blob: 8f5cabb3c5b055c41c0aa41fd4a37d227c06bfb0 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001#include <linux/init.h>
2#include <linux/bitops.h>
3#include <linux/mm.h>
Yinghai Lu8d71a2e2008-09-07 17:58:53 -07004
Alan Cox8bdbd962009-07-04 00:35:45 +01005#include <linux/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07006#include <asm/processor.h>
Andi Kleend3f7eae2007-08-10 22:31:07 +02007#include <asm/apic.h>
Yinghai Lu1f442d72009-03-07 23:46:26 -08008#include <asm/cpu.h>
Andreas Herrmann42937e82009-06-08 15:55:09 +02009#include <asm/pci-direct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010
Yinghai Lu8d71a2e2008-09-07 17:58:53 -070011#ifdef CONFIG_X86_64
12# include <asm/numa_64.h>
13# include <asm/mmconfig.h>
14# include <asm/cacheflush.h>
15#endif
16
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include "cpu.h"
18
Yinghai Lu6c62aa42008-09-07 17:58:54 -070019#ifdef CONFIG_X86_32
Linus Torvalds1da177e2005-04-16 15:20:36 -070020/*
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.
24 *
25 * See http://www.multimania.com/poulot/k6bug.html
26 * http://www.amd.com/K6/k6docs/revgd.html
27 *
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..
31 */
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +010032
Linus Torvalds1da177e2005-04-16 15:20:36 -070033extern void vide(void);
34__asm__(".align 4\nvide: ret");
35
Yinghai Lu11fdd252008-09-07 17:58:50 -070036static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
37{
38/*
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.
43 */
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) {
Alan Cox8bdbd962009-07-04 00:35:45 +010048 if (inl(CBAR) & CBAR_ENB)
49 outl(0 | CBAR_KEY, CBAR);
Yinghai Lu11fdd252008-09-07 17:58:50 -070050 }
51}
52
53
54static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
55{
56 u32 l, h;
57 int mbytes = num_physpages >> (20-PAGE_SHIFT);
58
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);
64 }
65 return;
66 }
67
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;
73
74 printk(KERN_INFO "AMD K6 stepping B detected - ");
75
76 /*
77 * It looks like AMD fixed the 2.6.2 bug and improved indirect
78 * calls at the same time.
79 */
80
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;
88
89 if (d > 20*K6_BUG_LOOP)
Alan Cox8bdbd962009-07-04 00:35:45 +010090 printk(KERN_CONT
91 "system stability may be impaired when more than 32 MB are used.\n");
Yinghai Lu11fdd252008-09-07 17:58:50 -070092 else
Alan Cox8bdbd962009-07-04 00:35:45 +010093 printk(KERN_CONT "probably OK (after B9730xxxx).\n");
Yinghai Lu11fdd252008-09-07 17:58:50 -070094 printk(KERN_INFO "Please see http://membres.lycos.fr/poulot/k6bug.html\n");
95 }
96
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;
103
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);
114 }
115 return;
116 }
117
118 if ((c->x86_model == 8 && c->x86_mask > 7) ||
119 c->x86_model == 9 || c->x86_model == 13) {
120 /* The more serious chips .. */
121
122 if (mbytes > 4092)
123 mbytes = 4092;
124
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);
135 }
136
137 return;
138 }
139
140 if (c->x86_model == 10) {
141 /* AMD Geode LX is model 10 */
142 /* placeholder for any needed mods */
143 return;
144 }
145}
146
Yinghai Lu1f442d72009-03-07 23:46:26 -0800147static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
148{
149#ifdef CONFIG_SMP
150 /* calling is from identify_secondary_cpu() ? */
Robert Richterf6e9456c2010-07-21 19:03:58 +0200151 if (!c->cpu_index)
Yinghai Lu1f442d72009-03-07 23:46:26 -0800152 return;
153
154 /*
155 * Certain Athlons might work (for various values of 'work') in SMP
156 * but they are not certified as MP capable.
157 */
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;
162
163 /* Duron 670 is valid */
164 if ((c->x86_model == 7) && (c->x86_mask == 0))
165 goto valid_k7;
166
167 /*
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.
