xref: /openbmc/linux/arch/x86/kernel/cpu/mtrr/generic.c (revision 7583028d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * This only handles 32bit MTRR on 32bit hosts. This is strictly wrong
4  * because MTRRs can span up to 40 bits (36bits on most modern x86)
5  */
6 
7 #include <linux/export.h>
8 #include <linux/init.h>
9 #include <linux/io.h>
10 #include <linux/mm.h>
11 #include <linux/cc_platform.h>
12 #include <asm/processor-flags.h>
13 #include <asm/cacheinfo.h>
14 #include <asm/cpufeature.h>
15 #include <asm/hypervisor.h>
16 #include <asm/mshyperv.h>
17 #include <asm/tlbflush.h>
18 #include <asm/mtrr.h>
19 #include <asm/msr.h>
20 #include <asm/memtype.h>
21 
22 #include "mtrr.h"
23 
24 struct fixed_range_block {
25 	int base_msr;		/* start address of an MTRR block */
26 	int ranges;		/* number of MTRRs in this block  */
27 };
28 
29 static struct fixed_range_block fixed_range_blocks[] = {
30 	{ MSR_MTRRfix64K_00000, 1 }, /* one   64k MTRR  */
31 	{ MSR_MTRRfix16K_80000, 2 }, /* two   16k MTRRs */
32 	{ MSR_MTRRfix4K_C0000,  8 }, /* eight  4k MTRRs */
33 	{}
34 };
35 
36 struct cache_map {
37 	u64 start;
38 	u64 end;
39 	u64 flags;
40 	u64 type:8;
41 	u64 fixed:1;
42 };
43 
44 bool mtrr_debug;
45 
46 static int __init mtrr_param_setup(char *str)
47 {
48 	int rc = 0;
49 
50 	if (!str)
51 		return -EINVAL;
52 	if (!strcmp(str, "debug"))
53 		mtrr_debug = true;
54 	else
55 		rc = -EINVAL;
56 
57 	return rc;
58 }
59 early_param("mtrr", mtrr_param_setup);
60 
61 /*
62  * CACHE_MAP_MAX is the maximum number of memory ranges in cache_map, where
63  * no 2 adjacent ranges have the same cache mode (those would be merged).
64  * The number is based on the worst case:
65  * - no two adjacent fixed MTRRs share the same cache mode
66  * - one variable MTRR is spanning a huge area with mode WB
67  * - 255 variable MTRRs with mode UC all overlap with the WB MTRR, creating 2
68  *   additional ranges each (result like "ababababa...aba" with a = WB, b = UC),
69  *   accounting for MTRR_MAX_VAR_RANGES * 2 - 1 range entries
70  * - a TOP_MEM2 area (even with overlapping an UC MTRR can't add 2 range entries
71  *   to the possible maximum, as it always starts at 4GB, thus it can't be in
72  *   the middle of that MTRR, unless that MTRR starts at 0, which would remove
73  *   the initial "a" from the "abababa" pattern above)
74  * The map won't contain ranges with no matching MTRR (those fall back to the
75  * default cache mode).
76  */
77 #define CACHE_MAP_MAX	(MTRR_NUM_FIXED_RANGES + MTRR_MAX_VAR_RANGES * 2)
78 
79 static struct cache_map init_cache_map[CACHE_MAP_MAX] __initdata;
80 static struct cache_map *cache_map __refdata = init_cache_map;
81 static unsigned int cache_map_size = CACHE_MAP_MAX;
82 static unsigned int cache_map_n;
83 static unsigned int cache_map_fixed;
84 
85 static unsigned long smp_changes_mask;
86 static int mtrr_state_set;
87 u64 mtrr_tom2;
88 
89 struct mtrr_state_type mtrr_state;
90 EXPORT_SYMBOL_GPL(mtrr_state);
91 
92 /* Reserved bits in the high portion of the MTRRphysBaseN MSR. */
93 u32 phys_hi_rsvd;
94 
95 /*
96  * BIOS is expected to clear MtrrFixDramModEn bit, see for example
97  * "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
98  * Opteron Processors" (26094 Rev. 3.30 February 2006), section
99  * "13.2.1.2 SYSCFG Register": "The MtrrFixDramModEn bit should be set
100  * to 1 during BIOS initialization of the fixed MTRRs, then cleared to
101  * 0 for operation."
