xref: /linux/arch/x86/kernel/cpu/mtrr/generic.c (revision 6e7fd890f1d6ac83805409e9c346240de2705584)
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 	if (cc_platform_has(CC_ATTR_HOST_SEV_SNP))
112 		return;
113 
114 	rdmsr(MSR_AMD64_SYSCFG, lo, hi);
115 	if (lo & K8_MTRRFIXRANGE_DRAM_MODIFY) {
116 		pr_err(FW_WARN "MTRR: CPU %u: SYSCFG[MtrrFixDramModEn]"
117 		       " not cleared by BIOS, clearing this bit\n",
118 		       smp_processor_id());
119 		lo &= ~K8_MTRRFIXRANGE_DRAM_MODIFY;
120 		mtrr_wrmsr(MSR_AMD64_SYSCFG, lo, hi);
121 	}
122 }
123 
124 /* Get the size of contiguous MTRR range */
125 static u64 get_mtrr_size(u64 mask)
126 {
127 	u64 size;
128 
129 	mask |= (u64)phys_hi_rsvd << 32;
130 	size = -mask;
131 
132 	return size;
133 }
134 
135 static u8 get_var_mtrr_state(unsigned int reg, u64 *start, u64 *size)
136 {
137 	struct mtrr_var_range *mtrr = mtrr_state.var_ranges + reg;
138 
139 	if (!(mtrr->mask_lo & MTRR_PHYSMASK_V))
140 		return MTRR_TYPE_INVALID;
141 
142 	*start = (((u64)mtrr->base_hi) << 32) + (mtrr->base_lo & PAGE_MASK);
143 	*size = get_mtrr_size((((u64)mtrr->mask_hi) << 32) +
144 			      (mtrr->mask_lo & PAGE_MASK));
145 
146 	return mtrr->base_lo & MTRR_PHYSBASE_TYPE;
147 }
148 
149 static u8 get_effective_type(u8 type1, u8 type2)
150 {
151 	if (type1 == MTRR_TYPE_UNCACHABLE || type2 == MTRR_TYPE_UNCACHABLE)
152 		return MTRR_TYPE_UNCACHABLE;
153 
154 	if ((type1 == MTRR_TYPE_WRBACK && type2 == MTRR_TYPE_WRTHROUGH) ||
155 	    (type1 == MTRR_TYPE_WRTHROUGH && type2 == MTRR_TYPE_WRBACK))
156 		return MTRR_TYPE_WRTHROUGH;
157 
158 	if (type1 != type2)
159 		return MTRR_TYPE_UNCACHABLE;
160 
161 	return type1;
162 }
163 
164 static void rm_map_entry_at(int idx)
165 {
166 	cache_map_n--;
167 	if (cache_map_n > idx) {
168 		memmove(cache_map + idx, cache_map + idx + 1,
169 			sizeof(*cache_map) * (cache_map_n - idx));
170 	}
171 }
172 
173 /*
174  * Add an entry into cache_map at a specific index.  Merges adjacent entries if
175  * appropriate.  Return the number of merges for correcting the scan index
176  * (this is needed as merging will reduce the number of entries, which will
177  * result in skipping entries in future iterations if the scan index isn't
178  * corrected).
179  * Note that the corrected index can never go below -1 (resulting in being 0 in
180  * the next scan iteration), as "2" is returned only if the current index is
181  * larger than zero.
182  */
183 static int add_map_entry_at(u64 start, u64 end, u8 type, int idx)
184 {
185 	bool merge_prev = false, merge_next = false;
186 
187 	if (start >= end)
188 		return 0;
189 
190 	if (idx > 0) {
191 		struct cache_map *prev = cache_map + idx - 1;
192 
193 		if (!prev->fixed && start == prev->end && type == prev->type)
194 			merge_prev = true;
195 	}
196 
197 	if (idx < cache_map_n) {
198 		struct cache_map *next = cache_map + idx;
199 
200 		if (!next->fixed && end == next->start && type == next->type)
201 			merge_next = true;
202 	}
203 
204 	if (merge_prev && merge_next) {
205 		cache_map[idx - 1].end = cache_map[idx].end;
206 		rm_map_entry_at(idx);
207 		return 2;
208 	}
209 	if (merge_prev) {
210 		cache_map[idx - 1].end = end;
211 		return 1;
212 	}
213 	if (merge_next) {
214 		cache_map[idx].start = start;
215 		return 1;
216 	}
217 
218 	/* Sanity check: the array should NEVER be too small! */
219 	if (cache_map_n == cache_map_size) {
220 		WARN(1, "MTRR cache mode memory map exhausted!\n");
221 		cache_map_n = cache_map_fixed;
222 		return 0;
223 	}
224 
225 	if (cache_map_n > idx) {
226 		memmove(cache_map + idx + 1, cache_map + idx,
227 			sizeof(*cache_map) * (cache_map_n - idx));
228 	}
229 
230 	cache_map[idx].start = start;
231 	cache_map[idx].end = end;
232 	cache_map[idx].type = type;
233 	cache_map[idx].fixed = 0;
234 	cache_map_n++;
235 
236 	return 0;
237 }
238 
239 /* Clear a part of an entry. Return 1 if start of entry is still valid. */
240 static int clr_map_range_at(u64 start, u64 end, int idx)
241 {
242 	int ret = start != cache_map[idx].start;
243 	u64 tmp;
244 
245 	if (start == cache_map[idx].start && end == cache_map[idx].end) {
246 		rm_map_entry_at(idx);
247 	} else if (start == cache_map[idx].start) {
248 		cache_map[idx].start = end;
249 	} else if (end == cache_map[idx].end) {
250 		cache_map[idx].end = start;
251 	} else {
252 		tmp = cache_map[idx].end;
253 		cache_map[idx].end = start;
254 		add_map_entry_at(end, tmp, cache_map[idx].type, idx + 1);
255 	}
256 
257 	return ret;
258 }
259 
260 /*
261  * Add MTRR to the map.  The current map is scanned and each part of the MTRR
262  * either overlapping with an existing entry or with a hole in the map is
263  * handled separately.
