Line data Source code
1 : #ifndef HEADER_fd_src_disco_topo_fd_topo_h
2 : #define HEADER_fd_src_disco_topo_fd_topo_h
3 :
4 : #include "../stem/fd_stem.h"
5 : #include "../quic/fd_tpu.h"
6 : #include "../../tango/fd_tango.h"
7 : #include "../../waltz/xdp/fd_xdp1.h"
8 :
9 : /* Maximum number of workspaces that may be present in a topology. */
10 : #define FD_TOPO_MAX_WKSPS (256UL)
11 : /* Maximum number of links that may be present in a topology. */
12 : #define FD_TOPO_MAX_LINKS (256UL)
13 : /* Maximum number of tiles that may be present in a topology. */
14 : #define FD_TOPO_MAX_TILES (256UL)
15 : /* Maximum number of objects that may be present in a topology. */
16 : #define FD_TOPO_MAX_OBJS (4096UL)
17 : /* Maximum number of links that may go into any one tile in the
18 : topology. */
19 : #define FD_TOPO_MAX_TILE_IN_LINKS ( 128UL)
20 : /* Maximum number of links that a tile may write to. */
21 : #define FD_TOPO_MAX_TILE_OUT_LINKS ( 32UL)
22 : /* Maximum number of objects that a tile can use. */
23 : #define FD_TOPO_MAX_TILE_OBJS ( 256UL)
24 :
25 : /* Maximum number of additional ip addresses */
26 : #define FD_NET_MAX_SRC_ADDR 4
27 :
28 : /* A workspace is a Firedance specific memory management structure that
29 : sits on top of 1 or more memory mapped gigantic or huge pages mounted
30 : to the hugetlbfs. */
31 : typedef struct {
32 : ulong id; /* The ID of this workspace. Indexed from [0, wksp_cnt). When placed in a topology, the ID must be the index of the workspace in the workspaces list. */
33 : char name[ 13UL ]; /* The name of this workspace, like "pack". There can be at most one of each workspace name in a topology. */
34 :
35 : ulong numa_idx; /* The index of the NUMA node on the system that this workspace should be allocated from. */
36 :
37 : /* Computed fields. These are not supplied as configuration but calculated as needed. */
38 : struct {
39 : ulong page_sz; /* The size of the pages that this workspace is backed by. One of FD_PAGE_SIZE_*. */
40 : ulong page_cnt; /* The number of pages that must be mapped to this workspace to store all the data needed by consumers. */
41 : ulong part_max; /* The maximum number of partitions in the underlying workspace. There can only be this many allocations made at any one time. */
42 :
43 : fd_wksp_t * wksp; /* The workspace memory in the local process. */
44 : ulong known_footprint; /* Total size in bytes of all data in Firedancer that will be stored in this workspace at startup. */
45 : ulong total_footprint; /* Total size in bytes of all data in Firedancer that could be stored in this workspace, includes known data and loose data. */
46 : };
47 : } fd_topo_wksp_t;
48 :
49 : /* A link is an mcache in a workspace that has one producer and one or
50 : more consumers. A link may optionally also have a dcache, that holds
51 : fragments referred to by the mcache entries.
52 :
53 : A link belongs to exactly one workspace. A link has exactly one
54 : producer, and 1 or more consumers. Each consumer is either reliable
55 : or not reliable. A link has a depth and a MTU, which correspond to
56 : the depth and MTU of the mcache and dcache respectively. A MTU of
57 : zero means no dcache is needed, as there is no data. */
58 : typedef struct {
59 : ulong id; /* The ID of this link. Indexed from [0, link_cnt). When placed in a topology, the ID must be the index of the link in the links list. */
60 : char name[ 13UL ]; /* The name of this link, like "pack_bank". There can be multiple of each link name in a topology. */
61 : ulong kind_id; /* The ID of this link within its name. If there are N links of a particular name, they have IDs [0, N). The pair (name, kind_id) uniquely identifies a link, as does "id" on its own. */
62 :
63 : int is_reasm; /* If the link is a reassembly buffer. */
64 : ulong depth; /* The depth of the mcache representing the link. */
65 : ulong mtu; /* The MTU of data fragments in the mcache. A value of 0 means there is no dcache. */
66 : ulong burst; /* The max amount of MTU sized data fragments that might be bursted to the dcache. */
67 :
68 : ulong mcache_obj_id;
69 : ulong dcache_obj_id;
70 : ulong reasm_obj_id;
71 :
72 : /* Computed fields. These are not supplied as configuration but calculated as needed. */
73 : struct {
74 : fd_frag_meta_t * mcache; /* The mcache of this link. */
75 : void * dcache; /* The dcache of this link, if it has one. */
76 : fd_tpu_reasm_t * reasm; /* The reassembly buffer of this link, if it has one. */
77 : };
78 : } fd_topo_link_t;
79 :
80 : /* A tile is a unique process that is spawned by Firedancer to represent
81 : one thread of execution. Firedancer sandboxes all tiles to their own
82 : process for security reasons.