173 */
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;
179
180 /* If we get here, not a certified SMP capable AMD system. */
181
182 /*
183 * Don't taint if we are running SMP kernel on a single non-MP
184 * approved Athlon
185 */
186 WARN_ONCE(1, "WARNING: This combination of AMD"
Michael Tokarev7da8b6d2009-07-22 17:50:23 +0400187 " processors is not suitable for SMP.\n");
Yinghai Lu1f442d72009-03-07 23:46:26 -0800188 if (!test_taint(TAINT_UNSAFE_SMP))
189 add_taint(TAINT_UNSAFE_SMP);
190
191valid_k7:
192 ;
193#endif
194}
195
Yinghai Lu11fdd252008-09-07 17:58:50 -0700196static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
197{
198 u32 l, h;
199
200 /*
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.
204 */
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);
212 }
213 }
214
215 /*
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
219 */
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) {
Alan Cox8bdbd962009-07-04 00:35:45 +0100223 printk(KERN_INFO
224 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
225 l, ((l & 0x000fffff)|0x20000000));
Yinghai Lu11fdd252008-09-07 17:58:50 -0700226 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
227 }
228 }
229
230 set_cpu_cap(c, X86_FEATURE_K7);
Yinghai Lu1f442d72009-03-07 23:46:26 -0800231
232 amd_k7_smp_check(c);
Yinghai Lu11fdd252008-09-07 17:58:50 -0700233}
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700234#endif
235
Tejun Heo645a7912011-01-23 14:37:40 +0100236#ifdef CONFIG_NUMA
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100237/*
238 * To workaround broken NUMA config. Read the comment in
239 * srat_detect_node().
240 */
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700241static int __cpuinit nearby_node(int apicid)
242{
243 int i, node;
244
245 for (i = apicid - 1; i >= 0; i--) {
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100246 node = __apicid_to_node[i];
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700247 if (node != NUMA_NO_NODE && node_online(node))
248 return node;
249 }
250 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100251 node = __apicid_to_node[i];
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700252 if (node != NUMA_NO_NODE && node_online(node))
253 return node;
254 }
255 return first_node(node_online_map); /* Shouldn't happen */
256}
257#endif
Yinghai Lu11fdd252008-09-07 17:58:50 -0700258
259/*
Andreas Herrmann23588c32010-09-30 14:36:28 +0200260 * Fixup core topology information for
261 * (1) AMD multi-node processors
262 * Assumption: Number of cores in each internal node is the same.
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200263 * (2) AMD processors supporting compute units
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200264 */
265#ifdef CONFIG_X86_HT
Andreas Herrmann23588c32010-09-30 14:36:28 +0200266static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200267{
Borislav Petkov9e815092011-02-14 18:14:51 +0100268 u32 nodes, cores_per_cu = 1;
Andreas Herrmann23588c32010-09-30 14:36:28 +0200269 u8 node_id;
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200270 int cpu = smp_processor_id();
271
Andreas Herrmann23588c32010-09-30 14:36:28 +0200272 /* get information required for multi-node processors */
273 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200274 u32 eax, ebx, ecx, edx;
275
276 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
277 nodes = ((ecx >> 8) & 7) + 1;
278 node_id = ecx & 7;
279
280 /* get compute unit information */
281 smp_num_siblings = ((ebx >> 8) & 3) + 1;
282 c->compute_unit_id = ebx & 0xff;
Borislav Petkov9e815092011-02-14 18:14:51 +0100283 cores_per_cu += ((ebx >> 8) & 3);
Andreas Herrmann23588c32010-09-30 14:36:28 +0200284 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200285 u64 value;
286
Andreas Herrmann23588c32010-09-30 14:36:28 +0200287 rdmsrl(MSR_FAM10H_NODE_ID, value);
288 nodes = ((value >> 3) & 7) + 1;
289 node_id = value & 7;
290 } else
Andreas Herrmann9d260eb2009-12-16 15:43:55 +0100291 return;
292
Andreas Herrmann23588c32010-09-30 14:36:28 +0200293 /* fixup multi-node processor information */
294 if (nodes > 1) {
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200295 u32 cores_per_node;
Andreas Herrmannd5185732011-01-24 16:05:40 +0100296 u32 cus_per_node;
Andreas Herrmann6057b4d2010-09-30 14:38:57 +0200297
Andreas Herrmann23588c32010-09-30 14:36:28 +0200298 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
299 cores_per_node = c->x86_max_cores / nodes;
Andreas Herrmannd5185732011-01-24 16:05:40 +0100300 cus_per_node = cores_per_node / cores_per_cu;
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200301
Andreas Herrmann23588c32010-09-30 14:36:28 +0200302 /* store NodeID, use llc_shared_map to store sibling info */
303 per_cpu(cpu_llc_id, cpu) = node_id;
Andreas Herrmann9d260eb2009-12-16 15:43:55 +0100304
Borislav Petkov9e815092011-02-14 18:14:51 +0100305 /* core id has to be in the [0 .. cores_per_node - 1] range */
Andreas Herrmannd5185732011-01-24 16:05:40 +0100306 c->cpu_core_id %= cores_per_node;
307 c->compute_unit_id %= cus_per_node;
Andreas Herrmann23588c32010-09-30 14:36:28 +0200308 }
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200309}
310#endif
311
312/*
Yinghai Lu11fdd252008-09-07 17:58:50 -0700313 * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
314 * Assumes number of cores is a power of two.
