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 <linux/futex.h>
19 : #include "generated/fd_resolv_tile_seccomp.h"
20 :
21 : #if FD_HAS_AVX
22 : #include "../../util/simd/fd_avx.h"
23 : #endif
24 :
25 0 : #define IN_KIND_DEDUP (0)
26 0 : #define IN_KIND_REPLAY (1)
27 :
28 : struct blockhash {
29 : uchar b[ 32 ];
30 : };
31 :
32 : typedef struct blockhash blockhash_t;
33 :
34 : struct blockhash_map {
35 : blockhash_t key;
36 : ulong slot;
37 : ulong block_height;
38 : /* The map_chain variables: */
39 : uint next;
40 : /* It's probably not worth 4 bytes of cache per element to store
41 : chain_prev, since the query/delete ratio is probably about 1000:1,
42 : but alignof is 8, so we'd just store padding here instead */
43 : uint prev;
44 : };
45 :
46 : typedef struct blockhash_map blockhash_map_t;
47 :
48 : /* The blockhash ring holds recent blockhashes, so we can identify when
49 : a transaction arrives, what slot it will expire (and can no longer be
50 : packed) in. This is useful so we don't send transactions to pack
51 : that are no longer packable.
52 :
53 : Similarly, we also store the nonce version of the blockhash, which is
54 : stored in nonce accounts when a transaction executes in the following
55 : slot. Since nonce transactions stay valid for arbitrarily long, this
56 : just helps us order nonce transactions that try to advance the same
57 : account.
58 :
59 : Unfortunately, poorly written transaction senders frequently send
60 : transactions from millions of slots ago, so we need a large ring to
61 : be able to determine and evict these. The highest practically useful
62 : value here is around 22, which works out to 19 days of blockhash
63 : history. Beyond this, the validator is likely to be restarted, and
64 : lose the history anyway. */
65 :
66 0 : #define BLOCKHASH_LG_RING_CNT 22UL
67 0 : #define BLOCKHASH_RING_LEN (1UL<<BLOCKHASH_LG_RING_CNT)
68 :
69 : #define MAP_NAME map
70 0 : #define MAP_ELE_T blockhash_map_t
71 : #define MAP_KEY_T blockhash_t
72 0 : #define MAP_IDX_T uint
73 : #define MAP_OPTIMIZE_RANDOM_ACCESS_REMOVAL 1
74 0 : #define MAP_KEY_EQ(k0,k1) (!memcmp((k0)->b,(k1)->b, 32UL))
75 0 : #define MAP_KEY_HASH(key,seed) ((uint)fd_hash32( (key)->b, (seed) ))
76 :
77 : #include "../../util/tmpl/fd_map_chain.c"
78 :
79 : typedef struct {
80 : union {
81 : ulong pool_next; /* Used when it's released */
82 : ulong lru_next; /* Used when it's acquired */
83 : }; /* .. so it's okay to store them in the same memory */
84 : ulong lru_prev;
85 :
86 : ulong map_next;
87 : ulong map_prev;
88 :
89 : blockhash_t * blockhash;
90 : uchar _[ FD_TPU_PARSED_MTU ] __attribute__((aligned(alignof(fd_txn_m_t))));
91 : } fd_stashed_txn_m_t;
92 :
93 : #define POOL_NAME pool
94 0 : #define POOL_T fd_stashed_txn_m_t
95 0 : #define POOL_NEXT pool_next
96 : #define POOL_IDX_T ulong
97 :
98 : #include "../../util/tmpl/fd_pool.c"
99 :
100 : /* We'll push at the head, which means the tail is the oldest. */
101 : #define DLIST_NAME lru_list
102 : #define DLIST_ELE_T fd_stashed_txn_m_t
103 0 : #define DLIST_PREV lru_prev
104 0 : #define DLIST_NEXT lru_next
105 :
106 : #include "../../util/tmpl/fd_dlist.c"
107 :
108 : #define MAP_NAME map_chain
109 0 : #define MAP_ELE_T fd_stashed_txn_m_t
110 : #define MAP_KEY_T blockhash_t *
111 0 : #define MAP_KEY blockhash
112 0 : #define MAP_IDX_T ulong
113 0 : #define MAP_NEXT map_next
114 0 : #define MAP_PREV map_prev
115 0 : #define MAP_KEY_HASH(k,s) fd_hash32( (*(k))->b, (s) )
116 0 : #define MAP_KEY_EQ(k0,k1) (!memcmp((*(k0))->b, (*(k1))->b, 32UL))
117 : #define MAP_OPTIMIZE_RANDOM_ACCESS_REMOVAL 1
118 : #define MAP_MULTI 1
119 :
120 : #include "../../util/tmpl/fd_map_chain.c"
121 :
122 : typedef struct {
123 : int kind;
124 :
125 : fd_wksp_t * mem;
126 : ulong chunk0;
127 : ulong wmark;
128 : ulong mtu;
129 : } fd_resolv_in_ctx_t;
130 :
131 : typedef struct {
132 : fd_wksp_t * mem;
133 : ulong chunk0;
134 : ulong wmark;
135 : ulong chunk;
136 : } fd_resolv_out_ctx_t;
137 :
138 : typedef struct {
139 : ulong round_robin_idx;
140 : ulong round_robin_cnt;
141 : ulong map_seed;
142 :
143 : int bundle_failed;
144 : ulong bundle_id;
145 :
146 : map_t * blockhash_map;
147 : map_t * nonce_blockhash_map;
148 :
149 : ulong flushing_block_height;
150 : ulong flush_pool_idx;
151 :
152 : /* In the full client, the resolv tile is passed only a rooted bank
153 : index from replay whenever the root is advanced.
