1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * (C) Copyright 2008-2011
4 * Graeme Russ, <graeme.russ@gmail.com>
5 *
6 * (C) Copyright 2002
7 * Daniel Engström, Omicron Ceti AB, <daniel@omicron.se>
8 *
9 * (C) Copyright 2002
10 * Sysgo Real-Time Solutions, GmbH <www.elinos.com>
11 * Marius Groeger <mgroeger@sysgo.de>
12 *
13 * (C) Copyright 2002
14 * Sysgo Real-Time Solutions, GmbH <www.elinos.com>
15 * Alex Zuepke <azu@sysgo.de>
16 *
17 * Part of this file is adapted from coreboot
18 * src/arch/x86/lib/cpu.c
19 */
20
21 #include <common.h>
22 #include <malloc.h>
23 #include <asm/control_regs.h>
24 #include <asm/cpu.h>
25 #include <asm/mp.h>
26 #include <asm/msr.h>
27 #include <asm/mtrr.h>
28 #include <asm/processor-flags.h>
29
30 DECLARE_GLOBAL_DATA_PTR;
31
32 /*
33 * Constructor for a conventional segment GDT (or LDT) entry
34 * This is a macro so it can be used in initialisers
35 */
36 #define GDT_ENTRY(flags, base, limit) \
37 ((((base) & 0xff000000ULL) << (56-24)) | \
38 (((flags) & 0x0000f0ffULL) << 40) | \
39 (((limit) & 0x000f0000ULL) << (48-16)) | \
40 (((base) & 0x00ffffffULL) << 16) | \
41 (((limit) & 0x0000ffffULL)))
42
43 struct gdt_ptr {
44 u16 len;
45 u32 ptr;
46 } __packed;
47
48 struct cpu_device_id {
49 unsigned vendor;
50 unsigned device;
51 };
52
53 struct cpuinfo_x86 {
54 uint8_t x86; /* CPU family */
55 uint8_t x86_vendor; /* CPU vendor */
56 uint8_t x86_model;
57 uint8_t x86_mask;
58 };
59
60 /*
61 * List of cpu vendor strings along with their normalized
62 * id values.
63 */
64 static const struct {
65 int vendor;
66 const char *name;
67 } x86_vendors[] = {
68 { X86_VENDOR_INTEL, "GenuineIntel", },
69 { X86_VENDOR_CYRIX, "CyrixInstead", },
70 { X86_VENDOR_AMD, "AuthenticAMD", },
71 { X86_VENDOR_UMC, "UMC UMC UMC ", },
72 { X86_VENDOR_NEXGEN, "NexGenDriven", },
73 { X86_VENDOR_CENTAUR, "CentaurHauls", },
74 { X86_VENDOR_RISE, "RiseRiseRise", },
75 { X86_VENDOR_TRANSMETA, "GenuineTMx86", },
76 { X86_VENDOR_TRANSMETA, "TransmetaCPU", },
77 { X86_VENDOR_NSC, "Geode by NSC", },
78 { X86_VENDOR_SIS, "SiS SiS SiS ", },
79 };
80
load_ds(u32 segment)81 static void load_ds(u32 segment)
82 {
83 asm volatile("movl %0, %%ds" : : "r" (segment * X86_GDT_ENTRY_SIZE));
84 }
85
load_es(u32 segment)86 static void load_es(u32 segment)
87 {
88 asm volatile("movl %0, %%es" : : "r" (segment * X86_GDT_ENTRY_SIZE));
89 }
90
load_fs(u32 segment)91 static void load_fs(u32 segment)
92 {
93 asm volatile("movl %0, %%fs" : : "r" (segment * X86_GDT_ENTRY_SIZE));
94 }
95
load_gs(u32 segment)96 static void load_gs(u32 segment)
97 {
98 asm volatile("movl %0, %%gs" : : "r" (segment * X86_GDT_ENTRY_SIZE));
99 }
100
