x86, tsc: Remove CPU frequency calibration on AMD
[linux-2.6/btrfs-unstable.git] / arch / x86 / kernel / cpu / amd.c
blobfc563fabde6731d0dded6ffe6532fedfbb539174
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 if ((c->x86 == 0x10) && (c->x86_model == 9))
309 amd_fixup_dcm(c);
310 #endif
313 int amd_get_nb_id(int cpu)
315 int id = 0;
316 #ifdef CONFIG_SMP
317 id = per_cpu(cpu_llc_id, cpu);
318 #endif
319 return id;
321 EXPORT_SYMBOL_GPL(amd_get_nb_id);
323 static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
325 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
326 int cpu = smp_processor_id();
327 int node;
328 unsigned apicid = c->apicid;
330 node = per_cpu(cpu_llc_id, cpu);
332 if (apicid_to_node[apicid] != NUMA_NO_NODE)
333 node = apicid_to_node[apicid];
334 if (!node_online(node)) {
335 /* Two possibilities here:
336 - The CPU is missing memory and no node was created.
337 In that case try picking one from a nearby CPU
338 - The APIC IDs differ from the HyperTransport node IDs
339 which the K8 northbridge parsing fills in.
340 Assume they are all increased by a constant offset,
341 but in the same order as the HT nodeids.
342 If that doesn't result in a usable node fall back to the
343 path for the previous case. */
345 int ht_nodeid = c->initial_apicid;
347 if (ht_nodeid >= 0 &&
348 apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
349 node = apicid_to_node[ht_nodeid];
350 /* Pick a nearby node */
351 if (!node_online(node))
352 node = nearby_node(apicid);
354 numa_set_node(cpu, node);
355 #endif
358 static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
360 #ifdef CONFIG_X86_HT
361 unsigned bits, ecx;
363 /* Multi core CPU? */
364 if (c->extended_cpuid_level < 0x80000008)
365 return;
367 ecx = cpuid_ecx(0x80000008);
369 c->x86_max_cores = (ecx & 0xff) + 1;
371 /* CPU telling us the core id bits shift? */
372 bits = (ecx >> 12) & 0xF;
374 /* Otherwise recompute */
375 if (bits == 0) {
376 while ((1 << bits) < c->x86_max_cores)
377 bits++;
380 c->x86_coreid_bits = bits;
381 #endif
384 static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
386 early_init_amd_mc(c);
389 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
390 * with P/T states and does not stop in deep C-states
392 if (c->x86_power & (1 << 8)) {
393 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
394 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
397 #ifdef CONFIG_X86_64
398 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
399 #else
400 /* Set MTRR capability flag if appropriate */
401 if (c->x86 == 5)
402 if (c->x86_model == 13 || c->x86_model == 9 ||
403 (c->x86_model == 8 && c->x86_mask >= 8))
404 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
405 #endif
406 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
407 /* check CPU config space for extended APIC ID */
408 if (cpu_has_apic && c->x86 >= 0xf) {
409 unsigned int val;
410 val = read_pci_config(0, 24, 0, 0x68);
411 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
412 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
414 #endif
416 /* We need to do the following only once */
417 if (c != &boot_cpu_data)
418 return;
420 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
422 if (c->x86 > 0x10 ||
423 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
424 u64 val;
426 rdmsrl(MSR_K7_HWCR, val);
427 if (!(val & BIT(24)))
428 printk(KERN_WARNING FW_BUG "TSC doesn't count "
429 "with P0 frequency!\n");
434 static void __cpuinit init_amd(struct cpuinfo_x86 *c)
436 #ifdef CONFIG_SMP
437 unsigned long long value;
440 * Disable TLB flush filter by setting HWCR.FFDIS on K8
441 * bit 6 of msr C001_0015
443 * Errata 63 for SH-B3 steppings
444 * Errata 122 for all steppings (F+ have it disabled by default)
446 if (c->x86 == 0xf) {
447 rdmsrl(MSR_K7_HWCR, value);
448 value |= 1 << 6;
449 wrmsrl(MSR_K7_HWCR, value);
451 #endif
453 early_init_amd(c);
456 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
457 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
459 clear_cpu_cap(c, 0*32+31);
461 #ifdef CONFIG_X86_64
462 /* On C+ stepping K8 rep microcode works well for copy/memset */
463 if (c->x86 == 0xf) {
464 u32 level;
466 level = cpuid_eax(1);
467 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
468 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
471 * Some BIOSes incorrectly force this feature, but only K8
472 * revision D (model = 0x14) and later actually support it.
473 * (AMD Erratum #110, docId: 25759).