102  */
103 static inline void k8_check_syscfg_dram_mod_en(void)
104 {
105 	u32 lo, hi;
106 
107 	if (!((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) &&
108 	      (boot_cpu_data.x86 >= 0x0f)))
109 		return;
110 
111 	rdmsr(MSR_AMD64_SYSCFG, lo, hi);
112 	if (lo & K8_MTRRFIXRANGE_DRAM_MODIFY) {
113 		pr_err(FW_WARN "MTRR: CPU %u: SYSCFG[MtrrFixDramModEn]"
114 		       " not cleared by BIOS, clearing this bit\n",
115 		       smp_processor_id());
116 		lo &= ~K8_MTRRFIXRANGE_DRAM_MODIFY;
117 		mtrr_wrmsr(MSR_AMD64_SYSCFG, lo, hi);
118 	}
119 }
120 
121 /* Get the size of contiguous MTRR range */
122 static u64 get_mtrr_size(u64 mask)
123 {
124 	u64 size;
125 
126 	mask |= (u64)phys_hi_rsvd << 32;
127 	size = -mask;
128 
129 	return size;
130 }
131 
132 static u8 get_var_mtrr_state(unsigned int reg, u64 *start, u64 *size)
133 {
134 	struct mtrr_var_range *mtrr = mtrr_state.var_ranges + reg;
135 
136 	if (!(mtrr->mask_lo & MTRR_PHYSMASK_V))
137 		return MTRR_TYPE_INVALID;
138 
139 	*start = (((u64)mtrr->base_hi) << 32) + (mtrr->base_lo & PAGE_MASK);
140 	*size = get_mtrr_size((((u64)mtrr->mask_hi) << 32) +
141 			      (mtrr->mask_lo & PAGE_MASK));
142 
143 	return mtrr->base_lo & MTRR_PHYSBASE_TYPE;
144 }
145 
146 static u8 get_effective_type(u8 type1, u8 type2)
147 {
148 	if (type1 == MTRR_TYPE_UNCACHABLE || type2 == MTRR_TYPE_UNCACHABLE)
149 		return MTRR_TYPE_UNCACHABLE;
150 
151 	if ((type1 == MTRR_TYPE_WRBACK && type2 == MTRR_TYPE_WRTHROUGH) ||
152 	    (type1 == MTRR_TYPE_WRTHROUGH && type2 == MTRR_TYPE_WRBACK))
153 		return MTRR_TYPE_WRTHROUGH;
154 
155 	if (type1 != type2)
156 		return MTRR_TYPE_UNCACHABLE;
157 
158 	return type1;
159 }
160 
161 static void rm_map_entry_at(int idx)
162 {
163 	cache_map_n--;
164 	if (cache_map_n > idx) {
165 		memmove(cache_map + idx, cache_map + idx + 1,
166 			sizeof(*cache_map) * (cache_map_n - idx));
167 	}
168 }
169 
170 /*
171  * Add an entry into cache_map at a specific index.  Merges adjacent entries if
172  * appropriate.  Return the number of merges for correcting the scan index
173  * (this is needed as merging will reduce the number of entries, which will
174  * result in skipping entries in future iterations if the scan index isn't
175  * corrected).
176  * Note that the corrected index can never go below -1 (resulting in being 0 in
177  * the next scan iteration), as "2" is returned only if the current index is
178  * larger than zero.
179  */
180 static int add_map_entry_at(u64 start, u64 end, u8 type, int idx)
181 {
182 	bool merge_prev = false, merge_next = false;
183 
184 	if (start >= end)
185 		return 0;
186 
187 	if (idx > 0) {
188 		struct cache_map *prev = cache_map + idx - 1;
189 
190 		if (!prev->fixed && start == prev->end && type == prev->type)
191 			merge_prev = true;
192 	}
193 
194 	if (idx < cache_map_n) {
195 		struct cache_map *next = cache_map + idx;
196 
197 		if (!next->fixed && end == next->start && type == next->type)
198 			merge_next = true;
199 	}
200 
201 	if (merge_prev && merge_next) {
202 		cache_map[idx - 1].end = cache_map[idx].end;
203 		rm_map_entry_at(idx);
204 		return 2;
205 	}
206 	if (merge_prev) {
207 		cache_map[idx - 1].end = end;
208 		return 1;
209 	}
210 	if (merge_next) {
211 		cache_map[idx].start = start;
212 		return 1;
213 	}
214 
215 	/* Sanity check: the array should NEVER be too small! */
216 	if (cache_map_n == cache_map_size) {
217 		WARN(1, "MTRR cache mode memory map exhausted!\n");
218 		cache_map_n = cache_map_fixed;
219 		return 0;
220 	}
221 
222 	if (cache_map_n > idx) {
223 		memmove(cache_map + idx + 1, cache_map + idx,
224 			sizeof(*cache_map) * (cache_map_n - idx));
225 	}
226 
227 	cache_map[idx].start = start;
228 	cache_map[idx].end = end;
229 	cache_map[idx].type = type;
230 	cache_map[idx].fixed = 0;
231 	cache_map_n++;
232 
233 	return 0;
234 }
235 
236 /* Clear a part of an entry. Return 1 if start of entry is still valid. */
237 static int clr_map_range_at(u64 start, u64 end, int idx)
238 {
239 	int ret = start != cache_map[idx].start;
240 	u64 tmp;
241 
242 	if (start == cache_map[idx].start && end == cache_map[idx].end) {
243 		rm_map_entry_at(idx);
244 	} else if (start == cache_map[idx].start) {
245 		cache_map[idx].start = end;
246 	} else if (end == cache_map[idx].end) {
247 		cache_map[idx].end = start;
248 	} else {
249 		tmp = cache_map[idx].end;
250 		cache_map[idx].end = start;
251 		add_map_entry_at(end, tmp, cache_map[idx].type, idx + 1);
252 	}
253 
254 	return ret;
255 }
256 
257 /*
258  * Add MTRR to the map.  The current map is scanned and each part of the MTRR
259  * either overlapping with an existing entry or with a hole in the map is
260  * handled separately.
