1
2
3
4
5
6
7
8
9
10
11
12
13#include "qemu/osdep.h"
14#include "qemu-common.h"
15#include "hw/ppc/spapr_numa.h"
16#include "hw/pci-host/spapr.h"
17#include "hw/ppc/fdt.h"
18
19
20#define SPAPR_GPU_NUMA_ID (cpu_to_be32(1))
21
22static bool spapr_machine_using_legacy_numa(SpaprMachineState *spapr)
23{
24 MachineState *machine = MACHINE(spapr);
25 SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(machine);
26
27 return smc->pre_5_2_numa_associativity ||
28 machine->numa_state->num_nodes <= 1;
29}
30
31static bool spapr_numa_is_symmetrical(MachineState *ms)
32{
33 int src, dst;
34 int nb_numa_nodes = ms->numa_state->num_nodes;
35 NodeInfo *numa_info = ms->numa_state->nodes;
36
37 for (src = 0; src < nb_numa_nodes; src++) {
38 for (dst = src; dst < nb_numa_nodes; dst++) {
39 if (numa_info[src].distance[dst] !=
40 numa_info[dst].distance[src]) {
41 return false;
42 }
43 }
44 }
45
46 return true;
47}
48
49
50
51
52
53
54
55
56
57
58unsigned int spapr_numa_initial_nvgpu_numa_id(MachineState *machine)
59{
60 return MAX(1, machine->numa_state->num_nodes);
61}
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80static uint8_t spapr_numa_get_numa_level(uint8_t distance)
81{
82 if (distance == 10) {
83 return 0x4;
84 } else if (distance > 11 && distance <= 30) {
85 return 0x3;
86 } else if (distance > 31 && distance <= 60) {
87 return 0x2;
88 } else if (distance > 61 && distance <= 120) {
89 return 0x1;
90 }
91
92 return 0;
93}
94
95static void spapr_numa_define_associativity_domains(SpaprMachineState *spapr)
96{
97 MachineState *ms = MACHINE(spapr);
98 NodeInfo *numa_info = ms->numa_state->nodes;
99 int nb_numa_nodes = ms->numa_state->num_nodes;
100 int src, dst, i;
101
102 for (src = 0; src < nb_numa_nodes; src++) {
103 for (dst = src; dst < nb_numa_nodes; dst++) {
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124 uint8_t distance = numa_info[src].distance[dst];
125 uint8_t n_level = spapr_numa_get_numa_level(distance);
126 uint32_t assoc_src;
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141 if (n_level == 0) {
142 continue;
143 }
144
145
146
147
148
149 for (i = n_level; i > 0; i--) {
150 assoc_src = spapr->numa_assoc_array[src][i];
151 spapr->numa_assoc_array[dst][i] = assoc_src;
152 }
153 }
154 }
155
156}
157
158void spapr_numa_associativity_init(SpaprMachineState *spapr,
159 MachineState *machine)
160{
161 SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(spapr);
162 int nb_numa_nodes = machine->numa_state->num_nodes;
163 int i, j, max_nodes_with_gpus;
164 bool using_legacy_numa = spapr_machine_using_legacy_numa(spapr);
165
166
167
168
169
170
171
172
173
174
175 for (i = 0; i < nb_numa_nodes; i++) {
176 spapr->numa_assoc_array[i][0] = cpu_to_be32(MAX_DISTANCE_REF_POINTS);
177 spapr->numa_assoc_array[i][MAX_DISTANCE_REF_POINTS] = cpu_to_be32(i);
178
179
180
181
182
183
184 if (!using_legacy_numa && i != 0) {
185 for (j = 1; j < MAX_DISTANCE_REF_POINTS; j++) {
186 spapr->numa_assoc_array[i][j] = cpu_to_be32(i);
187 }
188 }
189 }
190
191
192
193
194
195
196
197
198
199 max_nodes_with_gpus = nb_numa_nodes + NVGPU_MAX_NUM;
200
201 for (i = nb_numa_nodes; i < max_nodes_with_gpus; i++) {
202 spapr->numa_assoc_array[i][0] = cpu_to_be32(MAX_DISTANCE_REF_POINTS);
203
204 for (j = 1; j < MAX_DISTANCE_REF_POINTS; j++) {
205 uint32_t gpu_assoc = smc->pre_5_1_assoc_refpoints ?
