0 : return block->shred_cnt;
1276 0 : }
1277 :
1278 : char *
1279 0 : fd_sched_get_state_cstr( fd_sched_t * sched ) {
1280 0 : sched->print_buf_sz = 0UL;
1281 0 : print_metrics( sched );
1282 0 : print_sched( sched );
1283 0 : return sched->print_buf;
1284 0 : }
1285 :
1286 0 : void * fd_sched_leave ( fd_sched_t * sched ) { return sched; }
1287 0 : void * fd_sched_delete( void * mem ) { return mem; }
1288 :
1289 :
1290 : /* Internal helpers. */
1291 :
1292 : static void
1293 : add_block( fd_sched_t * sched,
1294 : ulong bank_idx,
1295 0 : ulong parent_bank_idx ) {
1296 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1297 0 : FD_TEST( !block->in_sched );
1298 0 : sched->block_pool_popcnt++;
1299 :
1300 0 : block->txn_parsed_cnt = 0U;
1301 0 : block->txn_exec_in_flight_cnt = 0U;
1302 0 : block->txn_exec_done_cnt = 0U;
1303 0 : block->txn_sigverify_in_flight_cnt = 0U;
1304 0 : block->txn_sigverify_done_cnt = 0U;
1305 0 : block->txn_done_cnt = 0U;
1306 0 : block->txn_pool_max_popcnt = FD_SCHED_MAX_DEPTH-sched->txn_pool_free_cnt;
1307 0 : block->block_pool_max_popcnt = sched->block_pool_popcnt;
1308 0 : block->shred_cnt = 0U;
1309 0 : block->fec_cnt = 0U;
1310 :
1311 0 : block->mblks_rem = 0UL;
1312 0 : block->txns_rem = 0UL;
1313 0 : block->fec_buf_sz = 0U;
1314 0 : block->fec_buf_boff = 0U;
1315 0 : block->fec_buf_soff = 0U;
1316 0 : block->fec_eob = 0;
1317 0 : block->fec_sob = 1;
1318 :
1319 0 : block->fec_eos = 0;
1320 0 : block->rooted = 0;
1321 0 : block->dying = 0;
1322 0 : block->in_sched = 1;
1323 0 : block->in_rdisp = 0;
1324 0 : block->block_start_signaled = 0;
1325 0 : block->block_end_signaled = 0;
1326 0 : block->block_start_done = 0;
1327 0 : block->block_end_done = 0;
1328 0 : block->staged = 0;
1329 :
1330 0 : block->luf_depth = 0UL;
1331 :
1332 : /* New leaf node, no child, no sibling. */
1333 0 : block->child_idx = ULONG_MAX;
1334 0 : block->sibling_idx = ULONG_MAX;
1335 0 : block->parent_idx = ULONG_MAX;
1336 :
1337 0 : if( FD_UNLIKELY( parent_bank_idx==ULONG_MAX ) ) {
1338 0 : return;
1339 0 : }
1340 :
1341 : /* node->parent link */
1342 0 : fd_sched_block_t * parent_block = block_pool_ele( sched, parent_bank_idx );
1343 0 : block->parent_idx = parent_bank_idx;
1344 :
1345 : /* parent->node and sibling->node links */
1346 0 : ulong child_idx = bank_idx;
1347 0 : if( FD_LIKELY( parent_block->child_idx==ULONG_MAX ) ) { /* Optimize for no forking. */
1348 0 : parent_block->child_idx = child_idx;
1349 0 : } else {
1350 0 : fd_sched_block_t * curr_block = block_pool_ele( sched, parent_block->child_idx );
1351 0 : while( curr_block->sibling_idx!=ULONG_MAX ) {
1352 0 : curr_block = block_pool_ele( sched, curr_block->sibling_idx );
1353 0 : }
1354 0 : curr_block->sibling_idx = child_idx;
1355 0 : }
1356 :
1357 0 : if( FD_UNLIKELY( parent_block->dying ) ) {
1358 0 : block->dying = 1;
1359 0 : }
1360 0 : }
1361 :
1362 0 : #define CHECK( cond ) do { \
1363 0 : if( FD_UNLIKELY( !(cond) ) ) { \
1364 0 : return FD_SCHED_PARSER_AGAIN_LATER; \
1365 0 : } \
1366 0 : } while( 0 )
1367 :
1368 : /* CHECK that it is safe to read at least n more bytes. */
1369 0 : #define CHECK_LEFT( n ) CHECK( (n)<=(block->fec_buf_sz-block->fec_buf_soff) )
1370 :
1371 : /* Consume as much as possible from the buffer. By the end of this
1372 : function, we will either have residual data that is unparseable only
1373 : because it is a batch that straddles FEC set boundaries, or we will
