Loading...
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 <linux/sched/clock.h>
9#include <linux/random.h>
10#include <asm/processor.h>
11#include <asm/apic.h>
12#include <asm/cpu.h>
13#include <asm/spec-ctrl.h>
14#include <asm/smp.h>
15#include <asm/pci-direct.h>
16#include <asm/delay.h>
17
18#ifdef CONFIG_X86_64
19# include <asm/mmconfig.h>
20# include <asm/set_memory.h>
21#endif
22
23#include "cpu.h"
24
25static const int amd_erratum_383[];
26static const int amd_erratum_400[];
27static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
28
29/*
30 * nodes_per_socket: Stores the number of nodes per socket.
31 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
32 * Node Identifiers[10:8]
33 */
34static u32 nodes_per_socket = 1;
35
36static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
37{
38 u32 gprs[8] = { 0 };
39 int err;
40
41 WARN_ONCE((boot_cpu_data.x86 != 0xf),
42 "%s should only be used on K8!\n", __func__);
43
44 gprs[1] = msr;
45 gprs[7] = 0x9c5a203a;
46
47 err = rdmsr_safe_regs(gprs);
48
49 *p = gprs[0] | ((u64)gprs[2] << 32);
50
51 return err;
52}
53
54static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
55{
56 u32 gprs[8] = { 0 };
57
58 WARN_ONCE((boot_cpu_data.x86 != 0xf),
59 "%s should only be used on K8!\n", __func__);
60
61 gprs[0] = (u32)val;
62 gprs[1] = msr;
63 gprs[2] = val >> 32;
64 gprs[7] = 0x9c5a203a;
65
66 return wrmsr_safe_regs(gprs);
67}
68
69/*
70 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
71 * misexecution of code under Linux. Owners of such processors should
72 * contact AMD for precise details and a CPU swap.
73 *
74 * See http://www.multimania.com/poulot/k6bug.html
75 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
76 * (Publication # 21266 Issue Date: August 1998)
77 *
78 * The following test is erm.. interesting. AMD neglected to up
79 * the chip setting when fixing the bug but they also tweaked some
80 * performance at the same time..
81 */
82
83extern __visible void vide(void);
84__asm__(".globl vide\n"
85 ".type vide, @function\n"
86 ".align 4\n"
87 "vide: ret\n");
88
89static void init_amd_k5(struct cpuinfo_x86 *c)
90{
91#ifdef CONFIG_X86_32
92/*
93 * General Systems BIOSen alias the cpu frequency registers
94 * of the Elan at 0x000df000. Unfortunately, one of the Linux
95 * drivers subsequently pokes it, and changes the CPU speed.
96 * Workaround : Remove the unneeded alias.
97 */
98#define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
99#define CBAR_ENB (0x80000000)
100#define CBAR_KEY (0X000000CB)
101 if (c->x86_model == 9 || c->x86_model == 10) {
102 if (inl(CBAR) & CBAR_ENB)
103 outl(0 | CBAR_KEY, CBAR);
104 }
105#endif
106}
107
108static void init_amd_k6(struct cpuinfo_x86 *c)
109{
110#ifdef CONFIG_X86_32
111 u32 l, h;
112 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
113
114 if (c->x86_model < 6) {
115 /* Based on AMD doc 20734R - June 2000 */
116 if (c->x86_model == 0) {
117 clear_cpu_cap(c, X86_FEATURE_APIC);
118 set_cpu_cap(c, X86_FEATURE_PGE);
119 }
120 return;
121 }
122
123 if (c->x86_model == 6 && c->x86_stepping == 1) {
124 const int K6_BUG_LOOP = 1000000;
125 int n;
126 void (*f_vide)(void);
127 u64 d, d2;
128
129 pr_info("AMD K6 stepping B detected - ");
130
131 /*
132 * It looks like AMD fixed the 2.6.2 bug and improved indirect
133 * calls at the same time.
134 */
135
136 n = K6_BUG_LOOP;
137 f_vide = vide;
138 OPTIMIZER_HIDE_VAR(f_vide);
139 d = rdtsc();
140 while (n--)
141 f_vide();
142 d2 = rdtsc();
143 d = d2-d;
144
145 if (d > 20*K6_BUG_LOOP)
146 pr_cont("system stability may be impaired when more than 32 MB are used.\n");
147 else
148 pr_cont("probably OK (after B9730xxxx).\n");
149 }
150
151 /* K6 with old style WHCR */
152 if (c->x86_model < 8 ||
153 (c->x86_model == 8 && c->x86_stepping < 8)) {
154 /* We can only write allocate on the low 508Mb */
155 if (mbytes > 508)
156 mbytes = 508;
157
158 rdmsr(MSR_K6_WHCR, l, h);
159 if ((l&0x0000FFFF) == 0) {
160 unsigned long flags;
161 l = (1<<0)|((mbytes/4)<<1);
162 local_irq_save(flags);
163 wbinvd();
164 wrmsr(MSR_K6_WHCR, l, h);
165 local_irq_restore(flags);
166 pr_info("Enabling old style K6 write allocation for %d Mb\n",
167 mbytes);
168 }
169 return;
170 }
171
172 if ((c->x86_model == 8 && c->x86_stepping > 7) ||
173 c->x86_model == 9 || c->x86_model == 13) {
174 /* The more serious chips .. */
175
176 if (mbytes > 4092)
177 mbytes = 4092;
178
179 rdmsr(MSR_K6_WHCR, l, h);
180 if ((l&0xFFFF0000) == 0) {
181 unsigned long flags;
182 l = ((mbytes>>2)<<22)|(1<<16);
183 local_irq_save(flags);
184 wbinvd();
185 wrmsr(MSR_K6_WHCR, l, h);
186 local_irq_restore(flags);
187 pr_info("Enabling new style K6 write allocation for %d Mb\n",
188 mbytes);
189 }
190
191 return;
192 }
193
194 if (c->x86_model == 10) {
195 /* AMD Geode LX is model 10 */
196 /* placeholder for any needed mods */
197 return;
198 }
199#endif
200}
201
202static void init_amd_k7(struct cpuinfo_x86 *c)
203{
204#ifdef CONFIG_X86_32
205 u32 l, h;
206
207 /*
208 * Bit 15 of Athlon specific MSR 15, needs to be 0
209 * to enable SSE on Palomino/Morgan/Barton CPU's.
210 * If the BIOS didn't enable it already, enable it here.