315 */
316static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
317{
318#ifdef CONFIG_X86_HT
319 unsigned bits;
Andreas Herrmann99bd0c02009-06-19 10:59:09 +0200320 int cpu = smp_processor_id();
Yinghai Lu11fdd252008-09-07 17:58:50 -0700321
322 bits = c->x86_coreid_bits;
Yinghai Lu11fdd252008-09-07 17:58:50 -0700323 /* Low order bits define the core id (index of core in socket) */
324 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
325 /* Convert the initial APIC ID into the socket ID */
326 c->phys_proc_id = c->initial_apicid >> bits;
Andreas Herrmann99bd0c02009-06-19 10:59:09 +0200327 /* use socket ID also for last level cache */
328 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
Andreas Herrmann23588c32010-09-30 14:36:28 +0200329 amd_get_topology(c);
Yinghai Lu11fdd252008-09-07 17:58:50 -0700330#endif
331}
332
Andreas Herrmann6a812692009-09-16 11:33:40 +0200333int amd_get_nb_id(int cpu)
334{
335 int id = 0;
336#ifdef CONFIG_SMP
337 id = per_cpu(cpu_llc_id, cpu);
338#endif
339 return id;
340}
341EXPORT_SYMBOL_GPL(amd_get_nb_id);
342
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700343static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
344{
Tejun Heo645a7912011-01-23 14:37:40 +0100345#ifdef CONFIG_NUMA
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700346 int cpu = smp_processor_id();
347 int node;
Yinghai Lu0d96b9f2009-08-29 13:17:14 -0700348 unsigned apicid = c->apicid;
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700349
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100350 node = numa_cpu_node(cpu);
351 if (node == NUMA_NO_NODE)
352 node = per_cpu(cpu_llc_id, cpu);
Andreas Herrmann4a376ec2009-09-03 09:40:21 +0200353
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700354 if (!node_online(node)) {
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100355 /*
356 * Two possibilities here:
357 *
358 * - The CPU is missing memory and no node was created. In
359 * that case try picking one from a nearby CPU.
360 *
361 * - The APIC IDs differ from the HyperTransport node IDs
362 * which the K8 northbridge parsing fills in. Assume
363 * they are all increased by a constant offset, but in
364 * the same order as the HT nodeids. If that doesn't
365 * result in a usable node fall back to the path for the
366 * previous case.
367 *
368 * This workaround operates directly on the mapping between
369 * APIC ID and NUMA node, assuming certain relationship
370 * between APIC ID, HT node ID and NUMA topology. As going
371 * through CPU mapping may alter the outcome, directly
372 * access __apicid_to_node[].