154 :
155 : This is enough to query the accounts database for that bank and
156 : retrieve the address lookup tables. Because of lifetime concerns
157 : around bank ownership, the replay tile is solely responsible for
158 : freeing the bank when it is no longer needed. To facilitate this,
159 : the resolv tile sends a message to replay when it is done with a
160 : rooted bank (after exchanging it for a new rooted bank). */
161 : fd_banks_t * banks;
162 : fd_bank_t * bank;
163 : fd_accdb_t * accdb;
164 :
165 : fd_stashed_txn_m_t * pool;
166 : map_chain_t * map_chain;
167 : lru_list_t lru_list[1];
168 :
169 : fd_startup_gate_t startup_gate[1];
170 :
171 : ulong completed_slot;
172 : ulong completed_block_height;
173 : /* The total number of blockhashes that have been inserted in
174 : blockhash_ring (and since nonce_blockhash_ring is parallel, also in
175 : nonce_blockhash_ring. */
176 : ulong blockhash_ring_idx;
177 :
178 : /* These are the pools used the blockhash_map */
179 : blockhash_map_t blockhash_ring [ BLOCKHASH_RING_LEN ];
180 : blockhash_map_t nonce_blockhash_ring[ BLOCKHASH_RING_LEN ];
181 :
182 : fd_replay_root_advanced_t _rooted_slot_msg;
183 : fd_replay_slot_completed_t _completed_slot_msg;
184 :
185 : struct {
186 : ulong lut[ FD_METRICS_COUNTER_RESOLV_TXN_LUT_RESOLVE_ATTEMPTED_CNT ];
187 : ulong txn_dropped[ FD_METRICS_ENUM_RESOLV_TXN_DROP_REASON_CNT ];
188 : ulong stash_inserted;
189 : ulong stash_removed[ FD_METRICS_COUNTER_RESOLV_STASH_TXN_REMOVED_CNT ];
190 : } metrics;
191 :
192 : fd_resolv_in_ctx_t in[ 64UL ];
193 :
194 : fd_resolv_out_ctx_t out_pack[ 1UL ];
195 : fd_resolv_out_ctx_t out_replay[ 1UL ];
196 :
197 : /* Scratch buffers for fd_accdb_read_one_nocache. RO accdb joiners
198 : must use the nocache API (see fd_accdb.h), which writes the account
199 : data into caller-provided buffers rather than returning a pointer
200 : into the cache. Reused across alut reads; peek_alut consumes the
201 : bytes synchronously inside fd_alut_interp_next. */
202 : uchar alut_owner[ 32UL ];
203 : uchar alut_data[ FD_RUNTIME_ACC_SZ_MAX ];
204 : } fd_resolv_ctx_t;
205 :
206 : FD_FN_CONST static inline ulong
207 0 : scratch_align( void ) {
208 0 : return fd_ulong_max( fd_ulong_max( alignof( fd_resolv_ctx_t ), pool_align() ), fd_ulong_max( map_chain_align(), map_align() ) );
209 0 : }
210 :
211 : FD_FN_PURE static inline ulong
212 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
213 : /* The map is normally full, so make the chain cnt a bit bigger */
214 0 : ulong map_chain_cnt = 2UL*map_chain_cnt_est( BLOCKHASH_RING_LEN );
215 0 : ulong l = FD_LAYOUT_INIT;
216 0 : l = FD_LAYOUT_APPEND( l, alignof( fd_resolv_ctx_t ), sizeof( fd_resolv_ctx_t ) );
217 0 : l = FD_LAYOUT_APPEND( l, pool_align(), pool_footprint ( 1UL<<16UL ) );
218 0 : l = FD_LAYOUT_APPEND( l, map_chain_align(), map_chain_footprint( 8192UL ) );
219 0 : l = FD_LAYOUT_APPEND( l, map_align(), map_footprint( map_chain_cnt ) );
220 0 : l = FD_LAYOUT_APPEND( l, map_align(), map_footprint( map_chain_cnt ) );
221 0 : l = FD_LAYOUT_APPEND( l, fd_accdb_align(), fd_accdb_footprint( tile->resolv.max_live_slots, 0 ) );
222 0 : return FD_LAYOUT_FINI( l, scratch_align() );
223 0 : }
224 :
225 : static inline void
226 0 : metrics_write( fd_resolv_ctx_t * ctx ) {
227 0 : FD_MCNT_ENUM_COPY( RESOLV, TXN_DROPPED, ctx->metrics.txn_dropped );
228 0 : FD_MCNT_ENUM_COPY( RESOLV, TXN_LUT_RESOLVE_ATTEMPTED, ctx->metrics.lut );