261  */
262 static void add_map_entry(u64 start, u64 end, u8 type)
263 {
264 	u8 new_type, old_type;
265 	u64 tmp;
266 	int i;
267 
268 	for (i = 0; i < cache_map_n && start < end; i++) {
269 		if (start >= cache_map[i].end)
270 			continue;
271 
272 		if (start < cache_map[i].start) {
273 			/* Region start has no overlap. */
274 			tmp = min(end, cache_map[i].start);
275 			i -= add_map_entry_at(start, tmp,  type, i);
276 			start = tmp;
277 			continue;
278 		}
279 
280 		new_type = get_effective_type(type, cache_map[i].type);
281 		old_type = cache_map[i].type;
282 
283 		if (cache_map[i].fixed || new_type == old_type) {
284 			/* Cut off start of new entry. */
285 			start = cache_map[i].end;
286 			continue;
287 		}
288 
289 		/* Handle only overlapping part of region. */
290 		tmp = min(end, cache_map[i].end);
291 		i += clr_map_range_at(start, tmp, i);
292 		i -= add_map_entry_at(start, tmp, new_type, i);
293 		start = tmp;
294 	}
295 
296 	/* Add rest of region after last map entry (rest might be empty). */
297 	add_map_entry_at(start, end, type, i);
298 }
299 
300 /* Add variable MTRRs to cache map. */
301 static void map_add_var(void)
302 {
303 	u64 start, size;
304 	unsigned int i;
305 	u8 type;
306 
307 	/*
308 	 * Add AMD TOP_MEM2 area.  Can't be added in mtrr_build_map(), as it
309 	 * needs to be added again when rebuilding the map due to potentially
310 	 * having moved as a result of variable MTRRs for memory below 4GB.
311 	 */
312 	if (mtrr_tom2) {
313 		add_map_entry(BIT_ULL(32), mtrr_tom2, MTRR_TYPE_WRBACK);
314 		cache_map[cache_map_n - 1].fixed = 1;
315 	}
316 
317 	for (i = 0; i < num_var_ranges; i++) {
318 		type = get_var_mtrr_state(i, &start, &size);
319 		if (type != MTRR_TYPE_INVALID)
320 			add_map_entry(start, start + size, type);
321 	}
322 }
323 
324 /*
325  * Rebuild map by replacing variable entries.  Needs to be called when MTRR
326  * registers are being changed after boot, as such changes could include
327  * removals of registers, which are complicated to handle without rebuild of
328  * the map.
329  */
330 void generic_rebuild_map(void)
331 {
332 	if (mtrr_if != &generic_mtrr_ops)
333 		return;
334 
335 	cache_map_n = cache_map_fixed;
336 
337 	map_add_var();
338 }
339 
340 static unsigned int __init get_cache_map_size(void)
341 {
342 	return cache_map_fixed + 2 * num_var_ranges + (mtrr_tom2 != 0);
343 }
344 
345 /* Build the cache_map containing the cache modes per memory range. */
346 void __init mtrr_build_map(void)
347 {
348 	u64 start, end, size;
349 	unsigned int i;
350 	u8 type;
351 
352 	/* Add fixed MTRRs, optimize for adjacent entries with same type. */
353 	if (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED) {
354 		/*
355 		 * Start with 64k size fixed entries, preset 1st one (hence the
356 		 * loop below is starting with index 1).
357 		 */
358 		start = 0;
359 		end = size = 0x10000;
360 		type = mtrr_state.fixed_ranges[0];
361 
362 		for (i = 1; i < MTRR_NUM_FIXED_RANGES; i++) {
363 			/* 8 64k entries, then 16 16k ones, rest 4k. */
364 			if (i == 8 || i == 24)
365 				size >>= 2;
366 
367 			if (mtrr_state.fixed_ranges[i] != type) {
368 				add_map_entry(start, end, type);
369 				start = end;
370 				type = mtrr_state.fixed_ranges[i];
371 			}
372 			end += size;
373 		}
374 		add_map_entry(start, end, type);
375 	}
376 
377 	/* Mark fixed, they take precedence. */
378 	for (i = 0; i < cache_map_n; i++)
379 		cache_map[i].fixed = 1;
380 	cache_map_fixed = cache_map_n;
381 
382 	map_add_var();
383 
384 	pr_info("MTRR map: %u entries (%u fixed + %u variable; max %u), built from %u variable MTRRs\n",
385 		cache_map_n, cache_map_fixed, cache_map_n - cache_map_fixed,
386 		get_cache_map_size(), num_var_ranges + (mtrr_tom2 != 0));
387 
388 	if (mtrr_debug) {
389 		for (i = 0; i < cache_map_n; i++) {
390 			pr_info("%3u: %016llx-%016llx %s\n", i,
391 				cache_map[i].start, cache_map[i].end - 1,
392 				mtrr_attrib_to_str(cache_map[i].type));
393 		}
394 	}
395 }
396 
397 /* Copy the cache_map from __initdata memory to dynamically allocated one. */
398 void __init mtrr_copy_map(void)
399 {
400 	unsigned int new_size = get_cache_map_size();
401 
402 	if (!mtrr_state.enabled || !new_size) {
403 		cache_map = NULL;
404 		return;
405 	}
406 
407 	mutex_lock(&mtrr_mutex);
408 
409 	cache_map = kcalloc(new_size, sizeof(*cache_map), GFP_KERNEL);
410 	if (cache_map) {
411 		memmove(cache_map, init_cache_map,
412 			cache_map_n * sizeof(*cache_map));
413 		cache_map_size = new_size;
414 	} else {
415 		mtrr_state.enabled = 0;
416 		pr_err("MTRRs disabled due to allocation failure for lookup map.\n");
417 	}
418 
419 	mutex_unlock(&mtrr_mutex);
420 }
421 
422 /**
423  * mtrr_overwrite_state - set static MTRR state
424  *
425  * Used to set MTRR state via different means (e.g. with data obtained from
426  * a hypervisor).
