Line data Source code
1 : #include "../tiles.h"
2 :
3 : #include "generated/fd_dedup_tile_seccomp.h"
4 :
5 : #include "../verify/fd_verify_tile.h"
6 : #include "../metrics/fd_metrics.h"
7 :
8 : #include <linux/unistd.h>
9 :
10 : /* fd_dedup provides services to deduplicate multiple streams of input
11 : fragments and present them to a mix of reliable and unreliable
12 : consumers as though they were generated by a single multi-stream
13 : producer.
14 :
15 : The dedup tile is simply a wrapper around the mux tile, that also
16 : checks the transaction signature field for duplicates and filters
17 : them out. */
18 :
19 0 : #define IN_KIND_GOSSIP (0UL)
20 0 : #define IN_KIND_VOTER (1UL)
21 0 : #define IN_KIND_VERIFY (2UL)
22 :
23 : /* fd_dedup_in_ctx_t is a context object for each in (producer) mcache
24 : connected to the dedup tile. */
25 :
26 : typedef struct {
27 : fd_wksp_t * mem;
28 : ulong chunk0;
29 : ulong wmark;
30 : } fd_dedup_in_ctx_t;
31 :
32 : /* fd_dedup_ctx_t is the context object provided to callbacks from the
33 : mux tile, and contains all state needed to progress the tile. */
34 :
35 : typedef struct {
36 : ulong tcache_depth; /* == fd_tcache_depth( tcache ), depth of this dedups's tcache (const) */
37 : ulong tcache_map_cnt; /* == fd_tcache_map_cnt( tcache ), number of slots to use for tcache map (const) */
38 : ulong * tcache_sync; /* == fd_tcache_oldest_laddr( tcache ), local join to the oldest key in the tcache */
39 : ulong * tcache_ring;
40 : ulong * tcache_map;
41 :
42 : ulong in_kind[ 64UL ];
43 : fd_dedup_in_ctx_t in[ 64UL ];
44 :
45 : int bundle_failed;
46 : ulong bundle_id;
47 : ulong bundle_idx;
48 : uchar bundle_signatures[ 4UL ][ 64UL ];
49 :
50 : fd_wksp_t * out_mem;
51 : ulong out_chunk0;
52 : ulong out_wmark;
53 : ulong out_chunk;
54 :
55 : ulong hashmap_seed;
56 :
57 : struct {
58 : ulong bundle_peer_failure_cnt;
59 : ulong dedup_fail_cnt;
60 : } metrics;
61 : } fd_dedup_ctx_t;
62 :
63 : FD_FN_CONST static inline ulong
64 3 : scratch_align( void ) {
65 3 : return alignof( fd_dedup_ctx_t );
66 3 : }
67 :
68 : FD_FN_PURE static inline ulong
69 3 : scratch_footprint( fd_topo_tile_t const * tile ) {
70 3 : ulong l = FD_LAYOUT_INIT;
71 3 : l = FD_LAYOUT_APPEND( l, alignof( fd_dedup_ctx_t ), sizeof( fd_dedup_ctx_t ) );
72 3 : l = FD_LAYOUT_APPEND( l, fd_tcache_align(), fd_tcache_footprint( tile->dedup.tcache_depth, 0UL ) );
73 3 : return FD_LAYOUT_FINI( l, scratch_align() );
74 3 : }
75 :
76 : static inline void
77 0 : metrics_write( fd_dedup_ctx_t * ctx ) {
78 0 : FD_MCNT_SET( DEDUP, TRANSACTION_BUNDLE_PEER_FAILURE, ctx->metrics.bundle_peer_failure_cnt );
79 0 : FD_MCNT_SET( DEDUP, TRANSACTION_DEDUP_FAILURE, ctx->metrics.dedup_fail_cnt );
80 0 : }
81 :
82 : /* during_frag is called between pairs for sequence number checks, as
83 : we are reading incoming frags. We don't actually need to copy the
84 : fragment here, flow control prevents it getting overrun, and
85 : downstream consumers could reuse the same chunk and workspace to
86 : improve performance.
87 :
88 : The bounds checking and copying here are defensive measures,
89 :
90 : * In a functioning system, the bounds checking should never fail,
91 : but we want to prevent an attacker with code execution on a producer
92 : tile from trivially being able to jump to a consumer tile with
93 : out of bounds chunks.
