]>
Commit | Line | Data |
---|---|---|
69c60c88 | 1 | #include <linux/export.h> |
1da177e4 | 2 | #include <linux/bitops.h> |
5cdd174f | 3 | #include <linux/elf.h> |
1da177e4 | 4 | #include <linux/mm.h> |
8d71a2ea | 5 | |
8bdbd962 | 6 | #include <linux/io.h> |
c98fdeaa | 7 | #include <linux/sched.h> |
4e26d11f | 8 | #include <linux/random.h> |
1da177e4 | 9 | #include <asm/processor.h> |
d3f7eae1 | 10 | #include <asm/apic.h> |
1f442d70 | 11 | #include <asm/cpu.h> |
26bfa5f8 | 12 | #include <asm/smp.h> |
42937e81 | 13 | #include <asm/pci-direct.h> |
b466bdb6 | 14 | #include <asm/delay.h> |
1da177e4 | 15 | |
8d71a2ea | 16 | #ifdef CONFIG_X86_64 |
8d71a2ea YL |
17 | # include <asm/mmconfig.h> |
18 | # include <asm/cacheflush.h> | |
19 | #endif | |
20 | ||
1da177e4 LT |
21 | #include "cpu.h" |
22 | ||
cc2749e4 AG |
23 | /* |
24 | * nodes_per_socket: Stores the number of nodes per socket. | |
25 | * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX | |
26 | * Node Identifiers[10:8] | |
27 | */ | |
28 | static u32 nodes_per_socket = 1; | |
29 | ||
2c929ce6 BP |
30 | static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p) |
31 | { | |
2c929ce6 BP |
32 | u32 gprs[8] = { 0 }; |
33 | int err; | |
34 | ||
682469a5 BP |
35 | WARN_ONCE((boot_cpu_data.x86 != 0xf), |
36 | "%s should only be used on K8!\n", __func__); | |
2c929ce6 BP |
37 | |
38 | gprs[1] = msr; | |
39 | gprs[7] = 0x9c5a203a; | |
40 | ||
41 | err = rdmsr_safe_regs(gprs); | |
42 | ||
43 | *p = gprs[0] | ((u64)gprs[2] << 32); | |
44 | ||
45 | return err; | |
46 | } | |
47 | ||
48 | static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val) | |
49 | { | |
2c929ce6 BP |
50 | u32 gprs[8] = { 0 }; |
51 | ||
682469a5 BP |
52 | WARN_ONCE((boot_cpu_data.x86 != 0xf), |
53 | "%s should only be used on K8!\n", __func__); | |
2c929ce6 BP |
54 | |
55 | gprs[0] = (u32)val; | |
56 | gprs[1] = msr; | |
57 | gprs[2] = val >> 32; | |
58 | gprs[7] = 0x9c5a203a; | |
59 | ||
60 | return wrmsr_safe_regs(gprs); | |
61 | } | |
62 | ||
1da177e4 LT |
63 | /* |
64 | * B step AMD K6 before B 9730xxxx have hardware bugs that can cause | |
65 | * misexecution of code under Linux. Owners of such processors should | |
66 | * contact AMD for precise details and a CPU swap. | |
67 | * | |
68 | * See http://www.multimania.com/poulot/k6bug.html | |
d7de8649 AH |
69 | * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6" |
70 | * (Publication # 21266 Issue Date: August 1998) | |
1da177e4 LT |
71 | * |
72 | * The following test is erm.. interesting. AMD neglected to up | |
73 | * the chip setting when fixing the bug but they also tweaked some | |
74 | * performance at the same time.. | |
75 | */ | |
fb87a298 | 76 | |
277d5b40 | 77 | extern __visible void vide(void); |
de642faf JP |
78 | __asm__(".globl vide\n" |
79 | ".type vide, @function\n" | |
80 | ".align 4\n" | |
81 | "vide: ret\n"); | |
1da177e4 | 82 | |
148f9bb8 | 83 | static void init_amd_k5(struct cpuinfo_x86 *c) |
11fdd252 | 84 | { |
26bfa5f8 | 85 | #ifdef CONFIG_X86_32 |
11fdd252 YL |
86 | /* |
87 | * General Systems BIOSen alias the cpu frequency registers | |
88 | * of the Elan at 0x000df000. Unfortuantly, one of the Linux | |
89 | * drivers subsequently pokes it, and changes the CPU speed. | |
90 | * Workaround : Remove the unneeded alias. | |
91 | */ | |
92 | #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */ | |
93 | #define CBAR_ENB (0x80000000) | |
94 | #define CBAR_KEY (0X000000CB) | |
95 | if (c->x86_model == 9 || c->x86_model == 10) { | |
8bdbd962 AC |
96 | if (inl(CBAR) & CBAR_ENB) |
97 | outl(0 | CBAR_KEY, CBAR); | |
11fdd252 | 98 | } |
26bfa5f8 | 99 | #endif |
11fdd252 YL |
100 | } |
101 | ||
148f9bb8 | 102 | static void init_amd_k6(struct cpuinfo_x86 *c) |
11fdd252 | 103 | { |
26bfa5f8 | 104 | #ifdef CONFIG_X86_32 |
11fdd252 | 105 | u32 l, h; |
46a84132 | 106 | int mbytes = get_num_physpages() >> (20-PAGE_SHIFT); |
11fdd252 YL |
107 | |
108 | if (c->x86_model < 6) { | |
109 | /* Based on AMD doc 20734R - June 2000 */ | |
110 | if (c->x86_model == 0) { | |
111 | clear_cpu_cap(c, X86_FEATURE_APIC); | |
112 | set_cpu_cap(c, X86_FEATURE_PGE); | |
113 | } | |
114 | return; | |
115 | } | |
116 | ||
117 | if (c->x86_model == 6 && c->x86_mask == 1) { | |
118 | const int K6_BUG_LOOP = 1000000; | |
119 | int n; | |
120 | void (*f_vide)(void); | |
37963666 | 121 | u64 d, d2; |
11fdd252 YL |
122 | |