264  */
265 static void add_map_entry(u64 start, u64 end, u8 type)
266 {
267 	u8 new_type, old_type;
268 	u64 tmp;
269 	int i;
270 
271 	for (i = 0; i < cache_map_n && start < end; i++) {
272 		if (start >= cache_map[i].end)
273 			continue;
274 
275 		if (start < cache_map[i].start) {
276 			/* Region start has no overlap. */
277 			tmp = min(end, cache_map[i].start);
278 			i -= add_map_entry_at(start, tmp,  type, i);
279 			start = tmp;
280 			continue;
281 		}
282 
283 		new_type = get_effective_type(type, cache_map[i].type);
284 		old_type = cache_map[i].type;
285 
286 		if (cache_map[i].fixed || new_type == old_type) {
287 			/* Cut off start of new entry. */
288 			start = cache_map[i].end;
289 			continue;
290 		}
291 
292 		/* Handle only overlapping part of region. */
293 		tmp = min(end, cache_map[i].end);
294 		i += clr_map_range_at(start, tmp, i);
295 		i -= add_map_entry_at(start, tmp, new_type, i);
296 		start = tmp;
297 	}
298 
299 	/* Add rest of region after last map entry (rest might be empty). */
300 	add_map_entry_at(start, end, type, i);
301 }
302 
303 /* Add variable MTRRs to cache map. */
304 static void map_add_var(void)
305 {
306 	u64 start, size;
307 	unsigned int i;
308 	u8 type;
309 
310 	/*
311 	 * Add AMD TOP_MEM2 area.  Can't be added in mtrr_build_map(), as it
312 	 * needs to be added again when rebuilding the map due to potentially
313 	 * having moved as a result of variable MTRRs for memory below 4GB.
314 	 */
315 	if (mtrr_tom2) {
316 		add_map_entry(BIT_ULL(32), mtrr_tom2, MTRR_TYPE_WRBACK);
317 		cache_map[cache_map_n - 1].fixed = 1;
318 	}
319 
320 	for (i = 0; i < num_var_ranges; i++) {
321 		type = get_var_mtrr_state(i, &start, &size);
322 		if (type != MTRR_TYPE_INVALID)
323 			add_map_entry(start, start + size, type);
324 	}
325 }
326 
327 /*
328  * Rebuild map by replacing variable entries.  Needs to be called when MTRR
329  * registers are being changed after boot, as such changes could include
330  * removals of registers, which are complicated to handle without rebuild of
331  * the map.
332  */
333 void generic_rebuild_map(void)
334 {
335 	if (mtrr_if != &generic_mtrr_ops)
336 		return;
337 
338 	cache_map_n = cache_map_fixed;
339 
340 	map_add_var();
341 }
342 
343 static unsigned int __init get_cache_map_size(void)
344 {
345 	return cache_map_fixed + 2 * num_var_ranges + (mtrr_tom2 != 0);
346 }
347 
348 /* Build the cache_map containing the cache modes per memory range. */
349 void __init mtrr_build_map(void)
350 {
351 	u64 start, end, size;
352 	unsigned int i;
353 	u8 type;
354 
355 	/* Add fixed MTRRs, optimize for adjacent entries with same type. */
356 	if (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED) {
357 		/*
358 		 * Start with 64k size fixed entries, preset 1st one (hence the
359 		 * loop below is starting with index 1).
