Line data Source code
1 : #include "fd_resolv_tile.h"
2 : #include "../../disco/fd_txn_m.h"
3 : #include "../../disco/topo/fd_topo.h"
4 : #include "../replay/fd_replay_tile.h"
5 : #include "../../discof/fd_startup.h"
6 : #include "../../disco/metrics/fd_metrics.h"
7 : #include "../../flamenco/accdb/fd_accdb.h"
8 : #include "../../flamenco/accdb/fd_accdb_shmem.h"
9 : #include "../../flamenco/runtime/fd_alut.h"
10 : #include "../../flamenco/runtime/fd_runtime_const.h"
11 : #include "../../flamenco/runtime/fd_system_ids_pp.h"
12 : #include "../../flamenco/runtime/fd_bank.h"
13 : #include "../../tango/fseq/fd_fseq.h"
14 : #include "../../util/pod/fd_pod_format.h"
15 : #include "../../util/fd_hash32.h"
16 :
17 : #include <time.h>
18 : #include "generated/fd_resolv_tile_seccomp.h"
19 :
20 : #if FD_HAS_AVX
21 : #include "../../util/simd/fd_avx.h"
22 : #endif
23 :
24 0 : #define IN_KIND_DEDUP (0)
25 0 : #define IN_KIND_REPLAY (1)
26 :
27 : struct blockhash {
28 : uchar b[ 32 ];
29 : };
30 :
31 : typedef struct blockhash blockhash_t;
32 :
33 : struct blockhash_map {
34 : blockhash_t key;
35 : ulong slot;
36 : };
37 :
38 : typedef struct blockhash_map blockhash_map_t;
39 :
40 : static const blockhash_t null_blockhash = { 0 };
41 :
42 : /* The blockhash ring holds recent blockhashes, so we can identify when
43 : a transaction arrives, what slot it will expire (and can no longer be
44 : packed) in. This is useful so we don't send transactions to pack
45 : that are no longer packable.
46 :
47 : Unfortunately, poorly written transaction senders frequently send
48 : transactions from millions of slots ago, so we need a large ring to
49 : be able to determine and evict these. The highest practically useful
50 : value here is around 22, which works out to 19 days of blockhash
51 : history. Beyond this, the validator is likely to be restarted, and
52 : lose the history anyway. */
53 :
54 0 : #define BLOCKHASH_LG_RING_CNT 22UL
55 0 : #define BLOCKHASH_RING_LEN (1UL<<BLOCKHASH_LG_RING_CNT)
56 :
57 : #define MAP_NAME map
58 0 : #define MAP_T blockhash_map_t
59 0 : #define MAP_KEY_T blockhash_t
60 0 : #define MAP_LG_SLOT_CNT (BLOCKHASH_LG_RING_CNT+1UL)
61 0 : #define MAP_KEY_NULL null_blockhash
62 : #if FD_HAS_AVX
63 0 : # define MAP_KEY_INVAL(k) _mm256_testz_si256( wb_ldu( (k).b ), wb_ldu( (k).b ) )
64 : #else
65 : # define MAP_KEY_INVAL(k) MAP_KEY_EQUAL(k, null_blockhash)
66 : #endif
67 0 : #define MAP_KEY_EQUAL(k0,k1) (!memcmp((k0).b,(k1).b, 32UL))
68 : #define MAP_MEMOIZE 0
69 : #define MAP_KEY_EQUAL_IS_SLOW 1
70 0 : #define MAP_KEY_HASH(key,seed) ((uint)fd_hash32( (key).b, (seed) ))
71 : #define MAP_QUERY_OPT 1
72 :
73 : #include "../../util/tmpl/fd_map_dynamic.c"
74 :
75 : typedef struct {
76 : union {
77 : ulong pool_next; /* Used when it's released */
78 : ulong lru_next; /* Used when it's acquired */
79 : }; /* .. so it's okay to store them in the same memory */
80 : ulong lru_prev;
81 :
82 : ulong map_next;
83 : ulong map_prev;
84 :
85 : blockhash_t * blockhash;
86 : uchar _[ FD_TPU_PARSED_MTU ] __attribute__((aligned(alignof(fd_txn_m_t))));
87 : } fd_stashed_txn_m_t;
88 :
89 : #define POOL_NAME pool
90 0 : #define POOL_T fd_stashed_txn_m_t
91 0 : #define POOL_NEXT pool_next
92 : #define POOL_IDX_T ulong
93 :
94 : #include "../../util/tmpl/fd_pool.c"
95 :
96 : /* We'll push at the head, which means the tail is the oldest. */
97 : #define DLIST_NAME lru_list
98 : #define DLIST_ELE_T fd_stashed_txn_m_t
99 0 : #define DLIST_PREV lru_prev
100 0 : #define DLIST_NEXT lru_next
101 :
102 : #include "../../util/tmpl/fd_dlist.c"
103 :
104 : #define MAP_NAME map_chain
105 0 : #define MAP_ELE_T fd_stashed_txn_m_t
106 : #define MAP_KEY_T blockhash_t *
107 0 : #define MAP_KEY blockhash
108 0 : #define MAP_IDX_T ulong
109 0 : #define MAP_NEXT map_next
110 0 : #define MAP_PREV map_prev
111 0 : #define MAP_KEY_HASH(k,s) fd_hash32( (*(k))->b, (s) )
112 0 : #define MAP_KEY_EQ(k0,k1) (!memcmp((*(k0))->b, (*(k1))->b, 32UL))
113 : #define MAP_OPTIMIZE_RANDOM_ACCESS_REMOVAL 1
114 : #define MAP_MULTI 1
115 :
116 : #include "../../util/tmpl/fd_map_chain.c"
117 :
118 : typedef struct {
119 : int kind;
120 :
121 : fd_wksp_t * mem;
122 : ulong chunk0;
123 : ulong wmark;
124 : ulong mtu;
125 : } fd_resolv_in_ctx_t;
126 :
127 : typedef struct {
128 : fd_wksp_t * mem;
129 : ulong chunk0;
130 : ulong wmark;
131 : ulong chunk;
132 : } fd_resolv_out_ctx_t;
133 :
134 : typedef struct {
135 : ulong round_robin_idx;
136 : ulong round_robin_cnt;
137 : ulong map_seed;
138 :
139 : int bundle_failed;
140 : ulong bundle_id;
141 :
142 : blockhash_map_t * blockhash_map;
143 :
144 : ulong flushing_slot;
145 : ulong flush_pool_idx;
146 :
147 : /* In the full client, the resolv tile is passed only a rooted bank
148 : index from replay whenever the root is advanced.