123 | printk(KERN_INFO "AMD K6 stepping B detected - "); | |
124 | ||
125 | /* | |
126 | * It looks like AMD fixed the 2.6.2 bug and improved indirect | |
127 | * calls at the same time. | |
128 | */ | |
129 | ||
130 | n = K6_BUG_LOOP; | |
131 | f_vide = vide; | |
4ea1636b | 132 | d = rdtsc(); |
11fdd252 YL |
133 | while (n--) |
134 | f_vide(); | |
4ea1636b | 135 | d2 = rdtsc(); |
11fdd252 YL |
136 | d = d2-d; |
137 | ||
138 | if (d > 20*K6_BUG_LOOP) | |
8bdbd962 AC |
139 | printk(KERN_CONT |
140 | "system stability may be impaired when more than 32 MB are used.\n"); | |
11fdd252 | 141 | else |
8bdbd962 | 142 | printk(KERN_CONT "probably OK (after B9730xxxx).\n"); |
11fdd252 YL |
143 | } |
144 | ||
145 | /* K6 with old style WHCR */ | |
146 | if (c->x86_model < 8 || | |
147 | (c->x86_model == 8 && c->x86_mask < 8)) { | |
148 | /* We can only write allocate on the low 508Mb */ | |
149 | if (mbytes > 508) | |
150 | mbytes = 508; | |
151 | ||
152 | rdmsr(MSR_K6_WHCR, l, h); | |
153 | if ((l&0x0000FFFF) == 0) { | |
154 | unsigned long flags; | |
155 | l = (1<<0)|((mbytes/4)<<1); | |
156 | local_irq_save(flags); | |
157 | wbinvd(); | |
158 | wrmsr(MSR_K6_WHCR, l, h); | |
159 | local_irq_restore(flags); | |
160 | printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n", | |
161 | mbytes); | |
162 | } | |
163 | return; | |
164 | } | |
165 | ||
166 | if ((c->x86_model == 8 && c->x86_mask > 7) || | |
167 | c->x86_model == 9 || c->x86_model == 13) { | |
168 | /* The more serious chips .. */ | |
169 | ||
170 | if (mbytes > 4092) | |
171 | mbytes = 4092; | |
172 | ||
173 | rdmsr(MSR_K6_WHCR, l, h); | |
174 | if ((l&0xFFFF0000) == 0) { | |
175 | unsigned long flags; | |
176 | l = ((mbytes>>2)<<22)|(1<<16); | |
177 | local_irq_save(flags); | |
178 | wbinvd(); | |
179 | wrmsr(MSR_K6_WHCR, l, h); | |
180 | local_irq_restore(flags); | |
181 | printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n", | |
182 | mbytes); | |
183 | } | |
184 | ||
185 | return; | |
186 | } | |
187 | ||
188 | if (c->x86_model == 10) { | |
189 | /* AMD Geode LX is model 10 */ | |
190 | /* placeholder for any needed mods */ | |
191 | return; | |
192 | } | |
26bfa5f8 | 193 | #endif |
11fdd252 YL |
194 | } |
195 | ||
26bfa5f8 | 196 | static void init_amd_k7(struct cpuinfo_x86 *c) |
1f442d70 | 197 | { |
26bfa5f8 BP |
198 | #ifdef CONFIG_X86_32 |
199 | u32 l, h; | |
200 | ||
201 | /* | |
202 | * Bit 15 of Athlon specific MSR 15, needs to be 0 | |
203 | * to enable SSE on Palomino/Morgan/Barton CPU's. | |
204 | * If the BIOS didn't enable it already, enable it here. | |
205 | */ | |
206 | if (c->x86_model >= 6 && c->x86_model <= 10) { | |
207 | if (!cpu_has(c, X86_FEATURE_XMM)) { | |
208 | printk(KERN_INFO "Enabling disabled K7/SSE Support.\n"); | |
209 | msr_clear_bit(MSR_K7_HWCR, 15); | |
210 | set_cpu_cap(c, X86_FEATURE_XMM); | |
211 | } | |
212 | } | |
213 | ||
214 | /* | |
215 | * It's been determined by AMD that Athlons since model 8 stepping 1 | |
216 | * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx | |
217 | * As per AMD technical note 27212 0.2 | |
218 | */ | |
219 | if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) { | |
220 | rdmsr(MSR_K7_CLK_CTL, l, h); | |
221 | if ((l & 0xfff00000) != 0x20000000) { | |
222 | printk(KERN_INFO | |
223 | "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n", | |
224 | l, ((l & 0x000fffff)|0x20000000)); | |
225 | wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h); | |
226 | } | |
227 | } | |
228 | ||
229 | set_cpu_cap(c, X86_FEATURE_K7); | |
230 | ||
1f442d70 | 231 | /* calling is from identify_secondary_cpu() ? */ |
f6e9456c | 232 | if (!c->cpu_index) |
1f442d70 YL |
233 | return; |
234 | ||
235 | /* | |
236 | * Certain Athlons might work (for various values of 'work') in SMP | |
237 | * but they are not certified as MP capable. | |
238 | */ | |
239 | /* Athlon 660/661 is valid. */ | |
240 | if ((c->x86_model == 6) && ((c->x86_mask == 0) || | |
241 | (c->x86_mask == 1))) | |
1077c932 | 242 | return; |
1f442d70 YL |
243 | |
244 | /* Duron 670 is valid */ | |
245 | if ((c->x86_model == 7) && (c->x86_mask == 0)) | |
1077c932 | 246 | return; |
1f442d70 YL |
247 | |
248 | /* | |
249 | * Athlon 662, Duron 671, and Athlon >model 7 have capability | |
250 | * bit. It's worth noting that the A5 stepping (662) of some | |
251 | * Athlon XP's have the MP bit set. | |
