Line data Source code
1 : #include "fd_topob.h"
2 :
3 : #include "../../util/pod/fd_pod_format.h"
4 : #include "fd_cpu_topo.h"
5 :
6 : fd_topo_t *
7 : fd_topob_new( void * mem,
8 3 : char const * app_name ) {
9 3 : fd_topo_t * topo = (fd_topo_t *)mem;
10 :
11 3 : if( FD_UNLIKELY( !topo ) ) {
12 0 : FD_LOG_WARNING( ( "NULL topo" ) );
13 0 : return NULL;
14 0 : }
15 :
16 3 : if( FD_UNLIKELY( !fd_ulong_is_aligned( (ulong)topo, alignof(fd_topo_t) ) ) ) {
17 0 : FD_LOG_WARNING( ( "misaligned topo" ) );
18 0 : return NULL;
19 0 : }
20 :
21 3 : fd_memset( topo, 0, sizeof(fd_topo_t) );
22 :
23 3 : FD_TEST( fd_pod_new( topo->props, sizeof(topo->props) ) );
24 :
25 3 : if( FD_UNLIKELY( strlen( app_name )>=sizeof(topo->app_name) ) ) FD_LOG_ERR(( "app_name too long: %s", app_name ));
26 3 : strncpy( topo->app_name, app_name, sizeof(topo->app_name) );
27 :
28 3 : topo->max_page_size = FD_SHMEM_GIGANTIC_PAGE_SZ;
29 3 : topo->gigantic_page_threshold = 4 * FD_SHMEM_HUGE_PAGE_SZ;
30 :
31 3 : return topo;
32 3 : }
33 :
34 : fd_topo_wksp_t *
35 : fd_topob_wksp( fd_topo_t * topo,
36 6 : char const * name ) {
37 6 : if( FD_UNLIKELY( !topo || !name || !strlen( name ) ) ) FD_LOG_ERR(( "NULL args" ));
38 6 : if( FD_UNLIKELY( strlen( name )>=sizeof(topo->workspaces[ topo->wksp_cnt ].name ) ) ) FD_LOG_ERR(( "wksp name too long: %s", name ));
39 6 : if( FD_UNLIKELY( topo->wksp_cnt>=FD_TOPO_MAX_WKSPS ) ) FD_LOG_ERR(( "too many workspaces" ));
40 :
41 6 : fd_topo_wksp_t * wksp = &topo->workspaces[ topo->wksp_cnt ];
42 6 : strncpy( wksp->name, name, sizeof(wksp->name) );
43 6 : wksp->id = topo->wksp_cnt;
44 6 : topo->wksp_cnt++;
45 6 : return wksp;
46 6 : }
47 :
48 : fd_topo_obj_t *
49 : fd_topob_obj( fd_topo_t * topo,
50 : char const * obj_name,
51 66 : char const * wksp_name ) {
52 66 : if( FD_UNLIKELY( !topo || !obj_name || !wksp_name ) ) FD_LOG_ERR(( "NULL args" ));
53 66 : if( FD_UNLIKELY( strlen( obj_name )>=sizeof(topo->objs[ topo->obj_cnt ].name ) ) ) FD_LOG_ERR(( "obj name too long: %s", obj_name ));
54 66 : if( FD_UNLIKELY( topo->obj_cnt>=FD_TOPO_MAX_OBJS ) ) FD_LOG_ERR(( "too many objects" ));
55 :
56 66 : ulong wksp_id = fd_topo_find_wksp( topo, wksp_name );
57 66 : if( FD_UNLIKELY( wksp_id==ULONG_MAX ) ) FD_LOG_ERR(( "workspace not found: %s", wksp_name ));
58 :
59 66 : fd_topo_obj_t * obj = &topo->objs[ topo->obj_cnt ];
60 66 : memset( obj, 0, sizeof(fd_topo_obj_t) );
61 66 : strncpy( obj->name, obj_name, sizeof(obj->name) );
62 66 : obj->id = topo->obj_cnt;
63 66 : obj->wksp_id = wksp_id;
64 66 : obj->label_idx = ULONG_MAX;
65 66 : topo->obj_cnt++;
66 :
67 66 : return obj;
68 66 : }
69 :
70 : fd_topo_obj_t *
71 : fd_topob_obj_named( fd_topo_t * topo,
72 : char const * obj_type,
73 : char const * wksp_name,
74 0 : char const * label ) {
75 0 : if( FD_UNLIKELY( !label ) ) FD_LOG_ERR(( "NULL args" ));
76 0 : if( FD_UNLIKELY( strlen( label )>=sizeof(topo->objs[ topo->obj_cnt ].label ) ) ) FD_LOG_ERR(( "obj label too long: %s", label ));
77 0 : fd_topo_obj_t * obj = fd_topob_obj( topo, obj_type, wksp_name );
78 0 : if( FD_UNLIKELY( !obj ) ) return NULL;
79 :
80 0 : fd_cstr_ncpy( obj->label, label, sizeof(obj->label) );
81 0 : obj->label_idx = fd_topo_obj_cnt( topo, obj_type, label );
82 :
83 0 : return obj;
84 0 : }
85 :
86 : fd_topo_link_t *
87 : fd_topob_link( fd_topo_t * topo,
88 : char const * link_name,
89 : char const * wksp_name,
90 : ulong depth,
91 : ulong mtu,
92 18 : ulong burst ) {
93 18 : if( FD_UNLIKELY( !topo || !link_name || !wksp_name ) ) FD_LOG_ERR(( "NULL args" ));
94 18 : if( FD_UNLIKELY( strlen( link_name )>=sizeof(topo->links[ topo->link_cnt ].name ) ) ) FD_LOG_ERR(( "link name too long: %s", link_name ));
