x86, cpu: Fix regression in AMD errata checking code
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / arch / x86 / kernel / cpu / amd.c
blobb203d0dbfc0494f0e9fb8f8d22e286bceb096e3a
1 #include <linux/init.h>
2 #include <linux/bitops.h>
3 #include <linux/mm.h>
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>
11 #ifdef CONFIG_X86_64
12 # include <asm/numa_64.h>
13 # include <asm/mmconfig.h>
14 # include <asm/cacheflush.h>
15 #endif
17 #include "cpu.h"
19 #ifdef CONFIG_X86_32
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.
25 * See http://www.multimania.com/poulot/k6bug.html
26 * http://www.amd.com/K6/k6docs/revgd.html
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..
33 extern void vide(void);
34 __asm__(".align 4\nvide: ret");
36 static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
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.
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);
54 static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
56 u32 l, h;
57 int mbytes = num_physpages >> (20-PAGE_SHIFT);
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);
65 return;
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;
74 printk(KERN_INFO "AMD K6 stepping B detected - ");
77 * It looks like AMD fixed the 2.6.2 bug and improved indirect
78 * calls at the same time.
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;
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");
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;
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);
115 return;
118 if ((c->x86_model == 8 && c->x86_mask > 7) ||
119 c->x86_model == 9 || c->x86_model == 13) {
120 /* The more serious chips .. */
122 if (mbytes > 4092)
123 mbytes = 4092;
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);
137 return;
140 if (c->x86_model == 10) {
141 /* AMD Geode LX is model 10 */
142 /* placeholder for any needed mods */
143 return;
147 static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
149 #ifdef CONFIG_SMP
150 /* calling is from identify_secondary_cpu() ? */
151 if (c->cpu_index == boot_cpu_id)
152 return;
155 * Certain Athlons might work (for various values of 'work') in SMP
156 * but they are not certified as MP capable.
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;
163 /* Duron 670 is valid */
164 if ((c->x86_model == 7) && (c->x86_mask == 0))
165 goto valid_k7;
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.
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;
180 /* If we get here, not a certified SMP capable AMD system. */
183 * Don't taint if we are running SMP kernel on a single non-MP
184 * approved Athlon
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);
191 valid_k7:
193 #endif
196 static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
198 u32 l, h;
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.
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);
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
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);
230 set_cpu_cap(c, X86_FEATURE_K7);
232 amd_k7_smp_check(c);
234 #endif
236 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
237 static int __cpuinit nearby_node(int apicid)
239 int i, node;
241 for (i = apicid - 1; i >= 0; i--) {
242 node = apicid_to_node[i];
243 if (node != NUMA_NO_NODE && node_online(node))
244 return node;
246 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
247 node = apicid_to_node[i];
248 if (node != NUMA_NO_NODE && node_online(node))
249 return node;
251 return first_node(node_online_map); /* Shouldn't happen */
253 #endif
256 * Fixup core topology information for AMD multi-node processors.
257 * Assumption: Number of cores in each internal node is the same.
259 #ifdef CONFIG_X86_HT
260 static void __cpuinit amd_fixup_dcm(struct cpuinfo_x86 *c)
262 unsigned long long value;
263 u32 nodes, cores_per_node;
264 int cpu = smp_processor_id();
266 if (!cpu_has(c, X86_FEATURE_NODEID_MSR))
267 return;
269 /* fixup topology information only once for a core */
270 if (cpu_has(c, X86_FEATURE_AMD_DCM))
271 return;
273 rdmsrl(MSR_FAM10H_NODE_ID, value);
275 nodes = ((value >> 3) & 7) + 1;
276 if (nodes == 1)
277 return;
279 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
280 cores_per_node = c->x86_max_cores / nodes;
282 /* store NodeID, use llc_shared_map to store sibling info */
283 per_cpu(cpu_llc_id, cpu) = value & 7;
285 /* fixup core id to be in range from 0 to (cores_per_node - 1) */
286 c->cpu_core_id = c->cpu_core_id % cores_per_node;
288 #endif
291 * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
292 * Assumes number of cores is a power of two.
294 static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
296 #ifdef CONFIG_X86_HT
297 unsigned bits;
298 int cpu = smp_processor_id();
300 bits = c->x86_coreid_bits;
301 /* Low order bits define the core id (index of core in socket) */
302 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
303 /* Convert the initial APIC ID into the socket ID */
304 c->phys_proc_id = c->initial_apicid >> bits;
305 /* use socket ID also for last level cache */
306 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
307 /* fixup topology information on multi-node processors */
308 amd_fixup_dcm(c);
309 #endif
312 int amd_get_nb_id(int cpu)
314 int id = 0;
315 #ifdef CONFIG_SMP
316 id = per_cpu(cpu_llc_id, cpu);
317 #endif
318 return id;
320 EXPORT_SYMBOL_GPL(amd_get_nb_id);
322 static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
324 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
325 int cpu = smp_processor_id();
326 int node;
327 unsigned apicid = c->apicid;
329 node = per_cpu(cpu_llc_id, cpu);
331 if (apicid_to_node[apicid] != NUMA_NO_NODE)
332 node = apicid_to_node[apicid];
333 if (!node_online(node)) {
334 /* Two possibilities here:
335 - The CPU is missing memory and no node was created.
