Merge 3.4-rc6 into usb-next
[deliverable/linux.git] / arch / x86 / kernel / cpu / amd.c
CommitLineData
69c60c88 1#include <linux/export.h>
1da177e4
LT
2#include <linux/init.h>
3#include <linux/bitops.h>
5cdd174f 4#include <linux/elf.h>
1da177e4 5#include <linux/mm.h>
8d71a2ea 6
8bdbd962 7#include <linux/io.h>
c98fdeaa 8#include <linux/sched.h>
1da177e4 9#include <asm/processor.h>
d3f7eae1 10#include <asm/apic.h>
1f442d70 11#include <asm/cpu.h>
42937e81 12#include <asm/pci-direct.h>
1da177e4 13
8d71a2ea
YL
14#ifdef CONFIG_X86_64
15# include <asm/numa_64.h>
16# include <asm/mmconfig.h>
17# include <asm/cacheflush.h>
18#endif
19
1da177e4
LT
20#include "cpu.h"
21
6c62aa4a 22#ifdef CONFIG_X86_32
1da177e4
LT
23/*
24 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
25 * misexecution of code under Linux. Owners of such processors should
26 * contact AMD for precise details and a CPU swap.
27 *
28 * See http://www.multimania.com/poulot/k6bug.html
d7de8649
AH
29 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
30 * (Publication # 21266 Issue Date: August 1998)
1da177e4
LT
31 *
32 * The following test is erm.. interesting. AMD neglected to up
33 * the chip setting when fixing the bug but they also tweaked some
34 * performance at the same time..
35 */
fb87a298 36
1da177e4
LT
37extern void vide(void);
38__asm__(".align 4\nvide: ret");
39
11fdd252
YL
40static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
41{
42/*
43 * General Systems BIOSen alias the cpu frequency registers
44 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
45 * drivers subsequently pokes it, and changes the CPU speed.
46 * Workaround : Remove the unneeded alias.
47 */
48#define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
49#define CBAR_ENB (0x80000000)
50#define CBAR_KEY (0X000000CB)
51 if (c->x86_model == 9 || c->x86_model == 10) {
8bdbd962
AC
52 if (inl(CBAR) & CBAR_ENB)
53 outl(0 | CBAR_KEY, CBAR);
11fdd252
YL
54 }
55}
56
57
58static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
59{
60 u32 l, h;
61 int mbytes = num_physpages >> (20-PAGE_SHIFT);
62
63 if (c->x86_model < 6) {
64 /* Based on AMD doc 20734R - June 2000 */
65 if (c->x86_model == 0) {
66 clear_cpu_cap(c, X86_FEATURE_APIC);
67 set_cpu_cap(c, X86_FEATURE_PGE);
68 }
69 return;
70 }
71
72 if (c->x86_model == 6 && c->x86_mask == 1) {
73 const int K6_BUG_LOOP = 1000000;
74 int n;
75 void (*f_vide)(void);
76 unsigned long d, d2;
77
78 printk(KERN_INFO "AMD K6 stepping B detected - ");
79
80 /*
81 * It looks like AMD fixed the 2.6.2 bug and improved indirect
82 * calls at the same time.