1374 : have reached the end of a batch. In the former case, any remaining
1375 : bytes should be concatenated with the next FEC set for further
1376 : parsing. In the latter case, any remaining bytes should be thrown
1377 : away. */
1378 : FD_WARN_UNUSED static int
1379 0 : fd_sched_parse( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx ) {
1380 0 : while( 1 ) {
1381 0 : while( block->txns_rem>0UL ) {
1382 0 : int err;
1383 0 : if( FD_UNLIKELY( (err=fd_sched_parse_txn( sched, block, alut_ctx ))!=FD_SCHED_PARSER_OK ) ) {
1384 0 : return err;
1385 0 : }
1386 0 : }
1387 0 : if( block->txns_rem==0UL && block->mblks_rem>0UL ) {
1388 0 : CHECK_LEFT( sizeof(fd_microblock_hdr_t) );
1389 0 : fd_microblock_hdr_t * hdr = (fd_microblock_hdr_t *)fd_type_pun( block->fec_buf+block->fec_buf_soff );
1390 0 : block->fec_buf_soff += (uint)sizeof(fd_microblock_hdr_t);
1391 :
1392 0 : memcpy( block->poh.hash, hdr->hash, sizeof(block->poh.hash) );
1393 0 : block->txns_rem = hdr->txn_cnt;
1394 0 : block->mblks_rem--;
1395 0 : continue;
1396 0 : }
1397 0 : if( block->txns_rem==0UL && block->mblks_rem==0UL && block->fec_sob ) {
1398 0 : CHECK_LEFT( sizeof(ulong) );
1399 0 : FD_TEST( block->fec_buf_soff==0U );
1400 0 : block->mblks_rem = FD_LOAD( ulong, block->fec_buf );
1401 0 : block->fec_buf_soff += (uint)sizeof(ulong);
1402 : /* FIXME what happens if someone sends us mblks_rem==0UL here? */
1403 :
1404 0 : block->fec_sob = 0;
1405 0 : continue;
1406 0 : }
1407 0 : if( block->txns_rem==0UL && block->mblks_rem==0UL ) {
1408 0 : break;
1409 0 : }
1410 0 : }
1411 0 : if( block->fec_eob ) {
1412 : /* Ignore trailing bytes at the end of a batch. */
1413 0 : sched->metrics->bytes_ingested_unparsed_cnt += block->fec_buf_sz-block->fec_buf_soff;
1414 0 : block->fec_buf_boff += block->fec_buf_sz;
1415 0 : block->fec_buf_soff = 0U;
1416 0 : block->fec_buf_sz = 0U;
1417 0 : block->fec_sob = 1;
1418 0 : block->fec_eob = 0;
1419 0 : }
1420 0 : return FD_SCHED_PARSER_OK;
1421 0 : }
1422 :
1423 : FD_WARN_UNUSED static int
1424 0 : fd_sched_parse_txn( fd_sched_t * sched, fd_sched_block_t * block, fd_sched_alut_ctx_t * alut_ctx ) {
1425 0 : fd_txn_t * txn = fd_type_pun( block->txn );
1426 :
1427 : /* FIXME: For the replay pipeline, we allow up to 128 instructions per
1428 : transaction. Note that we are not concomitantly bumping the size
1429 : of fd_txn_t. We allow this because transactions like that do get
1430 : packed by other validators, so we have to replay them. Those
1431 : transactions will eventually fail in the runtime, which imposes a
1432 : limit of 64 instructions, but unfortunately they are not tossed out
1433 : at parse time and they land on chain. static_instruction_limit is
1434 : going to enforece this limit at parse time, and transactions like
1435 : that would not land on chain. Then this short term change should
1436 : be rolled back. */
1437 0 : ulong pay_sz = 0UL;
1438 0 : ulong txn_sz = fd_txn_parse_core( block->fec_buf+block->fec_buf_soff,
1439 0 : fd_ulong_min( FD_TXN_MTU, block->fec_buf_sz-block->fec_buf_soff ),
1440 0 : txn,
1441 0 : NULL,
1442 0 : &pay_sz,
1443 0 : FD_TXN_INSTR_MAX*2UL );
1444 :
1445 0 : if( FD_UNLIKELY( !pay_sz || !txn_sz ) ) {
1446 : /* Can't parse out a full transaction. */
1447 0 : return FD_SCHED_PARSER_AGAIN_LATER;
1448 0 : }
1449 :
1450 0 : if( FD_UNLIKELY( block->txn_parsed_cnt>=FD_MAX_TXN_PER_SLOT ) ) {
1451 : /* The block contains more transactions than a valid block would.