360 		 */
361 		start = 0;
362 		end = size = 0x10000;
363 		type = mtrr_state.fixed_ranges[0];
364 
365 		for (i = 1; i < MTRR_NUM_FIXED_RANGES; i++) {
366 			/* 8 64k entries, then 16 16k ones, rest 4k. */
367 			if (i == 8 || i == 24)
368 				size >>= 2;
369 
370 			if (mtrr_state.fixed_ranges[i] != type) {
371 				add_map_entry(start, end, type);
372 				start = end;
373 				type = mtrr_state.fixed_ranges[i];
374 			}
375 			end += size;
376 		}
377 		add_map_entry(start, end, type);
378 	}
379 
380 	/* Mark fixed, they take precedence. */
381 	for (i = 0; i < cache_map_n; i++)
382 		cache_map[i].fixed = 1;
383 	cache_map_fixed = cache_map_n;
384 
385 	map_add_var();
386 
387 	pr_info("MTRR map: %u entries (%u fixed + %u variable; max %u), built from %u variable MTRRs\n",
388 		cache_map_n, cache_map_fixed, cache_map_n - cache_map_fixed,
389 		get_cache_map_size(), num_var_ranges + (mtrr_tom2 != 0));
390 
391 	if (mtrr_debug) {
392 		for (i = 0; i < cache_map_n; i++) {
393 			pr_info("%3u: %016llx-%016llx %s\n", i,
394 				cache_map[i].start, cache_map[i].end - 1,
395 				mtrr_attrib_to_str(cache_map[i].type));
396 		}
397 	}
398 }
399 
400 /* Copy the cache_map from __initdata memory to dynamically allocated one. */
401 void __init mtrr_copy_map(void)
402 {
403 	unsigned int new_size = get_cache_map_size();
404 
405 	if (!mtrr_state.enabled || !new_size) {
406 		cache_map = NULL;
407 		return;
408 	}
409 
410 	mutex_lock(&mtrr_mutex);
411 
412 	cache_map = kcalloc(new_size, sizeof(*cache_map), GFP_KERNEL);
413 	if (cache_map) {
414 		memmove(cache_map, init_cache_map,
415 			cache_map_n * sizeof(*cache_map));
416 		cache_map_size = new_size;
417 	} else {
418 		mtrr_state.enabled = 0;
419 		pr_err("MTRRs disabled due to allocation failure for lookup map.\n");
420 	}
421 
422 	mutex_unlock(&mtrr_mutex);
423 }
424 
425 /**
426  * mtrr_overwrite_state - set static MTRR state
427  *
428  * Used to set MTRR state via different means (e.g. with data obtained from
429  * a hypervisor).
430  * Is allowed only for special cases when running virtualized. Must be called
431  * from the x86_init.hyper.init_platform() hook.  It can be called only once.
432  * The MTRR state can't be changed afterwards.  To ensure that, X86_FEATURE_MTRR
433  * is cleared.
434  *
435  * @var: MTRR variable range array to use
436  * @num_var: length of the @var array
437  * @def_type: default caching type
438  */
439 void mtrr_overwrite_state(struct mtrr_var_range *var, unsigned int num_var,
440 			  mtrr_type def_type)
441 {
442 	unsigned int i;
443 
444 	/* Only allowed to be called once before mtrr_bp_init(). */
445 	if (WARN_ON_ONCE(mtrr_state_set))
446 		return;
447 
448 	/* Only allowed when running virtualized. */
449 	if (!cpu_feature_enabled(X86_FEATURE_HYPERVISOR))
450 		return;
451 
452 	/*
453 	 * Only allowed for special virtualization cases:
454 	 * - when running as Hyper-V, SEV-SNP guest using vTOM
455 	 * - when running as Xen PV guest
456 	 * - when running as SEV-SNP or TDX guest to avoid unnecessary
457 	 *   VMM communication/Virtualization exceptions (#VC, #VE)
458 	 */
459 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) &&
460 	    !hv_is_isolation_supported() &&
461 	    !cpu_feature_enabled(X86_FEATURE_XENPV) &&
462 	    !cpu_feature_enabled(X86_FEATURE_TDX_GUEST))
463 		return;
464 
465 	/* Disable MTRR in order to disable MTRR modifications. */
466 	setup_clear_cpu_cap(X86_FEATURE_MTRR);
467 
468 	if (var) {
469 		if (num_var > MTRR_MAX_VAR_RANGES) {
470 			pr_warn("Trying to overwrite MTRR state with %u variable entries\n",
471 				num_var);
472 			num_var = MTRR_MAX_VAR_RANGES;
473 		}
474 		for (i = 0; i < num_var; i++)
475 			mtrr_state.var_ranges[i] = var[i];
476 		num_var_ranges = num_var;
477 	}
478 
479 	mtrr_state.def_type = def_type;
480 	mtrr_state.enabled |= MTRR_STATE_MTRR_ENABLED;
481 
482 	mtrr_state_set = 1;
483 }
484 
485 static u8 type_merge(u8 type, u8 new_type, u8 *uniform)
486 {
487 	u8 effective_type;
488 
489 	if (type == MTRR_TYPE_INVALID)
490 		return new_type;
491 
492 	effective_type = get_effective_type(type, new_type);
493 	if (type != effective_type)
494 		*uniform = 0;
495 
496 	return effective_type;
497 }
498 
499 /**
500  * mtrr_type_lookup - look up memory type in MTRR
501  *
502  * @start: Begin of the physical address range
503  * @end: End of the physical address range