427  * Is allowed only for special cases when running virtualized. Must be called
428  * from the x86_init.hyper.init_platform() hook.  It can be called only once.
429  * The MTRR state can't be changed afterwards.  To ensure that, X86_FEATURE_MTRR
430  * is cleared.
431  */
432 void mtrr_overwrite_state(struct mtrr_var_range *var, unsigned int num_var,
433 			  mtrr_type def_type)
434 {
435 	unsigned int i;
436 
437 	/* Only allowed to be called once before mtrr_bp_init(). */
438 	if (WARN_ON_ONCE(mtrr_state_set))
439 		return;
440 
441 	/* Only allowed when running virtualized. */
442 	if (!cpu_feature_enabled(X86_FEATURE_HYPERVISOR))
443 		return;
444 
445 	/*
446 	 * Only allowed for special virtualization cases:
447 	 * - when running as Hyper-V, SEV-SNP guest using vTOM
448 	 * - when running as Xen PV guest
449 	 * - when running as SEV-SNP or TDX guest to avoid unnecessary
450 	 *   VMM communication/Virtualization exceptions (#VC, #VE)
451 	 */
452 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) &&
453 	    !hv_is_isolation_supported() &&
454 	    !cpu_feature_enabled(X86_FEATURE_XENPV) &&
455 	    !cpu_feature_enabled(X86_FEATURE_TDX_GUEST))
456 		return;
457 
458 	/* Disable MTRR in order to disable MTRR modifications. */
459 	setup_clear_cpu_cap(X86_FEATURE_MTRR);
460 
461 	if (var) {
462 		if (num_var > MTRR_MAX_VAR_RANGES) {
463 			pr_warn("Trying to overwrite MTRR state with %u variable entries\n",
464 				num_var);
465 			num_var = MTRR_MAX_VAR_RANGES;
466 		}
467 		for (i = 0; i < num_var; i++)
468 			mtrr_state.var_ranges[i] = var[i];
469 		num_var_ranges = num_var;
470 	}
471 
472 	mtrr_state.def_type = def_type;
473 	mtrr_state.enabled |= MTRR_STATE_MTRR_ENABLED;
474 
475 	mtrr_state_set = 1;
476 }
477 
478 static u8 type_merge(u8 type, u8 new_type, u8 *uniform)
479 {
480 	u8 effective_type;
481 
482 	if (type == MTRR_TYPE_INVALID)
483 		return new_type;
484 
485 	effective_type = get_effective_type(type, new_type);
486 	if (type != effective_type)
487 		*uniform = 0;
488 
489 	return effective_type;
490 }
491 
492 /**
493  * mtrr_type_lookup - look up memory type in MTRR
494  *
495  * Return Values:
496  * MTRR_TYPE_(type)  - The effective MTRR type for the region
497  * MTRR_TYPE_INVALID - MTRR is disabled
498  *
499  * Output Argument:
500  * uniform - Set to 1 when the returned MTRR type is valid for the whole
501  *	     region, set to 0 else.
502  */
503 u8 mtrr_type_lookup(u64 start, u64 end, u8 *uniform)
504 {
505 	u8 type = MTRR_TYPE_INVALID;
506 	unsigned int i;
507 
508 	if (!mtrr_state_set) {
509 		/* Uniformity is unknown. */
510 		*uniform = 0;
511 		return MTRR_TYPE_UNCACHABLE;
512 	}
513 
514 	*uniform = 1;
515 
516 	if (!(mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED))
517 		return MTRR_TYPE_UNCACHABLE;
518 
519 	for (i = 0; i < cache_map_n && start < end; i++) {
520 		/* Region after current map entry? -> continue with next one. */
521 		if (start >= cache_map[i].end)
522 			continue;
523 
524 		/* Start of region not covered by current map entry? */
525 		if (start < cache_map[i].start) {
526 			/* At least some part of region has default type. */
527 			type = type_merge(type, mtrr_state.def_type, uniform);
528 			/* End of region not covered, too? -> lookup done. */
529 			if (end <= cache_map[i].start)
530 				return type;
531 		}
532 
533 		/* At least part of region covered by map entry. */
534 		type = type_merge(type, cache_map[i].type, uniform);
535 
536 		start = cache_map[i].end;
537 	}
538 
539 	/* End of region past last entry in map? -> use default type. */
540 	if (start < end)
541 		type = type_merge(type, mtrr_state.def_type, uniform);
542 
543 	return type;
544 }
545 
546 /* Get the MSR pair relating to a var range */
547 static void
548 get_mtrr_var_range(unsigned int index, struct mtrr_var_range *vr)
549 {
550 	rdmsr(MTRRphysBase_MSR(index), vr->base_lo, vr->base_hi);