229 0 : FD_MCNT_SET( RESOLV, STASH_TXN_INSERTED, ctx->metrics.stash_inserted );
230 0 : FD_MCNT_ENUM_COPY( RESOLV, STASH_TXN_REMOVED, ctx->metrics.stash_removed );
231 :
232 0 : FD_ACCDB_METRICS_WRITE_RO( RESOLV, fd_accdb_metrics( ctx->accdb ) );
233 0 : }
234 :
235 : static int
236 : before_frag( fd_resolv_ctx_t * ctx,
237 : ulong in_idx,
238 : ulong seq,
239 0 : ulong sig ) {
240 0 : fd_startup_gate_busy( ctx->startup_gate );
241 :
242 0 : if( FD_UNLIKELY( ctx->in[in_idx].kind==IN_KIND_REPLAY ) )
243 0 : return sig!=REPLAY_SIG_SLOT_COMPLETED && sig!=REPLAY_SIG_ROOT_ADVANCED;
244 :
245 : /* Bundle transactions (sig==1) must arrive at pack in order. Route
246 : all bundle traffic to resolv:0. */
247 0 : if( FD_UNLIKELY( sig ) ) return ctx->round_robin_idx!=0UL;
248 :
249 0 : return (seq % ctx->round_robin_cnt) != ctx->round_robin_idx;
250 0 : }
251 :
252 : static inline void
253 : during_frag( fd_resolv_ctx_t * ctx,
254 : ulong in_idx,
255 : ulong seq FD_PARAM_UNUSED,
256 : ulong sig FD_PARAM_UNUSED,
257 : ulong chunk,
258 : ulong sz,
259 0 : ulong ctl FD_PARAM_UNUSED ) {
260 :
261 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz>ctx->in[ in_idx ].mtu ) )
262 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 ));
263 :
264 0 : switch( ctx->in[in_idx].kind ) {
265 0 : case IN_KIND_DEDUP: {
266 0 : uchar * src = (uchar *)fd_chunk_to_laddr( ctx->in[in_idx].mem, chunk );
267 0 : uchar * dst = (uchar *)fd_chunk_to_laddr( ctx->out_pack->mem, ctx->out_pack->chunk );
268 0 : fd_memcpy( dst, src, sz );
269 0 : break;
270 0 : }
271 0 : case IN_KIND_REPLAY: {
272 0 : if( FD_UNLIKELY( sig==REPLAY_SIG_ROOT_ADVANCED ) ) {
273 0 : ctx->_rooted_slot_msg = *(fd_replay_root_advanced_t *)fd_chunk_to_laddr_const( ctx->in[in_idx].mem, chunk );
274 0 : } else if( FD_UNLIKELY( sig==REPLAY_SIG_SLOT_COMPLETED ) ) {
275 0 : ctx->_completed_slot_msg = *(fd_replay_slot_completed_t *)fd_chunk_to_laddr_const( ctx->in[in_idx].mem, chunk );
276 0 : }
277 0 : break;
278 0 : }
279 0 : default:
280 0 : FD_LOG_ERR(( "unknown in kind %d", ctx->in[in_idx].kind ));
281 0 : }
282 0 : }
283 :
284 : /* peek_alut reads a single address lookup table from database cache. */
285 :
286 : static int
287 : peek_alut( fd_resolv_ctx_t * ctx,
288 : fd_txn_m_t * txnm,
289 : fd_alut_interp_t * interp,
290 0 : ulong alut_idx ) {
291 0 : fd_txn_t const * txn = fd_txn_m_txn_t_const ( txnm );
292 0 : uchar const * txn_payload = fd_txn_m_payload_const( txnm );
293 0 : fd_txn_acct_addr_lut_t const * addr_lut = &fd_txn_get_address_tables_const( txn )[ alut_idx ];
294 0 : fd_pubkey_t addr_lut_acc = FD_LOAD( fd_pubkey_t, txn_payload+addr_lut->addr_off );
295 :
296 : /* https://github.com/anza-xyz/agave/blob/368ea563c423b0a85cc317891187e15c9a321521/accounts-db/src/accounts.rs#L90-L94
297 :
298 : The resolv tile maps accdb read-only and so must use the nocache
299 : read API; fd_accdb_read_one would mutate writer-only shmem. */
300 0 : ulong lamports;
301 0 : int executable;
302 0 : ulong data_len;
303 0 : fd_accdb_read_one_nocache( ctx->accdb, ctx->bank->accdb_fork_id, addr_lut_acc.uc,
304 0 : &lamports, &executable, ctx->alut_owner, ctx->alut_data, &data_len );
305 0 : if( FD_UNLIKELY( !lamports ) ) return FD_RUNTIME_TXN_ERR_ADDRESS_LOOKUP_TABLE_NOT_FOUND;
306 :
307 0 : return fd_alut_interp_next( interp, &addr_lut_acc, ctx->alut_owner, ctx->alut_data, data_len );
308 0 : }
309 :
310 : /* peek_aluts reads address lookup tables from database cache.