373 */
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700374 int ht_nodeid = c->initial_apicid;
375
376 if (ht_nodeid >= 0 &&
Tejun Heobbc9e2f2011-01-23 14:37:39 +0100377 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
378 node = __apicid_to_node[ht_nodeid];
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700379 /* Pick a nearby node */
380 if (!node_online(node))
381 node = nearby_node(apicid);
382 }
383 numa_set_node(cpu, node);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700384#endif
385}
386
Yinghai Lu11fdd252008-09-07 17:58:50 -0700387static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
388{
389#ifdef CONFIG_X86_HT
390 unsigned bits, ecx;
391
392 /* Multi core CPU? */
393 if (c->extended_cpuid_level < 0x80000008)
394 return;
395
396 ecx = cpuid_ecx(0x80000008);
397
398 c->x86_max_cores = (ecx & 0xff) + 1;
399
400 /* CPU telling us the core id bits shift? */
401 bits = (ecx >> 12) & 0xF;
402
403 /* Otherwise recompute */
404 if (bits == 0) {
405 while ((1 << bits) < c->x86_max_cores)
406 bits++;
407 }
408
409 c->x86_coreid_bits = bits;
410#endif
411}
412
Thomas Petazzoni03ae5762008-02-15 12:00:23 +0100413static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
Andi Kleen2b16a232008-01-30 13:32:40 +0100414{
Yinghai Lu11fdd252008-09-07 17:58:50 -0700415 early_init_amd_mc(c);
416
Venki Pallipadi40fb1712008-11-17 16:11:37 -0800417 /*
418 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
419 * with P/T states and does not stop in deep C-states
420 */
421 if (c->x86_power & (1 << 8)) {
Yinghai Lue3224232008-09-06 01:52:28 -0700422 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
Venki Pallipadi40fb1712008-11-17 16:11:37 -0800423 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
424 }
Yinghai Lu5fef55f2008-09-04 21:09:43 +0200425
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700426#ifdef CONFIG_X86_64
427 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
428#else
Yinghai Lu5fef55f2008-09-04 21:09:43 +0200429 /* Set MTRR capability flag if appropriate */
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700430 if (c->x86 == 5)
431 if (c->x86_model == 13 || c->x86_model == 9 ||
432 (c->x86_model == 8 && c->x86_mask >= 8))
433 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
434#endif
Andreas Herrmann42937e82009-06-08 15:55:09 +0200435#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
436 /* check CPU config space for extended APIC ID */
Jeremy Fitzhardinge2cb07862009-07-22 09:59:35 -0700437 if (cpu_has_apic && c->x86 >= 0xf) {
Andreas Herrmann42937e82009-06-08 15:55:09 +0200438 unsigned int val;
439 val = read_pci_config(0, 24, 0, 0x68);
440 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
441 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
442 }
443#endif
Borislav Petkovacf01732010-08-25 18:28:23 +0200444
445 /* We need to do the following only once */
446 if (c != &boot_cpu_data)
447 return;
448
449 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
450
451 if (c->x86 > 0x10 ||
452 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
453 u64 val;
454
455 rdmsrl(MSR_K7_HWCR, val);
456 if (!(val & BIT(24)))
457 printk(KERN_WARNING FW_BUG "TSC doesn't count "
458 "with P0 frequency!\n");
459 }
460 }
Andi Kleen2b16a232008-01-30 13:32:40 +0100461}
462
Magnus Dammb4af3f72006-09-26 10:52:36 +0200463static void __cpuinit init_amd(struct cpuinfo_x86 *c)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464{
Andi Kleen7d318d72005-09-29 22:05:55 +0200465#ifdef CONFIG_SMP
Andi Kleen3c92c2b2005-10-11 01:28:33 +0200466 unsigned long long value;
Andi Kleen7d318d72005-09-29 22:05:55 +0200467
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100468 /*
469 * Disable TLB flush filter by setting HWCR.FFDIS on K8
Andi Kleen7d318d72005-09-29 22:05:55 +0200470 * bit 6 of msr C001_0015
471 *
472 * Errata 63 for SH-B3 steppings
473 * Errata 122 for all steppings (F+ have it disabled by default)
474 */
Yinghai Lu11fdd252008-09-07 17:58:50 -0700475 if (c->x86 == 0xf) {
Andi Kleen7d318d72005-09-29 22:05:55 +0200476 rdmsrl(MSR_K7_HWCR, value);
477 value |= 1 << 6;
478 wrmsrl(MSR_K7_HWCR, value);
479 }
480#endif
481
Andi Kleen2b16a232008-01-30 13:32:40 +0100482 early_init_amd(c);
483
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 /*
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100485 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
Ingo Molnar16282a82008-02-26 08:49:57 +0100486 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100487 */
Ingo Molnar16282a82008-02-26 08:49:57 +0100488 clear_cpu_cap(c, 0*32+31);
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100489
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700490#ifdef CONFIG_X86_64
491 /* On C+ stepping K8 rep microcode works well for copy/memset */
492 if (c->x86 == 0xf) {
493 u32 level;
494
495 level = cpuid_eax(1);
Alan Cox8bdbd962009-07-04 00:35:45 +0100496 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700497 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
Kevin Winchesterfbd8b182009-08-10 19:56:45 -0300498
499 /*
500 * Some BIOSes incorrectly force this feature, but only K8
501 * revision D (model = 0x14) and later actually support it.