149 :
150 : This is enough to query the accounts database for that bank and
151 : retrieve the address lookup tables. Because of lifetime concerns
152 : around bank ownership, the replay tile is solely responsible for
153 : freeing the bank when it is no longer needed. To facilitate this,
154 : the resolv tile sends a message to replay when it is done with a
155 : rooted bank (after exchanging it for a new rooted bank). */
156 : fd_banks_t * banks;
157 : fd_bank_t * bank;
158 : fd_accdb_t * accdb;
159 :
160 : fd_stashed_txn_m_t * pool;
161 : map_chain_t * map_chain;
162 : lru_list_t lru_list[1];
163 :
164 : fd_startup_gate_t startup_gate[1];
165 :
166 : ulong completed_slot;
167 : ulong blockhash_ring_idx;
168 : blockhash_t blockhash_ring[ BLOCKHASH_RING_LEN ];
169 :
170 : fd_replay_root_advanced_t _rooted_slot_msg;
171 : fd_replay_slot_completed_t _completed_slot_msg;
172 :
173 : struct {
174 : ulong lut[ FD_METRICS_COUNTER_RESOLV_LUT_RESOLVED_CNT ];
175 : ulong blockhash_expired;
176 : ulong bundle_peer_failure;
177 : ulong stash[ FD_METRICS_COUNTER_RESOLV_STASH_OPERATION_CNT ];
178 : } metrics;
179 :
180 : fd_resolv_in_ctx_t in[ 64UL ];
181 :
182 : fd_resolv_out_ctx_t out_pack[ 1UL ];
183 : fd_resolv_out_ctx_t out_replay[ 1UL ];
184 :
185 : /* Scratch buffers for fd_accdb_read_one_nocache. RO accdb joiners
186 : must use the nocache API (see fd_accdb.h), which writes the account
187 : data into caller-provided buffers rather than returning a pointer
188 : into the cache. Reused across alut reads; peek_alut consumes the
189 : bytes synchronously inside fd_alut_interp_next. */
190 : uchar alut_owner[ 32UL ];
191 : uchar alut_data[ FD_RUNTIME_ACC_SZ_MAX ];
192 : } fd_resolv_ctx_t;
193 :
194 : FD_FN_CONST static inline ulong
195 0 : scratch_align( void ) {
196 0 : return fd_ulong_max( fd_ulong_max( alignof( fd_resolv_ctx_t ), pool_align() ), fd_ulong_max( map_chain_align(), map_align() ) );
197 0 : }
198 :
199 : FD_FN_PURE static inline ulong
200 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
201 0 : ulong l = FD_LAYOUT_INIT;
202 0 : l = FD_LAYOUT_APPEND( l, alignof( fd_resolv_ctx_t ), sizeof( fd_resolv_ctx_t ) );
203 0 : l = FD_LAYOUT_APPEND( l, pool_align(), pool_footprint ( 1UL<<16UL ) );
204 0 : l = FD_LAYOUT_APPEND( l, map_chain_align(), map_chain_footprint( 8192UL ) );
205 0 : l = FD_LAYOUT_APPEND( l, map_align(), map_footprint( MAP_LG_SLOT_CNT ) );
206 0 : l = FD_LAYOUT_APPEND( l, fd_accdb_align(), fd_accdb_footprint( tile->resolv.max_live_slots ) );
207 0 : return FD_LAYOUT_FINI( l, scratch_align() );
208 0 : }
209 :
210 : static inline void
211 0 : metrics_write( fd_resolv_ctx_t * ctx ) {
212 0 : FD_MCNT_SET( RESOLV, BLOCKHASH_EXPIRED, ctx->metrics.blockhash_expired );
213 0 : FD_MCNT_ENUM_COPY( RESOLV, LUT_RESOLVED, ctx->metrics.lut );
214 0 : FD_MCNT_ENUM_COPY( RESOLV, STASH_OPERATION, ctx->metrics.stash );
215 0 : FD_MCNT_SET( RESOLV, TXN_BUNDLE_PEER_FAILED, ctx->metrics.bundle_peer_failure );
216 :
217 0 : FD_ACCDB_METRICS_WRITE_RO( RESOLV, fd_accdb_metrics( ctx->accdb ) );
218 0 : }
219 :
220 : static int
221 : before_frag( fd_resolv_ctx_t * ctx,
222 : ulong in_idx,
223 : ulong seq,
