1 #include <linux/export.h>
2 #include <linux/init.h>
3 #include <linux/bitops.h>
8 #include <asm/processor.h>
11 #include <asm/pci-direct.h>
14 # include <asm/numa_64.h>
15 # include <asm/mmconfig.h>
16 # include <asm/cacheflush.h>
23 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
24 * misexecution of code under Linux. Owners of such processors should
25 * contact AMD for precise details and a CPU swap.
27 * See http://www.multimania.com/poulot/k6bug.html
28 * http://www.amd.com/K6/k6docs/revgd.html
30 * The following test is erm.. interesting. AMD neglected to up
31 * the chip setting when fixing the bug but they also tweaked some
32 * performance at the same time..
35 extern void vide(void);
36 __asm__(".align 4\nvide: ret");
38 static void __cpuinit
init_amd_k5(struct cpuinfo_x86
*c
)
41 * General Systems BIOSen alias the cpu frequency registers
42 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
43 * drivers subsequently pokes it, and changes the CPU speed.
44 * Workaround : Remove the unneeded alias.
46 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
47 #define CBAR_ENB (0x80000000)
48 #define CBAR_KEY (0X000000CB)
49 if (c
->x86_model
== 9 || c
->x86_model
== 10) {
50 if (inl(CBAR
) & CBAR_ENB
)
51 outl(0 | CBAR_KEY
, CBAR
);
56 static void __cpuinit
init_amd_k6(struct cpuinfo_x86
*c
)
59 int mbytes
= num_physpages
>> (20-PAGE_SHIFT
);
61 if (c
->x86_model
< 6) {
62 /* Based on AMD doc 20734R - June 2000 */
63 if (c
->x86_model
== 0) {
64 clear_cpu_cap(c
, X86_FEATURE_APIC
);
65 set_cpu_cap(c
, X86_FEATURE_PGE
);
70 if (c
->x86_model
== 6 && c
->x86_mask
== 1) {
71 const int K6_BUG_LOOP
= 1000000;
76 printk(KERN_INFO
"AMD K6 stepping B detected - ");
79 * It looks like AMD fixed the 2.6.2 bug and improved indirect
80 * calls at the same time.
91 if (d
> 20*K6_BUG_LOOP
)
93 "system stability may be impaired when more than 32 MB are used.\n");
95 printk(KERN_CONT
"probably OK (after B9730xxxx).\n");
96 printk(KERN_INFO
"Please see http://membres.lycos.fr/poulot/k6bug.html\n");
99 /* K6 with old style WHCR */
100 if (c
->x86_model
< 8 ||
101 (c
->x86_model
== 8 && c
->x86_mask
< 8)) {
102 /* We can only write allocate on the low 508Mb */
106 rdmsr(MSR_K6_WHCR
, l
, h
);
107 if ((l
&0x0000FFFF) == 0) {
109 l
= (1<<0)|((mbytes
/4)<<1);
110 local_irq_save(flags
);
112 wrmsr(MSR_K6_WHCR
, l
, h
);
113 local_irq_restore(flags
);
114 printk(KERN_INFO
"Enabling old style K6 write allocation for %d Mb\n",
120 if ((c
->x86_model
== 8 && c
->x86_mask
> 7) ||
121 c
->x86_model
== 9 || c
->x86_model
== 13) {
122 /* The more serious chips .. */
127 rdmsr(MSR_K6_WHCR
, l
, h
);
128 if ((l
&0xFFFF0000) == 0) {
130 l
= ((mbytes
>>2)<<22)|(1<<16);
131 local_irq_save(flags
);
133 wrmsr(MSR_K6_WHCR
, l
, h
);
134 local_irq_restore(flags
);
135 printk(KERN_INFO
"Enabling new style K6 write allocation for %d Mb\n",
142 if (c
->x86_model
== 10) {
143 /* AMD Geode LX is model 10 */
144 /* placeholder for any needed mods */
149 static void __cpuinit
amd_k7_smp_check(struct cpuinfo_x86
*c
)
152 /* calling is from identify_secondary_cpu() ? */
157 * Certain Athlons might work (for various values of 'work') in SMP
158 * but they are not certified as MP capable.
