Pull ec into release branch
[linux-2.6] / arch / i386 / kernel / cpu / centaur.c
1 #include <linux/kernel.h>
2 #include <linux/init.h>
3 #include <linux/bitops.h>
4 #include <asm/processor.h>
5 #include <asm/msr.h>
6 #include <asm/e820.h>
7 #include <asm/mtrr.h>
8 #include "cpu.h"
9
10 #ifdef CONFIG_X86_OOSTORE
11
12 static u32 __cpuinit power2(u32 x)
13 {
14         u32 s=1;
15         while(s<=x)
16                 s<<=1;
17         return s>>=1;
18 }
19
20
21 /*
22  *      Set up an actual MCR
23  */
24  
25 static void __cpuinit centaur_mcr_insert(int reg, u32 base, u32 size, int key)
26 {
27         u32 lo, hi;
28         
29         hi = base & ~0xFFF;
30         lo = ~(size-1);         /* Size is a power of 2 so this makes a mask */
31         lo &= ~0xFFF;           /* Remove the ctrl value bits */
32         lo |= key;              /* Attribute we wish to set */
33         wrmsr(reg+MSR_IDT_MCR0, lo, hi);
34         mtrr_centaur_report_mcr(reg, lo, hi);   /* Tell the mtrr driver */
35 }
36
37 /*
38  *      Figure what we can cover with MCR's
39  *
40  *      Shortcut: We know you can't put 4Gig of RAM on a winchip
41  */
42
43 static u32 __cpuinit ramtop(void)               /* 16388 */
44 {
45         int i;
46         u32 top = 0;
47         u32 clip = 0xFFFFFFFFUL;
48         
49         for (i = 0; i < e820.nr_map; i++) {
50                 unsigned long start, end;
51
52                 if (e820.map[i].addr > 0xFFFFFFFFUL)
53                         continue;
54                 /*
55                  *      Don't MCR over reserved space. Ignore the ISA hole
56                  *      we frob around that catastrophy already
57                  */
58                                         
59                 if (e820.map[i].type == E820_RESERVED)
60                 {
61                         if(e820.map[i].addr >= 0x100000UL && e820.map[i].addr < clip)
62                                 clip = e820.map[i].addr;
63                         continue;
64                 }
65                 start = e820.map[i].addr;
66                 end = e820.map[i].addr + e820.map[i].size;
67                 if (start >= end)
68                         continue;
69                 if (end > top)
70                         top = end;
71         }
72         /* Everything below 'top' should be RAM except for the ISA hole.
73            Because of the limited MCR's we want to map NV/ACPI into our
74            MCR range for gunk in RAM 
75            
76            Clip might cause us to MCR insufficient RAM but that is an
77            acceptable failure mode and should only bite obscure boxes with
78            a VESA hole at 15Mb
79            
80            The second case Clip sometimes kicks in is when the EBDA is marked
81            as reserved. Again we fail safe with reasonable results
82         */
83         
84         if(top>clip)
85                 top=clip;
86                 
87         return top;
88 }
89
90 /*
91  *      Compute a set of MCR's to give maximum coverage
92  */
93
94 static int __cpuinit centaur_mcr_compute(int nr, int key)
95 {
96         u32 mem = ramtop();
97         u32 root = power2(mem);
98         u32 base = root;
99         u32 top = root;
100         u32 floor = 0;
101         int ct = 0;
102         
103         while(ct<nr)
104         {
105                 u32 fspace = 0;
106
107                 /*
108                  *      Find the largest block we will fill going upwards
109                  */
110
111                 u32 high = power2(mem-top);     
112
113                 /*
114                  *      Find the largest block we will fill going downwards
115                  */
116
117                 u32 low = base/2;
118
119                 /*
120                  *      Don't fill below 1Mb going downwards as there
121                  *      is an ISA hole in the way.