336 In that case try picking one from a nearby CPU
337 - The APIC IDs differ from the HyperTransport node IDs
338 which the K8 northbridge parsing fills in.
339 Assume they are all increased by a constant offset,
340 but in the same order as the HT nodeids.
341 If that doesn't result in a usable node fall back to the
342 path for the previous case. */
344 int ht_nodeid = c->initial_apicid;
346 if (ht_nodeid >= 0 &&
347 apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
348 node = apicid_to_node[ht_nodeid];
349 /* Pick a nearby node */
350 if (!node_online(node))
351 node = nearby_node(apicid);
353 numa_set_node(cpu, node);
354 #endif
357 static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
359 #ifdef CONFIG_X86_HT
360 unsigned bits, ecx;
362 /* Multi core CPU? */
363 if (c->extended_cpuid_level < 0x80000008)
364 return;
366 ecx = cpuid_ecx(0x80000008);
368 c->x86_max_cores = (ecx & 0xff) + 1;
370 /* CPU telling us the core id bits shift? */
371 bits = (ecx >> 12) & 0xF;
373 /* Otherwise recompute */
374 if (bits == 0) {
375 while ((1 << bits) < c->x86_max_cores)
376 bits++;
379 c->x86_coreid_bits = bits;
380 #endif
383 static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
385 early_init_amd_mc(c);
388 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
389 * with P/T states and does not stop in deep C-states
391 if (c->x86_power & (1 << 8)) {
392 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
393 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
396 #ifdef CONFIG_X86_64
397 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
398 #else
399 /* Set MTRR capability flag if appropriate */
400 if (c->x86 == 5)
401 if (c->x86_model == 13 || c->x86_model == 9 ||
402 (c->x86_model == 8 && c->x86_mask >= 8))
403 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
404 #endif
405 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
406 /* check CPU config space for extended APIC ID */
407 if (cpu_has_apic && c->x86 >= 0xf) {
408 unsigned int val;
409 val = read_pci_config(0, 24, 0, 0x68);
410 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
411 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
413 #endif
416 static void __cpuinit init_amd(struct cpuinfo_x86 *c)
418 #ifdef CONFIG_SMP
419 unsigned long long value;
422 * Disable TLB flush filter by setting HWCR.FFDIS on K8
423 * bit 6 of msr C001_0015
425 * Errata 63 for SH-B3 steppings
426 * Errata 122 for all steppings (F+ have it disabled by default)
428 if (c->x86 == 0xf) {
429 rdmsrl(MSR_K7_HWCR, value);
430 value |= 1 << 6;
431 wrmsrl(MSR_K7_HWCR, value);
433 #endif
435 early_init_amd(c);
438 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
439 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
441 clear_cpu_cap(c, 0*32+31);
443 #ifdef CONFIG_X86_64
444 /* On C+ stepping K8 rep microcode works well for copy/memset */
445 if (c->x86 == 0xf) {
446 u32 level;
448 level = cpuid_eax(1);
449 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
450 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
453 * Some BIOSes incorrectly force this feature, but only K8
454 * revision D (model = 0x14) and later actually support it.
455 * (AMD Erratum #110, docId: 25759).
457 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
458 u64 val;
460 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
461 if (!rdmsrl_amd_safe(0xc001100d, &val)) {
462 val &= ~(1ULL << 32);
463 wrmsrl_amd_safe(0xc001100d, val);
468 if (c->x86 == 0x10 || c->x86 == 0x11)
469 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
471 /* get apicid instead of initial apic id from cpuid */
472 c->apicid = hard_smp_processor_id();
473 #else
476 * FIXME: We should handle the K5 here. Set up the write
477 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
478 * no bus pipeline)
481 switch (c->x86) {
482 case 4:
483 init_amd_k5(c);
484 break;
485 case 5:
486 init_amd_k6(c);
487 break;
488 case 6: /* An Athlon/Duron */
489 init_amd_k7(c);
490 break;
493 /* K6s reports MCEs but don't actually have all the MSRs */
494 if (c->x86 < 6)
495 clear_cpu_cap(c, X86_FEATURE_MCE);
496 #endif
498 /* Enable workaround for FXSAVE leak */
499 if (c->x86 >= 6)
500 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
502 if (!c->x86_model_id[0]) {
503 switch (c->x86) {
504 case 0xf:
505 /* Should distinguish Models here, but this is only
506 a fallback anyways. */
507 strcpy(c->x86_model_id, "Hammer");
508 break;
512 cpu_detect_cache_sizes(c);
514 /* Multi core CPU? */
515 if (c->extended_cpuid_level >= 0x80000008) {
516 amd_detect_cmp(c);
517 srat_detect_node(c);
520 #ifdef CONFIG_X86_32
521 detect_ht(c);
522 #endif
524 if (c->extended_cpuid_level >= 0x80000006) {
525 if ((c->x86 >= 0x0f) && (cpuid_edx(0x80000006) & 0xf000))
526 num_cache_leaves = 4;
527 else
528 num_cache_leaves = 3;
531 if (c->x86 >= 0xf && c->x86 <= 0x11)
532 set_cpu_cap(c, X86_FEATURE_K8);
534 if (cpu_has_xmm2) {
535 /* MFENCE stops RDTSC speculation */
536 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
539 #ifdef CONFIG_X86_64
540 if (c->x86 == 0x10) {
541 /* do this for boot cpu */
542 if (c == &boot_cpu_data)
543 check_enable_amd_mmconf_dmi();
545 fam10h_check_enable_mmcfg();
548 if (c == &boot_cpu_data && c->x86 >= 0xf && c->x86 <= 0x11) {
549 unsigned long long tseg;
552 * Split up direct mapping around the TSEG SMM area.