94 :
95 : * For security reasons, we have chosen to isolate all workspaces from
96 : one another, so for example, if the QUIC tile is compromised with
97 : RCE, it cannot wait until the sigverify tile has verified a transaction,
98 : and then overwrite the transaction while it's being processed by the
99 : banking stage. */
100 :
101 : static inline void
102 : during_frag( fd_dedup_ctx_t * ctx,
103 : ulong in_idx,
104 : ulong seq FD_PARAM_UNUSED,
105 : ulong sig FD_PARAM_UNUSED,
106 : ulong chunk,
107 : ulong sz,
108 0 : ulong ctl FD_PARAM_UNUSED ) {
109 :
110 0 : if( FD_UNLIKELY( chunk<ctx->in[ in_idx ].chunk0 || chunk>ctx->in[ in_idx ].wmark || sz>FD_TPU_PARSED_MTU ) )
111 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 ));
112 :
113 0 : uchar * src = (uchar *)fd_chunk_to_laddr( ctx->in[ in_idx ].mem, chunk );
114 0 : uchar * dst = (uchar *)fd_chunk_to_laddr( ctx->out_mem, ctx->out_chunk );
115 :
116 0 : if( FD_UNLIKELY( ctx->in_kind[ in_idx ]==IN_KIND_GOSSIP || ctx->in_kind[ in_idx ]==IN_KIND_VOTER ) ) {
117 0 : if( FD_UNLIKELY( sz>FD_TPU_MTU ) ) FD_LOG_ERR(( "received a gossip or voter transaction that was too large" ));
118 :
119 0 : fd_txn_m_t * txnm = (fd_txn_m_t *)dst;
120 0 : txnm->payload_sz = (ushort)sz;
121 0 : fd_memcpy( fd_txn_m_payload( txnm ), src, sz );
122 0 : txnm->block_engine.bundle_id = 0UL;
123 0 : } else {
124 0 : fd_memcpy( dst, src, sz );
125 0 : }
126 0 : }
127 :
128 : /* After the transaction has been fully received, and we know we were
129 : not overrun while reading it, check if it's a duplicate of a prior
130 : transaction.
131 :
132 : If the transaction came in from the gossip link, then it hasn't been
133 : parsed by us. So parse it here if necessary. */
134 :
135 : static inline void
136 : after_frag( fd_dedup_ctx_t * ctx,
137 : ulong in_idx,
138 : ulong seq,
139 : ulong sig,
140 : ulong sz,
141 : ulong tsorig,
142 : ulong _tspub,
143 0 : fd_stem_context_t * stem ) {
144 0 : (void)seq;
145 0 : (void)sig;
146 0 : (void)sz;
147 0 : (void)_tspub;
148 :
149 0 : fd_txn_m_t * txnm = (fd_txn_m_t *)fd_chunk_to_laddr( ctx->out_mem, ctx->out_chunk );
150 0 : FD_TEST( txnm->payload_sz<=FD_TPU_MTU );
151 0 : fd_txn_t * txn = fd_txn_m_txn_t( txnm );
152 :
153 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id && (txnm->block_engine.bundle_id!=ctx->bundle_id) ) ) {
154 0 : ctx->bundle_failed = 0;
155 0 : ctx->bundle_id = txnm->block_engine.bundle_id;
156 0 : ctx->bundle_idx = 0UL;
157 0 : }
158 :
159 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id && ctx->bundle_failed ) ) {
160 0 : ctx->metrics.bundle_peer_failure_cnt++;
161 0 : return;
162 0 : }
163 :
164 0 : if( FD_UNLIKELY( ctx->in_kind[ in_idx ]==IN_KIND_GOSSIP || ctx->in_kind[ in_idx]==IN_KIND_VOTER ) ) {
165 : /* Transactions coming in from these links are not parsed.
166 :
167 : We'll need to parse it so it's ready for downstream consumers.
