blob: 9d92e19039f05f84b4c172de7c1c243a4dad72e0 [file] [log] [blame]
Paul Gortmaker69c60c82011-05-26 12:22:53 -04001#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07002#include <linux/init.h>
3#include <linux/bitops.h>
Stephen Rothwell5cdd1742011-08-10 11:49:56 +10004#include <linux/elf.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/mm.h>
Yinghai Lu8d71a2e2008-09-07 17:58:53 -07006
Alan Cox8bdbd962009-07-04 00:35:45 +01007#include <linux/io.h>
Borislav Petkovc98fdea2012-02-07 13:08:52 +01008#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <asm/processor.h>
Andi Kleend3f7eae2007-08-10 22:31:07 +020010#include <asm/apic.h>
Yinghai Lu1f442d72009-03-07 23:46:26 -080011#include <asm/cpu.h>
Andreas Herrmann42937e82009-06-08 15:55:09 +020012#include <asm/pci-direct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013
Yinghai Lu8d71a2e2008-09-07 17:58:53 -070014#ifdef CONFIG_X86_64
15# include <asm/numa_64.h>
16# include <asm/mmconfig.h>
17# include <asm/cacheflush.h>
18#endif
19
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include "cpu.h"
21
Borislav Petkov2c929ce2012-06-01 16:52:38 +020022static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
23{
24 struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
25 u32 gprs[8] = { 0 };
26 int err;
27
28 WARN_ONCE((c->x86 != 0xf), "%s should only be used on K8!\n", __func__);
29
30 gprs[1] = msr;
31 gprs[7] = 0x9c5a203a;
32
33 err = rdmsr_safe_regs(gprs);
34
35 *p = gprs[0] | ((u64)gprs[2] << 32);
36
37 return err;
38}
39
40static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
41{
42 struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
43 u32 gprs[8] = { 0 };
44
45 WARN_ONCE((c->x86 != 0xf), "%s should only be used on K8!\n", __func__);
46
47 gprs[0] = (u32)val;
48 gprs[1] = msr;
49 gprs[2] = val >> 32;
50 gprs[7] = 0x9c5a203a;
51
52 return wrmsr_safe_regs(gprs);
53}
54
Yinghai Lu6c62aa42008-09-07 17:58:54 -070055#ifdef CONFIG_X86_32
Linus Torvalds1da177e2005-04-16 15:20:36 -070056/*
57 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
58 * misexecution of code under Linux. Owners of such processors should
59 * contact AMD for precise details and a CPU swap.
60 *
61 * See http://www.multimania.com/poulot/k6bug.html
Andreas Herrmannd7de8642012-04-11 17:12:38 +020062 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
63 * (Publication # 21266 Issue Date: August 1998)
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 *
65 * The following test is erm.. interesting. AMD neglected to up
66 * the chip setting when fixing the bug but they also tweaked some
67 * performance at the same time..
68 */
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +010069
Linus Torvalds1da177e2005-04-16 15:20:36 -070070extern void vide(void);
71__asm__(".align 4\nvide: ret");
72
Yinghai Lu11fdd252008-09-07 17:58:50 -070073static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
74{
75/*
76 * General Systems BIOSen alias the cpu frequency registers
77 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
78 * drivers subsequently pokes it, and changes the CPU speed.
79 * Workaround : Remove the unneeded alias.
80 */
81#define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
82#define CBAR_ENB (0x80000000)
83#define CBAR_KEY (0X000000CB)
84 if (c->x86_model == 9 || c->x86_model == 10) {
Alan Cox8bdbd962009-07-04 00:35:45 +010085 if (inl(CBAR) & CBAR_ENB)
86 outl(0 | CBAR_KEY, CBAR);
Yinghai Lu11fdd252008-09-07 17:58:50 -070087 }
88}
89
90
91static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
92{
93 u32 l, h;
94 int mbytes = num_physpages >> (20-PAGE_SHIFT);
95
96 if (c->x86_model < 6) {
97 /* Based on AMD doc 20734R - June 2000 */
98 if (c->x86_model == 0) {
99 clear_cpu_cap(c, X86_FEATURE_APIC);
100 set_cpu_cap(c, X86_FEATURE_PGE);
101 }
102 return;
103 }
104
105 if (c->x86_model == 6 && c->x86_mask == 1) {
106 const int K6_BUG_LOOP = 1000000;
107 int n;
108 void (*f_vide)(void);
109 unsigned long d, d2;
110
111 printk(KERN_INFO "AMD K6 stepping B detected - ");
112
113 /*
114 * It looks like AMD fixed the 2.6.2 bug and improved indirect
115 * calls at the same time.