Borislav Petkov6b0f43d2009-08-31 09:50:11 +0200502 * (AMD Erratum #110, docId: 25759).
Kevin Winchesterfbd8b182009-08-10 19:56:45 -0300503 */
Borislav Petkov6b0f43d2009-08-31 09:50:11 +0200504 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
505 u64 val;
506
Kevin Winchesterfbd8b182009-08-10 19:56:45 -0300507 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
Borislav Petkov6b0f43d2009-08-31 09:50:11 +0200508 if (!rdmsrl_amd_safe(0xc001100d, &val)) {
509 val &= ~(1ULL << 32);
510 wrmsrl_amd_safe(0xc001100d, val);
511 }
512 }
513
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700514 }
Borislav Petkov12d8a962010-06-02 20:29:21 +0200515 if (c->x86 >= 0x10)
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700516 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
Yinghai Lu0d96b9f2009-08-29 13:17:14 -0700517
518 /* get apicid instead of initial apic id from cpuid */
519 c->apicid = hard_smp_processor_id();
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700520#else
521
522 /*
523 * FIXME: We should handle the K5 here. Set up the write
524 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
525 * no bus pipeline)
526 */
527
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100528 switch (c->x86) {
529 case 4:
Yinghai Lu11fdd252008-09-07 17:58:50 -0700530 init_amd_k5(c);
531 break;
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100532 case 5:
Yinghai Lu11fdd252008-09-07 17:58:50 -0700533 init_amd_k6(c);
534 break;
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100535 case 6: /* An Athlon/Duron */
Yinghai Lu11fdd252008-09-07 17:58:50 -0700536 init_amd_k7(c);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 break;
Andi Kleen67cddd92007-07-21 17:10:03 +0200538 }
Andi Kleen3556ddf2007-04-02 12:14:12 +0200539
Andi Kleenc12ceb72007-05-21 14:31:47 +0200540 /* K6s reports MCEs but don't actually have all the MSRs */
541 if (c->x86 < 6)
Ingo Molnar16282a82008-02-26 08:49:57 +0100542 clear_cpu_cap(c, X86_FEATURE_MCE);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700543#endif
Andi Kleende421862008-01-30 13:32:37 +0100544
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700545 /* Enable workaround for FXSAVE leak */
Yinghai Lu11fdd252008-09-07 17:58:50 -0700546 if (c->x86 >= 6)
547 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
548
549 if (!c->x86_model_id[0]) {
550 switch (c->x86) {
551 case 0xf:
552 /* Should distinguish Models here, but this is only
553 a fallback anyways. */
554 strcpy(c->x86_model_id, "Hammer");
555 break;
556 }
557 }
558
Borislav Petkov27c13ec2009-11-21 14:01:45 +0100559 cpu_detect_cache_sizes(c);
Yinghai Lu11fdd252008-09-07 17:58:50 -0700560
561 /* Multi core CPU? */
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700562 if (c->extended_cpuid_level >= 0x80000008) {
Yinghai Lu11fdd252008-09-07 17:58:50 -0700563 amd_detect_cmp(c);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700564 srat_detect_node(c);
565 }
Yinghai Lu11fdd252008-09-07 17:58:50 -0700566
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700567#ifdef CONFIG_X86_32
Yinghai Lu11fdd252008-09-07 17:58:50 -0700568 detect_ht(c);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700569#endif
Yinghai Lu11fdd252008-09-07 17:58:50 -0700570
571 if (c->extended_cpuid_level >= 0x80000006) {
Andreas Herrmannd9fadd72010-09-02 15:37:10 +0200572 if (cpuid_edx(0x80000006) & 0xf000)
Yinghai Lu11fdd252008-09-07 17:58:50 -0700573 num_cache_leaves = 4;
574 else
575 num_cache_leaves = 3;
576 }
577
Borislav Petkov12d8a962010-06-02 20:29:21 +0200578 if (c->x86 >= 0xf)
Yinghai Lu11fdd252008-09-07 17:58:50 -0700579 set_cpu_cap(c, X86_FEATURE_K8);
580
581 if (cpu_has_xmm2) {
582 /* MFENCE stops RDTSC speculation */
Ingo Molnar16282a82008-02-26 08:49:57 +0100583 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
Yinghai Lu11fdd252008-09-07 17:58:50 -0700584 }
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700585
586#ifdef CONFIG_X86_64
587 if (c->x86 == 0x10) {
588 /* do this for boot cpu */
589 if (c == &boot_cpu_data)
590 check_enable_amd_mmconf_dmi();
591
592 fam10h_check_enable_mmcfg();
593 }
594
Borislav Petkov12d8a962010-06-02 20:29:21 +0200595 if (c == &boot_cpu_data && c->x86 >= 0xf) {
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700596 unsigned long long tseg;
597
598 /*
599 * Split up direct mapping around the TSEG SMM area.
