Loading...
1#include <linux/init.h>
2#include <linux/bitops.h>
3#include <linux/mm.h>
4
5#include <linux/io.h>
6#include <asm/processor.h>
7#include <asm/apic.h>
8#include <asm/cpu.h>
9#include <asm/pci-direct.h>
10
11#ifdef CONFIG_X86_64
12# include <asm/numa_64.h>
13# include <asm/mmconfig.h>
14# include <asm/cacheflush.h>
15#endif
16
17#include "cpu.h"
18
19#ifdef CONFIG_X86_32
20/*
21 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
22 * misexecution of code under Linux. Owners of such processors should
23 * contact AMD for precise details and a CPU swap.
24 *
25 * See http://www.multimania.com/poulot/k6bug.html
26 * http://www.amd.com/K6/k6docs/revgd.html
27 *
28 * The following test is erm.. interesting. AMD neglected to up
29 * the chip setting when fixing the bug but they also tweaked some
30 * performance at the same time..
31 */
32
33extern void vide(void);
34__asm__(".align 4\nvide: ret");
35
36static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
37{
38/*
39 * General Systems BIOSen alias the cpu frequency registers
40 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
41 * drivers subsequently pokes it, and changes the CPU speed.
42 * Workaround : Remove the unneeded alias.
43 */
44#define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
45#define CBAR_ENB (0x80000000)
46#define CBAR_KEY (0X000000CB)
47 if (c->x86_model == 9 || c->x86_model == 10) {
48 if (inl(CBAR) & CBAR_ENB)
49 outl(0 | CBAR_KEY, CBAR);
50 }
51}
52
53
54static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
55{
56 u32 l, h;
57 int mbytes = num_physpages >> (20-PAGE_SHIFT);
58
59 if (c->x86_model < 6) {
60 /* Based on AMD doc 20734R - June 2000 */
61 if (c->x86_model == 0) {
62 clear_cpu_cap(c, X86_FEATURE_APIC);
63 set_cpu_cap(c, X86_FEATURE_PGE);
64 }
65 return;
66 }
67
68 if (c->x86_model == 6 && c->x86_mask == 1) {
69 const int K6_BUG_LOOP = 1000000;
70 int n;
71 void (*f_vide)(void);
72 unsigned long d, d2;
73
74 printk(KERN_INFO "AMD K6 stepping B detected - ");
75
76 /*
77 * It looks like AMD fixed the 2.6.2 bug and improved indirect
78 * calls at the same time.
79 */
80
81 n = K6_BUG_LOOP;
82 f_vide = vide;
83 rdtscl(d);
84 while (n--)
85 f_vide();
86 rdtscl(d2);
87 d = d2-d;
88
89 if (d > 20*K6_BUG_LOOP)
90 printk(KERN_CONT
91 "system stability may be impaired when more than 32 MB are used.\n");
92 else
93 printk(KERN_CONT "probably OK (after B9730xxxx).\n");
94 printk(KERN_INFO "Please see http://membres.lycos.fr/poulot/k6bug.html\n");
95 }
96
97 /* K6 with old style WHCR */
98 if (c->x86_model < 8 ||
99 (c->x86_model == 8 && c->x86_mask < 8)) {
100 /* We can only write allocate on the low 508Mb */
101 if (mbytes > 508)
102 mbytes = 508;
103
104 rdmsr(MSR_K6_WHCR, l, h);
105 if ((l&0x0000FFFF) == 0) {
106 unsigned long flags;
107 l = (1<<0)|((mbytes/4)<<1);
108 local_irq_save(flags);
109 wbinvd();
110 wrmsr(MSR_K6_WHCR, l, h);
111 local_irq_restore(flags);
112 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
113 mbytes);
114 }
115 return;
116 }
117
118 if ((c->x86_model == 8 && c->x86_mask > 7) ||
119 c->x86_model == 9 || c->x86_model == 13) {
120 /* The more serious chips .. */
121
122 if (mbytes > 4092)
123 mbytes = 4092;
124
125 rdmsr(MSR_K6_WHCR, l, h);
126 if ((l&0xFFFF0000) == 0) {
127 unsigned long flags;
128 l = ((mbytes>>2)<<22)|(1<<16);
129 local_irq_save(flags);
130 wbinvd();
131 wrmsr(MSR_K6_WHCR, l, h);
132 local_irq_restore(flags);
133 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
134 mbytes);
135 }
136
137 return;
138 }
139
140 if (c->x86_model == 10) {
141 /* AMD Geode LX is model 10 */
142 /* placeholder for any needed mods */
143 return;
144 }
145}
146
147static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
148{
149#ifdef CONFIG_SMP
150 /* calling is from identify_secondary_cpu() ? */
151 if (!c->cpu_index)
152 return;
153
154 /*
155 * Certain Athlons might work (for various values of 'work') in SMP
156 * but they are not certified as MP capable.
157 */
158 /* Athlon 660/661 is valid. */
159 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
160 (c->x86_mask == 1)))
161 goto valid_k7;
162
163 /* Duron 670 is valid */
164 if ((c->x86_model == 7) && (c->x86_mask == 0))
165 goto valid_k7;
166
167 /*
168 * Athlon 662, Duron 671, and Athlon >model 7 have capability
169 * bit. It's worth noting that the A5 stepping (662) of some
170 * Athlon XP's have the MP bit set.
171 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
172 * more.
173 */
174 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
175 ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
176 (c->x86_model > 7))
177 if (cpu_has_mp)
178 goto valid_k7;
179
180 /* If we get here, not a certified SMP capable AMD system. */
181
182 /*
183 * Don't taint if we are running SMP kernel on a single non-MP
184 * approved Athlon
185 */
186 WARN_ONCE(1, "WARNING: This combination of AMD"
187 " processors is not suitable for SMP.\n");
188 if (!test_taint(TAINT_UNSAFE_SMP))
189 add_taint(TAINT_UNSAFE_SMP);
190
191valid_k7:
192 ;
193#endif
194}
195
196static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
197{
198 u32 l, h;
199
200 /*
201 * Bit 15 of Athlon specific MSR 15, needs to be 0
202 * to enable SSE on Palomino/Morgan/Barton CPU's.
