1 #include <linux/init.h>
2 #include <linux/bitops.h>
7 #include <asm/processor.h>
10 #include <asm/pci-direct.h>
13 # include <asm/numa_64.h>
14 # include <asm/mmconfig.h>
15 # include <asm/cacheflush.h>
22 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
23 * misexecution of code under Linux. Owners of such processors should
24 * contact AMD for precise details and a CPU swap.
26 * See http://www.multimania.com/poulot/k6bug.html
27 * http://www.amd.com/K6/k6docs/revgd.html
29 * The following test is erm.. interesting. AMD neglected to up
30 * the chip setting when fixing the bug but they also tweaked some
31 * performance at the same time..
34 extern void vide(void);
35 __asm__(".align 4\nvide: ret");
37 static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
40 * General Systems BIOSen alias the cpu frequency registers
41 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
42 * drivers subsequently pokes it, and changes the CPU speed.
43 * Workaround : Remove the unneeded alias.
45 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
46 #define CBAR_ENB (0x80000000)
47 #define CBAR_KEY (0X000000CB)
48 if (c->x86_model == 9 || c->x86_model == 10) {
49 if (inl(CBAR) & CBAR_ENB)
50 outl(0 | CBAR_KEY, CBAR);
55 static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
58 int mbytes = num_physpages >> (20-PAGE_SHIFT);
60 if (c->x86_model < 6) {
61 /* Based on AMD doc 20734R - June 2000 */
62 if (c->x86_model == 0) {
63 clear_cpu_cap(c, X86_FEATURE_APIC);
64 set_cpu_cap(c, X86_FEATURE_PGE);
69 if (c->x86_model == 6 && c->x86_mask == 1) {
70 const int K6_BUG_LOOP = 1000000;
75 printk(KERN_INFO "AMD K6 stepping B detected - ");
78 * It looks like AMD fixed the 2.6.2 bug and improved indirect
79 * calls at the same time.
90 if (d > 20*K6_BUG_LOOP)
92 "system stability may be impaired when more than 32 MB are used.\n");
94 printk(KERN_CONT "probably OK (after B9730xxxx).\n");
95 printk(KERN_INFO "Please see http://membres.lycos.fr/poulot/k6bug.html\n");
98 /* K6 with old style WHCR */
99 if (c->x86_model < 8 ||
100 (c->x86_model == 8 && c->x86_mask < 8)) {
101 /* We can only write allocate on the low 508Mb */
105 rdmsr(MSR_K6_WHCR, l, h);
106 if ((l&0x0000FFFF) == 0) {
108 l = (1<<0)|((mbytes/4)<<1);
109 local_irq_save(flags);
111 wrmsr(MSR_K6_WHCR, l, h);
112 local_irq_restore(flags);
113 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
119 if ((c->x86_model == 8 && c->x86_mask > 7) ||
120 c->x86_model == 9 || c->x86_model == 13) {
121 /* The more serious chips .. */
126 rdmsr(MSR_K6_WHCR, l, h);
127 if ((l&0xFFFF0000) == 0) {
129 l = ((mbytes>>2)<<22)|(1<<16);
130 local_irq_save(flags);
132 wrmsr(MSR_K6_WHCR, l, h);
133 local_irq_restore(flags);
134 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
141 if (c->x86_model == 10) {
142 /* AMD Geode LX is model 10 */
143 /* placeholder for any needed mods */
148 static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
151 /* calling is from identify_secondary_cpu() ? */
156 * Certain Athlons might work (for various values of 'work') in SMP
157 * but they are not certified as MP capable.
159 /* Athlon 660/661 is valid. */
160 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
164 /* Duron 670 is valid */
165 if ((c->x86_model == 7) && (c->x86_mask == 0))
169 * Athlon 662, Duron 671, and Athlon >model 7 have capability
170 * bit. It's worth noting that the A5 stepping (662) of some
171 * Athlon XP's have the MP bit set.