95 18 : if( FD_UNLIKELY( topo->link_cnt>=FD_TOPO_MAX_LINKS ) ) FD_LOG_ERR(( "too many links" ));
96 :
97 18 : ulong kind_id = 0UL;
98 39 : for( ulong i=0UL; i<topo->link_cnt; i++ ) {
99 21 : if( !strcmp( topo->links[ i ].name, link_name ) ) kind_id++;
100 21 : }
101 :
102 18 : fd_topo_link_t * link = &topo->links[ topo->link_cnt ];
103 18 : strncpy( link->name, link_name, sizeof(link->name) );
104 18 : link->id = topo->link_cnt;
105 18 : link->kind_id = kind_id;
106 18 : link->depth = depth;
107 18 : link->mtu = mtu;
108 18 : link->burst = burst;
109 :
110 18 : fd_topo_obj_t * obj = fd_topob_obj( topo, "mcache", wksp_name );
111 18 : link->mcache_obj_id = obj->id;
112 18 : FD_TEST( fd_pod_insertf_ulong( topo->props, depth, "obj.%lu.depth", obj->id ) );
113 :
114 18 : if( mtu ) {
115 6 : obj = fd_topob_obj( topo, "dcache", wksp_name );
116 6 : link->dcache_obj_id = obj->id;
117 6 : FD_TEST( fd_pod_insertf_ulong( topo->props, depth, "obj.%lu.depth", obj->id ) );
118 6 : FD_TEST( fd_pod_insertf_ulong( topo->props, burst, "obj.%lu.burst", obj->id ) );
119 6 : FD_TEST( fd_pod_insertf_ulong( topo->props, mtu, "obj.%lu.mtu", obj->id ) );
120 6 : }
121 18 : topo->link_cnt++;
122 :
123 18 : return link;
124 18 : }
125 :
126 : void
127 : fd_topob_tile_uses( fd_topo_t * topo,
128 : fd_topo_tile_t * tile,
129 : fd_topo_obj_t const * obj,
130 51 : int mode ) {
131 51 : (void)topo;
132 :
133 51 : if( FD_UNLIKELY( tile->uses_obj_cnt>=FD_TOPO_MAX_TILE_OBJS ) ) FD_LOG_ERR(( "tile `%s` uses too many objects", tile->name ));
134 :
135 51 : tile->uses_obj_id[ tile->uses_obj_cnt ] = obj->id;
136 51 : tile->uses_obj_mode[ tile->uses_obj_cnt ] = mode;
137 51 : tile->uses_obj_cnt++;
138 51 : }
139 :
140 : fd_topo_tile_t *
141 : fd_topob_tile( fd_topo_t * topo,
142 : char const * tile_name,
143 : char const * tile_wksp,
144 : char const * metrics_wksp,
145 : ulong cpu_idx,
146 : int is_agave,
147 6 : int uses_keyswitch ) {
148 6 : if( FD_UNLIKELY( !topo || !tile_name || !tile_wksp || !metrics_wksp ) ) FD_LOG_ERR(( "NULL args" ));
149 6 : if( FD_UNLIKELY( strlen( tile_name )>=sizeof(topo->tiles[ topo->tile_cnt ].name ) ) ) FD_LOG_ERR(( "tile name too long: %s", tile_name ));
150 6 : if( FD_UNLIKELY( topo->tile_cnt>=FD_TOPO_MAX_TILES ) ) FD_LOG_ERR(( "too many tiles %lu", topo->tile_cnt ));
151 :
152 6 : ulong kind_id = 0UL;
153 6 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
154 0 : if( !strcmp( topo->tiles[ i ].name, tile_name ) ) kind_id++;
155 0 : }
156 :
157 6 : fd_topo_tile_t * tile = &topo->tiles[ topo->tile_cnt ];
158 6 : strncpy( tile->name, tile_name, sizeof(tile->name) );
159 6 : tile->metrics_name[ 0 ] = 0;
160 6 : tile->id = topo->tile_cnt;
161 6 : tile->kind_id = kind_id;
162 6 : tile->is_agave = is_agave;
163 6 : tile->cpu_idx = cpu_idx;
164 6 : tile->in_cnt = 0UL;
165 6 : tile->out_cnt = 0UL;
166 6 : tile->uses_obj_cnt = 0UL;
167 :
168 6 : fd_topo_obj_t * tile_obj = fd_topob_obj( topo, "tile", tile_wksp );
169 6 : tile->tile_obj_id = tile_obj->id;
170 6 : fd_topob_tile_uses( topo, tile, tile_obj, FD_SHMEM_JOIN_MODE_READ_WRITE );
171 :
172 6 : fd_topo_obj_t * obj = fd_topob_obj( topo, "metrics", metrics_wksp );
173 6 : tile->metrics_obj_id = obj->id;
174 6 : fd_topob_tile_uses( topo, tile, obj, FD_SHMEM_JOIN_MODE_READ_WRITE );
175 :
176 6 : if( FD_LIKELY( uses_keyswitch ) ) {
177 0 : obj = fd_topob_obj( topo, "keyswitch", tile_wksp );
178 0 : tile->keyswitch_obj_id = obj->id;
179 0 : fd_topob_tile_uses( topo, tile, obj, FD_SHMEM_JOIN_MODE_READ_WRITE );
180 6 : } else {
181 6 : tile->keyswitch_obj_id = ULONG_MAX;
182 6 : }
183 :
184 6 : topo->tile_cnt++;
185 6 : return tile;
186 6 : }
187 :
188 : void