83 */
84
85 n = K6_BUG_LOOP;
86 f_vide = vide;
87 rdtscl(d);
88 while (n--)
89 f_vide();
90 rdtscl(d2);
91 d = d2-d;
92
93 if (d > 20*K6_BUG_LOOP)
8bdbd962
AC
94 printk(KERN_CONT
95 "system stability may be impaired when more than 32 MB are used.\n");
11fdd252 96 else
8bdbd962 97 printk(KERN_CONT "probably OK (after B9730xxxx).\n");
11fdd252
YL
98 }
99
100 /* K6 with old style WHCR */
101 if (c->x86_model < 8 ||
102 (c->x86_model == 8 && c->x86_mask < 8)) {
103 /* We can only write allocate on the low 508Mb */
104 if (mbytes > 508)
105 mbytes = 508;
106
107 rdmsr(MSR_K6_WHCR, l, h);
108 if ((l&0x0000FFFF) == 0) {
109 unsigned long flags;
110 l = (1<<0)|((mbytes/4)<<1);
111 local_irq_save(flags);
112 wbinvd();
113 wrmsr(MSR_K6_WHCR, l, h);
114 local_irq_restore(flags);
115 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
116 mbytes);
117 }
118 return;
119 }
120
121 if ((c->x86_model == 8 && c->x86_mask > 7) ||
122 c->x86_model == 9 || c->x86_model == 13) {
123 /* The more serious chips .. */
124
125 if (mbytes > 4092)
126 mbytes = 4092;
127
128 rdmsr(MSR_K6_WHCR, l, h);
129 if ((l&0xFFFF0000) == 0) {
130 unsigned long flags;
131 l = ((mbytes>>2)<<22)|(1<<16);
132 local_irq_save(flags);
133 wbinvd();
134 wrmsr(MSR_K6_WHCR, l, h);
135 local_irq_restore(flags);
136 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
137 mbytes);
138 }
139
140 return;
141 }
142
143 if (c->x86_model == 10) {
144 /* AMD Geode LX is model 10 */
145 /* placeholder for any needed mods */
146 return;
147 }
148}
149
1f442d70
YL
150static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
151{
1f442d70 152 /* calling is from identify_secondary_cpu() ? */
f6e9456c 153 if (!c->cpu_index)
1f442d70
YL
154 return;
155
156 /*
157 * Certain Athlons might work (for various values of 'work') in SMP
158 * but they are not certified as MP capable.
159 */
160 /* Athlon 660/661 is valid. */
161 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
162 (c->x86_mask == 1)))
163 goto valid_k7;
164
165 /* Duron 670 is valid */
166 if ((c->x86_model == 7) && (c->x86_mask == 0))
167 goto valid_k7;
168
169 /*
170 * Athlon 662, Duron 671, and Athlon >model 7 have capability
171 * bit. It's worth noting that the A5 stepping (662) of some
172 * Athlon XP's have the MP bit set.
173 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
174 * more.
175 */
176 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
177 ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
178 (c->x86_model > 7))
179 if (cpu_has_mp)
180 goto valid_k7;
181
182 /* If we get here, not a certified SMP capable AMD system. */
183
184 /*
185 * Don't taint if we are running SMP kernel on a single non-MP
186 * approved Athlon
187 */
188 WARN_ONCE(1, "WARNING: This combination of AMD"
7da8b6dd 189 " processors is not suitable for SMP.\n");
1f442d70
YL
190 if (!test_taint(TAINT_UNSAFE_SMP))
191 add_taint(TAINT_UNSAFE_SMP);
192
193valid_k7:
194 ;
1f442d70
YL
195}
196
11fdd252
YL
197static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
198{
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 rdmsr(MSR_K7_HWCR, l, h);
210 l &= ~0x00008000;
211 wrmsr(MSR_K7_HWCR, l, h);
212 set_cpu_cap(c, X86_FEATURE_XMM);
213 }
214 }
215
216 /*
217 * It's been determined by AMD that Athlons since model 8 stepping 1
218 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
219 * As per AMD technical note 27212 0.2
220 */
221 if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
222 rdmsr(MSR_K7_CLK_CTL, l, h);
223 if ((l & 0xfff00000) != 0x20000000) {
8bdbd962
AC
224 printk(KERN_INFO
225 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
226 l, ((l & 0x000fffff)|0x20000000));
11fdd252
YL
227 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
228 }
229 }
230
231 set_cpu_cap(c, X86_FEATURE_K7);
1f442d70
YL
232
233 amd_k7_smp_check(c);
11fdd252 234}
6c62aa4a
YL
235#endif
236
645a7919 237#ifdef CONFIG_NUMA
bbc9e2f4
TH
238/*
239 * To workaround broken NUMA config. Read the comment in
240 * srat_detect_node().