252 | * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for | |
253 | * more. | |
254 | */ | |
255 | if (((c->x86_model == 6) && (c->x86_mask >= 2)) || | |
256 | ((c->x86_model == 7) && (c->x86_mask >= 1)) || | |
257 | (c->x86_model > 7)) | |
26bfa5f8 | 258 | if (cpu_has(c, X86_FEATURE_MP)) |
1077c932 | 259 | return; |
1f442d70 YL |
260 | |
261 | /* If we get here, not a certified SMP capable AMD system. */ | |
262 | ||
263 | /* | |
264 | * Don't taint if we are running SMP kernel on a single non-MP | |
265 | * approved Athlon | |
266 | */ | |
267 | WARN_ONCE(1, "WARNING: This combination of AMD" | |
7da8b6dd | 268 | " processors is not suitable for SMP.\n"); |
8c90487c | 269 | add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE); |
6c62aa4a | 270 | #endif |
26bfa5f8 | 271 | } |
6c62aa4a | 272 | |
645a7919 | 273 | #ifdef CONFIG_NUMA |
bbc9e2f4 TH |
274 | /* |
275 | * To workaround broken NUMA config. Read the comment in | |
276 | * srat_detect_node(). | |
277 | */ | |
148f9bb8 | 278 | static int nearby_node(int apicid) |
6c62aa4a YL |
279 | { |
280 | int i, node; | |
281 | ||
282 | for (i = apicid - 1; i >= 0; i--) { | |
bbc9e2f4 | 283 | node = __apicid_to_node[i]; |
6c62aa4a YL |
284 | if (node != NUMA_NO_NODE && node_online(node)) |
285 | return node; | |
286 | } | |
287 | for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) { | |
bbc9e2f4 | 288 | node = __apicid_to_node[i]; |
6c62aa4a YL |
289 | if (node != NUMA_NO_NODE && node_online(node)) |
290 | return node; | |
291 | } | |
292 | return first_node(node_online_map); /* Shouldn't happen */ | |
293 | } | |
294 | #endif | |
11fdd252 | 295 | |
4a376ec3 | 296 | /* |
23588c38 AH |
297 | * Fixup core topology information for |
298 | * (1) AMD multi-node processors | |
299 | * Assumption: Number of cores in each internal node is the same. | |
6057b4d3 | 300 | * (2) AMD processors supporting compute units |
4a376ec3 | 301 | */ |
c8e56d20 | 302 | #ifdef CONFIG_SMP |
148f9bb8 | 303 | static void amd_get_topology(struct cpuinfo_x86 *c) |
4a376ec3 | 304 | { |
cc2749e4 | 305 | u32 cores_per_cu = 1; |
23588c38 | 306 | u8 node_id; |
4a376ec3 AH |
307 | int cpu = smp_processor_id(); |
308 | ||
23588c38 | 309 | /* get information required for multi-node processors */ |
362f924b | 310 | if (boot_cpu_has(X86_FEATURE_TOPOEXT)) { |
6057b4d3 AH |
311 | u32 eax, ebx, ecx, edx; |
312 | ||
313 | cpuid(0x8000001e, &eax, &ebx, &ecx, &edx); | |
cc2749e4 | 314 | nodes_per_socket = ((ecx >> 8) & 7) + 1; |
6057b4d3 AH |
315 | node_id = ecx & 7; |
316 | ||
317 | /* get compute unit information */ | |
318 | smp_num_siblings = ((ebx >> 8) & 3) + 1; | |
319 | c->compute_unit_id = ebx & 0xff; | |
9e81509e | 320 | cores_per_cu += ((ebx >> 8) & 3); |
23588c38 | 321 | } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) { |
6057b4d3 AH |
322 | u64 value; |
323 | ||
23588c38 | 324 | rdmsrl(MSR_FAM10H_NODE_ID, value); |
cc2749e4 | 325 | nodes_per_socket = ((value >> 3) & 7) + 1; |
23588c38 AH |
326 | node_id = value & 7; |
327 | } else | |
4a376ec3 AH |
328 | return; |
329 | ||
23588c38 | 330 | /* fixup multi-node processor information */ |
cc2749e4 | 331 | if (nodes_per_socket > 1) { |
6057b4d3 | 332 | u32 cores_per_node; |
d518573d | 333 | u32 cus_per_node; |
6057b4d3 | 334 | |
23588c38 | 335 | set_cpu_cap(c, X86_FEATURE_AMD_DCM); |
cc2749e4 | 336 | cores_per_node = c->x86_max_cores / nodes_per_socket; |
d518573d | 337 | cus_per_node = cores_per_node / cores_per_cu; |
9d260ebc | 338 | |
23588c38 AH |
339 | /* store NodeID, use llc_shared_map to store sibling info */ |
340 | per_cpu(cpu_llc_id, cpu) = node_id; | |
4a376ec3 | 341 | |
9e81509e | 342 | /* core id has to be in the [0 .. cores_per_node - 1] range */ |
d518573d AH |
343 | c->cpu_core_id %= cores_per_node; |
344 | c->compute_unit_id %= cus_per_node; | |
23588c38 | 345 | } |
4a376ec3 AH |
346 | } |
347 | #endif | |
348 | ||
11fdd252 | 349 | /* |
aa5e5dc2 | 350 | * On a AMD dual core setup the lower bits of the APIC id distinguish the cores. |
11fdd252 YL |
351 | * Assumes number of cores is a power of two. |
352 | */ | |
148f9bb8 | 353 | static void amd_detect_cmp(struct cpuinfo_x86 *c) |
11fdd252 | 354 | { |
c8e56d20 | 355 | #ifdef CONFIG_SMP |
11fdd252 | 356 | unsigned bits; |
99bd0c0f | 357 | int cpu = smp_processor_id(); |
3849e91f | 358 | unsigned int socket_id, core_complex_id; |
11fdd252 YL |
359 | |