311 : Gracefully recovers from data races and missing accounts. */
312 :
313 : static int
314 : peek_aluts( fd_resolv_ctx_t * ctx,
315 0 : fd_txn_m_t * txnm ) {
316 : /* Unpack context */
317 0 : fd_txn_t const * txn = fd_txn_m_txn_t_const ( txnm );
318 0 : uchar const * txn_payload = fd_txn_m_payload_const( txnm );
319 0 : ulong const alut_cnt = txn->addr_table_lookup_cnt;
320 0 : ulong const slot = ctx->bank->f.slot;
321 0 : fd_sysvar_cache_t const * sysvar_cache = &ctx->bank->f.sysvar_cache;
322 0 : fd_slot_hashes_t slot_hashes_view[1];
323 0 : if( FD_UNLIKELY( !fd_sysvar_cache_slot_hashes_view( sysvar_cache, slot_hashes_view ) ) ) {
324 0 : ctx->metrics.lut[ FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_ACCOUNT_NOT_FOUND_IDX ]++;
325 0 : return FD_RUNTIME_TXN_ERR_ACCOUNT_NOT_FOUND;
326 0 : }
327 :
328 : /* Write indirect addrs into here */
329 0 : fd_acct_addr_t * indir_addrs = fd_txn_m_alut( txnm );
330 :
331 0 : int err = FD_RUNTIME_EXECUTE_SUCCESS;
332 0 : fd_alut_interp_t interp[1];
333 0 : fd_alut_interp_new( interp, indir_addrs, txn, txn_payload, slot_hashes_view, slot );
334 0 : for( ulong i=0UL; i<alut_cnt; i++ ) {
335 0 : err = peek_alut( ctx, txnm, interp, i );
336 0 : if( FD_UNLIKELY( err ) ) break;
337 0 : }
338 :
339 0 : ulong ctr_idx;
340 0 : switch( err ) {
341 0 : case FD_RUNTIME_EXECUTE_SUCCESS: ctr_idx = FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_SUCCESS_IDX; break;
342 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;
343 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;
344 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;
345 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;
346 0 : default: ctr_idx = FD_METRICS_ENUM_LUT_RESOLVE_RESULT_V_ACCOUNT_UNINITIALIZED_IDX; break;
347 0 : }
348 0 : ctx->metrics.lut[ ctr_idx ]++;
349 0 : return err;
350 0 : }
351 :
352 : static int
353 : publish_txn( fd_resolv_ctx_t * ctx,
354 : fd_stem_context_t * stem,
355 0 : fd_stashed_txn_m_t const * stashed ) {
356 0 : fd_txn_m_t * txnm = fd_chunk_to_laddr( ctx->out_pack->mem, ctx->out_pack->chunk );
357 0 : fd_memcpy( txnm, stashed->_, fd_txn_m_realized_footprint( (fd_txn_m_t *)stashed->_, 1, 0 ) );
358 :
359 0 : fd_txn_t const * txnt = fd_txn_m_txn_t( txnm );
360 :
361 0 : txnm->reference_block_height = ctx->flushing_block_height;
362 :
363 0 : if( FD_UNLIKELY( txnt->addr_table_adtl_cnt ) ) {
364 0 : if( FD_UNLIKELY( !ctx->bank ) ) {
365 0 : ctx->metrics.txn_dropped[ FD_METRICS_ENUM_RESOLV_TXN_DROP_REASON_V_NO_BANK_IDX ]++;
366 0 : return 0;
367 0 : }
368 0 : int err = peek_aluts( ctx, txnm );
369 0 : if( FD_UNLIKELY( err ) ) return 0;
370 0 : }
371 :
372 0 : ulong realized_sz = fd_txn_m_realized_footprint( txnm, 1, 1 );
373 0 : ulong tspub = fd_frag_meta_ts_comp( fd_tickcount() );
374 0 : fd_stem_publish( stem, 0UL, txnm->reference_block_height, ctx->out_pack->chunk, realized_sz, 0UL, 0UL, tspub );
375 0 : ctx->out_pack->chunk = fd_dcache_compact_next( ctx->out_pack->chunk, realized_sz, ctx->out_pack->chunk0, ctx->out_pack->wmark );
376 :
377 0 : return 1;
378 0 : }
379 :
380 : static inline void
381 : after_credit( fd_resolv_ctx_t * ctx,
382 : fd_stem_context_t * stem,
383 : int * opt_poll_in,
384 0 : int * charge_busy ) {
385 0 : if( FD_UNLIKELY( !fd_startup_gate_idle( ctx->startup_gate ) ) ) return;
386 :
387 0 : if( FD_LIKELY( ctx->flush_pool_idx==ULONG_MAX ) ) return;
388 :
389 0 : *charge_busy = 1;
390 0 : *opt_poll_in = 0;
391 :
392 0 : ulong next = map_chain_idx_next_const( ctx->flush_pool_idx, ULONG_MAX, ctx->pool );
393 0 : map_chain_idx_remove_fast( ctx->map_chain, ctx->flush_pool_idx, ctx->pool );
394 0 : if( FD_LIKELY( publish_txn( ctx, stem, pool_ele( ctx->pool, ctx->flush_pool_idx ) ) ) ) {
395 0 : ctx->metrics.stash_removed[ FD_METRICS_ENUM_STASH_REMOVE_REASON_V_PUBLISHED_IDX ]++;
396 0 : } else {
397 0 : ctx->metrics.stash_removed[ FD_METRICS_ENUM_STASH_REMOVE_REASON_V_FAILED_IDX ]++;
398 0 : }
399 0 : lru_list_idx_remove( ctx->lru_list, ctx->flush_pool_idx, ctx->pool );
400 0 : pool_idx_release( ctx->pool, ctx->flush_pool_idx );
401 0 : ctx->flush_pool_idx = next;
402 0 : }
403 :
404 : static inline void
405 : after_frag( fd_resolv_ctx_t * ctx,
406 : ulong in_idx,
407 : ulong seq,
408 : ulong sig,
409 : ulong sz,
410 : ulong tsorig,
411 : ulong _tspub,
412 0 : fd_stem_context_t * stem ) {
413 0 : (void)seq;
414 0 : (void)sz;
415 0 : (void)_tspub;
416 :
417 0 : if( FD_UNLIKELY( ctx->in[in_idx].kind==IN_KIND_REPLAY ) ) {
418 0 : switch( sig ) {
419 0 : case REPLAY_SIG_SLOT_COMPLETED: {
420 0 : fd_replay_slot_completed_t const * msg = &ctx->_completed_slot_msg;
421 :
422 : /* Equivocating slot with same blockhash, ignore. See fd_txncache.h on how this is possible.