475 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
476 u64 val;
478 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
479 if (!rdmsrl_amd_safe(0xc001100d, &val)) {
480 val &= ~(1ULL << 32);
481 wrmsrl_amd_safe(0xc001100d, val);
486 if (c->x86 >= 0x10)
487 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
489 /* get apicid instead of initial apic id from cpuid */
490 c->apicid = hard_smp_processor_id();
491 #else
494 * FIXME: We should handle the K5 here. Set up the write
495 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
496 * no bus pipeline)
499 switch (c->x86) {
500 case 4:
501 init_amd_k5(c);
502 break;
503 case 5:
504 init_amd_k6(c);
505 break;
506 case 6: /* An Athlon/Duron */
507 init_amd_k7(c);
508 break;
511 /* K6s reports MCEs but don't actually have all the MSRs */
512 if (c->x86 < 6)
513 clear_cpu_cap(c, X86_FEATURE_MCE);
514 #endif
516 /* Enable workaround for FXSAVE leak */
517 if (c->x86 >= 6)
518 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
520 if (!c->x86_model_id[0]) {
521 switch (c->x86) {
522 case 0xf:
523 /* Should distinguish Models here, but this is only
524 a fallback anyways. */
525 strcpy(c->x86_model_id, "Hammer");
526 break;
530 cpu_detect_cache_sizes(c);
532 /* Multi core CPU? */
533 if (c->extended_cpuid_level >= 0x80000008) {
534 amd_detect_cmp(c);
535 srat_detect_node(c);
538 #ifdef CONFIG_X86_32
539 detect_ht(c);
540 #endif
542 if (c->extended_cpuid_level >= 0x80000006) {
543 if ((c->x86 >= 0x0f) && (cpuid_edx(0x80000006) & 0xf000))
544 num_cache_leaves = 4;
545 else
546 num_cache_leaves = 3;
549 if (c->x86 >= 0xf)
550 set_cpu_cap(c, X86_FEATURE_K8);
552 if (cpu_has_xmm2) {
553 /* MFENCE stops RDTSC speculation */
554 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
557 #ifdef CONFIG_X86_64
558 if (c->x86 == 0x10) {
559 /* do this for boot cpu */
560 if (c == &boot_cpu_data)
561 check_enable_amd_mmconf_dmi();
563 fam10h_check_enable_mmcfg();
566 if (c == &boot_cpu_data && c->x86 >= 0xf) {
567 unsigned long long tseg;
570 * Split up direct mapping around the TSEG SMM area.
571 * Don't do it for gbpages because there seems very little
572 * benefit in doing so.
574 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
575 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
576 if ((tseg>>PMD_SHIFT) <
577 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
578 ((tseg>>PMD_SHIFT) <
579 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
580 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
581 set_memory_4k((unsigned long)__va(tseg), 1);
584 #endif
587 #ifdef CONFIG_X86_32
588 static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
589 unsigned int size)
591 /* AMD errata T13 (order #21922) */
592 if ((c->x86 == 6)) {
593 /* Duron Rev A0 */
594 if (c->x86_model == 3 && c->x86_mask == 0)
595 size = 64;
596 /* Tbird rev A1/A2 */
597 if (c->x86_model == 4 &&
598 (c->x86_mask == 0 || c->x86_mask == 1))
599 size = 256;
601 return size;
603 #endif
605 static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
606 .c_vendor = "AMD",
607 .c_ident = { "AuthenticAMD" },
608 #ifdef CONFIG_X86_32
609 .c_models = {
610 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
612 [3] = "486 DX/2",
613 [7] = "486 DX/2-WB",
614 [8] = "486 DX/4",
615 [9] = "486 DX/4-WB",
616 [14] = "Am5x86-WT",
617 [15] = "Am5x86-WB"
621 .c_size_cache = amd_size_cache,
622 #endif
623 .c_early_init = early_init_amd,
624 .c_init = init_amd,
625 .c_x86_vendor = X86_VENDOR_AMD,
628 cpu_dev_register(amd_cpu_dev);
631 * AMD errata checking
633 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
634 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
635 * have an OSVW id assigned, which it takes as first argument. Both take a
636 * variable number of family-specific model-stepping ranges created by
637 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
638 * int[] in arch/x86/include/asm/processor.h.
640 * Example:
642 * const int amd_erratum_319[] =
643 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
644 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
645 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
648 const int amd_erratum_400[] =
649 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
650 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
651 EXPORT_SYMBOL_GPL(amd_erratum_400);
653 const int amd_erratum_383[] =
654 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
655 EXPORT_SYMBOL_GPL(amd_erratum_383);
657 bool cpu_has_amd_erratum(const int *erratum)
659 struct cpuinfo_x86 *cpu = &current_cpu_data;
660 int osvw_id = *erratum++;
661 u32 range;
662 u32 ms;
665 * If called early enough that current_cpu_data hasn't been initialized
666 * yet, fall back to boot_cpu_data.
668 if (cpu->x86 == 0)
669 cpu = &boot_cpu_data;
671 if (cpu->x86_vendor != X86_VENDOR_AMD)
672 return false;
674 if (osvw_id >= 0 && osvw_id < 65536 &&
675 cpu_has(cpu, X86_FEATURE_OSVW)) {
676 u64 osvw_len;
678 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
679 if (osvw_id < osvw_len) {
680 u64 osvw_bits;
682 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
683 osvw_bits);
684 return osvw_bits & (1ULL << (osvw_id & 0x3f));
688 /* OSVW unavailable or ID unknown, match family-model-stepping range */
689 ms = (cpu->x86_model << 4) | cpu->x86_mask;
690 while ((range = *erratum++))
691 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
692 (ms >= AMD_MODEL_RANGE_START(range)) &&
693 (ms <= AMD_MODEL_RANGE_END(range)))
694 return true;
696 return false;
699 EXPORT_SYMBOL_GPL(cpu_has_amd_erratum);