504  * @uniform: output argument:
505  *  - 1: the returned MTRR type is valid for the whole region
506  *  - 0: otherwise
507  *
508  * Return Values:
509  * MTRR_TYPE_(type)  - The effective MTRR type for the region
510  * MTRR_TYPE_INVALID - MTRR is disabled
511  */
512 u8 mtrr_type_lookup(u64 start, u64 end, u8 *uniform)
513 {
514 	u8 type = MTRR_TYPE_INVALID;
515 	unsigned int i;
516 
517 	if (!mtrr_state_set) {
518 		/* Uniformity is unknown. */
519 		*uniform = 0;
520 		return MTRR_TYPE_UNCACHABLE;
521 	}
522 
523 	*uniform = 1;
524 
525 	if (!(mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED))
526 		return MTRR_TYPE_UNCACHABLE;
527 
528 	for (i = 0; i < cache_map_n && start < end; i++) {
529 		/* Region after current map entry? -> continue with next one. */
530 		if (start >= cache_map[i].end)
531 			continue;
532 
533 		/* Start of region not covered by current map entry? */
534 		if (start < cache_map[i].start) {
535 			/* At least some part of region has default type. */
536 			type = type_merge(type, mtrr_state.def_type, uniform);
537 			/* End of region not covered, too? -> lookup done. */
538 			if (end <= cache_map[i].start)
539 				return type;
540 		}
541 
542 		/* At least part of region covered by map entry. */
543 		type = type_merge(type, cache_map[i].type, uniform);
544 
545 		start = cache_map[i].end;
546 	}
547 
548 	/* End of region past last entry in map? -> use default type. */
549 	if (start < end)
550 		type = type_merge(type, mtrr_state.def_type, uniform);
551 
552 	return type;
553 }
554 
555 /* Get the MSR pair relating to a var range */
556 static void
557 get_mtrr_var_range(unsigned int index, struct mtrr_var_range *vr)
558 {
559 	rdmsr(MTRRphysBase_MSR(index), vr->base_lo, vr->base_hi);
560 	rdmsr(MTRRphysMask_MSR(index), vr->mask_lo, vr->mask_hi);
561 }
562 
563 /* Fill the MSR pair relating to a var range */
564 void fill_mtrr_var_range(unsigned int index,
565 		u32 base_lo, u32 base_hi, u32 mask_lo, u32 mask_hi)
566 {
567 	struct mtrr_var_range *vr;
568 
569 	vr = mtrr_state.var_ranges;
570 
571 	vr[index].base_lo = base_lo;
572 	vr[index].base_hi = base_hi;
573 	vr[index].mask_lo = mask_lo;
574 	vr[index].mask_hi = mask_hi;
575 }
576 
577 static void get_fixed_ranges(mtrr_type *frs)
578 {
579 	unsigned int *p = (unsigned int *)frs;
580 	int i;
581 
582 	k8_check_syscfg_dram_mod_en();
583 
584 	rdmsr(MSR_MTRRfix64K_00000, p[0], p[1]);
585 
586 	for (i = 0; i < 2; i++)
587 		rdmsr(MSR_MTRRfix16K_80000 + i, p[2 + i * 2], p[3 + i * 2]);
588 	for (i = 0; i < 8; i++)
589 		rdmsr(MSR_MTRRfix4K_C0000 + i, p[6 + i * 2], p[7 + i * 2]);
590 }
591 
592 void mtrr_save_fixed_ranges(void *info)
593 {
594 	if (boot_cpu_has(X86_FEATURE_MTRR))
595 		get_fixed_ranges(mtrr_state.fixed_ranges);
596 }
597 
598 static unsigned __initdata last_fixed_start;
599 static unsigned __initdata last_fixed_end;
600 static mtrr_type __initdata last_fixed_type;
601 
602 static void __init print_fixed_last(void)
603 {
604 	if (!last_fixed_end)
605 		return;
606 
607 	pr_info("  %05X-%05X %s\n", last_fixed_start,
608 		last_fixed_end - 1, mtrr_attrib_to_str(last_fixed_type));
609 
610 	last_fixed_end = 0;
611 }
612 
613 static void __init update_fixed_last(unsigned base, unsigned end,
614 				     mtrr_type type)
615 {
616 	last_fixed_start = base;
617 	last_fixed_end = end;
618 	last_fixed_type = type;
619 }
620 
621 static void __init
622 print_fixed(unsigned base, unsigned step, const mtrr_type *types)
623 {
624 	unsigned i;
625 
626 	for (i = 0; i < 8; ++i, ++types, base += step) {
627 		if (last_fixed_end == 0) {
628 			update_fixed_last(base, base + step, *types);
629 			continue;
630 		}
631 		if (last_fixed_end == base && last_fixed_type == *types) {
632 			last_fixed_end = base + step;
633 			continue;
634 		}
635 		/* new segments: gap or different type */
636 		print_fixed_last();
637 		update_fixed_last(base, base + step, *types);
638 	}
639 }
640 
641 static void __init print_mtrr_state(void)
642 {
643 	unsigned int i;
644 	int high_width;
645 
646 	pr_info("MTRR default type: %s\n",
647 		mtrr_attrib_to_str(mtrr_state.def_type));
648 	if (mtrr_state.have_fixed) {
649 		pr_info("MTRR fixed ranges %sabled:\n",
650 			((mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED) &&
651 			 (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED)) ?