122                  */             
123                  
124                 if(base <= 1024*1024)
125                         low = 0;
126                         
127                 /*
128                  *      See how much space we could cover by filling below
129                  *      the ISA hole
130                  */
131                  
132                 if(floor == 0)
133                         fspace = 512*1024;
134                 else if(floor ==512*1024)
135                         fspace = 128*1024;
136
137                 /* And forget ROM space */
138                 
139                 /*
140                  *      Now install the largest coverage we get
141                  */
142                  
143                 if(fspace > high && fspace > low)
144                 {
145                         centaur_mcr_insert(ct, floor, fspace, key);
146                         floor += fspace;
147                 }
148                 else if(high > low)
149                 {
150                         centaur_mcr_insert(ct, top, high, key);
151                         top += high;
152                 }
153                 else if(low > 0)
154                 {
155                         base -= low;
156                         centaur_mcr_insert(ct, base, low, key);
157                 }
158                 else break;
159                 ct++;
160         }
161         /*
162          *      We loaded ct values. We now need to set the mask. The caller
163          *      must do this bit.
164          */
165          
166         return ct;
167 }
168
169 static void __cpuinit centaur_create_optimal_mcr(void)
170 {
171         int i;
172         /*
173          *      Allocate up to 6 mcrs to mark as much of ram as possible
174          *      as write combining and weak write ordered.
175          *
176          *      To experiment with: Linux never uses stack operations for 
177          *      mmio spaces so we could globally enable stack operation wc
178          *
179          *      Load the registers with type 31 - full write combining, all
180          *      writes weakly ordered.
181          */
182         int used = centaur_mcr_compute(6, 31);
183
184         /*
185          *      Wipe unused MCRs
186          */
187          
188         for(i=used;i<8;i++)
189                 wrmsr(MSR_IDT_MCR0+i, 0, 0);
190 }
191
192 static void __cpuinit winchip2_create_optimal_mcr(void)
193 {
194         u32 lo, hi;
195         int i;
196
197         /*
198          *      Allocate up to 6 mcrs to mark as much of ram as possible
199          *      as write combining, weak store ordered.
200          *
201          *      Load the registers with type 25
202          *              8       -       weak write ordering
203          *              16      -       weak read ordering
204          *              1       -       write combining
205          */
206
207         int used = centaur_mcr_compute(6, 25);
208         
209         /*
210          *      Mark the registers we are using.
211          */
212          
213         rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
214         for(i=0;i<used;i++)
215                 lo|=1<<(9+i);
216         wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
217         
218         /*
219          *      Wipe unused MCRs
220          */
221          
222         for(i=used;i<8;i++)
223                 wrmsr(MSR_IDT_MCR0+i, 0, 0);
224 }
225
226 /*
227  *      Handle the MCR key on the Winchip 2.
228  */
229
230 static void __cpuinit winchip2_unprotect_mcr(void)
231 {
232         u32 lo, hi;
233         u32 key;
234         
235         rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
236         lo&=~0x1C0;     /* blank bits 8-6 */
237         key = (lo>>17) & 7;
238         lo |= key<<6;   /* replace with unlock key */
239         wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
240 }
241
242 static void __cpuinit winchip2_protect_mcr(void)
243 {
244         u32 lo, hi;
245         
246         rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
247         lo&=~0x1C0;     /* blank bits 8-6 */
248         wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
249 }
250 #endif /* CONFIG_X86_OOSTORE */
251
252 #define ACE_PRESENT     (1 << 6)
253 #define ACE_ENABLED     (1 << 7)
254 #define ACE_FCR         (1 << 28)       /* MSR_VIA_FCR */
255
256 #define RNG_PRESENT     (1 << 2)
257 #define RNG_ENABLED     (1 << 3)
258 #define RNG_ENABLE      (1 << 6)        /* MSR_VIA_RNG */
259
260 static void __cpuinit init_c3(struct cpuinfo_x86 *c)
261 {
262         u32  lo, hi;
263
264         /* Test for Centaur Extended Feature Flags presence */
265         if (cpuid_eax(0xC0000000) >= 0xC0000001) {
266                 u32 tmp = cpuid_edx(0xC0000001);
267