203 * If the BIOS didn't enable it already, enable it here.
204 */
205 if (c->x86_model >= 6 && c->x86_model <= 10) {
206 if (!cpu_has(c, X86_FEATURE_XMM)) {
207 printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
208 rdmsr(MSR_K7_HWCR, l, h);
209 l &= ~0x00008000;
210 wrmsr(MSR_K7_HWCR, l, h);
211 set_cpu_cap(c, X86_FEATURE_XMM);
212 }
213 }
214
215 /*
216 * It's been determined by AMD that Athlons since model 8 stepping 1
217 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
218 * As per AMD technical note 27212 0.2
219 */
220 if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
221 rdmsr(MSR_K7_CLK_CTL, l, h);
222 if ((l & 0xfff00000) != 0x20000000) {
223 printk(KERN_INFO
224 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
225 l, ((l & 0x000fffff)|0x20000000));
226 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
227 }
228 }
229
230 set_cpu_cap(c, X86_FEATURE_K7);
231
232 amd_k7_smp_check(c);
233}
234#endif
235
236#ifdef CONFIG_NUMA
237/*
238 * To workaround broken NUMA config. Read the comment in
239 * srat_detect_node().
240 */
241static int __cpuinit nearby_node(int apicid)
242{
243 int i, node;
244
245 for (i = apicid - 1; i >= 0; i--) {
246 node = __apicid_to_node[i];
247 if (node != NUMA_NO_NODE && node_online(node))
248 return node;
249 }
250 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
251 node = __apicid_to_node[i];
252 if (node != NUMA_NO_NODE && node_online(node))
253 return node;
254 }
255 return first_node(node_online_map); /* Shouldn't happen */
256}
257#endif
258
259/*
260 * Fixup core topology information for
261 * (1) AMD multi-node processors
262 * Assumption: Number of cores in each internal node is the same.
263 * (2) AMD processors supporting compute units
264 */
265#ifdef CONFIG_X86_HT
266static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
267{
268 u32 nodes, cores_per_cu = 1;
269 u8 node_id;
270 int cpu = smp_processor_id();
271
272 /* get information required for multi-node processors */
273 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
274 u32 eax, ebx, ecx, edx;
275
276 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
277 nodes = ((ecx >> 8) & 7) + 1;
278 node_id = ecx & 7;
279
280 /* get compute unit information */
281 smp_num_siblings = ((ebx >> 8) & 3) + 1;
282 c->compute_unit_id = ebx & 0xff;
283 cores_per_cu += ((ebx >> 8) & 3);
284 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
285 u64 value;
286
287 rdmsrl(MSR_FAM10H_NODE_ID, value);
288 nodes = ((value >> 3) & 7) + 1;
289 node_id = value & 7;
290 } else
291 return;
292
293 /* fixup multi-node processor information */
294 if (nodes > 1) {
295 u32 cores_per_node;
296 u32 cus_per_node;
297
298 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
299 cores_per_node = c->x86_max_cores / nodes;
300 cus_per_node = cores_per_node / cores_per_cu;
301
302 /* store NodeID, use llc_shared_map to store sibling info */
303 per_cpu(cpu_llc_id, cpu) = node_id;
304
305 /* core id has to be in the [0 .. cores_per_node - 1] range */
306 c->cpu_core_id %= cores_per_node;
307 c->compute_unit_id %= cus_per_node;
308 }
309}
310#endif
311
312/*
313 * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
314 * Assumes number of cores is a power of two.
315 */
316static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
317{
318#ifdef CONFIG_X86_HT
319 unsigned bits;
320 int cpu = smp_processor_id();
321
322 bits = c->x86_coreid_bits;
323 /* Low order bits define the core id (index of core in socket) */
324 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
325 /* Convert the initial APIC ID into the socket ID */
326 c->phys_proc_id = c->initial_apicid >> bits;
327 /* use socket ID also for last level cache */
328 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
329 amd_get_topology(c);
330#endif
331}
332
333int amd_get_nb_id(int cpu)
334{
335 int id = 0;
336#ifdef CONFIG_SMP
337 id = per_cpu(cpu_llc_id, cpu);
338#endif
339 return id;
340}
341EXPORT_SYMBOL_GPL(amd_get_nb_id);
342
343static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
344{
345#ifdef CONFIG_NUMA
346 int cpu = smp_processor_id();
347 int node;
348 unsigned apicid = c->apicid;
349
350 node = numa_cpu_node(cpu);
351 if (node == NUMA_NO_NODE)
352 node = per_cpu(cpu_llc_id, cpu);
353
354 if (!node_online(node)) {
355 /*
356 * Two possibilities here:
357 *
358 * - The CPU is missing memory and no node was created. In
359 * that case try picking one from a nearby CPU.
