xref: /openbmc/linux/arch/x86/kernel/cpu/mtrr/generic.c (revision a4316603534cf7f4dcd9f9f1ed170257b987cca9)
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 static 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 /*
162  * Check and return the effective type for MTRR-MTRR type overlap.
163  * Returns true if the effective type is UNCACHEABLE, else returns false
164  */
165 static bool check_type_overlap(u8 *prev, u8 *curr)
166 {
167 	*prev = *curr = get_effective_type(*curr, *prev);
168 
169 	return *prev == MTRR_TYPE_UNCACHABLE;
170 }
171 
172 /**
173  * mtrr_type_lookup_fixed - look up memory type in MTRR fixed entries
174  *
175  * Return the MTRR fixed memory type of 'start'.
176  *
177  * MTRR fixed entries are divided into the following ways:
178  *  0x00000 - 0x7FFFF : This range is divided into eight 64KB sub-ranges
179  *  0x80000 - 0xBFFFF : This range is divided into sixteen 16KB sub-ranges
180  *  0xC0000 - 0xFFFFF : This range is divided into sixty-four 4KB sub-ranges
181  *
182  * Return Values:
183  * MTRR_TYPE_(type)  - Matched memory type
184  * MTRR_TYPE_INVALID - Unmatched
185  */
186 static u8 mtrr_type_lookup_fixed(u64 start, u64 end)
187 {
188 	int idx;
189 
190 	if (start >= 0x100000)
191 		return MTRR_TYPE_INVALID;
192 
193 	/* 0x0 - 0x7FFFF */
194 	if (start < 0x80000) {
195 		idx = 0;
196 		idx += (start >> 16);
197 		return mtrr_state.fixed_ranges[idx];
198 	/* 0x80000 - 0xBFFFF */
199 	} else if (start < 0xC0000) {
200 		idx = 1 * 8;
201 		idx += ((start - 0x80000) >> 14);
202 		return mtrr_state.fixed_ranges[idx];
203 	}
204 
205 	/* 0xC0000 - 0xFFFFF */
206 	idx = 3 * 8;
207 	idx += ((start - 0xC0000) >> 12);
208 	return mtrr_state.fixed_ranges[idx];
209 }
210 
211 /**
212  * mtrr_type_lookup_variable - look up memory type in MTRR variable entries
213  *
214  * Return Value:
215  * MTRR_TYPE_(type) - Matched memory type or default memory type (unmatched)
216  *
217  * Output Arguments:
218  * repeat - Set to 1 when [start:end] spanned across MTRR range and type
219  *	    returned corresponds only to [start:*partial_end].  Caller has
220  *	    to lookup again for [*partial_end:end].
221  *
222  * uniform - Set to 1 when an MTRR covers the region uniformly, i.e. the
223  *	     region is fully covered by a single MTRR entry or the default
224  *	     type.
225  */
226 static u8 mtrr_type_lookup_variable(u64 start, u64 end, u64 *partial_end,
227 				    int *repeat, u8 *uniform)
228 {
229 	int i;
230 	u64 base, mask;
231 	u8 prev_match, curr_match;
232 
233 	*repeat = 0;
234 	*uniform = 1;
235 
236 	prev_match = MTRR_TYPE_INVALID;
237 	for (i = 0; i < num_var_ranges; ++i) {
238 		unsigned short start_state, end_state, inclusive;
239 
240 		if (!(mtrr_state.var_ranges[i].mask_lo & MTRR_PHYSMASK_V))
241 			continue;
242 
243 		base = (((u64)mtrr_state.var_ranges[i].base_hi) << 32) +
244 		       (mtrr_state.var_ranges[i].base_lo & PAGE_MASK);
245 		mask = (((u64)mtrr_state.var_ranges[i].mask_hi) << 32) +
246 		       (mtrr_state.var_ranges[i].mask_lo & PAGE_MASK);
247 
248 		start_state = ((start & mask) == (base & mask));
249 		end_state = ((end & mask) == (base & mask));
250 		inclusive = ((start < base) && (end > base));
251 
252 		if ((start_state != end_state) || inclusive) {
253 			/*
254 			 * We have start:end spanning across an MTRR.
255 			 * We split the region into either
256 			 *
257 			 * - start_state:1
258 			 * (start:mtrr_end)(mtrr_end:end)
259 			 * - end_state:1
260 			 * (start:mtrr_start)(mtrr_start:end)
261 			 * - inclusive:1
262 			 * (start:mtrr_start)(mtrr_start:mtrr_end)(mtrr_end:end)
263 			 *
264 			 * depending on kind of overlap.
265 			 *
266 			 * Return the type of the first region and a pointer
267 			 * to the start of next region so that caller will be
268 			 * advised to lookup again after having adjusted start
269 			 * and end.
270 			 *
271 			 * Note: This way we handle overlaps with multiple
272 			 * entries and the default type properly.
