x86: add PCI extended config space access for AMD Barcelona
[cascardo/linux.git] / arch / x86 / kernel / cpu / amd.c
1 #include <linux/init.h>
2 #include <linux/bitops.h>
3 #include <linux/mm.h>
4 #include <asm/io.h>
5 #include <asm/processor.h>
6 #include <asm/apic.h>
7
8 #include <mach_apic.h>
9 #include "../setup.h"
10 #include "cpu.h"
11
12 /*
13  *      B step AMD K6 before B 9730xxxx have hardware bugs that can cause
14  *      misexecution of code under Linux. Owners of such processors should
15  *      contact AMD for precise details and a CPU swap.
16  *
17  *      See     http://www.multimania.com/poulot/k6bug.html
18  *              http://www.amd.com/K6/k6docs/revgd.html
19  *
20  *      The following test is erm.. interesting. AMD neglected to up
21  *      the chip setting when fixing the bug but they also tweaked some
22  *      performance at the same time..
23  */
24
25 extern void vide(void);
26 __asm__(".align 4\nvide: ret");
27
28 #ifdef CONFIG_X86_LOCAL_APIC
29 #define ENABLE_C1E_MASK         0x18000000
30 #define CPUID_PROCESSOR_SIGNATURE       1
31 #define CPUID_XFAM              0x0ff00000
32 #define CPUID_XFAM_K8           0x00000000
33 #define CPUID_XFAM_10H          0x00100000
34 #define CPUID_XFAM_11H          0x00200000
35 #define CPUID_XMOD              0x000f0000
36 #define CPUID_XMOD_REV_F        0x00040000
37
38 /* AMD systems with C1E don't have a working lAPIC timer. Check for that. */
39 static __cpuinit int amd_apic_timer_broken(void)
40 {
41         u32 lo, hi;
42         u32 eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
43         switch (eax & CPUID_XFAM) {
44         case CPUID_XFAM_K8:
45                 if ((eax & CPUID_XMOD) < CPUID_XMOD_REV_F)
46                         break;
47         case CPUID_XFAM_10H:
48         case CPUID_XFAM_11H:
49                 rdmsr(MSR_K8_ENABLE_C1E, lo, hi);
50                 if (lo & ENABLE_C1E_MASK) {
51                         if (smp_processor_id() != boot_cpu_physical_apicid)
52                                 printk(KERN_INFO "AMD C1E detected late. "
53                                        "        Force timer broadcast.\n");
54                         return 1;
55                 }
56                 break;
57         default:
58                 /* err on the side of caution */
59                 return 1;
60         }
61         return 0;
62 }
63 #endif
64
65 int force_mwait __cpuinitdata;
66
67 static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
68 {
69         if (cpuid_eax(0x80000000) >= 0x80000007) {
70                 c->x86_power = cpuid_edx(0x80000007);
71                 if (c->x86_power & (1<<8))
72                         set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
73         }
74 }
75
76 static void __cpuinit init_amd(struct cpuinfo_x86 *c)
77 {
78         u32 l, h;
79         int mbytes = num_physpages >> (20-PAGE_SHIFT);
80         int r;
81
82 #ifdef CONFIG_SMP
83         unsigned long long value;
84
85         /*
86          * Disable TLB flush filter by setting HWCR.FFDIS on K8
87          * bit 6 of msr C001_0015
88          *
89          * Errata 63 for SH-B3 steppings
90          * Errata 122 for all steppings (F+ have it disabled by default)
91          */
92         if (c->x86 == 15) {
93                 rdmsrl(MSR_K7_HWCR, value);
94                 value |= 1 << 6;
95                 wrmsrl(MSR_K7_HWCR, value);
96         }
97 #endif
98
99         early_init_amd(c);
100
101         /*
102          *      FIXME: We should handle the K5 here. Set up the write
103          *      range and also turn on MSR 83 bits 4 and 31 (write alloc,
104          *      no bus pipeline)
105          */
106
107         /*
108          * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
109          * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
110          */
111         clear_cpu_cap(c, 0*32+31);
112
113         r = get_model_name(c);
114
115         switch (c->x86) {
116         case 4:
117                 /*
118                  * General Systems BIOSen alias the cpu frequency registers
119                  * of the Elan at 0x000df000. Unfortuantly, one of the Linux
120                  * drivers subsequently pokes it, and changes the CPU speed.