360 *
361 * - The APIC IDs differ from the HyperTransport node IDs
362 * which the K8 northbridge parsing fills in. Assume
363 * they are all increased by a constant offset, but in
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 */
374 int ht_nodeid = c->initial_apicid;
375
376 if (ht_nodeid >= 0 &&
377 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
378 node = __apicid_to_node[ht_nodeid];
379 /* Pick a nearby node */
380 if (!node_online(node))
381 node = nearby_node(apicid);
382 }
383 numa_set_node(cpu, node);
384#endif
385}
386
387static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
388{
389#ifdef CONFIG_X86_HT
390 unsigned bits, ecx;
391
392 /* Multi core CPU? */
393 if (c->extended_cpuid_level < 0x80000008)
394 return;
395
396 ecx = cpuid_ecx(0x80000008);
397
398 c->x86_max_cores = (ecx & 0xff) + 1;
399
400 /* CPU telling us the core id bits shift? */
401 bits = (ecx >> 12) & 0xF;
402
403 /* Otherwise recompute */
404 if (bits == 0) {
405 while ((1 << bits) < c->x86_max_cores)
406 bits++;
407 }
408
409 c->x86_coreid_bits = bits;
410#endif
411}
412
413static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
414{
415 early_init_amd_mc(c);
416
417 /*
418 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
419 * with P/T states and does not stop in deep C-states
420 */
421 if (c->x86_power & (1 << 8)) {
422 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
423 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
424 }
425
426#ifdef CONFIG_X86_64
427 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
428#else
429 /* Set MTRR capability flag if appropriate */
430 if (c->x86 == 5)
431 if (c->x86_model == 13 || c->x86_model == 9 ||
432 (c->x86_model == 8 && c->x86_mask >= 8))
433 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
434#endif
435#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
436 /* check CPU config space for extended APIC ID */
437 if (cpu_has_apic && c->x86 >= 0xf) {
438 unsigned int val;
439 val = read_pci_config(0, 24, 0, 0x68);
440 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
441 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
442 }
443#endif
444
445 /* We need to do the following only once */
446 if (c != &boot_cpu_data)
447 return;
448
449 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
450
451 if (c->x86 > 0x10 ||
452 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
453 u64 val;
454
455 rdmsrl(MSR_K7_HWCR, val);
456 if (!(val & BIT(24)))
457 printk(KERN_WARNING FW_BUG "TSC doesn't count "
458 "with P0 frequency!\n");
459 }
460 }
461}
462
463static void __cpuinit init_amd(struct cpuinfo_x86 *c)
464{
465#ifdef CONFIG_SMP
466 unsigned long long value;
467
468 /*
469 * Disable TLB flush filter by setting HWCR.FFDIS on K8
470 * bit 6 of msr C001_0015
471 *
472 * Errata 63 for SH-B3 steppings
473 * Errata 122 for all steppings (F+ have it disabled by default)
474 */
475 if (c->x86 == 0xf) {
476 rdmsrl(MSR_K7_HWCR, value);
477 value |= 1 << 6;
478 wrmsrl(MSR_K7_HWCR, value);
479 }
480#endif
481
482 early_init_amd(c);
483
484 /*
485 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
486 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
487 */
488 clear_cpu_cap(c, 0*32+31);
489
490#ifdef CONFIG_X86_64
491 /* On C+ stepping K8 rep microcode works well for copy/memset */
492 if (c->x86 == 0xf) {
493 u32 level;
494
495 level = cpuid_eax(1);
496 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
497 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
498
499 /*
500 * Some BIOSes incorrectly force this feature, but only K8
501 * revision D (model = 0x14) and later actually support it.
502 * (AMD Erratum #110, docId: 25759).
503 */
504 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
505 u64 val;
506
507 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
508 if (!rdmsrl_amd_safe(0xc001100d, &val)) {
509 val &= ~(1ULL << 32);
510 wrmsrl_amd_safe(0xc001100d, val);
511 }
512 }
513
514 }
515 if (c->x86 >= 0x10)
516 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
517
518 /* get apicid instead of initial apic id from cpuid */
519 c->apicid = hard_smp_processor_id();
520#else
521
522 /*
523 * FIXME: We should handle the K5 here. Set up the write
524 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
525 * no bus pipeline)
526 */
527
528 switch (c->x86) {
529 case 4:
530 init_amd_k5(c);
531 break;
532 case 5:
533 init_amd_k6(c);
534 break;
535 case 6: /* An Athlon/Duron */
536 init_amd_k7(c);
537 break;
538 }
539
540 /* K6s reports MCEs but don't actually have all the MSRs */
541 if (c->x86 < 6)
542 clear_cpu_cap(c, X86_FEATURE_MCE);
543#endif
544
545 /* Enable workaround for FXSAVE leak */
546 if (c->x86 >= 6)
547 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
548
549 if (!c->x86_model_id[0]) {
550 switch (c->x86) {
551 case 0xf:
552 /* Should distinguish Models here, but this is only
553 a fallback anyways. */
554 strcpy(c->x86_model_id, "Hammer");
555 break;
556 }
557 }
558
559 cpu_detect_cache_sizes(c);
560
561 /* Multi core CPU? */
562 if (c->extended_cpuid_level >= 0x80000008) {
563 amd_detect_cmp(c);
564 srat_detect_node(c);
565 }
566
567#ifdef CONFIG_X86_32
568 detect_ht(c);
569#endif
570
571 if (c->extended_cpuid_level >= 0x80000006) {
572 if (cpuid_edx(0x80000006) & 0xf000)
573 num_cache_leaves = 4;
574 else
575 num_cache_leaves = 3;
576 }
577
578 if (c->x86 >= 0xf)
579 set_cpu_cap(c, X86_FEATURE_K8);
580
581 if (cpu_has_xmm2) {
582 /* MFENCE stops RDTSC speculation */
583 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
584 }
585
586#ifdef CONFIG_X86_64
587 if (c->x86 == 0x10) {
588 /* do this for boot cpu */
589 if (c == &boot_cpu_data)
590 check_enable_amd_mmconf_dmi();
591
592 fam10h_check_enable_mmcfg();
593 }
594
595 if (c == &boot_cpu_data && c->x86 >= 0xf) {
596 unsigned long long tseg;
597
598 /*
599 * Split up direct mapping around the TSEG SMM area.
600 * Don't do it for gbpages because there seems very little
601 * benefit in doing so.
602 */
603 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
604 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
605 if ((tseg>>PMD_SHIFT) <
606 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
607 ((tseg>>PMD_SHIFT) <
608 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
609 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
610 set_memory_4k((unsigned long)__va(tseg), 1);
611 }
612 }
613#endif
614
615 /*
616 * Family 0x12 and above processors have APIC timer
617 * running in deep C states.
618 */
619 if (c->x86 > 0x11)
620 set_cpu_cap(c, X86_FEATURE_ARAT);
621
622 /*
623 * Disable GART TLB Walk Errors on Fam10h. We do this here
624 * because this is always needed when GART is enabled, even in a
625 * kernel which has no MCE support built in.