360 | bits = c->x86_coreid_bits; | |
11fdd252 YL |
361 | /* Low order bits define the core id (index of core in socket) */ |
362 | c->cpu_core_id = c->initial_apicid & ((1 << bits)-1); | |
363 | /* Convert the initial APIC ID into the socket ID */ | |
364 | c->phys_proc_id = c->initial_apicid >> bits; | |
99bd0c0f AH |
365 | /* use socket ID also for last level cache */ |
366 | per_cpu(cpu_llc_id, cpu) = c->phys_proc_id; | |
23588c38 | 367 | amd_get_topology(c); |
3849e91f AG |
368 | |
369 | /* | |
370 | * Fix percpu cpu_llc_id here as LLC topology is different | |
371 | * for Fam17h systems. | |
372 | */ | |
373 | if (c->x86 != 0x17 || !cpuid_edx(0x80000006)) | |
374 | return; | |
375 | ||
376 | socket_id = (c->apicid >> bits) - 1; | |
377 | core_complex_id = (c->apicid & ((1 << bits) - 1)) >> 3; | |
378 | ||
379 | per_cpu(cpu_llc_id, cpu) = (socket_id << 3) | core_complex_id; | |
11fdd252 YL |
380 | #endif |
381 | } | |
382 | ||
8b84c8df | 383 | u16 amd_get_nb_id(int cpu) |
6a812691 | 384 | { |
8b84c8df | 385 | u16 id = 0; |
6a812691 AH |
386 | #ifdef CONFIG_SMP |
387 | id = per_cpu(cpu_llc_id, cpu); | |
388 | #endif | |
389 | return id; | |
390 | } | |
391 | EXPORT_SYMBOL_GPL(amd_get_nb_id); | |
392 | ||
cc2749e4 AG |
393 | u32 amd_get_nodes_per_socket(void) |
394 | { | |
395 | return nodes_per_socket; | |
396 | } | |
397 | EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket); | |
398 | ||
148f9bb8 | 399 | static void srat_detect_node(struct cpuinfo_x86 *c) |
6c62aa4a | 400 | { |
645a7919 | 401 | #ifdef CONFIG_NUMA |
6c62aa4a YL |
402 | int cpu = smp_processor_id(); |
403 | int node; | |
0d96b9ff | 404 | unsigned apicid = c->apicid; |
6c62aa4a | 405 | |
bbc9e2f4 TH |
406 | node = numa_cpu_node(cpu); |
407 | if (node == NUMA_NO_NODE) | |
408 | node = per_cpu(cpu_llc_id, cpu); | |
6c62aa4a | 409 | |
64be4c1c | 410 | /* |
68894632 AH |
411 | * On multi-fabric platform (e.g. Numascale NumaChip) a |
412 | * platform-specific handler needs to be called to fixup some | |
413 | * IDs of the CPU. | |
64be4c1c | 414 | */ |
68894632 | 415 | if (x86_cpuinit.fixup_cpu_id) |
64be4c1c DB |
416 | x86_cpuinit.fixup_cpu_id(c, node); |
417 | ||
6c62aa4a | 418 | if (!node_online(node)) { |
bbc9e2f4 TH |
419 | /* |
420 | * Two possibilities here: | |
421 | * | |
422 | * - The CPU is missing memory and no node was created. In | |
423 | * that case try picking one from a nearby CPU. | |
424 | * | |
425 | * - The APIC IDs differ from the HyperTransport node IDs | |
426 | * which the K8 northbridge parsing fills in. Assume | |
427 | * they are all increased by a constant offset, but in | |
428 | * the same order as the HT nodeids. If that doesn't | |
429 | * result in a usable node fall back to the path for the | |
430 | * previous case. | |
431 | * | |
432 | * This workaround operates directly on the mapping between | |
433 | * APIC ID and NUMA node, assuming certain relationship | |
434 | * between APIC ID, HT node ID and NUMA topology. As going | |
435 | * through CPU mapping may alter the outcome, directly | |
436 | * access __apicid_to_node[]. | |
437 | */ | |
6c62aa4a YL |
438 | int ht_nodeid = c->initial_apicid; |
439 | ||
7030a7e9 | 440 | if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE) |
bbc9e2f4 | 441 | node = __apicid_to_node[ht_nodeid]; |
6c62aa4a YL |
442 | /* Pick a nearby node */ |
443 | if (!node_online(node)) | |
444 | node = nearby_node(apicid); | |
445 | } | |
446 | numa_set_node(cpu, node); | |
6c62aa4a YL |
447 | #endif |
448 | } | |
449 | ||
148f9bb8 | 450 | static void early_init_amd_mc(struct cpuinfo_x86 *c) |
11fdd252 | 451 | { |
c8e56d20 | 452 | #ifdef CONFIG_SMP |
11fdd252 YL |
453 | unsigned bits, ecx; |
454 | ||
455 | /* Multi core CPU? */ | |
456 | if (c->extended_cpuid_level < 0x80000008) | |
457 | return; | |
458 | ||
459 | ecx = cpuid_ecx(0x80000008); | |
460 | ||
461 | c->x86_max_cores = (ecx & 0xff) + 1; | |
462 | ||
463 | /* CPU telling us the core id bits shift? */ | |
464 | bits = (ecx >> 12) & 0xF; | |
465 | ||
466 | /* Otherwise recompute */ | |
467 | if (bits == 0) { | |
468 | while ((1 << bits) < c->x86_max_cores) | |
469 | bits++; | |
470 | } | |
471 | ||
472 | c->x86_coreid_bits = bits; | |
473 | #endif | |
474 | } | |
475 | ||
148f9bb8 | 476 | static void bsp_init_amd(struct cpuinfo_x86 *c) |
8fa8b035 | 477 | { |
26bfa5f8 BP |
478 | |
479 | #ifdef CONFIG_X86_64 | |