172 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
175 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
176 ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
181 /* If we get here, not a certified SMP capable AMD system. */
184 * Don't taint if we are running SMP kernel on a single non-MP
187 WARN_ONCE(1, "WARNING: This combination of AMD"
188 " processors is not suitable for SMP.\n");
189 if (!test_taint(TAINT_UNSAFE_SMP))
190 add_taint(TAINT_UNSAFE_SMP);
197 static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
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.
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);
211 wrmsr(MSR_K7_HWCR, l, h);
212 set_cpu_cap(c, X86_FEATURE_XMM);
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
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) {
225 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
226 l, ((l & 0x000fffff)|0x20000000));
227 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
231 set_cpu_cap(c, X86_FEATURE_K7);
239 * To workaround broken NUMA config. Read the comment in
240 * srat_detect_node().
242 static int __cpuinit nearby_node(int apicid)
246 for (i = apicid - 1; i >= 0; i--) {
247 node = __apicid_to_node[i];
248 if (node != NUMA_NO_NODE && node_online(node))
251 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
252 node = __apicid_to_node[i];
253 if (node != NUMA_NO_NODE && node_online(node))
256 return first_node(node_online_map); /* Shouldn't happen */
261 * Fixup core topology information for
262 * (1) AMD multi-node processors
263 * Assumption: Number of cores in each internal node is the same.
264 * (2) AMD processors supporting compute units
267 static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
269 u32 nodes, cores_per_cu = 1;
271 int cpu = smp_processor_id();
273 /* get information required for multi-node processors */
274 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
275 u32 eax, ebx, ecx, edx;
277 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
278 nodes = ((ecx >> 8) & 7) + 1;
281 /* get compute unit information */
282 smp_num_siblings = ((ebx >> 8) & 3) + 1;
283 c->compute_unit_id = ebx & 0xff;
284 cores_per_cu += ((ebx >> 8) & 3);
285 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
288 rdmsrl(MSR_FAM10H_NODE_ID, value);
289 nodes = ((value >> 3) & 7) + 1;
294 /* fixup multi-node processor information */
299 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
300 cores_per_node = c->x86_max_cores / nodes;
301 cus_per_node = cores_per_node / cores_per_cu;
303 /* store NodeID, use llc_shared_map to store sibling info */
304 per_cpu(cpu_llc_id, cpu) = node_id;
306 /* core id has to be in the [0 .. cores_per_node - 1] range */
307 c->cpu_core_id %= cores_per_node;
308 c->compute_unit_id %= cus_per_node;
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.
317 static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
321 int cpu = smp_processor_id();
323 bits = c->x86_coreid_bits;
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;
328 /* use socket ID also for last level cache */
329 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
334 int amd_get_nb_id(int cpu)
338 id = per_cpu(cpu_llc_id, cpu);
342 EXPORT_SYMBOL_GPL(amd_get_nb_id);
344 static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
347 int cpu = smp_processor_id();
349 unsigned apicid = c->apicid;
351 node = numa_cpu_node(cpu);
352 if (node == NUMA_NO_NODE)
353 node = per_cpu(cpu_llc_id, cpu);
355 if (!node_online(node)) {
357 * Two possibilities here:
359 * - The CPU is missing memory and no node was created. In
360 * that case try picking one from a nearby CPU.
362 * - The APIC IDs differ from the HyperTransport node IDs
363 * which the K8 northbridge parsing fills in. Assume
364 * they are all increased by a constant offset, but in
365 * the same order as the HT nodeids. If that doesn't
366 * result in a usable node fall back to the path for the
369 * This workaround operates directly on the mapping between
370 * APIC ID and NUMA node, assuming certain relationship
371 * between APIC ID, HT node ID and NUMA topology. As going
372 * through CPU mapping may alter the outcome, directly
373 * access __apicid_to_node[].