626 */
627 if (c->x86 == 0x10) {
628 /*
629 * BIOS should disable GartTlbWlk Errors themself. If
630 * it doesn't do it here as suggested by the BKDG.
631 *
632 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
633 */
634 u64 mask;
635 int err;
636
637 err = rdmsrl_safe(MSR_AMD64_MCx_MASK(4), &mask);
638 if (err == 0) {
639 mask |= (1 << 10);
640 checking_wrmsrl(MSR_AMD64_MCx_MASK(4), mask);
641 }
642 }
643}
644
645#ifdef CONFIG_X86_32
646static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
647 unsigned int size)
648{
649 /* AMD errata T13 (order #21922) */
650 if ((c->x86 == 6)) {
651 /* Duron Rev A0 */
652 if (c->x86_model == 3 && c->x86_mask == 0)
653 size = 64;
654 /* Tbird rev A1/A2 */
655 if (c->x86_model == 4 &&
656 (c->x86_mask == 0 || c->x86_mask == 1))
657 size = 256;
658 }
659 return size;
660}
661#endif
662
663static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
664 .c_vendor = "AMD",
665 .c_ident = { "AuthenticAMD" },
666#ifdef CONFIG_X86_32
667 .c_models = {
668 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
669 {
670 [3] = "486 DX/2",
671 [7] = "486 DX/2-WB",
672 [8] = "486 DX/4",
673 [9] = "486 DX/4-WB",
674 [14] = "Am5x86-WT",
675 [15] = "Am5x86-WB"
676 }
677 },
678 },
679 .c_size_cache = amd_size_cache,
680#endif
681 .c_early_init = early_init_amd,
682 .c_init = init_amd,
683 .c_x86_vendor = X86_VENDOR_AMD,
684};
685
686cpu_dev_register(amd_cpu_dev);
687
688/*
689 * AMD errata checking
690 *
691 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
692 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
693 * have an OSVW id assigned, which it takes as first argument. Both take a
694 * variable number of family-specific model-stepping ranges created by
695 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
696 * int[] in arch/x86/include/asm/processor.h.
697 *
698 * Example:
699 *
700 * const int amd_erratum_319[] =
701 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
702 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
703 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
704 */
705
706const int amd_erratum_400[] =
707 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
708 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
709EXPORT_SYMBOL_GPL(amd_erratum_400);
710
711const int amd_erratum_383[] =
712 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
713EXPORT_SYMBOL_GPL(amd_erratum_383);
714
715bool cpu_has_amd_erratum(const int *erratum)
716{
717 struct cpuinfo_x86 *cpu = __this_cpu_ptr(&cpu_info);
718 int osvw_id = *erratum++;
719 u32 range;
720 u32 ms;
721
722 /*
723 * If called early enough that current_cpu_data hasn't been initialized
724 * yet, fall back to boot_cpu_data.
725 */
726 if (cpu->x86 == 0)
727 cpu = &boot_cpu_data;
728
729 if (cpu->x86_vendor != X86_VENDOR_AMD)
730 return false;
731
732 if (osvw_id >= 0 && osvw_id < 65536 &&
733 cpu_has(cpu, X86_FEATURE_OSVW)) {
734 u64 osvw_len;
735
736 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
737 if (osvw_id < osvw_len) {
738 u64 osvw_bits;
739
740 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
741 osvw_bits);
742 return osvw_bits & (1ULL << (osvw_id & 0x3f));
743 }
744 }
745
746 /* OSVW unavailable or ID unknown, match family-model-stepping range */
747 ms = (cpu->x86_model << 4) | cpu->x86_mask;
748 while ((range = *erratum++))
749 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
750 (ms >= AMD_MODEL_RANGE_START(range)) &&
751 (ms <= AMD_MODEL_RANGE_END(range)))
752 return true;
753
754 return false;
755}
756
757EXPORT_SYMBOL_GPL(cpu_has_amd_erratum);
1#include <linux/export.h>
2#include <linux/bitops.h>
3#include <linux/elf.h>
4#include <linux/mm.h>
5
6#include <linux/io.h>
7#include <linux/sched.h>
8#include <asm/processor.h>
9#include <asm/apic.h>
10#include <asm/cpu.h>
11#include <asm/pci-direct.h>
12
13#ifdef CONFIG_X86_64
14# include <asm/mmconfig.h>
15# include <asm/cacheflush.h>
16#endif
17
18#include "cpu.h"
19
20static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
21{
22 u32 gprs[8] = { 0 };
23 int err;
24
25 WARN_ONCE((boot_cpu_data.x86 != 0xf),
26 "%s should only be used on K8!\n", __func__);
27
28 gprs[1] = msr;
29 gprs[7] = 0x9c5a203a;
30
31 err = rdmsr_safe_regs(gprs);
32
33 *p = gprs[0] | ((u64)gprs[2] << 32);
34
35 return err;
36}
37
38static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
39{
40 u32 gprs[8] = { 0 };
41
42 WARN_ONCE((boot_cpu_data.x86 != 0xf),
43 "%s should only be used on K8!\n", __func__);
44
45 gprs[0] = (u32)val;
46 gprs[1] = msr;
47 gprs[2] = val >> 32;
48 gprs[7] = 0x9c5a203a;
49
50 return wrmsr_safe_regs(gprs);
51}
52
53#ifdef CONFIG_X86_32
54/*
55 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
56 * misexecution of code under Linux. Owners of such processors should
57 * contact AMD for precise details and a CPU swap.
58 *
59 * See http://www.multimania.com/poulot/k6bug.html
60 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
61 * (Publication # 21266 Issue Date: August 1998)
62 *
63 * The following test is erm.. interesting. AMD neglected to up
64 * the chip setting when fixing the bug but they also tweaked some
65 * performance at the same time..
66 */
67
68extern __visible void vide(void);
69__asm__(".globl vide\n\t.align 4\nvide: ret");
70
71static void init_amd_k5(struct cpuinfo_x86 *c)
72{
73/*
74 * General Systems BIOSen alias the cpu frequency registers
75 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
76 * drivers subsequently pokes it, and changes the CPU speed.