189 : fd_topob_tile_in( fd_topo_t * topo,
190 : char const * tile_name,
191 : ulong tile_kind_id,
192 : char const * fseq_wksp,
193 : char const * link_name,
194 : ulong link_kind_id,
195 : int reliable,
196 15 : int polled ) {
197 15 : if( FD_UNLIKELY( !topo || !tile_name || !fseq_wksp || !link_name ) ) FD_LOG_ERR(( "NULL args" ));
198 :
199 15 : ulong tile_id = fd_topo_find_tile( topo, tile_name, tile_kind_id );
200 15 : if( FD_UNLIKELY( tile_id==ULONG_MAX ) ) FD_LOG_ERR(( "tile not found: %s:%lu", tile_name, tile_kind_id ));
201 15 : fd_topo_tile_t * tile = &topo->tiles[ tile_id ];
202 :
203 15 : ulong link_id = fd_topo_find_link( topo, link_name, link_kind_id );
204 15 : if( FD_UNLIKELY( link_id==ULONG_MAX ) ) FD_LOG_ERR(( "link not found: %s:%lu", link_name, link_kind_id ));
205 15 : fd_topo_link_t * link = &topo->links[ link_id ];
206 :
207 15 : if( FD_UNLIKELY( tile->in_cnt>=FD_TOPO_MAX_TILE_IN_LINKS ) ) FD_LOG_ERR(( "too many in links: %s:%lu", tile_name, tile_kind_id ) );
208 15 : tile->in_link_id[ tile->in_cnt ] = link->id;
209 15 : tile->in_link_reliable[ tile->in_cnt ] = reliable;
210 15 : tile->in_link_poll[ tile->in_cnt ] = polled;
211 15 : fd_topo_obj_t * obj = fd_topob_obj( topo, "fseq", fseq_wksp );
212 15 : fd_topob_tile_uses( topo, tile, obj, FD_SHMEM_JOIN_MODE_READ_WRITE );
213 15 : tile->in_link_fseq_obj_id[ tile->in_cnt ] = obj->id;
214 15 : tile->in_cnt++;
215 :
216 15 : fd_topob_tile_uses( topo, tile, &topo->objs[ link->mcache_obj_id ], FD_SHMEM_JOIN_MODE_READ_ONLY );
217 15 : if( FD_LIKELY( link->mtu ) ) {
218 3 : fd_topob_tile_uses( topo, tile, &topo->objs[ link->dcache_obj_id ], FD_SHMEM_JOIN_MODE_READ_ONLY );
219 3 : }
220 15 : }
221 :
222 : void
223 : fd_topob_tile_out( fd_topo_t * topo,
224 : char const * tile_name,
225 : ulong tile_kind_id,
226 : char const * link_name,
227 3 : ulong link_kind_id ) {
228 3 : ulong tile_id = fd_topo_find_tile( topo, tile_name, tile_kind_id );
229 3 : if( FD_UNLIKELY( tile_id==ULONG_MAX ) ) FD_LOG_ERR(( "tile not found: %s:%lu", tile_name, tile_kind_id ));
230 3 : fd_topo_tile_t * tile = &topo->tiles[ tile_id ];
231 :
232 3 : ulong link_id = fd_topo_find_link( topo, link_name, link_kind_id );
233 3 : if( FD_UNLIKELY( link_id==ULONG_MAX ) ) FD_LOG_ERR(( "link not found: %s:%lu", link_name, link_kind_id ));
234 3 : fd_topo_link_t * link = &topo->links[ link_id ];
235 :
236 3 : if( FD_UNLIKELY( tile->out_cnt>=FD_TOPO_MAX_TILE_OUT_LINKS ) ) FD_LOG_ERR(( "too many out links: %s", tile_name ));
237 3 : tile->out_link_id[ tile->out_cnt ] = link->id;
238 3 : tile->out_cnt++;
239 :
240 3 : fd_topob_tile_uses( topo, tile, &topo->objs[ link->mcache_obj_id ], FD_SHMEM_JOIN_MODE_READ_WRITE );
241 3 : if( FD_LIKELY( link->mtu ) ) {
242 3 : fd_topob_tile_uses( topo, tile, &topo->objs[ link->dcache_obj_id ], FD_SHMEM_JOIN_MODE_READ_WRITE );
243 3 : }
244 3 : }
245 :
246 : static void
247 0 : validate( fd_topo_t const * topo ) {
248 : /* Objects have valid wksp_ids */
249 0 : for( ulong i=0UL; i<topo->obj_cnt; i++ ) {
250 0 : if( FD_UNLIKELY( topo->objs[ i ].wksp_id>=topo->wksp_cnt ) )
251 0 : FD_LOG_ERR(( "invalid workspace id %lu", topo->objs[ i ].wksp_id ));
252 0 : }
253 :
254 : /* Tile ins are valid */
255 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
256 0 : for( ulong j=0UL; j<topo->tiles[ i ].in_cnt; j++ ) {
257 0 : if( FD_UNLIKELY( topo->tiles[ i ].in_link_id[ j ]>=topo->link_cnt ) )
258 0 : FD_LOG_ERR(( "tile %lu (%s) has invalid in link %lu", i, topo->tiles[ i ].name, topo->tiles[ i ].in_link_id[ j ] ));
259 0 : }
260 0 : }
261 :
262 : /* Tile does not have duplicated ins */
263 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