1452 : Mark the block dead instead of keep processing it. */
1453 0 : return FD_SCHED_PARSER_BAD_BLOCK;
1454 0 : }
1455 :
1456 : /* Try to expand ALUTs. */
1457 0 : int has_aluts = txn->transaction_version==FD_TXN_V0 && txn->addr_table_adtl_cnt>0;
1458 0 : int serializing = 0;
1459 0 : if( has_aluts ) {
1460 0 : fd_funk_t * funk = fd_accdb_user_v1_funk( alut_ctx->accdb );
1461 0 : uchar __attribute__((aligned(FD_SLOT_HASHES_GLOBAL_ALIGN))) slot_hashes_mem[ FD_SYSVAR_SLOT_HASHES_FOOTPRINT ];
1462 0 : fd_slot_hashes_global_t const * slot_hashes_global = fd_sysvar_slot_hashes_read( funk, alut_ctx->xid, slot_hashes_mem );
1463 0 : if( FD_LIKELY( slot_hashes_global ) ) {
1464 0 : fd_slot_hash_t * slot_hash = deq_fd_slot_hash_t_join( (uchar *)slot_hashes_global + slot_hashes_global->hashes_offset );
1465 0 : serializing = !!fd_runtime_load_txn_address_lookup_tables( txn, block->fec_buf+block->fec_buf_soff, funk, alut_ctx->xid, alut_ctx->els, slot_hash, block->aluts );
1466 0 : sched->metrics->alut_success_cnt += (uint)!serializing;
1467 0 : } else {
1468 0 : serializing = 1;
1469 0 : }
1470 0 : }
1471 :
1472 0 : ulong bank_idx = (ulong)(block-sched->block_pool);
1473 0 : ulong txn_idx = fd_rdisp_add_txn( sched->rdisp, bank_idx, txn, block->fec_buf+block->fec_buf_soff, serializing ? NULL : block->aluts, serializing );
1474 0 : FD_TEST( txn_idx!=0UL );
1475 0 : sched->metrics->txn_parsed_cnt++;
1476 0 : sched->metrics->alut_serializing_cnt += (uint)serializing;
1477 0 : sched->txn_pool_free_cnt--;
1478 0 : fd_txn_p_t * txn_p = sched->txn_pool + txn_idx;
1479 0 : txn_p->payload_sz = pay_sz;
1480 :
1481 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 );
1482 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 );
1483 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 );
1484 :
1485 0 : fd_memcpy( txn_p->payload, block->fec_buf+block->fec_buf_soff, pay_sz );
1486 0 : fd_memcpy( TXN(txn_p), txn, txn_sz );
1487 0 : txn_bitset_remove( sched->exec_done_set, txn_idx );
1488 0 : txn_bitset_remove( sched->sigverify_done_set, txn_idx );
1489 0 : block->txn_idx[ block->txn_parsed_cnt ] = txn_idx;
1490 0 : block->fec_buf_soff += (uint)pay_sz;
1491 0 : block->txn_parsed_cnt++;
1492 : #if FD_SCHED_SKIP_SIGVERIFY
1493 : txn_bitset_insert( sched->sigverify_done_set, txn_idx );
1494 : block->txn_sigverify_done_cnt++;
1495 : #endif
1496 0 : block->txns_rem--;
1497 0 : return FD_SCHED_PARSER_OK;
1498 0 : }
1499 :
1500 : #undef CHECK
1501 : #undef CHECK_LEFT
1502 :
1503 : static void
1504 0 : try_activate_block( fd_sched_t * sched ) {
1505 :
1506 : /* See if there are any allocated staging lanes that we can activate
1507 : for scheduling ... */
1508 0 : ulong staged_bitset = sched->staged_bitset;
1509 0 : while( staged_bitset ) {
1510 0 : int lane_idx = fd_ulong_find_lsb( staged_bitset );
1511 0 : staged_bitset = fd_ulong_pop_lsb( staged_bitset );
1512 :
1513 0 : ulong head_idx = sched->staged_head_bank_idx[ lane_idx ];
1514 0 : fd_sched_block_t * head_block = block_pool_ele( sched, head_idx );
1515 0 : fd_sched_block_t * parent_block = block_pool_ele( sched, head_block->parent_idx );
1516 0 : if( FD_UNLIKELY( parent_block->dying ) ) {
1517 : /* Invariant: no child of a dying block should be staged. */
1518 0 : FD_LOG_CRIT(( "invariant violation: staged_head_bank_idx %lu, slot %lu, parent slot %lu on lane %d has parent_block->dying set, slot %lu, parent slot %lu",