273 			 */
274 			if (start_state)
275 				*partial_end = base + get_mtrr_size(mask);
276 			else
277 				*partial_end = base;
278 
279 			if (unlikely(*partial_end <= start)) {
280 				WARN_ON(1);
281 				*partial_end = start + PAGE_SIZE;
282 			}
283 
284 			end = *partial_end - 1; /* end is inclusive */
285 			*repeat = 1;
286 			*uniform = 0;
287 		}
288 
289 		if ((start & mask) != (base & mask))
290 			continue;
291 
292 		curr_match = mtrr_state.var_ranges[i].base_lo & MTRR_PHYSBASE_TYPE;
293 		if (prev_match == MTRR_TYPE_INVALID) {
294 			prev_match = curr_match;
295 			continue;
296 		}
297 
298 		*uniform = 0;
299 		if (check_type_overlap(&prev_match, &curr_match))
300 			return curr_match;
301 	}
302 
303 	if (prev_match != MTRR_TYPE_INVALID)
304 		return prev_match;
305 
306 	return mtrr_state.def_type;
307 }
308 
309 static void rm_map_entry_at(int idx)
310 {
311 	cache_map_n--;
312 	if (cache_map_n > idx) {
313 		memmove(cache_map + idx, cache_map + idx + 1,
314 			sizeof(*cache_map) * (cache_map_n - idx));
315 	}
316 }
317 
318 /*
319  * Add an entry into cache_map at a specific index.  Merges adjacent entries if
320  * appropriate.  Return the number of merges for correcting the scan index
321  * (this is needed as merging will reduce the number of entries, which will
322  * result in skipping entries in future iterations if the scan index isn't
323  * corrected).
324  * Note that the corrected index can never go below -1 (resulting in being 0 in
325  * the next scan iteration), as "2" is returned only if the current index is
326  * larger than zero.
327  */
328 static int add_map_entry_at(u64 start, u64 end, u8 type, int idx)
329 {
330 	bool merge_prev = false, merge_next = false;
331 
332 	if (start >= end)
333 		return 0;
334 
335 	if (idx > 0) {
336 		struct cache_map *prev = cache_map + idx - 1;
337 
338 		if (!prev->fixed && start == prev->end && type == prev->type)
339 			merge_prev = true;
340 	}
341 
342 	if (idx < cache_map_n) {
343 		struct cache_map *next = cache_map + idx;
344 
345 		if (!next->fixed && end == next->start && type == next->type)
346 			merge_next = true;
347 	}
348 
349 	if (merge_prev && merge_next) {
350 		cache_map[idx - 1].end = cache_map[idx].end;
351 		rm_map_entry_at(idx);
352 		return 2;
353 	}
354 	if (merge_prev) {
355 		cache_map[idx - 1].end = end;
356 		return 1;
357 	}
358 	if (merge_next) {
359 		cache_map[idx].start = start;
360 		return 1;
361 	}
362 
363 	/* Sanity check: the array should NEVER be too small! */
364 	if (cache_map_n == cache_map_size) {
365 		WARN(1, "MTRR cache mode memory map exhausted!\n");
366 		cache_map_n = cache_map_fixed;
367 		return 0;
368 	}
369 
370 	if (cache_map_n > idx) {
371 		memmove(cache_map + idx + 1, cache_map + idx,
372 			sizeof(*cache_map) * (cache_map_n - idx));
373 	}
374 
375 	cache_map[idx].start = start;
376 	cache_map[idx].end = end;
377 	cache_map[idx].type = type;
378 	cache_map[idx].fixed = 0;
379 	cache_map_n++;
380 
381 	return 0;
382 }
383 
384 /* Clear a part of an entry. Return 1 if start of entry is still valid. */
385 static int clr_map_range_at(u64 start, u64 end, int idx)
386 {
387 	int ret = start != cache_map[idx].start;
388 	u64 tmp;
389 
390 	if (start == cache_map[idx].start && end == cache_map[idx].end) {
391 		rm_map_entry_at(idx);
392 	} else if (start == cache_map[idx].start) {
393 		cache_map[idx].start = end;
394 	} else if (end == cache_map[idx].end) {
395 		cache_map[idx].end = start;
396 	} else {
397 		tmp = cache_map[idx].end;
398 		cache_map[idx].end = start;
399 		add_map_entry_at(end, tmp, cache_map[idx].type, idx + 1);
400 	}
401 
402 	return ret;
403 }
404 
405 /*
406  * Add MTRR to the map.  The current map is scanned and each part of the MTRR
407  * either overlapping with an existing entry or with a hole in the map is
408  * handled separately.
409  */
410 static void add_map_entry(u64 start, u64 end, u8 type)
411 {
412 	u8 new_type, old_type;
413 	u64 tmp;
414 	int i;
415 
416 	for (i = 0; i < cache_map_n && start < end; i++) {
417 		if (start >= cache_map[i].end)
418 			continue;
419 
420 		if (start < cache_map[i].start) {
421 			/* Region start has no overlap. */
422 			tmp = min(end, cache_map[i].start);
423 			i -= add_map_entry_at(start, tmp,  type, i);
424 			start = tmp;
425 			continue;
426 		}
427 
428 		new_type = get_effective_type(type, cache_map[i].type);
429 		old_type = cache_map[i].type;
430 
431 		if (cache_map[i].fixed || new_type == old_type) {
432 			/* Cut off start of new entry. */
433 			start = cache_map[i].end;
434 			continue;
435 		}
436 
437 		/* Handle only overlapping part of region. */
438 		tmp = min(end, cache_map[i].end);
439 		i += clr_map_range_at(start, tmp, i);
440 		i -= add_map_entry_at(start, tmp, new_type, i);
441 		start = tmp;
442 	}
443 
444 	/* Add rest of region after last map entry (rest might be empty). */
445 	add_map_entry_at(start, end, type, i);
446 }
447 
448 /* Add variable MTRRs to cache map. */
449 static void map_add_var(void)
450 {
451 	u64 start, size;
452 	unsigned int i;
453 	u8 type;
454 
455 	/*
456 	 * Add AMD TOP_MEM2 area.  Can't be added in mtrr_build_map(), as it
457 	 * needs to be added again when rebuilding the map due to potentially
458 	 * having moved as a result of variable MTRRs for memory below 4GB.