264 0 : for( ulong j=0UL; j<topo->tiles[ i ].in_cnt; j++ ) {
265 0 : for( ulong k=0UL; k<topo->tiles[ i ].in_cnt; k++ ) {
266 0 : if( FD_UNLIKELY( j==k ) ) continue;
267 0 : if( FD_UNLIKELY( topo->tiles[ i ].in_link_id[ j ] == topo->tiles[ i ].in_link_id[ k ] ) )
268 0 : FD_LOG_ERR(( "tile %lu (%s) has duplicated in link %lu (%s)", i, topo->tiles[ i ].name,
269 0 : topo->tiles[ i ].in_link_id[ j ], topo->links[ topo->tiles[ i ].in_link_id[ j ] ].name ));
270 0 : }
271 0 : }
272 0 : }
273 :
274 : /* Tile does not have duplicated outs */
275 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
276 0 : for( ulong j=0UL; j<topo->tiles[ i ].out_cnt; j++ ) {
277 0 : for( ulong k=0UL; k<topo->tiles[ i ].out_cnt; k++ ) {
278 0 : if( FD_UNLIKELY( j==k ) ) continue;
279 0 : if( FD_UNLIKELY( topo->tiles[ i ].out_link_id[ j ] == topo->tiles[ i ].out_link_id[ k ] ) )
280 0 : FD_LOG_ERR(( "tile %lu (%s) has duplicated out link %lu (%s)", i, topo->tiles[ i ].name,
281 0 : topo->tiles[ i ].out_link_id[ j ], topo->links[ topo->tiles[ i ].out_link_id[ j ] ].name ));
282 0 : }
283 0 : }
284 0 : }
285 :
286 : /* Tile outs are different than ins */
287 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
288 0 : for( ulong j=0UL; j<topo->tiles[ i ].out_cnt; j++ ) {
289 0 : for( ulong k=0UL; k<topo->tiles[ i ].in_cnt; k++ ) {
290 0 : char const * link_name = topo->links[ topo->tiles[ i ].out_link_id[ j ] ].name;
291 : /* PoH tile "publishes" this on behalf of Agave, so it's not
292 : a real circular link. */
293 0 : if( FD_UNLIKELY( !strcmp( link_name, "stake_out" ) ||
294 0 : !strcmp( link_name, "crds_shred" ) ) ) continue;
295 :
296 0 : if( FD_UNLIKELY( topo->tiles[ i ].out_link_id[ j ] == topo->tiles[ i ].in_link_id[ k ] ) )
297 0 : FD_LOG_ERR(( "tile %lu has out link %lu same as in", i, topo->tiles[ i ].out_link_id[ j ] ));
298 0 : }
299 0 : }
300 0 : }
301 :
302 : /* Non polling tile ins are also not reliable */
303 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
304 0 : for( ulong j=0UL; j<topo->tiles[ i ].in_cnt; j++ ) {
305 0 : if( FD_UNLIKELY( !topo->tiles[ i ].in_link_poll[ j ] && topo->tiles[ i ].in_link_reliable[ j ] ) )
306 0 : FD_LOG_ERR(( "tile %lu has in link %lu which is not polled but reliable", i, topo->tiles[ i ].in_link_id[ j ] ));
307 0 : }
308 0 : }
309 :
310 : /* Tile outs are valid */
311 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
312 0 : for( ulong j=0UL; j<topo->tiles[ i ].out_cnt; j++ ) {
313 0 : if( FD_UNLIKELY( topo->tiles[ i ].out_link_id[ j ] >= topo->link_cnt ) )
314 0 : FD_LOG_ERR(( "tile %lu has invalid out link %lu", i, topo->tiles[ i ].out_link_id[ j ] ));
315 0 : }
316 0 : }
317 :
318 : /* Workspace names are unique */
319 0 : for( ulong i=0UL; i<topo->wksp_cnt; i++ ) {
320 0 : for( ulong j=0UL; j<topo->wksp_cnt; j++ ) {
321 0 : if( FD_UNLIKELY( i==j ) ) continue;
322 0 : if( FD_UNLIKELY( !strcmp( topo->workspaces[ i ].name, topo->workspaces[ j ].name ) ) )
323 0 : FD_LOG_ERR(( "duplicate workspace name %s", topo->workspaces[ i ].name ));
324 0 : }
325 0 : }
326 :
327 : /* Each workspace is identified correctly */
328 0 : for( ulong i=0UL; i<topo->wksp_cnt; i++ ) {
329 0 : if( FD_UNLIKELY( topo->workspaces[ i ].id != i ) )
330 0 : FD_LOG_ERR(( "workspace %lu has id %lu", i, topo->workspaces[ i ].id ));
331 0 : }
332 :
333 : /* Each link has exactly one producer */
334 0 : for( ulong i=0UL; i<topo->link_cnt; i++ ) {
335 0 : ulong producer_cnt = 0;
336 0 : for( ulong j=0UL; j<topo->tile_cnt; j++ ) {
337 0 : for( ulong k=0UL; k<topo->tiles[ j ].out_cnt; k++ ) {
338 0 : if( topo->tiles[ j ].out_link_id[ k ]==i ) producer_cnt++;
339 0 : }
340 0 : }