83 :
84 : A tile belongs to exactly one workspace. A tile is a consumer of 0
85 : or more links, it's inputs. A tile is a producer of 0 or more output
86 : links.
87 :
88 : All input links will be automatically polled by the tile
89 : infrastructure, and output links will automatically source and manage
90 : credits from consumers. */
91 : typedef struct {
92 : ulong id; /* The ID of this tile. Indexed from [0, tile_cnt). When placed in a topology, the ID must be the index of the tile in the tiles list. */
93 : char name[ 7UL ]; /* The name of this tile. There can be multiple of each tile name in a topology. */
94 : ulong kind_id; /* The ID of this tile within its name. If there are n tile of a particular name, they have IDs [0, N). The pair (name, kind_id) uniquely identifies a tile, as does "id" on its own. */
95 : int is_agave; /* If the tile needs to run in the Agave (Anza) address space or not. */
96 :
97 : ulong cpu_idx; /* The CPU index to pin the tile on. A value of ULONG_MAX or more indicates the tile should be floating and not pinned to a core. */
98 :
99 : ulong in_cnt; /* The number of links that this tile reads from. */
100 : ulong in_link_id[ FD_TOPO_MAX_TILE_IN_LINKS ]; /* The link_id of each link that this tile reads from, indexed in [0, in_cnt). */
101 : int in_link_reliable[ FD_TOPO_MAX_TILE_IN_LINKS ]; /* If each link that this tile reads from is a reliable or unreliable consumer, indexed in [0, in_cnt). */
102 : int in_link_poll[ FD_TOPO_MAX_TILE_IN_LINKS ]; /* If each link that this tile reads from should be polled by the tile infrastructure, indexed in [0, in_cnt).
103 : If the link is not polled, the tile will not receive frags for it and the tile writer is responsible for
104 : reading from the link. The link must be marked as unreliable as it is not flow controlled. */
105 :
106 : ulong out_cnt; /* The number of links that this tile writes to. */
107 : ulong out_link_id[ FD_TOPO_MAX_TILE_OUT_LINKS ]; /* The link_id of each link that this tile writes to, indexed in [0, link_cnt). */
108 :
109 : ulong tile_obj_id;
110 : ulong metrics_obj_id;
111 : ulong in_link_fseq_obj_id[ FD_TOPO_MAX_TILE_IN_LINKS ];
112 :
113 : ulong uses_obj_cnt;
114 : ulong uses_obj_id[ FD_TOPO_MAX_TILE_OBJS ];
115 : int uses_obj_mode[ FD_TOPO_MAX_TILE_OBJS ];
116 :
117 : /* Computed fields. These are not supplied as configuration but calculated as needed. */
118 : struct {
119 : ulong * metrics; /* The shared memory for metrics that this tile should write. Consumer by monitoring and metrics writing tiles. */
120 :
121 : /* The fseq of each link that this tile reads from. Multiple fseqs
122 : may point to the link, if there are multiple consumers. An fseq
123 : can be uniquely identified via (link_id, tile_id), or (link_kind,
124 : link_kind_id, tile_kind, tile_kind_id) */
125 : ulong * in_link_fseq[ FD_TOPO_MAX_TILE_IN_LINKS ];
126 : };
127 :
128 : /* Configuration fields. These are required to be known by the topology so it can determine the
129 : total size of Firedancer in memory. */
130 : union {
131 : struct {
132 : char interface[ 16 ];
133 : ulong xdp_rx_queue_size;
134 : ulong xdp_tx_queue_size;
135 : ulong xdp_aio_depth;
136 : char xdp_mode[4];
137 : int zero_copy;
138 : uint src_ip_addr;
139 : uchar src_mac_addr[6];
140 :
141 : ushort shred_listen_port;
142 : ushort quic_transaction_listen_port;
143 : ushort legacy_transaction_listen_port;
144 : ushort gossip_listen_port;
145 : ushort repair_intake_listen_port;
146 : ushort repair_serve_listen_port;
147 :
148 : /* multihoming support */
149 : ulong multihome_ip_addrs_cnt;
150 : uint multihome_ip_addrs[FD_NET_MAX_SRC_ADDR];
151 : } net;
152 :
153 : struct {
154 : ulong depth;
155 : uint reasm_cnt;
156 : ulong max_concurrent_connections;
157 : ulong max_concurrent_handshakes;