224 0 : ulong sig ) {
225 0 : fd_startup_gate_busy( ctx->startup_gate );
226 :
227 0 : if( FD_UNLIKELY( ctx->in[in_idx].kind==IN_KIND_REPLAY ) )
228 0 : return sig!=REPLAY_SIG_SLOT_COMPLETED && sig!=REPLAY_SIG_ROOT_ADVANCED;
229 :
230 : /* Bundle transactions (sig==1) must arrive at pack in order. Route
231 : all bundle traffic to resolv:0. */
232 0 : if( FD_UNLIKELY( sig ) ) return ctx->round_robin_idx!=0UL;
233 :
234 0 : return (seq % ctx->round_robin_cnt) != ctx->round_robin_idx;
235 0 : }
236 :
237 : static inline void
238 : during_frag( fd_resolv_ctx_t * ctx,
239 : ulong in_idx,
240 : ulong seq FD_PARAM_UNUSED,
241 : ulong sig FD_PARAM_UNUSED,
242 : ulong chunk,
243 : ulong sz,
244 0 : ulong ctl FD_PARAM_UNUSED ) {
245 :
246 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz>ctx->in[ in_idx ].mtu ) )
247 0 : FD_LOG_ERR(( "chunk %lu %lu corrupt, not in range [%lu,%lu]", chunk, sz, ctx->in[ in_idx ].chunk0, ctx->in[ in_idx ].wmark ));
248 :
249 0 : switch( ctx->in[in_idx].kind ) {
250 0 : case IN_KIND_DEDUP: {
251 0 : uchar * src = (uchar *)fd_chunk_to_laddr( ctx->in[in_idx].mem, chunk );
252 0 : uchar * dst = (uchar *)fd_chunk_to_laddr( ctx->out_pack->mem, ctx->out_pack->chunk );
253 0 : fd_memcpy( dst, src, sz );
254 0 : break;
255 0 : }
256 0 : case IN_KIND_REPLAY: {
257 0 : if( FD_UNLIKELY( sig==REPLAY_SIG_ROOT_ADVANCED ) ) {
258 0 : ctx->_rooted_slot_msg = *(fd_replay_root_advanced_t *)fd_chunk_to_laddr_const( ctx->in[in_idx].mem, chunk );
259 0 : } else if( FD_UNLIKELY( sig==REPLAY_SIG_SLOT_COMPLETED ) ) {
260 0 : ctx->_completed_slot_msg = *(fd_replay_slot_completed_t *)fd_chunk_to_laddr_const( ctx->in[in_idx].mem, chunk );
261 0 : }
262 0 : break;
263 0 : }
264 0 : default:
265 0 : FD_LOG_ERR(( "unknown in kind %d", ctx->in[in_idx].kind ));
266 0 : }
267 0 : }
268 :
269 : /* peek_alut reads a single address lookup table from database cache. */
270 :
271 : static int
272 : peek_alut( fd_resolv_ctx_t * ctx,
273 : fd_txn_m_t * txnm,
274 : fd_alut_interp_t * interp,
275 0 : ulong alut_idx ) {
276 0 : fd_txn_t const * txn = fd_txn_m_txn_t_const ( txnm );
277 0 : uchar const * txn_payload = fd_txn_m_payload_const( txnm );
278 0 : fd_txn_acct_addr_lut_t const * addr_lut = &fd_txn_get_address_tables_const( txn )[ alut_idx ];
279 0 : fd_pubkey_t addr_lut_acc = FD_LOAD( fd_pubkey_t, txn_payload+addr_lut->addr_off );
280 :
281 : /* https://github.com/anza-xyz/agave/blob/368ea563c423b0a85cc317891187e15c9a321521/accounts-db/src/accounts.rs#L90-L94
282 :
283 : The resolv tile maps accdb read-only and so must use the nocache
284 : read API; fd_accdb_read_one would mutate writer-only shmem. */
285 0 : ulong lamports;
286 0 : int executable;
287 0 : ulong data_len;
288 0 : fd_accdb_read_one_nocache( ctx->accdb, ctx->bank->accdb_fork_id, addr_lut_acc.uc,
289 0 : &lamports, &executable, ctx->alut_owner, ctx->alut_data, &data_len );
290 0 : if( FD_UNLIKELY( !lamports ) ) return FD_RUNTIME_TXN_ERR_ADDRESS_LOOKUP_TABLE_NOT_FOUND;
291 :
292 0 : return fd_alut_interp_next( interp, &addr_lut_acc, ctx->alut_owner, ctx->alut_data, data_len );
293 0 : }
294 :
295 : /* peek_aluts reads address lookup tables from database cache.