551 	rdmsr(MTRRphysMask_MSR(index), vr->mask_lo, vr->mask_hi);
552 }
553 
554 /* Fill the MSR pair relating to a var range */
555 void fill_mtrr_var_range(unsigned int index,
556 		u32 base_lo, u32 base_hi, u32 mask_lo, u32 mask_hi)
557 {
558 	struct mtrr_var_range *vr;
559 
560 	vr = mtrr_state.var_ranges;
561 
562 	vr[index].base_lo = base_lo;
563 	vr[index].base_hi = base_hi;
564 	vr[index].mask_lo = mask_lo;
565 	vr[index].mask_hi = mask_hi;
566 }
567 
568 static void get_fixed_ranges(mtrr_type *frs)
569 {
570 	unsigned int *p = (unsigned int *)frs;
571 	int i;
572 
573 	k8_check_syscfg_dram_mod_en();
574 
575 	rdmsr(MSR_MTRRfix64K_00000, p[0], p[1]);
576 
577 	for (i = 0; i < 2; i++)
578 		rdmsr(MSR_MTRRfix16K_80000 + i, p[2 + i * 2], p[3 + i * 2]);
579 	for (i = 0; i < 8; i++)
580 		rdmsr(MSR_MTRRfix4K_C0000 + i, p[6 + i * 2], p[7 + i * 2]);
581 }
582 
583 void mtrr_save_fixed_ranges(void *info)
584 {
585 	if (boot_cpu_has(X86_FEATURE_MTRR))
586 		get_fixed_ranges(mtrr_state.fixed_ranges);
587 }
588 
589 static unsigned __initdata last_fixed_start;
590 static unsigned __initdata last_fixed_end;
591 static mtrr_type __initdata last_fixed_type;
592 
593 static void __init print_fixed_last(void)
594 {
595 	if (!last_fixed_end)
596 		return;
597 
598 	pr_info("  %05X-%05X %s\n", last_fixed_start,
599 		last_fixed_end - 1, mtrr_attrib_to_str(last_fixed_type));
600 
601 	last_fixed_end = 0;
602 }
603 
604 static void __init update_fixed_last(unsigned base, unsigned end,
605 				     mtrr_type type)
606 {
607 	last_fixed_start = base;
608 	last_fixed_end = end;
609 	last_fixed_type = type;
610 }
611 
612 static void __init
613 print_fixed(unsigned base, unsigned step, const mtrr_type *types)
614 {
615 	unsigned i;
616 
617 	for (i = 0; i < 8; ++i, ++types, base += step) {
618 		if (last_fixed_end == 0) {
619 			update_fixed_last(base, base + step, *types);
620 			continue;
621 		}
622 		if (last_fixed_end == base && last_fixed_type == *types) {
623 			last_fixed_end = base + step;
624 			continue;
625 		}
626 		/* new segments: gap or different type */
627 		print_fixed_last();
628 		update_fixed_last(base, base + step, *types);
629 	}
630 }
631 
632 static void __init print_mtrr_state(void)
633 {
634 	unsigned int i;
635 	int high_width;
636 
637 	pr_info("MTRR default type: %s\n",
638 		mtrr_attrib_to_str(mtrr_state.def_type));
639 	if (mtrr_state.have_fixed) {
640 		pr_info("MTRR fixed ranges %sabled:\n",
641 			((mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED) &&
642 			 (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED)) ?
643 			 "en" : "dis");
644 		print_fixed(0x00000, 0x10000, mtrr_state.fixed_ranges + 0);
645 		for (i = 0; i < 2; ++i)
646 			print_fixed(0x80000 + i * 0x20000, 0x04000,
647 				    mtrr_state.fixed_ranges + (i + 1) * 8);
648 		for (i = 0; i < 8; ++i)
649 			print_fixed(0xC0000 + i * 0x08000, 0x01000,
650 				    mtrr_state.fixed_ranges + (i + 3) * 8);
651 
652 		/* tail */
653 		print_fixed_last();
654 	}
655 	pr_info("MTRR variable ranges %sabled:\n",
656 		mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED ? "en" : "dis");
657 	high_width = (boot_cpu_data.x86_phys_bits - (32 - PAGE_SHIFT) + 3) / 4;
658 
659 	for (i = 0; i < num_var_ranges; ++i) {
660 		if (mtrr_state.var_ranges[i].mask_lo & MTRR_PHYSMASK_V)
661 			pr_info("  %u base %0*X%05X000 mask %0*X%05X000 %s\n",
662 				i,
663 				high_width,
664 				mtrr_state.var_ranges[i].base_hi,
665 				mtrr_state.var_ranges[i].base_lo >> 12,
666 				high_width,
667 				mtrr_state.var_ranges[i].mask_hi,
668 				mtrr_state.var_ranges[i].mask_lo >> 12,
669 				mtrr_attrib_to_str(mtrr_state.var_ranges[i].base_lo &
670 						    MTRR_PHYSBASE_TYPE));
671 		else
672 			pr_info("  %u disabled\n", i);
673 	}
674 	if (mtrr_tom2)
675 		pr_info("TOM2: %016llx aka %lldM\n", mtrr_tom2, mtrr_tom2>>20);
676 }
677 
678 /* Grab all of the MTRR state for this CPU into *state */