77 * Workaround : Remove the unneeded alias.
78 */
79#define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
80#define CBAR_ENB (0x80000000)
81#define CBAR_KEY (0X000000CB)
82 if (c->x86_model == 9 || c->x86_model == 10) {
83 if (inl(CBAR) & CBAR_ENB)
84 outl(0 | CBAR_KEY, CBAR);
85 }
86}
87
88
89static void init_amd_k6(struct cpuinfo_x86 *c)
90{
91 u32 l, h;
92 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
93
94 if (c->x86_model < 6) {
95 /* Based on AMD doc 20734R - June 2000 */
96 if (c->x86_model == 0) {
97 clear_cpu_cap(c, X86_FEATURE_APIC);
98 set_cpu_cap(c, X86_FEATURE_PGE);
99 }
100 return;
101 }
102
103 if (c->x86_model == 6 && c->x86_mask == 1) {
104 const int K6_BUG_LOOP = 1000000;
105 int n;
106 void (*f_vide)(void);
107 unsigned long d, d2;
108
109 printk(KERN_INFO "AMD K6 stepping B detected - ");
110
111 /*
112 * It looks like AMD fixed the 2.6.2 bug and improved indirect
113 * calls at the same time.
114 */
115
116 n = K6_BUG_LOOP;
117 f_vide = vide;
118 rdtscl(d);
119 while (n--)
120 f_vide();
121 rdtscl(d2);
122 d = d2-d;
123
124 if (d > 20*K6_BUG_LOOP)
125 printk(KERN_CONT
126 "system stability may be impaired when more than 32 MB are used.\n");
127 else
128 printk(KERN_CONT "probably OK (after B9730xxxx).\n");
129 }
130
131 /* K6 with old style WHCR */
132 if (c->x86_model < 8 ||
133 (c->x86_model == 8 && c->x86_mask < 8)) {
134 /* We can only write allocate on the low 508Mb */
135 if (mbytes > 508)
136 mbytes = 508;
137
138 rdmsr(MSR_K6_WHCR, l, h);
139 if ((l&0x0000FFFF) == 0) {
140 unsigned long flags;
141 l = (1<<0)|((mbytes/4)<<1);
142 local_irq_save(flags);
143 wbinvd();
144 wrmsr(MSR_K6_WHCR, l, h);
145 local_irq_restore(flags);
146 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
147 mbytes);
148 }
149 return;
150 }
151
152 if ((c->x86_model == 8 && c->x86_mask > 7) ||
153 c->x86_model == 9 || c->x86_model == 13) {
154 /* The more serious chips .. */
155
156 if (mbytes > 4092)
157 mbytes = 4092;
158
159 rdmsr(MSR_K6_WHCR, l, h);
160 if ((l&0xFFFF0000) == 0) {
161 unsigned long flags;
162 l = ((mbytes>>2)<<22)|(1<<16);
163 local_irq_save(flags);
164 wbinvd();
165 wrmsr(MSR_K6_WHCR, l, h);
166 local_irq_restore(flags);
167 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
168 mbytes);
169 }
170
171 return;
172 }
173
174 if (c->x86_model == 10) {
175 /* AMD Geode LX is model 10 */
176 /* placeholder for any needed mods */
177 return;
178 }
179}
180
181static void amd_k7_smp_check(struct cpuinfo_x86 *c)
182{
183 /* calling is from identify_secondary_cpu() ? */
184 if (!c->cpu_index)
185 return;
186
187 /*
188 * Certain Athlons might work (for various values of 'work') in SMP
189 * but they are not certified as MP capable.
190 */
191 /* Athlon 660/661 is valid. */
192 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
193 (c->x86_mask == 1)))
194 return;
195
196 /* Duron 670 is valid */
197 if ((c->x86_model == 7) && (c->x86_mask == 0))
198 return;
199
200 /*
201 * Athlon 662, Duron 671, and Athlon >model 7 have capability
202 * bit. It's worth noting that the A5 stepping (662) of some
203 * Athlon XP's have the MP bit set.
204 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
205 * more.
206 */
207 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
208 ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
209 (c->x86_model > 7))
210 if (cpu_has_mp)
211 return;
212
213 /* If we get here, not a certified SMP capable AMD system. */
214
215 /*
216 * Don't taint if we are running SMP kernel on a single non-MP
217 * approved Athlon
218 */
219 WARN_ONCE(1, "WARNING: This combination of AMD"
220 " processors is not suitable for SMP.\n");
221 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
222}
223
224static void init_amd_k7(struct cpuinfo_x86 *c)
225{
226 u32 l, h;
227
228 /*
229 * Bit 15 of Athlon specific MSR 15, needs to be 0
230 * to enable SSE on Palomino/Morgan/Barton CPU's.
231 * If the BIOS didn't enable it already, enable it here.
232 */
233 if (c->x86_model >= 6 && c->x86_model <= 10) {
234 if (!cpu_has(c, X86_FEATURE_XMM)) {
235 printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
236 msr_clear_bit(MSR_K7_HWCR, 15);
237 set_cpu_cap(c, X86_FEATURE_XMM);
238 }
239 }
240
241 /*
242 * It's been determined by AMD that Athlons since model 8 stepping 1
243 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
244 * As per AMD technical note 27212 0.2
245 */
246 if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
247 rdmsr(MSR_K7_CLK_CTL, l, h);
248 if ((l & 0xfff00000) != 0x20000000) {
249 printk(KERN_INFO
250 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
251 l, ((l & 0x000fffff)|0x20000000));
252 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
253 }
254 }
255
256 set_cpu_cap(c, X86_FEATURE_K7);
257
258 amd_k7_smp_check(c);
259}
260#endif
261
262#ifdef CONFIG_NUMA
263/*
264 * To workaround broken NUMA config. Read the comment in
265 * srat_detect_node().