1519 0 : head_idx, head_block->slot, head_block->parent_slot, lane_idx, parent_block->slot, parent_block->parent_slot ));
1520 0 : }
1521 : //FIXME: restore this invariant check when we have immediate demotion of dying blocks
1522 : // if( FD_UNLIKELY( head_block->dying ) ) {
1523 : // /* Invariant: no dying block should be staged. */
1524 : // FD_LOG_CRIT(( "invariant violation: staged_head_bank_idx %lu, slot %lu, prime %lu on lane %u has head_block->dying set",
1525 : // head_idx, (ulong)head_block->block_id.slot, (ulong)head_block->block_id.prime, lane_idx ));
1526 : // }
1527 0 : if( block_is_done( parent_block ) && block_is_activatable( head_block ) ) {
1528 : /* ... Yes, on this staging lane the parent block is done. So we
1529 : can switch to the staged child. */
1530 0 : sched->active_bank_idx = head_idx;
1531 0 : sched->metrics->lane_switch_cnt++;
1532 0 : return;
1533 0 : }
1534 0 : }
1535 :
1536 : /* ... No, promote unstaged blocks. */
1537 0 : ulong root_idx = sched->root_idx;
1538 0 : if( FD_UNLIKELY( root_idx==ULONG_MAX ) ) {
1539 0 : FD_LOG_CRIT(( "invariant violation: root_idx==ULONG_MAX indicating fd_sched is unintialized" ));
1540 0 : }
1541 : /* Find and stage the longest stageable unstaged fork. This is a
1542 : policy decision. */
1543 0 : ulong depth = compute_longest_unstaged_fork( sched, root_idx );
1544 0 : if( FD_LIKELY( depth>0UL ) ) {
1545 0 : if( FD_UNLIKELY( sched->staged_bitset==fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) ) ) {
1546 : /* No more staging lanes available. All of them are occupied by
1547 : slow squatters. Demote one of them. */
1548 : //FIXME implement this, note that only empty blocks can be
1549 : //demoted, and so blocks with in-flight transactions, including
1550 : //dying in-flight blocks, shouldn't be demoted
1551 0 : FD_LOG_CRIT(( "unimplemented" ));
1552 0 : sched->metrics->lane_demoted_cnt++;
1553 : // sched->metrics->block_demoted_cnt++; for every demoted block
1554 0 : }
1555 0 : FD_TEST( sched->staged_bitset!=fd_ulong_mask_lsb( FD_SCHED_MAX_STAGING_LANES ) );
1556 0 : int lane_idx = fd_ulong_find_lsb( ~sched->staged_bitset );
1557 0 : if( FD_UNLIKELY( lane_idx>=(int)FD_SCHED_MAX_STAGING_LANES ) ) {
1558 0 : FD_LOG_CRIT(( "invariant violation: lane_idx %d, sched->staged_bitset %lx",
1559 0 : lane_idx, sched->staged_bitset ));
1560 0 : }
1561 0 : ulong head_bank_idx = stage_longest_unstaged_fork( sched, root_idx, lane_idx );
1562 0 : if( FD_UNLIKELY( head_bank_idx==ULONG_MAX ) ) {
1563 : /* We found a promotable fork depth>0. This should not happen. */
1564 0 : FD_LOG_CRIT(( "invariant violation: head_bank_idx==ULONG_MAX" ));
1565 0 : }
1566 : /* We don't bother with promotion unless the block is immediately
1567 : dispatchable. So it's okay to set the active block here. */
1568 0 : sched->active_bank_idx = head_bank_idx;
1569 0 : return;
1570 0 : }
1571 : /* No unstaged blocks to promote. So we're done. Yay. */
1572 0 : }
1573 :
1574 : static void
1575 0 : check_or_set_active_block( fd_sched_t * sched ) {
1576 0 : if( FD_UNLIKELY( sched->active_bank_idx==ULONG_MAX ) ) {
1577 0 : try_activate_block( sched );
1578 0 : } else {
1579 0 : fd_sched_block_t * active_block = block_pool_ele( sched, sched->active_bank_idx );
1580 0 : if( FD_UNLIKELY( block_should_deactivate( active_block ) ) ) {