679 bool __init get_mtrr_state(void)
680 {
681 	struct mtrr_var_range *vrs;
682 	unsigned lo, dummy;
683 	unsigned int i;
684 
685 	vrs = mtrr_state.var_ranges;
686 
687 	rdmsr(MSR_MTRRcap, lo, dummy);
688 	mtrr_state.have_fixed = lo & MTRR_CAP_FIX;
689 
690 	for (i = 0; i < num_var_ranges; i++)
691 		get_mtrr_var_range(i, &vrs[i]);
692 	if (mtrr_state.have_fixed)
693 		get_fixed_ranges(mtrr_state.fixed_ranges);
694 
695 	rdmsr(MSR_MTRRdefType, lo, dummy);
696 	mtrr_state.def_type = lo & MTRR_DEF_TYPE_TYPE;
697 	mtrr_state.enabled = (lo & MTRR_DEF_TYPE_ENABLE) >> MTRR_STATE_SHIFT;
698 
699 	if (amd_special_default_mtrr()) {
700 		unsigned low, high;
701 
702 		/* TOP_MEM2 */
703 		rdmsr(MSR_K8_TOP_MEM2, low, high);
704 		mtrr_tom2 = high;
705 		mtrr_tom2 <<= 32;
706 		mtrr_tom2 |= low;
707 		mtrr_tom2 &= 0xffffff800000ULL;
708 	}
709 
710 	if (mtrr_debug)
711 		print_mtrr_state();
712 
713 	mtrr_state_set = 1;
714 
715 	return !!(mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED);
716 }
717 
718 /* Some BIOS's are messed up and don't set all MTRRs the same! */
719 void __init mtrr_state_warn(void)
720 {
721 	unsigned long mask = smp_changes_mask;
722 
723 	if (!mask)
724 		return;
725 	if (mask & MTRR_CHANGE_MASK_FIXED)
726 		pr_warn("mtrr: your CPUs had inconsistent fixed MTRR settings\n");
727 	if (mask & MTRR_CHANGE_MASK_VARIABLE)
728 		pr_warn("mtrr: your CPUs had inconsistent variable MTRR settings\n");
729 	if (mask & MTRR_CHANGE_MASK_DEFTYPE)
730 		pr_warn("mtrr: your CPUs had inconsistent MTRRdefType settings\n");
731 
732 	pr_info("mtrr: probably your BIOS does not setup all CPUs.\n");
733 	pr_info("mtrr: corrected configuration.\n");
734 }
735 
736 /*
737  * Doesn't attempt to pass an error out to MTRR users
738  * because it's quite complicated in some cases and probably not
739  * worth it because the best error handling is to ignore it.
740  */
741 void mtrr_wrmsr(unsigned msr, unsigned a, unsigned b)
742 {
743 	if (wrmsr_safe(msr, a, b) < 0) {
744 		pr_err("MTRR: CPU %u: Writing MSR %x to %x:%x failed\n",
745 			smp_processor_id(), msr, a, b);
746 	}
747 }
748 
749 /**
750  * set_fixed_range - checks & updates a fixed-range MTRR if it
751  *		     differs from the value it should have
752  * @msr: MSR address of the MTTR which should be checked and updated
753  * @changed: pointer which indicates whether the MTRR needed to be changed
754  * @msrwords: pointer to the MSR values which the MSR should have
755  */
756 static void set_fixed_range(int msr, bool *changed, unsigned int *msrwords)
757 {
758 	unsigned lo, hi;
759 
760 	rdmsr(msr, lo, hi);
761 
762 	if (lo != msrwords[0] || hi != msrwords[1]) {
763 		mtrr_wrmsr(msr, msrwords[0], msrwords[1]);
764 		*changed = true;
765 	}
766 }
767 
768 /**
769  * generic_get_free_region - Get a free MTRR.
770  * @base: The starting (base) address of the region.
771  * @size: The size (in bytes) of the region.
772  * @replace_reg: mtrr index to be replaced; set to invalid value if none.
773  *
774  * Returns: The index of the region on success, else negative on error.
775  */
776 int
777 generic_get_free_region(unsigned long base, unsigned long size, int replace_reg)
778 {
779 	unsigned long lbase, lsize;
780 	mtrr_type ltype;
781 	int i, max;
782 
783 	max = num_var_ranges;
784 	if (replace_reg >= 0 && replace_reg < max)
785 		return replace_reg;
786 
787 	for (i = 0; i < max; ++i) {
788 		mtrr_if->get(i, &lbase, &lsize, &ltype);
789 		if (lsize == 0)
790 			return i;
791 	}
792 
793 	return -ENOSPC;
794 }
795 
796 static void generic_get_mtrr(unsigned int reg, unsigned long *base,
797 			     unsigned long *size, mtrr_type *type)
798 {
799 	u32 mask_lo, mask_hi, base_lo, base_hi;
800 	unsigned int hi;
801 	u64 tmp, mask;
802 
803 	/*
804 	 * get_mtrr doesn't need to update mtrr_state, also it could be called
805 	 * from any cpu, so try to print it out directly.