241 */
6c62aa4a
YL
242static int __cpuinit nearby_node(int apicid)
243{
244 int i, node;
245
246 for (i = apicid - 1; i >= 0; i--) {
bbc9e2f4 247 node = __apicid_to_node[i];
6c62aa4a
YL
248 if (node != NUMA_NO_NODE && node_online(node))
249 return node;
250 }
251 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
bbc9e2f4 252 node = __apicid_to_node[i];
6c62aa4a
YL
253 if (node != NUMA_NO_NODE && node_online(node))
254 return node;
255 }
256 return first_node(node_online_map); /* Shouldn't happen */
257}
258#endif
11fdd252 259
4a376ec3 260/*
23588c38
AH
261 * Fixup core topology information for
262 * (1) AMD multi-node processors
263 * Assumption: Number of cores in each internal node is the same.
6057b4d3 264 * (2) AMD processors supporting compute units
4a376ec3
AH
265 */
266#ifdef CONFIG_X86_HT
23588c38 267static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
4a376ec3 268{
9e81509e 269 u32 nodes, cores_per_cu = 1;
23588c38 270 u8 node_id;
4a376ec3
AH
271 int cpu = smp_processor_id();
272
23588c38
AH
273 /* get information required for multi-node processors */
274 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
6057b4d3
AH
275 u32 eax, ebx, ecx, edx;
276
277 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
278 nodes = ((ecx >> 8) & 7) + 1;
279 node_id = ecx & 7;
280
281 /* get compute unit information */
282 smp_num_siblings = ((ebx >> 8) & 3) + 1;
283 c->compute_unit_id = ebx & 0xff;
9e81509e 284 cores_per_cu += ((ebx >> 8) & 3);
23588c38 285 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
6057b4d3
AH
286 u64 value;
287
23588c38
AH
288 rdmsrl(MSR_FAM10H_NODE_ID, value);
289 nodes = ((value >> 3) & 7) + 1;
290 node_id = value & 7;
291 } else
4a376ec3
AH
292 return;
293
23588c38
AH
294 /* fixup multi-node processor information */
295 if (nodes > 1) {
6057b4d3 296 u32 cores_per_node;
d518573d 297 u32 cus_per_node;
6057b4d3 298
23588c38
AH
299 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
300 cores_per_node = c->x86_max_cores / nodes;
d518573d 301 cus_per_node = cores_per_node / cores_per_cu;
9d260ebc 302
23588c38
AH
303 /* store NodeID, use llc_shared_map to store sibling info */
304 per_cpu(cpu_llc_id, cpu) = node_id;
4a376ec3 305
9e81509e 306 /* core id has to be in the [0 .. cores_per_node - 1] range */
d518573d
AH
307 c->cpu_core_id %= cores_per_node;
308 c->compute_unit_id %= cus_per_node;
23588c38 309 }
4a376ec3
AH
310}
311#endif
312
11fdd252
YL
313/*
314 * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
315 * Assumes number of cores is a power of two.
316 */
317static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
318{
319#ifdef CONFIG_X86_HT
320 unsigned bits;
99bd0c0f 321 int cpu = smp_processor_id();
11fdd252
YL
322
323 bits = c->x86_coreid_bits;
11fdd252
YL
324 /* Low order bits define the core id (index of core in socket) */
325 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
326 /* Convert the initial APIC ID into the socket ID */
327 c->phys_proc_id = c->initial_apicid >> bits;
99bd0c0f
AH
328 /* use socket ID also for last level cache */
329 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
23588c38 330 amd_get_topology(c);
11fdd252
YL
331#endif
332}
333
6a812691
AH
334int amd_get_nb_id(int cpu)
335{
336 int id = 0;
337#ifdef CONFIG_SMP
338 id = per_cpu(cpu_llc_id, cpu);
339#endif
340 return id;
341}
342EXPORT_SYMBOL_GPL(amd_get_nb_id);
343
6c62aa4a
YL
344static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
345{
645a7919 346#ifdef CONFIG_NUMA
6c62aa4a
YL
347 int cpu = smp_processor_id();
348 int node;
0d96b9ff 349 unsigned apicid = c->apicid;
6c62aa4a 350
bbc9e2f4
TH
351 node = numa_cpu_node(cpu);
352 if (node == NUMA_NO_NODE)
353 node = per_cpu(cpu_llc_id, cpu);
6c62aa4a 354
64be4c1c 355 /*
68894632
AH
356 * On multi-fabric platform (e.g. Numascale NumaChip) a
357 * platform-specific handler needs to be called to fixup some
358 * IDs of the CPU.