211 */
212 if (c->x86_model >= 6 && c->x86_model <= 10) {
213 if (!cpu_has(c, X86_FEATURE_XMM)) {
214 pr_info("Enabling disabled K7/SSE Support.\n");
215 msr_clear_bit(MSR_K7_HWCR, 15);
216 set_cpu_cap(c, X86_FEATURE_XMM);
217 }
218 }
219
220 /*
221 * It's been determined by AMD that Athlons since model 8 stepping 1
222 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
223 * As per AMD technical note 27212 0.2
224 */
225 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
226 rdmsr(MSR_K7_CLK_CTL, l, h);
227 if ((l & 0xfff00000) != 0x20000000) {
228 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
229 l, ((l & 0x000fffff)|0x20000000));
230 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
231 }
232 }
233
234 set_cpu_cap(c, X86_FEATURE_K7);
235
236 /* calling is from identify_secondary_cpu() ? */
237 if (!c->cpu_index)
238 return;
239
240 /*
241 * Certain Athlons might work (for various values of 'work') in SMP
242 * but they are not certified as MP capable.
243 */
244 /* Athlon 660/661 is valid. */
245 if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
246 (c->x86_stepping == 1)))
247 return;
248
249 /* Duron 670 is valid */
250 if ((c->x86_model == 7) && (c->x86_stepping == 0))
251 return;
252
253 /*
254 * Athlon 662, Duron 671, and Athlon >model 7 have capability
255 * bit. It's worth noting that the A5 stepping (662) of some
256 * Athlon XP's have the MP bit set.
257 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
258 * more.
259 */
260 if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
261 ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
262 (c->x86_model > 7))
263 if (cpu_has(c, X86_FEATURE_MP))
264 return;
265
266 /* If we get here, not a certified SMP capable AMD system. */
267
268 /*
269 * Don't taint if we are running SMP kernel on a single non-MP
270 * approved Athlon
271 */
272 WARN_ONCE(1, "WARNING: This combination of AMD"
273 " processors is not suitable for SMP.\n");
274 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
275#endif
276}
277
278#ifdef CONFIG_NUMA
279/*
280 * To workaround broken NUMA config. Read the comment in
281 * srat_detect_node().
282 */
283static int nearby_node(int apicid)
284{
285 int i, node;
286
287 for (i = apicid - 1; i >= 0; i--) {
288 node = __apicid_to_node[i];
289 if (node != NUMA_NO_NODE && node_online(node))
290 return node;
291 }
292 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
293 node = __apicid_to_node[i];
294 if (node != NUMA_NO_NODE && node_online(node))
295 return node;
296 }
297 return first_node(node_online_map); /* Shouldn't happen */
298}
299#endif
300
301#ifdef CONFIG_SMP
302/*
303 * Fix up cpu_core_id for pre-F17h systems to be in the
304 * [0 .. cores_per_node - 1] range. Not really needed but
305 * kept so as not to break existing setups.
306 */
307static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
308{
309 u32 cus_per_node;
310
311 if (c->x86 >= 0x17)
312 return;
313
314 cus_per_node = c->x86_max_cores / nodes_per_socket;
315 c->cpu_core_id %= cus_per_node;
316}
317
318/*
319 * Fixup core topology information for
320 * (1) AMD multi-node processors
321 * Assumption: Number of cores in each internal node is the same.
322 * (2) AMD processors supporting compute units
323 */
324static void amd_get_topology(struct cpuinfo_x86 *c)
325{
326 u8 node_id;
327 int cpu = smp_processor_id();
328
329 /* get information required for multi-node processors */
330 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
331 u32 eax, ebx, ecx, edx;
332
333 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
334
335 node_id = ecx & 0xff;
336 smp_num_siblings = ((ebx >> 8) & 0xff) + 1;
337
338 if (c->x86 == 0x15)
339 c->cu_id = ebx & 0xff;
340
341 if (c->x86 >= 0x17) {
342 c->cpu_core_id = ebx & 0xff;
343
344 if (smp_num_siblings > 1)
345 c->x86_max_cores /= smp_num_siblings;
346 }
347
348 /*
349 * We may have multiple LLCs if L3 caches exist, so check if we
350 * have an L3 cache by looking at the L3 cache CPUID leaf.
351 */
352 if (cpuid_edx(0x80000006)) {
353 if (c->x86 == 0x17) {
354 /*
355 * LLC is at the core complex level.
356 * Core complex id is ApicId[3].
357 */
358 per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
359 } else {
360 /* LLC is at the node level. */
361 per_cpu(cpu_llc_id, cpu) = node_id;
362 }
363 }
364 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
365 u64 value;
366
367 rdmsrl(MSR_FAM10H_NODE_ID, value);
368 node_id = value & 7;
369
370 per_cpu(cpu_llc_id, cpu) = node_id;
371 } else
372 return;
373
374 if (nodes_per_socket > 1) {
375 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
376 legacy_fixup_core_id(c);
377 }
378}
379#endif
380
381/*
382 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
383 * Assumes number of cores is a power of two.
384 */
385static void amd_detect_cmp(struct cpuinfo_x86 *c)
386{
387#ifdef CONFIG_SMP
388 unsigned bits;
389 int cpu = smp_processor_id();
390
391 bits = c->x86_coreid_bits;
392 /* Low order bits define the core id (index of core in socket) */
393 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
394 /* Convert the initial APIC ID into the socket ID */
395 c->phys_proc_id = c->initial_apicid >> bits;
396 /* use socket ID also for last level cache */
397 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
398 amd_get_topology(c);
399#endif
400}
401
402u16 amd_get_nb_id(int cpu)
403{
404 u16 id = 0;
405#ifdef CONFIG_SMP
406 id = per_cpu(cpu_llc_id, cpu);
407#endif
408 return id;
409}
410EXPORT_SYMBOL_GPL(amd_get_nb_id);
411
412u32 amd_get_nodes_per_socket(void)
413{
414 return nodes_per_socket;
415}
416EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
417
418static void srat_detect_node(struct cpuinfo_x86 *c)
419{
420#ifdef CONFIG_NUMA
421 int cpu = smp_processor_id();
422 int node;
423 unsigned apicid = c->apicid;
424
425 node = numa_cpu_node(cpu);
426 if (node == NUMA_NO_NODE)
427 node = per_cpu(cpu_llc_id, cpu);
428
429 /*
430 * On multi-fabric platform (e.g. Numascale NumaChip) a
431 * platform-specific handler needs to be called to fixup some
432 * IDs of the CPU.
433 */
434 if (x86_cpuinit.fixup_cpu_id)
435 x86_cpuinit.fixup_cpu_id(c, node);
436
437 if (!node_online(node)) {
438 /*
439 * Two possibilities here:
440 *
441 * - The CPU is missing memory and no node was created. In
442 * that case try picking one from a nearby CPU.
443 *
444 * - The APIC IDs differ from the HyperTransport node IDs
445 * which the K8 northbridge parsing fills in. Assume
446 * they are all increased by a constant offset, but in
447 * the same order as the HT nodeids. If that doesn't
448 * result in a usable node fall back to the path for the
449 * previous case.