160 /* Athlon 660/661 is valid. */
161 if ((c
->x86_model
== 6) && ((c
->x86_mask
== 0) ||
165 /* Duron 670 is valid */
166 if ((c
->x86_model
== 7) && (c
->x86_mask
== 0))
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
176 if (((c
->x86_model
== 6) && (c
->x86_mask
>= 2)) ||
177 ((c
->x86_model
== 7) && (c
->x86_mask
>= 1)) ||
182 /* If we get here, not a certified SMP capable AMD system. */
185 * Don't taint if we are running SMP kernel on a single non-MP
188 WARN_ONCE(1, "WARNING: This combination of AMD"
189 " processors is not suitable for SMP.\n");
190 if (!test_taint(TAINT_UNSAFE_SMP
))
191 add_taint(TAINT_UNSAFE_SMP
);
198 static void __cpuinit
init_amd_k7(struct cpuinfo_x86
*c
)
203 * Bit 15 of Athlon specific MSR 15, needs to be 0
204 * to enable SSE on Palomino/Morgan/Barton CPU's.
205 * If the BIOS didn't enable it already, enable it here.
207 if (c
->x86_model
>= 6 && c
->x86_model
<= 10) {
208 if (!cpu_has(c
, X86_FEATURE_XMM
)) {
209 printk(KERN_INFO
"Enabling disabled K7/SSE Support.\n");
210 rdmsr(MSR_K7_HWCR
, l
, h
);
212 wrmsr(MSR_K7_HWCR
, l
, h
);
213 set_cpu_cap(c
, X86_FEATURE_XMM
);
218 * It's been determined by AMD that Athlons since model 8 stepping 1
219 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
220 * As per AMD technical note 27212 0.2
222 if ((c
->x86_model
== 8 && c
->x86_mask
>= 1) || (c
->x86_model
> 8)) {
223 rdmsr(MSR_K7_CLK_CTL
, l
, h
);
224 if ((l
& 0xfff00000) != 0x20000000) {
226 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
227 l
, ((l
& 0x000fffff)|0x20000000));
228 wrmsr(MSR_K7_CLK_CTL
, (l
& 0x000fffff)|0x20000000, h
);
232 set_cpu_cap(c
, X86_FEATURE_K7
);
240 * To workaround broken NUMA config. Read the comment in
241 * srat_detect_node().
243 static int __cpuinit
nearby_node(int apicid
)
247 for (i
= apicid
- 1; i
>= 0; i
--) {
248 node
= __apicid_to_node
[i
];
249 if (node
!= NUMA_NO_NODE
&& node_online(node
))
252 for (i
= apicid
+ 1; i
< MAX_LOCAL_APIC
; i
++) {
253 node
= __apicid_to_node
[i
];
254 if (node
!= NUMA_NO_NODE
&& node_online(node
))
257 return first_node(node_online_map
); /* Shouldn't happen */
262 * Fixup core topology information for
263 * (1) AMD multi-node processors
264 * Assumption: Number of cores in each internal node is the same.
265 * (2) AMD processors supporting compute units
268 static void __cpuinit
amd_get_topology(struct cpuinfo_x86
*c
)
270 u32 nodes
, cores_per_cu
= 1;
272 int cpu
= smp_processor_id();
274 /* get information required for multi-node processors */
275 if (cpu_has(c
, X86_FEATURE_TOPOEXT
)) {
276 u32 eax
, ebx
, ecx
, edx
;
278 cpuid(0x8000001e, &eax
, &ebx
, &ecx
, &edx
);
279 nodes
= ((ecx
>> 8) & 7) + 1;
282 /* get compute unit information */
283 smp_num_siblings
= ((ebx
>> 8) & 3) + 1;
284 c
->compute_unit_id
= ebx
& 0xff;
285 cores_per_cu
+= ((ebx
>> 8) & 3);
286 } else if (cpu_has(c
, X86_FEATURE_NODEID_MSR
)) {
289 rdmsrl(MSR_FAM10H_NODE_ID
, value
);
290 nodes
= ((value
>> 3) & 7) + 1;
295 /* fixup multi-node processor information */
300 set_cpu_cap(c
, X86_FEATURE_AMD_DCM
);
301 cores_per_node
= c
->x86_max_cores
/ nodes
;
302 cus_per_node
= cores_per_node
/ cores_per_cu
;
304 /* store NodeID, use llc_shared_map to store sibling info */
305 per_cpu(cpu_llc_id
, cpu
) = node_id
;
307 /* core id has to be in the [0 .. cores_per_node - 1] range */
308 c
->cpu_core_id
%= cores_per_node
;
309 c
->compute_unit_id
%= cus_per_node
;
315 * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
316 * Assumes number of cores is a power of two.