296 : Gracefully recovers from data races and missing accounts. */
297 :
298 : static int
299 : peek_aluts( fd_resolv_ctx_t * ctx,
300 0 : fd_txn_m_t * txnm ) {
301 : /* Unpack context */
302 0 : fd_txn_t const * txn = fd_txn_m_txn_t_const ( txnm );
303 0 : uchar const * txn_payload = fd_txn_m_payload_const( txnm );
304 0 : ulong const alut_cnt = txn->addr_table_lookup_cnt;
305 0 : ulong const slot = ctx->bank->f.slot;
306 0 : fd_sysvar_cache_t const * sysvar_cache = &ctx->bank->f.sysvar_cache;
307 0 : fd_slot_hashes_t slot_hashes_view[1];
308 0 : if( FD_UNLIKELY( !fd_sysvar_cache_slot_hashes_view( sysvar_cache, slot_hashes_view ) ) ) {
309 0 : FD_LOG_ERR(( "slot hashes sysvar cache is invalid" ));
310 0 : }
311 :
312 : /* Write indirect addrs into here */
313 0 : fd_acct_addr_t * indir_addrs = fd_txn_m_alut( txnm );
314 :
315 0 : int err = FD_RUNTIME_EXECUTE_SUCCESS;
316 0 : fd_alut_interp_t interp[1];
317 0 : fd_alut_interp_new( interp, indir_addrs, txn, txn_payload, slot_hashes_view, slot );
318 0 : for( ulong i=0UL; i<alut_cnt; i++ ) {
319 0 : err = peek_alut( ctx, txnm, interp, i );
320 0 : if( FD_UNLIKELY( err ) ) break;
321 0 : }
322 :
323 0 : ulong ctr_idx;
324 0 : switch( err ) {
325 0 : case FD_RUNTIME_EXECUTE_SUCCESS: ctr_idx = FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_SUCCESS_IDX; break;
326 0 : case FD_RUNTIME_TXN_ERR_ADDRESS_LOOKUP_TABLE_NOT_FOUND: ctr_idx = FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_ACCOUNT_NOT_FOUND_IDX; break;
327 0 : case FD_RUNTIME_TXN_ERR_INVALID_ADDRESS_LOOKUP_TABLE_OWNER: ctr_idx = FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_INVALID_ACCOUNT_OWNER_IDX; break;
328 0 : case FD_RUNTIME_TXN_ERR_INVALID_ADDRESS_LOOKUP_TABLE_DATA: ctr_idx = FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_INVALID_ACCOUNT_DATA_IDX; break;
329 0 : case FD_RUNTIME_TXN_ERR_INVALID_ADDRESS_LOOKUP_TABLE_INDEX: ctr_idx = FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_INVALID_LOOKUP_INDEX_IDX; break;
330 0 : default: ctr_idx = FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_ACCOUNT_UNINITIALIZED_IDX; break;
331 0 : }
332 0 : ctx->metrics.lut[ ctr_idx ]++;
333 0 : return err;
334 0 : }
335 :
336 : static int
337 : publish_txn( fd_resolv_ctx_t * ctx,
338 : fd_stem_context_t * stem,
339 0 : fd_stashed_txn_m_t const * stashed ) {
340 0 : fd_txn_m_t * txnm = fd_chunk_to_laddr( ctx->out_pack->mem, ctx->out_pack->chunk );
341 0 : fd_memcpy( txnm, stashed->_, fd_txn_m_realized_footprint( (fd_txn_m_t *)stashed->_, 1, 0 ) );
342 :
343 0 : fd_txn_t const * txnt = fd_txn_m_txn_t( txnm );
344 :
345 0 : txnm->reference_slot = ctx->flushing_slot;
346 :
347 0 : if( FD_UNLIKELY( txnt->addr_table_adtl_cnt ) ) {
348 0 : if( FD_UNLIKELY( !ctx->bank ) ) {
349 0 : FD_MCNT_INC( RESOLV, TXN_NO_BANK, 1 );
350 0 : return 0;
351 0 : }
352 0 : int err = peek_aluts( ctx, txnm );
353 0 : if( FD_UNLIKELY( err ) ) return 0;
354 0 : }
355 :
356 0 : ulong realized_sz = fd_txn_m_realized_footprint( txnm, 1, 1 );
357 0 : ulong tspub = fd_frag_meta_ts_comp( fd_tickcount() );
358 0 : fd_stem_publish( stem, 0UL, txnm->reference_slot, ctx->out_pack->chunk, realized_sz, 0UL, 0UL, tspub );
359 0 : ctx->out_pack->chunk = fd_dcache_compact_next( ctx->out_pack->chunk, realized_sz, ctx->out_pack->chunk0, ctx->out_pack->wmark );
360 :
361 0 : return 1;
362 0 : }
363 :
364 : static inline void
365 : after_credit( fd_resolv_ctx_t * ctx,
366 : fd_stem_context_t * stem,
367 : int * opt_poll_in,
368 0 : int * charge_busy ) {
369 0 : if( FD_UNLIKELY( !fd_startup_gate_idle( ctx->startup_gate ) ) ) return;
370 :
371 0 : if( FD_LIKELY( ctx->flush_pool_idx==ULONG_MAX ) ) return;
372 :
373 0 : *charge_busy = 1;
374 0 : *opt_poll_in = 0;
375 :
376 0 : ulong next = map_chain_idx_next_const( ctx->flush_pool_idx, ULONG_MAX, ctx->pool );
377 0 : map_chain_idx_remove_fast( ctx->map_chain, ctx->flush_pool_idx, ctx->pool );
378 0 : if( FD_LIKELY( publish_txn( ctx, stem, pool_ele( ctx->pool, ctx->flush_pool_idx ) ) ) ) {
379 0 : ctx->metrics.stash[ FD_METRICS_ENUM_RESOLVE_STASH_OPERATION_V_PUBLISHED_IDX ]++;
380 0 : } else {
381 0 : ctx->metrics.stash[ FD_METRICS_ENUM_RESOLVE_STASH_OPERATION_V_REMOVED_IDX ]++;
382 0 : }
383 0 : lru_list_idx_remove( ctx->lru_list, ctx->flush_pool_idx, ctx->pool );
384 0 : pool_idx_release( ctx->pool, ctx->flush_pool_idx );
385 0 : ctx->flush_pool_idx = next;
386 0 : }
387 :
388 : /* Returns 0 if not a durable nonce transaction and 1 if it may be a
