diff options
Diffstat (limited to 'arch/i386/kernel/cpu/amd.c')
-rw-r--r-- | arch/i386/kernel/cpu/amd.c | 249 |
1 files changed, 249 insertions, 0 deletions
diff --git a/arch/i386/kernel/cpu/amd.c b/arch/i386/kernel/cpu/amd.c new file mode 100644 index 000000000000..ae94585d0445 --- /dev/null +++ b/arch/i386/kernel/cpu/amd.c | |||
@@ -0,0 +1,249 @@ | |||
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 | |||
7 | #include "cpu.h" | ||
8 | |||
9 | /* | ||
10 | * B step AMD K6 before B 9730xxxx have hardware bugs that can cause | ||
11 | * misexecution of code under Linux. Owners of such processors should | ||
12 | * contact AMD for precise details and a CPU swap. | ||
13 | * | ||
14 | * See http://www.multimania.com/poulot/k6bug.html | ||
15 | * http://www.amd.com/K6/k6docs/revgd.html | ||
16 | * | ||
17 | * The following test is erm.. interesting. AMD neglected to up | ||
18 | * the chip setting when fixing the bug but they also tweaked some | ||
19 | * performance at the same time.. | ||
20 | */ | ||
21 | |||
22 | extern void vide(void); | ||
23 | __asm__(".align 4\nvide: ret"); | ||
24 | |||
25 | static void __init init_amd(struct cpuinfo_x86 *c) | ||
26 | { | ||
27 | u32 l, h; | ||
28 | int mbytes = num_physpages >> (20-PAGE_SHIFT); | ||
29 | int r; | ||
30 | |||
31 | /* | ||
32 | * FIXME: We should handle the K5 here. Set up the write | ||
33 | * range and also turn on MSR 83 bits 4 and 31 (write alloc, | ||
34 | * no bus pipeline) | ||
35 | */ | ||
36 | |||
37 | /* Bit 31 in normal CPUID used for nonstandard 3DNow ID; | ||
38 | 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway */ | ||
39 | clear_bit(0*32+31, c->x86_capability); | ||
40 | |||
41 | r = get_model_name(c); | ||
42 | |||
43 | switch(c->x86) | ||
44 | { | ||
45 | case 4: | ||
46 | /* | ||
47 | * General Systems BIOSen alias the cpu frequency registers | ||
48 | * of the Elan at 0x000df000. Unfortuantly, one of the Linux | ||
49 | * drivers subsequently pokes it, and changes the CPU speed. | ||
50 | * Workaround : Remove the unneeded alias. | ||
51 | */ | ||
52 | #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */ | ||
53 | #define CBAR_ENB (0x80000000) | ||
54 | #define CBAR_KEY (0X000000CB) | ||
55 | if (c->x86_model==9 || c->x86_model == 10) { | ||
56 | if (inl (CBAR) & CBAR_ENB) | ||
57 | outl (0 | CBAR_KEY, CBAR); | ||
58 | } | ||
59 | break; | ||
60 | case 5: | ||
61 | if( c->x86_model < 6 ) | ||
62 | { | ||
63 | /* Based on AMD doc 20734R - June 2000 */ | ||
64 | if ( c->x86_model == 0 ) { | ||
65 | clear_bit(X86_FEATURE_APIC, c->x86_capability); | ||
66 | set_bit(X86_FEATURE_PGE, c->x86_capability); | ||
67 | } | ||
68 | break; | ||
69 | } | ||
70 | |||
71 | if ( c->x86_model == 6 && c->x86_mask == 1 ) { | ||
72 | const int K6_BUG_LOOP = 1000000; | ||
73 | int n; | ||
74 | void (*f_vide)(void); | ||
75 | unsigned long d, d2; | ||
76 | |||
77 | printk(KERN_INFO "AMD K6 stepping B detected - "); | ||