1581 0 : sched->print_buf_sz = 0UL;
1582 0 : print_all( sched, active_block );
1583 0 : FD_LOG_NOTICE(( "%s", sched->print_buf ));
1584 0 : FD_LOG_CRIT(( "invariant violation: should have been deactivated" ));
1585 0 : }
1586 0 : }
1587 0 : }
1588 :
1589 : /* It's safe to call this function more than once on the same block. */
1590 : static void
1591 0 : subtree_abandon( fd_sched_t * sched, fd_sched_block_t * block ) {
1592 0 : if( FD_UNLIKELY( block->rooted ) ) {
1593 0 : FD_LOG_CRIT(( "invariant violation: rooted block should not be abandoned, slot %lu, parent slot %lu",
1594 0 : block->slot, block->parent_slot ));
1595 0 : }
1596 : /* All minority fork nodes pass through this function eventually. So
1597 : this is a good point to check per-node invariants for minority
1598 : forks. */
1599 0 : if( FD_UNLIKELY( block->staged && !block->in_rdisp ) ) {
1600 0 : FD_LOG_CRIT(( "invariant violation: staged block is not in the dispatcher, slot %lu, parent slot %lu",
1601 0 : block->slot, block->parent_slot ));
1602 0 : }
1603 :
1604 : /* Setting the flag is non-optional and can happen more than once. */
1605 0 : block->dying = 1;
1606 :
1607 : /* Removal from dispatcher should only happen once. */
1608 0 : if( block->in_rdisp ) {
1609 0 : fd_sched_block_t * parent = block_pool_ele( sched, block->parent_idx );
1610 0 : if( FD_UNLIKELY( !parent ) ) {
1611 : /* Only the root has no parent. Abandon should never be called on
1612 : the root. So any block we are trying to abandon should have a
1613 : parent. */
1614 0 : FD_LOG_CRIT(( "invariant violation: parent not found slot %lu, parent slot %lu",
1615 0 : block->slot, block->parent_slot ));
1616 0 : }
1617 :
1618 : /* The dispatcher expects blocks to be abandoned in the same order
1619 : that they were added on each lane. There are no requirements on
1620 : the order of abandoning if two blocks are not on the same lane,
1621 : or if a block is unstaged. This means that in general we
1622 : shouldn't abandon a child block if the parent hasn't been
1623 : abandoned yet, if and only if they are on the same lane. So wait
1624 : until we can abandon the parent, and then descend down the fork
1625 : tree to ensure orderly abandoning. */
1626 0 : int in_order = !parent->in_rdisp || /* parent is not in the dispatcher */
1627 0 : !parent->staged || /* parent is in the dispatcher but not staged */
1628 0 : !block->staged || /* parent is in the dispatcher and staged but this block is unstaged */
1629 0 : block->staging_lane!=parent->staging_lane; /* this block is on a different staging lane than its parent */
1630 :
1631 0 : 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 ) ) {
1632 0 : FD_TEST( block_pool_ele( sched, sched->active_bank_idx )==block );
1633 0 : FD_LOG_DEBUG(( "reset active_bank_idx %lu", sched->active_bank_idx ));
1634 0 : sched->active_bank_idx = ULONG_MAX;
1635 0 : }
1636 :
1637 : /* We inform the dispatcher of an abandon only when there are no
1638 : more in-flight transactions. Otherwise, if the dispatcher
1639 : recycles the same txn_id that was just abandoned, and we receive
1640 : completion of an in-flight transaction whose txn_id was just
1641 : recycled. */
1642 : // FIXME The recycling might be fine now that we no longer use
1643 : // txn_id to index into anything. We might be able to just drop
1644 : // txn_id on abandoned blocks.