389 : durable nonce transaction */
390 :
391 : FD_FN_PURE static inline int
392 : fd_resolv_is_durable_nonce( fd_txn_t const * txn,
393 0 : uchar const * payload ) {
394 0 : if( FD_UNLIKELY( txn->instr_cnt==0 ) ) return 0;
395 :
396 0 : fd_txn_instr_t const * ix0 = &txn->instr[ 0 ];
397 0 : fd_acct_addr_t const * prog0 = fd_txn_get_acct_addrs( txn, payload ) + ix0->program_id;
398 : /* First instruction must be SystemProgram nonceAdvance instruction */
399 0 : fd_acct_addr_t const system_program[1] = { { { SYS_PROG_ID } } };
400 0 : if( FD_LIKELY( memcmp( prog0, system_program, sizeof(fd_acct_addr_t) ) ) ) return 0;
401 :
402 : /* instruction with three accounts and a four byte instruction data, a
403 : little-endian uint value 4 */
404 0 : if( FD_UNLIKELY( (ix0->data_sz!=4) | (ix0->acct_cnt!=3) ) ) return 0;
405 :
406 0 : return fd_uint_load_4( payload + ix0->data_off )==4U;
407 0 : }
408 :
409 : static inline void
410 : after_frag( fd_resolv_ctx_t * ctx,
411 : ulong in_idx,
412 : ulong seq,
413 : ulong sig,
414 : ulong sz,
415 : ulong tsorig,
416 : ulong _tspub,
417 0 : fd_stem_context_t * stem ) {
418 0 : (void)seq;
419 0 : (void)sz;
420 0 : (void)_tspub;
421 :
422 0 : if( FD_UNLIKELY( ctx->in[in_idx].kind==IN_KIND_REPLAY ) ) {
423 0 : switch( sig ) {
424 0 : case REPLAY_SIG_SLOT_COMPLETED: {
425 0 : fd_replay_slot_completed_t const * msg = &ctx->_completed_slot_msg;
426 :
427 : /* Equivocating slot with same blockhash, ignore. See fd_txncache.h on how this is possible.
428 : TODO make sure matches how agave handles it */
429 0 : if( FD_UNLIKELY( map_query( ctx->blockhash_map, *(blockhash_t *)msg->block_hash.uc, NULL ) ) ) {
430 0 : FD_LOG_WARNING(( "slot with same blockhash, ignoring: %lu", msg->slot ));
431 0 : return;
432 0 : }
433 :
434 : /* blockhash_ring is initialized to all zeros. blockhash=0 is an illegal map query */
435 0 : if( FD_UNLIKELY( memcmp( &ctx->blockhash_ring[ ctx->blockhash_ring_idx%BLOCKHASH_RING_LEN ], (uchar[ 32UL ]){ 0UL }, sizeof(blockhash_t) ) ) ) {
436 0 : blockhash_map_t * entry = map_query( ctx->blockhash_map, ctx->blockhash_ring[ ctx->blockhash_ring_idx%BLOCKHASH_RING_LEN ], NULL );
437 0 : if( FD_LIKELY( entry ) ) map_remove( ctx->blockhash_map, entry );
438 0 : }
439 :
440 0 : memcpy( ctx->blockhash_ring[ ctx->blockhash_ring_idx%BLOCKHASH_RING_LEN ].b, msg->block_hash.uc, 32UL );
441 0 : ctx->blockhash_ring_idx++;
442 :
443 0 : blockhash_map_t * blockhash = map_insert( ctx->blockhash_map, *(blockhash_t *)msg->block_hash.uc );
444 0 : blockhash->slot = msg->slot;
445 :
446 0 : blockhash_t * hash = (blockhash_t *)msg->block_hash.uc;
447 0 : ctx->flush_pool_idx = map_chain_idx_query_const( ctx->map_chain, &hash, ULONG_MAX, ctx->pool );
448 0 : ctx->flushing_slot = msg->slot;
449 :
450 0 : ctx->completed_slot = msg->slot;
451 0 : break;
452 0 : }
453 0 : case REPLAY_SIG_ROOT_ADVANCED: {
454 0 : fd_replay_root_advanced_t const * msg = &ctx->_rooted_slot_msg;
455 :
456 : /* Replace current bank with new bank */
457 0 : fd_bank_t * prev_bank = ctx->bank;
458 :
459 0 : ctx->bank = fd_banks_bank_query( ctx->banks, msg->bank_idx );
460 0 : FD_TEST( ctx->bank );
461 :
462 : /* Send slot completed message back to replay, so it can
463 : decrement the reference count of the previous bank. */
464 0 : if( FD_LIKELY( prev_bank ) ) {
465 0 : ulong tspub = fd_frag_meta_ts_comp( fd_tickcount() );
466 0 : fd_resolv_slot_exchanged_t * slot_exchanged =
467 0 : fd_type_pun( fd_chunk_to_laddr( ctx->out_replay->mem, ctx->out_replay->chunk ) );
468 0 : slot_exchanged->bank_idx = prev_bank->idx;
469 0 : fd_stem_publish( stem, 1UL, 0UL, ctx->out_replay->chunk, sizeof(fd_resolv_slot_exchanged_t), 0UL, tsorig, tspub );
470 0 : ctx->out_replay->chunk = fd_dcache_compact_next( ctx->out_replay->chunk, sizeof(fd_resolv_slot_exchanged_t), ctx->out_replay->chunk0, ctx->out_replay->wmark );
471 0 : }
472 :
473 0 : break;
474 0 : }
475 0 : default: break;
476 0 : }
477 0 : return;
478 0 : }
479 :
480 0 : fd_txn_m_t * txnm = (fd_txn_m_t *)fd_chunk_to_laddr( ctx->out_pack->mem, ctx->out_pack->chunk );
481 0 : FD_TEST( txnm->payload_sz<=FD_TPU_MTU );
482 0 : FD_TEST( txnm->txn_t_sz<=FD_TXN_MAX_SZ );
483 0 : fd_txn_t const * txnt = fd_txn_m_txn_t( txnm );
484 :
485 : /* If we find the recent blockhash, life is simple. We drop
486 : transactions that couldn't possibly execute any more, and forward
487 : to pack ones that could.