318 static void __cpuinit
amd_detect_cmp(struct cpuinfo_x86
*c
)
322 int cpu
= smp_processor_id();
324 bits
= c
->x86_coreid_bits
;
325 /* Low order bits define the core id (index of core in socket) */
326 c
->cpu_core_id
= c
->initial_apicid
& ((1 << bits
)-1);
327 /* Convert the initial APIC ID into the socket ID */
328 c
->phys_proc_id
= c
->initial_apicid
>> bits
;
329 /* use socket ID also for last level cache */
330 per_cpu(cpu_llc_id
, cpu
) = c
->phys_proc_id
;
335 int amd_get_nb_id(int cpu
)
339 id
= per_cpu(cpu_llc_id
, cpu
);
343 EXPORT_SYMBOL_GPL(amd_get_nb_id
);
345 static void __cpuinit
srat_detect_node(struct cpuinfo_x86
*c
)
348 int cpu
= smp_processor_id();
350 unsigned apicid
= c
->apicid
;
352 node
= numa_cpu_node(cpu
);
353 if (node
== NUMA_NO_NODE
)
354 node
= per_cpu(cpu_llc_id
, cpu
);
356 if (!node_online(node
)) {
358 * Two possibilities here:
360 * - The CPU is missing memory and no node was created. In
361 * that case try picking one from a nearby CPU.
363 * - The APIC IDs differ from the HyperTransport node IDs
364 * which the K8 northbridge parsing fills in. Assume
365 * they are all increased by a constant offset, but in
366 * the same order as the HT nodeids. If that doesn't
367 * result in a usable node fall back to the path for the
370 * This workaround operates directly on the mapping between
371 * APIC ID and NUMA node, assuming certain relationship
372 * between APIC ID, HT node ID and NUMA topology. As going
373 * through CPU mapping may alter the outcome, directly
374 * access __apicid_to_node[].
376 int ht_nodeid
= c
->initial_apicid
;
378 if (ht_nodeid
>= 0 &&
379 __apicid_to_node
[ht_nodeid
] != NUMA_NO_NODE
)
380 node
= __apicid_to_node
[ht_nodeid
];
381 /* Pick a nearby node */
382 if (!node_online(node
))
383 node
= nearby_node(apicid
);
385 numa_set_node(cpu
, node
);
389 static void __cpuinit
early_init_amd_mc(struct cpuinfo_x86
*c
)
394 /* Multi core CPU? */
395 if (c
->extended_cpuid_level
< 0x80000008)
398 ecx
= cpuid_ecx(0x80000008);
400 c
->x86_max_cores
= (ecx
& 0xff) + 1;
402 /* CPU telling us the core id bits shift? */
403 bits
= (ecx
>> 12) & 0xF;
405 /* Otherwise recompute */
407 while ((1 << bits
) < c
->x86_max_cores
)
411 c
->x86_coreid_bits
= bits
;
415 static void __cpuinit
bsp_init_amd(struct cpuinfo_x86
*c
)
417 if (cpu_has(c
, X86_FEATURE_CONSTANT_TSC
)) {
420 (c
->x86
== 0x10 && c
->x86_model
>= 0x2)) {
423 rdmsrl(MSR_K7_HWCR
, val
);
424 if (!(val
& BIT(24)))
425 printk(KERN_WARNING FW_BUG
"TSC doesn't count "
426 "with P0 frequency!\n");
430 if (c
->x86
== 0x15) {
431 unsigned long upperbit
;
434 cpuid
= cpuid_edx(0x80000005);
435 assoc
= cpuid
>> 16 & 0xff;
436 upperbit
= ((cpuid
>> 24) << 10) / assoc
;
438 va_align
.mask
= (upperbit
- 1) & PAGE_MASK
;
439 va_align
.flags
= ALIGN_VA_32
| ALIGN_VA_64
;
443 static void __cpuinit
early_init_amd(struct cpuinfo_x86
*c
)
445 early_init_amd_mc(c
);
448 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
449 * with P/T states and does not stop in deep C-states
451 if (c
->x86_power
& (1 << 8)) {