423 : TODO make sure matches how agave handles it */
424 0 : if( FD_UNLIKELY( map_ele_query( ctx->blockhash_map, (blockhash_t *)msg->block_hash.uc, NULL, ctx->blockhash_ring ) ) ) {
425 0 : FD_LOG_WARNING(( "slot with same blockhash, ignoring: %lu", msg->slot ));
426 0 : return;
427 0 : }
428 :
429 : /* if we've already inserted BLOCKHASH_RING_LEN, then we need to
430 : delete the one we're overwriting */
431 0 : if( FD_UNLIKELY( ctx->blockhash_ring_idx>=BLOCKHASH_RING_LEN ) ) {
432 0 : map_idx_remove_fast( ctx->blockhash_map, ctx->blockhash_ring_idx%BLOCKHASH_RING_LEN, ctx->blockhash_ring );
433 0 : map_idx_remove_fast( ctx->nonce_blockhash_map, ctx->blockhash_ring_idx%BLOCKHASH_RING_LEN, ctx->nonce_blockhash_ring );
434 0 : }
435 0 : blockhash_map_t * entry = ctx->blockhash_ring +(ctx->blockhash_ring_idx%BLOCKHASH_RING_LEN);
436 0 : blockhash_map_t * nonce_entry = ctx->nonce_blockhash_ring+(ctx->blockhash_ring_idx%BLOCKHASH_RING_LEN);
437 :
438 : /* See fd_durable_nonce_from_blockhash */
439 0 : struct {
440 0 : char tag[13];
441 0 : fd_hash_t bh[1];
442 0 : } hash_buf[1] = {{ .tag = "DURABLE_NONCE", .bh = { msg->block_hash } }};
443 :
444 0 : memcpy( entry->key.b, msg->block_hash.uc, 32UL ); fd_sha256_hash( hash_buf, sizeof(hash_buf), nonce_entry->key.b );
445 0 : entry->slot = msg->slot; nonce_entry->slot = msg->slot;
446 0 : entry->block_height = msg->block_height; nonce_entry->block_height = msg->block_height;
447 :
448 0 : map_ele_insert( ctx->blockhash_map, entry, ctx->blockhash_ring );
449 0 : map_ele_insert( ctx->nonce_blockhash_map, nonce_entry, ctx->nonce_blockhash_ring );
450 0 : ctx->blockhash_ring_idx++;
451 :
452 0 : blockhash_t * hash = (blockhash_t *)msg->block_hash.uc;
453 0 : ctx->flush_pool_idx = map_chain_idx_query_const( ctx->map_chain, &hash, ULONG_MAX, ctx->pool );
454 0 : ctx->flushing_block_height = msg->block_height;
455 :
456 0 : ctx->completed_slot = msg->slot;
457 0 : ctx->completed_block_height = msg->block_height;
458 0 : break;
459 0 : }
460 0 : case REPLAY_SIG_ROOT_ADVANCED: {
461 0 : fd_replay_root_advanced_t const * msg = &ctx->_rooted_slot_msg;
462 :
463 : /* Replace current bank with new bank */
464 0 : fd_bank_t * prev_bank = ctx->bank;
465 :
466 0 : ctx->bank = fd_banks_bank_query( ctx->banks, msg->bank_idx );
467 0 : FD_TEST( ctx->bank );
468 :
469 : /* Send slot completed message back to replay, so it can
470 : decrement the reference count of the previous bank. */
471 0 : if( FD_LIKELY( prev_bank ) ) {
472 0 : ulong tspub = fd_frag_meta_ts_comp( fd_tickcount() );
473 0 : fd_resolv_slot_exchanged_t * slot_exchanged =
474 0 : fd_type_pun( fd_chunk_to_laddr( ctx->out_replay->mem, ctx->out_replay->chunk ) );
475 0 : slot_exchanged->bank_idx = prev_bank->idx;
476 0 : fd_stem_publish( stem, 1UL, 0UL, ctx->out_replay->chunk, sizeof(fd_resolv_slot_exchanged_t), 0UL, tsorig, tspub );
477 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 );
478 0 : }
479 :
480 0 : break;
481 0 : }
482 0 : default: break;
483 0 : }
484 0 : return;
485 0 : }
486 :
487 0 : fd_txn_m_t * txnm = (fd_txn_m_t *)fd_chunk_to_laddr( ctx->out_pack->mem, ctx->out_pack->chunk );
488 0 : FD_TEST( txnm->payload_sz<=FD_TPU_MTU );
489 0 : FD_TEST( txnm->txn_t_sz<=FD_TXN_MAX_SZ );
490 0 : fd_txn_t const * txnt = fd_txn_m_txn_t( txnm );
491 :
492 : /* If the transaction doesn't look like a nonce transaction and we
493 : find the recent blockhash, life is simple. We drop transactions
494 : that couldn't possibly execute any more, and forward to pack ones
495 : that could.