488 :
489 : If we can't find the recent blockhash ... it means one of four
490 : things,
491 :
492 : (1) The blockhash is really old (more than 19 days) or just
493 : non-existent.
494 : (2) The blockhash is not that old, but was created before this
495 : validator was started.
496 : (3) It's really new (we haven't seen the bank yet).
497 : (4) It's a durable nonce transaction, or part of a bundle (just let
498 : it pass).
499 :
500 : For durable nonce transactions, there isn't much we can do except
501 : pass them along and see if they execute.
502 :
503 : For the other three cases ... we don't want to flood pack with what
504 : might be junk transactions, so we accumulate them into a local
505 : buffer. If we later see the blockhash come to exist, we forward any
506 : buffered transactions to back. */
507 :
508 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id && (txnm->block_engine.bundle_id!=ctx->bundle_id) ) ) {
509 0 : ctx->bundle_failed = 0;
510 0 : ctx->bundle_id = txnm->block_engine.bundle_id;
511 0 : }
512 :
513 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id && ctx->bundle_failed ) ) {
514 0 : ctx->metrics.bundle_peer_failure++;
515 0 : return;
516 0 : }
517 :
518 0 : txnm->reference_slot = ctx->completed_slot;
519 :
520 0 : blockhash_t const * recent_blockhash = (blockhash_t const *)( fd_txn_m_payload( txnm )+txnt->recent_blockhash_off );
521 0 : blockhash_map_t const * blockhash = NULL;
522 0 : if( FD_LIKELY( !map_key_inval( *recent_blockhash ) ) ) {
523 0 : blockhash = map_query( ctx->blockhash_map, *recent_blockhash, NULL );
524 0 : }
525 0 : if( FD_LIKELY( blockhash ) ) {
526 0 : txnm->reference_slot = blockhash->slot;
527 0 : if( FD_UNLIKELY( txnm->reference_slot+151UL<ctx->completed_slot ) ) {
528 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id ) ) ctx->bundle_failed = 1;
529 0 : ctx->metrics.blockhash_expired++;
530 0 : return;
531 0 : }
532 0 : }
533 :
534 0 : int is_bundle_member = !!txnm->block_engine.bundle_id;
535 0 : int is_durable_nonce = fd_resolv_is_durable_nonce( txnt, fd_txn_m_payload( txnm ) );
536 :
537 0 : if( FD_UNLIKELY( !is_bundle_member && !is_durable_nonce && !blockhash ) ) {
538 0 : ulong pool_idx;
539 0 : if( FD_UNLIKELY( !pool_free( ctx->pool ) ) ) {
540 0 : pool_idx = lru_list_idx_pop_tail( ctx->lru_list, ctx->pool );
541 0 : map_chain_idx_remove_fast( ctx->map_chain, pool_idx, ctx->pool );
542 0 : ctx->metrics.stash[ FD_METRICS_ENUM_RESOLVE_STASH_OPERATION_V_OVERRUN_IDX ]++;
543 0 : } else {
544 0 : pool_idx = pool_idx_acquire( ctx->pool );
545 0 : }
546 :
547 0 : fd_stashed_txn_m_t * stash_txn = pool_ele( ctx->pool, pool_idx );
548 : /* There's a compiler bug in GCC version 12 (at least 12.1, 12.3 and
549 : 12.4) that cause it to think stash_txn is a null pointer. It
550 : then complains that the memcpy is bad and refuses to compile the
551 : memcpy below. It is possible for pool_ele to return NULL, but
552 : that can't happen because if pool_free is 0, then all the pool
553 : elements must be in the LRU list, so idx_pop_tail won't return
554 : IDX_NULL; and if pool_free returns non-zero, then
555 : pool_idx_acquire won't return POOL_IDX_NULL. */
556 0 : FD_COMPILER_FORGET( stash_txn );
557 0 : fd_memcpy( stash_txn->_, txnm, fd_txn_m_realized_footprint( txnm, 1, 0 ) );
558 0 : stash_txn->blockhash = (blockhash_t *)(fd_txn_m_payload( (fd_txn_m_t *)(stash_txn->_) ) + txnt->recent_blockhash_off);
559 0 : ctx->metrics.stash[ FD_METRICS_ENUM_RESOLVE_STASH_OPERATION_V_INSERTED_IDX ]++;
560 :
561 0 : map_chain_ele_insert( ctx->map_chain, stash_txn, ctx->pool );
562 0 : lru_list_idx_push_head( ctx->lru_list, pool_idx, ctx->pool );
563 :
564 0 : return;
565 0 : }
566 :
567 0 : if( FD_UNLIKELY( txnt->addr_table_adtl_cnt ) ) {
568 0 : if( FD_UNLIKELY( !ctx->bank ) ) {
569 0 : FD_MCNT_INC( RESOLV, TXN_NO_BANK, 1 );
570 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id ) ) ctx->bundle_failed = 1;
571 0 : return;
572 0 : }
573 :
574 0 : int result = peek_aluts( ctx, txnm );
575 0 : if( FD_UNLIKELY( result ) ) {