268                 /* enable ACE unit, if present and disabled */
269                 if ((tmp & (ACE_PRESENT | ACE_ENABLED)) == ACE_PRESENT) {
270                         rdmsr (MSR_VIA_FCR, lo, hi);
271                         lo |= ACE_FCR;          /* enable ACE unit */
272                         wrmsr (MSR_VIA_FCR, lo, hi);
273                         printk(KERN_INFO "CPU: Enabled ACE h/w crypto\n");
274                 }
275
276                 /* enable RNG unit, if present and disabled */
277                 if ((tmp & (RNG_PRESENT | RNG_ENABLED)) == RNG_PRESENT) {
278                         rdmsr (MSR_VIA_RNG, lo, hi);
279                         lo |= RNG_ENABLE;       /* enable RNG unit */
280                         wrmsr (MSR_VIA_RNG, lo, hi);
281                         printk(KERN_INFO "CPU: Enabled h/w RNG\n");
282                 }
283
284                 /* store Centaur Extended Feature Flags as
285                  * word 5 of the CPU capability bit array
286                  */
287                 c->x86_capability[5] = cpuid_edx(0xC0000001);
288         }
289
290         /* Cyrix III family needs CX8 & PGE explicity enabled. */
291         if (c->x86_model >=6 && c->x86_model <= 9) {
292                 rdmsr (MSR_VIA_FCR, lo, hi);
293                 lo |= (1<<1 | 1<<7);
294                 wrmsr (MSR_VIA_FCR, lo, hi);
295                 set_bit(X86_FEATURE_CX8, c->x86_capability);
296         }
297
298         /* Before Nehemiah, the C3's had 3dNOW! */
299         if (c->x86_model >=6 && c->x86_model <9)
300                 set_bit(X86_FEATURE_3DNOW, c->x86_capability);
301
302         get_model_name(c);
303         display_cacheinfo(c);
304 }
305
306 static void __cpuinit init_centaur(struct cpuinfo_x86 *c)
307 {
308         enum {
309                 ECX8=1<<1,
310                 EIERRINT=1<<2,
311                 DPM=1<<3,
312                 DMCE=1<<4,
313                 DSTPCLK=1<<5,
314                 ELINEAR=1<<6,
315                 DSMC=1<<7,
316                 DTLOCK=1<<8,
317                 EDCTLB=1<<8,
318                 EMMX=1<<9,
319                 DPDC=1<<11,
320                 EBRPRED=1<<12,
321                 DIC=1<<13,
322                 DDC=1<<14,
323                 DNA=1<<15,
324                 ERETSTK=1<<16,
325                 E2MMX=1<<19,
326                 EAMD3D=1<<20,
327         };
328
329         char *name;
330         u32  fcr_set=0;
331         u32  fcr_clr=0;
332         u32  lo,hi,newlo;
333         u32  aa,bb,cc,dd;
334
335         /* Bit 31 in normal CPUID used for nonstandard 3DNow ID;
336            3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway */
337         clear_bit(0*32+31, c->x86_capability);
338
339         switch (c->x86) {
340
341                 case 5:
342                         switch(c->x86_model) {
343                         case 4:
344                                 name="C6";
345                                 fcr_set=ECX8|DSMC|EDCTLB|EMMX|ERETSTK;
346                                 fcr_clr=DPDC;
347                                 printk(KERN_NOTICE "Disabling bugged TSC.\n");
348                                 clear_bit(X86_FEATURE_TSC, c->x86_capability);
349 #ifdef CONFIG_X86_OOSTORE
350                                 centaur_create_optimal_mcr();
351                                 /* Enable
352                                         write combining on non-stack, non-string
353                                         write combining on string, all types
354                                         weak write ordering 
355                                         
356                                    The C6 original lacks weak read order 
357                                    
358                                    Note 0x120 is write only on Winchip 1 */
359                                    
360                                 wrmsr(MSR_IDT_MCR_CTRL, 0x01F0001F, 0);
361 #endif                          
362                                 break;
363                         case 8:
364                                 switch(c->x86_mask) {
365                                 default:
366                                         name="2";
367                                         break;
368                                 case 7 ... 9:
369                                         name="2A";
370                                         break;
371                                 case 10 ... 15:
372                                         name="2B";
373                                         break;
374                                 }
375                                 fcr_set=ECX8|DSMC|DTLOCK|EMMX|EBRPRED|ERETSTK|E2MMX|EAMD3D;
376                                 fcr_clr=DPDC;
377 #ifdef CONFIG_X86_OOSTORE
378                                 winchip2_unprotect_mcr();