480 | if (c->x86 >= 0xf) { | |
481 | unsigned long long tseg; | |
482 | ||
483 | /* | |
484 | * Split up direct mapping around the TSEG SMM area. | |
485 | * Don't do it for gbpages because there seems very little | |
486 | * benefit in doing so. | |
487 | */ | |
488 | if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) { | |
489 | unsigned long pfn = tseg >> PAGE_SHIFT; | |
490 | ||
491 | printk(KERN_DEBUG "tseg: %010llx\n", tseg); | |
492 | if (pfn_range_is_mapped(pfn, pfn + 1)) | |
493 | set_memory_4k((unsigned long)__va(tseg), 1); | |
494 | } | |
495 | } | |
496 | #endif | |
497 | ||
8fa8b035 BP |
498 | if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) { |
499 | ||
500 | if (c->x86 > 0x10 || | |
501 | (c->x86 == 0x10 && c->x86_model >= 0x2)) { | |
502 | u64 val; | |
503 | ||
504 | rdmsrl(MSR_K7_HWCR, val); | |
505 | if (!(val & BIT(24))) | |
506 | printk(KERN_WARNING FW_BUG "TSC doesn't count " | |
507 | "with P0 frequency!\n"); | |
508 | } | |
509 | } | |
510 | ||
511 | if (c->x86 == 0x15) { | |
512 | unsigned long upperbit; | |
513 | u32 cpuid, assoc; | |
514 | ||
515 | cpuid = cpuid_edx(0x80000005); | |
516 | assoc = cpuid >> 16 & 0xff; | |
517 | upperbit = ((cpuid >> 24) << 10) / assoc; | |
518 | ||
519 | va_align.mask = (upperbit - 1) & PAGE_MASK; | |
520 | va_align.flags = ALIGN_VA_32 | ALIGN_VA_64; | |
4e26d11f HMG |
521 | |
522 | /* A random value per boot for bit slice [12:upper_bit) */ | |
523 | va_align.bits = get_random_int() & va_align.mask; | |
8fa8b035 | 524 | } |
b466bdb6 HR |
525 | |
526 | if (cpu_has(c, X86_FEATURE_MWAITX)) | |
527 | use_mwaitx_delay(); | |
8fa8b035 BP |
528 | } |
529 | ||
148f9bb8 | 530 | static void early_init_amd(struct cpuinfo_x86 *c) |
2b16a235 | 531 | { |
11fdd252 YL |
532 | early_init_amd_mc(c); |
533 | ||
40fb1715 VP |
534 | /* |
535 | * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate | |
536 | * with P/T states and does not stop in deep C-states | |
537 | */ | |
538 | if (c->x86_power & (1 << 8)) { | |
e3224234 | 539 | set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); |
40fb1715 | 540 | set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); |
c98fdeaa | 541 | if (!check_tsc_unstable()) |
35af99e6 | 542 | set_sched_clock_stable(); |
40fb1715 | 543 | } |
5fef55fd | 544 | |
6c62aa4a YL |
545 | #ifdef CONFIG_X86_64 |
546 | set_cpu_cap(c, X86_FEATURE_SYSCALL32); | |
547 | #else | |
5fef55fd | 548 | /* Set MTRR capability flag if appropriate */ |
6c62aa4a YL |
549 | if (c->x86 == 5) |
550 | if (c->x86_model == 13 || c->x86_model == 9 || | |
551 | (c->x86_model == 8 && c->x86_mask >= 8)) | |
552 | set_cpu_cap(c, X86_FEATURE_K6_MTRR); | |
553 | #endif | |
42937e81 | 554 | #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI) |
b9d16a2a AG |
555 | /* |
556 | * ApicID can always be treated as an 8-bit value for AMD APIC versions | |
557 | * >= 0x10, but even old K8s came out of reset with version 0x10. So, we | |
558 | * can safely set X86_FEATURE_EXTD_APICID unconditionally for families | |
559 | * after 16h. | |
560 | */ | |
561 | if (cpu_has_apic && c->x86 > 0x16) { | |
562 | set_cpu_cap(c, X86_FEATURE_EXTD_APICID); | |
563 | } else if (cpu_has_apic && c->x86 >= 0xf) { | |
564 | /* check CPU config space for extended APIC ID */ | |
42937e81 AH |
565 | unsigned int val; |
566 | val = read_pci_config(0, 24, 0, 0x68); | |
567 | if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18))) | |
568 | set_cpu_cap(c, X86_FEATURE_EXTD_APICID); | |
569 | } | |
570 | #endif | |
3b564968 | 571 | |
c1118b36 PB |
572 | /* |
573 | * This is only needed to tell the kernel whether to use VMCALL | |
574 | * and VMMCALL. VMMCALL is never executed except under virt, so | |
575 | * we can set it unconditionally. | |
576 | */ | |
577 | set_cpu_cap(c, X86_FEATURE_VMMCALL); | |
578 | ||
3b564968 | 579 | /* F16h erratum 793, CVE-2013-6885 */ |
8f86a737 BP |
580 | if (c->x86 == 0x16 && c->x86_model <= 0xf) |
581 | msr_set_bit(MSR_AMD64_LS_CFG, 15); | |
2b16a235 AK |
582 | } |
583 | ||
e6ee94d5 | 584 | static const int amd_erratum_383[]; |
7d7dc116 | 585 | static const int amd_erratum_400[]; |
8c6b79bb | 586 | static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum); |
e6ee94d5 | 587 | |
26bfa5f8 BP |
588 | static void init_amd_k8(struct cpuinfo_x86 *c) |
589 | { | |
590 | u32 level; | |
591 | u64 value; | |
592 | ||