158 : ulong max_inflight_quic_packets;
159 : ulong max_concurrent_streams_per_connection;
160 : uint ip_addr;
161 : uchar src_mac_addr[ 6 ];
162 : ushort quic_transaction_listen_port;
163 : ulong idle_timeout_millis;
164 : uint ack_delay_millis;
165 : char identity_key_path[ PATH_MAX ];
166 : int retry;
167 : } quic;
168 :
169 : struct {
170 : ulong tcache_depth;
171 : } dedup;
172 :
173 : struct {
174 : ulong max_pending_transactions;
175 : ulong bank_tile_count;
176 : int larger_max_cost_per_block;
177 : int larger_shred_limits_per_block;
178 : int use_consumed_cus;
179 : char identity_key_path[ PATH_MAX ];
180 : } pack;
181 :
182 : struct {
183 : int plugins_enabled;
184 : ulong bank_cnt;
185 : char identity_key_path[ PATH_MAX ];
186 : } poh;
187 :
188 : struct {
189 : ulong depth;
190 : uint ip_addr;
191 : uchar src_mac_addr[ 6 ];
192 : ulong fec_resolver_depth;
193 : char identity_key_path[ PATH_MAX ];
194 : ushort shred_listen_port;
195 : int larger_shred_limits_per_block;
196 : ulong expected_shred_version;
197 : } shred;
198 :
199 : struct {
200 : ulong disable_blockstore_from_slot;
201 : } store;
202 :
203 : struct {
204 : char identity_key_path[ PATH_MAX ];
205 : } sign;
206 :
207 : struct {
208 : uint listen_addr;
209 : ushort listen_port;
210 :
211 : int is_voting;
212 :
213 : char cluster[ 32 ];
214 : char identity_key_path[ PATH_MAX ];
215 : } gui;
216 :
217 : struct {
218 : uint prometheus_listen_addr;
219 : ushort prometheus_listen_port;
220 : } metric;
221 :
222 : struct {
223 :
224 : /* specified by [tiles.replay] */
225 :
226 : char blockstore_checkpt[ PATH_MAX ];
227 : int blockstore_publish;
228 : int tx_metadata_storage;
229 : char capture[ PATH_MAX ];
230 : char funk_checkpt[ PATH_MAX ];
231 : ulong funk_rec_max;
232 : ulong funk_sz_gb;
233 : ulong funk_txn_max;
234 : char funk_file[ PATH_MAX ];
235 : char genesis[ PATH_MAX ];
236 : char incremental[ PATH_MAX ];
237 : char slots_replayed[ PATH_MAX ];
238 : char snapshot[ PATH_MAX ];
239 : char status_cache[ PATH_MAX ];
240 : ulong tpool_thread_count;
241 : char cluster_version[ 32 ];
242 :
243 : /* not specified by [tiles.replay] */
244 :
245 : char identity_key_path[ PATH_MAX ];
246 : uint ip_addr;
247 : uchar src_mac_addr[ 6 ];
248 : int vote;
249 : char vote_account_path[ PATH_MAX ];
250 : ulong bank_tile_count;
251 : } replay;
252 :
253 : struct {
254 : ushort send_to_port;
255 : uint send_to_ip_addr;
256 : ulong conn_cnt;
257 : int no_quic;
258 : } benchs;
259 :
260 : struct {
261 : ushort rpc_port;
262 : uint rpc_ip_addr;
263 : } bencho;
264 :
265 : struct {
266 : ulong accounts_cnt;
267 : int mode;
268 : float contending_fraction;
269 : float cu_price_spread;
270 : } benchg;
271 :
272 : /* Firedancer-only tile configs */
273 :
274 : struct {
275 : ushort gossip_listen_port;
276 : ulong entrypoints_cnt;
277 : uint entrypoints[16];
278 : ulong peer_ports_cnt;
279 : ushort peer_ports[16];
280 :
281 : uint ip_addr;
282 : uchar src_mac_addr[ 6 ];
283 : char identity_key_path[ PATH_MAX ];
284 : ushort tvu_port;
285 : ushort tvu_fwd_port;
286 : ushort tpu_port;
287 : ushort tpu_vote_port;
288 : ushort repair_serve_port;
289 : ulong expected_shred_version;
290 : } gossip;
291 :
292 : struct {
293 : ushort repair_intake_listen_port;
294 : ushort repair_serve_listen_port;
295 :
296 : /* non-config */
297 :
298 : uint ip_addr;
299 : uchar src_mac_addr[ 6 ];
300 : char identity_key_path[ PATH_MAX ];
301 : } repair;
302 :
303 : struct {
304 : char blockstore_restore[ PATH_MAX ];
305 : char slots_pending[PATH_MAX];
306 :
307 : ulong expected_shred_version;
308 :
309 : /* non-config */
310 :
311 : char identity_key_path[ PATH_MAX ];
312 : char shred_cap_archive[ PATH_MAX ];
313 : char shred_cap_replay[ PATH_MAX ];
314 : } store_int;
315 :
316 : struct {