load_ss(u32 segment)101 static void load_ss(u32 segment)
102 {
103 asm volatile("movl %0, %%ss" : : "r" (segment * X86_GDT_ENTRY_SIZE));
104 }
105
load_gdt(const u64 * boot_gdt,u16 num_entries)106 static void load_gdt(const u64 *boot_gdt, u16 num_entries)
107 {
108 struct gdt_ptr gdt;
109
110 gdt.len = (num_entries * X86_GDT_ENTRY_SIZE) - 1;
111 gdt.ptr = (ulong)boot_gdt;
112
113 asm volatile("lgdtl %0\n" : : "m" (gdt));
114 }
115
arch_setup_gd(gd_t * new_gd)116 void arch_setup_gd(gd_t *new_gd)
117 {
118 u64 *gdt_addr;
119
120 gdt_addr = new_gd->arch.gdt;
121
122 /*
123 * CS: code, read/execute, 4 GB, base 0
124 *
125 * Some OS (like VxWorks) requires GDT entry 1 to be the 32-bit CS
126 */
127 gdt_addr[X86_GDT_ENTRY_UNUSED] = GDT_ENTRY(0xc09b, 0, 0xfffff);
128 gdt_addr[X86_GDT_ENTRY_32BIT_CS] = GDT_ENTRY(0xc09b, 0, 0xfffff);
129
130 /* DS: data, read/write, 4 GB, base 0 */
131 gdt_addr[X86_GDT_ENTRY_32BIT_DS] = GDT_ENTRY(0xc093, 0, 0xfffff);
132
133 /* FS: data, read/write, 4 GB, base (Global Data Pointer) */
134 new_gd->arch.gd_addr = new_gd;
135 gdt_addr[X86_GDT_ENTRY_32BIT_FS] = GDT_ENTRY(0xc093,
136 (ulong)&new_gd->arch.gd_addr, 0xfffff);
137
138 /* 16-bit CS: code, read/execute, 64 kB, base 0 */
139 gdt_addr[X86_GDT_ENTRY_16BIT_CS] = GDT_ENTRY(0x009b, 0, 0x0ffff);
140
141 /* 16-bit DS: data, read/write, 64 kB, base 0 */
142 gdt_addr[X86_GDT_ENTRY_16BIT_DS] = GDT_ENTRY(0x0093, 0, 0x0ffff);
143
144 gdt_addr[X86_GDT_ENTRY_16BIT_FLAT_CS] = GDT_ENTRY(0x809b, 0, 0xfffff);
145 gdt_addr[X86_GDT_ENTRY_16BIT_FLAT_DS] = GDT_ENTRY(0x8093, 0, 0xfffff);
146
147 load_gdt(gdt_addr, X86_GDT_NUM_ENTRIES);
148 load_ds(X86_GDT_ENTRY_32BIT_DS);
149 load_es(X86_GDT_ENTRY_32BIT_DS);
150 load_gs(X86_GDT_ENTRY_32BIT_DS);
151 load_ss(X86_GDT_ENTRY_32BIT_DS);
152 load_fs(X86_GDT_ENTRY_32BIT_FS);
153 }
154
155 #ifdef CONFIG_HAVE_FSP
156 /*
157 * Setup FSP execution environment GDT
158 *
159 * Per Intel FSP external architecture specification, before calling any FSP
160 * APIs, we need make sure the system is in flat 32-bit mode and both the code
161 * and data selectors should have full 4GB access range. Here we reuse the one
162 * we used in arch/x86/cpu/start16.S, and reload the segement registers.
163 */
setup_fsp_gdt(void)164 void setup_fsp_gdt(void)
165 {
166 load_gdt((const u64 *)(gdt_rom + CONFIG_RESET_SEG_START), 4);
167 load_ds(X86_GDT_ENTRY_32BIT_DS);
168 load_ss(X86_GDT_ENTRY_32BIT_DS);
169 load_es(X86_GDT_ENTRY_32BIT_DS);
170 load_fs(X86_GDT_ENTRY_32BIT_DS);
171 load_gs(X86_GDT_ENTRY_32BIT_DS);
172 }
173 #endif
174
175 /*
176 * Cyrix CPUs without cpuid or with cpuid not yet enabled can be detected
177 * by the fact that they preserve the flags across the division of 5/2.
