1 #include <linux/bitops.h>
2 #include <linux/kernel.h>
3 #include <linux/init.h>
5 #include <asm/processor.h>
12 #ifdef CONFIG_X86_OOSTORE
14 static u32 power2(u32 x)
26 * Set up an actual MCR
28 static void centaur_mcr_insert(int reg, u32 base, u32 size, int key)
33 lo = ~(size-1); /* Size is a power of 2 so this makes a mask */
34 lo &= ~0xFFF; /* Remove the ctrl value bits */
35 lo |= key; /* Attribute we wish to set */
36 wrmsr(reg+MSR_IDT_MCR0, lo, hi);
37 mtrr_centaur_report_mcr(reg, lo, hi); /* Tell the mtrr driver */
41 * Figure what we can cover with MCR's
43 * Shortcut: We know you can't put 4Gig of RAM on a winchip
45 static u32 ramtop(void)
47 u32 clip = 0xFFFFFFFFUL;
51 for (i = 0; i < e820.nr_map; i++) {
52 unsigned long start, end;
54 if (e820.map[i].addr > 0xFFFFFFFFUL)
57 * Don't MCR over reserved space. Ignore the ISA hole
58 * we frob around that catastrophe already
60 if (e820.map[i].type == E820_RESERVED) {
61 if (e820.map[i].addr >= 0x100000UL &&
62 e820.map[i].addr < clip)
63 clip = e820.map[i].addr;
66 start = e820.map[i].addr;
67 end = e820.map[i].addr + e820.map[i].size;
74 * Everything below 'top' should be RAM except for the ISA hole.
75 * Because of the limited MCR's we want to map NV/ACPI into our
76 * MCR range for gunk in RAM
78 * Clip might cause us to MCR insufficient RAM but that is an
79 * acceptable failure mode and should only bite obscure boxes with
82 * The second case Clip sometimes kicks in is when the EBDA is marked
83 * as reserved. Again we fail safe with reasonable results
92 * Compute a set of MCR's to give maximum coverage
94 static int centaur_mcr_compute(int nr, int key)
97 u32 root = power2(mem);
109 * Find the largest block we will fill going upwards
111 high = power2(mem-top);
114 * Find the largest block we will fill going downwards
119 * Don't fill below 1Mb going downwards as there
120 * is an ISA hole in the way.
122 if (base <= 1024*1024)
126 * See how much space we could cover by filling below
132 else if (floor == 512*1024)
135 /* And forget ROM space */
138 * Now install the largest coverage we get
140 if (fspace > high && fspace > low) {
141 centaur_mcr_insert(ct, floor, fspace, key);
143 } else if (high > low) {
144 centaur_mcr_insert(ct, top, high, key);
146 } else if (low > 0) {
148 centaur_mcr_insert(ct, base, low, key);
154 * We loaded ct values. We now need to set the mask. The caller
160 static void centaur_create_optimal_mcr(void)
166 * Allocate up to 6 mcrs to mark as much of ram as possible
167 * as write combining and weak write ordered.
169 * To experiment with: Linux never uses stack operations for
170 * mmio spaces so we could globally enable stack operation wc
172 * Load the registers with type 31 - full write combining, all
173 * writes weakly ordered.
175 used = centaur_mcr_compute(6, 31);
180 for (i = used; i < 8; i++)
181 wrmsr(MSR_IDT_MCR0+i, 0, 0);
184 static void winchip2_create_optimal_mcr(void)
191 * Allocate up to 6 mcrs to mark as much of ram as possible
192 * as write combining, weak store ordered.
194 * Load the registers with type 25
195 * 8 - weak write ordering
196 * 16 - weak read ordering
197 * 1 - write combining
199 used = centaur_mcr_compute(6, 25);
202 * Mark the registers we are using.
204 rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
205 for (i = 0; i < used; i++)
207 wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
213 for (i = used; i < 8; i++)
214 wrmsr(MSR_IDT_MCR0+i, 0, 0);
218 * Handle the MCR key on the Winchip 2.