64be4c1c 359 */
68894632 360 if (x86_cpuinit.fixup_cpu_id)
64be4c1c
DB
361 x86_cpuinit.fixup_cpu_id(c, node);
362
6c62aa4a 363 if (!node_online(node)) {
bbc9e2f4
TH
364 /*
365 * Two possibilities here:
366 *
367 * - The CPU is missing memory and no node was created. In
368 * that case try picking one from a nearby CPU.
369 *
370 * - The APIC IDs differ from the HyperTransport node IDs
371 * which the K8 northbridge parsing fills in. Assume
372 * they are all increased by a constant offset, but in
373 * the same order as the HT nodeids. If that doesn't
374 * result in a usable node fall back to the path for the
375 * previous case.
376 *
377 * This workaround operates directly on the mapping between
378 * APIC ID and NUMA node, assuming certain relationship
379 * between APIC ID, HT node ID and NUMA topology. As going
380 * through CPU mapping may alter the outcome, directly
381 * access __apicid_to_node[].
382 */
6c62aa4a
YL
383 int ht_nodeid = c->initial_apicid;
384
385 if (ht_nodeid >= 0 &&
bbc9e2f4
TH
386 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
387 node = __apicid_to_node[ht_nodeid];
6c62aa4a
YL
388 /* Pick a nearby node */
389 if (!node_online(node))
390 node = nearby_node(apicid);
391 }
392 numa_set_node(cpu, node);
6c62aa4a
YL
393#endif
394}
395
11fdd252
YL
396static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
397{
398#ifdef CONFIG_X86_HT
399 unsigned bits, ecx;
400
401 /* Multi core CPU? */
402 if (c->extended_cpuid_level < 0x80000008)
403 return;
404
405 ecx = cpuid_ecx(0x80000008);
406
407 c->x86_max_cores = (ecx & 0xff) + 1;
408
409 /* CPU telling us the core id bits shift? */
410 bits = (ecx >> 12) & 0xF;
411
412 /* Otherwise recompute */
413 if (bits == 0) {
414 while ((1 << bits) < c->x86_max_cores)
415 bits++;
416 }
417
418 c->x86_coreid_bits = bits;
419#endif
420}
421
8fa8b035
BP
422static void __cpuinit bsp_init_amd(struct cpuinfo_x86 *c)
423{
424 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
425
426 if (c->x86 > 0x10 ||
427 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
428 u64 val;
429
430 rdmsrl(MSR_K7_HWCR, val);
431 if (!(val & BIT(24)))
432 printk(KERN_WARNING FW_BUG "TSC doesn't count "
433 "with P0 frequency!\n");
434 }
435 }
436
437 if (c->x86 == 0x15) {
438 unsigned long upperbit;
439 u32 cpuid, assoc;
440
441 cpuid = cpuid_edx(0x80000005);
442 assoc = cpuid >> 16 & 0xff;
443 upperbit = ((cpuid >> 24) << 10) / assoc;
444
445 va_align.mask = (upperbit - 1) & PAGE_MASK;
446 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
447 }
448}
449
03ae5768 450static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
2b16a235 451{
11fdd252
YL
452 early_init_amd_mc(c);
453
40fb1715
VP
454 /*
455 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
456 * with P/T states and does not stop in deep C-states
457 */
458 if (c->x86_power & (1 << 8)) {
e3224234 459 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
40fb1715 460 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
c98fdeaa
BP
461 if (!check_tsc_unstable())
462 sched_clock_stable = 1;
40fb1715 463 }
5fef55fd 464
6c62aa4a
YL
465#ifdef CONFIG_X86_64
466 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
467#else