450 *
451 * This workaround operates directly on the mapping between
452 * APIC ID and NUMA node, assuming certain relationship
453 * between APIC ID, HT node ID and NUMA topology. As going
454 * through CPU mapping may alter the outcome, directly
455 * access __apicid_to_node[].
456 */
457 int ht_nodeid = c->initial_apicid;
458
459 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
460 node = __apicid_to_node[ht_nodeid];
461 /* Pick a nearby node */
462 if (!node_online(node))
463 node = nearby_node(apicid);
464 }
465 numa_set_node(cpu, node);
466#endif
467}
468
469static void early_init_amd_mc(struct cpuinfo_x86 *c)
470{
471#ifdef CONFIG_SMP
472 unsigned bits, ecx;
473
474 /* Multi core CPU? */
475 if (c->extended_cpuid_level < 0x80000008)
476 return;
477
478 ecx = cpuid_ecx(0x80000008);
479
480 c->x86_max_cores = (ecx & 0xff) + 1;
481
482 /* CPU telling us the core id bits shift? */
483 bits = (ecx >> 12) & 0xF;
484
485 /* Otherwise recompute */
486 if (bits == 0) {
487 while ((1 << bits) < c->x86_max_cores)
488 bits++;
489 }
490
491 c->x86_coreid_bits = bits;
492#endif
493}
494
495static void bsp_init_amd(struct cpuinfo_x86 *c)
496{
497
498#ifdef CONFIG_X86_64
499 if (c->x86 >= 0xf) {
500 unsigned long long tseg;
501
502 /*
503 * Split up direct mapping around the TSEG SMM area.
504 * Don't do it for gbpages because there seems very little
505 * benefit in doing so.
506 */
507 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
508 unsigned long pfn = tseg >> PAGE_SHIFT;
509
510 pr_debug("tseg: %010llx\n", tseg);
511 if (pfn_range_is_mapped(pfn, pfn + 1))
512 set_memory_4k((unsigned long)__va(tseg), 1);
513 }
514 }
515#endif
516
517 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
518
519 if (c->x86 > 0x10 ||
520 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
521 u64 val;
522
523 rdmsrl(MSR_K7_HWCR, val);
524 if (!(val & BIT(24)))
525 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
526 }
527 }
528
529 if (c->x86 == 0x15) {
530 unsigned long upperbit;
531 u32 cpuid, assoc;
532
533 cpuid = cpuid_edx(0x80000005);
534 assoc = cpuid >> 16 & 0xff;
535 upperbit = ((cpuid >> 24) << 10) / assoc;
536
537 va_align.mask = (upperbit - 1) & PAGE_MASK;
538 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
539
540 /* A random value per boot for bit slice [12:upper_bit) */
541 va_align.bits = get_random_int() & va_align.mask;
542 }
543
544 if (cpu_has(c, X86_FEATURE_MWAITX))
545 use_mwaitx_delay();
546
547 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
548 u32 ecx;
549
550 ecx = cpuid_ecx(0x8000001e);
551 nodes_per_socket = ((ecx >> 8) & 7) + 1;
552 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
553 u64 value;
554
555 rdmsrl(MSR_FAM10H_NODE_ID, value);
556 nodes_per_socket = ((value >> 3) & 7) + 1;
557 }
558
559 if (c->x86 >= 0x15 && c->x86 <= 0x17) {
560 unsigned int bit;
561
562 switch (c->x86) {
563 case 0x15: bit = 54; break;
564 case 0x16: bit = 33; break;
565 case 0x17: bit = 10; break;
566 default: return;
567 }
568 /*
569 * Try to cache the base value so further operations can
570 * avoid RMW. If that faults, do not enable SSBD.
571 */
572 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
573 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
574 setup_force_cpu_cap(X86_FEATURE_SSBD);
575 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
576 }
577 }
578}
579
580static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
581{
582 u64 msr;
583
584 /*
585 * BIOS support is required for SME and SEV.
586 * For SME: If BIOS has enabled SME then adjust x86_phys_bits by
587 * the SME physical address space reduction value.
588 * If BIOS has not enabled SME then don't advertise the
589 * SME feature (set in scattered.c).
590 * For SEV: If BIOS has not enabled SEV then don't advertise the
591 * SEV feature (set in scattered.c).
592 *
593 * In all cases, since support for SME and SEV requires long mode,
594 * don't advertise the feature under CONFIG_X86_32.
595 */
596 if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
597 /* Check if memory encryption is enabled */
598 rdmsrl(MSR_K8_SYSCFG, msr);
599 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
600 goto clear_all;
601
602 /*
603 * Always adjust physical address bits. Even though this
604 * will be a value above 32-bits this is still done for
605 * CONFIG_X86_32 so that accurate values are reported.
606 */
607 c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
608
609 if (IS_ENABLED(CONFIG_X86_32))
610 goto clear_all;
611
612 rdmsrl(MSR_K7_HWCR, msr);
613 if (!(msr & MSR_K7_HWCR_SMMLOCK))
614 goto clear_sev;
615
616 return;
617
618clear_all:
619 clear_cpu_cap(c, X86_FEATURE_SME);
620clear_sev:
621 clear_cpu_cap(c, X86_FEATURE_SEV);
622 }
623}
624
625static void early_init_amd(struct cpuinfo_x86 *c)
626{
627 u32 dummy;
628
629 early_init_amd_mc(c);
630
631 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
632
633 /*
634 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
635 * with P/T states and does not stop in deep C-states
636 */
637 if (c->x86_power & (1 << 8)) {
638 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
639 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
640 }
641
642 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
643 if (c->x86_power & BIT(12))
644 set_cpu_cap(c, X86_FEATURE_ACC_POWER);
645
646#ifdef CONFIG_X86_64
647 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
648#else
649 /* Set MTRR capability flag if appropriate */
650 if (c->x86 == 5)
651 if (c->x86_model == 13 || c->x86_model == 9 ||
652 (c->x86_model == 8 && c->x86_stepping >= 8))
653 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
654#endif
655#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
656 /*
657 * ApicID can always be treated as an 8-bit value for AMD APIC versions
658 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
659 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
660 * after 16h.
661 */
662 if (boot_cpu_has(X86_FEATURE_APIC)) {
663 if (c->x86 > 0x16)
664 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
665 else if (c->x86 >= 0xf) {
666 /* check CPU config space for extended APIC ID */
667 unsigned int val;
668
669 val = read_pci_config(0, 24, 0, 0x68);
670 if ((val >> 17 & 0x3) == 0x3)
671 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
672 }
673 }
674#endif
675
676 /*
677 * This is only needed to tell the kernel whether to use VMCALL
678 * and VMMCALL. VMMCALL is never executed except under virt, so
679 * we can set it unconditionally.