341 0 : if( FD_UNLIKELY( producer_cnt>1UL || ( producer_cnt==0UL && !topo->links[ i ].permit_no_producers ) ) )
342 0 : FD_LOG_ERR(( "link %lu (%s:%lu) has %lu producers", i, topo->links[ i ].name, topo->links[ i ].kind_id, producer_cnt ));
343 0 : }
344 :
345 : /* Each link has at least one consumer */
346 0 : for( ulong i=0UL; i<topo->link_cnt; i++ ) {
347 0 : ulong cnt = fd_topo_link_consumer_cnt( topo, &topo->links[ i ] );
348 0 : if( FD_UNLIKELY( cnt < 1UL && !topo->links[ i ].permit_no_consumers ) ) {
349 0 : FD_LOG_ERR(( "link %lu (%s:%lu) has 0 consumers", i, topo->links[ i ].name, topo->links[ i ].kind_id ));
350 0 : }
351 0 : }
352 0 : }
353 :
354 : void
355 : fd_topob_auto_layout( fd_topo_t * topo,
356 0 : int reserve_agave_cores ) {
357 : /* Incredibly simple automatic layout system for now ... just assign
358 : tiles to CPU cores in NUMA sequential order, except for a few tiles
359 : which should be floating. */
360 :
361 0 : char const * FLOATING[] = {
362 0 : "netlnk",
363 0 : "metric",
364 0 : "cswtch",
365 0 : "bencho",
366 0 : "genesi", /* FIREDANCER ONLY */
367 0 : "ipecho", /* FIREDANCER ONLY */
368 0 : "snapwr", /* FIREDANCER ONLY */
369 0 : };
370 :
371 0 : char const * ORDERED[] = {
372 0 : "backt",
373 0 : "benchg",
374 0 : "benchs",
375 0 : "net",
376 0 : "sock",
377 0 : "quic",
378 0 : "bundle",
379 0 : "verify",
380 0 : "dedup",
381 0 : "resolv", /* FRANK only */
382 0 : "pack",
383 0 : "bank", /* FRANK only */
384 0 : "poh", /* FRANK only */
385 0 : "pohi", /* FIREDANCER only */
386 0 : "shred",
387 0 : "store", /* FRANK only */
388 0 : "storei", /* FIREDANCER only */
389 0 : "sign",
390 0 : "plugin",
391 0 : "gui",
392 0 : "rpc", /* FIREDANCER only */
393 0 : "gossvf", /* FIREDANCER only */
394 0 : "gossip", /* FIREDANCER only */
395 0 : "repair", /* FIREDANCER only */
396 0 : "replay", /* FIREDANCER only */
397 0 : "exec", /* FIREDANCER only */
398 0 : "send", /* FIREDANCER only */
399 0 : "tower", /* FIREDANCER only */
400 0 : "rpcsrv", /* FIREDANCER only */
401 0 : "pktgen",
402 0 : "snapct", /* FIREDANCER only */
403 0 : "snapld", /* FIREDANCER only */
404 0 : "snapdc", /* FIREDANCER only */
405 0 : "snapin", /* FIREDANCER only */
406 0 : "snapwm", /* FIREDANCER only */
407 0 : "snapwh", /* FIREDANCER only */
408 0 : "snapla", /* FIREDANCER only */
409 0 : "snapls", /* FIREDANCER only */
410 0 : "arch_f", /* FIREDANCER only */
411 0 : "arch_w", /* FIREDANCER only */
412 0 : "vinyl", /* FIREDANCER only */
413 0 : };
414 :
415 0 : char const * CRITICAL_TILES[] = {
416 0 : "pack",
417 0 : "poh",
418 0 : "gui",
419 0 : "snapld", /* TODO: Snapshot loading speed depends on having full core */
420 0 : "snapdc", /* TODO: Snapshot loading speed depends on having full core */
421 0 : "snapin", /* TODO: Snapshot loading speed depends on having full core */
422 0 : "snapwm", /* TODO: Snapshot loading speed depends on having full core */
423 0 : "snapwh", /* TODO: Snapshot loading speed depends on having full core */
424 0 : };
425 :
426 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
427 0 : fd_topo_tile_t * tile = &topo->tiles[ i ];
428 0 : tile->cpu_idx = ULONG_MAX;
429 0 : }
430 :
431 0 : fd_topo_cpus_t cpus[1];
432 0 : fd_topo_cpus_init( cpus );
433 :
434 0 : ulong cpu_ordering[ FD_TILE_MAX ] = { 0UL };
435 0 : int pairs_assigned[ FD_TILE_MAX ] = { 0 };
436 :
437 0 : ulong next_cpu_idx = 0UL;
438 0 : for( ulong i=0UL; i<cpus->numa_node_cnt; i++ ) {
439 0 : for( ulong j=0UL; j<cpus->cpu_cnt; j++ ) {
440 0 : fd_topo_cpu_t * cpu = &cpus->cpu[ j ];
441 :
442 0 : if( FD_UNLIKELY( pairs_assigned[ j ] || cpu->numa_node!=i ) ) continue;
443 :
444 0 : FD_TEST( next_cpu_idx<FD_TILE_MAX );
445 0 : cpu_ordering[ next_cpu_idx++ ] = j;