496 :
497 : If we can't find the recent blockhash ... it means one of three
498 : things,
499 :
500 : (1) The blockhash is really old (more than 19 days) or just
501 : non-existent.
502 : (2) The blockhash is not that old, but was created before this
503 : validator was started.
504 : (3) It's really new (we haven't seen the bank yet).
505 :
506 : For durable nonce transactions, we map the value in the recent
507 : blockhash field to a slot and its block height. This doesn't
508 : immediately let us discard any transactions, but it lets pack order
509 : nonce transactions, which allows it to throw out old ones.
510 :
511 : For the other three cases ... we don't want to flood pack with what
512 : might be junk transactions, so we accumulate them into a local
513 : buffer. If we later see the blockhash come to exist, we forward any
514 : buffered transactions to back. */
515 :
516 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id && (txnm->block_engine.bundle_id!=ctx->bundle_id) ) ) {
517 0 : ctx->bundle_failed = 0;
518 0 : ctx->bundle_id = txnm->block_engine.bundle_id;
519 0 : }
520 :
521 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id && ctx->bundle_failed ) ) {
522 0 : ctx->metrics.txn_dropped[ FD_METRICS_ENUM_RESOLV_TXN_DROP_REASON_V_BUNDLE_PEER_FAILED_IDX ]++;
523 0 : return;
524 0 : }
525 :
526 0 : txnm->reference_block_height = ctx->completed_block_height;
527 :
528 0 : int is_durable_nonce = fd_disco_tpu_is_durable_nonce( txnt, fd_txn_m_payload( txnm ) );
529 0 : map_t const * map = fd_ptr_if( is_durable_nonce, ctx->nonce_blockhash_map, ctx->blockhash_map );
530 0 : blockhash_map_t const * pool = is_durable_nonce ? ctx->nonce_blockhash_ring : ctx->blockhash_ring;
531 :
532 0 : blockhash_t const * recent_blockhash = (blockhash_t const *)( fd_txn_m_payload( txnm )+txnt->recent_blockhash_off );
533 :
534 0 : blockhash_map_t const * blockhash = map_ele_query_const( map, recent_blockhash, NULL, pool );
535 0 : if( FD_LIKELY( blockhash ) ) {
536 0 : txnm->reference_block_height = blockhash->block_height;
537 0 : if( FD_UNLIKELY( (!is_durable_nonce) & (txnm->reference_block_height+151UL<ctx->completed_block_height) ) ) {
538 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id ) ) ctx->bundle_failed = 1;
539 0 : ctx->metrics.txn_dropped[ FD_METRICS_ENUM_RESOLV_TXN_DROP_REASON_V_BLOCKHASH_EXPIRED_IDX ]++;
540 0 : return;
541 0 : }
542 0 : }
543 :
544 0 : int is_bundle_member = !!txnm->block_engine.bundle_id;
545 :
546 0 : if( FD_UNLIKELY( !is_bundle_member && !is_durable_nonce && !blockhash ) ) {
547 0 : ulong pool_idx;
548 0 : if( FD_UNLIKELY( !pool_free( ctx->pool ) ) ) {
549 0 : pool_idx = lru_list_idx_pop_tail( ctx->lru_list, ctx->pool );
550 0 : map_chain_idx_remove_fast( ctx->map_chain, pool_idx, ctx->pool );
551 0 : ctx->metrics.stash_removed[ FD_METRICS_ENUM_STASH_REMOVE_REASON_V_EVICTED_IDX ]++;
552 0 : } else {
553 0 : pool_idx = pool_idx_acquire( ctx->pool );
554 0 : }
555 :
556 0 : fd_stashed_txn_m_t * stash_txn = pool_ele( ctx->pool, pool_idx );
557 : /* There's a compiler bug in GCC version 12 (at least 12.1, 12.3 and
558 : 12.4) that cause it to think stash_txn is a null pointer. It
559 : then complains that the memcpy is bad and refuses to compile the
560 : memcpy below. It is possible for pool_ele to return NULL, but
561 : that can't happen because if pool_free is 0, then all the pool
562 : elements must be in the LRU list, so idx_pop_tail won't return
563 : IDX_NULL; and if pool_free returns non-zero, then
564 : pool_idx_acquire won't return POOL_IDX_NULL. */
565 0 : FD_COMPILER_FORGET( stash_txn );
566 0 : fd_memcpy( stash_txn->_, txnm, fd_txn_m_realized_footprint( txnm, 1, 0 ) );
567 0 : stash_txn->blockhash = (blockhash_t *)(fd_txn_m_payload( (fd_txn_m_t *)(stash_txn->_) ) + txnt->recent_blockhash_off);
568 0 : ctx->metrics.stash_inserted++;
569 :
570 0 : map_chain_ele_insert( ctx->map_chain, stash_txn, ctx->pool );
571 0 : lru_list_idx_push_head( ctx->lru_list, pool_idx, ctx->pool );
572 :
573 0 : return;
574 0 : }
575 :
576 0 : if( FD_UNLIKELY( txnt->addr_table_adtl_cnt ) ) {
577 0 : if( FD_UNLIKELY( !ctx->bank ) ) {
578 0 : ctx->metrics.txn_dropped[ FD_METRICS_ENUM_RESOLV_TXN_DROP_REASON_V_NO_BANK_IDX ]++;