168 : Equally importantly, we need to parse to extract the signature
169 : for dedup. Just parse it right into the output dcache. */
170 0 : txnm->txn_t_sz = (ushort)fd_txn_parse( fd_txn_m_payload( txnm ), txnm->payload_sz, txn, NULL );
171 0 : if( FD_UNLIKELY( !txnm->txn_t_sz ) ) FD_LOG_ERR(( "fd_txn_parse failed for vote transactions that should have been sigverified" ));
172 :
173 0 : if( FD_UNLIKELY( ctx->in_kind[ in_idx ]==IN_KIND_GOSSIP ) ) FD_MCNT_INC( DEDUP, GOSSIPED_VOTES_RECEIVED, 1UL );
174 0 : }
175 :
176 0 : int is_dup = 0;
177 0 : if( FD_LIKELY( !txnm->block_engine.bundle_id ) ) {
178 : /* Compute fd_hash(signature) for dedup. */
179 0 : ulong ha_dedup_tag = fd_hash( ctx->hashmap_seed, fd_txn_m_payload( txnm )+txn->signature_off, 64UL );
180 :
181 0 : FD_TCACHE_INSERT( is_dup, *ctx->tcache_sync, ctx->tcache_ring, ctx->tcache_depth, ctx->tcache_map, ctx->tcache_map_cnt, ha_dedup_tag );
182 0 : } else {
183 : /* Make sure bundles don't contain a duplicate transaction inside
184 : the bundle, which would not be valid. */
185 :
186 0 : for( ulong i=0UL; i<ctx->bundle_idx; i++ ) {
187 0 : if( !memcmp( ctx->bundle_signatures[ i ], fd_txn_m_payload( txnm )+txn->signature_off, 64UL ) ) {
188 0 : is_dup = 1;
189 0 : break;
190 0 : }
191 0 : }
192 :
193 0 : if( FD_UNLIKELY( ctx->bundle_idx>4UL ) ) FD_LOG_ERR(( "bundle_idx %lu > 4", ctx->bundle_idx ));
194 0 : else if( FD_UNLIKELY( ctx->bundle_idx==4UL ) ) ctx->bundle_idx++;
195 0 : else fd_memcpy( ctx->bundle_signatures[ ctx->bundle_idx++ ], fd_txn_m_payload( txnm )+txn->signature_off, 64UL );
196 0 : }
197 :
198 0 : if( FD_LIKELY( is_dup ) ) {
199 0 : if( FD_UNLIKELY( txnm->block_engine.bundle_id ) ) ctx->bundle_failed = 1;
200 :
201 0 : ctx->metrics.dedup_fail_cnt++;
202 0 : } else {
203 0 : ulong realized_sz = fd_txn_m_realized_footprint( txnm, 1, 0 );
204 0 : ulong tspub = (ulong)fd_frag_meta_ts_comp( fd_tickcount() );
205 0 : fd_stem_publish( stem, 0UL, 0, ctx->out_chunk, realized_sz, 0UL, tsorig, tspub );
206 0 : ctx->out_chunk = fd_dcache_compact_next( ctx->out_chunk, realized_sz, ctx->out_chunk0, ctx->out_wmark );
207 0 : }
208 0 : }
209 :
210 : static void
211 : privileged_init( fd_topo_t * topo,
212 0 : fd_topo_tile_t * tile ) {
213 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
214 :
215 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
216 0 : fd_dedup_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_dedup_ctx_t ), sizeof( fd_dedup_ctx_t ) );
217 0 : FD_TEST( fd_rng_secure( &ctx->hashmap_seed, 8U ) );
218 0 : }
219 :
220 : static void
221 : unprivileged_init( fd_topo_t * topo,
222 0 : fd_topo_tile_t * tile ) {
223 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
224 :
225 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
226 0 : fd_dedup_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_dedup_ctx_t ), sizeof( fd_dedup_ctx_t ) );
227 0 : fd_tcache_t * tcache = fd_tcache_join( fd_tcache_new( FD_SCRATCH_ALLOC_APPEND( l, fd_tcache_align(), fd_tcache_footprint( tile->dedup.tcache_depth, 0) ), tile->dedup.tcache_depth, 0 ) );
228 0 : if( FD_UNLIKELY( !tcache ) ) FD_LOG_ERR(( "fd_tcache_new failed" ));
229 :
230 0 : ctx->bundle_failed = 0;
231 0 : ctx->bundle_id = 0UL;
232 0 : ctx->bundle_idx = 0UL;
233 :
234 0 : memset( &ctx->metrics, 0, sizeof( ctx->metrics ) );
235 :
236 0 : ctx->tcache_depth = fd_tcache_depth ( tcache );
237 0 : ctx->tcache_map_cnt = fd_tcache_map_cnt ( tcache );
238 0 : ctx->tcache_sync = fd_tcache_oldest_laddr( tcache );