680 */
681 set_cpu_cap(c, X86_FEATURE_VMMCALL);
682
683 /* F16h erratum 793, CVE-2013-6885 */
684 if (c->x86 == 0x16 && c->x86_model <= 0xf)
685 msr_set_bit(MSR_AMD64_LS_CFG, 15);
686
687 /*
688 * Check whether the machine is affected by erratum 400. This is
689 * used to select the proper idle routine and to enable the check
690 * whether the machine is affected in arch_post_acpi_init(), which
691 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
692 */
693 if (cpu_has_amd_erratum(c, amd_erratum_400))
694 set_cpu_bug(c, X86_BUG_AMD_E400);
695
696 early_detect_mem_encrypt(c);
697}
698
699static void init_amd_k8(struct cpuinfo_x86 *c)
700{
701 u32 level;
702 u64 value;
703
704 /* On C+ stepping K8 rep microcode works well for copy/memset */
705 level = cpuid_eax(1);
706 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
707 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
708
709 /*
710 * Some BIOSes incorrectly force this feature, but only K8 revision D
711 * (model = 0x14) and later actually support it.
712 * (AMD Erratum #110, docId: 25759).
713 */
714 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
715 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
716 if (!rdmsrl_amd_safe(0xc001100d, &value)) {
717 value &= ~BIT_64(32);
718 wrmsrl_amd_safe(0xc001100d, value);
719 }
720 }
721
722 if (!c->x86_model_id[0])
723 strcpy(c->x86_model_id, "Hammer");
724
725#ifdef CONFIG_SMP
726 /*
727 * Disable TLB flush filter by setting HWCR.FFDIS on K8
728 * bit 6 of msr C001_0015
729 *
730 * Errata 63 for SH-B3 steppings
731 * Errata 122 for all steppings (F+ have it disabled by default)
732 */
733 msr_set_bit(MSR_K7_HWCR, 6);
734#endif
735 set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
736}
737
738static void init_amd_gh(struct cpuinfo_x86 *c)
739{
740#ifdef CONFIG_MMCONF_FAM10H
741 /* do this for boot cpu */
742 if (c == &boot_cpu_data)
743 check_enable_amd_mmconf_dmi();
744
745 fam10h_check_enable_mmcfg();
746#endif
747
748 /*
749 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
750 * is always needed when GART is enabled, even in a kernel which has no
751 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
752 * If it doesn't, we do it here as suggested by the BKDG.
753 *
754 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
755 */
756 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
757
758 /*
759 * On family 10h BIOS may not have properly enabled WC+ support, causing
760 * it to be converted to CD memtype. This may result in performance
761 * degradation for certain nested-paging guests. Prevent this conversion
762 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
763 *
764 * NOTE: we want to use the _safe accessors so as not to #GP kvm
765 * guests on older kvm hosts.
766 */
767 msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
768
769 if (cpu_has_amd_erratum(c, amd_erratum_383))
770 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
771}
772
773#define MSR_AMD64_DE_CFG 0xC0011029
774
775static void init_amd_ln(struct cpuinfo_x86 *c)
776{
777 /*
778 * Apply erratum 665 fix unconditionally so machines without a BIOS
779 * fix work.
780 */
781 msr_set_bit(MSR_AMD64_DE_CFG, 31);
782}
783
784static void init_amd_bd(struct cpuinfo_x86 *c)
785{
786 u64 value;
787
788 /* re-enable TopologyExtensions if switched off by BIOS */
789 if ((c->x86_model >= 0x10) && (c->x86_model <= 0x6f) &&
790 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
791
792 if (msr_set_bit(0xc0011005, 54) > 0) {
793 rdmsrl(0xc0011005, value);
794 if (value & BIT_64(54)) {
795 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
796 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
797 }
798 }
799 }
800
801 /*
802 * The way access filter has a performance penalty on some workloads.
803 * Disable it on the affected CPUs.
804 */
805 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
806 if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
807 value |= 0x1E;
808 wrmsrl_safe(MSR_F15H_IC_CFG, value);
809 }
810 }
811}
812
813static void init_amd_zn(struct cpuinfo_x86 *c)
814{
815 set_cpu_cap(c, X86_FEATURE_ZEN);
816 /*
817 * Fix erratum 1076: CPB feature bit not being set in CPUID. It affects
818 * all up to and including B1.
819 */
820 if (c->x86_model <= 1 && c->x86_stepping <= 1)
821 set_cpu_cap(c, X86_FEATURE_CPB);
822}
823
824static void init_amd(struct cpuinfo_x86 *c)
825{
826 early_init_amd(c);
827
828 /*
829 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
830 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
831 */
832 clear_cpu_cap(c, 0*32+31);
833
834 if (c->x86 >= 0x10)
835 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
836
837 /* get apicid instead of initial apic id from cpuid */
838 c->apicid = hard_smp_processor_id();
839
840 /* K6s reports MCEs but don't actually have all the MSRs */
841 if (c->x86 < 6)
842 clear_cpu_cap(c, X86_FEATURE_MCE);
843
844 switch (c->x86) {
845 case 4: init_amd_k5(c); break;
846 case 5: init_amd_k6(c); break;
847 case 6: init_amd_k7(c); break;
848 case 0xf: init_amd_k8(c); break;
849 case 0x10: init_amd_gh(c); break;
850 case 0x12: init_amd_ln(c); break;
851 case 0x15: init_amd_bd(c); break;
852 case 0x17: init_amd_zn(c); break;
853 }
854
855 /*
856 * Enable workaround for FXSAVE leak on CPUs
857 * without a XSaveErPtr feature
858 */
859 if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
860 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
861
862 cpu_detect_cache_sizes(c);
863
864 /* Multi core CPU? */
865 if (c->extended_cpuid_level >= 0x80000008) {
866 amd_detect_cmp(c);
867 srat_detect_node(c);
868 }
869
870#ifdef CONFIG_X86_32
871 detect_ht(c);
872#endif
873
874 init_amd_cacheinfo(c);
875
876 if (c->x86 >= 0xf)
877 set_cpu_cap(c, X86_FEATURE_K8);
878
879 if (cpu_has(c, X86_FEATURE_XMM2)) {
880 unsigned long long val;
881 int ret;
882
883 /*
884 * A serializing LFENCE has less overhead than MFENCE, so
885 * use it for execution serialization. On families which
886 * don't have that MSR, LFENCE is already serializing.
887 * msr_set_bit() uses the safe accessors, too, even if the MSR
888 * is not present.