576 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id ) ) ctx->bundle_failed = 1;
577 0 : return;
578 0 : }
579 0 : }
580 :
581 0 : ulong realized_sz = fd_txn_m_realized_footprint( txnm, 1, 1 );
582 0 : ulong tspub = fd_frag_meta_ts_comp( fd_tickcount() );
583 0 : fd_stem_publish( stem, 0UL, txnm->reference_slot, ctx->out_pack->chunk, realized_sz, 0UL, tsorig, tspub );
584 0 : ctx->out_pack->chunk = fd_dcache_compact_next( ctx->out_pack->chunk, realized_sz, ctx->out_pack->chunk0, ctx->out_pack->wmark );
585 0 : }
586 :
587 : static void
588 : privileged_init( fd_topo_t const * topo,
589 0 : fd_topo_tile_t const * tile ) {
590 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
591 :
592 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
593 0 : fd_resolv_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_resolv_ctx_t ), sizeof( fd_resolv_ctx_t ) );
594 0 : FD_TEST( fd_rng_secure( &ctx->map_seed, sizeof(ctx->map_seed) ) );
595 0 : }
596 :
597 : static void
598 : unprivileged_init( fd_topo_t const * topo,
599 0 : fd_topo_tile_t const * tile ) {
600 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
601 :
602 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
603 0 : fd_resolv_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_resolv_ctx_t ), sizeof( fd_resolv_ctx_t ) );
604 :
605 0 : ctx->round_robin_cnt = fd_topo_tile_name_cnt( topo, tile->name );
606 0 : ctx->round_robin_idx = tile->kind_id;
607 :
608 0 : ctx->bundle_failed = 0;
609 0 : ctx->bundle_id = 0UL;
610 :
611 0 : ctx->completed_slot = 0UL;
612 0 : ctx->blockhash_ring_idx = 0UL;
613 :
614 0 : ctx->flush_pool_idx = ULONG_MAX;
615 :
616 0 : ctx->pool = pool_join( pool_new( FD_SCRATCH_ALLOC_APPEND( l, pool_align(), pool_footprint( 1UL<<16UL ) ), 1UL<<16UL ) );
617 0 : FD_TEST( ctx->pool );
618 :
619 0 : ctx->map_chain = map_chain_join( map_chain_new( FD_SCRATCH_ALLOC_APPEND( l, map_chain_align(), map_chain_footprint( 8192ULL ) ), 8192UL, ctx->map_seed ) );
620 0 : FD_TEST( ctx->map_chain );
621 :
622 0 : FD_TEST( ctx->lru_list==lru_list_join( lru_list_new( ctx->lru_list ) ) );
623 :
624 0 : memset( ctx->blockhash_ring, 0, sizeof( ctx->blockhash_ring ) );
625 0 : memset( &ctx->metrics, 0, sizeof( ctx->metrics ) );
626 :
627 0 : ctx->blockhash_map = map_join( map_new( FD_SCRATCH_ALLOC_APPEND( l, map_align(), map_footprint( MAP_LG_SLOT_CNT ) ), MAP_LG_SLOT_CNT, ctx->map_seed ) );
628 0 : FD_TEST( ctx->blockhash_map );
629 :
630 0 : FD_TEST( tile->in_cnt<=sizeof( ctx->in )/sizeof( ctx->in[ 0 ] ) );
631 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
632 0 : fd_topo_link_t const * link = &topo->links[ tile->in_link_id[ i ] ];
633 0 : fd_topo_wksp_t const * link_wksp = &topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ];
634 :
635 0 : if( FD_LIKELY( !strcmp( link->name, "replay_out" ) ) ) ctx->in[ i ].kind = IN_KIND_REPLAY;
636 0 : else if( FD_LIKELY( !strcmp( link->name, "dedup_resolv" ) ) ) ctx->in[ i ].kind = IN_KIND_DEDUP;
637 0 : else FD_LOG_ERR(( "unknown in link name '%s'", link->name ));
638 :
639 0 : ctx->in[i].mem = link_wksp->wksp;
640 0 : ctx->in[i].chunk0 = fd_dcache_compact_chunk0( ctx->in[i].mem, link->dcache );
641 0 : ctx->in[i].wmark = fd_dcache_compact_wmark ( ctx->in[i].mem, link->dcache, link->mtu );
642 0 : ctx->in[i].mtu = link->mtu;
643 0 : }
644 :
645 0 : ctx->out_pack->mem = topo->workspaces[ topo->objs[ topo->links[ tile->out_link_id[ 0 ] ].dcache_obj_id ].wksp_id ].wksp;
646 0 : ctx->out_pack->chunk0 = fd_dcache_compact_chunk0( ctx->out_pack->mem, topo->links[ tile->out_link_id[ 0 ] ].dcache );
647 0 : ctx->out_pack->wmark = fd_dcache_compact_wmark ( ctx->out_pack->mem, topo->links[ tile->out_link_id[ 0 ] ].dcache, topo->links[ tile->out_link_id[ 0 ] ].mtu );
648 0 : ctx->out_pack->chunk = ctx->out_pack->chunk0;
649 :
650 0 : ctx->out_replay->mem = topo->workspaces[ topo->objs[ topo->links[ tile->out_link_id[ 1 ] ].dcache_obj_id ].wksp_id ].wksp;
651 0 : ctx->out_replay->chunk0 = fd_dcache_compact_chunk0( ctx->out_replay->mem, topo->links[ tile->out_link_id[ 1 ] ].dcache );