116 */
117
118 n = K6_BUG_LOOP;
119 f_vide = vide;
120 rdtscl(d);
121 while (n--)
122 f_vide();
123 rdtscl(d2);
124 d = d2-d;
125
126 if (d > 20*K6_BUG_LOOP)
Alan Cox8bdbd962009-07-04 00:35:45 +0100127 printk(KERN_CONT
128 "system stability may be impaired when more than 32 MB are used.\n");
Yinghai Lu11fdd252008-09-07 17:58:50 -0700129 else
Alan Cox8bdbd962009-07-04 00:35:45 +0100130 printk(KERN_CONT "probably OK (after B9730xxxx).\n");
Yinghai Lu11fdd252008-09-07 17:58:50 -0700131 }
132
133 /* K6 with old style WHCR */
134 if (c->x86_model < 8 ||
135 (c->x86_model == 8 && c->x86_mask < 8)) {
136 /* We can only write allocate on the low 508Mb */
137 if (mbytes > 508)
138 mbytes = 508;
139
140 rdmsr(MSR_K6_WHCR, l, h);
141 if ((l&0x0000FFFF) == 0) {
142 unsigned long flags;
143 l = (1<<0)|((mbytes/4)<<1);
144 local_irq_save(flags);
145 wbinvd();
146 wrmsr(MSR_K6_WHCR, l, h);
147 local_irq_restore(flags);
148 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
149 mbytes);
150 }
151 return;
152 }
153
154 if ((c->x86_model == 8 && c->x86_mask > 7) ||
155 c->x86_model == 9 || c->x86_model == 13) {
156 /* The more serious chips .. */
157
158 if (mbytes > 4092)
159 mbytes = 4092;
160
161 rdmsr(MSR_K6_WHCR, l, h);
162 if ((l&0xFFFF0000) == 0) {
163 unsigned long flags;
164 l = ((mbytes>>2)<<22)|(1<<16);
165 local_irq_save(flags);
166 wbinvd();
167 wrmsr(MSR_K6_WHCR, l, h);
168 local_irq_restore(flags);
169 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
170 mbytes);
171 }
172
173 return;
174 }
175
176 if (c->x86_model == 10) {
177 /* AMD Geode LX is model 10 */
178 /* placeholder for any needed mods */
179 return;
180 }
181}
182
Yinghai Lu1f442d72009-03-07 23:46:26 -0800183static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
184{
Yinghai Lu1f442d72009-03-07 23:46:26 -0800185 /* calling is from identify_secondary_cpu() ? */
Robert Richterf6e9456c2010-07-21 19:03:58 +0200186 if (!c->cpu_index)
Yinghai Lu1f442d72009-03-07 23:46:26 -0800187 return;
188
189 /*
190 * Certain Athlons might work (for various values of 'work') in SMP
191 * but they are not certified as MP capable.
192 */
193 /* Athlon 660/661 is valid. */
194 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
195 (c->x86_mask == 1)))
196 goto valid_k7;
197
198 /* Duron 670 is valid */
199 if ((c->x86_model == 7) && (c->x86_mask == 0))
200 goto valid_k7;
201
202 /*
203 * Athlon 662, Duron 671, and Athlon >model 7 have capability
204 * bit. It's worth noting that the A5 stepping (662) of some
205 * Athlon XP's have the MP bit set.
206 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
207 * more.
208 */
209 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
210 ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
211 (c->x86_model > 7))
212 if (cpu_has_mp)
213 goto valid_k7;
214
215 /* If we get here, not a certified SMP capable AMD system. */
216
217 /*
218 * Don't taint if we are running SMP kernel on a single non-MP
219 * approved Athlon
220 */
221 WARN_ONCE(1, "WARNING: This combination of AMD"
Michael Tokarev7da8b6d2009-07-22 17:50:23 +0400222 " processors is not suitable for SMP.\n");
Yinghai Lu1f442d72009-03-07 23:46:26 -0800223 if (!test_taint(TAINT_UNSAFE_SMP))
224 add_taint(TAINT_UNSAFE_SMP);
225
226valid_k7:
227 ;
Yinghai Lu1f442d72009-03-07 23:46:26 -0800228}
229
Yinghai Lu11fdd252008-09-07 17:58:50 -0700230static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
231{
232 u32 l, h;
233
234 /*
235 * Bit 15 of Athlon specific MSR 15, needs to be 0
236 * to enable SSE on Palomino/Morgan/Barton CPU's.
237 * If the BIOS didn't enable it already, enable it here.