593 | /* On C+ stepping K8 rep microcode works well for copy/memset */ | |
594 | level = cpuid_eax(1); | |
595 | if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58) | |
596 | set_cpu_cap(c, X86_FEATURE_REP_GOOD); | |
597 | ||
598 | /* | |
599 | * Some BIOSes incorrectly force this feature, but only K8 revision D | |
600 | * (model = 0x14) and later actually support it. | |
601 | * (AMD Erratum #110, docId: 25759). | |
602 | */ | |
603 | if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) { | |
604 | clear_cpu_cap(c, X86_FEATURE_LAHF_LM); | |
605 | if (!rdmsrl_amd_safe(0xc001100d, &value)) { | |
606 | value &= ~BIT_64(32); | |
607 | wrmsrl_amd_safe(0xc001100d, value); | |
608 | } | |
609 | } | |
610 | ||
611 | if (!c->x86_model_id[0]) | |
612 | strcpy(c->x86_model_id, "Hammer"); | |
6f9b63a0 BP |
613 | |
614 | #ifdef CONFIG_SMP | |
615 | /* | |
616 | * Disable TLB flush filter by setting HWCR.FFDIS on K8 | |
617 | * bit 6 of msr C001_0015 | |
618 | * | |
619 | * Errata 63 for SH-B3 steppings | |
620 | * Errata 122 for all steppings (F+ have it disabled by default) | |
621 | */ | |
622 | msr_set_bit(MSR_K7_HWCR, 6); | |
623 | #endif | |
26bfa5f8 BP |
624 | } |
625 | ||
626 | static void init_amd_gh(struct cpuinfo_x86 *c) | |
627 | { | |
628 | #ifdef CONFIG_X86_64 | |
629 | /* do this for boot cpu */ | |
630 | if (c == &boot_cpu_data) | |
631 | check_enable_amd_mmconf_dmi(); | |
632 | ||
633 | fam10h_check_enable_mmcfg(); | |
634 | #endif | |
635 | ||
636 | /* | |
637 | * Disable GART TLB Walk Errors on Fam10h. We do this here because this | |
638 | * is always needed when GART is enabled, even in a kernel which has no | |
639 | * MCE support built in. BIOS should disable GartTlbWlk Errors already. | |
640 | * If it doesn't, we do it here as suggested by the BKDG. | |
641 | * | |
642 | * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012 | |
643 | */ | |
644 | msr_set_bit(MSR_AMD64_MCx_MASK(4), 10); | |
645 | ||
646 | /* | |
647 | * On family 10h BIOS may not have properly enabled WC+ support, causing | |
648 | * it to be converted to CD memtype. This may result in performance | |
649 | * degradation for certain nested-paging guests. Prevent this conversion | |
650 | * by clearing bit 24 in MSR_AMD64_BU_CFG2. | |
651 | * | |
652 | * NOTE: we want to use the _safe accessors so as not to #GP kvm | |
653 | * guests on older kvm hosts. | |
654 | */ | |
655 | msr_clear_bit(MSR_AMD64_BU_CFG2, 24); | |
656 | ||
657 | if (cpu_has_amd_erratum(c, amd_erratum_383)) | |
658 | set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH); | |
659 | } | |
660 | ||
661 | static void init_amd_bd(struct cpuinfo_x86 *c) | |
662 | { | |
663 | u64 value; | |
664 | ||
665 | /* re-enable TopologyExtensions if switched off by BIOS */ | |
666 | if ((c->x86_model >= 0x10) && (c->x86_model <= 0x1f) && | |
667 | !cpu_has(c, X86_FEATURE_TOPOEXT)) { | |
668 | ||
669 | if (msr_set_bit(0xc0011005, 54) > 0) { | |
670 | rdmsrl(0xc0011005, value); | |
671 | if (value & BIT_64(54)) { | |
672 | set_cpu_cap(c, X86_FEATURE_TOPOEXT); | |
673 | pr_info(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n"); | |
674 | } | |
675 | } | |
676 | } | |
677 | ||
678 | /* | |
679 | * The way access filter has a performance penalty on some workloads. | |
680 | * Disable it on the affected CPUs. | |
681 | */ | |
682 | if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) { | |
ae8b7875 | 683 | if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) { |
26bfa5f8 | 684 | value |= 0x1E; |
ae8b7875 | 685 | wrmsrl_safe(MSR_F15H_IC_CFG, value); |
26bfa5f8 BP |
686 | } |
687 | } | |
688 | } | |
689 | ||
148f9bb8 | 690 | static void init_amd(struct cpuinfo_x86 *c) |
1da177e4 | 691 | { |
8e8da023 | 692 | u32 dummy; |
7d318d77 | 693 | |
2b16a235 AK |
694 | early_init_amd(c); |
695 | ||
fb87a298 PC |
696 | /* |
697 | * Bit 31 in normal CPUID used for nonstandard 3DNow ID; | |
16282a8e | 698 | * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway |
fb87a298 | 699 | */ |
16282a8e | 700 | clear_cpu_cap(c, 0*32+31); |
fb87a298 | 701 | |
12d8a961 | 702 | if (c->x86 >= 0x10) |
6c62aa4a | 703 | set_cpu_cap(c, X86_FEATURE_REP_GOOD); |
0d96b9ff YL |
704 | |
705 | /* get apicid instead of initial apic id from cpuid */ | |
706 | c->apicid = hard_smp_processor_id(); | |
11fdd252 YL |
707 | |
708 | /* K6s reports MCEs but don't actually have all the MSRs */ | |