375 int ht_nodeid = c->initial_apicid;
377 if (ht_nodeid >= 0 &&
378 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
379 node = __apicid_to_node[ht_nodeid];
380 /* Pick a nearby node */
381 if (!node_online(node))
382 node = nearby_node(apicid);
384 numa_set_node(cpu, node);
388 static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
393 /* Multi core CPU? */
394 if (c->extended_cpuid_level < 0x80000008)
397 ecx = cpuid_ecx(0x80000008);
399 c->x86_max_cores = (ecx & 0xff) + 1;
401 /* CPU telling us the core id bits shift? */
402 bits = (ecx >> 12) & 0xF;
404 /* Otherwise recompute */
406 while ((1 << bits) < c->x86_max_cores)
410 c->x86_coreid_bits = bits;
414 static void __cpuinit bsp_init_amd(struct cpuinfo_x86 *c)
416 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
419 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
422 rdmsrl(MSR_K7_HWCR, val);
423 if (!(val & BIT(24)))
424 printk(KERN_WARNING FW_BUG "TSC doesn't count "
425 "with P0 frequency!\n");
429 if (c->x86 == 0x15) {
430 unsigned long upperbit;
433 cpuid = cpuid_edx(0x80000005);
434 assoc = cpuid >> 16 & 0xff;
435 upperbit = ((cpuid >> 24) << 10) / assoc;
437 va_align.mask = (upperbit - 1) & PAGE_MASK;
438 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
442 static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
444 early_init_amd_mc(c);
447 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
448 * with P/T states and does not stop in deep C-states
450 if (c->x86_power & (1 << 8)) {
451 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
452 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
456 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
458 /* Set MTRR capability flag if appropriate */
460 if (c->x86_model == 13 || c->x86_model == 9 ||
461 (c->x86_model == 8 && c->x86_mask >= 8))
462 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
464 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
465 /* check CPU config space for extended APIC ID */
466 if (cpu_has_apic && c->x86 >= 0xf) {
468 val = read_pci_config(0, 24, 0, 0x68);
469 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
470 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
475 static void __cpuinit init_amd(struct cpuinfo_x86 *c)
478 unsigned long long value;
481 * Disable TLB flush filter by setting HWCR.FFDIS on K8
482 * bit 6 of msr C001_0015
484 * Errata 63 for SH-B3 steppings
485 * Errata 122 for all steppings (F+ have it disabled by default)
488 rdmsrl(MSR_K7_HWCR, value);
490 wrmsrl(MSR_K7_HWCR, value);
497 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
498 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
500 clear_cpu_cap(c, 0*32+31);
503 /* On C+ stepping K8 rep microcode works well for copy/memset */
507 level = cpuid_eax(1);
508 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
509 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
512 * Some BIOSes incorrectly force this feature, but only K8
513 * revision D (model = 0x14) and later actually support it.
514 * (AMD Erratum #110, docId: 25759).
516 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
519 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
520 if (!rdmsrl_amd_safe(0xc001100d, &val)) {
521 val &= ~(1ULL << 32);
522 wrmsrl_amd_safe(0xc001100d, val);
528 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
530 /* get apicid instead of initial apic id from cpuid */
531 c->apicid = hard_smp_processor_id();
535 * FIXME: We should handle the K5 here. Set up the write
536 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
547 case 6: /* An Athlon/Duron */
552 /* K6s reports MCEs but don't actually have all the MSRs */
554 clear_cpu_cap(c, X86_FEATURE_MCE);
557 /* Enable workaround for FXSAVE leak */
559 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
561 if (!c->x86_model_id[0]) {
564 /* Should distinguish Models here, but this is only
565 a fallback anyways. */
566 strcpy(c->x86_model_id, "Hammer");
571 cpu_detect_cache_sizes(c);
573 /* Multi core CPU? */
574 if (c->extended_cpuid_level >= 0x80000008) {
583 if (c->extended_cpuid_level >= 0x80000006) {
584 if (cpuid_edx(0x80000006) & 0xf000)
585 num_cache_leaves = 4;
587 num_cache_leaves = 3;
591 set_cpu_cap(c, X86_FEATURE_K8);
594 /* MFENCE stops RDTSC speculation */
595 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
599 if (c->x86 == 0x10) {
600 /* do this for boot cpu */
601 if (c == &boot_cpu_data)
602 check_enable_amd_mmconf_dmi();
604 fam10h_check_enable_mmcfg();
607 if (c == &boot_cpu_data && c->x86 >= 0xf) {
608 unsigned long long tseg;
611 * Split up direct mapping around the TSEG SMM area.