317 : ushort tpu_listen_port;
318 :
319 : /* non-config */
320 :
321 : uint ip_addr;
322 : uchar src_mac_addr[ 6 ];
323 : char identity_key_path[ PATH_MAX ];
324 : } sender;
325 :
326 : struct {
327 : char identity_key_path[ PATH_MAX ];
328 : } eqvoc;
329 :
330 : struct {
331 : ushort rpc_port;
332 : ushort tpu_port;
333 : uint tpu_ip_addr;
334 : char identity_key_path[ PATH_MAX ];
335 : } rpcserv;
336 : };
337 : } fd_topo_tile_t;
338 :
339 : typedef struct {
340 : ulong id;
341 : char name[ 13UL ];
342 : ulong wksp_id;
343 :
344 : ulong offset;
345 : ulong footprint;
346 : } fd_topo_obj_t;
347 :
348 : /* An fd_topo_t represents the overall structure of a Firedancer
349 : configuration, describing all the workspaces, tiles, and links
350 : between them. */
351 : typedef struct fd_topo_t {
352 : char app_name[ 256UL ];
353 : uchar props[ 16384UL ];
354 :
355 : ulong wksp_cnt;
356 : ulong link_cnt;
357 : ulong tile_cnt;
358 : ulong obj_cnt;
359 :
360 : fd_topo_wksp_t workspaces[ FD_TOPO_MAX_WKSPS ];
361 : fd_topo_link_t links[ FD_TOPO_MAX_LINKS ];
362 : fd_topo_tile_t tiles[ FD_TOPO_MAX_TILES ];
363 : fd_topo_obj_t objs[ FD_TOPO_MAX_OBJS ];
364 :
365 : ulong agave_affinity_cnt;
366 : ulong agave_affinity_cpu_idx[ FD_TILE_MAX ];
367 : } fd_topo_t;
368 :
369 : typedef struct {
370 : char const * name;
371 :
372 : int keep_host_networking;
373 : ulong rlimit_file_cnt;
374 : int for_tpool;
375 :
376 : ulong (*populate_allowed_seccomp)( fd_topo_t const * topo, fd_topo_tile_t const * tile, ulong out_cnt, struct sock_filter * out );
377 : ulong (*populate_allowed_fds )( fd_topo_t const * topo, fd_topo_tile_t const * tile, ulong out_fds_sz, int * out_fds );
378 : ulong (*scratch_align )( void );
379 : ulong (*scratch_footprint )( fd_topo_tile_t const * tile );
380 : ulong (*loose_footprint )( fd_topo_tile_t const * tile );
381 : void (*privileged_init )( fd_topo_t * topo, fd_topo_tile_t * tile );
382 : void (*unprivileged_init )( fd_topo_t * topo, fd_topo_tile_t * tile );
383 : void (*run )( fd_topo_t * topo, fd_topo_tile_t * tile );
384 : } fd_topo_run_tile_t;
385 :
386 : FD_PROTOTYPES_BEGIN
387 :
388 : FD_FN_CONST static inline ulong
389 0 : fd_topo_workspace_align( void ) {
390 : /* This needs to be the max( align ) of all the child members that
391 : could be aligned into this workspace, otherwise our footprint
392 : calculation will not be correct. For now just set to 4096 but this
393 : should probably be calculated dynamically, or we should reduce
394 : those child aligns if we can. */
395 0 : return 4096UL;
396 0 : }
397 :
398 : FD_FN_PURE static inline void *
399 : fd_topo_obj_laddr( fd_topo_t const * topo,
400 0 : ulong obj_id ) {
401 0 : fd_topo_obj_t const * obj = &topo->objs[ obj_id ];
402 0 : return (void *)((ulong)topo->workspaces[ obj->wksp_id ].wksp + obj->offset);
403 0 : }
404 :
405 : FD_FN_PURE static inline ulong
406 : fd_topo_tile_name_cnt( fd_topo_t const * topo,
407 0 : char const * name ) {
408 0 : ulong cnt = 0;
409 0 : for( ulong i=0; i<topo->tile_cnt; i++ ) {
410 0 : if( FD_UNLIKELY( !strcmp( topo->tiles[ i ].name, name ) ) ) cnt++;
411 0 : }
412 0 : return cnt;
413 0 : }
414 :
415 : /* Finds the workspace of a given name in the topology. Returns
416 : ULONG_MAX if there is no such workspace. There can be at most one
417 : workspace of a given name. */
418 :
419 : FD_FN_PURE static inline ulong
420 : fd_topo_find_wksp( fd_topo_t const * topo,
421 342 : char const * name ) {
422 4209 : for( ulong i=0; i<topo->wksp_cnt; i++ ) {
423 4209 : if( FD_UNLIKELY( !strcmp( topo->workspaces[ i ].name, name ) ) ) return i;
424 4209 : }
425 0 : return ULONG_MAX;
426 342 : }
427 :
428 : /* Find the tile of a given name and kind_id in the topology, there will
429 : be at most one such tile, since kind_id is unique among the name.