553 * Don't do it for gbpages because there seems very little
554 * benefit in doing so.
556 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
557 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
558 if ((tseg>>PMD_SHIFT) <
559 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
560 ((tseg>>PMD_SHIFT) <
561 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
562 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
563 set_memory_4k((unsigned long)__va(tseg), 1);
566 #endif
568 /* As a rule processors have APIC timer running in deep C states */
569 if (c->x86 >= 0xf && !cpu_has_amd_erratum(amd_erratum_400))
570 set_cpu_cap(c, X86_FEATURE_ARAT);
573 * Disable GART TLB Walk Errors on Fam10h. We do this here
574 * because this is always needed when GART is enabled, even in a
575 * kernel which has no MCE support built in.
577 if (c->x86 == 0x10) {
579 * BIOS should disable GartTlbWlk Errors themself. If
580 * it doesn't do it here as suggested by the BKDG.
582 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
584 u64 mask;
586 rdmsrl(MSR_AMD64_MCx_MASK(4), mask);
587 mask |= (1 << 10);
588 wrmsrl(MSR_AMD64_MCx_MASK(4), mask);
592 #ifdef CONFIG_X86_32
593 static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
594 unsigned int size)
596 /* AMD errata T13 (order #21922) */
597 if ((c->x86 == 6)) {
598 /* Duron Rev A0 */
599 if (c->x86_model == 3 && c->x86_mask == 0)
600 size = 64;
601 /* Tbird rev A1/A2 */
602 if (c->x86_model == 4 &&
603 (c->x86_mask == 0 || c->x86_mask == 1))
604 size = 256;
606 return size;
608 #endif
610 static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
611 .c_vendor = "AMD",
612 .c_ident = { "AuthenticAMD" },
613 #ifdef CONFIG_X86_32
614 .c_models = {
615 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
617 [3] = "486 DX/2",
618 [7] = "486 DX/2-WB",
619 [8] = "486 DX/4",
620 [9] = "486 DX/4-WB",
621 [14] = "Am5x86-WT",
622 [15] = "Am5x86-WB"
626 .c_size_cache = amd_size_cache,
627 #endif
628 .c_early_init = early_init_amd,
629 .c_init = init_amd,
630 .c_x86_vendor = X86_VENDOR_AMD,
633 cpu_dev_register(amd_cpu_dev);
636 * AMD errata checking
638 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
639 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
640 * have an OSVW id assigned, which it takes as first argument. Both take a
641 * variable number of family-specific model-stepping ranges created by
642 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
643 * int[] in arch/x86/include/asm/processor.h.
645 * Example:
647 * const int amd_erratum_319[] =
648 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
649 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
650 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
653 const int amd_erratum_400[] =
654 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
655 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
658 bool cpu_has_amd_erratum(const int *erratum)
660 struct cpuinfo_x86 *cpu = &current_cpu_data;
661 int osvw_id = *erratum++;
662 u32 range;
663 u32 ms;
666 * If called early enough that current_cpu_data hasn't been initialized
667 * yet, fall back to boot_cpu_data.
669 if (cpu->x86 == 0)
670 cpu = &boot_cpu_data;
672 if (cpu->x86_vendor != X86_VENDOR_AMD)
673 return false;
675 if (osvw_id >= 0 && osvw_id < 65536 &&
676 cpu_has(cpu, X86_FEATURE_OSVW)) {
677 u64 osvw_len;
679 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
680 if (osvw_id < osvw_len) {
681 u64 osvw_bits;
683 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
684 osvw_bits);
685 return osvw_bits & (1ULL << (osvw_id & 0x3f));
689 /* OSVW unavailable or ID unknown, match family-model-stepping range */
690 ms = (cpu->x86_model << 4) | cpu->x86_mask;
691 while ((range = *erratum++))
692 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
693 (ms >= AMD_MODEL_RANGE_START(range)) &&
694 (ms <= AMD_MODEL_RANGE_END(range)))
695 return true;
697 return false;