78 | |||
79 | /* | ||
80 | * It looks like AMD fixed the 2.6.2 bug and improved indirect | ||
81 | * calls at the same time. | ||
82 | */ | ||
83 | |||
84 | n = K6_BUG_LOOP; | ||
85 | f_vide = vide; | ||
86 | rdtscl(d); | ||
87 | while (n--) | ||
88 | f_vide(); | ||
89 | rdtscl(d2); | ||
90 | d = d2-d; | ||
91 | |||
92 | /* Knock these two lines out if it debugs out ok */ | ||
93 | printk(KERN_INFO "AMD K6 stepping B detected - "); | ||
94 | /* -- cut here -- */ | ||
95 | if (d > 20*K6_BUG_LOOP) | ||
96 | printk("system stability may be impaired when more than 32 MB are used.\n"); | ||
97 | else | ||
98 | printk("probably OK (after B9730xxxx).\n"); | ||
99 | printk(KERN_INFO "Please see http://membres.lycos.fr/poulot/k6bug.html\n"); | ||
100 | } | ||
101 | |||
102 | /* K6 with old style WHCR */ | ||
103 | if (c->x86_model < 8 || | ||
104 | (c->x86_model== 8 && c->x86_mask < 8)) { | ||
105 | /* We can only write allocate on the low 508Mb */ | ||
106 | if(mbytes>508) | ||
107 | mbytes=508; | ||
108 | |||
109 | rdmsr(MSR_K6_WHCR, l, h); | ||
110 | if ((l&0x0000FFFF)==0) { | ||
111 | unsigned long flags; | ||
112 | l=(1<<0)|((mbytes/4)<<1); | ||
113 | local_irq_save(flags); | ||
114 | wbinvd(); | ||
115 | wrmsr(MSR_K6_WHCR, l, h); | ||
116 | local_irq_restore(flags); | ||
117 | printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n", | ||
118 | mbytes); | ||
119 | } | ||
120 | break; | ||
121 | } | ||
122 | |||
123 | if ((c->x86_model == 8 && c->x86_mask >7) || | ||
124 | c->x86_model == 9 || c->x86_model == 13) { | ||
125 | /* The more serious chips .. */ | ||
126 | |||
127 | if(mbytes>4092) | ||
128 | mbytes=4092; | ||
129 | |||
130 | rdmsr(MSR_K6_WHCR, l, h); | ||
131 | if ((l&0xFFFF0000)==0) { | ||
132 | unsigned long flags; | ||
133 | l=((mbytes>>2)<<22)|(1<<16); | ||
134 | local_irq_save(flags); | ||
135 | wbinvd(); | ||
136 | wrmsr(MSR_K6_WHCR, l, h); | ||
137 | local_irq_restore(flags); | ||
138 | printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n", | ||
139 | mbytes); | ||
140 | } | ||
141 | |||
142 | /* Set MTRR capability flag if appropriate */ | ||
143 | if (c->x86_model == 13 || c->x86_model == 9 || | ||
144 | (c->x86_model == 8 && c->x86_mask >= 8)) | ||
145 | set_bit(X86_FEATURE_K6_MTRR, c->x86_capability); | ||
146 | break; | ||
147 | } | ||
148 | break; | ||
149 | |||
150 | case 6: /* An Athlon/Duron */ | ||
151 | |||
152 | /* Bit 15 of Athlon specific MSR 15, needs to be 0 | ||
153 | * to enable SSE on Palomino/Morgan/Barton CPU's. | ||
154 | * If the BIOS didn't enable it already, enable it here. | ||
155 | */ | ||
156 | if (c->x86_model >= 6 && c->x86_model <= 10) { | ||
157 | if (!cpu_has(c, X86_FEATURE_XMM)) { | ||
158 | printk(KERN_INFO "Enabling disabled K7/SSE Support.\n"); | ||
159 | rdmsr(MSR_K7_HWCR, l, h); | ||
160 | l &= ~0x00008000; | ||
161 | wrmsr(MSR_K7_HWCR, l, h); | ||
162 | set_bit(X86_FEATURE_XMM, c->x86_capability); | ||
163 | } | ||
164 | } | ||