238 */
239 if (c->x86_model >= 6 && c->x86_model <= 10) {
240 if (!cpu_has(c, X86_FEATURE_XMM)) {
241 printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
242 rdmsr(MSR_K7_HWCR, l, h);
243 l &= ~0x00008000;
244 wrmsr(MSR_K7_HWCR, l, h);
245 set_cpu_cap(c, X86_FEATURE_XMM);
246 }
247 }
248
249 /*
250 * It's been determined by AMD that Athlons since model 8 stepping 1
251 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
252 * As per AMD technical note 27212 0.2
253 */
254 if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
255 rdmsr(MSR_K7_CLK_CTL, l, h);
256 if ((l & 0xfff00000) != 0x20000000) {
Alan Cox8bdbd962009-07-04 00:35:45 +0100257 printk(KERN_INFO
258 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
259 l, ((l & 0x000fffff)|0x20000000));
Yinghai Lu11fdd252008-09-07 17:58:50 -0700260 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
261 }
262 }
263
264 set_cpu_cap(c, X86_FEATURE_K7);
Yinghai Lu1f442d72009-03-07 23:46:26 -0800265
266 amd_k7_smp_check(c);
Yinghai Lu11fdd252008-09-07 17:58:50 -0700267}
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700268#endif
269
Tejun Heo645a7912011-01-23 14:37:40 +0100270#ifdef CONFIG_NUMA
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100271/*
272 * To workaround broken NUMA config. Read the comment in
273 * srat_detect_node().
274 */
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700275static int __cpuinit nearby_node(int apicid)
276{
277 int i, node;
278
279 for (i = apicid - 1; i >= 0; i--) {
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100280 node = __apicid_to_node[i];
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700281 if (node != NUMA_NO_NODE && node_online(node))
282 return node;
283 }
284 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100285 node = __apicid_to_node[i];
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700286 if (node != NUMA_NO_NODE && node_online(node))
287 return node;
288 }
289 return first_node(node_online_map); /* Shouldn't happen */
290}
291#endif
Yinghai Lu11fdd252008-09-07 17:58:50 -0700292
293/*
Andreas Herrmann23588c32010-09-30 14:36:28 +0200294 * Fixup core topology information for
295 * (1) AMD multi-node processors
296 * Assumption: Number of cores in each internal node is the same.
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200297 * (2) AMD processors supporting compute units
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200298 */
299#ifdef CONFIG_X86_HT
Andreas Herrmann23588c32010-09-30 14:36:28 +0200300static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200301{
Borislav Petkov9e815092011-02-14 18:14:51 +0100302 u32 nodes, cores_per_cu = 1;
Andreas Herrmann23588c32010-09-30 14:36:28 +0200303 u8 node_id;
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200304 int cpu = smp_processor_id();
305
Andreas Herrmann23588c32010-09-30 14:36:28 +0200306 /* get information required for multi-node processors */
307 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200308 u32 eax, ebx, ecx, edx;
309
310 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
311 nodes = ((ecx >> 8) & 7) + 1;
312 node_id = ecx & 7;
313
314 /* get compute unit information */
315 smp_num_siblings = ((ebx >> 8) & 3) + 1;
316 c->compute_unit_id = ebx & 0xff;
Borislav Petkov9e815092011-02-14 18:14:51 +0100317 cores_per_cu += ((ebx >> 8) & 3);
Andreas Herrmann23588c32010-09-30 14:36:28 +0200318 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200319 u64 value;
320
Andreas Herrmann23588c32010-09-30 14:36:28 +0200321 rdmsrl(MSR_FAM10H_NODE_ID, value);
322 nodes = ((value >> 3) & 7) + 1;
323 node_id = value & 7;
324 } else
Andreas Herrmann9d260eb2009-12-16 15:43:55 +0100325 return;
326
Andreas Herrmann23588c32010-09-30 14:36:28 +0200327 /* fixup multi-node processor information */
328 if (nodes > 1) {
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200329 u32 cores_per_node;
Andreas Herrmannd5185732011-01-24 16:05:40 +0100330 u32 cus_per_node;
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200331
Andreas Herrmann23588c32010-09-30 14:36:28 +0200332 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
333 cores_per_node = c->x86_max_cores / nodes;
Andreas Herrmannd5185732011-01-24 16:05:40 +0100334 cus_per_node = cores_per_node / cores_per_cu;
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200335
Andreas Herrmann23588c32010-09-30 14:36:28 +0200336 /* store NodeID, use llc_shared_map to store sibling info */
337 per_cpu(cpu_llc_id, cpu) = node_id;
Andreas Herrmann9d260eb2009-12-16 15:43:55 +0100338
Borislav Petkov9e815092011-02-14 18:14:51 +0100339 /* core id has to be in the [0 .. cores_per_node - 1] range */
Andreas Herrmannd5185732011-01-24 16:05:40 +0100340 c->cpu_core_id %= cores_per_node;
341 c->compute_unit_id %= cus_per_node;
Andreas Herrmann23588c32010-09-30 14:36:28 +0200342 }
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200343}
344#endif
345
346/*
Yinghai Lu11fdd252008-09-07 17:58:50 -0700347 * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
348 * Assumes number of cores is a power of two.