709 | if (c->x86 < 6) | |
710 | clear_cpu_cap(c, X86_FEATURE_MCE); | |
26bfa5f8 BP |
711 | |
712 | switch (c->x86) { | |
713 | case 4: init_amd_k5(c); break; | |
714 | case 5: init_amd_k6(c); break; | |
715 | case 6: init_amd_k7(c); break; | |
716 | case 0xf: init_amd_k8(c); break; | |
717 | case 0x10: init_amd_gh(c); break; | |
718 | case 0x15: init_amd_bd(c); break; | |
719 | } | |
11fdd252 | 720 | |
6c62aa4a | 721 | /* Enable workaround for FXSAVE leak */ |
18bd057b | 722 | if (c->x86 >= 6) |
9b13a93d | 723 | set_cpu_bug(c, X86_BUG_FXSAVE_LEAK); |
1da177e4 | 724 | |
27c13ece | 725 | cpu_detect_cache_sizes(c); |
3dd9d514 | 726 | |
11fdd252 | 727 | /* Multi core CPU? */ |
6c62aa4a | 728 | if (c->extended_cpuid_level >= 0x80000008) { |
11fdd252 | 729 | amd_detect_cmp(c); |
6c62aa4a YL |
730 | srat_detect_node(c); |
731 | } | |
faee9a5d | 732 | |
6c62aa4a | 733 | #ifdef CONFIG_X86_32 |
11fdd252 | 734 | detect_ht(c); |
6c62aa4a | 735 | #endif |
39b3a791 | 736 | |
04a15418 | 737 | init_amd_cacheinfo(c); |
3556ddfa | 738 | |
12d8a961 | 739 | if (c->x86 >= 0xf) |
11fdd252 | 740 | set_cpu_cap(c, X86_FEATURE_K8); |
de421863 | 741 | |
11fdd252 YL |
742 | if (cpu_has_xmm2) { |
743 | /* MFENCE stops RDTSC speculation */ | |
16282a8e | 744 | set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC); |
11fdd252 | 745 | } |
6c62aa4a | 746 | |
e9cdd343 BO |
747 | /* |
748 | * Family 0x12 and above processors have APIC timer | |
749 | * running in deep C states. | |
750 | */ | |
751 | if (c->x86 > 0x11) | |
b87cf80a | 752 | set_cpu_cap(c, X86_FEATURE_ARAT); |
5bbc097d | 753 | |
8c6b79bb | 754 | if (cpu_has_amd_erratum(c, amd_erratum_400)) |
7d7dc116 BP |
755 | set_cpu_bug(c, X86_BUG_AMD_APIC_C1E); |
756 | ||
8e8da023 | 757 | rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy); |
a930dc45 BP |
758 | |
759 | /* 3DNow or LM implies PREFETCHW */ | |
760 | if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH)) | |
761 | if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM)) | |
762 | set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH); | |
61f01dd9 AL |
763 | |
764 | /* AMD CPUs don't reset SS attributes on SYSRET */ | |
765 | set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS); | |
1da177e4 LT |
766 | } |
767 | ||
6c62aa4a | 768 | #ifdef CONFIG_X86_32 |
148f9bb8 | 769 | static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size) |
1da177e4 LT |
770 | { |
771 | /* AMD errata T13 (order #21922) */ | |
772 | if ((c->x86 == 6)) { | |
8bdbd962 AC |
773 | /* Duron Rev A0 */ |
774 | if (c->x86_model == 3 && c->x86_mask == 0) | |
1da177e4 | 775 | size = 64; |
8bdbd962 | 776 | /* Tbird rev A1/A2 */ |
1da177e4 | 777 | if (c->x86_model == 4 && |
8bdbd962 | 778 | (c->x86_mask == 0 || c->x86_mask == 1)) |
1da177e4 LT |
779 | size = 256; |
780 | } | |
781 | return size; | |
782 | } | |
6c62aa4a | 783 | #endif |
1da177e4 | 784 | |
148f9bb8 | 785 | static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c) |
b46882e4 BP |
786 | { |
787 | u32 ebx, eax, ecx, edx; | |
788 | u16 mask = 0xfff; | |
789 | ||
790 | if (c->x86 < 0xf) | |
791 | return; | |
792 | ||
793 | if (c->extended_cpuid_level < 0x80000006) | |
794 | return; | |
795 | ||
796 | cpuid(0x80000006, &eax, &ebx, &ecx, &edx); | |
797 | ||
798 | tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask; | |
799 | tlb_lli_4k[ENTRIES] = ebx & mask; | |
800 | ||
801 | /* | |
802 | * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB | |
803 | * characteristics from the CPUID function 0x80000005 instead. | |
804 | */ | |
805 | if (c->x86 == 0xf) { | |
806 | cpuid(0x80000005, &eax, &ebx, &ecx, &edx); | |
807 | mask = 0xff; | |
808 | } | |
809 | ||
810 | /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ | |
d1393367 BP |
811 | if (!((eax >> 16) & mask)) |
812 | tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff; | |
813 | else | |
b46882e4 | 814 | tlb_lld_2m[ENTRIES] = (eax >> 16) & mask; |
b46882e4 BP |
815 | |
816 | /* a 4M entry uses two 2M entries */ | |
817 | tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1; | |
818 | ||
819 | /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ | |
820 | if (!(eax & mask)) { | |
821 | /* Erratum 658 */ | |
822 | if (c->x86 == 0x15 && c->x86_model <= 0x1f) { | |
823 | tlb_lli_2m[ENTRIES] = 1024; | |
824 | } else { | |
825 | cpuid(0x80000005, &eax, &ebx, &ecx, &edx); | |