459 	 */
460 	if (mtrr_tom2) {
461 		add_map_entry(BIT_ULL(32), mtrr_tom2, MTRR_TYPE_WRBACK);
462 		cache_map[cache_map_n - 1].fixed = 1;
463 	}
464 
465 	for (i = 0; i < num_var_ranges; i++) {
466 		type = get_var_mtrr_state(i, &start, &size);
467 		if (type != MTRR_TYPE_INVALID)
468 			add_map_entry(start, start + size, type);
469 	}
470 }
471 
472 /*
473  * Rebuild map by replacing variable entries.  Needs to be called when MTRR
474  * registers are being changed after boot, as such changes could include
475  * removals of registers, which are complicated to handle without rebuild of
476  * the map.
477  */
478 void generic_rebuild_map(void)
479 {
480 	if (mtrr_if != &generic_mtrr_ops)
481 		return;
482 
483 	cache_map_n = cache_map_fixed;
484 
485 	map_add_var();
486 }
487 
488 static unsigned int __init get_cache_map_size(void)
489 {
490 	return cache_map_fixed + 2 * num_var_ranges + (mtrr_tom2 != 0);
491 }
492 
493 /* Build the cache_map containing the cache modes per memory range. */
494 void __init mtrr_build_map(void)
495 {
496 	u64 start, end, size;
497 	unsigned int i;
498 	u8 type;
499 
500 	/* Add fixed MTRRs, optimize for adjacent entries with same type. */
501 	if (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED) {
502 		/*
503 		 * Start with 64k size fixed entries, preset 1st one (hence the
504 		 * loop below is starting with index 1).
505 		 */
506 		start = 0;
507 		end = size = 0x10000;
508 		type = mtrr_state.fixed_ranges[0];
509 
510 		for (i = 1; i < MTRR_NUM_FIXED_RANGES; i++) {
511 			/* 8 64k entries, then 16 16k ones, rest 4k. */
512 			if (i == 8 || i == 24)
513 				size >>= 2;
514 
515 			if (mtrr_state.fixed_ranges[i] != type) {
516 				add_map_entry(start, end, type);
517 				start = end;
518 				type = mtrr_state.fixed_ranges[i];
519 			}
520 			end += size;
521 		}
522 		add_map_entry(start, end, type);
523 	}
524 
525 	/* Mark fixed, they take precedence. */
526 	for (i = 0; i < cache_map_n; i++)
527 		cache_map[i].fixed = 1;
528 	cache_map_fixed = cache_map_n;
529 
530 	map_add_var();
531 
532 	pr_info("MTRR map: %u entries (%u fixed + %u variable; max %u), built from %u variable MTRRs\n",
533 		cache_map_n, cache_map_fixed, cache_map_n - cache_map_fixed,
534 		get_cache_map_size(), num_var_ranges + (mtrr_tom2 != 0));
535 
536 	if (mtrr_debug) {
537 		for (i = 0; i < cache_map_n; i++) {
538 			pr_info("%3u: %016llx-%016llx %s\n", i,
539 				cache_map[i].start, cache_map[i].end - 1,
540 				mtrr_attrib_to_str(cache_map[i].type));
541 		}
542 	}
543 }
544 
545 /* Copy the cache_map from __initdata memory to dynamically allocated one. */
546 void __init mtrr_copy_map(void)
547 {
548 	unsigned int new_size = get_cache_map_size();
549 
550 	if (!mtrr_state.enabled || !new_size) {
551 		cache_map = NULL;
552 		return;
553 	}
554 
555 	mutex_lock(&mtrr_mutex);
556 
557 	cache_map = kcalloc(new_size, sizeof(*cache_map), GFP_KERNEL);
558 	if (cache_map) {
559 		memmove(cache_map, init_cache_map,
560 			cache_map_n * sizeof(*cache_map));
561 		cache_map_size = new_size;
562 	} else {
563 		mtrr_state.enabled = 0;
564 		pr_err("MTRRs disabled due to allocation failure for lookup map.\n");
565 	}
566 
567 	mutex_unlock(&mtrr_mutex);
568 }
569 
570 /**
571  * mtrr_overwrite_state - set static MTRR state
572  *
573  * Used to set MTRR state via different means (e.g. with data obtained from
574  * a hypervisor).
575  * Is allowed only for special cases when running virtualized. Must be called
576  * from the x86_init.hyper.init_platform() hook.  It can be called only once.
577  * The MTRR state can't be changed afterwards.  To ensure that, X86_FEATURE_MTRR
578  * is cleared.