430 : Returns ULONG_MAX if there is no such tile. */
431 :
432 : FD_FN_PURE static inline ulong
433 : fd_topo_find_tile( fd_topo_t const * topo,
434 : char const * name,
435 180 : ulong kind_id ) {
436 1602 : for( ulong i=0; i<topo->tile_cnt; i++ ) {
437 1602 : if( FD_UNLIKELY( !strcmp( topo->tiles[ i ].name, name ) ) && topo->tiles[ i ].kind_id == kind_id ) return i;
438 1602 : }
439 0 : return ULONG_MAX;
440 180 : }
441 :
442 : /* Find the link of a given name and kind_id in the topology, there will
443 : be at most one such link, since kind_id is unique among the name.
444 : Returns ULONG_MAX if there is no such link. */
445 :
446 : FD_FN_PURE static inline ulong
447 : fd_topo_find_link( fd_topo_t const * topo,
448 : char const * name,
449 162 : ulong kind_id ) {
450 1917 : for( ulong i=0; i<topo->link_cnt; i++ ) {
451 1917 : if( FD_UNLIKELY( !strcmp( topo->links[ i ].name, name ) ) && topo->links[ i ].kind_id == kind_id ) return i;
452 1917 : }
453 0 : return ULONG_MAX;
454 162 : }
455 :
456 : FD_FN_PURE static inline ulong
457 : fd_topo_find_tile_in_link( fd_topo_t const * topo,
458 : fd_topo_tile_t const * tile,
459 : char const * name,
460 0 : ulong kind_id ) {
461 0 : for( ulong i=0; i<tile->in_cnt; i++ ) {
462 0 : if( FD_UNLIKELY( !strcmp( topo->links[ tile->in_link_id[ i ] ].name, name ) )
463 0 : && topo->links[ tile->in_link_id[ i ] ].kind_id == kind_id ) return i;
464 0 : }
465 0 : return ULONG_MAX;
466 0 : }
467 :
468 : FD_FN_PURE static inline ulong
469 : fd_topo_find_tile_out_link( fd_topo_t const * topo,
470 : fd_topo_tile_t const * tile,
471 : char const * name,
472 0 : ulong kind_id ) {
473 0 : for( ulong i=0; i<tile->out_cnt; i++ ) {
474 0 : if( FD_UNLIKELY( !strcmp( topo->links[ tile->out_link_id[ i ] ].name, name ) )
475 0 : && topo->links[ tile->out_link_id[ i ] ].kind_id == kind_id ) return i;
476 0 : }
477 0 : return ULONG_MAX;
478 0 : }
479 :
480 : /* Find the id of the tile which is a producer for the given link. If
481 : no tile is a producer for the link, returns ULONG_MAX. This should
482 : not be possible for a well formed and validated topology. */
483 : FD_FN_PURE static inline ulong
484 : fd_topo_find_link_producer( fd_topo_t const * topo,
485 0 : fd_topo_link_t const * link ) {
486 0 : for( ulong i=0; i<topo->tile_cnt; i++ ) {
487 0 : fd_topo_tile_t const * tile = &topo->tiles[ i ];
488 :
489 0 : for( ulong j=0; j<tile->out_cnt; j++ ) {
490 0 : if( FD_UNLIKELY( tile->out_link_id[ j ] == link->id ) ) return i;
491 0 : }
492 0 : }
493 0 : return ULONG_MAX;
494 0 : }
495 :
496 : /* Given a link, count the number of consumers of that link among all
497 : the tiles in the topology. */
498 : FD_FN_PURE static inline ulong
499 : fd_topo_link_consumer_cnt( fd_topo_t const * topo,
500 69 : fd_topo_link_t const * link ) {
501 69 : ulong cnt = 0;
502 1242 : for( ulong i=0; i<topo->tile_cnt; i++ ) {
503 1173 : fd_topo_tile_t const * tile = &topo->tiles[ i ];
504 3312 : for( ulong j=0; j<tile->in_cnt; j++ ) {
505 2139 : if( FD_UNLIKELY( tile->in_link_id[ j ] == link->id ) ) cnt++;
506 2139 : }
507 1173 : }
508 :