165 | |||
166 | /* It's been determined by AMD that Athlons since model 8 stepping 1 | ||
167 | * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx | ||
168 | * As per AMD technical note 27212 0.2 | ||
169 | */ | ||
170 | if ((c->x86_model == 8 && c->x86_mask>=1) || (c->x86_model > 8)) { | ||
171 | rdmsr(MSR_K7_CLK_CTL, l, h); | ||
172 | if ((l & 0xfff00000) != 0x20000000) { | ||
173 | printk ("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n", l, | ||
174 | ((l & 0x000fffff)|0x20000000)); | ||
175 | wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h); | ||
176 | } | ||
177 | } | ||
178 | break; | ||
179 | } | ||
180 | |||
181 | switch (c->x86) { | ||
182 | case 15: | ||
183 | set_bit(X86_FEATURE_K8, c->x86_capability); | ||
184 | break; | ||
185 | case 6: | ||
186 | set_bit(X86_FEATURE_K7, c->x86_capability); | ||
187 | break; | ||
188 | } | ||
189 | |||
190 | display_cacheinfo(c); | ||
191 | detect_ht(c); | ||
192 | |||
193 | #ifdef CONFIG_X86_HT | ||
194 | /* AMD dual core looks like HT but isn't really. Hide it from the | ||
195 | scheduler. This works around problems with the domain scheduler. | ||
196 | Also probably gives slightly better scheduling and disables | ||
197 | SMT nice which is harmful on dual core. | ||
198 | TBD tune the domain scheduler for dual core. */ | ||
199 | if (cpu_has(c, X86_FEATURE_CMP_LEGACY)) | ||
200 | smp_num_siblings = 1; | ||
201 | #endif | ||
202 | |||
203 | if (cpuid_eax(0x80000000) >= 0x80000008) { | ||
204 | c->x86_num_cores = (cpuid_ecx(0x80000008) & 0xff) + 1; | ||
205 | if (c->x86_num_cores & (c->x86_num_cores - 1)) | ||
206 | c->x86_num_cores = 1; | ||
207 | } | ||
208 | } | ||
209 | |||
210 | static unsigned int amd_size_cache(struct cpuinfo_x86 * c, unsigned int size) | ||
211 | { | ||
212 | /* AMD errata T13 (order #21922) */ | ||
213 | if ((c->x86 == 6)) { | ||
214 | if (c->x86_model == 3 && c->x86_mask == 0) /* Duron Rev A0 */ | ||
215 | size = 64; | ||
216 | if (c->x86_model == 4 && | ||
217 | (c->x86_mask==0 || c->x86_mask==1)) /* Tbird rev A1/A2 */ | ||
218 | size = 256; | ||
219 | } | ||
220 | return size; | ||
221 | } | ||
222 | |||
223 | static struct cpu_dev amd_cpu_dev __initdata = { | ||
224 | .c_vendor = "AMD", | ||
225 | .c_ident = { "AuthenticAMD" }, | ||
226 | .c_models = { | ||
227 | { .vendor = X86_VENDOR_AMD, .family = 4, .model_names = | ||
228 | { | ||
229 | [3] = "486 DX/2", | ||
230 | [7] = "486 DX/2-WB", | ||
231 | [8] = "486 DX/4", | ||
232 | [9] = "486 DX/4-WB", | ||
233 | [14] = "Am5x86-WT", | ||
234 | [15] = "Am5x86-WB" | ||
235 | } | ||
236 | }, | ||
237 | }, | ||
238 | .c_init = init_amd, | ||
239 | .c_identify = generic_identify, | ||
240 | .c_size_cache = amd_size_cache, | ||
241 | }; | ||
242 | |||
243 | int __init amd_init_cpu(void) | ||
244 | { | ||
245 | cpu_devs[X86_VENDOR_AMD] = &amd_cpu_dev; | ||
246 | return 0; | ||
247 | } | ||
248 | |||
249 | //early_arch_initcall(amd_init_cpu); | ||