652 			 "en" : "dis");
653 		print_fixed(0x00000, 0x10000, mtrr_state.fixed_ranges + 0);
654 		for (i = 0; i < 2; ++i)
655 			print_fixed(0x80000 + i * 0x20000, 0x04000,
656 				    mtrr_state.fixed_ranges + (i + 1) * 8);
657 		for (i = 0; i < 8; ++i)
658 			print_fixed(0xC0000 + i * 0x08000, 0x01000,
659 				    mtrr_state.fixed_ranges + (i + 3) * 8);
660 
661 		/* tail */
662 		print_fixed_last();
663 	}
664 	pr_info("MTRR variable ranges %sabled:\n",
665 		mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED ? "en" : "dis");
666 	high_width = (boot_cpu_data.x86_phys_bits - (32 - PAGE_SHIFT) + 3) / 4;
667 
668 	for (i = 0; i < num_var_ranges; ++i) {
669 		if (mtrr_state.var_ranges[i].mask_lo & MTRR_PHYSMASK_V)
670 			pr_info("  %u base %0*X%05X000 mask %0*X%05X000 %s\n",
671 				i,
672 				high_width,
673 				mtrr_state.var_ranges[i].base_hi,
674 				mtrr_state.var_ranges[i].base_lo >> 12,
675 				high_width,
676 				mtrr_state.var_ranges[i].mask_hi,
677 				mtrr_state.var_ranges[i].mask_lo >> 12,
678 				mtrr_attrib_to_str(mtrr_state.var_ranges[i].base_lo &
679 						    MTRR_PHYSBASE_TYPE));
680 		else
681 			pr_info("  %u disabled\n", i);
682 	}
683 	if (mtrr_tom2)
684 		pr_info("TOM2: %016llx aka %lldM\n", mtrr_tom2, mtrr_tom2>>20);
685 }
686 
687 /* Grab all of the MTRR state for this CPU into *state */
688 bool __init get_mtrr_state(void)
689 {
690 	struct mtrr_var_range *vrs;
691 	unsigned lo, dummy;
692 	unsigned int i;
693 
694 	vrs = mtrr_state.var_ranges;
695 
696 	rdmsr(MSR_MTRRcap, lo, dummy);
697 	mtrr_state.have_fixed = lo & MTRR_CAP_FIX;
698 
699 	for (i = 0; i < num_var_ranges; i++)
700 		get_mtrr_var_range(i, &vrs[i]);
701 	if (mtrr_state.have_fixed)
702 		get_fixed_ranges(mtrr_state.fixed_ranges);
703 
704 	rdmsr(MSR_MTRRdefType, lo, dummy);
705 	mtrr_state.def_type = lo & MTRR_DEF_TYPE_TYPE;
706 	mtrr_state.enabled = (lo & MTRR_DEF_TYPE_ENABLE) >> MTRR_STATE_SHIFT;
707 
708 	if (amd_special_default_mtrr()) {
709 		unsigned low, high;
710 
711 		/* TOP_MEM2 */
712 		rdmsr(MSR_K8_TOP_MEM2, low, high);
713 		mtrr_tom2 = high;
714 		mtrr_tom2 <<= 32;
715 		mtrr_tom2 |= low;
716 		mtrr_tom2 &= 0xffffff800000ULL;
717 	}
718 
719 	if (mtrr_debug)
720 		print_mtrr_state();
721 
722 	mtrr_state_set = 1;
723 
724 	return !!(mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED);
725 }
726 
727 /* Some BIOS's are messed up and don't set all MTRRs the same! */
728 void __init mtrr_state_warn(void)
729 {
730 	unsigned long mask = smp_changes_mask;
731 
732 	if (!mask)
733 		return;
734 	if (mask & MTRR_CHANGE_MASK_FIXED)
735 		pr_warn("mtrr: your CPUs had inconsistent fixed MTRR settings\n");
736 	if (mask & MTRR_CHANGE_MASK_VARIABLE)
737 		pr_warn("mtrr: your CPUs had inconsistent variable MTRR settings\n");
738 	if (mask & MTRR_CHANGE_MASK_DEFTYPE)
739 		pr_warn("mtrr: your CPUs had inconsistent MTRRdefType settings\n");
740 
741 	pr_info("mtrr: probably your BIOS does not setup all CPUs.\n");
742 	pr_info("mtrr: corrected configuration.\n");
743 }
744 
745 /*
746  * Doesn't attempt to pass an error out to MTRR users
747  * because it's quite complicated in some cases and probably not
748  * worth it because the best error handling is to ignore it.