889 */
890 msr_set_bit(MSR_F10H_DECFG,
891 MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT);
892
893 /*
894 * Verify that the MSR write was successful (could be running
895 * under a hypervisor) and only then assume that LFENCE is
896 * serializing.
897 */
898 ret = rdmsrl_safe(MSR_F10H_DECFG, &val);
899 if (!ret && (val & MSR_F10H_DECFG_LFENCE_SERIALIZE)) {
900 /* A serializing LFENCE stops RDTSC speculation */
901 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
902 } else {
903 /* MFENCE stops RDTSC speculation */
904 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
905 }
906 }
907
908 /*
909 * Family 0x12 and above processors have APIC timer
910 * running in deep C states.
911 */
912 if (c->x86 > 0x11)
913 set_cpu_cap(c, X86_FEATURE_ARAT);
914
915 /* 3DNow or LM implies PREFETCHW */
916 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
917 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
918 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
919
920 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
921 if (!cpu_has(c, X86_FEATURE_XENPV))
922 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
923}
924
925#ifdef CONFIG_X86_32
926static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
927{
928 /* AMD errata T13 (order #21922) */
929 if ((c->x86 == 6)) {
930 /* Duron Rev A0 */
931 if (c->x86_model == 3 && c->x86_stepping == 0)
932 size = 64;
933 /* Tbird rev A1/A2 */
934 if (c->x86_model == 4 &&
935 (c->x86_stepping == 0 || c->x86_stepping == 1))
936 size = 256;
937 }
938 return size;
939}
940#endif
941
942static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
943{
944 u32 ebx, eax, ecx, edx;
945 u16 mask = 0xfff;
946
947 if (c->x86 < 0xf)
948 return;
949
950 if (c->extended_cpuid_level < 0x80000006)
951 return;
952
953 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
954
955 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
956 tlb_lli_4k[ENTRIES] = ebx & mask;
957
958 /*
959 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
960 * characteristics from the CPUID function 0x80000005 instead.
961 */
962 if (c->x86 == 0xf) {
963 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
964 mask = 0xff;
965 }
966
967 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
968 if (!((eax >> 16) & mask))
969 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
970 else
971 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
972
973 /* a 4M entry uses two 2M entries */
974 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
975
976 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
977 if (!(eax & mask)) {
978 /* Erratum 658 */
979 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
980 tlb_lli_2m[ENTRIES] = 1024;
981 } else {
982 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
983 tlb_lli_2m[ENTRIES] = eax & 0xff;
984 }
985 } else
986 tlb_lli_2m[ENTRIES] = eax & mask;
987
988 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
989}
990
991static const struct cpu_dev amd_cpu_dev = {
992 .c_vendor = "AMD",
993 .c_ident = { "AuthenticAMD" },
994#ifdef CONFIG_X86_32
995 .legacy_models = {
996 { .family = 4, .model_names =
997 {
998 [3] = "486 DX/2",
999 [7] = "486 DX/2-WB",
1000 [8] = "486 DX/4",
1001 [9] = "486 DX/4-WB",
1002 [14] = "Am5x86-WT",
1003 [15] = "Am5x86-WB"
1004 }
1005 },
1006 },
1007 .legacy_cache_size = amd_size_cache,
1008#endif
1009 .c_early_init = early_init_amd,
1010 .c_detect_tlb = cpu_detect_tlb_amd,
1011 .c_bsp_init = bsp_init_amd,
1012 .c_init = init_amd,
1013 .c_x86_vendor = X86_VENDOR_AMD,
1014};
1015
1016cpu_dev_register(amd_cpu_dev);
1017
1018/*
1019 * AMD errata checking
1020 *
1021 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1022 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1023 * have an OSVW id assigned, which it takes as first argument. Both take a
1024 * variable number of family-specific model-stepping ranges created by
1025 * AMD_MODEL_RANGE().
1026 *
1027 * Example:
1028 *
1029 * const int amd_erratum_319[] =
1030 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1031 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1032 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1033 */
1034
1035#define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
1036#define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
1037#define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1038 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1039#define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
1040#define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
1041#define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
1042
1043static const int amd_erratum_400[] =
1044 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1045 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1046
1047static const int amd_erratum_383[] =
1048 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1049
1050
1051static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1052{
1053 int osvw_id = *erratum++;
1054 u32 range;
1055 u32 ms;
1056
1057 if (osvw_id >= 0 && osvw_id < 65536 &&
1058 cpu_has(cpu, X86_FEATURE_OSVW)) {
1059 u64 osvw_len;
1060
1061 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1062 if (osvw_id < osvw_len) {
1063 u64 osvw_bits;
1064
1065 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1066 osvw_bits);
1067 return osvw_bits & (1ULL << (osvw_id & 0x3f));
1068 }
1069 }
1070
1071 /* OSVW unavailable or ID unknown, match family-model-stepping range */
1072 ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1073 while ((range = *erratum++))
1074 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1075 (ms >= AMD_MODEL_RANGE_START(range)) &&
1076 (ms <= AMD_MODEL_RANGE_END(range)))
1077 return true;
1078
1079 return false;
1080}
1081
1082void set_dr_addr_mask(unsigned long mask, int dr)
1083{
1084 if (!boot_cpu_has(X86_FEATURE_BPEXT))
1085 return;
1086
1087 switch (dr) {
1088 case 0:
1089 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1090 break;
1091 case 1:
1092 case 2:
1093 case 3:
1094 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);
1095 break;
1096 default:
1097 break;
1098 }
1099}
1#include <linux/export.h>
2#include <linux/init.h>
3#include <linux/bitops.h>
4#include <linux/elf.h>
5#include <linux/mm.h>
6
7#include <linux/io.h>
8#include <linux/sched.h>
9#include <asm/processor.h>
10#include <asm/apic.h>
11#include <asm/cpu.h>
12#include <asm/pci-direct.h>
13
14#ifdef CONFIG_X86_64
15# include <asm/numa_64.h>
16# include <asm/mmconfig.h>
17# include <asm/cacheflush.h>
18#endif
19
20#include "cpu.h"
21
22#ifdef CONFIG_X86_32
23/*
24 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
25 * misexecution of code under Linux. Owners of such processors should
26 * contact AMD for precise details and a CPU swap.
27 *
28 * See http://www.multimania.com/poulot/k6bug.html
29 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
30 * (Publication # 21266 Issue Date: August 1998)
31 *
32 * The following test is erm.. interesting. AMD neglected to up
33 * the chip setting when fixing the bug but they also tweaked some
34 * performance at the same time..
35 */
36
37extern void vide(void);
38__asm__(".align 4\nvide: ret");
39
40static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
41{
42/*
43 * General Systems BIOSen alias the cpu frequency registers
44 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
45 * drivers subsequently pokes it, and changes the CPU speed.