806 	 */
807 	get_cpu();
808 
809 	rdmsr(MTRRphysMask_MSR(reg), mask_lo, mask_hi);
810 
811 	if (!(mask_lo & MTRR_PHYSMASK_V)) {
812 		/*  Invalid (i.e. free) range */
813 		*base = 0;
814 		*size = 0;
815 		*type = 0;
816 		goto out_put_cpu;
817 	}
818 
819 	rdmsr(MTRRphysBase_MSR(reg), base_lo, base_hi);
820 
821 	/* Work out the shifted address mask: */
822 	tmp = (u64)mask_hi << 32 | (mask_lo & PAGE_MASK);
823 	mask = (u64)phys_hi_rsvd << 32 | tmp;
824 
825 	/* Expand tmp with high bits to all 1s: */
826 	hi = fls64(tmp);
827 	if (hi > 0) {
828 		tmp |= ~((1ULL<<(hi - 1)) - 1);
829 
830 		if (tmp != mask) {
831 			pr_warn("mtrr: your BIOS has configured an incorrect mask, fixing it.\n");
832 			add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
833 			mask = tmp;
834 		}
835 	}
836 
837 	/*
838 	 * This works correctly if size is a power of two, i.e. a
839 	 * contiguous range:
840 	 */
841 	*size = -mask >> PAGE_SHIFT;
842 	*base = (u64)base_hi << (32 - PAGE_SHIFT) | base_lo >> PAGE_SHIFT;
843 	*type = base_lo & MTRR_PHYSBASE_TYPE;
844 
845 out_put_cpu:
846 	put_cpu();
847 }
848 
849 /**
850  * set_fixed_ranges - checks & updates the fixed-range MTRRs if they
851  *		      differ from the saved set
852  * @frs: pointer to fixed-range MTRR values, saved by get_fixed_ranges()
853  */
854 static int set_fixed_ranges(mtrr_type *frs)
855 {
856 	unsigned long long *saved = (unsigned long long *)frs;
857 	bool changed = false;
858 	int block = -1, range;
859 
860 	k8_check_syscfg_dram_mod_en();
861 
862 	while (fixed_range_blocks[++block].ranges) {
863 		for (range = 0; range < fixed_range_blocks[block].ranges; range++)
864 			set_fixed_range(fixed_range_blocks[block].base_msr + range,
865 					&changed, (unsigned int *)saved++);
866 	}
867 
868 	return changed;
869 }
870 
871 /*
872  * Set the MSR pair relating to a var range.
873  * Returns true if changes are made.
874  */
875 static bool set_mtrr_var_ranges(unsigned int index, struct mtrr_var_range *vr)
876 {
877 	unsigned int lo, hi;
878 	bool changed = false;
879 
880 	rdmsr(MTRRphysBase_MSR(index), lo, hi);
881 	if ((vr->base_lo & ~MTRR_PHYSBASE_RSVD) != (lo & ~MTRR_PHYSBASE_RSVD)
882 	    || (vr->base_hi & ~phys_hi_rsvd) != (hi & ~phys_hi_rsvd)) {
883 
884 		mtrr_wrmsr(MTRRphysBase_MSR(index), vr->base_lo, vr->base_hi);
885 		changed = true;
886 	}
887 
888 	rdmsr(MTRRphysMask_MSR(index), lo, hi);
889 
890 	if ((vr->mask_lo & ~MTRR_PHYSMASK_RSVD) != (lo & ~MTRR_PHYSMASK_RSVD)
891 	    || (vr->mask_hi & ~phys_hi_rsvd) != (hi & ~phys_hi_rsvd)) {
892 		mtrr_wrmsr(MTRRphysMask_MSR(index), vr->mask_lo, vr->mask_hi);
893 		changed = true;
894 	}
895 	return changed;
896 }
897 
898 static u32 deftype_lo, deftype_hi;
899 
900 /**
901  * set_mtrr_state - Set the MTRR state for this CPU.
902  *
903  * NOTE: The CPU must already be in a safe state for MTRR changes, including
904  *       measures that only a single CPU can be active in set_mtrr_state() in
905  *       order to not be subject to races for usage of deftype_lo. This is
906  *       accomplished by taking cache_disable_lock.
907  * RETURNS: 0 if no changes made, else a mask indicating what was changed.
908  */
909 static unsigned long set_mtrr_state(void)
910 {
911 	unsigned long change_mask = 0;
912 	unsigned int i;
913 
914 	for (i = 0; i < num_var_ranges; i++) {
915 		if (set_mtrr_var_ranges(i, &mtrr_state.var_ranges[i]))
916 			change_mask |= MTRR_CHANGE_MASK_VARIABLE;
917 	}
918 
919 	if (mtrr_state.have_fixed && set_fixed_ranges(mtrr_state.fixed_ranges))
920 		change_mask |= MTRR_CHANGE_MASK_FIXED;
921 
922 	/*
923 	 * Set_mtrr_restore restores the old value of MTRRdefType,
924 	 * so to set it we fiddle with the saved value:
925 	 */
926 	if ((deftype_lo & MTRR_DEF_TYPE_TYPE) != mtrr_state.def_type ||
927 	    ((deftype_lo & MTRR_DEF_TYPE_ENABLE) >> MTRR_STATE_SHIFT) != mtrr_state.enabled) {
928 
929 		deftype_lo = (deftype_lo & MTRR_DEF_TYPE_DISABLE) |
930 			     mtrr_state.def_type |
931 			     (mtrr_state.enabled << MTRR_STATE_SHIFT);
932 		change_mask |= MTRR_CHANGE_MASK_DEFTYPE;
933 	}
934 
935 	return change_mask;
936 }
937 
938 void mtrr_disable(void)
939 {
940 	/* Save MTRR state */
941 	rdmsr(MSR_MTRRdefType, deftype_lo, deftype_hi);
942 
943 	/* Disable MTRRs, and set the default type to uncached */
944 	mtrr_wrmsr(MSR_MTRRdefType, deftype_lo & MTRR_DEF_TYPE_DISABLE, deftype_hi);
945 }
946 
947 void mtrr_enable(void)
948 {
949 	/* Intel (P6) standard MTRRs */
950 	mtrr_wrmsr(MSR_MTRRdefType, deftype_lo, deftype_hi);
951 }
952 
953 void mtrr_generic_set_state(void)
954 {
955 	unsigned long mask, count;
956 
957 	/* Actually set the state */
958 	mask = set_mtrr_state();
959 
960 	/* Use the atomic bitops to update the global mask */
961 	for (count = 0; count < sizeof(mask) * 8; ++count) {
962 		if (mask & 0x01)
963 			set_bit(count, &smp_changes_mask);
964 		mask >>= 1;
965 	}
966 }
967 
968 /**
969  * generic_set_mtrr - set variable MTRR register on the local CPU.