579  */
580 void mtrr_overwrite_state(struct mtrr_var_range *var, unsigned int num_var,
581 			  mtrr_type def_type)
582 {
583 	unsigned int i;
584 
585 	/* Only allowed to be called once before mtrr_bp_init(). */
586 	if (WARN_ON_ONCE(mtrr_state_set))
587 		return;
588 
589 	/* Only allowed when running virtualized. */
590 	if (!cpu_feature_enabled(X86_FEATURE_HYPERVISOR))
591 		return;
592 
593 	/*
594 	 * Only allowed for special virtualization cases:
595 	 * - when running as Hyper-V, SEV-SNP guest using vTOM
596 	 * - when running as Xen PV guest
597 	 * - when running as SEV-SNP or TDX guest to avoid unnecessary
598 	 *   VMM communication/Virtualization exceptions (#VC, #VE)
599 	 */
600 	if (!cc_platform_has(CC_ATTR_GUEST_SEV_SNP) &&
601 	    !hv_is_isolation_supported() &&
602 	    !cpu_feature_enabled(X86_FEATURE_XENPV) &&
603 	    !cpu_feature_enabled(X86_FEATURE_TDX_GUEST))
604 		return;
605 
606 	/* Disable MTRR in order to disable MTRR modifications. */
607 	setup_clear_cpu_cap(X86_FEATURE_MTRR);
608 
609 	if (var) {
610 		if (num_var > MTRR_MAX_VAR_RANGES) {
611 			pr_warn("Trying to overwrite MTRR state with %u variable entries\n",
612 				num_var);
613 			num_var = MTRR_MAX_VAR_RANGES;
614 		}
615 		for (i = 0; i < num_var; i++)
616 			mtrr_state.var_ranges[i] = var[i];
617 		num_var_ranges = num_var;
618 	}
619 
620 	mtrr_state.def_type = def_type;
621 	mtrr_state.enabled |= MTRR_STATE_MTRR_ENABLED;
622 
623 	mtrr_state_set = 1;
624 }
625 
626 /**
627  * mtrr_type_lookup - look up memory type in MTRR
628  *
629  * Return Values:
630  * MTRR_TYPE_(type)  - The effective MTRR type for the region
631  * MTRR_TYPE_INVALID - MTRR is disabled
632  *
633  * Output Argument:
634  * uniform - Set to 1 when an MTRR covers the region uniformly, i.e. the
635  *	     region is fully covered by a single MTRR entry or the default
636  *	     type.
637  */
638 u8 mtrr_type_lookup(u64 start, u64 end, u8 *uniform)
639 {
640 	u8 type, prev_type, is_uniform = 1, dummy;
641 	int repeat;
642 	u64 partial_end;
643 
644 	/* Make end inclusive instead of exclusive */
645 	end--;
646 
647 	if (!mtrr_state_set)
648 		return MTRR_TYPE_INVALID;
649 
650 	if (!(mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED))
651 		return MTRR_TYPE_INVALID;
652 
653 	/*
654 	 * Look up the fixed ranges first, which take priority over
655 	 * the variable ranges.
656 	 */
657 	if ((start < 0x100000) &&
658 	    (mtrr_state.have_fixed) &&
659 	    (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED)) {
660 		is_uniform = 0;
661 		type = mtrr_type_lookup_fixed(start, end);
662 		goto out;
663 	}
664 
665 	/*
666 	 * Look up the variable ranges.  Look of multiple ranges matching
667 	 * this address and pick type as per MTRR precedence.
668 	 */
669 	type = mtrr_type_lookup_variable(start, end, &partial_end,
670 					 &repeat, &is_uniform);
671 
672 	/*
673 	 * Common path is with repeat = 0.
674 	 * However, we can have cases where [start:end] spans across some
675 	 * MTRR ranges and/or the default type.  Do repeated lookups for
676 	 * that case here.
677 	 */
678 	while (repeat) {
679 		prev_type = type;
680 		start = partial_end;
681 		is_uniform = 0;
682 		type = mtrr_type_lookup_variable(start, end, &partial_end,
683 						 &repeat, &dummy);
684 
685 		if (check_type_overlap(&prev_type, &type))
686 			goto out;
687 	}
688 
689 	if (mtrr_tom2 && (start >= (1ULL<<32)) && (end < mtrr_tom2))
690 		type = MTRR_TYPE_WRBACK;
691 
692 out:
693 	*uniform = is_uniform;
694 	return type;
695 }
696 
697 /* Get the MSR pair relating to a var range */
698 static void
699 get_mtrr_var_range(unsigned int index, struct mtrr_var_range *vr)
700 {
701 	rdmsr(MTRRphysBase_MSR(index), vr->base_lo, vr->base_hi);
702 	rdmsr(MTRRphysMask_MSR(index), vr->mask_lo, vr->mask_hi);
703 }
704 
705 /* Fill the MSR pair relating to a var range */