220 static void winchip2_unprotect_mcr(void)
225 rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
226 lo &= ~0x1C0; /* blank bits 8-6 */
228 lo |= key<<6; /* replace with unlock key */
229 wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
232 static void winchip2_protect_mcr(void)
236 rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
237 lo &= ~0x1C0; /* blank bits 8-6 */
238 wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
240 #endif /* CONFIG_X86_OOSTORE */
242 #define ACE_PRESENT (1 << 6)
243 #define ACE_ENABLED (1 << 7)
244 #define ACE_FCR (1 << 28) /* MSR_VIA_FCR */
246 #define RNG_PRESENT (1 << 2)
247 #define RNG_ENABLED (1 << 3)
248 #define RNG_ENABLE (1 << 6) /* MSR_VIA_RNG */
250 static void init_c3(struct cpuinfo_x86 *c)
254 /* Test for Centaur Extended Feature Flags presence */
255 if (cpuid_eax(0xC0000000) >= 0xC0000001) {
256 u32 tmp = cpuid_edx(0xC0000001);
258 /* enable ACE unit, if present and disabled */
259 if ((tmp & (ACE_PRESENT | ACE_ENABLED)) == ACE_PRESENT) {
260 rdmsr(MSR_VIA_FCR, lo, hi);
261 lo |= ACE_FCR; /* enable ACE unit */
262 wrmsr(MSR_VIA_FCR, lo, hi);
263 printk(KERN_INFO "CPU: Enabled ACE h/w crypto\n");
266 /* enable RNG unit, if present and disabled */
267 if ((tmp & (RNG_PRESENT | RNG_ENABLED)) == RNG_PRESENT) {
268 rdmsr(MSR_VIA_RNG, lo, hi);
269 lo |= RNG_ENABLE; /* enable RNG unit */
270 wrmsr(MSR_VIA_RNG, lo, hi);
271 printk(KERN_INFO "CPU: Enabled h/w RNG\n");
274 /* store Centaur Extended Feature Flags as
275 * word 5 of the CPU capability bit array
277 c->x86_capability[5] = cpuid_edx(0xC0000001);
280 /* Cyrix III family needs CX8 & PGE explicitly enabled. */
281 if (c->x86_model >= 6 && c->x86_model <= 13) {
282 rdmsr(MSR_VIA_FCR, lo, hi);
284 wrmsr(MSR_VIA_FCR, lo, hi);
285 set_cpu_cap(c, X86_FEATURE_CX8);
288 /* Before Nehemiah, the C3's had 3dNOW! */
289 if (c->x86_model >= 6 && c->x86_model < 9)
290 set_cpu_cap(c, X86_FEATURE_3DNOW);
292 if (c->x86 == 0x6 && c->x86_model >= 0xf) {
293 c->x86_cache_alignment = c->x86_clflush_size * 2;
294 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
297 cpu_detect_cache_sizes(c);
321 static void early_init_centaur(struct cpuinfo_x86 *c)
326 /* Emulate MTRRs using Centaur's MCR. */
327 set_cpu_cap(c, X86_FEATURE_CENTAUR_MCR);
331 if (c->x86_model >= 0xf)
332 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
336 set_cpu_cap(c, X86_FEATURE_SYSENTER32);
340 static void init_centaur(struct cpuinfo_x86 *c)
350 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
351 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
353 clear_cpu_cap(c, 0*32+31);
355 early_init_centaur(c);
359 switch (c->x86_model) {
362 fcr_set = ECX8|DSMC|EDCTLB|EMMX|ERETSTK;
364 printk(KERN_NOTICE "Disabling bugged TSC.\n");
365 clear_cpu_cap(c, X86_FEATURE_TSC);
366 #ifdef CONFIG_X86_OOSTORE
367 centaur_create_optimal_mcr();
370 * write combining on non-stack, non-string
371 * write combining on string, all types
372 * weak write ordering
374 * The C6 original lacks weak read order
376 * Note 0x120 is write only on Winchip 1
378 wrmsr(MSR_IDT_MCR_CTRL, 0x01F0001F, 0);
382 switch (c->x86_mask) {
393 fcr_set = ECX8|DSMC|DTLOCK|EMMX|EBRPRED|ERETSTK|
396 #ifdef CONFIG_X86_OOSTORE
397 winchip2_unprotect_mcr();
398 winchip2_create_optimal_mcr();
399 rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
402 * write combining on non-stack, non-string
403 * write combining on string, all types
404 * weak write ordering
407 wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
408 winchip2_protect_mcr();
413 fcr_set = ECX8|DSMC|DTLOCK|EMMX|EBRPRED|ERETSTK|
416 #ifdef CONFIG_X86_OOSTORE
417 winchip2_unprotect_mcr();
418 winchip2_create_optimal_mcr();
419 rdmsr(MSR_IDT_MCR_CTRL, lo, hi);
422 * write combining on non-stack, non-string
423 * write combining on string, all types
424 * weak write ordering
427 wrmsr(MSR_IDT_MCR_CTRL, lo, hi);
428 winchip2_protect_mcr();
435 rdmsr(MSR_IDT_FCR1, lo, hi);
436 newlo = (lo|fcr_set) & (~fcr_clr);
439 printk(KERN_INFO "Centaur FCR was 0x%X now 0x%X\n",
441 wrmsr(MSR_IDT_FCR1, newlo, hi);
443 printk(KERN_INFO "Centaur FCR is 0x%X\n", lo);
445 /* Emulate MTRRs using Centaur's MCR. */
446 set_cpu_cap(c, X86_FEATURE_CENTAUR_MCR);
448 set_cpu_cap(c, X86_FEATURE_CX8);
449 /* Set 3DNow! on Winchip 2 and above. */
450 if (c->x86_model >= 8)
451 set_cpu_cap(c, X86_FEATURE_3DNOW);
452 /* See if we can find out some more. */
453 if (cpuid_eax(0x80000000) >= 0x80000005) {
455 cpuid(0x80000005, &aa, &bb, &cc, &dd);
456 /* Add L1 data and code cache sizes. */
457 c->x86_cache_size = (cc>>24)+(dd>>24);
459 sprintf(c->x86_model_id, "WinChip %s", name);
467 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
472 centaur_size_cache(struct cpuinfo_x86 *c, unsigned int size)
475 /* VIA C3 CPUs (670-68F) need further shifting. */
476 if ((c->x86 == 6) && ((c->x86_model == 7) || (c->x86_model == 8)))
480 * There's also an erratum in Nehemiah stepping 1, which
481 * returns '65KB' instead of '64KB'
482 * - Note, it seems this may only be in engineering samples.
484 if ((c->x86 == 6) && (c->x86_model == 9) &&
485 (c->x86_mask == 1) && (size == 65))
491 static const struct cpu_dev centaur_cpu_dev = {
492 .c_vendor = "Centaur",
493 .c_ident = { "CentaurHauls" },
494 .c_early_init = early_init_centaur,
495 .c_init = init_centaur,
496 .c_size_cache = centaur_size_cache,
497 .c_x86_vendor = X86_VENDOR_CENTAUR,
500 cpu_dev_register(centaur_cpu_dev);