349 */
350static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
351{
352#ifdef CONFIG_X86_HT
353 unsigned bits;
Andreas Herrmann99bd0c02009-06-19 10:59:09 +0200354 int cpu = smp_processor_id();
Yinghai Lu11fdd252008-09-07 17:58:50 -0700355
356 bits = c->x86_coreid_bits;
Yinghai Lu11fdd252008-09-07 17:58:50 -0700357 /* Low order bits define the core id (index of core in socket) */
358 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
359 /* Convert the initial APIC ID into the socket ID */
360 c->phys_proc_id = c->initial_apicid >> bits;
Andreas Herrmann99bd0c02009-06-19 10:59:09 +0200361 /* use socket ID also for last level cache */
362 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
Andreas Herrmann23588c32010-09-30 14:36:28 +0200363 amd_get_topology(c);
Yinghai Lu11fdd252008-09-07 17:58:50 -0700364#endif
365}
366
Andreas Herrmann6a812692009-09-16 11:33:40 +0200367int amd_get_nb_id(int cpu)
368{
369 int id = 0;
370#ifdef CONFIG_SMP
371 id = per_cpu(cpu_llc_id, cpu);
372#endif
373 return id;
374}
375EXPORT_SYMBOL_GPL(amd_get_nb_id);
376
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700377static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
378{
Tejun Heo645a7912011-01-23 14:37:40 +0100379#ifdef CONFIG_NUMA
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700380 int cpu = smp_processor_id();
381 int node;
Yinghai Lu0d96b9f2009-08-29 13:17:14 -0700382 unsigned apicid = c->apicid;
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700383
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100384 node = numa_cpu_node(cpu);
385 if (node == NUMA_NO_NODE)
386 node = per_cpu(cpu_llc_id, cpu);
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200387
Daniel J Blueman64be4c12011-12-05 16:20:37 +0800388 /*
Andreas Herrmann68894632012-04-02 18:06:48 +0200389 * On multi-fabric platform (e.g. Numascale NumaChip) a
390 * platform-specific handler needs to be called to fixup some
391 * IDs of the CPU.
Daniel J Blueman64be4c12011-12-05 16:20:37 +0800392 */
Andreas Herrmann68894632012-04-02 18:06:48 +0200393 if (x86_cpuinit.fixup_cpu_id)
Daniel J Blueman64be4c12011-12-05 16:20:37 +0800394 x86_cpuinit.fixup_cpu_id(c, node);
395
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700396 if (!node_online(node)) {
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100397 /*
398 * Two possibilities here:
399 *
400 * - The CPU is missing memory and no node was created. In
401 * that case try picking one from a nearby CPU.
402 *
403 * - The APIC IDs differ from the HyperTransport node IDs
404 * which the K8 northbridge parsing fills in. Assume
405 * they are all increased by a constant offset, but in
406 * the same order as the HT nodeids. If that doesn't
407 * result in a usable node fall back to the path for the
408 * previous case.
409 *
410 * This workaround operates directly on the mapping between
411 * APIC ID and NUMA node, assuming certain relationship
412 * between APIC ID, HT node ID and NUMA topology. As going
413 * through CPU mapping may alter the outcome, directly
414 * access __apicid_to_node[].