446 :
447 0 : if( FD_UNLIKELY( cpu->sibling!=ULONG_MAX ) ) {
448 : /* If the CPU has a HT pair, place it immediately after so they
449 : are sequentially assigned. */
450 0 : FD_TEST( next_cpu_idx<FD_TILE_MAX );
451 0 : cpu_ordering[ next_cpu_idx++ ] = cpu->sibling;
452 0 : pairs_assigned[ cpu->sibling ] = 1;
453 0 : }
454 0 : }
455 0 : }
456 :
457 0 : FD_TEST( next_cpu_idx==cpus->cpu_cnt );
458 :
459 0 : int cpu_assigned[ FD_TILE_MAX ] = {0};
460 :
461 0 : ulong cpu_idx = 0UL;
462 0 : for( ulong i=0UL; i<sizeof(ORDERED)/sizeof(ORDERED[0]); i++ ) {
463 0 : for( ulong j=0UL; j<topo->tile_cnt; j++ ) {
464 0 : fd_topo_tile_t * tile = &topo->tiles[ j ];
465 0 : if( !strcmp( tile->name, ORDERED[ i ] ) ) {
466 0 : if( FD_UNLIKELY( cpu_idx>=cpus->cpu_cnt ) ) {
467 0 : FD_LOG_ERR(( "auto layout cannot set affinity for tile `%s:%lu` because all the CPUs are already assigned", tile->name, tile->kind_id ));
468 0 : } else {
469 : /* Certain tiles are latency and throughput critical and
470 : should not get a HT pair assigned. */
471 0 : fd_topo_cpu_t const * cpu = &cpus->cpu[ cpu_ordering[ cpu_idx ] ];
472 :
473 0 : int is_ht_critical = 0;
474 0 : if( FD_UNLIKELY( cpu->sibling!=ULONG_MAX ) ) {
475 0 : for( ulong k=0UL; k<sizeof(CRITICAL_TILES)/sizeof(CRITICAL_TILES[0]); k++ ) {
476 0 : if( !strcmp( tile->name, CRITICAL_TILES[ k ] ) ) {
477 0 : is_ht_critical = 1;
478 0 : break;
479 0 : }
480 0 : }
481 0 : }
482 :
483 0 : if( FD_UNLIKELY( is_ht_critical ) ) {
484 0 : ulong try_assign = cpu_idx;
485 0 : while( cpu_assigned[ cpu_ordering[ try_assign ] ] || (cpus->cpu[ cpu_ordering[ try_assign ] ].sibling!=ULONG_MAX && cpu_assigned[ cpus->cpu[ cpu_ordering[ try_assign ] ].sibling ]) ) {
486 0 : try_assign++;
487 0 : if( FD_UNLIKELY( try_assign>=cpus->cpu_cnt ) ) FD_LOG_ERR(( "auto layout cannot set affinity for tile `%s:%lu` because all the CPUs are already assigned or have a HT pair assigned", tile->name, tile->kind_id ));
488 0 : }
489 :
490 0 : ulong sibling = cpus->cpu[ cpu_ordering[ try_assign ] ].sibling;
491 0 : cpu_assigned[ cpu_ordering[ try_assign ] ] = 1;
492 0 : if( sibling!=ULONG_MAX ) {
493 0 : cpu_assigned[ sibling ] = 1;
494 0 : }
495 0 : tile->cpu_idx = cpu_ordering[ try_assign ];
496 0 : while( cpu_assigned[ cpu_ordering[ cpu_idx ] ] ) cpu_idx++;
497 0 : } else {
498 0 : cpu_assigned[ cpu_ordering[ cpu_idx ] ] = 1;
499 0 : tile->cpu_idx = cpu_ordering[ cpu_idx ];
500 0 : while( cpu_assigned[ cpu_ordering[ cpu_idx ] ] ) cpu_idx++;
501 0 : }
502 0 : }
503 0 : }
504 0 : }
505 0 : }
506 :
507 : /* Make sure all the tiles we haven't set are supposed to be floating. */
508 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
509 0 : fd_topo_tile_t * tile = &topo->tiles[ i ];
510 0 : if( tile->cpu_idx!=ULONG_MAX ) continue;
511 :
512 0 : int found = 0;
513 0 : for( ulong j=0UL; j<sizeof(FLOATING)/sizeof(FLOATING[0]); j++ ) {
514 0 : if( !strcmp( tile->name, FLOATING[ j ] ) ) {
515 0 : found = 1;
516 0 : break;
517 0 : }
518 0 : }
519 :
520 0 : if( FD_UNLIKELY( !found ) ) FD_LOG_WARNING(( "auto layout cannot affine tile `%s:%lu` because it is unknown. Leaving it floating", tile->name, tile->kind_id ));
521 0 : }
522 :
523 0 : if( FD_UNLIKELY( reserve_agave_cores ) ) {
524 0 : for( ulong i=cpu_idx; i<cpus->cpu_cnt; i++ ) {
525 0 : if( FD_UNLIKELY( !cpus->cpu[ cpu_ordering[ i ] ].online ) ) continue;
526 :
527 0 : if( FD_LIKELY( topo->agave_affinity_cnt<sizeof(topo->agave_affinity_cpu_idx)/sizeof(topo->agave_affinity_cpu_idx[0]) ) ) {
528 0 : topo->agave_affinity_cpu_idx[ topo->agave_affinity_cnt++ ] = cpu_ordering[ i ];
529 0 : }
530 0 : }
531 0 : }
532 0 : }
533 :
534 : ulong