1645 0 : int abandon = in_order && block->txn_exec_in_flight_cnt==0 && block->txn_sigverify_in_flight_cnt==0;
1646 :
1647 0 : if( abandon ) {
1648 0 : block->in_rdisp = 0;
1649 0 : fd_rdisp_abandon_block( sched->rdisp, (ulong)(block-sched->block_pool) );
1650 0 : sched->txn_pool_free_cnt += block->txn_parsed_cnt-block->txn_done_cnt; /* in_flight_cnt==0 */
1651 0 : sched->metrics->block_abandoned_cnt++;
1652 0 : sched->metrics->txn_abandoned_parsed_cnt += block->txn_parsed_cnt;
1653 0 : sched->metrics->txn_abandoned_exec_done_cnt += block->txn_exec_done_cnt;
1654 0 : sched->metrics->txn_abandoned_done_cnt += block->txn_done_cnt;
1655 :
1656 : /* Now release the staging lane. */
1657 : //FIXME when demote supports non-empty blocks, we should demote
1658 : //the block from the lane unconditionally and immediately,
1659 : //regardles of whether it's safe to abandon or not. So a block
1660 : //would go immediately from staged to unstaged and eventually to
1661 : //abandoned.
1662 0 : if( FD_LIKELY( block->staged ) ) {
1663 0 : block->staged = 0;
1664 0 : sched->staged_bitset = fd_ulong_clear_bit( sched->staged_bitset, (int)block->staging_lane );
1665 0 : sched->staged_head_bank_idx[ block->staging_lane ] = ULONG_MAX;
1666 0 : }
1667 0 : }
1668 0 : }
1669 :
1670 : /* Abandon the entire fork chaining off of this block. */
1671 0 : ulong child_idx = block->child_idx;
1672 0 : while( child_idx!=ULONG_MAX ) {
1673 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1674 0 : subtree_abandon( sched, child );
1675 0 : child_idx = child->sibling_idx;
1676 0 : }
1677 0 : }
1678 :
1679 : static void
1680 0 : maybe_switch_block( fd_sched_t * sched, ulong bank_idx ) {
1681 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1682 0 : if( FD_UNLIKELY( block_is_done( block ) ) ) {
1683 0 : block->in_rdisp = 0;
1684 0 : block->staged = 0;
1685 0 : fd_rdisp_remove_block( sched->rdisp, bank_idx );
1686 0 : sched->metrics->block_removed_cnt++;
1687 :
1688 : /* See if there is a child block down the same staging lane. This
1689 : is a policy decision to minimize fork churn. We could in theory
1690 : reevaluate staging lane allocation here and do promotion/demotion
1691 : as needed. */
1692 0 : ulong child_idx = block->child_idx;
1693 0 : while( child_idx!=ULONG_MAX ) {
1694 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1695 0 : if( FD_LIKELY( child->staged && child->staging_lane==block->staging_lane ) ) {
1696 : /* There is a child block down the same staging lane ... */
1697 0 : if( FD_LIKELY( !child->dying ) ) {
1698 : /* ... and the child isn't dead */
1699 0 : if( FD_UNLIKELY( !block_is_activatable( child ) ) ) {
1700 : /* ... but the child is not activatable, likely because
1701 : there are no transactions available yet. */
1702 0 : FD_LOG_DEBUG(( "reset active_bank_idx %lu", sched->active_bank_idx ));
1703 0 : sched->active_bank_idx = ULONG_MAX;
1704 0 : sched->metrics->deactivate_no_txn_cnt++;
1705 0 : try_activate_block( sched );
1706 0 : return;
1707 0 : }
1708 : /* ... and it's immediately dispatchable, so switch the active