178 * PII and PPro exhibit this behavior too, but they have cpuid available.
179 */
180
181 /*
182 * Perform the Cyrix 5/2 test. A Cyrix won't change
183 * the flags, while other 486 chips will.
184 */
test_cyrix_52div(void)185 static inline int test_cyrix_52div(void)
186 {
187 unsigned int test;
188
189 __asm__ __volatile__(
190 "sahf\n\t" /* clear flags (%eax = 0x0005) */
191 "div %b2\n\t" /* divide 5 by 2 */
192 "lahf" /* store flags into %ah */
193 : "=a" (test)
194 : "0" (5), "q" (2)
195 : "cc");
196
197 /* AH is 0x02 on Cyrix after the divide.. */
198 return (unsigned char) (test >> 8) == 0x02;
199 }
200
201 /*
202 * Detect a NexGen CPU running without BIOS hypercode new enough
203 * to have CPUID. (Thanks to Herbert Oppmann)
204 */
deep_magic_nexgen_probe(void)205 static int deep_magic_nexgen_probe(void)
206 {
207 int ret;
208
209 __asm__ __volatile__ (
210 " movw $0x5555, %%ax\n"
211 " xorw %%dx,%%dx\n"
212 " movw $2, %%cx\n"
213 " divw %%cx\n"
214 " movl $0, %%eax\n"
215 " jnz 1f\n"
216 " movl $1, %%eax\n"
217 "1:\n"
218 : "=a" (ret) : : "cx", "dx");
219 return ret;
220 }
221
has_cpuid(void)222 static bool has_cpuid(void)
223 {
224 return flag_is_changeable_p(X86_EFLAGS_ID);
225 }
226
has_mtrr(void)227 static bool has_mtrr(void)
228 {
229 return cpuid_edx(0x00000001) & (1 << 12) ? true : false;
230 }
231
build_vendor_name(char * vendor_name)232 static int build_vendor_name(char *vendor_name)
233 {
234 struct cpuid_result result;
235 result = cpuid(0x00000000);
236 unsigned int *name_as_ints = (unsigned int *)vendor_name;
237
238 name_as_ints[0] = result.ebx;
239 name_as_ints[1] = result.edx;
240 name_as_ints[2] = result.ecx;
241
242 return result.eax;
243 }
244
identify_cpu(struct cpu_device_id * cpu)245 static void identify_cpu(struct cpu_device_id *cpu)
246 {
247 char vendor_name[16];
248 int i;
249
250 vendor_name[0] = '\0'; /* Unset */
251 cpu->device = 0; /* fix gcc 4.4.4 warning */
252
253 /* Find the id and vendor_name */
254 if (!has_cpuid()) {
255 /* Its a 486 if we can modify the AC flag */
256 if (flag_is_changeable_p(X86_EFLAGS_AC))
257 cpu->device = 0x00000400; /* 486 */
258 else
259 cpu->device = 0x00000300; /* 386 */
260 if ((cpu->device == 0x00000400) && test_cyrix_52div()) {
261 memcpy(vendor_name, "CyrixInstead", 13);
262 /* If we ever care we can enable cpuid here */
263 }
264 /* Detect NexGen with old hypercode */
265 else if (deep_magic_nexgen_probe())
266 memcpy(vendor_name, "NexGenDriven", 13);
267 }
268 if (has_cpuid()) {
269 int cpuid_level;
270