5fef55fd 468 /* Set MTRR capability flag if appropriate */
6c62aa4a
YL
469 if (c->x86 == 5)
470 if (c->x86_model == 13 || c->x86_model == 9 ||
471 (c->x86_model == 8 && c->x86_mask >= 8))
472 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
473#endif
42937e81
AH
474#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
475 /* check CPU config space for extended APIC ID */
2cb07860 476 if (cpu_has_apic && c->x86 >= 0xf) {
42937e81
AH
477 unsigned int val;
478 val = read_pci_config(0, 24, 0, 0x68);
479 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
480 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
481 }
482#endif
2b16a235
AK
483}
484
b4af3f7c 485static void __cpuinit init_amd(struct cpuinfo_x86 *c)
1da177e4 486{
8e8da023
LT
487 u32 dummy;
488
7d318d77 489#ifdef CONFIG_SMP
3c92c2ba 490 unsigned long long value;
7d318d77 491
fb87a298
PC
492 /*
493 * Disable TLB flush filter by setting HWCR.FFDIS on K8
7d318d77
AK
494 * bit 6 of msr C001_0015
495 *
496 * Errata 63 for SH-B3 steppings
497 * Errata 122 for all steppings (F+ have it disabled by default)
498 */
11fdd252 499 if (c->x86 == 0xf) {
7d318d77
AK
500 rdmsrl(MSR_K7_HWCR, value);
501 value |= 1 << 6;
502 wrmsrl(MSR_K7_HWCR, value);
503 }
504#endif
505
2b16a235
AK
506 early_init_amd(c);
507
fb87a298
PC
508 /*
509 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
16282a8e 510 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
fb87a298 511 */
16282a8e 512 clear_cpu_cap(c, 0*32+31);
fb87a298 513
6c62aa4a
YL
514#ifdef CONFIG_X86_64
515 /* On C+ stepping K8 rep microcode works well for copy/memset */
516 if (c->x86 == 0xf) {
517 u32 level;
518
519 level = cpuid_eax(1);
8bdbd962 520 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
6c62aa4a 521 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
fbd8b181
KW
522
523 /*
524 * Some BIOSes incorrectly force this feature, but only K8
525 * revision D (model = 0x14) and later actually support it.
6b0f43dd 526 * (AMD Erratum #110, docId: 25759).
fbd8b181 527 */
6b0f43dd
BP
528 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
529 u64 val;
530
fbd8b181 531 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
6b0f43dd
BP
532 if (!rdmsrl_amd_safe(0xc001100d, &val)) {
533 val &= ~(1ULL << 32);
534 wrmsrl_amd_safe(0xc001100d, val);
535 }
536 }
537
6c62aa4a 538 }
12d8a961 539 if (c->x86 >= 0x10)
6c62aa4a 540 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
0d96b9ff
YL
541
542 /* get apicid instead of initial apic id from cpuid */
543 c->apicid = hard_smp_processor_id();
6c62aa4a
YL
544#else
545
546 /*
547 * FIXME: We should handle the K5 here. Set up the write
548 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
549 * no bus pipeline)
550 */
551
fb87a298
PC
552 switch (c->x86) {
553 case 4:
11fdd252
YL
554 init_amd_k5(c);
555 break;
fb87a298 556 case 5:
11fdd252 557 init_amd_k6(c);
1da177e4 558 break;
11fdd252
YL
559 case 6: /* An Athlon/Duron */
560 init_amd_k7(c);
1da177e4
LT
561 break;
562 }
11fdd252
YL
563
564 /* K6s reports MCEs but don't actually have all the MSRs */
565 if (c->x86 < 6)
566 clear_cpu_cap(c, X86_FEATURE_MCE);
6c62aa4a 567#endif
11fdd252 568
6c62aa4a 569 /* Enable workaround for FXSAVE leak */