206 SPAPR_GPU_NUMA_ID : cpu_to_be32(i);
207 spapr->numa_assoc_array[i][j] = gpu_assoc;
208 }
209
210 spapr->numa_assoc_array[i][MAX_DISTANCE_REF_POINTS] = cpu_to_be32(i);
211 }
212
213
214
215
216
217
218 if (using_legacy_numa) {
219 return;
220 }
221
222 if (!spapr_numa_is_symmetrical(machine)) {
223 error_report("Asymmetrical NUMA topologies aren't supported "
224 "in the pSeries machine");
225 exit(EXIT_FAILURE);
226 }
227
228 spapr_numa_define_associativity_domains(spapr);
229}
230
231void spapr_numa_write_associativity_dt(SpaprMachineState *spapr, void *fdt,
232 int offset, int nodeid)
233{
234 _FDT((fdt_setprop(fdt, offset, "ibm,associativity",
235 spapr->numa_assoc_array[nodeid],
236 sizeof(spapr->numa_assoc_array[nodeid]))));
237}
238
239static uint32_t *spapr_numa_get_vcpu_assoc(SpaprMachineState *spapr,
240 PowerPCCPU *cpu)
241{
242 uint32_t *vcpu_assoc = g_new(uint32_t, VCPU_ASSOC_SIZE);
243 int index = spapr_get_vcpu_id(cpu);
244
245
246
247
248
249
250
251 vcpu_assoc[0] = cpu_to_be32(MAX_DISTANCE_REF_POINTS + 1);
252 vcpu_assoc[VCPU_ASSOC_SIZE - 1] = cpu_to_be32(index);
253 memcpy(vcpu_assoc + 1, spapr->numa_assoc_array[cpu->node_id] + 1,
254 (VCPU_ASSOC_SIZE - 2) * sizeof(uint32_t));
255
256 return vcpu_assoc;
257}
258
259int spapr_numa_fixup_cpu_dt(SpaprMachineState *spapr, void *fdt,
260 int offset, PowerPCCPU *cpu)
261{
262 g_autofree uint32_t *vcpu_assoc = NULL;
263
264 vcpu_assoc = spapr_numa_get_vcpu_assoc(spapr, cpu);
265
266
267 return fdt_setprop(fdt, offset, "ibm,associativity", vcpu_assoc,
268 VCPU_ASSOC_SIZE * sizeof(uint32_t));
269}
270
271
272int spapr_numa_write_assoc_lookup_arrays(SpaprMachineState *spapr, void *fdt,
273 int offset)
274{
275 MachineState *machine = MACHINE(spapr);
276 int nb_numa_nodes = machine->numa_state->num_nodes;
277 int nr_nodes = nb_numa_nodes ? nb_numa_nodes : 1;
278 uint32_t *int_buf, *cur_index, buf_len;
279 int ret, i;
280
281
282 buf_len = (nr_nodes * MAX_DISTANCE_REF_POINTS + 2) * sizeof(uint32_t);
283 cur_index = int_buf = g_malloc0(buf_len);
284 int_buf[0] = cpu_to_be32(nr_nodes);
285
286 int_buf[1] = cpu_to_be32(MAX_DISTANCE_REF_POINTS);
287 cur_index += 2;
288 for (i = 0; i < nr_nodes; i++) {
289
290
291
292
293 uint32_t *associativity = spapr->numa_assoc_array[i];
294 memcpy(cur_index, ++associativity,
295 sizeof(uint32_t) * MAX_DISTANCE_REF_POINTS);
296 cur_index += MAX_DISTANCE_REF_POINTS;
297 }
298 ret = fdt_setprop(fdt, offset, "ibm,associativity-lookup-arrays", int_buf,
299 (cur_index - int_buf) * sizeof(uint32_t));
300 g_free(int_buf);
301
302 return ret;
303}
304
305
306
307
308
309
310void spapr_numa_write_rtas_dt(SpaprMachineState *spapr, void *fdt, int rtas)
311{
312 MachineState *ms = MACHINE(spapr);
313 SpaprMachineClass *smc = SPAPR_MACHINE_GET_CLASS(spapr);
314 uint32_t number_nvgpus_nodes = spapr->gpu_numa_id -