579 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id ) ) ctx->bundle_failed = 1;
580 0 : return;
581 0 : }
582 :
583 0 : int result = peek_aluts( ctx, txnm );
584 0 : if( FD_UNLIKELY( result ) ) {
585 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id ) ) ctx->bundle_failed = 1;
586 0 : return;
587 0 : }
588 0 : }
589 :
590 0 : ulong realized_sz = fd_txn_m_realized_footprint( txnm, 1, 1 );
591 0 : ulong tspub = fd_frag_meta_ts_comp( fd_tickcount() );
592 0 : fd_stem_publish( stem, 0UL, txnm->reference_block_height, ctx->out_pack->chunk, realized_sz, 0UL, tsorig, tspub );
593 0 : ctx->out_pack->chunk = fd_dcache_compact_next( ctx->out_pack->chunk, realized_sz, ctx->out_pack->chunk0, ctx->out_pack->wmark );
594 0 : }
595 :
596 : static void
597 : privileged_init( fd_topo_t const * topo,
598 0 : fd_topo_tile_t const * tile ) {
599 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
600 :
601 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
602 0 : fd_resolv_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_resolv_ctx_t ), sizeof( fd_resolv_ctx_t ) );
603 0 : FD_TEST( fd_rng_secure( &ctx->map_seed, sizeof(ctx->map_seed) ) );
604 0 : }
605 :
606 : static void
607 : unprivileged_init( fd_topo_t const * topo,
608 0 : fd_topo_tile_t const * tile ) {
609 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
610 :
611 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
612 0 : fd_resolv_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_resolv_ctx_t ), sizeof( fd_resolv_ctx_t ) );
613 :
614 0 : ctx->round_robin_cnt = fd_topo_tile_name_cnt( topo, tile->name );
615 0 : ctx->round_robin_idx = tile->kind_id;
616 :
617 0 : ctx->bundle_failed = 0;
618 0 : ctx->bundle_id = 0UL;
619 :
620 0 : ctx->completed_slot = 0UL;
621 0 : ctx->completed_block_height = 0UL;
622 0 : ctx->blockhash_ring_idx = 0UL;
623 :
624 0 : ctx->flush_pool_idx = ULONG_MAX;
625 :
626 0 : ctx->pool = pool_join( pool_new( FD_SCRATCH_ALLOC_APPEND( l, pool_align(), pool_footprint( 1UL<<16UL ) ), 1UL<<16UL ) );
627 0 : FD_TEST( ctx->pool );
628 :
629 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 ) );
630 0 : FD_TEST( ctx->map_chain );
631 :
632 0 : FD_TEST( ctx->lru_list==lru_list_join( lru_list_new( ctx->lru_list ) ) );
633 :
634 0 : memset( ctx->blockhash_ring, 0, sizeof( ctx->blockhash_ring ) );
635 0 : memset( ctx->nonce_blockhash_ring, 0, sizeof( ctx->nonce_blockhash_ring ) );
636 0 : memset( &ctx->metrics, 0, sizeof( ctx->metrics ) );
637 :
638 0 : ulong map_chain_cnt = 2UL*map_chain_cnt_est( BLOCKHASH_RING_LEN );
639 0 : ulong footprint = map_footprint( map_chain_cnt );
640 0 : ctx->blockhash_map = map_join( map_new( FD_SCRATCH_ALLOC_APPEND( l, map_align(), footprint ), map_chain_cnt, ctx->map_seed ) );
641 0 : ctx->nonce_blockhash_map = map_join( map_new( FD_SCRATCH_ALLOC_APPEND( l, map_align(), footprint ), map_chain_cnt, ctx->map_seed ) );
642 0 : FD_TEST( ctx->blockhash_map );
643 0 : FD_TEST( ctx->nonce_blockhash_map );
644 :
645 0 : FD_TEST( tile->in_cnt<=sizeof( ctx->in )/sizeof( ctx->in[ 0 ] ) );
646 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
647 0 : fd_topo_link_t const * link = &topo->links[ tile->in_link_id[ i ] ];
648 0 : fd_topo_wksp_t const * link_wksp = &topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ];
649 :
650 0 : if( FD_LIKELY( !strcmp( link->name, "replay_slot" ) ) ) ctx->in[ i ].kind = IN_KIND_REPLAY;
651 0 : else if( FD_LIKELY( !strcmp( link->name, "dedup_resolv" ) ) ) ctx->in[ i ].kind = IN_KIND_DEDUP;
652 0 : else FD_LOG_ERR(( "unknown in link name '%s'", link->name ));
653 :
654 0 : ctx->in[i].mem = link_wksp->wksp;
655 0 : ctx->in[i].chunk0 = fd_dcache_compact_chunk0( ctx->in[i].mem, link->dcache );
656 0 : ctx->in[i].wmark = fd_dcache_compact_wmark ( ctx->in[i].mem, link->dcache, link->mtu );
657 0 : ctx->in[i].mtu = link->mtu;
658 0 : }
659 :
660 0 : ctx->out_pack->mem = topo->workspaces[ topo->objs[ topo->links[ tile->out_link_id[ 0 ] ].dcache_obj_id ].wksp_id ].wksp;
661 0 : ctx->out_pack->chunk0 = fd_dcache_compact_chunk0( ctx->out_pack->mem, topo->links[ tile->out_link_id[ 0 ] ].dcache );
662 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 );
663 0 : ctx->out_pack->chunk = ctx->out_pack->chunk0;