749  */
750 void mtrr_wrmsr(unsigned msr, unsigned a, unsigned b)
751 {
752 	if (wrmsr_safe(msr, a, b) < 0) {
753 		pr_err("MTRR: CPU %u: Writing MSR %x to %x:%x failed\n",
754 			smp_processor_id(), msr, a, b);
755 	}
756 }
757 
758 /**
759  * set_fixed_range - checks & updates a fixed-range MTRR if it
760  *		     differs from the value it should have
761  * @msr: MSR address of the MTTR which should be checked and updated
762  * @changed: pointer which indicates whether the MTRR needed to be changed
763  * @msrwords: pointer to the MSR values which the MSR should have
764  */
765 static void set_fixed_range(int msr, bool *changed, unsigned int *msrwords)
766 {
767 	unsigned lo, hi;
768 
769 	rdmsr(msr, lo, hi);
770 
771 	if (lo != msrwords[0] || hi != msrwords[1]) {
772 		mtrr_wrmsr(msr, msrwords[0], msrwords[1]);
773 		*changed = true;
774 	}
775 }
776 
777 /**
778  * generic_get_free_region - Get a free MTRR.
779  * @base: The starting (base) address of the region.
780  * @size: The size (in bytes) of the region.
781  * @replace_reg: mtrr index to be replaced; set to invalid value if none.
782  *
783  * Returns: The index of the region on success, else negative on error.
784  */
785 int
786 generic_get_free_region(unsigned long base, unsigned long size, int replace_reg)
787 {
788 	unsigned long lbase, lsize;
789 	mtrr_type ltype;
790 	int i, max;
791 
792 	max = num_var_ranges;
793 	if (replace_reg >= 0 && replace_reg < max)
794 		return replace_reg;
795 
796 	for (i = 0; i < max; ++i) {
797 		mtrr_if->get(i, &lbase, &lsize, &ltype);
798 		if (lsize == 0)
799 			return i;
800 	}
801 
802 	return -ENOSPC;
803 }
804 
805 static void generic_get_mtrr(unsigned int reg, unsigned long *base,
806 			     unsigned long *size, mtrr_type *type)
807 {
808 	u32 mask_lo, mask_hi, base_lo, base_hi;
809 	unsigned int hi;
810 	u64 tmp, mask;
811 
812 	/*
813 	 * get_mtrr doesn't need to update mtrr_state, also it could be called
814 	 * from any cpu, so try to print it out directly.
815 	 */
816 	get_cpu();
817 
818 	rdmsr(MTRRphysMask_MSR(reg), mask_lo, mask_hi);
819 
820 	if (!(mask_lo & MTRR_PHYSMASK_V)) {
821 		/*  Invalid (i.e. free) range */
822 		*base = 0;
823 		*size = 0;
824 		*type = 0;
825 		goto out_put_cpu;
826 	}
827 
828 	rdmsr(MTRRphysBase_MSR(reg), base_lo, base_hi);
829 
830 	/* Work out the shifted address mask: */
831 	tmp = (u64)mask_hi << 32 | (mask_lo & PAGE_MASK);
832 	mask = (u64)phys_hi_rsvd << 32 | tmp;
833 
834 	/* Expand tmp with high bits to all 1s: */
835 	hi = fls64(tmp);
836 	if (hi > 0) {
837 		tmp |= ~((1ULL<<(hi - 1)) - 1);
838 
839 		if (tmp != mask) {
840 			pr_warn("mtrr: your BIOS has configured an incorrect mask, fixing it.\n");
841 			add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
842 			mask = tmp;
843 		}
844 	}
845 
846 	/*
847 	 * This works correctly if size is a power of two, i.e. a
848 	 * contiguous range:
849 	 */
850 	*size = -mask >> PAGE_SHIFT;
851 	*base = (u64)base_hi << (32 - PAGE_SHIFT) | base_lo >> PAGE_SHIFT;
852 	*type = base_lo & MTRR_PHYSBASE_TYPE;
853 
854 out_put_cpu:
855 	put_cpu();
856 }
857 
858 /**
859  * set_fixed_ranges - checks & updates the fixed-range MTRRs if they
860  *		      differ from the saved set
861  * @frs: pointer to fixed-range MTRR values, saved by get_fixed_ranges()
862  */
863 static int set_fixed_ranges(mtrr_type *frs)
864 {
865 	unsigned long long *saved = (unsigned long long *)frs;
866 	bool changed = false;
867 	int block = -1, range;
868 
869 	k8_check_syscfg_dram_mod_en();
870 
871 	while (fixed_range_blocks[++block].ranges) {
872 		for (range = 0; range < fixed_range_blocks[block].ranges; range++)
873 			set_fixed_range(fixed_range_blocks[block].base_msr + range,
874 					&changed, (unsigned int *)saved++);
875 	}
876 
877 	return changed;
878 }
879 
880 /*
881  * Set the MSR pair relating to a var range.