415 */
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700416 int ht_nodeid = c->initial_apicid;
417
418 if (ht_nodeid >= 0 &&
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100419 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
420 node = __apicid_to_node[ht_nodeid];
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700421 /* Pick a nearby node */
422 if (!node_online(node))
423 node = nearby_node(apicid);
424 }
425 numa_set_node(cpu, node);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700426#endif
427}
428
Yinghai Lu11fdd252008-09-07 17:58:50 -0700429static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
430{
431#ifdef CONFIG_X86_HT
432 unsigned bits, ecx;
433
434 /* Multi core CPU? */
435 if (c->extended_cpuid_level < 0x80000008)
436 return;
437
438 ecx = cpuid_ecx(0x80000008);
439
440 c->x86_max_cores = (ecx & 0xff) + 1;
441
442 /* CPU telling us the core id bits shift? */
443 bits = (ecx >> 12) & 0xF;
444
445 /* Otherwise recompute */
446 if (bits == 0) {
447 while ((1 << bits) < c->x86_max_cores)
448 bits++;
449 }
450
451 c->x86_coreid_bits = bits;
452#endif
453}
454
Borislav Petkov8fa8b032011-08-05 20:04:09 +0200455static void __cpuinit bsp_init_amd(struct cpuinfo_x86 *c)
456{
457 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
458
459 if (c->x86 > 0x10 ||
460 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
461 u64 val;
462
463 rdmsrl(MSR_K7_HWCR, val);
464 if (!(val & BIT(24)))
465 printk(KERN_WARNING FW_BUG "TSC doesn't count "
466 "with P0 frequency!\n");
467 }
468 }
469
470 if (c->x86 == 0x15) {
471 unsigned long upperbit;
472 u32 cpuid, assoc;
473
474 cpuid = cpuid_edx(0x80000005);
475 assoc = cpuid >> 16 & 0xff;
476 upperbit = ((cpuid >> 24) << 10) / assoc;
477
478 va_align.mask = (upperbit - 1) & PAGE_MASK;
479 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
480 }
481}
482
Thomas Petazzoni03ae5762008-02-15 12:00:23 +0100483static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
Andi Kleen2b16a232008-01-30 13:32:40 +0100484{
Yinghai Lu11fdd252008-09-07 17:58:50 -0700485 early_init_amd_mc(c);
486
Venki Pallipadi40fb1712008-11-17 16:11:37 -0800487 /*
488 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
489 * with P/T states and does not stop in deep C-states
490 */
491 if (c->x86_power & (1 << 8)) {
Yinghai Lue3224232008-09-06 01:52:28 -0700492 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
Venki Pallipadi40fb1712008-11-17 16:11:37 -0800493 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
Borislav Petkovc98fdea2012-02-07 13:08:52 +0100494 if (!check_tsc_unstable())
495 sched_clock_stable = 1;
Venki Pallipadi40fb1712008-11-17 16:11:37 -0800496 }
Yinghai Lu5fef55f2008-09-04 21:09:43 +0200497
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700498#ifdef CONFIG_X86_64
499 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
500#else
Yinghai Lu5fef55f2008-09-04 21:09:43 +0200501 /* Set MTRR capability flag if appropriate */
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700502 if (c->x86 == 5)
503 if (c->x86_model == 13 || c->x86_model == 9 ||
504 (c->x86_model == 8 && c->x86_mask >= 8))
505 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
506#endif
Andreas Herrmann42937e82009-06-08 15:55:09 +0200507#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
508 /* check CPU config space for extended APIC ID */
Jeremy Fitzhardinge2cb07862009-07-22 09:59:35 -0700509 if (cpu_has_apic && c->x86 >= 0xf) {
Andreas Herrmann42937e82009-06-08 15:55:09 +0200510 unsigned int val;
511 val = read_pci_config(0, 24, 0, 0x68);
512 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
513 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
514 }
515#endif
Andi Kleen2b16a232008-01-30 13:32:40 +0100516}
517
Magnus Dammb4af3f72006-09-26 10:52:36 +0200518static void __cpuinit init_amd(struct cpuinfo_x86 *c)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519{
Linus Torvalds8e8da022011-12-04 11:57:09 -0800520 u32 dummy;
521
Andi Kleen7d318d72005-09-29 22:05:55 +0200522#ifdef CONFIG_SMP
Andi Kleen3c92c2b2005-10-11 01:28:33 +0200523 unsigned long long value;
Andi Kleen7d318d72005-09-29 22:05:55 +0200524
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100525 /*
526 * Disable TLB flush filter by setting HWCR.FFDIS on K8
Andi Kleen7d318d72005-09-29 22:05:55 +0200527 * bit 6 of msr C001_0015
528 *
529 * Errata 63 for SH-B3 steppings
530 * Errata 122 for all steppings (F+ have it disabled by default)
531 */
Yinghai Lu11fdd252008-09-07 17:58:50 -0700532 if (c->x86 == 0xf) {
Andi Kleen7d318d72005-09-29 22:05:55 +0200533 rdmsrl(MSR_K7_HWCR, value);
534 value |= 1 << 6;
535 wrmsrl(MSR_K7_HWCR, value);
536 }
537#endif
538
Andi Kleen2b16a232008-01-30 13:32:40 +0100539 early_init_amd(c);
540
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 /*
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100542 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
Ingo Molnar16282a82008-02-26 08:49:57 +0100543 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100544 */
Ingo Molnar16282a82008-02-26 08:49:57 +0100545 clear_cpu_cap(c, 0*32+31);
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100546
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700547#ifdef CONFIG_X86_64
548 /* On C+ stepping K8 rep microcode works well for copy/memset */
549 if (c->x86 == 0xf) {
550 u32 level;
551
552 level = cpuid_eax(1);
Alan Cox8bdbd962009-07-04 00:35:45 +0100553 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700554 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
Kevin Winchesterfbd8b182009-08-10 19:56:45 -0300555
556 /*
557 * Some BIOSes incorrectly force this feature, but only K8
558 * revision D (model = 0x14) and later actually support it.