509 69 : return cnt;
510 69 : }
511 :
512 : /* Given a link, count the number of reliable consumers of that link
513 : among all the tiles in the topology. */
514 : FD_FN_PURE static inline ulong
515 : fd_topo_link_reliable_consumer_cnt( fd_topo_t const * topo,
516 0 : fd_topo_link_t const * link ) {
517 0 : ulong cnt = 0;
518 0 : for( ulong i=0; i<topo->tile_cnt; i++ ) {
519 0 : fd_topo_tile_t const * tile = &topo->tiles[ i ];
520 0 : for( ulong j=0; j<tile->in_cnt; j++ ) {
521 0 : if( FD_UNLIKELY( tile->in_link_id[ j ] == link->id && tile->in_link_reliable[ j ] ) ) cnt++;
522 0 : }
523 0 : }
524 0 :
525 0 : return cnt;
526 0 : }
527 :
528 : /* Join (map into the process) all shared memory (huge/gigantic pages)
529 : needed by the tile, in the given topology. All memory associated
530 : with the tile (aka. used by links that the tile either produces to or
531 : consumes from, or used by the tile itself for its cnc) will be
532 : attached (mapped into the process).
533 :
534 : This is needed to play nicely with the sandbox. Once a process is
535 : sandboxed we can no longer map any memory. */
536 : void
537 : fd_topo_join_tile_workspaces( fd_topo_t * topo,
538 : fd_topo_tile_t * tile );
539 :
540 : /* Join (map into the process) the shared memory (huge/gigantic pages)
541 : for the given workspace. Mode is one of
542 : FD_SHMEM_JOIN_MODE_READ_WRITE or FD_SHMEM_JOIN_MODE_READ_ONLY and
543 : determines the prot argument that will be passed to mmap when mapping
544 : the pages in (PROT_WRITE or PROT_READ respectively). */
545 : void
546 : fd_topo_join_workspace( fd_topo_t * topo,
547 : fd_topo_wksp_t * wksp,
548 : int mode );
549 :
550 : /* Join (map into the process) all shared memory (huge/gigantic pages)
551 : needed by all tiles in the topology. Mode is one of
552 : FD_SHMEM_JOIN_MODE_READ_WRITE or FD_SHMEM_JOIN_MODE_READ_ONLY and
553 : determines the prot argument that will be passed to mmap when
554 : mapping the pages in (PROT_WRITE or PROT_READ respectively). */
555 : void
556 : fd_topo_join_workspaces( fd_topo_t * topo,
557 : int mode );
558 :
559 : /* Leave (unmap from the process) the shared memory needed for the
560 : given workspace in the topology, if it was previously mapped.
561 :
562 : topo and wksp are assumed non-NULL. It is OK if the workspace
563 : has not been previously joined, in which case this is a no-op. */
564 :
565 : void
566 : fd_topo_leave_workspace( fd_topo_t * topo,
567 : fd_topo_wksp_t * wksp );
568 :
569 : /* Leave (unmap from the process) all shared memory needed by all
570 : tiles in the topology, if each of them was mapped.
571 :
572 : topo is assumed non-NULL. Only workspaces which were previously
573 : joined are unmapped. */
574 :
575 : void
576 : fd_topo_leave_workspaces( fd_topo_t * topo );
577 :
578 : /* Create the given workspace needed by the topology on the system.
579 : This does not "join" the workspaces (map their memory into the
580 : process), but only creates the .wksp file and formats it correctly
581 : as a workspace.
582 :
583 : Returns 0 on success and -1 on failure, with errno set to the error.