18bd057b 570 if (c->x86 >= 6)
16282a8e 571 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
1da177e4 572
11fdd252
YL
573 if (!c->x86_model_id[0]) {
574 switch (c->x86) {
575 case 0xf:
576 /* Should distinguish Models here, but this is only
577 a fallback anyways. */
578 strcpy(c->x86_model_id, "Hammer");
579 break;
580 }
581 }
3dd9d514 582
f7f286a9
AH
583 /* re-enable TopologyExtensions if switched off by BIOS */
584 if ((c->x86 == 0x15) &&
585 (c->x86_model >= 0x10) && (c->x86_model <= 0x1f) &&
586 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
587 u64 val;
588
589 if (!rdmsrl_amd_safe(0xc0011005, &val)) {
590 val |= 1ULL << 54;
591 wrmsrl_amd_safe(0xc0011005, val);
592 rdmsrl(0xc0011005, val);
593 if (val & (1ULL << 54)) {
594 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
595 printk(KERN_INFO FW_INFO "CPU: Re-enabling "
596 "disabled Topology Extensions Support\n");
597 }
598 }
599 }
600
27c13ece 601 cpu_detect_cache_sizes(c);
3dd9d514 602
11fdd252 603 /* Multi core CPU? */
6c62aa4a 604 if (c->extended_cpuid_level >= 0x80000008) {
11fdd252 605 amd_detect_cmp(c);
6c62aa4a
YL
606 srat_detect_node(c);
607 }
faee9a5d 608
6c62aa4a 609#ifdef CONFIG_X86_32
11fdd252 610 detect_ht(c);
6c62aa4a 611#endif
39b3a791 612
11fdd252 613 if (c->extended_cpuid_level >= 0x80000006) {
d9fadd7b 614 if (cpuid_edx(0x80000006) & 0xf000)
67cddd94
AK
615 num_cache_leaves = 4;
616 else
617 num_cache_leaves = 3;
618 }
3556ddfa 619
12d8a961 620 if (c->x86 >= 0xf)
11fdd252 621 set_cpu_cap(c, X86_FEATURE_K8);
de421863 622
11fdd252
YL
623 if (cpu_has_xmm2) {
624 /* MFENCE stops RDTSC speculation */
16282a8e 625 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
11fdd252 626 }
6c62aa4a
YL
627
628#ifdef CONFIG_X86_64
629 if (c->x86 == 0x10) {
630 /* do this for boot cpu */
631 if (c == &boot_cpu_data)
632 check_enable_amd_mmconf_dmi();
633
634 fam10h_check_enable_mmcfg();
635 }
636
12d8a961 637 if (c == &boot_cpu_data && c->x86 >= 0xf) {
6c62aa4a
YL
638 unsigned long long tseg;
639
640 /*
641 * Split up direct mapping around the TSEG SMM area.
642 * Don't do it for gbpages because there seems very little
643 * benefit in doing so.
644 */
645 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
8bdbd962
AC
646 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
647 if ((tseg>>PMD_SHIFT) <
6c62aa4a 648 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
8bdbd962 649 ((tseg>>PMD_SHIFT) <
6c62aa4a 650 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
8bdbd962
AC
651 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
652 set_memory_4k((unsigned long)__va(tseg), 1);
6c62aa4a
YL
653 }
654 }
655#endif
b87cf80a 656
e9cdd343
BO
657 /*
658 * Family 0x12 and above processors have APIC timer
659 * running in deep C states.
660 */
661 if (c->x86 > 0x11)
b87cf80a 662 set_cpu_cap(c, X86_FEATURE_ARAT);
5bbc097d
JR
663
664 /*
665 * Disable GART TLB Walk Errors on Fam10h. We do this here
666 * because this is always needed when GART is enabled, even in a
667 * kernel which has no MCE support built in.
668 */
669 if (c->x86 == 0x10) {
670 /*
671 * BIOS should disable GartTlbWlk Errors themself. If
672 * it doesn't do it here as suggested by the BKDG.