Borislav Petkov6b0f43d2009-08-31 09:50:11 +0200559 * (AMD Erratum #110, docId: 25759).
Kevin Winchesterfbd8b182009-08-10 19:56:45 -0300560 */
Borislav Petkov6b0f43d2009-08-31 09:50:11 +0200561 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
562 u64 val;
563
Kevin Winchesterfbd8b182009-08-10 19:56:45 -0300564 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
Borislav Petkov6b0f43d2009-08-31 09:50:11 +0200565 if (!rdmsrl_amd_safe(0xc001100d, &val)) {
566 val &= ~(1ULL << 32);
567 wrmsrl_amd_safe(0xc001100d, val);
568 }
569 }
570
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700571 }
Borislav Petkov12d8a962010-06-02 20:29:21 +0200572 if (c->x86 >= 0x10)
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700573 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
Yinghai Lu0d96b9f2009-08-29 13:17:14 -0700574
575 /* get apicid instead of initial apic id from cpuid */
576 c->apicid = hard_smp_processor_id();
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700577#else
578
579 /*
580 * FIXME: We should handle the K5 here. Set up the write
581 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
582 * no bus pipeline)
583 */
584
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100585 switch (c->x86) {
586 case 4:
Yinghai Lu11fdd252008-09-07 17:58:50 -0700587 init_amd_k5(c);
588 break;
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100589 case 5:
Yinghai Lu11fdd252008-09-07 17:58:50 -0700590 init_amd_k6(c);
591 break;
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100592 case 6: /* An Athlon/Duron */
Yinghai Lu11fdd252008-09-07 17:58:50 -0700593 init_amd_k7(c);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594 break;
Andi Kleen67cddd92007-07-21 17:10:03 +0200595 }
Andi Kleen3556ddf2007-04-02 12:14:12 +0200596
Andi Kleenc12ceb72007-05-21 14:31:47 +0200597 /* K6s reports MCEs but don't actually have all the MSRs */
598 if (c->x86 < 6)
Ingo Molnar16282a82008-02-26 08:49:57 +0100599 clear_cpu_cap(c, X86_FEATURE_MCE);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700600#endif
Andi Kleende421862008-01-30 13:32:37 +0100601
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700602 /* Enable workaround for FXSAVE leak */
Yinghai Lu11fdd252008-09-07 17:58:50 -0700603 if (c->x86 >= 6)
604 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
605
606 if (!c->x86_model_id[0]) {
607 switch (c->x86) {
608 case 0xf:
609 /* Should distinguish Models here, but this is only
610 a fallback anyways. */
611 strcpy(c->x86_model_id, "Hammer");
612 break;
613 }
614 }
615
Andreas Herrmannf7f286a2012-04-03 12:13:07 +0200616 /* re-enable TopologyExtensions if switched off by BIOS */
617 if ((c->x86 == 0x15) &&
618 (c->x86_model >= 0x10) && (c->x86_model <= 0x1f) &&
619 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
620 u64 val;
621
Andre Przywara169e9cb2012-06-01 16:52:37 +0200622 if (!rdmsrl_safe(0xc0011005, &val)) {
Andreas Herrmannf7f286a2012-04-03 12:13:07 +0200623 val |= 1ULL << 54;
H. Peter Anvin715c85b2012-06-07 13:32:04 -0700624 wrmsrl_safe(0xc0011005, val);
Andreas Herrmannf7f286a2012-04-03 12:13:07 +0200625 rdmsrl(0xc0011005, val);
626 if (val & (1ULL << 54)) {
627 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
628 printk(KERN_INFO FW_INFO "CPU: Re-enabling "
629 "disabled Topology Extensions Support\n");
630 }
631 }
632 }
633
Borislav Petkov27c13ec2009-11-21 14:01:45 +0100634 cpu_detect_cache_sizes(c);
Yinghai Lu11fdd252008-09-07 17:58:50 -0700635
636 /* Multi core CPU? */
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700637 if (c->extended_cpuid_level >= 0x80000008) {
Yinghai Lu11fdd252008-09-07 17:58:50 -0700638 amd_detect_cmp(c);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700639 srat_detect_node(c);
640 }
Yinghai Lu11fdd252008-09-07 17:58:50 -0700641
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700642#ifdef CONFIG_X86_32
Yinghai Lu11fdd252008-09-07 17:58:50 -0700643 detect_ht(c);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700644#endif
Yinghai Lu11fdd252008-09-07 17:58:50 -0700645
646 if (c->extended_cpuid_level >= 0x80000006) {
Andreas Herrmannd9fadd72010-09-02 15:37:10 +0200647 if (cpuid_edx(0x80000006) & 0xf000)