535 : fd_numa_node_idx( ulong cpu_idx );
536 :
537 : static void
538 0 : initialize_numa_assignments( fd_topo_t * topo ) {
539 : /* Assign workspaces to NUMA nodes. The heuristic here is pretty
540 : simple for now: workspaces go on the NUMA node of the first
541 : tile which maps the largest object in the workspace. */
542 :
543 0 : for( ulong i=0UL; i<topo->wksp_cnt; i++ ) {
544 0 : ulong max_footprint = 0UL;
545 0 : ulong max_obj = ULONG_MAX;
546 :
547 0 : for( ulong j=0UL; j<topo->obj_cnt; j++ ) {
548 0 : fd_topo_obj_t * obj = &topo->objs[ j ];
549 0 : if( obj->wksp_id!=i ) continue;
550 0 : if( FD_UNLIKELY( !obj->footprint ) ) FD_LOG_ERR(( "obj %lu (%s) has invalid parameters", j, obj->name ));
551 :
552 0 : if( FD_UNLIKELY( !max_footprint || obj->footprint>max_footprint ) ) {
553 0 : max_footprint = obj->footprint;
554 0 : max_obj = j;
555 0 : }
556 0 : }
557 :
558 0 : if( FD_UNLIKELY( max_obj==ULONG_MAX ) ) FD_LOG_ERR(( "no object found for workspace %s", topo->workspaces[ i ].name ));
559 :
560 0 : int found_strict = 0;
561 0 : int found_lazy = 0;
562 0 : for( ulong j=0UL; j<topo->tile_cnt; j++ ) {
563 0 : fd_topo_tile_t * tile = &topo->tiles[ j ];
564 0 : if( FD_UNLIKELY( tile->tile_obj_id==max_obj && tile->cpu_idx<FD_TILE_MAX ) ) {
565 0 : topo->workspaces[ i ].numa_idx = fd_numa_node_idx( tile->cpu_idx );
566 0 : FD_TEST( topo->workspaces[ i ].numa_idx!=ULONG_MAX );
567 0 : found_strict = 1;
568 0 : found_lazy = 1;
569 0 : break;
570 0 : } else if( FD_UNLIKELY( tile->tile_obj_id==max_obj && tile->cpu_idx>=FD_TILE_MAX ) ) {
571 0 : topo->workspaces[ i ].numa_idx = 0;
572 0 : found_lazy = 1;
573 0 : break;
574 0 : }
575 0 : }
576 :
577 0 : if( FD_LIKELY( !found_strict ) ) {
578 0 : for( ulong j=0UL; j<topo->tile_cnt; j++ ) {
579 0 : fd_topo_tile_t * tile = &topo->tiles[ j ];
580 0 : for( ulong k=0UL; k<tile->uses_obj_cnt; k++ ) {
581 0 : if( FD_LIKELY( tile->uses_obj_id[ k ]==max_obj && tile->cpu_idx<FD_TILE_MAX ) ) {
582 0 : topo->workspaces[ i ].numa_idx = fd_numa_node_idx( tile->cpu_idx );
583 0 : FD_TEST( topo->workspaces[ i ].numa_idx!=ULONG_MAX );
584 0 : found_lazy = 1;
585 0 : break;
586 0 : } else if( FD_UNLIKELY( tile->uses_obj_id[ k ]==max_obj ) && tile->cpu_idx>=FD_TILE_MAX ) {
587 0 : topo->workspaces[ i ].numa_idx = 0;
588 0 : found_lazy = 1;
589 : /* Don't break, keep looking -- a tile with a CPU assignment
590 : might also use object in which case we want to use that
591 : NUMA node. */
592 0 : }
593 0 : }
594 :
595 0 : if( FD_UNLIKELY( found_lazy ) ) break;
596 0 : }
597 0 : }
598 :
599 0 : if( FD_UNLIKELY( !found_lazy ) ) FD_LOG_ERR(( "no tile uses object %s for workspace %s", topo->objs[ max_obj ].name, topo->workspaces[ i ].name ));
600 0 : }
601 0 : }
602 :
603 : void
604 : fd_topob_finish( fd_topo_t * topo,
605 0 : fd_topo_obj_callbacks_t ** callbacks ) {
606 0 : for( ulong z=0UL; z<topo->tile_cnt; z++ ) {
607 0 : fd_topo_tile_t * tile = &topo->tiles[ z ];
608 :
609 0 : ulong in_cnt = 0UL;
610 0 : for( ulong i=0UL; i<tile->in_cnt; i++ ) {
611 0 : if( FD_UNLIKELY( !tile->in_link_poll[ i ] ) ) continue;
612 0 : in_cnt++;
613 0 : }
614 :
615 0 : ulong cons_cnt = 0UL;
616 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
617 0 : fd_topo_tile_t * consumer_tile = &topo->tiles[ i ];
618 0 : for( ulong j=0UL; j<consumer_tile->in_cnt; j++ ) {
619 0 : for( ulong k=0UL; k<tile->out_cnt; k++ ) {
620 0 : if( FD_UNLIKELY( consumer_tile->in_link_id[ j ]==tile->out_link_id[ k ] && consumer_tile->in_link_reliable[ j ] ) ) {
621 0 : cons_cnt++;
622 0 : }
623 0 : }
624 0 : }
625 0 : }
626 :
627 0 : FD_TEST( !fd_pod_replacef_ulong( topo->props, in_cnt, "obj.%lu.in_cnt", tile->metrics_obj_id ) );