584 : The only reason for failure currently that will be returned is
585 : ENOMEM, as other unexpected errors will cause the program to exit.
586 :
587 : If update_existing is 1, the workspace will not be created from
588 : scratch but it will be assumed that it already exists from a prior
589 : run and needs to be maybe resized and then have the header
590 : structures reinitialized. This can save a very expensive operation
591 : of zeroing all of the workspace pages. This is dangerous in
592 : production because it can leave stray memory from prior runs around,
593 : and should only be used in development environments. */
594 :
595 : int
596 : fd_topo_create_workspace( fd_topo_t * topo,
597 : fd_topo_wksp_t * wksp,
598 : int update_existing );
599 :
600 : /* Join the standard IPC objects needed by the topology of this particular
601 : tile */
602 :
603 : void
604 : fd_topo_fill_tile( fd_topo_t * topo,
605 : fd_topo_tile_t * tile );
606 :
607 : /* Same as fd_topo_fill_tile but fills in all the objects for a
608 : particular workspace with the given mode. */
609 : void
610 : fd_topo_workspace_fill( fd_topo_t * topo,
611 : fd_topo_wksp_t * wksp );
612 :
613 : /* Apply a function to every object that is resident in the given
614 : workspace in the topology. */
615 :
616 : void
617 : fd_topo_wksp_apply( fd_topo_t * topo,
618 : fd_topo_wksp_t * wksp,
619 : void (* fn )( fd_topo_t const * topo, fd_topo_obj_t const * obj ) );
620 :
621 : /* Same as fd_topo_fill_tile but fills in all tiles in the topology. */
622 :
623 : void
624 : fd_topo_fill( fd_topo_t * topo );
625 :
626 : /* fd_topo_tile_stack_join joins a huge page optimized stack for the
627 : provided tile. The stack is assumed to already exist at a known
628 : path in the hugetlbfs mount. */
629 :
630 : void *
631 : fd_topo_tile_stack_join( char const * app_name,
632 : char const * tile_name,
633 : ulong tile_kind_id );
634 :
635 : /* Install the XDP program needed by the net tiles into the local device
636 : and return the xsk_map_fd. */
637 :
638 : fd_xdp_fds_t
639 : fd_topo_install_xdp( fd_topo_t * topo );
640 :
641 : /* fd_topo_run_single_process runs all the tiles in a single process
642 : (the calling process). This spawns a thread for each tile, switches
643 : that thread to the given UID and GID and then runs the tile in it.
644 : Each thread will never exit, as tiles are expected to run forever.
645 : An error is logged and the application will exit if a tile exits.
646 : The function itself does return after spawning all the threads.
647 :
648 : The threads will not be sandboxed in any way, except switching to the
649 : provided UID and GID, so they will share the same address space, and
650 : not have any seccomp restrictions or use any Linux namespaces. The
651 : calling thread will also switch to the provided UID and GID before
652 : it returns.
653 :
654 : In production, when running with an Agave child process this is
655 : used for spawning certain tiles inside the Agave address space.
656 : It's also useful for tooling and debugging, but is not how the main
657 : production Firedancer process runs. For production, each tile is run
658 : in its own address space with a separate process and full security
659 : sandbox.
660 :
661 : The agave argument determines which tiles are started. If the
662 : argument is 0 or 1, only non-agave (or only agave) tiles are started.
663 : If the argument is any other value, all tiles in the topology are
664 : started regardless of if they are Agave tiles or not. */
665 :
666 : void
667 : fd_topo_run_single_process( fd_topo_t * topo,
668 : int agave,
669 : uint uid,
670 : uint gid,
671 : fd_topo_run_tile_t (* tile_run )( fd_topo_tile_t * tile ),
672 : int * done_futex );
673 :
674 : /* fd_topo_run_tile runs the given tile directly within the current
675 : process (and thread). The function will never return, as tiles are
676 : expected to run forever. An error is logged and the application will
677 : exit if the tile exits.
678 :
679 : The sandbox argument determines if the current process will be
680 : sandboxed fully before starting the tile. The thread will switch to
681 : the UID and GID provided before starting the tile, even if the thread
682 : is not being sandboxed. Although POSIX specifies that all threads in
683 : a process must share a UID and GID, this is not the case on Linux.
684 : The thread will switch to the provided UID and GID without switching
685 : the other threads in the process.
686 :
687 : If keep_controlling_terminal is set to 0, and the sandbox is enabled
688 : the controlling terminal will be detached as an additional sandbox
689 : measure, but you will not be able to send Ctrl+C or other signals
690 : from the terminal. See fd_sandbox.h for more information.