Yinghai Lu11fdd252008-09-07 17:58:50 -0700648 num_cache_leaves = 4;
649 else
650 num_cache_leaves = 3;
651 }
652
Borislav Petkov12d8a962010-06-02 20:29:21 +0200653 if (c->x86 >= 0xf)
Yinghai Lu11fdd252008-09-07 17:58:50 -0700654 set_cpu_cap(c, X86_FEATURE_K8);
655
656 if (cpu_has_xmm2) {
657 /* MFENCE stops RDTSC speculation */
Ingo Molnar16282a82008-02-26 08:49:57 +0100658 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
Yinghai Lu11fdd252008-09-07 17:58:50 -0700659 }
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700660
661#ifdef CONFIG_X86_64
662 if (c->x86 == 0x10) {
663 /* do this for boot cpu */
664 if (c == &boot_cpu_data)
665 check_enable_amd_mmconf_dmi();
666
667 fam10h_check_enable_mmcfg();
668 }
669
Borislav Petkov12d8a962010-06-02 20:29:21 +0200670 if (c == &boot_cpu_data && c->x86 >= 0xf) {
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700671 unsigned long long tseg;
672
673 /*
674 * Split up direct mapping around the TSEG SMM area.
675 * Don't do it for gbpages because there seems very little
676 * benefit in doing so.
677 */
678 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
Alan Cox8bdbd962009-07-04 00:35:45 +0100679 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
680 if ((tseg>>PMD_SHIFT) <
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700681 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
Alan Cox8bdbd962009-07-04 00:35:45 +0100682 ((tseg>>PMD_SHIFT) <
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700683 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
Alan Cox8bdbd962009-07-04 00:35:45 +0100684 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
685 set_memory_4k((unsigned long)__va(tseg), 1);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700686 }
687 }
688#endif
Boris Ostrovskyb87cf802011-03-15 12:13:44 -0400689
Boris Ostrovskye9cdd342011-05-26 11:19:52 -0400690 /*
691 * Family 0x12 and above processors have APIC timer
692 * running in deep C states.
693 */
694 if (c->x86 > 0x11)
Boris Ostrovskyb87cf802011-03-15 12:13:44 -0400695 set_cpu_cap(c, X86_FEATURE_ARAT);
Joerg Roedel5bbc0972011-04-15 14:47:40 +0200696
697 /*
698 * Disable GART TLB Walk Errors on Fam10h. We do this here
699 * because this is always needed when GART is enabled, even in a
700 * kernel which has no MCE support built in.
701 */
702 if (c->x86 == 0x10) {
703 /*
704 * BIOS should disable GartTlbWlk Errors themself. If
705 * it doesn't do it here as suggested by the BKDG.
706 *
707 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
708 */
709 u64 mask;
Roedel, Joergd47cc0d2011-05-19 11:13:39 +0200710 int err;
Joerg Roedel5bbc0972011-04-15 14:47:40 +0200711
Roedel, Joergd47cc0d2011-05-19 11:13:39 +0200712 err = rdmsrl_safe(MSR_AMD64_MCx_MASK(4), &mask);
713 if (err == 0) {
714 mask |= (1 << 10);
H. Peter Anvin715c85b2012-06-07 13:32:04 -0700715 wrmsrl_safe(MSR_AMD64_MCx_MASK(4), mask);
Roedel, Joergd47cc0d2011-05-19 11:13:39 +0200716 }
Joerg Roedel5bbc0972011-04-15 14:47:40 +0200717 }
Linus Torvalds8e8da022011-12-04 11:57:09 -0800718
719 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720}
721
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700722#ifdef CONFIG_X86_32
Alan Cox8bdbd962009-07-04 00:35:45 +0100723static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
724 unsigned int size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725{
726 /* AMD errata T13 (order #21922) */
727 if ((c->x86 == 6)) {
Alan Cox8bdbd962009-07-04 00:35:45 +0100728 /* Duron Rev A0 */
729 if (c->x86_model == 3 && c->x86_mask == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 size = 64;
Alan Cox8bdbd962009-07-04 00:35:45 +0100731 /* Tbird rev A1/A2 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 if (c->x86_model == 4 &&
Alan Cox8bdbd962009-07-04 00:35:45 +0100733 (c->x86_mask == 0 || c->x86_mask == 1))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 size = 256;
735 }
736 return size;
737}
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700738#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739