266 */
267static int nearby_node(int apicid)
268{
269 int i, node;
270
271 for (i = apicid - 1; i >= 0; i--) {
272 node = __apicid_to_node[i];
273 if (node != NUMA_NO_NODE && node_online(node))
274 return node;
275 }
276 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
277 node = __apicid_to_node[i];
278 if (node != NUMA_NO_NODE && node_online(node))
279 return node;
280 }
281 return first_node(node_online_map); /* Shouldn't happen */
282}
283#endif
284
285/*
286 * Fixup core topology information for
287 * (1) AMD multi-node processors
288 * Assumption: Number of cores in each internal node is the same.
289 * (2) AMD processors supporting compute units
290 */
291#ifdef CONFIG_X86_HT
292static void amd_get_topology(struct cpuinfo_x86 *c)
293{
294 u32 nodes, cores_per_cu = 1;
295 u8 node_id;
296 int cpu = smp_processor_id();
297
298 /* get information required for multi-node processors */
299 if (cpu_has_topoext) {
300 u32 eax, ebx, ecx, edx;
301
302 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
303 nodes = ((ecx >> 8) & 7) + 1;
304 node_id = ecx & 7;
305
306 /* get compute unit information */
307 smp_num_siblings = ((ebx >> 8) & 3) + 1;
308 c->compute_unit_id = ebx & 0xff;
309 cores_per_cu += ((ebx >> 8) & 3);
310 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
311 u64 value;
312
313 rdmsrl(MSR_FAM10H_NODE_ID, value);
314 nodes = ((value >> 3) & 7) + 1;
315 node_id = value & 7;
316 } else
317 return;
318
319 /* fixup multi-node processor information */
320 if (nodes > 1) {
321 u32 cores_per_node;
322 u32 cus_per_node;
323
324 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
325 cores_per_node = c->x86_max_cores / nodes;
326 cus_per_node = cores_per_node / cores_per_cu;
327
328 /* store NodeID, use llc_shared_map to store sibling info */
329 per_cpu(cpu_llc_id, cpu) = node_id;
330
331 /* core id has to be in the [0 .. cores_per_node - 1] range */
332 c->cpu_core_id %= cores_per_node;
333 c->compute_unit_id %= cus_per_node;
334 }
335}
336#endif
337
338/*
339 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
340 * Assumes number of cores is a power of two.
341 */
342static void amd_detect_cmp(struct cpuinfo_x86 *c)
343{
344#ifdef CONFIG_X86_HT
345 unsigned bits;
346 int cpu = smp_processor_id();
347
348 bits = c->x86_coreid_bits;
349 /* Low order bits define the core id (index of core in socket) */
350 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
351 /* Convert the initial APIC ID into the socket ID */
352 c->phys_proc_id = c->initial_apicid >> bits;
353 /* use socket ID also for last level cache */
354 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
355 amd_get_topology(c);
356#endif
357}
358
359u16 amd_get_nb_id(int cpu)
360{
361 u16 id = 0;
362#ifdef CONFIG_SMP
363 id = per_cpu(cpu_llc_id, cpu);
364#endif
365 return id;
366}
367EXPORT_SYMBOL_GPL(amd_get_nb_id);
368
369static void srat_detect_node(struct cpuinfo_x86 *c)
370{
371#ifdef CONFIG_NUMA
372 int cpu = smp_processor_id();
373 int node;
374 unsigned apicid = c->apicid;
375
376 node = numa_cpu_node(cpu);
377 if (node == NUMA_NO_NODE)
378 node = per_cpu(cpu_llc_id, cpu);
379
380 /*
381 * On multi-fabric platform (e.g. Numascale NumaChip) a
382 * platform-specific handler needs to be called to fixup some
383 * IDs of the CPU.
384 */
385 if (x86_cpuinit.fixup_cpu_id)
386 x86_cpuinit.fixup_cpu_id(c, node);
387
388 if (!node_online(node)) {
389 /*
390 * Two possibilities here:
391 *
392 * - The CPU is missing memory and no node was created. In
393 * that case try picking one from a nearby CPU.
394 *
395 * - The APIC IDs differ from the HyperTransport node IDs
396 * which the K8 northbridge parsing fills in. Assume
397 * they are all increased by a constant offset, but in
398 * the same order as the HT nodeids. If that doesn't
399 * result in a usable node fall back to the path for the
400 * previous case.
401 *
402 * This workaround operates directly on the mapping between
403 * APIC ID and NUMA node, assuming certain relationship
404 * between APIC ID, HT node ID and NUMA topology. As going
405 * through CPU mapping may alter the outcome, directly
406 * access __apicid_to_node[].