706 void fill_mtrr_var_range(unsigned int index,
707 		u32 base_lo, u32 base_hi, u32 mask_lo, u32 mask_hi)
708 {
709 	struct mtrr_var_range *vr;
710 
711 	vr = mtrr_state.var_ranges;
712 
713 	vr[index].base_lo = base_lo;
714 	vr[index].base_hi = base_hi;
715 	vr[index].mask_lo = mask_lo;
716 	vr[index].mask_hi = mask_hi;
717 }
718 
719 static void get_fixed_ranges(mtrr_type *frs)
720 {
721 	unsigned int *p = (unsigned int *)frs;
722 	int i;
723 
724 	k8_check_syscfg_dram_mod_en();
725 
726 	rdmsr(MSR_MTRRfix64K_00000, p[0], p[1]);
727 
728 	for (i = 0; i < 2; i++)
729 		rdmsr(MSR_MTRRfix16K_80000 + i, p[2 + i * 2], p[3 + i * 2]);
730 	for (i = 0; i < 8; i++)
731 		rdmsr(MSR_MTRRfix4K_C0000 + i, p[6 + i * 2], p[7 + i * 2]);
732 }
733 
734 void mtrr_save_fixed_ranges(void *info)
735 {
736 	if (boot_cpu_has(X86_FEATURE_MTRR))
737 		get_fixed_ranges(mtrr_state.fixed_ranges);
738 }
739 
740 static unsigned __initdata last_fixed_start;
741 static unsigned __initdata last_fixed_end;
742 static mtrr_type __initdata last_fixed_type;
743 
744 static void __init print_fixed_last(void)
745 {
746 	if (!last_fixed_end)
747 		return;
748 
749 	pr_info("  %05X-%05X %s\n", last_fixed_start,
750 		last_fixed_end - 1, mtrr_attrib_to_str(last_fixed_type));
751 
752 	last_fixed_end = 0;
753 }
754 
755 static void __init update_fixed_last(unsigned base, unsigned end,
756 				     mtrr_type type)
757 {
758 	last_fixed_start = base;
759 	last_fixed_end = end;
760 	last_fixed_type = type;
761 }
762 
763 static void __init
764 print_fixed(unsigned base, unsigned step, const mtrr_type *types)
765 {
766 	unsigned i;
767 
768 	for (i = 0; i < 8; ++i, ++types, base += step) {
769 		if (last_fixed_end == 0) {
770 			update_fixed_last(base, base + step, *types);
771 			continue;
772 		}
773 		if (last_fixed_end == base && last_fixed_type == *types) {
774 			last_fixed_end = base + step;
775 			continue;
776 		}
777 		/* new segments: gap or different type */
778 		print_fixed_last();
779 		update_fixed_last(base, base + step, *types);
780 	}
781 }
782 
783 static void __init print_mtrr_state(void)
784 {
785 	unsigned int i;
786 	int high_width;
787 
788 	pr_info("MTRR default type: %s\n",
789 		mtrr_attrib_to_str(mtrr_state.def_type));
790 	if (mtrr_state.have_fixed) {
791 		pr_info("MTRR fixed ranges %sabled:\n",
792 			((mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED) &&
793 			 (mtrr_state.enabled & MTRR_STATE_MTRR_FIXED_ENABLED)) ?
794 			 "en" : "dis");
795 		print_fixed(0x00000, 0x10000, mtrr_state.fixed_ranges + 0);
796 		for (i = 0; i < 2; ++i)
797 			print_fixed(0x80000 + i * 0x20000, 0x04000,
798 				    mtrr_state.fixed_ranges + (i + 1) * 8);
799 		for (i = 0; i < 8; ++i)
800 			print_fixed(0xC0000 + i * 0x08000, 0x01000,
801 				    mtrr_state.fixed_ranges + (i + 3) * 8);
802 
803 		/* tail */
804 		print_fixed_last();
805 	}
806 	pr_info("MTRR variable ranges %sabled:\n",
807 		mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED ? "en" : "dis");
808 	high_width = (boot_cpu_data.x86_phys_bits - (32 - PAGE_SHIFT) + 3) / 4;
809 
810 	for (i = 0; i < num_var_ranges; ++i) {
811 		if (mtrr_state.var_ranges[i].mask_lo & MTRR_PHYSMASK_V)
812 			pr_info("  %u base %0*X%05X000 mask %0*X%05X000 %s\n",
813 				i,
814 				high_width,
815 				mtrr_state.var_ranges[i].base_hi,
816 				mtrr_state.var_ranges[i].base_lo >> 12,
817 				high_width,
818 				mtrr_state.var_ranges[i].mask_hi,
819 				mtrr_state.var_ranges[i].mask_lo >> 12,
820 				mtrr_attrib_to_str(mtrr_state.var_ranges[i].base_lo &
821 						    MTRR_PHYSBASE_TYPE));
822 		else
823 			pr_info("  %u disabled\n", i);
824 	}
825 	if (mtrr_tom2)
826 		pr_info("TOM2: %016llx aka %lldM\n", mtrr_tom2, mtrr_tom2>>20);
827 }
828 
829 /* Grab all of the MTRR state for this CPU into *state */
830 bool __init get_mtrr_state(void)