691 :
692 : The allow_fd argument is only used if sandbox is true, and is a file
693 : descriptor which will be allowed to exist in the process. Normally
694 : the sandbox code rejects and aborts if there is an unexpected file
695 : descriptor present on boot. This is helpful to allow a parent
696 : process to be notified on termination of the tile by waiting for a
697 : pipe file descriptor to get closed.
698 :
699 : wait and debugger are both used in debugging. If wait is non-NULL,
700 : the runner will wait until the value pointed to by wait is non-zero
701 : before launching the tile. Likewise, if debugger is non-NULL, the
702 : runner will wait until a debugger is attached before setting the
703 : value pointed to by debugger to non-zero. These are intended to be
704 : used as a pair, where many tiles share a waiting reference, and then
705 : one of the tiles (a tile you want to attach the debugger to) has the
706 : same reference provided as the debugger, so all tiles will stop and
707 : wait for the debugger to attach to it before proceeding. */
708 :
709 : void
710 : fd_topo_run_tile( fd_topo_t * topo,
711 : fd_topo_tile_t * tile,
712 : int sandbox,
713 : int keep_controlling_terminal,
714 : uint uid,
715 : uint gid,
716 : int allow_fd,
717 : volatile int * wait,
718 : volatile int * debugger,
719 : fd_topo_run_tile_t * tile_run );
720 :
721 : /* This is for determining the value of RLIMIT_MLOCK that we need to
722 : successfully run all tiles in separate processes. The value returned
723 : is the maximum amount of memory that will be locked with mlock() by
724 : any individual process in the tree. Specifically, if we have three
725 : tile processes, and they each need to lock 5, 9, and 2 MiB of memory
726 : respectively, RLIMIT_MLOCK needs to be 9 MiB to allow all three
727 : process mlock() calls to succeed.
728 :
729 : Tiles lock memory in three ways. Any workspace they are using, they
730 : lock the entire workspace. Then each tile uses huge pages for the
731 : stack which are also locked, and finally some tiles use private
732 : locked mmaps outside the workspace for storing key material. The
733 : results here include all of this memory together.
734 :
735 : The result is not necessarily the amount of memory used by the tile
736 : process, although it will be quite close. Tiles could potentially
737 : allocate memory (eg, with brk) without needing to lock it, which
738 : would not need to included, and some kernel memory that tiles cause
739 : to be allocated (for example XSK buffers) is also not included. The
740 : actual amount of memory used will not be less than this value. */
741 : FD_FN_PURE ulong
742 : fd_topo_mlock_max_tile( fd_topo_t * topo );
743 :
744 : /* Same as fd_topo_mlock_max_tile, but for loading the entire topology
745 : topology into one process, rather than a separate process per tile.
746 : This is used, for example, by the configuration code when it creates
747 : all the workspaces, or the monitor that maps the entire system into
748 : one address space. */
749 : FD_FN_PURE ulong
750 : fd_topo_mlock( fd_topo_t * topo );
751 :
752 : /* This returns the number of gigantic pages needed by the topology on
753 : the provided numa node. It includes pages needed by the workspaces,
754 : as well as additional allocations like huge pages for process stacks
755 : and private key storage. */
756 :
757 : FD_FN_PURE ulong
758 : fd_topo_gigantic_page_cnt( fd_topo_t * topo,
759 : ulong numa_idx );
760 :
761 : /* This returns the number of huge pages in the application needed by
762 : the topology on the provided numa node. It includes pages needed by
763 : things placed in the hugetlbfs (workspaces, process stacks). If
764 : include_anonymous is true, it also includes anonymous hugepages which
765 : are needed but are not placed in the hugetlbfs. */
766 :
767 : FD_FN_PURE ulong
768 : fd_topo_huge_page_cnt( fd_topo_t * topo,
769 : ulong numa_idx,
770 : int include_anonymous );
771 :
772 : /* Check all invariants of the given topology to make sure it is valid.
773 : An invalid topology will cause the program to abort with an error
774 : message. */
775 : void
776 : fd_topo_validate( fd_topo_t const * topo );
777 :
778 : /* Prints a message describing the topology to an output stream. If
779 : stdout is true, will be written to stdout, otherwise will be written
780 : as a NOTICE log message to the log file. */
781 : void
782 : fd_topo_print_log( int stdout,
783 : fd_topo_t * topo );
784 :
785 : FD_PROTOTYPES_END
786 :
787 : #endif /* HEADER_fd_src_disco_topo_fd_topo_h */
|