121                  * Workaround : Remove the unneeded alias.
122                  */
123 #define CBAR            (0xfffc) /* Configuration Base Address  (32-bit) */
124 #define CBAR_ENB        (0x80000000)
125 #define CBAR_KEY        (0X000000CB)
126                         if (c->x86_model == 9 || c->x86_model == 10) {
127                                 if (inl (CBAR) & CBAR_ENB)
128                                         outl (0 | CBAR_KEY, CBAR);
129                         }
130                         break;
131         case 5:
132                         if (c->x86_model < 6) {
133                                 /* Based on AMD doc 20734R - June 2000 */
134                                 if (c->x86_model == 0) {
135                                         clear_cpu_cap(c, X86_FEATURE_APIC);
136                                         set_cpu_cap(c, X86_FEATURE_PGE);
137                                 }
138                                 break;
139                         }
140
141                         if (c->x86_model == 6 && c->x86_mask == 1) {
142                                 const int K6_BUG_LOOP = 1000000;
143                                 int n;
144                                 void (*f_vide)(void);
145                                 unsigned long d, d2;
146
147                                 printk(KERN_INFO "AMD K6 stepping B detected - ");
148
149                                 /*
150                                  * It looks like AMD fixed the 2.6.2 bug and improved indirect
151                                  * calls at the same time.
152                                  */
153
154                                 n = K6_BUG_LOOP;
155                                 f_vide = vide;
156                                 rdtscl(d);
157                                 while (n--)
158                                         f_vide();
159                                 rdtscl(d2);
160                                 d = d2-d;
161
162                                 if (d > 20*K6_BUG_LOOP)
163                                         printk("system stability may be impaired when more than 32 MB are used.\n");
164                                 else
165                                         printk("probably OK (after B9730xxxx).\n");
166                                 printk(KERN_INFO "Please see http://membres.lycos.fr/poulot/k6bug.html\n");
167                         }
168
169                         /* K6 with old style WHCR */
170                         if (c->x86_model < 8 ||
171                            (c->x86_model == 8 && c->x86_mask < 8)) {
172                                 /* We can only write allocate on the low 508Mb */
173                                 if (mbytes > 508)
174                                         mbytes = 508;
175
176                                 rdmsr(MSR_K6_WHCR, l, h);
177                                 if ((l&0x0000FFFF) == 0) {
178                                         unsigned long flags;
179                                         l = (1<<0)|((mbytes/4)<<1);
180                                         local_irq_save(flags);
181                                         wbinvd();
182                                         wrmsr(MSR_K6_WHCR, l, h);
183                                         local_irq_restore(flags);
184                                         printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
185                                                 mbytes);
186                                 }
187                                 break;
188                         }
189
190                         if ((c->x86_model == 8 && c->x86_mask > 7) ||
191                              c->x86_model == 9 || c->x86_model == 13) {
192                                 /* The more serious chips .. */
193
194                                 if (mbytes > 4092)
195                                         mbytes = 4092;
196
197                                 rdmsr(MSR_K6_WHCR, l, h);
198                                 if ((l&0xFFFF0000) == 0) {
199                                         unsigned long flags;
200                                         l = ((mbytes>>2)<<22)|(1<<16);
201                                         local_irq_save(flags);
202                                         wbinvd();
203                                         wrmsr(MSR_K6_WHCR, l, h);
204                                         local_irq_restore(flags);
205                                         printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
206                                                 mbytes);
207                                 }
208
209                                 /*  Set MTRR capability flag if appropriate */
210                                 if (c->x86_model == 13 || c->x86_model == 9 ||
211                                    (c->x86_model == 8 && c->x86_mask >= 8))
212                                         set_cpu_cap(c, X86_FEATURE_K6_MTRR);
213                                 break;
214                         }
215
216                         if (c->x86_model == 10) {
217                                 /* AMD Geode LX is model 10 */
218                                 /* placeholder for any needed mods */
219                                 break;
220                         }
221                         break;
222         case 6: /* An Athlon/Duron */
223
224                         /*
225                          * Bit 15 of Athlon specific MSR 15, needs to be 0
226                          * to enable SSE on Palomino/Morgan/Barton CPU's.
227                          * If the BIOS didn't enable it already, enable it here.