379                                 winchip2_create_optimal_mcr();
380                                 rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
381                                 /* Enable
382                                         write combining on non-stack, non-string
383                                         write combining on string, all types
384                                         weak write ordering 
385                                 */
386                                 lo|=31;                         
387                                 wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
388                                 winchip2_protect_mcr();
389 #endif
390                                 break;
391                         case 9:
392                                 name="3";
393                                 fcr_set=ECX8|DSMC|DTLOCK|EMMX|EBRPRED|ERETSTK|E2MMX|EAMD3D;
394                                 fcr_clr=DPDC;
395 #ifdef CONFIG_X86_OOSTORE
396                                 winchip2_unprotect_mcr();
397                                 winchip2_create_optimal_mcr();
398                                 rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
399                                 /* Enable
400                                         write combining on non-stack, non-string
401                                         write combining on string, all types
402                                         weak write ordering 
403                                 */
404                                 lo|=31;                         
405                                 wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
406                                 winchip2_protect_mcr();
407 #endif
408                                 break;
409                         default:
410                                 name="??";
411                         }
412
413                         rdmsr(MSR_IDT_FCR1, lo, hi);
414                         newlo=(lo|fcr_set) & (~fcr_clr);
415
416                         if (newlo!=lo) {
417                                 printk(KERN_INFO "Centaur FCR was 0x%X now 0x%X\n", lo, newlo );
418                                 wrmsr(MSR_IDT_FCR1, newlo, hi );
419                         } else {
420                                 printk(KERN_INFO "Centaur FCR is 0x%X\n",lo);
421                         }
422                         /* Emulate MTRRs using Centaur's MCR. */
423                         set_bit(X86_FEATURE_CENTAUR_MCR, c->x86_capability);
424                         /* Report CX8 */
425                         set_bit(X86_FEATURE_CX8, c->x86_capability);
426                         /* Set 3DNow! on Winchip 2 and above. */
427                         if (c->x86_model >=8)
428                                 set_bit(X86_FEATURE_3DNOW, c->x86_capability);
429                         /* See if we can find out some more. */
430                         if ( cpuid_eax(0x80000000) >= 0x80000005 ) {
431                                 /* Yes, we can. */
432                                 cpuid(0x80000005,&aa,&bb,&cc,&dd);
433                                 /* Add L1 data and code cache sizes. */
434                                 c->x86_cache_size = (cc>>24)+(dd>>24);
435                         }
436                         sprintf( c->x86_model_id, "WinChip %s", name );
437                         break;
438
439                 case 6:
440                         init_c3(c);
441                         break;
442         }
443 }
444
445 static unsigned int __cpuinit centaur_size_cache(struct cpuinfo_x86 * c, unsigned int size)
446 {
447         /* VIA C3 CPUs (670-68F) need further shifting. */
448         if ((c->x86 == 6) && ((c->x86_model == 7) || (c->x86_model == 8)))
449                 size >>= 8;
450
451         /* VIA also screwed up Nehemiah stepping 1, and made
452            it return '65KB' instead of '64KB'
453            - Note, it seems this may only be in engineering samples. */
454         if ((c->x86==6) && (c->x86_model==9) && (c->x86_mask==1) && (size==65))
455                 size -=1;
456
457         return size;
458 }
459
460 static struct cpu_dev centaur_cpu_dev __cpuinitdata = {
461         .c_vendor       = "Centaur",
462         .c_ident        = { "CentaurHauls" },
463         .c_init         = init_centaur,
464         .c_size_cache   = centaur_size_cache,
465 };
466
467 int __init centaur_init_cpu(void)
468 {
469         cpu_devs[X86_VENDOR_CENTAUR] = &centaur_cpu_dev;
470         return 0;
471 }
472
473 //early_arch_initcall(centaur_init_cpu);
474
475 static int __init centaur_exit_cpu(void)
476 {
477         cpu_devs[X86_VENDOR_CENTAUR] = NULL;
478         return 0;
479 }
480
481 late_initcall(centaur_exit_cpu);