882  * Returns true if changes are made.
883  */
884 static bool set_mtrr_var_ranges(unsigned int index, struct mtrr_var_range *vr)
885 {
886 	unsigned int lo, hi;
887 	bool changed = false;
888 
889 	rdmsr(MTRRphysBase_MSR(index), lo, hi);
890 	if ((vr->base_lo & ~MTRR_PHYSBASE_RSVD) != (lo & ~MTRR_PHYSBASE_RSVD)
891 	    || (vr->base_hi & ~phys_hi_rsvd) != (hi & ~phys_hi_rsvd)) {
892 
893 		mtrr_wrmsr(MTRRphysBase_MSR(index), vr->base_lo, vr->base_hi);
894 		changed = true;
895 	}
896 
897 	rdmsr(MTRRphysMask_MSR(index), lo, hi);
898 
899 	if ((vr->mask_lo & ~MTRR_PHYSMASK_RSVD) != (lo & ~MTRR_PHYSMASK_RSVD)
900 	    || (vr->mask_hi & ~phys_hi_rsvd) != (hi & ~phys_hi_rsvd)) {
901 		mtrr_wrmsr(MTRRphysMask_MSR(index), vr->mask_lo, vr->mask_hi);
902 		changed = true;
903 	}
904 	return changed;
905 }
906 
907 static u32 deftype_lo, deftype_hi;
908 
909 /**
910  * set_mtrr_state - Set the MTRR state for this CPU.
911  *
912  * NOTE: The CPU must already be in a safe state for MTRR changes, including
913  *       measures that only a single CPU can be active in set_mtrr_state() in
914  *       order to not be subject to races for usage of deftype_lo. This is
915  *       accomplished by taking cache_disable_lock.
916  * RETURNS: 0 if no changes made, else a mask indicating what was changed.
917  */
918 static unsigned long set_mtrr_state(void)
919 {
920 	unsigned long change_mask = 0;
921 	unsigned int i;
922 
923 	for (i = 0; i < num_var_ranges; i++) {
924 		if (set_mtrr_var_ranges(i, &mtrr_state.var_ranges[i]))
925 			change_mask |= MTRR_CHANGE_MASK_VARIABLE;
926 	}
927 
928 	if (mtrr_state.have_fixed && set_fixed_ranges(mtrr_state.fixed_ranges))
929 		change_mask |= MTRR_CHANGE_MASK_FIXED;
930 
931 	/*
932 	 * Set_mtrr_restore restores the old value of MTRRdefType,
933 	 * so to set it we fiddle with the saved value:
934 	 */
935 	if ((deftype_lo & MTRR_DEF_TYPE_TYPE) != mtrr_state.def_type ||
936 	    ((deftype_lo & MTRR_DEF_TYPE_ENABLE) >> MTRR_STATE_SHIFT) != mtrr_state.enabled) {
937 
938 		deftype_lo = (deftype_lo & MTRR_DEF_TYPE_DISABLE) |
939 			     mtrr_state.def_type |
940 			     (mtrr_state.enabled << MTRR_STATE_SHIFT);
941 		change_mask |= MTRR_CHANGE_MASK_DEFTYPE;
942 	}
943 
944 	return change_mask;
945 }
946 
947 void mtrr_disable(void)
948 {
949 	/* Save MTRR state */
950 	rdmsr(MSR_MTRRdefType, deftype_lo, deftype_hi);
951 
952 	/* Disable MTRRs, and set the default type to uncached */
953 	mtrr_wrmsr(MSR_MTRRdefType, deftype_lo & MTRR_DEF_TYPE_DISABLE, deftype_hi);
954 }
955 
956 void mtrr_enable(void)
957 {
958 	/* Intel (P6) standard MTRRs */
959 	mtrr_wrmsr(MSR_MTRRdefType, deftype_lo, deftype_hi);
960 }
961 
962 void mtrr_generic_set_state(void)
963 {
964 	unsigned long mask, count;
965 
966 	/* Actually set the state */
967 	mask = set_mtrr_state();
968 
969 	/* Use the atomic bitops to update the global mask */
970 	for (count = 0; count < sizeof(mask) * 8; ++count) {
971 		if (mask & 0x01)
972 			set_bit(count, &smp_changes_mask);
973 		mask >>= 1;
974 	}
975 }
976 
977 /**
978  * generic_set_mtrr - set variable MTRR register on the local CPU.