826 | tlb_lli_2m[ENTRIES] = eax & 0xff; | |
827 | } | |
828 | } else | |
829 | tlb_lli_2m[ENTRIES] = eax & mask; | |
830 | ||
831 | tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1; | |
832 | } | |
833 | ||
148f9bb8 | 834 | static const struct cpu_dev amd_cpu_dev = { |
1da177e4 | 835 | .c_vendor = "AMD", |
fb87a298 | 836 | .c_ident = { "AuthenticAMD" }, |
6c62aa4a | 837 | #ifdef CONFIG_X86_32 |
09dc68d9 JB |
838 | .legacy_models = { |
839 | { .family = 4, .model_names = | |
1da177e4 LT |
840 | { |
841 | [3] = "486 DX/2", | |
842 | [7] = "486 DX/2-WB", | |
fb87a298 PC |
843 | [8] = "486 DX/4", |
844 | [9] = "486 DX/4-WB", | |
1da177e4 | 845 | [14] = "Am5x86-WT", |
fb87a298 | 846 | [15] = "Am5x86-WB" |
1da177e4 LT |
847 | } |
848 | }, | |
849 | }, | |
09dc68d9 | 850 | .legacy_cache_size = amd_size_cache, |
6c62aa4a | 851 | #endif |
03ae5768 | 852 | .c_early_init = early_init_amd, |
b46882e4 | 853 | .c_detect_tlb = cpu_detect_tlb_amd, |
8fa8b035 | 854 | .c_bsp_init = bsp_init_amd, |
1da177e4 | 855 | .c_init = init_amd, |
10a434fc | 856 | .c_x86_vendor = X86_VENDOR_AMD, |
1da177e4 LT |
857 | }; |
858 | ||
10a434fc | 859 | cpu_dev_register(amd_cpu_dev); |
d78d671d HR |
860 | |
861 | /* | |
862 | * AMD errata checking | |
863 | * | |
864 | * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or | |
865 | * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that | |
866 | * have an OSVW id assigned, which it takes as first argument. Both take a | |
867 | * variable number of family-specific model-stepping ranges created by | |
7d7dc116 | 868 | * AMD_MODEL_RANGE(). |
d78d671d HR |
869 | * |
870 | * Example: | |
871 | * | |
872 | * const int amd_erratum_319[] = | |
873 | * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2), | |
874 | * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0), | |
875 | * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0)); | |
876 | */ | |
877 | ||
7d7dc116 BP |
878 | #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 } |
879 | #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 } | |
880 | #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \ | |
881 | ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end)) | |
882 | #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff) | |
883 | #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff) | |
884 | #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff) | |
885 | ||
886 | static const int amd_erratum_400[] = | |
328935e6 | 887 | AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf), |
9d8888c2 HR |
888 | AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf)); |
889 | ||
e6ee94d5 | 890 | static const int amd_erratum_383[] = |
1be85a6d | 891 | AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf)); |
9d8888c2 | 892 | |
8c6b79bb TK |
893 | |
894 | static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum) | |
d78d671d | 895 | { |
d78d671d HR |
896 | int osvw_id = *erratum++; |
897 | u32 range; | |
898 | u32 ms; | |
899 | ||
d78d671d HR |
900 | if (osvw_id >= 0 && osvw_id < 65536 && |
901 | cpu_has(cpu, X86_FEATURE_OSVW)) { | |
902 | u64 osvw_len; | |
903 | ||
904 | rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len); | |
905 | if (osvw_id < osvw_len) { | |
906 | u64 osvw_bits; | |
907 | ||
908 | rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6), | |
909 | osvw_bits); | |
910 | return osvw_bits & (1ULL << (osvw_id & 0x3f)); | |
911 | } | |
912 | } | |
913 | ||
914 | /* OSVW unavailable or ID unknown, match family-model-stepping range */ | |
07a7795c | 915 | ms = (cpu->x86_model << 4) | cpu->x86_mask; |
d78d671d HR |
916 | while ((range = *erratum++)) |
917 | if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) && | |
918 | (ms >= AMD_MODEL_RANGE_START(range)) && | |
919 | (ms <= AMD_MODEL_RANGE_END(range))) | |
920 | return true; | |
921 | ||
922 | return false; | |
923 | } | |
d6d55f0b JS |
924 | |
925 | void set_dr_addr_mask(unsigned long mask, int dr) | |
926 | { | |
362f924b | 927 | if (!boot_cpu_has(X86_FEATURE_BPEXT)) |
d6d55f0b JS |
928 | return; |
929 | ||
930 | switch (dr) { | |
931 | case 0: | |
932 | wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0); | |
933 | break; | |
934 | case 1: | |
935 | case 2: | |
936 | case 3: | |
937 | wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0); | |
938 | break; | |
939 | default: | |
940 | break; | |
941 | } | |
942 | } |