407 */
408 int ht_nodeid = c->initial_apicid;
409
410 if (ht_nodeid >= 0 &&
411 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
412 node = __apicid_to_node[ht_nodeid];
413 /* Pick a nearby node */
414 if (!node_online(node))
415 node = nearby_node(apicid);
416 }
417 numa_set_node(cpu, node);
418#endif
419}
420
421static void early_init_amd_mc(struct cpuinfo_x86 *c)
422{
423#ifdef CONFIG_X86_HT
424 unsigned bits, ecx;
425
426 /* Multi core CPU? */
427 if (c->extended_cpuid_level < 0x80000008)
428 return;
429
430 ecx = cpuid_ecx(0x80000008);
431
432 c->x86_max_cores = (ecx & 0xff) + 1;
433
434 /* CPU telling us the core id bits shift? */
435 bits = (ecx >> 12) & 0xF;
436
437 /* Otherwise recompute */
438 if (bits == 0) {
439 while ((1 << bits) < c->x86_max_cores)
440 bits++;
441 }
442
443 c->x86_coreid_bits = bits;
444#endif
445}
446
447static void bsp_init_amd(struct cpuinfo_x86 *c)
448{
449 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
450
451 if (c->x86 > 0x10 ||
452 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
453 u64 val;
454
455 rdmsrl(MSR_K7_HWCR, val);
456 if (!(val & BIT(24)))
457 printk(KERN_WARNING FW_BUG "TSC doesn't count "
458 "with P0 frequency!\n");
459 }
460 }
461
462 if (c->x86 == 0x15) {
463 unsigned long upperbit;
464 u32 cpuid, assoc;
465
466 cpuid = cpuid_edx(0x80000005);
467 assoc = cpuid >> 16 & 0xff;
468 upperbit = ((cpuid >> 24) << 10) / assoc;
469
470 va_align.mask = (upperbit - 1) & PAGE_MASK;
471 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
472 }
473}
474
475static void early_init_amd(struct cpuinfo_x86 *c)
476{
477 early_init_amd_mc(c);
478
479 /*
480 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
481 * with P/T states and does not stop in deep C-states
482 */
483 if (c->x86_power & (1 << 8)) {
484 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
485 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
486 if (!check_tsc_unstable())
487 set_sched_clock_stable();
488 }
489
490#ifdef CONFIG_X86_64
491 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
492#else
493 /* Set MTRR capability flag if appropriate */
494 if (c->x86 == 5)
495 if (c->x86_model == 13 || c->x86_model == 9 ||
496 (c->x86_model == 8 && c->x86_mask >= 8))
497 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
498#endif
499#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
500 /* check CPU config space for extended APIC ID */
501 if (cpu_has_apic && c->x86 >= 0xf) {
502 unsigned int val;
503 val = read_pci_config(0, 24, 0, 0x68);
504 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
505 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
506 }
507#endif
508
509 /* F16h erratum 793, CVE-2013-6885 */
510 if (c->x86 == 0x16 && c->x86_model <= 0xf)
511 msr_set_bit(MSR_AMD64_LS_CFG, 15);
512}
513
514static const int amd_erratum_383[];
515static const int amd_erratum_400[];
516static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
517
518static void init_amd(struct cpuinfo_x86 *c)
519{
520 u32 dummy;
521 unsigned long long value;
522
523#ifdef CONFIG_SMP
524 /*
525 * Disable TLB flush filter by setting HWCR.FFDIS on K8
526 * bit 6 of msr C001_0015
527 *
528 * Errata 63 for SH-B3 steppings
529 * Errata 122 for all steppings (F+ have it disabled by default)
530 */
531 if (c->x86 == 0xf)
532 msr_set_bit(MSR_K7_HWCR, 6);
533#endif
534
535 early_init_amd(c);
536
537 /*
538 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
539 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
540 */
541 clear_cpu_cap(c, 0*32+31);
542
543#ifdef CONFIG_X86_64
544 /* On C+ stepping K8 rep microcode works well for copy/memset */
545 if (c->x86 == 0xf) {
546 u32 level;
547
548 level = cpuid_eax(1);
549 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
550 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
551
552 /*
553 * Some BIOSes incorrectly force this feature, but only K8
554 * revision D (model = 0x14) and later actually support it.
555 * (AMD Erratum #110, docId: 25759).
556 */
557 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
558 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
559 if (!rdmsrl_amd_safe(0xc001100d, &value)) {
560 value &= ~(1ULL << 32);
561 wrmsrl_amd_safe(0xc001100d, value);
562 }
563 }
564
565 }
566 if (c->x86 >= 0x10)
567 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
568
569 /* get apicid instead of initial apic id from cpuid */
570 c->apicid = hard_smp_processor_id();
571#else
572
573 /*
574 * FIXME: We should handle the K5 here. Set up the write
575 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
576 * no bus pipeline)
577 */
578
579 switch (c->x86) {
580 case 4:
581 init_amd_k5(c);
582 break;
583 case 5:
584 init_amd_k6(c);
585 break;
586 case 6: /* An Athlon/Duron */
587 init_amd_k7(c);
588 break;
589 }
590
591 /* K6s reports MCEs but don't actually have all the MSRs */
592 if (c->x86 < 6)
593 clear_cpu_cap(c, X86_FEATURE_MCE);
594#endif
595
596 /* Enable workaround for FXSAVE leak */
597 if (c->x86 >= 6)
598 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
599
600 if (!c->x86_model_id[0]) {
601 switch (c->x86) {
602 case 0xf:
603 /* Should distinguish Models here, but this is only
604 a fallback anyways. */
605 strcpy(c->x86_model_id, "Hammer");
606 break;
607 }
608 }
609
610 /* re-enable TopologyExtensions if switched off by BIOS */
611 if ((c->x86 == 0x15) &&
612 (c->x86_model >= 0x10) && (c->x86_model <= 0x1f) &&
613 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
614
615 if (msr_set_bit(0xc0011005, 54) > 0) {
616 rdmsrl(0xc0011005, value);
617 if (value & BIT_64(54)) {
618 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
619 pr_info(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
620 }
621 }
622 }
623
624 /*
625 * The way access filter has a performance penalty on some workloads.
626 * Disable it on the affected CPUs.
627 */
628 if ((c->x86 == 0x15) &&
629 (c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
630
631 if (!rdmsrl_safe(0xc0011021, &value) && !(value & 0x1E)) {
632 value |= 0x1E;
633 wrmsrl_safe(0xc0011021, value);
634 }
635 }
636
637 cpu_detect_cache_sizes(c);
638
639 /* Multi core CPU? */
640 if (c->extended_cpuid_level >= 0x80000008) {
641 amd_detect_cmp(c);
642 srat_detect_node(c);
643 }
644
645#ifdef CONFIG_X86_32
646 detect_ht(c);
647#endif
648
649 init_amd_cacheinfo(c);
650
651 if (c->x86 >= 0xf)
652 set_cpu_cap(c, X86_FEATURE_K8);
653
654 if (cpu_has_xmm2) {
655 /* MFENCE stops RDTSC speculation */
656 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
657 }
658
659#ifdef CONFIG_X86_64
660 if (c->x86 == 0x10) {
661 /* do this for boot cpu */
662 if (c == &boot_cpu_data)
663 check_enable_amd_mmconf_dmi();
664
665 fam10h_check_enable_mmcfg();
666 }
667
668 if (c == &boot_cpu_data && c->x86 >= 0xf) {
669 unsigned long long tseg;
670
671 /*
672 * Split up direct mapping around the TSEG SMM area.