46 * Workaround : Remove the unneeded alias.
47 */
48#define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
49#define CBAR_ENB (0x80000000)
50#define CBAR_KEY (0X000000CB)
51 if (c->x86_model == 9 || c->x86_model == 10) {
52 if (inl(CBAR) & CBAR_ENB)
53 outl(0 | CBAR_KEY, CBAR);
54 }
55}
56
57
58static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
59{
60 u32 l, h;
61 int mbytes = num_physpages >> (20-PAGE_SHIFT);
62
63 if (c->x86_model < 6) {
64 /* Based on AMD doc 20734R - June 2000 */
65 if (c->x86_model == 0) {
66 clear_cpu_cap(c, X86_FEATURE_APIC);
67 set_cpu_cap(c, X86_FEATURE_PGE);
68 }
69 return;
70 }
71
72 if (c->x86_model == 6 && c->x86_mask == 1) {
73 const int K6_BUG_LOOP = 1000000;
74 int n;
75 void (*f_vide)(void);
76 unsigned long d, d2;
77
78 printk(KERN_INFO "AMD K6 stepping B detected - ");
79
80 /*
81 * It looks like AMD fixed the 2.6.2 bug and improved indirect
82 * calls at the same time.
83 */
84
85 n = K6_BUG_LOOP;
86 f_vide = vide;
87 rdtscl(d);
88 while (n--)
89 f_vide();
90 rdtscl(d2);
91 d = d2-d;
92
93 if (d > 20*K6_BUG_LOOP)
94 printk(KERN_CONT
95 "system stability may be impaired when more than 32 MB are used.\n");
96 else
97 printk(KERN_CONT "probably OK (after B9730xxxx).\n");
98 }
99
100 /* K6 with old style WHCR */
101 if (c->x86_model < 8 ||
102 (c->x86_model == 8 && c->x86_mask < 8)) {
103 /* We can only write allocate on the low 508Mb */
104 if (mbytes > 508)
105 mbytes = 508;
106
107 rdmsr(MSR_K6_WHCR, l, h);
108 if ((l&0x0000FFFF) == 0) {
109 unsigned long flags;
110 l = (1<<0)|((mbytes/4)<<1);
111 local_irq_save(flags);
112 wbinvd();
113 wrmsr(MSR_K6_WHCR, l, h);
114 local_irq_restore(flags);
115 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
116 mbytes);
117 }
118 return;
119 }
120
121 if ((c->x86_model == 8 && c->x86_mask > 7) ||
122 c->x86_model == 9 || c->x86_model == 13) {
123 /* The more serious chips .. */
124
125 if (mbytes > 4092)
126 mbytes = 4092;
127
128 rdmsr(MSR_K6_WHCR, l, h);
129 if ((l&0xFFFF0000) == 0) {
130 unsigned long flags;
131 l = ((mbytes>>2)<<22)|(1<<16);
132 local_irq_save(flags);
133 wbinvd();
134 wrmsr(MSR_K6_WHCR, l, h);
135 local_irq_restore(flags);
136 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
137 mbytes);
138 }
139
140 return;
141 }
142
143 if (c->x86_model == 10) {
144 /* AMD Geode LX is model 10 */
145 /* placeholder for any needed mods */
146 return;
147 }
148}
149
150static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
151{
152 /* calling is from identify_secondary_cpu() ? */
153 if (!c->cpu_index)
154 return;
155
156 /*
157 * Certain Athlons might work (for various values of 'work') in SMP
158 * but they are not certified as MP capable.
159 */
160 /* Athlon 660/661 is valid. */
161 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
162 (c->x86_mask == 1)))
163 goto valid_k7;
164
165 /* Duron 670 is valid */
166 if ((c->x86_model == 7) && (c->x86_mask == 0))
167 goto valid_k7;
168
169 /*
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
174 * more.
175 */
176 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
177 ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
178 (c->x86_model > 7))
179 if (cpu_has_mp)
180 goto valid_k7;
181
182 /* If we get here, not a certified SMP capable AMD system. */
183
184 /*
185 * Don't taint if we are running SMP kernel on a single non-MP
186 * approved Athlon
187 */
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);
192
193valid_k7:
194 ;
195}
196
197static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
198{
199 u32 l, h;
200
201 /*
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.
205 */
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);
210 l &= ~0x00008000;
211 wrmsr(MSR_K7_HWCR, l, h);
212 set_cpu_cap(c, X86_FEATURE_XMM);
213 }
214 }
215
216 /*
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
220 */
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) {
224 printk(KERN_INFO
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);
228 }
229 }
230
231 set_cpu_cap(c, X86_FEATURE_K7);
232
233 amd_k7_smp_check(c);
234}
235#endif
236
237#ifdef CONFIG_NUMA
238/*
239 * To workaround broken NUMA config. Read the comment in
240 * srat_detect_node().
241 */
242static int __cpuinit nearby_node(int apicid)
243{
244 int i, node;
245
246 for (i = apicid - 1; i >= 0; i--) {
247 node = __apicid_to_node[i];
248 if (node != NUMA_NO_NODE && node_online(node))
249 return node;
250 }
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))
254 return node;
255 }
256 return first_node(node_online_map); /* Shouldn't happen */
257}
258#endif
259
260/*
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
265 */
266#ifdef CONFIG_X86_HT
267static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
268{
269 u32 nodes, cores_per_cu = 1;
270 u8 node_id;
271 int cpu = smp_processor_id();
272
273 /* get information required for multi-node processors */
274 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
275 u32 eax, ebx, ecx, edx;
276
277 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
278 nodes = ((ecx >> 8) & 7) + 1;
279 node_id = ecx & 7;
280
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)) {
286 u64 value;
287
288 rdmsrl(MSR_FAM10H_NODE_ID, value);
289 nodes = ((value >> 3) & 7) + 1;
290 node_id = value & 7;
291 } else
292 return;
293
294 /* fixup multi-node processor information */
295 if (nodes > 1) {
296 u32 cores_per_node;
297 u32 cus_per_node;
298
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;
302
303 /* store NodeID, use llc_shared_map to store sibling info */
304 per_cpu(cpu_llc_id, cpu) = node_id;
305
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;
309 }
310}
311#endif
312
313/*
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.