Jan Beulich02dde8b2009-03-12 12:08:49 +0000740static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 .c_vendor = "AMD",
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100742 .c_ident = { "AuthenticAMD" },
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700743#ifdef CONFIG_X86_32
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 .c_models = {
745 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
746 {
747 [3] = "486 DX/2",
748 [7] = "486 DX/2-WB",
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100749 [8] = "486 DX/4",
750 [9] = "486 DX/4-WB",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 [14] = "Am5x86-WT",
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100752 [15] = "Am5x86-WB"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 }
754 },
755 },
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700756 .c_size_cache = amd_size_cache,
757#endif
Thomas Petazzoni03ae5762008-02-15 12:00:23 +0100758 .c_early_init = early_init_amd,
Borislav Petkov8fa8b032011-08-05 20:04:09 +0200759 .c_bsp_init = bsp_init_amd,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 .c_init = init_amd,
Yinghai Lu10a434f2008-09-04 21:09:45 +0200761 .c_x86_vendor = X86_VENDOR_AMD,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762};
763
Yinghai Lu10a434f2008-09-04 21:09:45 +0200764cpu_dev_register(amd_cpu_dev);
Hans Rosenfeldd78d6712010-07-28 19:09:30 +0200765
766/*
767 * AMD errata checking
768 *
769 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
770 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
771 * have an OSVW id assigned, which it takes as first argument. Both take a
772 * variable number of family-specific model-stepping ranges created by
773 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
774 * int[] in arch/x86/include/asm/processor.h.
775 *
776 * Example:
777 *
778 * const int amd_erratum_319[] =
779 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
780 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
781 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
782 */
783
Hans Rosenfeld9d8888c2010-07-28 19:09:31 +0200784const int amd_erratum_400[] =
Borislav Petkov328935e2011-05-17 14:55:18 +0200785 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
Hans Rosenfeld9d8888c2010-07-28 19:09:31 +0200786 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
H. Peter Anvina5b91602010-07-28 16:23:20 -0700787EXPORT_SYMBOL_GPL(amd_erratum_400);
Hans Rosenfeld9d8888c2010-07-28 19:09:31 +0200788
Hans Rosenfeld1be85a62010-07-28 19:09:32 +0200789const int amd_erratum_383[] =
790 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
H. Peter Anvina5b91602010-07-28 16:23:20 -0700791EXPORT_SYMBOL_GPL(amd_erratum_383);
Hans Rosenfeld9d8888c2010-07-28 19:09:31 +0200792
Hans Rosenfeldd78d6712010-07-28 19:09:30 +0200793bool cpu_has_amd_erratum(const int *erratum)
794{
Tejun Heo7b543a52010-12-18 16:30:05 +0100795 struct cpuinfo_x86 *cpu = __this_cpu_ptr(&cpu_info);
Hans Rosenfeldd78d6712010-07-28 19:09:30 +0200796 int osvw_id = *erratum++;
797 u32 range;
798 u32 ms;
799
800 /*
801 * If called early enough that current_cpu_data hasn't been initialized
802 * yet, fall back to boot_cpu_data.
803 */
804 if (cpu->x86 == 0)
805 cpu = &boot_cpu_data;
806
807 if (cpu->x86_vendor != X86_VENDOR_AMD)
808 return false;
809
810 if (osvw_id >= 0 && osvw_id < 65536 &&
811 cpu_has(cpu, X86_FEATURE_OSVW)) {
812 u64 osvw_len;
813
814 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
815 if (osvw_id < osvw_len) {
816 u64 osvw_bits;
817
818 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
819 osvw_bits);
820 return osvw_bits & (1ULL << (osvw_id & 0x3f));
821 }
822 }
823
824 /* OSVW unavailable or ID unknown, match family-model-stepping range */
Hans Rosenfeld07a77952010-08-18 16:19:50 +0200825 ms = (cpu->x86_model << 4) | cpu->x86_mask;
Hans Rosenfeldd78d6712010-07-28 19:09:30 +0200826 while ((range = *erratum++))
827 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
828 (ms >= AMD_MODEL_RANGE_START(range)) &&
829 (ms <= AMD_MODEL_RANGE_END(range)))
830 return true;
831
832 return false;
833}
H. Peter Anvina5b91602010-07-28 16:23:20 -0700834
835EXPORT_SYMBOL_GPL(cpu_has_amd_erratum);