228                          */
229                         if (c->x86_model >= 6 && c->x86_model <= 10) {
230                                 if (!cpu_has(c, X86_FEATURE_XMM)) {
231                                         printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
232                                         rdmsr(MSR_K7_HWCR, l, h);
233                                         l &= ~0x00008000;
234                                         wrmsr(MSR_K7_HWCR, l, h);
235                                         set_cpu_cap(c, X86_FEATURE_XMM);
236                                 }
237                         }
238
239                         /*
240                          * It's been determined by AMD that Athlons since model 8 stepping 1
241                          * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
242                          * As per AMD technical note 27212 0.2
243                          */
244                         if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
245                                 rdmsr(MSR_K7_CLK_CTL, l, h);
246                                 if ((l & 0xfff00000) != 0x20000000) {
247                                         printk ("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n", l,
248                                                 ((l & 0x000fffff)|0x20000000));
249                                         wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
250                                 }
251                         }
252                         break;
253         }
254
255         switch (c->x86) {
256         case 15:
257         /* Use K8 tuning for Fam10h and Fam11h */
258         case 0x10:
259         case 0x11:
260                 set_cpu_cap(c, X86_FEATURE_K8);
261                 break;
262         case 6:
263                 set_cpu_cap(c, X86_FEATURE_K7);
264                 break;
265         }
266         if (c->x86 >= 6)
267                 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
268
269         display_cacheinfo(c);
270
271         if (cpuid_eax(0x80000000) >= 0x80000008)
272                 c->x86_max_cores = (cpuid_ecx(0x80000008) & 0xff) + 1;
273
274 #ifdef CONFIG_X86_HT
275         /*
276          * On a AMD multi core setup the lower bits of the APIC id
277          * distinguish the cores.
278          */
279         if (c->x86_max_cores > 1) {
280                 int cpu = smp_processor_id();
281                 unsigned bits = (cpuid_ecx(0x80000008) >> 12) & 0xf;
282
283                 if (bits == 0) {
284                         while ((1 << bits) < c->x86_max_cores)
285                                 bits++;
286                 }
287                 c->cpu_core_id = c->phys_proc_id & ((1<<bits)-1);
288                 c->phys_proc_id >>= bits;
289                 printk(KERN_INFO "CPU %d(%d) -> Core %d\n",
290                        cpu, c->x86_max_cores, c->cpu_core_id);
291         }
292 #endif
293
294         if (cpuid_eax(0x80000000) >= 0x80000006) {
295                 if ((c->x86 == 0x10) && (cpuid_edx(0x80000006) & 0xf000))
296                         num_cache_leaves = 4;
297                 else
298                         num_cache_leaves = 3;
299         }
300
301 #ifdef CONFIG_X86_LOCAL_APIC
302         if (amd_apic_timer_broken())
303                 local_apic_timer_disabled = 1;
304 #endif
305
306         /* K6s reports MCEs but don't actually have all the MSRs */
307         if (c->x86 < 6)
308                 clear_cpu_cap(c, X86_FEATURE_MCE);
309
310         if (cpu_has_xmm2)
311                 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
312
313         if (c->x86 == 0x10)
314                 amd_enable_pci_ext_cfg(c);
315 }
316
317 static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
318 {
319         /* AMD errata T13 (order #21922) */
320         if ((c->x86 == 6)) {
321                 if (c->x86_model == 3 && c->x86_mask == 0)      /* Duron Rev A0 */
322                         size = 64;
323                 if (c->x86_model == 4 &&
324                     (c->x86_mask == 0 || c->x86_mask == 1))     /* Tbird rev A1/A2 */
325                         size = 256;
326         }
327         return size;
328 }
329
330 static struct cpu_dev amd_cpu_dev __cpuinitdata = {
331         .c_vendor       = "AMD",
332         .c_ident        = { "AuthenticAMD" },
333         .c_models = {
334                 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
335                   {
336                           [3] = "486 DX/2",
337                           [7] = "486 DX/2-WB",
338                           [8] = "486 DX/4",
339                           [9] = "486 DX/4-WB",
340                           [14] = "Am5x86-WT",
341                           [15] = "Am5x86-WB"
342                   }
343                 },
344         },
345         .c_early_init   = early_init_amd,
346         .c_init         = init_amd,
347         .c_size_cache   = amd_size_cache,
348 };
349
350 cpu_vendor_dev_register(X86_VENDOR_AMD, &amd_cpu_dev);