1709 : block to it, and have the child inherit the head status of
1710 : the lane. This is the common case. */
1711 0 : sched->active_bank_idx = child_idx;
1712 0 : sched->staged_head_bank_idx[ block->staging_lane ] = child_idx;
1713 0 : if( FD_UNLIKELY( !fd_ulong_extract_bit( sched->staged_bitset, (int)block->staging_lane ) ) ) {
1714 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",
1715 0 : sched->staged_bitset, block->staging_lane, block->slot, block->parent_slot, child->slot, child->parent_slot ));
1716 0 : }
1717 0 : return;
1718 0 : } else {
1719 : /* ... but the child block is considered dead, likely because
1720 : the parser considers it invalid. */
1721 0 : FD_LOG_INFO(( "child block %lu is already dead", child->slot ));
1722 0 : subtree_abandon( sched, child );
1723 0 : break;
1724 0 : }
1725 0 : }
1726 0 : child_idx = child->sibling_idx;
1727 0 : }
1728 : /* There isn't a child block down the same staging lane. This is
1729 : the last block in the staging lane. Release the staging lane. */
1730 0 : sched->staged_bitset = fd_ulong_clear_bit( sched->staged_bitset, (int)block->staging_lane );
1731 0 : sched->staged_head_bank_idx[ block->staging_lane ] = ULONG_MAX;
1732 :
1733 : /* Reset the active block. */
1734 0 : FD_LOG_DEBUG(( "reset active_bank_idx %lu", sched->active_bank_idx ));
1735 0 : sched->active_bank_idx = ULONG_MAX;
1736 0 : sched->metrics->deactivate_no_child_cnt++;
1737 0 : try_activate_block( sched );
1738 0 : } else if( block_should_deactivate( block ) ) {
1739 : /* We exhausted the active block, but it's not fully done yet. We
1740 : are just not getting FEC sets for it fast enough. This could
1741 : happen when the network path is congested, or when the leader
1742 : simply went down. Reset the active block. */
1743 0 : sched->active_bank_idx = ULONG_MAX;
1744 0 : sched->metrics->deactivate_no_txn_cnt++;
1745 0 : try_activate_block( sched );
1746 0 : }
1747 0 : }
1748 :
1749 : FD_FN_UNUSED static ulong
1750 0 : find_and_stage_longest_unstaged_fork( fd_sched_t * sched, int lane_idx ) {
1751 0 : ulong root_idx = sched->root_idx;
1752 0 :
1753 0 : if( FD_UNLIKELY( root_idx==ULONG_MAX ) ) {
1754 0 : FD_LOG_CRIT(( "invariant violation: root_idx==ULONG_MAX indicating fd_sched is unintialized" ));
1755 0 : }
1756 0 :
1757 0 : /* First pass: compute the longest unstaged fork depth for each node
1758 0 : in the fork tree. */
1759 0 : ulong depth = compute_longest_unstaged_fork( sched, root_idx );
1760 0 :
1761 0 : /* Second pass: stage blocks on the longest unstaged fork. */
1762 0 : ulong head_bank_idx = stage_longest_unstaged_fork( sched, root_idx, lane_idx );
1763 0 :
1764 0 : if( FD_UNLIKELY( (depth>0UL && head_bank_idx==ULONG_MAX) || (depth==0UL && head_bank_idx!=ULONG_MAX) ) ) {
1765 0 : FD_LOG_CRIT(( "invariant violation: depth %lu, head_bank_idx %lu",
1766 0 : depth, head_bank_idx ));
1767 0 : }
1768 0 :
1769 0 : return head_bank_idx;
1770 0 : }
1771 :
1772 : /* Returns length of the longest stageable unstaged fork, if there is
1773 : one, and 0 otherwise. */