628 0 : FD_TEST( !fd_pod_replacef_ulong( topo->props, cons_cnt, "obj.%lu.cons_cnt", tile->metrics_obj_id ) );
629 0 : }
630 :
631 0 : for( ulong i=0UL; i<topo->wksp_cnt; i++ ) {
632 0 : fd_topo_wksp_t * wksp = &topo->workspaces[ i ];
633 :
634 0 : ulong loose_sz = 0UL;
635 0 : for( ulong j=0UL; j<topo->obj_cnt; j++ ) {
636 0 : fd_topo_obj_t * obj = &topo->objs[ j ];
637 0 : if( FD_UNLIKELY( obj->wksp_id!=wksp->id ) ) continue;
638 :
639 0 : fd_topo_obj_callbacks_t * cb = NULL;
640 0 : for( ulong i=0UL; callbacks[ i ]; i++ ) {
641 0 : if( FD_UNLIKELY( !strcmp( callbacks[ i ]->name, obj->name ) ) ) {
642 0 : cb = callbacks[ i ];
643 0 : break;
644 0 : }
645 0 : }
646 0 : if( FD_UNLIKELY( !cb ) ) FD_LOG_ERR(( "no callbacks for object %s", obj->name ));
647 :
648 0 : if( FD_UNLIKELY( cb->loose ) ) loose_sz += cb->loose( topo, obj );
649 0 : }
650 :
651 0 : ulong part_max = wksp->part_max;
652 0 : if( !part_max ) part_max = (loose_sz / (64UL << 10)); /* alloc + residual padding */
653 0 : part_max += 3; /* for initial alignment */
654 0 : ulong offset = fd_ulong_align_up( fd_wksp_private_data_off( part_max ), fd_topo_workspace_align() );
655 :
656 0 : for( ulong j=0UL; j<topo->obj_cnt; j++ ) {
657 0 : fd_topo_obj_t * obj = &topo->objs[ j ];
658 0 : if( FD_UNLIKELY( obj->wksp_id!=wksp->id ) ) continue;
659 :
660 0 : fd_topo_obj_callbacks_t * cb = NULL;
661 0 : for( ulong i=0UL; callbacks[ i ]; i++ ) {
662 0 : if( FD_UNLIKELY( !strcmp( callbacks[ i ]->name, obj->name ) ) ) {
663 0 : cb = callbacks[ i ];
664 0 : break;
665 0 : }
666 0 : }
667 0 : if( FD_UNLIKELY( !cb ) ) FD_LOG_ERR(( "no callbacks for object %s", obj->name ));
668 :
669 0 : ulong align_ = cb->align( topo, obj );
670 0 : if( FD_UNLIKELY( !fd_ulong_is_pow2( align_ ) ) ) FD_LOG_ERR(( "Return value of fdctl_obj_align(%s,%lu) is not a power of 2", obj->name, obj->id ));
671 0 : offset = fd_ulong_align_up( offset, align_ );
672 0 : obj->offset = offset;
673 0 : obj->footprint = cb->footprint( topo, obj );
674 0 : if( FD_UNLIKELY( 0!=strcmp( obj->name, "tile" ) && (!obj->footprint || obj->footprint>LONG_MAX) ) ) {
675 0 : FD_LOG_ERR(( "fdctl_obj_footprint(%s,%lu) failed", obj->name, obj->id ));
676 0 : }
677 0 : offset += obj->footprint;
678 0 : }
679 :
680 0 : ulong footprint = fd_ulong_align_up( offset, fd_topo_workspace_align() );
681 :
682 0 : part_max = fd_ulong_max( part_max, wksp->min_part_max );
683 0 : loose_sz = fd_ulong_max( loose_sz, wksp->min_loose_sz );
684 :
685 : /* Compute footprint for a workspace that can store our footprint,
686 : with an extra align of padding incase gaddr_lo is not aligned. */
687 0 : ulong total_wksp_footprint = fd_wksp_footprint( part_max, footprint + fd_topo_workspace_align() + loose_sz );
688 :
689 0 : ulong page_sz = topo->max_page_size;
690 0 : if( total_wksp_footprint < topo->gigantic_page_threshold ) page_sz = FD_SHMEM_HUGE_PAGE_SZ;
691 0 : if( FD_UNLIKELY( page_sz!=FD_SHMEM_HUGE_PAGE_SZ && page_sz!=FD_SHMEM_GIGANTIC_PAGE_SZ ) ) FD_LOG_ERR(( "invalid page_sz" ));
692 :
693 0 : ulong wksp_aligned_footprint = fd_ulong_align_up( total_wksp_footprint, page_sz );
694 :
695 : /* Give any leftover space in the underlying shared memory to the
696 : data region of the workspace, since we might as well use it. */
697 0 : wksp->part_max = part_max;
698 0 : wksp->known_footprint = footprint;
699 0 : wksp->total_footprint = wksp_aligned_footprint - fd_ulong_align_up( fd_wksp_private_data_off( part_max ), fd_topo_workspace_align() );
700 0 : wksp->page_sz = page_sz;
701 0 : wksp->page_cnt = wksp_aligned_footprint / page_sz;
702 0 : }
703 :
704 0 : initialize_numa_assignments( topo );
705 :
706 0 : validate( topo );
707 0 : }
|