600 * Don't do it for gbpages because there seems very little
601 * benefit in doing so.
602 */
603 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
Alan Cox8bdbd962009-07-04 00:35:45 +0100604 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
605 if ((tseg>>PMD_SHIFT) <
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700606 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
Alan Cox8bdbd962009-07-04 00:35:45 +0100607 ((tseg>>PMD_SHIFT) <
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700608 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
Alan Cox8bdbd962009-07-04 00:35:45 +0100609 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
610 set_memory_4k((unsigned long)__va(tseg), 1);
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700611 }
612 }
613#endif
Boris Ostrovskyb87cf802011-03-15 12:13:44 -0400614
615 /* As a rule processors have APIC timer running in deep C states */
Borislav Petkov14fb57d2011-05-17 14:55:19 +0200616 if (c->x86 > 0xf && !cpu_has_amd_erratum(amd_erratum_400))
Boris Ostrovskyb87cf802011-03-15 12:13:44 -0400617 set_cpu_cap(c, X86_FEATURE_ARAT);
Joerg Roedel5bbc0972011-04-15 14:47:40 +0200618
619 /*
620 * Disable GART TLB Walk Errors on Fam10h. We do this here
621 * because this is always needed when GART is enabled, even in a
622 * kernel which has no MCE support built in.
623 */
624 if (c->x86 == 0x10) {
625 /*
626 * BIOS should disable GartTlbWlk Errors themself. If
627 * it doesn't do it here as suggested by the BKDG.
628 *
629 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
630 */
631 u64 mask;
Roedel, Joergd47cc0d2011-05-19 11:13:39 +0200632 int err;
Joerg Roedel5bbc0972011-04-15 14:47:40 +0200633
Roedel, Joergd47cc0d2011-05-19 11:13:39 +0200634 err = rdmsrl_safe(MSR_AMD64_MCx_MASK(4), &mask);
635 if (err == 0) {
636 mask |= (1 << 10);
637 checking_wrmsrl(MSR_AMD64_MCx_MASK(4), mask);
638 }
Joerg Roedel5bbc0972011-04-15 14:47:40 +0200639 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640}
641
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700642#ifdef CONFIG_X86_32
Alan Cox8bdbd962009-07-04 00:35:45 +0100643static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
644 unsigned int size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645{
646 /* AMD errata T13 (order #21922) */
647 if ((c->x86 == 6)) {
Alan Cox8bdbd962009-07-04 00:35:45 +0100648 /* Duron Rev A0 */
649 if (c->x86_model == 3 && c->x86_mask == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 size = 64;
Alan Cox8bdbd962009-07-04 00:35:45 +0100651 /* Tbird rev A1/A2 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 if (c->x86_model == 4 &&
Alan Cox8bdbd962009-07-04 00:35:45 +0100653 (c->x86_mask == 0 || c->x86_mask == 1))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 size = 256;
655 }
656 return size;
657}
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700658#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659
Jan Beulich02dde8b2009-03-12 12:08:49 +0000660static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 .c_vendor = "AMD",
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100662 .c_ident = { "AuthenticAMD" },
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700663#ifdef CONFIG_X86_32
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 .c_models = {
665 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
666 {
667 [3] = "486 DX/2",
668 [7] = "486 DX/2-WB",
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100669 [8] = "486 DX/4",