316 */
317static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
318{
319#ifdef CONFIG_X86_HT
320 unsigned bits;
321 int cpu = smp_processor_id();
322
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;
330 amd_get_topology(c);
331#endif
332}
333
334int amd_get_nb_id(int cpu)
335{
336 int id = 0;
337#ifdef CONFIG_SMP
338 id = per_cpu(cpu_llc_id, cpu);
339#endif
340 return id;
341}
342EXPORT_SYMBOL_GPL(amd_get_nb_id);
343
344static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
345{
346#ifdef CONFIG_NUMA
347 int cpu = smp_processor_id();
348 int node;
349 unsigned apicid = c->apicid;
350
351 node = numa_cpu_node(cpu);
352 if (node == NUMA_NO_NODE)
353 node = per_cpu(cpu_llc_id, cpu);
354
355 /*
356 * On multi-fabric platform (e.g. Numascale NumaChip) a
357 * platform-specific handler needs to be called to fixup some
358 * IDs of the CPU.
359 */
360 if (x86_cpuinit.fixup_cpu_id)
361 x86_cpuinit.fixup_cpu_id(c, node);
362
363 if (!node_online(node)) {
364 /*
365 * Two possibilities here:
366 *
367 * - The CPU is missing memory and no node was created. In
368 * that case try picking one from a nearby CPU.
369 *
370 * - The APIC IDs differ from the HyperTransport node IDs
371 * which the K8 northbridge parsing fills in. Assume
372 * they are all increased by a constant offset, but in
373 * the same order as the HT nodeids. If that doesn't
374 * result in a usable node fall back to the path for the
375 * previous case.
376 *
377 * This workaround operates directly on the mapping between
378 * APIC ID and NUMA node, assuming certain relationship
379 * between APIC ID, HT node ID and NUMA topology. As going
380 * through CPU mapping may alter the outcome, directly
381 * access __apicid_to_node[].
382 */
383 int ht_nodeid = c->initial_apicid;
384
385 if (ht_nodeid >= 0 &&
386 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
387 node = __apicid_to_node[ht_nodeid];
388 /* Pick a nearby node */
389 if (!node_online(node))
390 node = nearby_node(apicid);
391 }
392 numa_set_node(cpu, node);
393#endif
394}
395
396static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
397{
398#ifdef CONFIG_X86_HT
399 unsigned bits, ecx;
400
401 /* Multi core CPU? */
402 if (c->extended_cpuid_level < 0x80000008)
403 return;
404
405 ecx = cpuid_ecx(0x80000008);
406
407 c->x86_max_cores = (ecx & 0xff) + 1;
408
409 /* CPU telling us the core id bits shift? */
410 bits = (ecx >> 12) & 0xF;
411
412 /* Otherwise recompute */
413 if (bits == 0) {
414 while ((1 << bits) < c->x86_max_cores)
415 bits++;
416 }
417
418 c->x86_coreid_bits = bits;
419#endif
420}
421
422static void __cpuinit bsp_init_amd(struct cpuinfo_x86 *c)
423{
424 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
425
426 if (c->x86 > 0x10 ||
427 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
428 u64 val;
429
430 rdmsrl(MSR_K7_HWCR, val);
431 if (!(val & BIT(24)))
432 printk(KERN_WARNING FW_BUG "TSC doesn't count "
433 "with P0 frequency!\n");
434 }
435 }
436
437 if (c->x86 == 0x15) {
438 unsigned long upperbit;
439 u32 cpuid, assoc;
440
441 cpuid = cpuid_edx(0x80000005);
442 assoc = cpuid >> 16 & 0xff;
443 upperbit = ((cpuid >> 24) << 10) / assoc;
444
445 va_align.mask = (upperbit - 1) & PAGE_MASK;
446 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
447 }
448}
449
450static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
451{
452 early_init_amd_mc(c);
453
454 /*
455 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
456 * with P/T states and does not stop in deep C-states
457 */
458 if (c->x86_power & (1 << 8)) {
459 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
460 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
461 if (!check_tsc_unstable())
462 sched_clock_stable = 1;
463 }
464
465#ifdef CONFIG_X86_64
466 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
467#else
468 /* Set MTRR capability flag if appropriate */
469 if (c->x86 == 5)
470 if (c->x86_model == 13 || c->x86_model == 9 ||
471 (c->x86_model == 8 && c->x86_mask >= 8))
472 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
473#endif
474#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
475 /* check CPU config space for extended APIC ID */
476 if (cpu_has_apic && c->x86 >= 0xf) {
477 unsigned int val;
478 val = read_pci_config(0, 24, 0, 0x68);
479 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
480 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
481 }
482#endif
483}
484
485static void __cpuinit init_amd(struct cpuinfo_x86 *c)
486{
487 u32 dummy;
488
489#ifdef CONFIG_SMP
490 unsigned long long value;
491
492 /*
493 * Disable TLB flush filter by setting HWCR.FFDIS on K8
494 * bit 6 of msr C001_0015
495 *
496 * Errata 63 for SH-B3 steppings
497 * Errata 122 for all steppings (F+ have it disabled by default)
498 */
499 if (c->x86 == 0xf) {
500 rdmsrl(MSR_K7_HWCR, value);
501 value |= 1 << 6;
502 wrmsrl(MSR_K7_HWCR, value);
503 }
504#endif
505
506 early_init_amd(c);
507
508 /*
509 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
510 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
511 */
512 clear_cpu_cap(c, 0*32+31);
513
514#ifdef CONFIG_X86_64
515 /* On C+ stepping K8 rep microcode works well for copy/memset */
516 if (c->x86 == 0xf) {
517 u32 level;
518
519 level = cpuid_eax(1);
520 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
521 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
522
523 /*
524 * Some BIOSes incorrectly force this feature, but only K8
525 * revision D (model = 0x14) and later actually support it.
526 * (AMD Erratum #110, docId: 25759).