612 * Don't do it for gbpages because there seems very little
613 * benefit in doing so.
615 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
616 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
617 if ((tseg>>PMD_SHIFT) <
618 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
620 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
621 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
622 set_memory_4k((unsigned long)__va(tseg), 1);
628 * Family 0x12 and above processors have APIC timer
629 * running in deep C states.
632 set_cpu_cap(c, X86_FEATURE_ARAT);
635 * Disable GART TLB Walk Errors on Fam10h. We do this here
636 * because this is always needed when GART is enabled, even in a
637 * kernel which has no MCE support built in.
639 if (c->x86 == 0x10) {
641 * BIOS should disable GartTlbWlk Errors themself. If
642 * it doesn't do it here as suggested by the BKDG.
644 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
649 err = rdmsrl_safe(MSR_AMD64_MCx_MASK(4), &mask);
652 checking_wrmsrl(MSR_AMD64_MCx_MASK(4), mask);
658 static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
661 /* AMD errata T13 (order #21922) */
664 if (c->x86_model == 3 && c->x86_mask == 0)
666 /* Tbird rev A1/A2 */
667 if (c->x86_model == 4 &&
668 (c->x86_mask == 0 || c->x86_mask == 1))
675 static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
677 .c_ident = { "AuthenticAMD" },
680 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
691 .c_size_cache = amd_size_cache,
693 .c_early_init = early_init_amd,
694 .c_bsp_init = bsp_init_amd,
696 .c_x86_vendor = X86_VENDOR_AMD,
699 cpu_dev_register(amd_cpu_dev);
702 * AMD errata checking
704 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
705 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
706 * have an OSVW id assigned, which it takes as first argument. Both take a
707 * variable number of family-specific model-stepping ranges created by
708 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
709 * int[] in arch/x86/include/asm/processor.h.
713 * const int amd_erratum_319[] =
714 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
715 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
716 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
719 const int amd_erratum_400[] =
720 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
721 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
722 EXPORT_SYMBOL_GPL(amd_erratum_400);
724 const int amd_erratum_383[] =
725 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
726 EXPORT_SYMBOL_GPL(amd_erratum_383);
728 bool cpu_has_amd_erratum(const int *erratum)
730 struct cpuinfo_x86 *cpu = __this_cpu_ptr(&cpu_info);
731 int osvw_id = *erratum++;
736 * If called early enough that current_cpu_data hasn't been initialized
737 * yet, fall back to boot_cpu_data.
740 cpu = &boot_cpu_data;
742 if (cpu->x86_vendor != X86_VENDOR_AMD)
745 if (osvw_id >= 0 && osvw_id < 65536 &&
746 cpu_has(cpu, X86_FEATURE_OSVW)) {
749 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
750 if (osvw_id < osvw_len) {
753 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
755 return osvw_bits & (1ULL << (osvw_id & 0x3f));
759 /* OSVW unavailable or ID unknown, match family-model-stepping range */
760 ms = (cpu->x86_model << 4) | cpu->x86_mask;
761 while ((range = *erratum++))
762 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
763 (ms >= AMD_MODEL_RANGE_START(range)) &&
764 (ms <= AMD_MODEL_RANGE_END(range)))
770 EXPORT_SYMBOL_GPL(cpu_has_amd_erratum);