979  *
980  * @reg: The register to set.
981  * @base: The base address of the region.
982  * @size: The size of the region. If this is 0 the region is disabled.
983  * @type: The type of the region.
984  *
985  * Returns nothing.
986  */
987 static void generic_set_mtrr(unsigned int reg, unsigned long base,
988 			     unsigned long size, mtrr_type type)
989 {
990 	unsigned long flags;
991 	struct mtrr_var_range *vr;
992 
993 	vr = &mtrr_state.var_ranges[reg];
994 
995 	local_irq_save(flags);
996 	cache_disable();
997 
998 	if (size == 0) {
999 		/*
1000 		 * The invalid bit is kept in the mask, so we simply
1001 		 * clear the relevant mask register to disable a range.
1002 		 */
1003 		mtrr_wrmsr(MTRRphysMask_MSR(reg), 0, 0);
1004 		memset(vr, 0, sizeof(struct mtrr_var_range));
1005 	} else {
1006 		vr->base_lo = base << PAGE_SHIFT | type;
1007 		vr->base_hi = (base >> (32 - PAGE_SHIFT)) & ~phys_hi_rsvd;
1008 		vr->mask_lo = -size << PAGE_SHIFT | MTRR_PHYSMASK_V;
1009 		vr->mask_hi = (-size >> (32 - PAGE_SHIFT)) & ~phys_hi_rsvd;
1010 
1011 		mtrr_wrmsr(MTRRphysBase_MSR(reg), vr->base_lo, vr->base_hi);
1012 		mtrr_wrmsr(MTRRphysMask_MSR(reg), vr->mask_lo, vr->mask_hi);
1013 	}
1014 
1015 	cache_enable();
1016 	local_irq_restore(flags);
1017 }
1018 
1019 int generic_validate_add_page(unsigned long base, unsigned long size,
1020 			      unsigned int type)
1021 {
1022 	unsigned long lbase, last;
1023 
1024 	/*
1025 	 * For Intel PPro stepping <= 7
1026 	 * must be 4 MiB aligned and not touch 0x70000000 -> 0x7003FFFF
1027 	 */
1028 	if (mtrr_if == &generic_mtrr_ops && boot_cpu_data.x86 == 6 &&
1029 	    boot_cpu_data.x86_model == 1 &&
1030 	    boot_cpu_data.x86_stepping <= 7) {
1031 		if (base & ((1 << (22 - PAGE_SHIFT)) - 1)) {
1032 			pr_warn("mtrr: base(0x%lx000) is not 4 MiB aligned\n", base);
1033 			return -EINVAL;
1034 		}
1035 		if (!(base + size < 0x70000 || base > 0x7003F) &&
1036 		    (type == MTRR_TYPE_WRCOMB
1037 		     || type == MTRR_TYPE_WRBACK)) {
1038 			pr_warn("mtrr: writable mtrr between 0x70000000 and 0x7003FFFF may hang the CPU.\n");
1039 			return -EINVAL;
1040 		}
1041 	}
1042 
1043 	/*
1044 	 * Check upper bits of base and last are equal and lower bits are 0
1045 	 * for base and 1 for last
1046 	 */
1047 	last = base + size - 1;
1048 	for (lbase = base; !(lbase & 1) && (last & 1);
1049 	     lbase = lbase >> 1, last = last >> 1)
1050 		;
1051 	if (lbase != last) {
1052 		pr_warn("mtrr: base(0x%lx000) is not aligned on a size(0x%lx000) boundary\n", base, size);
1053 		return -EINVAL;
1054 	}
1055 	return 0;
1056 }
1057 
1058 static int generic_have_wrcomb(void)
1059 {
1060 	unsigned long config, dummy;
1061 	rdmsr(MSR_MTRRcap, config, dummy);
1062 	return config & MTRR_CAP_WC;
1063 }
1064 
1065 int positive_have_wrcomb(void)
1066 {
1067 	return 1;
1068 }
1069 
1070 /*
1071  * Generic structure...
1072  */
1073 const struct mtrr_ops generic_mtrr_ops = {
1074 	.get			= generic_get_mtrr,
1075 	.get_free_region	= generic_get_free_region,
1076 	.set			= generic_set_mtrr,
1077 	.validate_add_page	= generic_validate_add_page,
1078 	.have_wrcomb		= generic_have_wrcomb,
1079 };
1080