670 [9] = "486 DX/4-WB",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 [14] = "Am5x86-WT",
Paolo Ciarrocchifb87a292008-02-22 23:10:33 +0100672 [15] = "Am5x86-WB"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 }
674 },
675 },
Yinghai Lu6c62aa42008-09-07 17:58:54 -0700676 .c_size_cache = amd_size_cache,
677#endif
Thomas Petazzoni03ae5762008-02-15 12:00:23 +0100678 .c_early_init = early_init_amd,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 .c_init = init_amd,
Yinghai Lu10a434f2008-09-04 21:09:45 +0200680 .c_x86_vendor = X86_VENDOR_AMD,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681};
682
Yinghai Lu10a434f2008-09-04 21:09:45 +0200683cpu_dev_register(amd_cpu_dev);
Hans Rosenfeldd78d6712010-07-28 19:09:30 +0200684
685/*
686 * AMD errata checking
687 *
688 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
689 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
690 * have an OSVW id assigned, which it takes as first argument. Both take a
691 * variable number of family-specific model-stepping ranges created by
692 * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
693 * int[] in arch/x86/include/asm/processor.h.
694 *
695 * Example:
696 *
697 * const int amd_erratum_319[] =
698 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
699 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
700 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
701 */
702
Hans Rosenfeld9d8888c2010-07-28 19:09:31 +0200703const int amd_erratum_400[] =
Borislav Petkov328935e2011-05-17 14:55:18 +0200704 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
Hans Rosenfeld9d8888c2010-07-28 19:09:31 +0200705 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
H. Peter Anvina5b91602010-07-28 16:23:20 -0700706EXPORT_SYMBOL_GPL(amd_erratum_400);
Hans Rosenfeld9d8888c2010-07-28 19:09:31 +0200707
Hans Rosenfeld1be85a62010-07-28 19:09:32 +0200708const int amd_erratum_383[] =
709 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
H. Peter Anvina5b91602010-07-28 16:23:20 -0700710EXPORT_SYMBOL_GPL(amd_erratum_383);
Hans Rosenfeld9d8888c2010-07-28 19:09:31 +0200711
Hans Rosenfeldd78d6712010-07-28 19:09:30 +0200712bool cpu_has_amd_erratum(const int *erratum)
713{
Tejun Heo7b543a52010-12-18 16:30:05 +0100714 struct cpuinfo_x86 *cpu = __this_cpu_ptr(&cpu_info);
Hans Rosenfeldd78d6712010-07-28 19:09:30 +0200715 int osvw_id = *erratum++;
716 u32 range;
717 u32 ms;
718
719 /*
720 * If called early enough that current_cpu_data hasn't been initialized
721 * yet, fall back to boot_cpu_data.
722 */
723 if (cpu->x86 == 0)
724 cpu = &boot_cpu_data;
725
726 if (cpu->x86_vendor != X86_VENDOR_AMD)
727 return false;
728
729 if (osvw_id >= 0 && osvw_id < 65536 &&
730 cpu_has(cpu, X86_FEATURE_OSVW)) {
731 u64 osvw_len;
732
733 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
734 if (osvw_id < osvw_len) {
735 u64 osvw_bits;
736
737 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
738 osvw_bits);
739 return osvw_bits & (1ULL << (osvw_id & 0x3f));
740 }
741 }
742
743 /* OSVW unavailable or ID unknown, match family-model-stepping range */
Hans Rosenfeld07a77952010-08-18 16:19:50 +0200744 ms = (cpu->x86_model << 4) | cpu->x86_mask;
Hans Rosenfeldd78d6712010-07-28 19:09:30 +0200745 while ((range = *erratum++))
746 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
747 (ms >= AMD_MODEL_RANGE_START(range)) &&
748 (ms <= AMD_MODEL_RANGE_END(range)))
749 return true;
750
751 return false;
752}
H. Peter Anvina5b91602010-07-28 16:23:20 -0700753
754EXPORT_SYMBOL_GPL(cpu_has_amd_erratum);