652 0 : ctx->out_replay->wmark = fd_dcache_compact_wmark ( ctx->out_replay->mem, topo->links[ tile->out_link_id[ 1 ] ].dcache, topo->links[ tile->out_link_id[ 1 ] ].mtu );
653 0 : ctx->out_replay->chunk = ctx->out_replay->chunk0;
654 :
655 0 : ulong banks_obj_id = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "banks" );
656 0 : FD_TEST( banks_obj_id!=ULONG_MAX );
657 0 : ctx->banks = fd_banks_join( fd_topo_obj_laddr( topo, banks_obj_id ) );
658 0 : FD_TEST( ctx->banks );
659 0 : ctx->bank = NULL;
660 :
661 : /* Read-only join to accdb. The accdb workspace is mapped PROT_READ
662 : in this tile (see topology); the only writable external mapping
663 : is our private epoch fseq. FD_ACCDB_FD_RO is the O_RDONLY dup
664 : of the accdb data file. */
665 0 : void * _accdb_join = FD_SCRATCH_ALLOC_APPEND( l, fd_accdb_align(), fd_accdb_footprint( tile->resolv.max_live_slots ) );
666 0 : void * _accdb_shmem = fd_topo_obj_laddr( topo, tile->resolv.accdb_obj_id );
667 0 : fd_accdb_shmem_t * accdb_shmem_ro = fd_accdb_shmem_join( _accdb_shmem );
668 0 : FD_TEST( accdb_shmem_ro );
669 0 : ulong * epoch_fseq = fd_fseq_join( fd_topo_obj_laddr( topo, tile->resolv.accdb_epoch_fseq_obj_id ) );
670 0 : FD_TEST( epoch_fseq );
671 0 : ctx->accdb = fd_accdb_join_readonly( _accdb_join, accdb_shmem_ro, epoch_fseq, FD_ACCDB_FD_RO );
672 0 : FD_TEST( ctx->accdb );
673 :
674 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, scratch_align() );
675 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
676 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
677 :
678 0 : fd_startup_gate_init( ctx->startup_gate, topo, tile->in_cnt );
679 0 : }
680 :
681 : static ulong
682 : populate_allowed_seccomp( fd_topo_t const * topo,
683 : fd_topo_tile_t const * tile,
684 : ulong out_cnt,
685 0 : struct sock_filter * out ) {
686 0 : (void)topo;
687 0 : (void)tile;
688 :
689 0 : populate_sock_filter_policy_fd_resolv_tile( out_cnt, out, (uint)fd_log_private_logfile_fd(), (uint)FD_ACCDB_FD_RO );
690 0 : return sock_filter_policy_fd_resolv_tile_instr_cnt;
691 0 : }
692 :
693 : static ulong
694 : populate_allowed_fds( fd_topo_t const * topo,
695 : fd_topo_tile_t const * tile,
696 : ulong out_fds_cnt,
697 0 : int * out_fds ) {
698 0 : (void)topo;
699 0 : (void)tile;
700 :
701 0 : if( FD_UNLIKELY( out_fds_cnt<3UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
702 :
703 0 : ulong out_cnt = 0UL;
704 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
705 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
706 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
707 0 : out_fds[ out_cnt++ ] = FD_ACCDB_FD_RO; /* accounts db readonly fd */
708 0 : return out_cnt;
709 0 : }
710 :
711 0 : #define STEM_BURST (1UL)
712 :
713 : /* The default STEM_LAZY is derived from cr_max, which is the minimum
714 : depth among all reliably-consumed output links. The resolv_replay
715 : link (depth 4096) dominates this, even though it only carries ~2-3
716 : msgs/s. This makes housekeeping fire ~16x more often than necessary.
717 : We override with roughly what the default would be without accounting
718 : for it. */
719 0 : #define STEM_LAZY (128000L) /* 128 us */
720 :
721 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_resolv_ctx_t
722 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_resolv_ctx_t)
723 :
724 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
725 0 : #define STEM_CALLBACK_AFTER_CREDIT after_credit
726 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
727 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
728 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
729 :
730 : #include "../../disco/stem/fd_stem.c"
731 :
732 : fd_topo_run_tile_t fd_tile_resolv = {
733 : .name = "resolv",
734 : .populate_allowed_seccomp = populate_allowed_seccomp,
735 : .populate_allowed_fds = populate_allowed_fds,
736 : .scratch_align = scratch_align,
737 : .scratch_footprint = scratch_footprint,
738 : .privileged_init = privileged_init,
739 : .unprivileged_init = unprivileged_init,
740 : .run = stem_run,
741 : };
|