1774 : static ulong
1775 0 : compute_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx ) {
1776 0 : if( FD_UNLIKELY( bank_idx==ULONG_MAX ) ) {
1777 0 : FD_LOG_CRIT(( "invariant violation: bank_idx==ULONG_MAX" ));
1778 0 : }
1779 :
1780 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1781 :
1782 0 : ulong max_child_depth = 0UL;
1783 0 : ulong child_idx = block->child_idx;
1784 0 : while( child_idx!=ULONG_MAX ) {
1785 0 : ulong child_depth = compute_longest_unstaged_fork( sched, child_idx );
1786 0 : if( child_depth > max_child_depth ) {
1787 0 : max_child_depth = child_depth;
1788 0 : }
1789 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1790 0 : child_idx = child->sibling_idx;
1791 0 : }
1792 :
1793 0 : block->luf_depth = max_child_depth + fd_ulong_if( block_is_promotable( block ), 1UL, 0UL );
1794 0 : return block->luf_depth;
1795 0 : }
1796 :
1797 : static ulong
1798 0 : stage_longest_unstaged_fork_helper( fd_sched_t * sched, ulong bank_idx, int lane_idx ) {
1799 0 : if( FD_UNLIKELY( bank_idx==ULONG_MAX ) ) {
1800 0 : FD_LOG_CRIT(( "invariant violation: bank_idx==ULONG_MAX" ));
1801 0 : }
1802 :
1803 0 : fd_sched_block_t * block = block_pool_ele( sched, bank_idx );
1804 :
1805 0 : int stage_it = fd_int_if( block_is_promotable( block ), 1, 0 );
1806 0 : ulong rv = fd_ulong_if( stage_it, bank_idx, ULONG_MAX );
1807 0 : if( FD_LIKELY( stage_it ) ) {
1808 0 : block->staged = 1;
1809 0 : block->staging_lane = (ulong)lane_idx;
1810 0 : fd_rdisp_promote_block( sched->rdisp, bank_idx, block->staging_lane );
1811 0 : sched->metrics->block_promoted_cnt++;
1812 0 : }
1813 :
1814 : /* Base case: leaf node. */
1815 0 : if( block->child_idx==ULONG_MAX ) return rv;
1816 :
1817 0 : ulong max_depth = 0UL;
1818 0 : ulong best_child_idx = ULONG_MAX;
1819 0 : ulong child_idx = block->child_idx;
1820 0 : while( child_idx!=ULONG_MAX ) {
1821 0 : fd_sched_block_t * child = block_pool_ele( sched, child_idx );
1822 0 : if( child->luf_depth>max_depth ) {
1823 0 : max_depth = child->luf_depth;
1824 0 : best_child_idx = child_idx;
1825 0 : }
1826 0 : child_idx = child->sibling_idx;
1827 0 : }
1828 :
1829 : /* Recursively stage descendants. */
1830 0 : if( best_child_idx!=ULONG_MAX ) {
1831 0 : ulong head_bank_idx = stage_longest_unstaged_fork_helper( sched, best_child_idx, lane_idx );
1832 0 : rv = fd_ulong_if( rv!=ULONG_MAX, rv, head_bank_idx );
1833 0 : }
1834 :
1835 0 : return rv;
1836 0 : }
1837 :
1838 : /* Returns idx of head block of staged lane on success, idx_null
1839 : otherwise. */
1840 : static ulong
1841 0 : stage_longest_unstaged_fork( fd_sched_t * sched, ulong bank_idx, int lane_idx ) {
1842 0 : ulong head_bank_idx = stage_longest_unstaged_fork_helper( sched, bank_idx, lane_idx );
1843 0 : if( FD_LIKELY( head_bank_idx!=ULONG_MAX ) ) {
1844 0 : sched->metrics->lane_promoted_cnt++;
1845 0 : sched->staged_bitset = fd_ulong_set_bit( sched->staged_bitset, lane_idx );
1846 0 : sched->staged_head_bank_idx[ lane_idx ] = head_bank_idx;
1847 0 : }
1848 0 : return head_bank_idx;
1849 0 : }
|