673 *
674 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
675 */
676 u64 mask;
d47cc0db 677 int err;
5bbc097d 678
d47cc0db
RJ
679 err = rdmsrl_safe(MSR_AMD64_MCx_MASK(4), &mask);
680 if (err == 0) {
681 mask |= (1 << 10);
682 checking_wrmsrl(MSR_AMD64_MCx_MASK(4), mask);
683 }
5bbc097d 684 }
8e8da023
LT
685
686 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
1da177e4
LT
687}
688
6c62aa4a 689#ifdef CONFIG_X86_32
8bdbd962
AC
690static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
691 unsigned int size)
1da177e4
LT
692{
693 /* AMD errata T13 (order #21922) */
694 if ((c->x86 == 6)) {
8bdbd962
AC
695 /* Duron Rev A0 */
696 if (c->x86_model == 3 && c->x86_mask == 0)
1da177e4 697 size = 64;
8bdbd962 698 /* Tbird rev A1/A2 */
1da177e4 699 if (c->x86_model == 4 &&
8bdbd962 700 (c->x86_mask == 0 || c->x86_mask == 1))
1da177e4
LT
701 size = 256;
702 }
703 return size;
704}
6c62aa4a 705#endif
1da177e4 706
02dde8b4 707static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
1da177e4 708 .c_vendor = "AMD",
fb87a298 709 .c_ident = { "AuthenticAMD" },
6c62aa4a 710#ifdef CONFIG_X86_32
1da177e4
LT
711 .c_models = {
712 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
713 {
714 [3] = "486 DX/2",
715 [7] = "486 DX/2-WB",
fb87a298
PC
716 [8] = "486 DX/4",
717 [9] = "486 DX/4-WB",
1da177e4 718 [14] = "Am5x86-WT",
fb87a298 719 [15] = "Am5x86-WB"
1da177e4
LT
720 }
721 },
722 },
6c62aa4a
YL
723 .c_size_cache = amd_size_cache,
724#endif
03ae5768 725 .c_early_init = early_init_amd,
8fa8b035 726 .c_bsp_init = bsp_init_amd,
1da177e4 727 .c_init = init_amd,
10a434fc 728 .c_x86_vendor = X86_VENDOR_AMD,
1da177e4
LT
729};
730
10a434fc 731cpu_dev_register(amd_cpu_dev);
d78d671d
HR
732
733/*
734 * AMD errata checking
735 *
736 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
737 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
738 * have an OSVW id assigned, which it takes as first argument. Both take a
739 * variable number of family-specific model-stepping ranges created by
740 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
741 * int[] in arch/x86/include/asm/processor.h.
742 *
743 * Example:
744 *
745 * const int amd_erratum_319[] =
746 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
747 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
748 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
749 */
750
9d8888c2 751const int amd_erratum_400[] =
328935e6 752 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
9d8888c2 753 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
a5b91606 754EXPORT_SYMBOL_GPL(amd_erratum_400);
9d8888c2 755
1be85a6d
HR
756const int amd_erratum_383[] =
757 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
a5b91606 758EXPORT_SYMBOL_GPL(amd_erratum_383);
9d8888c2 759
d78d671d
HR
760bool cpu_has_amd_erratum(const int *erratum)
761{
7b543a53 762 struct cpuinfo_x86 *cpu = __this_cpu_ptr(&cpu_info);
d78d671d
HR
763 int osvw_id = *erratum++;
764 u32 range;
765 u32 ms;
766
767 /*
768 * If called early enough that current_cpu_data hasn't been initialized
769 * yet, fall back to boot_cpu_data.
770 */
771 if (cpu->x86 == 0)
772 cpu = &boot_cpu_data;
773
774 if (cpu->x86_vendor != X86_VENDOR_AMD)
775 return false;
776
777 if (osvw_id >= 0 && osvw_id < 65536 &&
778 cpu_has(cpu, X86_FEATURE_OSVW)) {
779 u64 osvw_len;
780
781 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
782 if (osvw_id < osvw_len) {
783 u64 osvw_bits;
784
785 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
786 osvw_bits);
787 return osvw_bits & (1ULL << (osvw_id & 0x3f));
788 }
789 }
790
791 /* OSVW unavailable or ID unknown, match family-model-stepping range */
07a7795c 792 ms = (cpu->x86_model << 4) | cpu->x86_mask;
d78d671d
HR
793 while ((range = *erratum++))
794 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
795 (ms >= AMD_MODEL_RANGE_START(range)) &&
796 (ms <= AMD_MODEL_RANGE_END(range)))
797 return true;
798
799 return false;
800}
a5b91606
PA
801
802EXPORT_SYMBOL_GPL(cpu_has_amd_erratum);
This page took 2.778922 seconds and 5 git commands to generate.