970  *
971  * @reg: The register to set.
972  * @base: The base address of the region.
973  * @size: The size of the region. If this is 0 the region is disabled.
974  * @type: The type of the region.
975  *
976  * Returns nothing.
977  */
978 static void generic_set_mtrr(unsigned int reg, unsigned long base,
979 			     unsigned long size, mtrr_type type)
980 {
981 	unsigned long flags;
982 	struct mtrr_var_range *vr;
983 
984 	vr = &mtrr_state.var_ranges[reg];
985 
986 	local_irq_save(flags);
987 	cache_disable();
988 
989 	if (size == 0) {
990 		/*
991 		 * The invalid bit is kept in the mask, so we simply
992 		 * clear the relevant mask register to disable a range.
993 		 */
994 		mtrr_wrmsr(MTRRphysMask_MSR(reg), 0, 0);
995 		memset(vr, 0, sizeof(struct mtrr_var_range));
996 	} else {
997 		vr->base_lo = base << PAGE_SHIFT | type;
998 		vr->base_hi = (base >> (32 - PAGE_SHIFT)) & ~phys_hi_rsvd;
999 		vr->mask_lo = -size << PAGE_SHIFT | MTRR_PHYSMASK_V;
1000 		vr->mask_hi = (-size >> (32 - PAGE_SHIFT)) & ~phys_hi_rsvd;
1001 
1002 		mtrr_wrmsr(MTRRphysBase_MSR(reg), vr->base_lo, vr->base_hi);
1003 		mtrr_wrmsr(MTRRphysMask_MSR(reg), vr->mask_lo, vr->mask_hi);
1004 	}
1005 
1006 	cache_enable();
1007 	local_irq_restore(flags);
1008 }
1009 
1010 int generic_validate_add_page(unsigned long base, unsigned long size,
1011 			      unsigned int type)
1012 {
1013 	unsigned long lbase, last;
1014 
1015 	/*
1016 	 * For Intel PPro stepping <= 7
1017 	 * must be 4 MiB aligned and not touch 0x70000000 -> 0x7003FFFF
1018 	 */
1019 	if (mtrr_if == &generic_mtrr_ops && boot_cpu_data.x86 == 6 &&
1020 	    boot_cpu_data.x86_model == 1 &&
1021 	    boot_cpu_data.x86_stepping <= 7) {
1022 		if (base & ((1 << (22 - PAGE_SHIFT)) - 1)) {
1023 			pr_warn("mtrr: base(0x%lx000) is not 4 MiB aligned\n", base);
1024 			return -EINVAL;
1025 		}
1026 		if (!(base + size < 0x70000 || base > 0x7003F) &&
1027 		    (type == MTRR_TYPE_WRCOMB
1028 		     || type == MTRR_TYPE_WRBACK)) {
1029 			pr_warn("mtrr: writable mtrr between 0x70000000 and 0x7003FFFF may hang the CPU.\n");
1030 			return -EINVAL;
1031 		}
1032 	}
1033 
1034 	/*
1035 	 * Check upper bits of base and last are equal and lower bits are 0
1036 	 * for base and 1 for last
1037 	 */
1038 	last = base + size - 1;
1039 	for (lbase = base; !(lbase & 1) && (last & 1);
1040 	     lbase = lbase >> 1, last = last >> 1)
1041 		;
1042 	if (lbase != last) {
1043 		pr_warn("mtrr: base(0x%lx000) is not aligned on a size(0x%lx000) boundary\n", base, size);
1044 		return -EINVAL;
1045 	}
1046 	return 0;
1047 }
1048 
1049 static int generic_have_wrcomb(void)
1050 {
1051 	unsigned long config, dummy;
1052 	rdmsr(MSR_MTRRcap, config, dummy);
1053 	return config & MTRR_CAP_WC;
1054 }
1055 
1056 int positive_have_wrcomb(void)
1057 {
1058 	return 1;
1059 }
1060 
1061 /*
1062  * Generic structure...
1063  */
1064 const struct mtrr_ops generic_mtrr_ops = {
1065 	.get			= generic_get_mtrr,
1066 	.get_free_region	= generic_get_free_region,
1067 	.set			= generic_set_mtrr,
1068 	.validate_add_page	= generic_validate_add_page,
1069 	.have_wrcomb		= generic_have_wrcomb,
1070 };
1071