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