831 {
832 	struct mtrr_var_range *vrs;
833 	unsigned lo, dummy;
834 	unsigned int i;
835 
836 	vrs = mtrr_state.var_ranges;
837 
838 	rdmsr(MSR_MTRRcap, lo, dummy);
839 	mtrr_state.have_fixed = lo & MTRR_CAP_FIX;
840 
841 	for (i = 0; i < num_var_ranges; i++)
842 		get_mtrr_var_range(i, &vrs[i]);
843 	if (mtrr_state.have_fixed)
844 		get_fixed_ranges(mtrr_state.fixed_ranges);
845 
846 	rdmsr(MSR_MTRRdefType, lo, dummy);
847 	mtrr_state.def_type = lo & MTRR_DEF_TYPE_TYPE;
848 	mtrr_state.enabled = (lo & MTRR_DEF_TYPE_ENABLE) >> MTRR_STATE_SHIFT;
849 
850 	if (amd_special_default_mtrr()) {
851 		unsigned low, high;
852 
853 		/* TOP_MEM2 */
854 		rdmsr(MSR_K8_TOP_MEM2, low, high);
855 		mtrr_tom2 = high;
856 		mtrr_tom2 <<= 32;
857 		mtrr_tom2 |= low;
858 		mtrr_tom2 &= 0xffffff800000ULL;
859 	}
860 
861 	if (mtrr_debug)
862 		print_mtrr_state();
863 
864 	mtrr_state_set = 1;
865 
866 	return !!(mtrr_state.enabled & MTRR_STATE_MTRR_ENABLED);
867 }
868 
869 /* Some BIOS's are messed up and don't set all MTRRs the same! */
870 void __init mtrr_state_warn(void)
871 {
872 	unsigned long mask = smp_changes_mask;
873 
874 	if (!mask)
875 		return;
876 	if (mask & MTRR_CHANGE_MASK_FIXED)
877 		pr_warn("mtrr: your CPUs had inconsistent fixed MTRR settings\n");
878 	if (mask & MTRR_CHANGE_MASK_VARIABLE)
879 		pr_warn("mtrr: your CPUs had inconsistent variable MTRR settings\n");
880 	if (mask & MTRR_CHANGE_MASK_DEFTYPE)
881 		pr_warn("mtrr: your CPUs had inconsistent MTRRdefType settings\n");
882 
883 	pr_info("mtrr: probably your BIOS does not setup all CPUs.\n");
884 	pr_info("mtrr: corrected configuration.\n");
885 }
886 
887 /*
888  * Doesn't attempt to pass an error out to MTRR users
889  * because it's quite complicated in some cases and probably not
890  * worth it because the best error handling is to ignore it.
891  */
892 void mtrr_wrmsr(unsigned msr, unsigned a, unsigned b)
893 {
894 	if (wrmsr_safe(msr, a, b) < 0) {
895 		pr_err("MTRR: CPU %u: Writing MSR %x to %x:%x failed\n",
896 			smp_processor_id(), msr, a, b);
897 	}
898 }
899 
900 /**
901  * set_fixed_range - checks & updates a fixed-range MTRR if it
902  *		     differs from the value it should have
903  * @msr: MSR address of the MTTR which should be checked and updated
904  * @changed: pointer which indicates whether the MTRR needed to be changed
905  * @msrwords: pointer to the MSR values which the MSR should have
906  */
907 static void set_fixed_range(int msr, bool *changed, unsigned int *msrwords)
908 {
909 	unsigned lo, hi;
910 
911 	rdmsr(msr, lo, hi);
912 
913 	if (lo != msrwords[0] || hi != msrwords[1]) {
914 		mtrr_wrmsr(msr, msrwords[0], msrwords[1]);
915 		*changed = true;
916 	}
917 }
918 
919 /**
920  * generic_get_free_region - Get a free MTRR.
921  * @base: The starting (base) address of the region.
922  * @size: The size (in bytes) of the region.
923  * @replace_reg: mtrr index to be replaced; set to invalid value if none.
924  *
925  * Returns: The index of the region on success, else negative on error.
926  */
927 int
928 generic_get_free_region(unsigned long base, unsigned long size, int replace_reg)
929 {
930 	unsigned long lbase, lsize;
931 	mtrr_type ltype;
932 	int i, max;
933 
934 	max = num_var_ranges;
935 	if (replace_reg >= 0 && replace_reg < max)
936 		return replace_reg;
937 
938 	for (i = 0; i < max; ++i) {
939 		mtrr_if->get(i, &lbase, &lsize, &ltype);
940 		if (lsize == 0)
941 			return i;
942 	}
943 
944 	return -ENOSPC;
945 }
946 
947 static void generic_get_mtrr(unsigned int reg, unsigned long *base,
948 			     unsigned long *size, mtrr_type *type)
949 {
950 	u32 mask_lo, mask_hi, base_lo, base_hi;
951 	unsigned int hi;
952 	u64 tmp, mask;
953 
954 	/*
955 	 * get_mtrr doesn't need to update mtrr_state, also it could be called
956 	 * from any cpu, so try to print it out directly.