315 spapr_numa_initial_nvgpu_numa_id(ms);
316 uint32_t refpoints[] = {
317 cpu_to_be32(0x4),
318 cpu_to_be32(0x3),
319 cpu_to_be32(0x2),
320 cpu_to_be32(0x1),
321 };
322 uint32_t nr_refpoints = ARRAY_SIZE(refpoints);
323 uint32_t maxdomain = ms->numa_state->num_nodes + number_nvgpus_nodes;
324 uint32_t maxdomains[] = {
325 cpu_to_be32(4),
326 cpu_to_be32(maxdomain),
327 cpu_to_be32(maxdomain),
328 cpu_to_be32(maxdomain),
329 cpu_to_be32(maxdomain)
330 };
331
332 if (spapr_machine_using_legacy_numa(spapr)) {
333 uint32_t legacy_refpoints[] = {
334 cpu_to_be32(0x4),
335 cpu_to_be32(0x4),
336 cpu_to_be32(0x2),
337 };
338 uint32_t legacy_maxdomain = spapr->gpu_numa_id > 1 ? 1 : 0;
339 uint32_t legacy_maxdomains[] = {
340 cpu_to_be32(4),
341 cpu_to_be32(legacy_maxdomain),
342 cpu_to_be32(legacy_maxdomain),
343 cpu_to_be32(legacy_maxdomain),
344 cpu_to_be32(spapr->gpu_numa_id),
345 };
346
347 G_STATIC_ASSERT(sizeof(legacy_refpoints) <= sizeof(refpoints));
348 G_STATIC_ASSERT(sizeof(legacy_maxdomains) <= sizeof(maxdomains));
349
350 nr_refpoints = 3;
351
352 memcpy(refpoints, legacy_refpoints, sizeof(legacy_refpoints));
353 memcpy(maxdomains, legacy_maxdomains, sizeof(legacy_maxdomains));
354
355
356 if (smc->pre_5_1_assoc_refpoints) {
357 nr_refpoints = 2;
358 }
359 }
360
361 _FDT(fdt_setprop(fdt, rtas, "ibm,associativity-reference-points",
362 refpoints, nr_refpoints * sizeof(refpoints[0])));
363
364 _FDT(fdt_setprop(fdt, rtas, "ibm,max-associativity-domains",
365 maxdomains, sizeof(maxdomains)));
366}
367
368static target_ulong h_home_node_associativity(PowerPCCPU *cpu,
369 SpaprMachineState *spapr,
370 target_ulong opcode,
371 target_ulong *args)
372{
373 g_autofree uint32_t *vcpu_assoc = NULL;
374 target_ulong flags = args[0];
375 target_ulong procno = args[1];
376 PowerPCCPU *tcpu;
377 int idx, assoc_idx;
378
379
380 if (flags != 0x1) {
381 return H_FUNCTION;
382 }
383
384 tcpu = spapr_find_cpu(procno);
385 if (tcpu == NULL) {
386 return H_P2;
387 }
388
389
390
391
392
393
394
395
396 G_STATIC_ASSERT((VCPU_ASSOC_SIZE - 1) <= 12);
397
398 vcpu_assoc = spapr_numa_get_vcpu_assoc(spapr, tcpu);
399
400 assoc_idx = 1;
401
402#define ASSOCIATIVITY(a, b) (((uint64_t)(a) << 32) | \
403 ((uint64_t)(b) & 0xffffffff))
404
405 for (idx = 0; idx < 6; idx++) {
406 int32_t a, b;
407
408
409
410
411
412
413
414
415
416
417 a = assoc_idx < VCPU_ASSOC_SIZE ?
418 be32_to_cpu(vcpu_assoc[assoc_idx++]) : -1;
419 b = assoc_idx < VCPU_ASSOC_SIZE ?
420 be32_to_cpu(vcpu_assoc[assoc_idx++]) : -1;
421
422 args[idx] = ASSOCIATIVITY(a, b);
423 }
424#undef ASSOCIATIVITY
425
426 return H_SUCCESS;
427}
428
429static void spapr_numa_register_types(void)
430{
431
432 spapr_register_hypercall(H_HOME_NODE_ASSOCIATIVITY,
433 h_home_node_associativity);
434}
435
436type_init(spapr_numa_register_types)
437