271 cpuid_level = build_vendor_name(vendor_name);
272 vendor_name[12] = '\0';
273
274 /* Intel-defined flags: level 0x00000001 */
275 if (cpuid_level >= 0x00000001) {
276 cpu->device = cpuid_eax(0x00000001);
277 } else {
278 /* Have CPUID level 0 only unheard of */
279 cpu->device = 0x00000400;
280 }
281 }
282 cpu->vendor = X86_VENDOR_UNKNOWN;
283 for (i = 0; i < ARRAY_SIZE(x86_vendors); i++) {
284 if (memcmp(vendor_name, x86_vendors[i].name, 12) == 0) {
285 cpu->vendor = x86_vendors[i].vendor;
286 break;
287 }
288 }
289 }
290
get_fms(struct cpuinfo_x86 * c,uint32_t tfms)291 static inline void get_fms(struct cpuinfo_x86 *c, uint32_t tfms)
292 {
293 c->x86 = (tfms >> 8) & 0xf;
294 c->x86_model = (tfms >> 4) & 0xf;
295 c->x86_mask = tfms & 0xf;
296 if (c->x86 == 0xf)
297 c->x86 += (tfms >> 20) & 0xff;
298 if (c->x86 >= 0x6)
299 c->x86_model += ((tfms >> 16) & 0xF) << 4;
300 }
301
cpu_get_family_model(void)302 u32 cpu_get_family_model(void)
303 {
304 return gd->arch.x86_device & 0x0fff0ff0;
305 }
306
cpu_get_stepping(void)307 u32 cpu_get_stepping(void)
308 {
309 return gd->arch.x86_mask;
310 }
311
x86_cpu_init_f(void)312 int x86_cpu_init_f(void)
313 {
314 const u32 em_rst = ~X86_CR0_EM;
315 const u32 mp_ne_set = X86_CR0_MP | X86_CR0_NE;
316
317 if (ll_boot_init()) {
318 /* initialize FPU, reset EM, set MP and NE */
319 asm ("fninit\n" \
320 "movl %%cr0, %%eax\n" \
321 "andl %0, %%eax\n" \
322 "orl %1, %%eax\n" \
323 "movl %%eax, %%cr0\n" \
324 : : "i" (em_rst), "i" (mp_ne_set) : "eax");
325 }
326
327 /* identify CPU via cpuid and store the decoded info into gd->arch */
328 if (has_cpuid()) {
329 struct cpu_device_id cpu;
330 struct cpuinfo_x86 c;
331
332 identify_cpu(&cpu);
333 get_fms(&c, cpu.device);
334 gd->arch.x86 = c.x86;
335 gd->arch.x86_vendor = cpu.vendor;
336 gd->arch.x86_model = c.x86_model;
337 gd->arch.x86_mask = c.x86_mask;
338 gd->arch.x86_device = cpu.device;
339
340 gd->arch.has_mtrr = has_mtrr();
341 }
342 /* Don't allow PCI region 3 to use memory in the 2-4GB memory hole */
343 gd->pci_ram_top = 0x80000000U;
344
345 /* Configure fixed range MTRRs for some legacy regions */
346 if (gd->arch.has_mtrr) {
347 u64 mtrr_cap;
348
349 mtrr_cap = native_read_msr(MTRR_CAP_MSR);
350 if (mtrr_cap & MTRR_CAP_FIX) {
351 /* Mark the VGA RAM area as uncacheable */
352 native_write_msr(MTRR_FIX_16K_A0000_MSR,
353 MTRR_FIX_TYPE(MTRR_TYPE_UNCACHEABLE),
354 MTRR_FIX_TYPE(MTRR_TYPE_UNCACHEABLE));
355
356 /*
357 * Mark the PCI ROM area as cacheable to improve ROM
358 * execution performance.