957 	 */
958 	get_cpu();
959 
960 	rdmsr(MTRRphysMask_MSR(reg), mask_lo, mask_hi);
961 
962 	if (!(mask_lo & MTRR_PHYSMASK_V)) {
963 		/*  Invalid (i.e. free) range */
964 		*base = 0;
965 		*size = 0;
966 		*type = 0;
967 		goto out_put_cpu;
968 	}
969 
970 	rdmsr(MTRRphysBase_MSR(reg), base_lo, base_hi);
971 
972 	/* Work out the shifted address mask: */
973 	tmp = (u64)mask_hi << 32 | (mask_lo & PAGE_MASK);
974 	mask = (u64)phys_hi_rsvd << 32 | tmp;
975 
976 	/* Expand tmp with high bits to all 1s: */
977 	hi = fls64(tmp);
978 	if (hi > 0) {
979 		tmp |= ~((1ULL<<(hi - 1)) - 1);
980 
981 		if (tmp != mask) {
982 			pr_warn("mtrr: your BIOS has configured an incorrect mask, fixing it.\n");
983 			add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
984 			mask = tmp;
985 		}
986 	}
987 
988 	/*
989 	 * This works correctly if size is a power of two, i.e. a
990 	 * contiguous range:
991 	 */
992 	*size = -mask >> PAGE_SHIFT;
993 	*base = (u64)base_hi << (32 - PAGE_SHIFT) | base_lo >> PAGE_SHIFT;
994 	*type = base_lo & MTRR_PHYSBASE_TYPE;
995 
996 out_put_cpu:
997 	put_cpu();
998 }
999 
1000 /**
1001  * set_fixed_ranges - checks & updates the fixed-range MTRRs if they
1002  *		      differ from the saved set
1003  * @frs: pointer to fixed-range MTRR values, saved by get_fixed_ranges()
1004  */
1005 static int set_fixed_ranges(mtrr_type *frs)
1006 {
1007 	unsigned long long *saved = (unsigned long long *)frs;
1008 	bool changed = false;
1009 	int block = -1, range;
1010 
1011 	k8_check_syscfg_dram_mod_en();
1012 
1013 	while (fixed_range_blocks[++block].ranges) {
1014 		for (range = 0; range < fixed_range_blocks[block].ranges; range++)
1015 			set_fixed_range(fixed_range_blocks[block].base_msr + range,
1016 					&changed, (unsigned int *)saved++);
1017 	}
1018 
1019 	return changed;
1020 }
1021 
1022 /*
1023  * Set the MSR pair relating to a var range.
1024  * Returns true if changes are made.
1025  */
1026 static bool set_mtrr_var_ranges(unsigned int index, struct mtrr_var_range *vr)
1027 {
1028 	unsigned int lo, hi;
1029 	bool changed = false;
1030 
1031 	rdmsr(MTRRphysBase_MSR(index), lo, hi);
1032 	if ((vr->base_lo & ~MTRR_PHYSBASE_RSVD) != (lo & ~MTRR_PHYSBASE_RSVD)
1033 	    || (vr->base_hi & ~phys_hi_rsvd) != (hi & ~phys_hi_rsvd)) {
1034 
1035 		mtrr_wrmsr(MTRRphysBase_MSR(index), vr->base_lo, vr->base_hi);
1036 		changed = true;
1037 	}
1038 
1039 	rdmsr(MTRRphysMask_MSR(index), lo, hi);
1040 
1041 	if ((vr->mask_lo & ~MTRR_PHYSMASK_RSVD) != (lo & ~MTRR_PHYSMASK_RSVD)
1042 	    || (vr->mask_hi & ~phys_hi_rsvd) != (hi & ~phys_hi_rsvd)) {
1043 		mtrr_wrmsr(MTRRphysMask_MSR(index), vr->mask_lo, vr->mask_hi);
1044 		changed = true;
1045 	}
1046 	return changed;
1047 }
1048 
1049 static u32 deftype_lo, deftype_hi;
1050 
1051 /**
1052  * set_mtrr_state - Set the MTRR state for this CPU.
1053  *
1054  * NOTE: The CPU must already be in a safe state for MTRR changes, including
1055  *       measures that only a single CPU can be active in set_mtrr_state() in
1056  *       order to not be subject to races for usage of deftype_lo. This is
1057  *       accomplished by taking cache_disable_lock.
1058  * RETURNS: 0 if no changes made, else a mask indicating what was changed.
1059  */
1060 static unsigned long set_mtrr_state(void)
1061 {
1062 	unsigned long change_mask = 0;
1063 	unsigned int i;
1064 
1065 	for (i = 0; i < num_var_ranges; i++) {
1066 		if (set_mtrr_var_ranges(i, &mtrr_state.var_ranges[i]))
1067 			change_mask |= MTRR_CHANGE_MASK_VARIABLE;
1068 	}
1069 
1070 	if (mtrr_state.have_fixed && set_fixed_ranges(mtrr_state.fixed_ranges))
1071 		change_mask |= MTRR_CHANGE_MASK_FIXED;
1072 
1073 	/*
1074 	 * Set_mtrr_restore restores the old value of MTRRdefType,
1075 	 * so to set it we fiddle with the saved value:
1076 	 */
1077 	if ((deftype_lo & MTRR_DEF_TYPE_TYPE) != mtrr_state.def_type ||
1078 	    ((deftype_lo & MTRR_DEF_TYPE_ENABLE) >> MTRR_STATE_SHIFT) != mtrr_state.enabled) {
1079 
1080 		deftype_lo = (deftype_lo & MTRR_DEF_TYPE_DISABLE) |
1081 			     mtrr_state.def_type |
1082 			     (mtrr_state.enabled << MTRR_STATE_SHIFT);
1083 		change_mask |= MTRR_CHANGE_MASK_DEFTYPE;
1084 	}
1085 
1086 	return change_mask;
1087 }
1088 
1089 void mtrr_disable(void)
1090 {
1091 	/* Save MTRR state */
1092 	rdmsr(MSR_MTRRdefType, deftype_lo, deftype_hi);
1093 
1094 	/* Disable MTRRs, and set the default type to uncached */
1095 	mtrr_wrmsr(MSR_MTRRdefType, deftype_lo & MTRR_DEF_TYPE_DISABLE, deftype_hi);
1096 }
1097 
1098 void mtrr_enable(void)
1099 {
1100 	/* Intel (P6) standard MTRRs */
1101 	mtrr_wrmsr(MSR_MTRRdefType, deftype_lo, deftype_hi);
1102 }
1103 
1104 void mtrr_generic_set_state(void)
1105 {
1106 	unsigned long mask, count;
1107 
1108 	/* Actually set the state */
1109 	mask = set_mtrr_state();
1110 
1111 	/* Use the atomic bitops to update the global mask */
1112 	for (count = 0; count < sizeof(mask) * 8; ++count) {
1113 		if (mask & 0x01)
1114 			set_bit(count, &smp_changes_mask);
1115 		mask >>= 1;
1116 	}
1117 }
1118 
1119 /**
1120  * generic_set_mtrr - set variable MTRR register on the local CPU.