664 :
665 0 : ctx->out_replay->mem = topo->workspaces[ topo->objs[ topo->links[ tile->out_link_id[ 1 ] ].dcache_obj_id ].wksp_id ].wksp;
666 0 : ctx->out_replay->chunk0 = fd_dcache_compact_chunk0( ctx->out_replay->mem, topo->links[ tile->out_link_id[ 1 ] ].dcache );
667 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 );
668 0 : ctx->out_replay->chunk = ctx->out_replay->chunk0;
669 :
670 0 : ulong banks_obj_id = fd_pod_queryf_ulong( topo->props, ULONG_MAX, "banks" );
671 0 : FD_TEST( banks_obj_id!=ULONG_MAX );
672 0 : ctx->banks = fd_banks_join( fd_topo_obj_laddr( topo, banks_obj_id ) );
673 0 : FD_TEST( ctx->banks );
674 0 : ctx->bank = NULL;
675 :
676 : /* Read-only join to accdb. The accdb workspace is mapped PROT_READ
677 : in this tile (see topology); the only writable external mapping
678 : is our private epoch fseq. FD_ACCDB_FD_RO is the O_RDONLY dup
679 : of the accdb data file. */
680 0 : void * _accdb_join = FD_SCRATCH_ALLOC_APPEND( l, fd_accdb_align(), fd_accdb_footprint( tile->resolv.max_live_slots, 0 ) );
681 0 : void * _accdb_shmem = fd_topo_obj_laddr( topo, tile->resolv.accdb_obj_id );
682 0 : fd_accdb_shmem_t * accdb_shmem_ro = fd_accdb_shmem_join( _accdb_shmem );
683 0 : FD_TEST( accdb_shmem_ro );
684 0 : ulong * epoch_fseq = fd_fseq_join( fd_topo_obj_laddr( topo, tile->resolv.accdb_epoch_fseq_obj_id ) );
685 0 : FD_TEST( epoch_fseq );
686 0 : ctx->accdb = fd_accdb_join_readonly( _accdb_join, accdb_shmem_ro, epoch_fseq, FD_ACCDB_FD_RO );
687 0 : FD_TEST( ctx->accdb );
688 :
689 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, scratch_align() );
690 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
691 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
692 :
693 0 : fd_startup_gate_init( ctx->startup_gate, topo, tile->in_cnt );
694 0 : }
695 :
696 : static ulong
697 : populate_allowed_seccomp( fd_topo_t const * topo,
698 : fd_topo_tile_t const * tile,
699 : ulong out_cnt,
700 0 : struct sock_filter * out ) {
701 0 : (void)topo;
702 0 : (void)tile;
703 :
704 0 : populate_sock_filter_policy_fd_resolv_tile( out_cnt, out, (uint)fd_log_private_logfile_fd(), (uint)FD_ACCDB_FD_RO );
705 0 : return sock_filter_policy_fd_resolv_tile_instr_cnt;
706 0 : }
707 :
708 : static ulong
709 : populate_allowed_fds( fd_topo_t const * topo,
710 : fd_topo_tile_t const * tile,
711 : ulong out_fds_cnt,
712 0 : int * out_fds ) {
713 0 : (void)topo;
714 0 : (void)tile;
715 :
716 0 : if( FD_UNLIKELY( out_fds_cnt<3UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
717 :
718 0 : ulong out_cnt = 0UL;
719 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
720 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
721 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
722 0 : out_fds[ out_cnt++ ] = FD_ACCDB_FD_RO; /* accounts db readonly fd */
723 0 : return out_cnt;
724 0 : }
725 :
726 0 : #define STEM_BURST (1UL)
727 :
728 : /* The default STEM_LAZY is derived from cr_max, which is the minimum
729 : depth among all reliably-consumed output links. The resolv_replay
730 : link (depth 4096) dominates this, even though it only carries ~2-3
731 : msgs/s. This makes housekeeping fire ~16x more often than necessary.
732 : We override with roughly what the default would be without accounting
733 : for it. */
734 0 : #define STEM_LAZY (128000L) /* 128 us */
735 :
736 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_resolv_ctx_t
737 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_resolv_ctx_t)
738 :
739 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
740 0 : #define STEM_CALLBACK_AFTER_CREDIT after_credit
741 0 : #define STEM_CALLBACK_BEFORE_FRAG before_frag
742 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
743 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
744 :
745 : #include "../../disco/stem/fd_stem.c"
746 :
747 : fd_topo_run_tile_t fd_tile_resolv = {
748 : .name = "resolv",
749 : .populate_allowed_seccomp = populate_allowed_seccomp,
750 : .populate_allowed_fds = populate_allowed_fds,
751 : .scratch_align = scratch_align,
752 : .scratch_footprint = scratch_footprint,
753 : .privileged_init = privileged_init,
754 : .unprivileged_init = unprivileged_init,
755 : .run = stem_run,
756 : };
|