359 */
360 native_write_msr(MTRR_FIX_4K_C0000_MSR,
361 MTRR_FIX_TYPE(MTRR_TYPE_WRBACK),
362 MTRR_FIX_TYPE(MTRR_TYPE_WRBACK));
363 native_write_msr(MTRR_FIX_4K_C8000_MSR,
364 MTRR_FIX_TYPE(MTRR_TYPE_WRBACK),
365 MTRR_FIX_TYPE(MTRR_TYPE_WRBACK));
366 native_write_msr(MTRR_FIX_4K_D0000_MSR,
367 MTRR_FIX_TYPE(MTRR_TYPE_WRBACK),
368 MTRR_FIX_TYPE(MTRR_TYPE_WRBACK));
369 native_write_msr(MTRR_FIX_4K_D8000_MSR,
370 MTRR_FIX_TYPE(MTRR_TYPE_WRBACK),
371 MTRR_FIX_TYPE(MTRR_TYPE_WRBACK));
372
373 /* Enable the fixed range MTRRs */
374 msr_setbits_64(MTRR_DEF_TYPE_MSR, MTRR_DEF_TYPE_FIX_EN);
375 }
376 }
377
378 #ifdef CONFIG_I8254_TIMER
379 /* Set up the i8254 timer if required */
380 i8254_init();
381 #endif
382
383 return 0;
384 }
385
x86_enable_caches(void)386 void x86_enable_caches(void)
387 {
388 unsigned long cr0;
389
390 cr0 = read_cr0();
391 cr0 &= ~(X86_CR0_NW | X86_CR0_CD);
392 write_cr0(cr0);
393 wbinvd();
394 }
395 void enable_caches(void) __attribute__((weak, alias("x86_enable_caches")));
396
x86_disable_caches(void)397 void x86_disable_caches(void)
398 {
399 unsigned long cr0;
400
401 cr0 = read_cr0();
402 cr0 |= X86_CR0_NW | X86_CR0_CD;
403 wbinvd();
404 write_cr0(cr0);
405 wbinvd();
406 }
407 void disable_caches(void) __attribute__((weak, alias("x86_disable_caches")));
408
dcache_status(void)409 int dcache_status(void)
410 {
411 return !(read_cr0() & X86_CR0_CD);
412 }
413
cpu_enable_paging_pae(ulong cr3)414 void cpu_enable_paging_pae(ulong cr3)
415 {
416 __asm__ __volatile__(
417 /* Load the page table address */
418 "movl %0, %%cr3\n"
419 /* Enable pae */
420 "movl %%cr4, %%eax\n"
421 "orl $0x00000020, %%eax\n"
422 "movl %%eax, %%cr4\n"
423 /* Enable paging */
424 "movl %%cr0, %%eax\n"
425 "orl $0x80000000, %%eax\n"
426 "movl %%eax, %%cr0\n"
427 :
428 : "r" (cr3)
429 : "eax");
430 }
431
cpu_disable_paging_pae(void)432 void cpu_disable_paging_pae(void)
433 {
434 /* Turn off paging */
435 __asm__ __volatile__ (
436 /* Disable paging */
437 "movl %%cr0, %%eax\n"
438 "andl $0x7fffffff, %%eax\n"
439 "movl %%eax, %%cr0\n"
440 /* Disable pae */
441 "movl %%cr4, %%eax\n"
442 "andl $0xffffffdf, %%eax\n"
443 "movl %%eax, %%cr4\n"
444 :
445 :
446 : "eax");
447 }
448
can_detect_long_mode(void)449 static bool can_detect_long_mode(void)
450 {
451 return cpuid_eax(0x80000000) > 0x80000000UL;
452 }
453
has_long_mode(void)454 static bool has_long_mode(void)
455 {
456 return cpuid_edx(0x80000001) & (1 << 29) ? true : false;
457 }
458
cpu_has_64bit(void)459 int cpu_has_64bit(void)
460 {
461 return has_cpuid() && can_detect_long_mode() &&
462 has_long_mode();
463 }
464
465 #define PAGETABLE_SIZE (6 * 4096)
466
467 /**
468 * build_pagetable() - build a flat 4GiB page table structure for 64-bti mode
469 *
470 * @pgtable: Pointer to a 24iKB block of memory
471 */
build_pagetable(uint32_t * pgtable)472 static void build_pagetable(uint32_t *pgtable)
473 {
474 uint i;
475
476 memset(pgtable, '\0', PAGETABLE_SIZE);
477
478 /* Level 4 needs a single entry */
479 pgtable[0] = (ulong)&pgtable[1024] + 7;
480
481 /* Level 3 has one 64-bit entry for each GiB of memory */
482 for (i = 0; i < 4; i++)
483 pgtable[1024 + i * 2] = (ulong)&pgtable[2048] + 0x1000 * i + 7;
484
485 /* Level 2 has 2048 64-bit entries, each repesenting 2MiB */
486 for (i = 0; i < 2048; i++)
487 pgtable[2048 + i * 2] = 0x183 + (i << 21UL);
488 }
489
cpu_jump_to_64bit(ulong setup_base,ulong target)490 int cpu_jump_to_64bit(ulong setup_base, ulong target)
491 {
492 uint32_t *pgtable;
493
494 pgtable = memalign(4096, PAGETABLE_SIZE);
495 if (!pgtable)
496 return -ENOMEM;
497
498 build_pagetable(pgtable);
499 cpu_call64((ulong)pgtable, setup_base, target);
500 free(pgtable);
501
502 return -EFAULT;
503 }
504
505 /*
506 * Jump from SPL to U-Boot
507 *
508 * This function is work-in-progress with many issues to resolve.