1121  *
1122  * @reg: The register to set.
1123  * @base: The base address of the region.
1124  * @size: The size of the region. If this is 0 the region is disabled.
1125  * @type: The type of the region.
1126  *
1127  * Returns nothing.
1128  */
1129 static void generic_set_mtrr(unsigned int reg, unsigned long base,
1130 			     unsigned long size, mtrr_type type)
1131 {
1132 	unsigned long flags;
1133 	struct mtrr_var_range *vr;
1134 
1135 	vr = &mtrr_state.var_ranges[reg];
1136 
1137 	local_irq_save(flags);
1138 	cache_disable();
1139 
1140 	if (size == 0) {
1141 		/*
1142 		 * The invalid bit is kept in the mask, so we simply
1143 		 * clear the relevant mask register to disable a range.
1144 		 */
1145 		mtrr_wrmsr(MTRRphysMask_MSR(reg), 0, 0);
1146 		memset(vr, 0, sizeof(struct mtrr_var_range));
1147 	} else {
1148 		vr->base_lo = base << PAGE_SHIFT | type;
1149 		vr->base_hi = (base >> (32 - PAGE_SHIFT)) & ~phys_hi_rsvd;
1150 		vr->mask_lo = -size << PAGE_SHIFT | MTRR_PHYSMASK_V;
1151 		vr->mask_hi = (-size >> (32 - PAGE_SHIFT)) & ~phys_hi_rsvd;
1152 
1153 		mtrr_wrmsr(MTRRphysBase_MSR(reg), vr->base_lo, vr->base_hi);
1154 		mtrr_wrmsr(MTRRphysMask_MSR(reg), vr->mask_lo, vr->mask_hi);
1155 	}
1156 
1157 	cache_enable();
1158 	local_irq_restore(flags);
1159 }
1160 
1161 int generic_validate_add_page(unsigned long base, unsigned long size,
1162 			      unsigned int type)
1163 {
1164 	unsigned long lbase, last;
1165 
1166 	/*
1167 	 * For Intel PPro stepping <= 7
1168 	 * must be 4 MiB aligned and not touch 0x70000000 -> 0x7003FFFF
1169 	 */
1170 	if (mtrr_if == &generic_mtrr_ops && boot_cpu_data.x86 == 6 &&
1171 	    boot_cpu_data.x86_model == 1 &&
1172 	    boot_cpu_data.x86_stepping <= 7) {
1173 		if (base & ((1 << (22 - PAGE_SHIFT)) - 1)) {
1174 			pr_warn("mtrr: base(0x%lx000) is not 4 MiB aligned\n", base);
1175 			return -EINVAL;
1176 		}
1177 		if (!(base + size < 0x70000 || base > 0x7003F) &&
1178 		    (type == MTRR_TYPE_WRCOMB
1179 		     || type == MTRR_TYPE_WRBACK)) {
1180 			pr_warn("mtrr: writable mtrr between 0x70000000 and 0x7003FFFF may hang the CPU.\n");
1181 			return -EINVAL;
1182 		}
1183 	}
1184 
1185 	/*
1186 	 * Check upper bits of base and last are equal and lower bits are 0
1187 	 * for base and 1 for last
1188 	 */
1189 	last = base + size - 1;
1190 	for (lbase = base; !(lbase & 1) && (last & 1);
1191 	     lbase = lbase >> 1, last = last >> 1)
1192 		;
1193 	if (lbase != last) {
1194 		pr_warn("mtrr: base(0x%lx000) is not aligned on a size(0x%lx000) boundary\n", base, size);
1195 		return -EINVAL;
1196 	}
1197 	return 0;
1198 }
1199 
1200 static int generic_have_wrcomb(void)
1201 {
1202 	unsigned long config, dummy;
1203 	rdmsr(MSR_MTRRcap, config, dummy);
1204 	return config & MTRR_CAP_WC;
1205 }
1206 
1207 int positive_have_wrcomb(void)
1208 {
1209 	return 1;
1210 }
1211 
1212 /*
1213  * Generic structure...
1214  */
1215 const struct mtrr_ops generic_mtrr_ops = {
1216 	.get			= generic_get_mtrr,
1217 	.get_free_region	= generic_get_free_region,
1218 	.set			= generic_set_mtrr,
1219 	.validate_add_page	= generic_validate_add_page,
1220 	.have_wrcomb		= generic_have_wrcomb,
1221 };
1222