aboutsummaryrefslogtreecommitdiffstats
path: root/arch/x86/kernel/cpu/amd.c
diff options
context:
space:
mode:
Diffstat (limited to 'arch/x86/kernel/cpu/amd.c')
-rw-r--r--arch/x86/kernel/cpu/amd.c61
1 files changed, 39 insertions, 22 deletions
diff --git a/arch/x86/kernel/cpu/amd.c b/arch/x86/kernel/cpu/amd.c
index 7c7bedb83c5a..f771ab6b49e9 100644
--- a/arch/x86/kernel/cpu/amd.c
+++ b/arch/x86/kernel/cpu/amd.c
@@ -233,18 +233,22 @@ static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
233} 233}
234#endif 234#endif
235 235
236#if defined(CONFIG_NUMA) && defined(CONFIG_X86_64) 236#ifdef CONFIG_NUMA
237/*
238 * To workaround broken NUMA config. Read the comment in
239 * srat_detect_node().
240 */
237static int __cpuinit nearby_node(int apicid) 241static int __cpuinit nearby_node(int apicid)
238{ 242{
239 int i, node; 243 int i, node;
240 244
241 for (i = apicid - 1; i >= 0; i--) { 245 for (i = apicid - 1; i >= 0; i--) {
242 node = apicid_to_node[i]; 246 node = __apicid_to_node[i];
243 if (node != NUMA_NO_NODE && node_online(node)) 247 if (node != NUMA_NO_NODE && node_online(node))
244 return node; 248 return node;
245 } 249 }
246 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) { 250 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
247 node = apicid_to_node[i]; 251 node = __apicid_to_node[i];
248 if (node != NUMA_NO_NODE && node_online(node)) 252 if (node != NUMA_NO_NODE && node_online(node))
249 return node; 253 return node;
250 } 254 }
@@ -261,7 +265,7 @@ static int __cpuinit nearby_node(int apicid)
261#ifdef CONFIG_X86_HT 265#ifdef CONFIG_X86_HT
262static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c) 266static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
263{ 267{
264 u32 nodes; 268 u32 nodes, cores_per_cu = 1;
265 u8 node_id; 269 u8 node_id;
266 int cpu = smp_processor_id(); 270 int cpu = smp_processor_id();
267 271
@@ -276,6 +280,7 @@ static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
276 /* get compute unit information */ 280 /* get compute unit information */
277 smp_num_siblings = ((ebx >> 8) & 3) + 1; 281 smp_num_siblings = ((ebx >> 8) & 3) + 1;
278 c->compute_unit_id = ebx & 0xff; 282 c->compute_unit_id = ebx & 0xff;
283 cores_per_cu += ((ebx >> 8) & 3);
279 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) { 284 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
280 u64 value; 285 u64 value;
281 286
@@ -288,15 +293,18 @@ static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
288 /* fixup multi-node processor information */ 293 /* fixup multi-node processor information */
289 if (nodes > 1) { 294 if (nodes > 1) {
290 u32 cores_per_node; 295 u32 cores_per_node;
296 u32 cus_per_node;
291 297
292 set_cpu_cap(c, X86_FEATURE_AMD_DCM); 298 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
293 cores_per_node = c->x86_max_cores / nodes; 299 cores_per_node = c->x86_max_cores / nodes;
300 cus_per_node = cores_per_node / cores_per_cu;
294 301
295 /* store NodeID, use llc_shared_map to store sibling info */ 302 /* store NodeID, use llc_shared_map to store sibling info */
296 per_cpu(cpu_llc_id, cpu) = node_id; 303 per_cpu(cpu_llc_id, cpu) = node_id;
297 304
298 /* core id to be in range from 0 to (cores_per_node - 1) */ 305 /* core id has to be in the [0 .. cores_per_node - 1] range */
299 c->cpu_core_id = c->cpu_core_id % cores_per_node; 306 c->cpu_core_id %= cores_per_node;
307 c->compute_unit_id %= cus_per_node;
300 } 308 }
301} 309}
302#endif 310#endif
@@ -334,31 +342,40 @@ EXPORT_SYMBOL_GPL(amd_get_nb_id);
334 342
335static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c) 343static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
336{ 344{
337#if defined(CONFIG_NUMA) && defined(CONFIG_X86_64) 345#ifdef CONFIG_NUMA
338 int cpu = smp_processor_id(); 346 int cpu = smp_processor_id();
339 int node; 347 int node;
340 unsigned apicid = c->apicid; 348 unsigned apicid = c->apicid;
341 349
342 node = per_cpu(cpu_llc_id, cpu); 350 node = numa_cpu_node(cpu);
351 if (node == NUMA_NO_NODE)
352 node = per_cpu(cpu_llc_id, cpu);
343 353
344 if (apicid_to_node[apicid] != NUMA_NO_NODE)
345 node = apicid_to_node[apicid];
346 if (!node_online(node)) { 354 if (!node_online(node)) {
347 /* Two possibilities here: 355 /*
348 - The CPU is missing memory and no node was created. 356 * Two possibilities here:
349 In that case try picking one from a nearby CPU 357 *
350 - The APIC IDs differ from the HyperTransport node IDs 358 * - The CPU is missing memory and no node was created. In
351 which the K8 northbridge parsing fills in. 359 * that case try picking one from a nearby CPU.
352 Assume they are all increased by a constant offset, 360 *
353 but in the same order as the HT nodeids. 361 * - The APIC IDs differ from the HyperTransport node IDs
354 If that doesn't result in a usable node fall back to the 362 * which the K8 northbridge parsing fills in. Assume
355 path for the previous case. */ 363 * they are all increased by a constant offset, but in
356 364 * the same order as the HT nodeids. If that doesn't
365 * result in a usable node fall back to the path for the
366 * previous case.
367 *
368 * This workaround operates directly on the mapping between
369 * APIC ID and NUMA node, assuming certain relationship
370 * between APIC ID, HT node ID and NUMA topology. As going
371 * through CPU mapping may alter the outcome, directly
372 * access __apicid_to_node[].
373 */
357 int ht_nodeid = c->initial_apicid; 374 int ht_nodeid = c->initial_apicid;
358 375
359 if (ht_nodeid >= 0 && 376 if (ht_nodeid >= 0 &&
360 apicid_to_node[ht_nodeid] != NUMA_NO_NODE) 377 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
361 node = apicid_to_node[ht_nodeid]; 378 node = __apicid_to_node[ht_nodeid];
362 /* Pick a nearby node */ 379 /* Pick a nearby node */
363 if (!node_online(node)) 380 if (!node_online(node))
364 node = nearby_node(apicid); 381 node = nearby_node(apicid);