452 set_cpu_cap(c
, X86_FEATURE_CONSTANT_TSC
);
453 set_cpu_cap(c
, X86_FEATURE_NONSTOP_TSC
);
457 set_cpu_cap(c
, X86_FEATURE_SYSCALL32
);
459 /* Set MTRR capability flag if appropriate */
461 if (c
->x86_model
== 13 || c
->x86_model
== 9 ||
462 (c
->x86_model
== 8 && c
->x86_mask
>= 8))
463 set_cpu_cap(c
, X86_FEATURE_K6_MTRR
);
465 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
466 /* check CPU config space for extended APIC ID */
467 if (cpu_has_apic
&& c
->x86
>= 0xf) {
469 val
= read_pci_config(0, 24, 0, 0x68);
470 if ((val
& ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
471 set_cpu_cap(c
, X86_FEATURE_EXTD_APICID
);
476 static void __cpuinit
init_amd(struct cpuinfo_x86
*c
)
479 unsigned long long value
;
482 * Disable TLB flush filter by setting HWCR.FFDIS on K8
483 * bit 6 of msr C001_0015
485 * Errata 63 for SH-B3 steppings
486 * Errata 122 for all steppings (F+ have it disabled by default)
489 rdmsrl(MSR_K7_HWCR
, value
);
491 wrmsrl(MSR_K7_HWCR
, value
);
498 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
499 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
501 clear_cpu_cap(c
, 0*32+31);
504 /* On C+ stepping K8 rep microcode works well for copy/memset */
508 level
= cpuid_eax(1);
509 if ((level
>= 0x0f48 && level
< 0x0f50) || level
>= 0x0f58)
510 set_cpu_cap(c
, X86_FEATURE_REP_GOOD
);
513 * Some BIOSes incorrectly force this feature, but only K8
514 * revision D (model = 0x14) and later actually support it.
515 * (AMD Erratum #110, docId: 25759).
517 if (c
->x86_model
< 0x14 && cpu_has(c
, X86_FEATURE_LAHF_LM
)) {
520 clear_cpu_cap(c
, X86_FEATURE_LAHF_LM
);
521 if (!rdmsrl_amd_safe(0xc001100d, &val
)) {
522 val
&= ~(1ULL << 32);
523 wrmsrl_amd_safe(0xc001100d, val
);
529 set_cpu_cap(c
, X86_FEATURE_REP_GOOD
);
531 /* get apicid instead of initial apic id from cpuid */
532 c
->apicid
= hard_smp_processor_id();
536 * FIXME: We should handle the K5 here. Set up the write
537 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
548 case 6: /* An Athlon/Duron */
553 /* K6s reports MCEs but don't actually have all the MSRs */
555 clear_cpu_cap(c
, X86_FEATURE_MCE
);
558 /* Enable workaround for FXSAVE leak */
560 set_cpu_cap(c
, X86_FEATURE_FXSAVE_LEAK
);
562 if (!c
->x86_model_id
[0]) {
565 /* Should distinguish Models here, but this is only
566 a fallback anyways. */
567 strcpy(c
->x86_model_id
, "Hammer");
572 cpu_detect_cache_sizes(c
);
574 /* Multi core CPU? */
575 if (c
->extended_cpuid_level
>= 0x80000008) {
584 if (c
->extended_cpuid_level
>= 0x80000006) {
585 if (cpuid_edx(0x80000006) & 0xf000)
586 num_cache_leaves
= 4;
588 num_cache_leaves
= 3;
592 set_cpu_cap(c
, X86_FEATURE_K8
);
595 /* MFENCE stops RDTSC speculation */
596 set_cpu_cap(c
, X86_FEATURE_MFENCE_RDTSC
);
600 if (c
->x86
== 0x10) {
601 /* do this for boot cpu */
602 if (c
== &boot_cpu_data
)
603 check_enable_amd_mmconf_dmi();
605 fam10h_check_enable_mmcfg();
608 if (c
== &boot_cpu_data
&& c
->x86
>= 0xf) {
609 unsigned long long tseg
;
612 * Split up direct mapping around the TSEG SMM area.