509 *
510 * It works by setting up several regions:
511 * ptr - a place to put the code that jumps into 64-bit mode
512 * gdt - a place to put the global descriptor table
513 * pgtable - a place to put the page tables
514 *
515 * The cpu_call64() code is copied from ROM and then manually patched so that
516 * it has the correct GDT address in RAM. U-Boot is copied from ROM into
517 * its pre-relocation address. Then we jump to the cpu_call64() code in RAM,
518 * which changes to 64-bit mode and starts U-Boot.
519 */
cpu_jump_to_64bit_uboot(ulong target)520 int cpu_jump_to_64bit_uboot(ulong target)
521 {
522 typedef void (*func_t)(ulong pgtable, ulong setup_base, ulong target);
523 uint32_t *pgtable;
524 func_t func;
525
526 /* TODO(sjg@chromium.org): Find a better place for this */
527 pgtable = (uint32_t *)0x1000000;
528 if (!pgtable)
529 return -ENOMEM;
530
531 build_pagetable(pgtable);
532
533 /* TODO(sjg@chromium.org): Find a better place for this */
534 char *ptr = (char *)0x3000000;
535 char *gdt = (char *)0x3100000;
536
537 extern char gdt64[];
538
539 memcpy(ptr, cpu_call64, 0x1000);
540 memcpy(gdt, gdt64, 0x100);
541
542 /*
543 * TODO(sjg@chromium.org): This manually inserts the pointers into
544 * the code. Tidy this up to avoid this.
545 */
546 func = (func_t)ptr;
547 ulong ofs = (ulong)cpu_call64 - (ulong)ptr;
548 *(ulong *)(ptr + 7) = (ulong)gdt;
549 *(ulong *)(ptr + 0xc) = (ulong)gdt + 2;
550 *(ulong *)(ptr + 0x13) = (ulong)gdt;
551 *(ulong *)(ptr + 0x117 - 0xd4) -= ofs;
552
553 /*
554 * Copy U-Boot from ROM
555 * TODO(sjg@chromium.org): Figure out a way to get the text base
556 * correctly here, and in the device-tree binman definition.
557 *
558 * Also consider using FIT so we get the correct image length and
559 * parameters.
560 */
561 memcpy((char *)target, (char *)0xfff00000, 0x100000);
562
563 /* Jump to U-Boot */
564 func((ulong)pgtable, 0, (ulong)target);
565
566 return -EFAULT;
567 }
568
569 #ifdef CONFIG_SMP
enable_smis(struct udevice * cpu,void * unused)570 static int enable_smis(struct udevice *cpu, void *unused)
571 {
572 return 0;
573 }
574
575 static struct mp_flight_record mp_steps[] = {
576 MP_FR_BLOCK_APS(mp_init_cpu, NULL, mp_init_cpu, NULL),
577 /* Wait for APs to finish initialization before proceeding */
578 MP_FR_BLOCK_APS(NULL, NULL, enable_smis, NULL),
579 };
580
x86_mp_init(void)581 int x86_mp_init(void)
582 {
583 struct mp_params mp_params;
584
585 mp_params.parallel_microcode_load = 0,
586 mp_params.flight_plan = &mp_steps[0];
587 mp_params.num_records = ARRAY_SIZE(mp_steps);
588 mp_params.microcode_pointer = 0;
589
590 if (mp_init(&mp_params)) {
591 printf("Warning: MP init failure\n");
592 return -EIO;
593 }
594
595 return 0;
596 }
597 #endif
598