527 */
528 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
529 u64 val;
530
531 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
532 if (!rdmsrl_amd_safe(0xc001100d, &val)) {
533 val &= ~(1ULL << 32);
534 wrmsrl_amd_safe(0xc001100d, val);
535 }
536 }
537
538 }
539 if (c->x86 >= 0x10)
540 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
541
542 /* get apicid instead of initial apic id from cpuid */
543 c->apicid = hard_smp_processor_id();
544#else
545
546 /*
547 * FIXME: We should handle the K5 here. Set up the write
548 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
549 * no bus pipeline)
550 */
551
552 switch (c->x86) {
553 case 4:
554 init_amd_k5(c);
555 break;
556 case 5:
557 init_amd_k6(c);
558 break;
559 case 6: /* An Athlon/Duron */
560 init_amd_k7(c);
561 break;
562 }
563
564 /* K6s reports MCEs but don't actually have all the MSRs */
565 if (c->x86 < 6)
566 clear_cpu_cap(c, X86_FEATURE_MCE);
567#endif
568
569 /* Enable workaround for FXSAVE leak */
570 if (c->x86 >= 6)
571 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
572
573 if (!c->x86_model_id[0]) {
574 switch (c->x86) {
575 case 0xf:
576 /* Should distinguish Models here, but this is only
577 a fallback anyways. */
578 strcpy(c->x86_model_id, "Hammer");
579 break;
580 }
581 }
582
583 /* re-enable TopologyExtensions if switched off by BIOS */
584 if ((c->x86 == 0x15) &&
585 (c->x86_model >= 0x10) && (c->x86_model <= 0x1f) &&
586 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
587 u64 val;
588
589 if (!rdmsrl_amd_safe(0xc0011005, &val)) {
590 val |= 1ULL << 54;
591 wrmsrl_amd_safe(0xc0011005, val);
592 rdmsrl(0xc0011005, val);
593 if (val & (1ULL << 54)) {
594 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
595 printk(KERN_INFO FW_INFO "CPU: Re-enabling "
596 "disabled Topology Extensions Support\n");
597 }
598 }
599 }
600
601 cpu_detect_cache_sizes(c);
602
603 /* Multi core CPU? */
604 if (c->extended_cpuid_level >= 0x80000008) {
605 amd_detect_cmp(c);
606 srat_detect_node(c);
607 }
608
609#ifdef CONFIG_X86_32
610 detect_ht(c);
611#endif
612
613 if (c->extended_cpuid_level >= 0x80000006) {
614 if (cpuid_edx(0x80000006) & 0xf000)
615 num_cache_leaves = 4;
616 else
617 num_cache_leaves = 3;
618 }
619
620 if (c->x86 >= 0xf)
621 set_cpu_cap(c, X86_FEATURE_K8);
622
623 if (cpu_has_xmm2) {
624 /* MFENCE stops RDTSC speculation */
625 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
626 }
627
628#ifdef CONFIG_X86_64
629 if (c->x86 == 0x10) {
630 /* do this for boot cpu */
631 if (c == &boot_cpu_data)
632 check_enable_amd_mmconf_dmi();
633
634 fam10h_check_enable_mmcfg();
635 }
636
637 if (c == &boot_cpu_data && c->x86 >= 0xf) {
638 unsigned long long tseg;
639
640 /*
641 * Split up direct mapping around the TSEG SMM area.
642 * Don't do it for gbpages because there seems very little
643 * benefit in doing so.
644 */
645 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
646 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
647 if ((tseg>>PMD_SHIFT) <
648 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
649 ((tseg>>PMD_SHIFT) <
650 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
651 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
652 set_memory_4k((unsigned long)__va(tseg), 1);
653 }
654 }
655#endif
656
657 /*
658 * Family 0x12 and above processors have APIC timer
659 * running in deep C states.
660 */
661 if (c->x86 > 0x11)
662 set_cpu_cap(c, X86_FEATURE_ARAT);
663
664 /*
665 * Disable GART TLB Walk Errors on Fam10h. We do this here
666 * because this is always needed when GART is enabled, even in a
667 * kernel which has no MCE support built in.
668 */
669 if (c->x86 == 0x10) {
670 /*
671 * BIOS should disable GartTlbWlk Errors themself. If
672 * it doesn't do it here as suggested by the BKDG.
673 *
674 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
675 */
676 u64 mask;
677 int err;
678
679 err = rdmsrl_safe(MSR_AMD64_MCx_MASK(4), &mask);
680 if (err == 0) {
681 mask |= (1 << 10);
682 checking_wrmsrl(MSR_AMD64_MCx_MASK(4), mask);
683 }
684 }
685
686 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
687}
688
689#ifdef CONFIG_X86_32
690static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
691 unsigned int size)
692{
693 /* AMD errata T13 (order #21922) */
694 if ((c->x86 == 6)) {
695 /* Duron Rev A0 */
696 if (c->x86_model == 3 && c->x86_mask == 0)
697 size = 64;
698 /* Tbird rev A1/A2 */
699 if (c->x86_model == 4 &&
700 (c->x86_mask == 0 || c->x86_mask == 1))
701 size = 256;
702 }
703 return size;
704}
705#endif
706
707static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
708 .c_vendor = "AMD",
709 .c_ident = { "AuthenticAMD" },
710#ifdef CONFIG_X86_32
711 .c_models = {
712 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
713 {
714 [3] = "486 DX/2",
715 [7] = "486 DX/2-WB",
716 [8] = "486 DX/4",
717 [9] = "486 DX/4-WB",
718 [14] = "Am5x86-WT",
719 [15] = "Am5x86-WB"
720 }
721 },
722 },
723 .c_size_cache = amd_size_cache,
724#endif
725 .c_early_init = early_init_amd,
726 .c_bsp_init = bsp_init_amd,
727 .c_init = init_amd,
728 .c_x86_vendor = X86_VENDOR_AMD,
729};
730
731cpu_dev_register(amd_cpu_dev);
732
733/*
734 * AMD errata checking
735 *
736 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
737 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
738 * have an OSVW id assigned, which it takes as first argument. Both take a
739 * variable number of family-specific model-stepping ranges created by
740 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
741 * int[] in arch/x86/include/asm/processor.h.
742 *
743 * Example:
744 *
745 * const int amd_erratum_319[] =
746 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
747 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
748 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
749 */
750
751const int amd_erratum_400[] =
752 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
753 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
754EXPORT_SYMBOL_GPL(amd_erratum_400);
755
756const int amd_erratum_383[] =
757 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
758EXPORT_SYMBOL_GPL(amd_erratum_383);
759
760bool cpu_has_amd_erratum(const int *erratum)
761{
762 struct cpuinfo_x86 *cpu = __this_cpu_ptr(&cpu_info);
763 int osvw_id = *erratum++;
764 u32 range;
765 u32 ms;
766
767 /*
768 * If called early enough that current_cpu_data hasn't been initialized
769 * yet, fall back to boot_cpu_data.
770 */
771 if (cpu->x86 == 0)
772 cpu = &boot_cpu_data;
773
774 if (cpu->x86_vendor != X86_VENDOR_AMD)
775 return false;
776
777 if (osvw_id >= 0 && osvw_id < 65536 &&
778 cpu_has(cpu, X86_FEATURE_OSVW)) {
779 u64 osvw_len;
780
781 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
782 if (osvw_id < osvw_len) {
783 u64 osvw_bits;
784
785 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
786 osvw_bits);
787 return osvw_bits & (1ULL << (osvw_id & 0x3f));
788 }
789 }
790
791 /* OSVW unavailable or ID unknown, match family-model-stepping range */
792 ms = (cpu->x86_model << 4) | cpu->x86_mask;
793 while ((range = *erratum++))
794 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
795 (ms >= AMD_MODEL_RANGE_START(range)) &&
796 (ms <= AMD_MODEL_RANGE_END(range)))
797 return true;
798
799 return false;
800}
801
802EXPORT_SYMBOL_GPL(cpu_has_amd_erratum);