239 0 : ctx->tcache_ring = fd_tcache_ring_laddr ( tcache );
240 0 : ctx->tcache_map = fd_tcache_map_laddr ( tcache );
241 :
242 0 : FD_TEST( tile->in_cnt<=sizeof( ctx->in )/sizeof( ctx->in[ 0 ] ) );
243 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
244 0 : fd_topo_link_t * link = &topo->links[ tile->in_link_id[ i ] ];
245 0 : fd_topo_wksp_t * link_wksp = &topo->workspaces[ topo->objs[ link->dcache_obj_id ].wksp_id ];
246 :
247 0 : ctx->in[i].mem = link_wksp->wksp;
248 0 : ctx->in[i].chunk0 = fd_dcache_compact_chunk0( ctx->in[i].mem, link->dcache );
249 0 : ctx->in[i].wmark = fd_dcache_compact_wmark ( ctx->in[i].mem, link->dcache, link->mtu );
250 :
251 0 : if( FD_UNLIKELY( !strcmp( link->name, "gossip_dedup" ) ) ) {
252 0 : ctx->in_kind[ i ] = IN_KIND_GOSSIP;
253 0 : } else if( FD_UNLIKELY( !strcmp( link->name, "voter_dedup" ) ) ) {
254 0 : ctx->in_kind[ i ] = IN_KIND_VOTER;
255 0 : } else if( FD_UNLIKELY( !strcmp( link->name, "verify_dedup" ) ) ) {
256 0 : ctx->in_kind[ i ] = IN_KIND_VERIFY;
257 0 : } else {
258 0 : FD_LOG_ERR(( "unexpected link name %s", link->name ));
259 0 : }
260 0 : }
261 :
262 0 : ctx->out_mem = topo->workspaces[ topo->objs[ topo->links[ tile->out_link_id[ 0 ] ].dcache_obj_id ].wksp_id ].wksp;
263 0 : ctx->out_chunk0 = fd_dcache_compact_chunk0( ctx->out_mem, topo->links[ tile->out_link_id[ 0 ] ].dcache );
264 0 : ctx->out_wmark = fd_dcache_compact_wmark ( ctx->out_mem, topo->links[ tile->out_link_id[ 0 ] ].dcache, topo->links[ tile->out_link_id[ 0 ] ].mtu );
265 0 : ctx->out_chunk = ctx->out_chunk0;
266 :
267 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, 1UL );
268 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
269 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
270 0 : }
271 :
272 : static ulong
273 : populate_allowed_seccomp( fd_topo_t const * topo,
274 : fd_topo_tile_t const * tile,
275 : ulong out_cnt,
276 0 : struct sock_filter * out ) {
277 0 : (void)topo;
278 0 : (void)tile;
279 :
280 0 : populate_sock_filter_policy_fd_dedup_tile( out_cnt, out, (uint)fd_log_private_logfile_fd() );
281 0 : return sock_filter_policy_fd_dedup_tile_instr_cnt;
282 0 : }
283 :
284 : static ulong
285 : populate_allowed_fds( fd_topo_t const * topo,
286 : fd_topo_tile_t const * tile,
287 : ulong out_fds_cnt,
288 0 : int * out_fds ) {
289 0 : (void)topo;
290 0 : (void)tile;
291 :
292 0 : if( FD_UNLIKELY( out_fds_cnt<2UL ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
293 :
294 0 : ulong out_cnt = 0UL;
295 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
296 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
297 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
298 0 : return out_cnt;
299 0 : }
300 :
301 0 : #define STEM_BURST (1UL)
302 :
303 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_dedup_ctx_t
304 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_dedup_ctx_t)
305 :
306 0 : #define STEM_CALLBACK_METRICS_WRITE metrics_write
307 0 : #define STEM_CALLBACK_DURING_FRAG during_frag
308 0 : #define STEM_CALLBACK_AFTER_FRAG after_frag
309 :
310 : #include "../stem/fd_stem.c"
311 :
312 : fd_topo_run_tile_t fd_tile_dedup = {
313 : .name = "dedup",
314 : .populate_allowed_seccomp = populate_allowed_seccomp,
315 : .populate_allowed_fds = populate_allowed_fds,
316 : .scratch_align = scratch_align,
317 : .scratch_footprint = scratch_footprint,
318 : .privileged_init = privileged_init,
319 : .unprivileged_init = unprivileged_init,
320 : .run = stem_run,
321 : };
|