673 * Don't do it for gbpages because there seems very little
674 * benefit in doing so.
675 */
676 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
677 unsigned long pfn = tseg >> PAGE_SHIFT;
678
679 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
680 if (pfn_range_is_mapped(pfn, pfn + 1))
681 set_memory_4k((unsigned long)__va(tseg), 1);
682 }
683 }
684#endif
685
686 /*
687 * Family 0x12 and above processors have APIC timer
688 * running in deep C states.
689 */
690 if (c->x86 > 0x11)
691 set_cpu_cap(c, X86_FEATURE_ARAT);
692
693 if (c->x86 == 0x10) {
694 /*
695 * Disable GART TLB Walk Errors on Fam10h. We do this here
696 * because this is always needed when GART is enabled, even in a
697 * kernel which has no MCE support built in.
698 * BIOS should disable GartTlbWlk Errors already. If
699 * it doesn't, do it here as suggested by the BKDG.
700 *
701 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
702 */
703 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
704
705 /*
706 * On family 10h BIOS may not have properly enabled WC+ support,
707 * causing it to be converted to CD memtype. This may result in
708 * performance degradation for certain nested-paging guests.
709 * Prevent this conversion by clearing bit 24 in
710 * MSR_AMD64_BU_CFG2.
711 *
712 * NOTE: we want to use the _safe accessors so as not to #GP kvm
713 * guests on older kvm hosts.
714 */
715 msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
716
717 if (cpu_has_amd_erratum(c, amd_erratum_383))
718 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
719 }
720
721 if (cpu_has_amd_erratum(c, amd_erratum_400))
722 set_cpu_bug(c, X86_BUG_AMD_APIC_C1E);
723
724 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
725}
726
727#ifdef CONFIG_X86_32
728static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
729{
730 /* AMD errata T13 (order #21922) */
731 if ((c->x86 == 6)) {
732 /* Duron Rev A0 */
733 if (c->x86_model == 3 && c->x86_mask == 0)
734 size = 64;
735 /* Tbird rev A1/A2 */
736 if (c->x86_model == 4 &&
737 (c->x86_mask == 0 || c->x86_mask == 1))
738 size = 256;
739 }
740 return size;
741}
742#endif
743
744static void cpu_set_tlb_flushall_shift(struct cpuinfo_x86 *c)
745{
746 tlb_flushall_shift = 6;
747}
748
749static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
750{
751 u32 ebx, eax, ecx, edx;
752 u16 mask = 0xfff;
753
754 if (c->x86 < 0xf)
755 return;
756
757 if (c->extended_cpuid_level < 0x80000006)
758 return;
759
760 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
761
762 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
763 tlb_lli_4k[ENTRIES] = ebx & mask;
764
765 /*
766 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
767 * characteristics from the CPUID function 0x80000005 instead.
768 */
769 if (c->x86 == 0xf) {
770 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
771 mask = 0xff;
772 }
773
774 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
775 if (!((eax >> 16) & mask))
776 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
777 else
778 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
779
780 /* a 4M entry uses two 2M entries */
781 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
782
783 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
784 if (!(eax & mask)) {
785 /* Erratum 658 */
786 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
787 tlb_lli_2m[ENTRIES] = 1024;
788 } else {
789 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
790 tlb_lli_2m[ENTRIES] = eax & 0xff;
791 }
792 } else
793 tlb_lli_2m[ENTRIES] = eax & mask;
794
795 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
796
797 cpu_set_tlb_flushall_shift(c);
798}
799
800static const struct cpu_dev amd_cpu_dev = {
801 .c_vendor = "AMD",
802 .c_ident = { "AuthenticAMD" },
803#ifdef CONFIG_X86_32
804 .legacy_models = {
805 { .family = 4, .model_names =
806 {
807 [3] = "486 DX/2",
808 [7] = "486 DX/2-WB",
809 [8] = "486 DX/4",
810 [9] = "486 DX/4-WB",
811 [14] = "Am5x86-WT",
812 [15] = "Am5x86-WB"
813 }
814 },
815 },
816 .legacy_cache_size = amd_size_cache,
817#endif
818 .c_early_init = early_init_amd,
819 .c_detect_tlb = cpu_detect_tlb_amd,
820 .c_bsp_init = bsp_init_amd,
821 .c_init = init_amd,
822 .c_x86_vendor = X86_VENDOR_AMD,
823};
824
825cpu_dev_register(amd_cpu_dev);
826
827/*
828 * AMD errata checking
829 *
830 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
831 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
832 * have an OSVW id assigned, which it takes as first argument. Both take a
833 * variable number of family-specific model-stepping ranges created by
834 * AMD_MODEL_RANGE().
835 *
836 * Example:
837 *
838 * const int amd_erratum_319[] =
839 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
840 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
841 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
842 */
843
844#define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
845#define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
846#define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
847 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
848#define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
849#define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
850#define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
851
852static const int amd_erratum_400[] =
853 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
854 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
855
856static const int amd_erratum_383[] =
857 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
858
859
860static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
861{
862 int osvw_id = *erratum++;
863 u32 range;
864 u32 ms;
865
866 if (osvw_id >= 0 && osvw_id < 65536 &&
867 cpu_has(cpu, X86_FEATURE_OSVW)) {
868 u64 osvw_len;
869
870 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
871 if (osvw_id < osvw_len) {
872 u64 osvw_bits;
873
874 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
875 osvw_bits);
876 return osvw_bits & (1ULL << (osvw_id & 0x3f));
877 }
878 }
879
880 /* OSVW unavailable or ID unknown, match family-model-stepping range */
881 ms = (cpu->x86_model << 4) | cpu->x86_mask;
882 while ((range = *erratum++))
883 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
884 (ms >= AMD_MODEL_RANGE_START(range)) &&
885 (ms <= AMD_MODEL_RANGE_END(range)))
886 return true;
887
888 return false;
889}