613 * Don't do it for gbpages because there seems very little
614 * benefit in doing so.
616 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR
, &tseg
)) {
617 printk(KERN_DEBUG
"tseg: %010llx\n", tseg
);
618 if ((tseg
>>PMD_SHIFT
) <
619 (max_low_pfn_mapped
>>(PMD_SHIFT
-PAGE_SHIFT
)) ||
621 (max_pfn_mapped
>>(PMD_SHIFT
-PAGE_SHIFT
)) &&
622 (tseg
>>PMD_SHIFT
) >= (1ULL<<(32 - PMD_SHIFT
))))
623 set_memory_4k((unsigned long)__va(tseg
), 1);
629 * Family 0x12 and above processors have APIC timer
630 * running in deep C states.
633 set_cpu_cap(c
, X86_FEATURE_ARAT
);
636 * Disable GART TLB Walk Errors on Fam10h. We do this here
637 * because this is always needed when GART is enabled, even in a
638 * kernel which has no MCE support built in.
640 if (c
->x86
== 0x10) {
642 * BIOS should disable GartTlbWlk Errors themself. If
643 * it doesn't do it here as suggested by the BKDG.
645 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
650 err
= rdmsrl_safe(MSR_AMD64_MCx_MASK(4), &mask
);
653 checking_wrmsrl(MSR_AMD64_MCx_MASK(4), mask
);
659 static unsigned int __cpuinit
amd_size_cache(struct cpuinfo_x86
*c
,
662 /* AMD errata T13 (order #21922) */
665 if (c
->x86_model
== 3 && c
->x86_mask
== 0)
667 /* Tbird rev A1/A2 */
668 if (c
->x86_model
== 4 &&
669 (c
->x86_mask
== 0 || c
->x86_mask
== 1))
676 static const struct cpu_dev __cpuinitconst amd_cpu_dev
= {
678 .c_ident
= { "AuthenticAMD" },
681 { .vendor
= X86_VENDOR_AMD
, .family
= 4, .model_names
=
692 .c_size_cache
= amd_size_cache
,
694 .c_early_init
= early_init_amd
,
695 .c_bsp_init
= bsp_init_amd
,
697 .c_x86_vendor
= X86_VENDOR_AMD
,
700 cpu_dev_register(amd_cpu_dev
);
703 * AMD errata checking
705 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
706 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
707 * have an OSVW id assigned, which it takes as first argument. Both take a
708 * variable number of family-specific model-stepping ranges created by
709 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
710 * int[] in arch/x86/include/asm/processor.h.
714 * const int amd_erratum_319[] =
715 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
716 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
717 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
720 const int amd_erratum_400
[] =
721 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
722 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
723 EXPORT_SYMBOL_GPL(amd_erratum_400
);
725 const int amd_erratum_383
[] =
726 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
727 EXPORT_SYMBOL_GPL(amd_erratum_383
);
729 bool cpu_has_amd_erratum(const int *erratum
)
731 struct cpuinfo_x86
*cpu
= __this_cpu_ptr(&cpu_info
);
732 int osvw_id
= *erratum
++;
737 * If called early enough that current_cpu_data hasn't been initialized
738 * yet, fall back to boot_cpu_data.
741 cpu
= &boot_cpu_data
;
743 if (cpu
->x86_vendor
!= X86_VENDOR_AMD
)
746 if (osvw_id
>= 0 && osvw_id
< 65536 &&
747 cpu_has(cpu
, X86_FEATURE_OSVW
)) {
750 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH
, osvw_len
);
751 if (osvw_id
< osvw_len
) {
754 rdmsrl(MSR_AMD64_OSVW_STATUS
+ (osvw_id
>> 6),
756 return osvw_bits
& (1ULL << (osvw_id
& 0x3f));
760 /* OSVW unavailable or ID unknown, match family-model-stepping range */
761 ms
= (cpu
->x86_model
<< 4) | cpu
->x86_mask
;
762 while ((range
= *erratum
++))
763 if ((cpu
->x86
== AMD_MODEL_RANGE_FAMILY(range
)) &&
764 (ms
>= AMD_MODEL_RANGE_START(range
)) &&
765 (ms
<= AMD_MODEL_RANGE_END(range
)))
771 EXPORT_SYMBOL_GPL(cpu_has_amd_erratum
);